def clone(self): for unit, attrs in self.reals.items(): for attr in attrs: value = getattr(unit, attr) if self.is_immutable(value): setattr(self, attr, value) continue if not isinstance(value, Array): cloned = getattr(self, attr, None) if cloned is None: setattr(self, attr, deepcopy(value)) continue if isinstance(value, list): del cloned[:] cloned.extend(value) elif isinstance(value, (dict, set)): cloned.clear() cloned.update(value) elif isinstance(value, Bool): cloned <<= value elif isinstance(value, numpy.ndarray): cloned[:] = value else: setattr(self, attr, deepcopy(value)) continue vec = getattr(self, attr, None) if vec is None: vec = Array() self.vectors[value] = vec setattr(self, attr, vec) else: assert isinstance(vec, Array) if not vec and value: vec.reset(value.mem.copy())
class Summator(AcceleratedUnit): """Multiplies two vectors pointwise. """ def __init__(self, workflow, **kwargs): super(Summator, self).__init__(workflow, **kwargs) self.output = Array() self.demand("x", "y") def initialize(self, device, **kwargs): super(Summator, self).initialize(device, **kwargs) if not self.output: self.output.reset(numpy.zeros_like(self.x.mem)) else: assert self.output.shape == self.x.shape self.init_vectors(self.x, self.y, self.output) def init_unpickled(self): super(Summator, self).init_unpickled() self.sources_["summator"] = {} def _gpu_init(self): self.build_program({"OUTPUT_SIZE": self.output.size}, "%s_%d" % (self.__class__.__name__, self.output.size), dtype=self.x.dtype) self.assign_kernel("add_forward") self.set_args(self.x, self.y, self.output) def cuda_init(self): self._gpu_init() block_size = self.device.suggest_block_size(self._kernel_) self._global_size = (int(numpy.ceil(self.output.size / block_size)), 1, 1) self._local_size = (block_size, 1, 1) def ocl_init(self): self._gpu_init() self._global_size = (self.output.size, 1, 1) self._local_size = None def numpy_init(self): pass # nothing to init def _gpu_run(self): self.unmap_vectors(self.x, self.y, self.output) self.execute_kernel(self._global_size, self._local_size) def cuda_run(self): self._gpu_run() def ocl_run(self): self._gpu_run() def numpy_run(self): self.x.map_read() self.y.map_read() self.output.map_invalidate() numpy.add(self.x.mem, self.y.mem, self.output.mem)
class Summator(AcceleratedUnit): """Multiplies two vectors pointwise. """ def __init__(self, workflow, **kwargs): super(Summator, self).__init__(workflow, **kwargs) self.output = Array() self.demand("x", "y") def initialize(self, device, **kwargs): super(Summator, self).initialize(device, **kwargs) if not self.output: self.output.reset(numpy.zeros_like(self.x.mem)) else: assert self.output.shape == self.x.shape self.init_vectors(self.x, self.y, self.output) def init_unpickled(self): super(Summator, self).init_unpickled() self.sources_["summator"] = {} def _gpu_init(self): self.build_program({"OUTPUT_SIZE": self.output.size}, "%s_%d" % (self.__class__.__name__, self.output.size), dtype=self.x.dtype) self.assign_kernel("add_forward") self.set_args(self.x, self.y, self.output) def cuda_init(self): self._gpu_init() block_size = self.device.suggest_block_size(self._kernel_) self._global_size = ( int(numpy.ceil(self.output.size / block_size)), 1, 1) self._local_size = (block_size, 1, 1) def ocl_init(self): self._gpu_init() self._global_size = (self.output.size, 1, 1) self._local_size = None def numpy_init(self): pass # nothing to init def _gpu_run(self): self.unmap_vectors(self.x, self.y, self.output) self.execute_kernel(self._global_size, self._local_size) def cuda_run(self): self._gpu_run() def ocl_run(self): self._gpu_run() def numpy_run(self): self.x.map_read() self.y.map_read() self.output.map_invalidate() numpy.add(self.x.mem, self.y.mem, self.output.mem)
class GDSummator(AcceleratedUnit): """Gradient descent for Multiplier. """ def __init__(self, workflow, **kwargs): super(GDSummator, self).__init__(workflow, **kwargs) self.err_x = Array() self.err_y = Array() self.demand("err_output") def initialize(self, device, **kwargs): super(GDSummator, self).initialize(device, **kwargs) if not self.err_x: self.err_x.reset(numpy.zeros_like(self.err_output.mem)) else: assert self.err_x.shape == self.err_output.shape if not self.err_y: self.err_y.reset(numpy.zeros_like(self.err_output.mem)) else: assert self.err_y.shape == self.err_output.shape self.init_vectors(self.err_x, self.err_y, self.err_output) def cuda_init(self): pass # nothing to init def ocl_init(self): pass # nothing to init def numpy_init(self): pass # nothing to init def cuda_run(self): self.unmap_vectors(self.err_output, self.err_x, self.err_y) self.err_x.devmem.from_device_async(self.err_output.devmem) self.err_y.devmem.from_device_async(self.err_output.devmem) def ocl_run(self): self.unmap_vectors(self.err_output, self.err_x, self.err_y) self.device.queue_.copy_buffer(self.err_output.devmem, self.err_x.devmem, 0, 0, self.err_output.nbytes, need_event=False) self.device.queue_.copy_buffer(self.err_output.devmem, self.err_y.devmem, 0, 0, self.err_output.nbytes, need_event=False) def numpy_run(self): self.err_output.map_read() self.err_x.map_invalidate() self.err_y.map_invalidate() self.err_x.mem[:] = self.err_output.mem[:] self.err_y.mem[:] = self.err_output.mem[:]
class GDSummator(AcceleratedUnit): """Gradient descent for Summator. """ def __init__(self, workflow, **kwargs): super(GDSummator, self).__init__(workflow, **kwargs) self.err_x = Array() self.err_y = Array() self.demand("err_output") def initialize(self, device, **kwargs): super(GDSummator, self).initialize(device, **kwargs) if self.err_x: assert self.err_x.shape[1:] == self.err_output.shape[1:] if not self.err_x or self.err_x.shape[0] != self.err_output.shape[0]: self.err_x.reset(numpy.zeros_like(self.err_output.mem)) if self.err_y: assert self.err_y.shape[1:] == self.err_output.shape[1:] if not self.err_y or self.err_y.shape[0] != self.err_output.shape[0]: self.err_y.reset(numpy.zeros_like(self.err_output.mem)) self.init_vectors(self.err_x, self.err_y, self.err_output) def cuda_init(self): pass # nothing to init def ocl_init(self): pass # nothing to init def numpy_init(self): pass # nothing to init def cuda_run(self): self.unmap_vectors(self.err_output, self.err_x, self.err_y) self.err_x.devmem.from_device_async(self.err_output.devmem) self.err_y.devmem.from_device_async(self.err_output.devmem) def ocl_run(self): self.unmap_vectors(self.err_output, self.err_x, self.err_y) self.device.queue_.copy_buffer( self.err_output.devmem, self.err_x.devmem, 0, 0, self.err_output.nbytes, need_event=False) self.device.queue_.copy_buffer( self.err_output.devmem, self.err_y.devmem, 0, 0, self.err_output.nbytes, need_event=False) def numpy_run(self): self.err_output.map_read() self.err_x.map_invalidate() self.err_y.map_invalidate() self.err_x.mem[:] = self.err_output.mem[:] self.err_y.mem[:] = self.err_output.mem[:]
class MemCpy(AcceleratedUnit): def __init__(self, workflow, **kwargs): super(MemCpy, self).__init__(workflow, **kwargs) self.output = Array() self.demand("input") def initialize(self, device, **kwargs): super(MemCpy, self).initialize(device, **kwargs) if (self.output.mem is None or self.output.mem.size != self.input.mem.size): self.output.reset() self.output.mem = numpy.zeros(self.input.mem.shape, dtype=self.input.mem.dtype) self.input.initialize(self.device) self.output.initialize(self.device) def cuda_init(self): pass def ocl_init(self): pass def _gpu_run(self): self.input.unmap() self.output.unmap() def ocl_run(self): self._gpu_run() self.device.queue_.copy_buffer(self.input.devmem, self.output.devmem, 0, 0, self.input.nbytes) def cuda_run(self): self._gpu_run() self.output.devmem.from_device_async(self.input.devmem) def numpy_run(self): self.input.map_read() self.output.map_invalidate() numpy.copyto(self.output.mem, self.input.mem)
class Cutter1D(AcceleratedUnit): """Cuts the specified interval from each 1D sample of input batch into output. y = alpha * x + beta * y """ def __init__(self, workflow, **kwargs): super(Cutter1D, self).__init__(workflow, **kwargs) self.alpha = kwargs.get("alpha") self.beta = kwargs.get("beta") self.output_offset = kwargs.get("output_offset", 0) self.output = Array() self.demand("alpha", "beta", "input") # TODO: add input_offset and length to demand and not to crash lstm # TODO: unit test def init_unpickled(self): super(Cutter1D, self).init_unpickled() self.sources_["cutter"] = {} def initialize(self, device, **kwargs): super(Cutter1D, self).initialize(device, **kwargs) if not self.output or self.output.shape[0] != self.input.shape[0]: self.output.reset( numpy.zeros( (self.input.shape[0], self.output_offset + self.length), dtype=self.input.dtype)) else: assert self.output.sample_size >= self.output_offset + self.length self.init_vectors(self.input, self.output) def cuda_init(self): dtype = self.input.dtype itemsize = self.input.itemsize limit = self.input.shape[0] * self.length self.build_program({}, "%s" % self.__class__.__name__, dtype=dtype) self.assign_kernel("cutter_1d_forward") self.set_args( int(self.input.devmem) + self.input_offset * itemsize, numpy.array([self.alpha], dtype=dtype), numpy.array([self.input.sample_size], dtype=numpy.int32), int(self.output.devmem) + self.output_offset * itemsize, numpy.array([self.beta], dtype=dtype), numpy.array([self.output.sample_size], dtype=numpy.int32), numpy.array([self.length], dtype=numpy.int32), numpy.array([limit], dtype=numpy.int32)) block_size = self.device.suggest_block_size(self._kernel_) self._global_size = (int(numpy.ceil(limit / block_size)), 1, 1) self._local_size = (block_size, 1, 1) def ocl_init(self): dtype = self.input.dtype self.build_program({}, "%s" % self.__class__.__name__, dtype=dtype) self.assign_kernel("cutter_1d_forward") self.set_args( self.input.devmem, numpy.array([self.input_offset], dtype=numpy.int32), numpy.array([self.alpha], dtype=dtype), numpy.array([self.input.sample_size], dtype=numpy.int32), self.output.devmem, numpy.array([self.output_offset], dtype=numpy.int32), numpy.array([self.beta], dtype=dtype), numpy.array([self.output.sample_size], dtype=numpy.int32)) self._global_size = (self.input.shape[0], self.length) self._local_size = None def _gpu_run(self): self.unmap_vectors(self.input, self.output) self.execute_kernel(self._global_size, self._local_size) def cuda_run(self): return self._gpu_run() def ocl_run(self): return self._gpu_run() def numpy_run(self): self.input.map_read() self.output.map_write() out = self.output.matrix[ :, self.output_offset:self.output_offset + self.length] if self.beta: out *= self.beta else: out[:] = 0 out += ( self.input.matrix[ :, self.input_offset:self.input_offset + self.length] * self.alpha)
class ImageLoader(LoaderWithValidationRatio): """Base class for all image loaders. It is generally used for loading large datasets. Attributes: color_space: the color space to which to convert images. Can be any of the values supported by OpenCV, e.g., GRAY or HSV. source_dtype: dtype to work with during various image operations. shape: image shape (tuple) - set after initialize(). Must be overriden in child classes: get_image_label() get_image_info() get_image_data() get_keys() """ def __init__(self, workflow, **kwargs): super(ImageLoader, self).__init__(workflow, **kwargs) self.color_space = kwargs.get("color_space", "RGB") self._source_dtype = numpy.float32 self._original_shape = tuple() self.class_keys = [[], [], []] self.verify_interface(IImageLoader) self.path_to_mean = kwargs.get("path_to_mean", None) self.add_sobel = kwargs.get("add_sobel", False) self.mirror = kwargs.get("mirror", False) # True, False, "random" self.scale = kwargs.get("scale", 1.0) self.scale_maintain_aspect_ratio = kwargs.get( "scale_maintain_aspect_ratio", True) self.rotations = kwargs.get("rotations", (0.0, )) # radians self.crop = kwargs.get("crop", None) self.crop_number = kwargs.get("crop_number", 1) self._background = None self.background_image = kwargs.get("background_image", None) self.background_color = kwargs.get("background_color", (0xff, 0x14, 0x93)) self.smart_crop = kwargs.get("smart_crop", True) self.minibatch_label_values = Array() @property def source_dtype(self): return self._source_dtype @property def color_space(self): return self._color_space @color_space.setter def color_space(self, value): self._validate_color_space(value) self._color_space = value @Loader.shape.getter def shape(self): """ :return: Final cropped image shape. """ if self.crop is not None: shape = self.crop else: shape = self.uncropped_shape if self.channels_number > 1: shape += (self.channels_number, ) return shape @property def uncropped_shape(self): """ :return: Uncropped (but scaled) image shape. """ if not isinstance(self.scale, tuple): if self._original_shape == tuple(): return tuple() return self._scale_shape(self._original_shape)[:2] else: return self.scale @property def original_shape(self): return self._original_shape @original_shape.setter def original_shape(self, value): if value is None: raise ValueError("shape must not be None") if not isinstance(value, tuple): raise TypeError("shape must be a tuple (got %s)" % (value, )) if len(value) not in (2, 3): raise ValueError("len(shape) must be equal to 2 or 3 (got %s)" % (value, )) for i, d in enumerate(value): if not isinstance(d, int): raise TypeError("shape[%d] is not an integer (= %s)" % (i, d)) if d < 1: raise ValueError("shape[%d] < 1 (= %s)" % (i, d)) self._original_shape = value @property def scale(self): return self._scale @scale.setter def scale(self, value): if not isinstance(value, (float, tuple)): raise TypeError("scale must be either float or tuple of two ints" " (got %s of type %s)" % (value, value.__class__)) if isinstance(value, tuple): if len(value) != 2: raise ValueError("scale must have length 2 (not %d in %s)" % (len(value), value)) if not isinstance(value[0], int) or not isinstance(value[1], int): raise ValueError("scale must consist of integers (got %s)" % value) self._scale = value @property def crop(self): return self._crop @crop.setter def crop(self, value): if value is None: self._crop = None return if not isinstance(value, tuple): raise TypeError( "crop must be a tuple of 2 integers or floats (got %s)" % value) if len(value) != 2: raise ValueError("invalid crop length (got %d for %s), must be 2" % (len(value), value)) for i, val in enumerate(value): if not isinstance(val, (int, float)): raise TypeError( "crop[%d] = %s is neither an integer nor a float" % (i, val[i])) if isinstance(val, int) and val < 1: raise ValueError("crop[%d] = %s is out of range" % (i, val)) if isinstance(val, float): if val <= 0 or val > 1: raise ValueError("Out of range crop %s: %s" % (("height", "width")[i], val)) self._crop = value @property def crop_number(self): return self._crop_number @crop_number.setter def crop_number(self, value): if not isinstance(value, int): raise TypeError("crop_number must be an integer (got %s)" % value) if value < 1: raise ValueError("crop_number must be greater than zero (got %d)" % value) if value > 1 and self.crop is None: raise ValueError( "crop parameter is None, refusing to set crop_number") self._crop_number = value @property def smart_crop(self): """ :return: Value indicating whether to crop only around bboxes. """ return self._smart_crop @smart_crop.setter def smart_crop(self, value): if not isinstance(value, bool): raise TypeError("smart_crop must be a boolean value") self._smart_crop = value @property def mirror(self): return self._mirror @mirror.setter def mirror(self, value): if value not in (False, True, "random"): raise ValueError( "mirror must be any of the following: False, True, \"random\"") self._mirror = value @property def rotations(self): return self._rotations @rotations.setter def rotations(self, value): if not isinstance(value, tuple): raise TypeError("rotations must be a tuple (got %s)" % value) for i, rot in enumerate(value): if not isinstance(rot, float): raise TypeError("rotations[%d] = %s is not a float" % (i, rot)) if rot >= numpy.pi * 2: raise ValueError("rotations[%d] = %s is greater than 2π" % (i, rot)) self._rotations = tuple(sorted(value)) @property def samples_inflation(self): return (1 if self.mirror is not True else 2) * len(self.rotations) * \ self.crop_number @property def background_image(self): return self._background_image @background_image.setter def background_image(self, value): if isinstance(value, str): with open(value, "rb") as fin: self.background_image = fin elif hasattr(value, "read") and hasattr(value, "seek"): self.background_image = numpy.array(Image.open(value)) elif isinstance(value, numpy.ndarray): if value.shape != self.shape: raise error.BadFormatError( "background_image's shape %s != sample's shape " "%s" % (value.shape, self.shape)) self._background_image = value if getattr(self, "background_color", None) is not None: self.warning( "background_color = %s is ignored in favor of " "background_image", self.background_color) elif value is None: self._background_image = None else: raise ValueError("background_image must be any of the following: " "file name, file object, numpy array or None") @property def background_color(self): return self._background_color @background_color.setter def background_color(self, value): if value is None: self._background_color = None return if not isinstance(value, tuple): raise TypeError("background_color must be a tuple (got %s)" % value) if len(value) != self.channels_number: raise ValueError( "background_color must have the same length as the number of " "channels = %d (got length %d for %s)" % (self.channels_number, len(value), value)) for i, col in enumerate(value): if not isinstance(col, int): raise TypeError("background_color[%d] = %s is not an integer" % (i, col)) if getattr(self, "background_image", None) is not None: self.warning( "background_color = %s is ignored in favor of " "background_image", value) self._background_color = value @property def background(self): if self._background is None: if self.background_image is not None: self._background = self.background_image else: self._background = numpy.zeros(self.shape) self._background[:] = self.background_color return self._background.copy() @property def channels_number(self): channels = COLOR_CHANNELS_MAP[self.color_space] if self.add_sobel: channels += 1 return channels def get_effective_image_info(self, key): info = self.get_image_info(key) if self.scale == 1.0: return info if isinstance(self.scale, tuple): return self.scale, info[1] else: return self._scale_shape(info[0]), info[1] def get_image_bbox(self, key, size): """ Override this method for custom label <-> bbox mapping. :param key: The image key. :param size: The image size (for optimization purposes). :return: (ymin, ymax, xmin, xmax). """ return 0, size[0], 0, size[1] def preprocess_image(self, data, color, crop, bbox): """ Transforms images before serving. :param data: the loaded image data. :param color: The loaded image color space. :param crop: True if must crop the scaled image; otherwise, False. :param bbox: The bounding box of the labeled object. Tuple (ymin, ymax, xmin, xmax). :return: The transformed image data, the label value (from 0 to 1). """ if color != self.color_space: method = getattr(cv2, "COLOR_%s2%s" % (color, self.color_space), None) if method is None: aux_method = getattr(cv2, "COLOR_%s2BGR" % color) try: data = cv2.cvtColor(data, aux_method) except cv2.error as e: self.error("Failed to perform '%s' conversion", aux_method) raise from_none(e) method = getattr(cv2, "COLOR_BGR2%s" % self.color_space) try: data = cv2.cvtColor(data, method) except cv2.error as e: self.error("Failed to perform '%s' conversion", method) raise from_none(e) if self.add_sobel: data = self.add_sobel_channel(data) if self.scale != 1.0: data, bbox = self.scale_image(data, bbox) if crop and self.crop is not None: data, label_value = self.crop_image(data, bbox) else: label_value = 1 return data, label_value, bbox def scale_image(self, data, bbox): bbox = numpy.array(bbox, float) if self.scale_maintain_aspect_ratio: if data.shape[1] >= data.shape[0]: dst_width = self.uncropped_shape[:2][1] dst_height = int( numpy.round( float(dst_width) * data.shape[0] / data.shape[1])) else: dst_height = self.uncropped_shape[:2][0] dst_width = int( numpy.round( float(dst_height) * data.shape[1] / data.shape[0])) dst_x_min = int( numpy.round(0.5 * (self.uncropped_shape[:2][1] - dst_width))) dst_y_min = int( numpy.round(0.5 * (self.uncropped_shape[:2][0] - dst_height))) data = cv2.resize(data, (dst_width, dst_height), interpolation=cv2.INTER_CUBIC) dst_x_max = dst_x_min + data.shape[1] dst_y_max = dst_y_min + data.shape[0] sample = self.background sample[dst_y_min:dst_y_max, dst_x_min:dst_x_max] = data data = sample.copy() bbox[:2] *= (dst_y_max - dst_y_min) / (bbox[1] - bbox[0]) bbox[:2] += dst_y_min bbox[2:] *= (dst_x_max - dst_x_min) / (bbox[3] - bbox[2]) bbox[2:] += dst_x_min else: data = cv2.resize(data, tuple(reversed(self.uncropped_shape[:2])), interpolation=cv2.INTER_CUBIC) bbox[:2] *= self.uncropped_shape[0] / (bbox[1] - bbox[0]) bbox[2:] *= self.uncropped_shape[1] / (bbox[3] - bbox[2]) return data, tuple(bbox.astype(numpy.int32)) def add_sobel_channel(self, data): original_data = data if self.channels_number == 1 + 1: original_data = original_data.reshape(original_data.shape[:2] + (1, )) elif self.color_space in ("RGB", "BGR", "RGBA", "BGRA"): data = cv2.cvtColor( data, getattr(cv2, "COLOR_%s2GRAY" % self.color_space)) elif self.color_space == "HSV": data = data[:, :, 2] elif self.color_space == "YCR_CB": data = data[:, :, 0] else: raise NotImplementedError( "Conversion from %s to GRAY is not ready" % self.color_space) sobel_xy = tuple( cv2.Sobel(data, cv2.CV_32F, *d, ksize=3) for d in ((1, 0), (0, 1))) sobel_data = numpy.zeros(shape=data.shape + (original_data.shape[2] + 1, ), dtype=original_data.dtype) sobel_data[:, :, -1] = numpy.linalg.norm(sobel_xy) sobel_data[:, :, :-1] = original_data return sobel_data def crop_image(self, data, bbox): """ Cuts a rectangular part of an image. :param data: The source image to crop. :param bbox: (ymin, ymax, xmin, xmax) :return: tuple (image part randomly cropped around the bbox,\ intersection ratio) """ crop_hw_yx = [[0, 0], [0, 0]] for i in 0, 1: crop_hw_yx[0][i] = self.crop[i] if isinstance(self.crop[i], int) \ else int(self.crop[i] * data.shape[i]) crop_size = crop_hw_yx[0][i] crop_hw_yx[1][i] = self.prng.randint( max(bbox[i * 2] - crop_size, 0), min(data.shape[i] - crop_size + 1, bbox[i * 2 + 1] + crop_size)) crop_first = crop_hw_yx[1] crop_last = tuple(crop_hw_yx[1][i] + crop_hw_yx[0][i] for i in (0, 1)) crop_bbox = crop_first[0], crop_last[0], crop_first[1], crop_last[1] return data[crop_bbox[0]:crop_bbox[1], crop_bbox[2]:crop_bbox[3]], \ self._intersection(bbox, crop_bbox) def distort(self, data, mirror, rot): if mirror: data = cv2.flip(data, 1) data = numpy.resize(data, data.shape[:2] + (data.shape[-1] + 1, )) data[:, :, -1] = 1 center = tuple(reversed(tuple(data.shape[i] // 2 for i in (0, 1)))) rot_matrix = cv2.getRotationMatrix2D(center, rot * 180 / numpy.pi, 1.0) data = cv2.warpAffine(data, rot_matrix, tuple(reversed(data.shape[:2]))) real = data[:, :, :-1] imag = data[:, :, -1] real *= imag[..., None] real += self.background * (1 - imag)[..., None] return real def get_distortion_by_index(self, index): index //= self.crop_number if self.mirror is True: return index % 2 == 1, self.rotations[index // 2] elif self.mirror == "random": mirror = bool(self.prng.randint(2)) else: mirror = False return mirror, self.rotations[index] def load_keys(self, keys, pbar, data, labels, label_values, crop=True): """Loads data from the specified keys. """ index = 0 has_labels = False for key in keys: obj, label_value, _ = self._load_image(key) label, has_labels = self._load_label(key, has_labels) if (self.crop is None or not crop) and \ obj.shape[:2] != self.uncropped_shape: self.warning("Ignored %s (label %s): shape %s", key, label, obj.shape[:2]) continue if data is not None: data[index] = obj if labels is not None: labels[index] = label if label_values is not None: label_values[index] = label_value index += 1 if pbar is not None: pbar.inc() return has_labels def load_labels(self): if not self.has_labels: return self.info("Reading labels...") different_labels = defaultdict(int), defaultdict(int), defaultdict(int) label_key_map = defaultdict(list), defaultdict(list), defaultdict(list) pb = ProgressBar(maxval=self.total_samples, term_width=40) pb.start() for class_index in range(3): for key in self.class_keys[class_index]: label, has_labels = self._load_label(key, True) assert has_labels different_labels[class_index][label] += 1 label_key_map[class_index][label].append(key) self._samples_mapping[label].add(key) pb.inc() pb.finish() return different_labels, label_key_map def initialize(self, **kwargs): self._restored_from_pickle_ = kwargs["snapshot"] super(ImageLoader, self).initialize(**kwargs) del self._restored_from_pickle_ def load_data(self): try: super(ImageLoader, self).load_data() except AttributeError: pass if self._restored_from_pickle_: self.info("Scanning for changes...") progress = ProgressBar(maxval=self.total_samples, term_width=40) progress.start() for keys in self.class_keys: for key in keys: progress.inc() size, _ = self.get_effective_image_info(key) if size != self.uncropped_shape: raise error.BadFormatError( "%s changed the effective size (now %s, was %s)" % (key, size, self.uncropped_shape)) progress.finish() return for keys in self.class_keys: del keys[:] for index, class_name in enumerate(CLASS_NAME): keys = set(self.get_keys(index)) self.class_keys[index].extend(keys) self.class_lengths[index] = len(keys) * self.samples_inflation self.class_keys[index].sort() if self.uncropped_shape == tuple(): raise error.BadFormatError( "original_shape was not initialized in get_keys()") self.info( "Found %d samples of shape %s (%d TEST, %d VALIDATION, %d TRAIN)", self.total_samples, self.shape, *self.class_lengths) # Perform a quick (unreliable) test to determine if we have labels keys = next(k for k in self.class_keys if len(k) > 0) self._has_labels = self.load_keys( (keys[RandomGenerator(None).randint(len(keys))], ), None, None, None, None) self._resize_validation_keys(self.load_labels()) def create_minibatch_data(self): self.minibatch_data.reset( numpy.zeros((self.max_minibatch_size, ) + self.shape, dtype=self.dtype)) self.minibatch_label_values.reset( numpy.zeros(self.max_minibatch_size, numpy.float32)) def keys_from_indices(self, indices): for index in indices: class_index, origin_index, _ = \ self._get_class_origin_distortion_from_index(index) yield self.class_keys[class_index][origin_index] def fill_minibatch(self): indices = self.minibatch_indices.mem[:self.minibatch_size] assert self.has_labels == self.load_keys( self.keys_from_indices(indices), None, self.minibatch_data.mem, self.raw_minibatch_labels, self.minibatch_label_values) if self.samples_inflation == 1: return for pos, index in enumerate(indices): _, _, dist_index = \ self._get_class_origin_distortion_from_index(index) self.minibatch_data[pos] = self.distort( self.minibatch_data[pos], *self.get_distortion_by_index(dist_index)) def _resize_validation_keys(self, label_analysis): if label_analysis is None: return different_labels, label_key_map = label_analysis if self.validation_ratio is None: self._setup_labels_mapping(different_labels) return if self.validation_ratio < 0: self.class_keys[TRAIN] += self.class_keys[VALID] self.class_lengths[TRAIN] += self.class_lengths[VALID] del self.class_keys[VALID][:] self.class_lengths[VALID] = 0 merged = { k: (different_labels[VALID][k] + different_labels)[TRAIN][k] for k in label_key_map[TRAIN] } self._setup_labels_mapping((different_labels[TEST], {}, merged)) return overall = sum(len(ck) for ck in self.class_keys[VALID:]) target_validation_length = int(overall * self.validation_ratio) if not self.has_labels: keys = list(chain.from_iterable(self.class_keys[VALID:])) keys.sort() self.prng.shuffle(keys) del self.class_keys[VALID][:] self.class_keys[VALID].extend(keys[:target_validation_length]) del self.class_keys[TRAIN][:] self.class_keys[TRAIN].extend(keys[target_validation_length:]) self._finalize_resizing_validation(different_labels, label_key_map) return # We must ensure that each set has the same labels # The first step is to pick two keys for each label and distribute them # into VALID and TRAIN evenly if len(label_key_map[TRAIN]) > target_validation_length: raise LoaderError( "Unable to set the new size of the validation set to %d (%.3f)" " since the number of labels is %d" % (target_validation_length * self.samples_inflation, self.validation_ratio, len(label_key_map[TRAIN]))) if overall - target_validation_length < len(label_key_map[TRAIN]): raise LoaderError( "Unable to set the new size of the training set to %d (%.3f) " "since the number of labels is %d" % ((overall - target_validation_length) * self.samples_inflation, 1.0 - self.validation_ratio, len(label_key_map[TRAIN]))) vt_label_key_map = { l: (label_key_map[VALID].get(l, []) + label_key_map[TRAIN].get(l, [])) for l in label_key_map[TRAIN] } for i in VALID, TRAIN: del self.class_keys[i][:] for label, keys in sorted(vt_label_key_map.items()): if len(keys) < 2: raise LoaderError("Label %s has less than 2 keys" % label) choice = self.prng.choice(len(keys), 2, replace=False) assert choice[0] != choice[1] for i in VALID, TRAIN: self.class_keys[i].append(keys[choice[i - 1]]) for c in sorted(choice, reverse=True): del keys[c] # Distribute the left keys randomly left_keys = list(sorted(chain.from_iterable( vt_label_key_map.values()))) self.prng.shuffle(left_keys) offset_val_length = \ target_validation_length - len(vt_label_key_map) self.class_keys[VALID].extend(left_keys[:offset_val_length]) self.class_keys[TRAIN].extend(left_keys[offset_val_length:]) self._finalize_resizing_validation(different_labels, label_key_map) def _finalize_resizing_validation(self, different_labels, label_key_map): for ck in self.class_keys[VALID:]: ck.sort() for i in VALID, TRAIN: self.class_lengths[i] = len(self.class_keys[i]) * \ self.samples_inflation new_diff = defaultdict(int), defaultdict(int) key_label_map = {} for ci in VALID, TRAIN: key_label_map.update( {k: l for l, keys in label_key_map[ci].items() for k in keys}) for ci in VALID, TRAIN: for key in self.class_keys[ci]: new_diff[ci - 1][key_label_map[key]] += 1 self._setup_labels_mapping((different_labels[TEST], ) + new_diff) def _get_class_origin_distortion_from_index(self, index): class_index, key_remainder = self.class_index_by_sample_index(index) key_index = self.class_lengths[class_index] - key_remainder return (class_index, ) + divmod(key_index, self.samples_inflation) def _load_image(self, key, crop=True): """Returns the data to serve corresponding to the given image key and the label value (from 0 to 1). """ data = self.get_image_data(key) size, color = self.get_image_info(key) bbox = self.get_image_bbox(key, size) return self.preprocess_image(data, color, crop, bbox) def _load_label(self, key, has_labels): label = self.get_image_label(key) if label is not None: has_labels = True if has_labels and label is None: raise error.BadFormatError( "%s does not have a label, but others do" % key) return label, has_labels def _intersection(self, bbox_a, bbox_b): ymin_a, ymax_a, xmin_a, xmax_a = bbox_a ymin_b, ymax_b, xmin_b, xmax_b = bbox_b x_intersection = min(xmax_a, xmax_b) - max(xmin_a, xmin_b) y_intersection = min(ymax_a, ymax_b) - max(ymin_a, ymin_b) if int(x_intersection) | int(y_intersection) <= 0: return 0 else: return x_intersection * y_intersection def _scale_shape(self, shape): return tuple(int(shape[i] * self.scale) for i in (0, 1)) + shape[2:] def _validate_color_space(self, value): if not isinstance(value, str): raise TypeError("db_colorpsace must be a string (got %s)" % type(value)) if value != "RGB" and not hasattr(cv2, "COLOR_%s2RGB" % value): raise ValueError("Unsupported color space: %s" % value)
class EvaluatorBase(AcceleratedUnit, TriviallyDistributable): hide_from_registry = True """Base class for evaluators. """ def __init__(self, workflow, **kwargs): kwargs["view_group"] = kwargs.get("view_group", "EVALUATOR") super(EvaluatorBase, self).__init__(workflow, **kwargs) self.mean = kwargs.get("mean", True) self.err_output = Array() self._merged_output = Array() self.krn_constants_i_ = None self.krn_constants_f_ = None self.demand("output", "batch_size") if self.testing: self.demand("class_lengths", "offset") @property def mean(self): """ :return: True if the error function averages values. Default is True. """ return self._mean @mean.setter def mean(self, value): if not isinstance(value, bool): raise TypeError("mean must be boolean (got %s)" % type(value)) self._mean = value @property def merged_output(self): assert self.testing return self._merged_output.mem def initialize(self, device, **kwargs): super(EvaluatorBase, self).initialize(device, **kwargs) dtype = self.output.dtype if self.testing: self._merged_output.reset(numpy.zeros( (self.class_lengths[TEST],) + self.output.shape[1:], dtype)) return self.krn_constants_i_ = numpy.zeros(1, numpy.int32) self.krn_constants_f_ = numpy.zeros(1, dtype) self.err_output.reset(numpy.zeros_like(self.output.mem, dtype)) for vec in self.output, self.err_output: vec.initialize(self.device) def run(self): if self.testing: self.output.map_read() self.merge_output() return return super(EvaluatorBase, self).run() def merge_output(self): self.merged_output[self.offset - self.batch_size:self.offset] = \ self.output[:self.batch_size] def get_metric_names(self): if self.testing: return {"Output"} return set() def get_metric_values(self): if self.testing: return {"Output": self.merged_output} return {}
class EvaluatorSoftmax(EvaluatorBase): MAPPING = "evaluator_softmax" LOSS = "softmax" """Evaluator for nn softmax output from the batch labels. Must be assigned before initialize(): output labels batch_size max_idx Updates after run(): err_output n_err confusion_matrix max_err_output_sum Creates within initialize(): err_output n_err confusion_matrix max_err_output_sum Attributes: labels: labels for Batch. output: output of the network_common as Batch. err_output: backpropagation errors based on labels. batch_size: number of elements in output to evaluate. confusion_matrix: confusion matrix for the output. compute_confusion_matrix: compute confusion matrix or not. max_idx: indexes of element with maximum real value for each sample. max_err_output_sum: maximum of backpropagated error sum by sample. """ def __init__(self, workflow, **kwargs): super(EvaluatorSoftmax, self).__init__(workflow, **kwargs) self.compute_confusion_matrix = kwargs.get( "compute_confusion_matrix", True) self.confusion_matrix = Array() self.n_err = Array() self.max_err_output_sum = Array() self.class_keys = None self.demand("labels", "max_idx") if self.testing: self.demand("labels_mapping") def initialize(self, device, **kwargs): super(EvaluatorSoftmax, self).initialize(device=device, **kwargs) if self.testing: return self.sources_["evaluator"] = {} dtype = self.output.dtype if not self.n_err: self.n_err.reset(numpy.zeros(2, dtype=numpy.int32)) else: assert self.n_err.size == 2 out_size = self.output.sample_size if self.compute_confusion_matrix: if not self.confusion_matrix: self.confusion_matrix.reset( numpy.zeros([out_size, out_size], numpy.int32)) else: assert self.confusion_matrix.size == out_size * out_size else: self.confusion_matrix.reset() if not self.max_err_output_sum: self.max_err_output_sum.reset(numpy.zeros(1, dtype)) else: assert self.max_err_output_sum.size == 1 self.init_vectors(self.confusion_matrix, self.n_err, self.max_idx, self.labels, self.max_err_output_sum) def _gpu_init(self): dtype = self.output.dtype block_size = min(self.err_output.shape[0], 256) self.build_program( cache_file_name="%s_%d_%d" % (self.__class__.__name__, self.output.shape[0], self.output.sample_size), dtype=dtype, block_size=block_size, max_batch_size=self.err_output.shape[0], output_size=self.err_output.sample_size) self.assign_kernel("evaluate_softmax") self.set_args(self.output, self.max_idx, self.labels, self.skip_args(2), self.n_err, self.confusion_matrix, self.max_err_output_sum, self.err_output) return block_size def ocl_init(self): if self.testing: return block_size = self._gpu_init() self._global_size = [block_size] self._local_size = [block_size] def cuda_init(self): if self.testing: return block_size = self._gpu_init() self._global_size = (1, 1, 1) self._local_size = (block_size, 1, 1) def _gpu_run(self): self.unmap_vectors( self.err_output, self.output, self.max_idx, self.labels, self.n_err, self.confusion_matrix, self.max_err_output_sum) self.krn_constants_i_[0] = self.batch_size self.set_arg(3, self.krn_constants_i_[0:1]) self.krn_constants_f_[0] = 1.0 / self.batch_size if self.mean else 1.0 self.set_arg(4, self.krn_constants_f_[0:1]) self.execute_kernel(self._global_size, self._local_size) def ocl_run(self): return self._gpu_run() def cuda_run(self): return self._gpu_run() def numpy_run(self): self.err_output.map_invalidate() for vec in self.output, self.max_idx, self.labels: vec.map_read() for vec in self.n_err, self.confusion_matrix, self.max_err_output_sum: vec.map_write() batch_size = self.batch_size labels = self.labels.mem confusion_matrix = self.confusion_matrix.mem n_ok = 0 n_total = 0 multiplier = 1.0 / batch_size if self.mean else 1.0 for i in range(batch_size): # loop by batch if labels[i] < 0: self.err_output.mem[i] = 0.0 continue output = ravel(self.output[i]) err_output = ravel(self.err_output[i]) max_idx = self.max_idx[i] confusion_matrix[max_idx, labels[i]] += 1 if max_idx == labels[i]: n_ok += 1 n_total += 1 # Compute softmax output error gradient err_output[:] = output[:] err_output[labels[i]] -= 1.0 err_output *= multiplier if err_output.dtype in (numpy.complex64, numpy.complex128): self.max_err_output_sum[0] = max( self.max_err_output_sum[0], numpy.linalg.norm(err_output)) else: self.max_err_output_sum[0] = max( self.max_err_output_sum[0], (numpy.fabs(err_output)).sum()) # Set errors for excessive samples to zero if batch_size < self.err_output.mem.shape[0]: self.err_output.mem[batch_size:] = 0.0 self.n_err[0] += batch_size - n_ok self.n_err[1] += n_total def get_metric_values(self): if self.testing: output_labels = {} class_keys = getattr(self, "class_keys", None) for index, labels in enumerate(self.merged_output[:]): max_value = 0 for label_index, value in enumerate(labels): if value >= max_value: max_value = value max_index = label_index if class_keys is not None: output_labels[self.class_keys[TEST][ index]] = self.labels_mapping[max_index] else: output_labels[index] = self.labels_mapping[max_index] return {"Output": output_labels} return {}
class Forward(ForwardBase): """Class for forward propagation units. Attributes: input: input layer values. output: output layer values. weights: weights. bias: bias. weights_stddev: magnitude of the random distribution for weights. bias_stddev: magnitude of the random distribution for bias. rand: prng.Rand() object for initial weights generation. """ hide_from_registry = True MAPPING = set() def __init__(self, workflow, **kwargs): kwargs["view_group"] = kwargs.get("view_group", "WORKER") super(Forward, self).__init__(workflow, **kwargs) self.weights_stddev = kwargs.get("weights_stddev") self.bias_stddev = kwargs.get("bias_stddev", self.weights_stddev) self.weights_filling = kwargs.get("weights_filling", "uniform") self.bias_filling = kwargs.get("bias_filling", "uniform") self.rand = kwargs.get("rand", prng.get()) self.weights_transposed = kwargs.get("weights_transposed", False) self.include_bias = kwargs.get("include_bias", True) self.demand("input") self.output = Array(shallow_pickle=True) self.weights = Array() self.bias = Array() self.forward_mode = False self.exports = [ "weights", "bias", "include_bias", "weights_transposed" ] def package_export(self): data = {} for attr in self.exports: value = getattr(self, attr) if value is not None: if isinstance(value, Array): value.map_read() value = value.mem data[attr] = value return data @property def forward_mode(self): return self._forward_mode @forward_mode.setter def forward_mode(self, value): if not isinstance(value, bool): raise TypeError("forward_mode must be boolean (got %s)" % type(value)) self._forward_mode = value def initialize(self, device, **kwargs): self.forward_mode = kwargs.get("forward_mode", False) super(Forward, self).initialize(device=device, **kwargs) def generate_data_for_slave(self, slave): if self.forward_mode: return None data = [None, None] if self.weights: self.weights.map_read() data[0] = self.weights.mem if self.bias: self.bias.map_read() data[1] = self.bias.mem return data def generate_data_for_master(self): return None def apply_data_from_master(self, data): if self.forward_mode: return if self.weights: self.weights.map_invalidate() numpy.copyto(self.weights.mem, data[0]) else: self.weights.reset(data[0]) if self.bias: self.bias.map_invalidate() numpy.copyto(self.bias.mem, data[1]) else: self.bias.reset(data[1]) def apply_data_from_slave(self, data, slave): pass def drop_slave(self, slave): pass
class Binarization(AcceleratedUnit, EmptyDeviceMethodsMixin): """ Input Binarization. Input and output is 2d arrays of the same size. Each element A(i,j) (in row i and column j) of input is a float number between 0 and 1. Each element B(i,j) of output is equal 1 with probability A(i,j) and 0 with 1 - A(i,j). Must be assigned before initialize(): * input Updates after run(): * output Creates within initialize(): * output Attributes: input: input as batch of samples. output: output as batch of samples. """ def __init__(self, workflow, **kwargs): super(Binarization, self).__init__(workflow, **kwargs) self.output = Array() self.rand = kwargs.get("rand", prng.get()) self.demand("input", "batch_size") def run(self): """Batch binarization on CPU only. """ self.output.map_invalidate() self.input.map_read() self.output.mem[:] = self.input.mem[:] self.output.mem[:self.batch_size, :] = self.matlab_binornd( 1, self.input.mem[:self.batch_size, :]) def initialize(self, device, **kwargs): super(Binarization, self).initialize(device=device, **kwargs) if not self.output or self.output.size != self.input.size: self.output.reset() self.output.mem = numpy.zeros_like(self.input.mem) self.output.initialize(self.device) def matlab_binornd(self, n, p_in): """ Analogue binornd in Matlab, but n must be scalar. The function generates a matrix of random variables, where the element at (i,j) position is generated from binomial distribution with the number of trials n and the probability of success p_in(i,j). Args: n (int): number of trials p_in (2 dimension numpy.array): success probability matrix Returns: res (2 dimension numpy.array): matrix of random variables generated from the binomial distribution """ p = numpy.copy(p_in) if len(p.shape) == 2: nrow = p.shape[0] ncol = p.shape[1] p = numpy.transpose(p) p = p.flatten() dim = p.shape[0] p = matlib.repmat(p, n, 1) f = self.rand.rand(n, dim) res = f < p res = numpy.sum(res, axis=0) res = numpy.transpose(res.reshape(ncol, nrow)).reshape(nrow, ncol) elif len(p.shape) == 1: p = matlib.repmat(p, n, 1) dim = p.shape[0] p = matlib.repmat(p, n, 1) f = self.rand.rand(n, dim) res = f < p res = numpy.sum(res, axis=0) else: # will make exeption raise ValueError("shape of input Binarization class " "must be 1 or 2 dimensions") return res
class GradientsCalculator(AcceleratedUnit, EmptyDeviceMethodsMixin): """ Making gradients for weights, hbias and vbias, using hbias0, vbias0 and vbias1, hbias1, which calculated with help BatchWeights. Must be assigned before initialize(): * hbias0 * vbias0 * hbias1 * vbias1 * weights1 * weights0 Updates after run(): * hbias_grad * vbias_grad * weights_grad Creates within initialize(): * hbias_grad * vbias_grad * weights_grad Attributes: vbias0: calculated with help BatchWeights from v0 hbias0: calculated with help BatchWeights from h0 vbias1: calculated with help BatchWeights from v1 hbias1: calculated with help BatchWeights from h1 weights1: calculated with help BatchWeights from v1. weights0: calculated with help BatchWeights from h1. hbias_grad: gradient for hbias vbias_grad: gradient for vbias weights_grad: gradient for weights """ def __init__(self, workflow, **kwargs): super(GradientsCalculator, self).__init__(workflow, **kwargs) self.vbias_grad = Array() self.hbias_grad = Array() self.weights_grad = Array() self.demand("hbias1", "vbias1", "hbias0", "vbias0", "weights0", "weights1") def initialize(self, device, **kwargs): super(GradientsCalculator, self).initialize(device=device, **kwargs) if not self.hbias_grad: self.hbias_grad.reset( numpy.zeros(self.hbias0.shape, dtype=self.hbias0.dtype)) else: assert self.hbias_grad.shape == self.hbias0.shape if not self.vbias_grad: self.vbias_grad.reset( numpy.zeros(self.vbias0.shape, dtype=self.vbias0.dtype)) else: assert self.vbias_grad.shape == self.vbias0.shape if not self.weights_grad: self.weights_grad.reset( numpy.zeros(self.weights0.shape, dtype=self.weights0.dtype)) else: assert self.weights_grad.shape == self.weights0.shape for v in (self.weights_grad, self.hbias_grad, self.vbias_grad, self.hbias0, self.vbias0, self.weights0, self.hbias1, self.vbias1, self.weights1): v.initialize(self.device) def run(self): for v in (self.hbias0, self.vbias0, self.weights0, self.hbias1, self.vbias1, self.weights1): v.map_read() for v in (self.weights_grad, self.vbias_grad, self.hbias_grad): v.map_invalidate() self.vbias_grad.mem[:] = self.vbias0.mem - self.vbias1.mem self.hbias_grad.mem[:] = self.hbias0.mem - self.hbias1.mem self.weights_grad.mem[:] = self.weights0.mem - self.weights1.mem
class InputJoiner(AcceleratedUnit): """Joins several minibatch inputs into one continuous minibatch output. Attributes: input_0, input_1, ...: inputs of type Array(), created via link_inputs offset_0, offset_1, ...: offsets of each input in elements, have valid values after initialize(). length_0, length_1, ...: lengths of each input in elements, have valid values after initialize. output: Array() minibatch_size: size of the minibatch (will be set to the minimum of the first shapes from the inputs if not provided prior to the initialize) """ def __init__(self, workflow, **kwargs): super(InputJoiner, self).__init__(workflow, **kwargs) self.output = Array() self._num_inputs = 0 self.inputs = kwargs.get("inputs") def init_unpickled(self): super(InputJoiner, self).init_unpickled() self.sources_["join"] = {} @property def num_inputs(self): return self._num_inputs @num_inputs.setter def num_inputs(self, value): try: value = int(value) except (ValueError, TypeError): raise ValueError("num_inputs must be copnvertible to int") for x in range(value, self._num_inputs): try: delattr(self, "input_%d" % x) delattr(self, "offset_%d" % x) delattr(self, "length_%d" % x) except AttributeError: pass for x in range(self._num_inputs, value): setattr(self, "input_%d" % x, None) setattr(self, "offset_%d" % x, None) setattr(self, "length_%d" % x, None) self._num_inputs = value @property def inputs(self): return list(getattr(self, "input_%d" % x) for x in range(self._num_inputs)) @property def offsets(self): return list(getattr(self, "offset_%d" % x) for x in range(self._num_inputs)) @property def lengths(self): return list(getattr(self, "length_%d" % x) for x in range(self._num_inputs)) @inputs.setter def inputs(self, value): if value is None: self.num_inputs = 0 return if not hasattr(value, "__iter__"): raise TypeError("inputs must be iterable") self.num_inputs = len(value) for i, inp in enumerate(value): setattr(self, "input_%d" % i, inp) def link_inputs(self, other, *args): """Adds more inputs and links them. It will link args to attributes named "input_0", "input_1", etc. Parameters: other: unit from which to link attributes. args: attribute names to link. """ if not len(args): raise ValueError("args may not be empty") num_inputs = self.num_inputs self.num_inputs = num_inputs + len(args) for arg in args: self.link_attrs(other, ("input_%d" % num_inputs, arg)) num_inputs += 1 def _init_offset_length_attributes(self): """Initializes offset_0, offset_1, ... length_0, length_1, ... """ offset = 0 for i in range(self.num_inputs): inp = getattr(self, "input_%d" % i) setattr(self, "offset_%d" % i, offset) setattr(self, "length_%d" % i, inp.sample_size) offset += inp.sample_size def initialize(self, device, **kwargs): if any(i.mem is None for i in self.inputs): # Not yet ready to initialize return True self._init_offset_length_attributes() super(InputJoiner, self).initialize(device=device, **kwargs) minibatch_size = min(i.shape[0] for i in self.inputs) if any(i.shape[0] > minibatch_size for i in self.inputs): self.warning("Detected inputs of different sizes. Sizes will be " "cut to the lowest value (%d)", minibatch_size) output_shape = (minibatch_size, sum(i.size // i.shape[0] for i in self.inputs)) if not self.output: self.output.reset(numpy.zeros(output_shape, self.inputs[0].dtype)) else: assert self.output.shape == output_shape self.init_vectors(self.output, *self.inputs) def _gpu_init(self): defines = { 'etype': opencl_types.numpy_dtype_to_opencl(self.output.dtype), } self.build_program( defines, "%s_%d_%s" % (type(self).__name__, self.output.shape[0], "_".join(map(str, self.output.shape[1:]))), inputs=self.inputs) self.assign_kernel("join") self.set_args(self.output, *self.inputs) def ocl_init(self): self._gpu_init() def cuda_init(self): self._gpu_init() def numpy_run(self): self.output.map_invalidate() # we will update output on CPU minibatch_size = self.output.shape[0] low = 0 for inp in self.inputs: inp.map_read() high = low + inp.size // inp.shape[0] if low >= high: break self.output.mem[:, low:high] = inp[:minibatch_size] low = high def ocl_run(self): for inp in self.inputs: inp.unmap() self.execute_kernel(*((self.output.shape[0],),) * 2) def cuda_run(self): for inp in self.inputs: inp.unmap() # TODO(a.kazantsev): rewrite CUDA kernel for proper grid size self.execute_kernel((1, 1, 1), (self.output.shape[0], 1, 1))
class All2AllSoftmax(All2All): """All2All with linear activation and softmax normalization. Must be assigned before initialize(): Updates after run(): max_idx Creates within initialize(): max_idx Attributes: krn_sm_: kernel for softmax activation calculation. max_idx: indexes of element with maximum value for each sample. """ __id__ = "420219fc-3e1a-45b1-87f8-aaa0c1540de4" MAPPING = {"softmax"} def __init__(self, workflow, **kwargs): super(All2AllSoftmax, self).__init__(workflow, **kwargs) self.max_idx = Array() self.reduce_size = 256 def init_unpickled(self): super(All2AllSoftmax, self).init_unpickled() self.krn_sm_ = None self._force_gpu_apply_exp = False def initialize(self, device, **kwargs): self.reduce_size = min(self.reduce_size, int(numpy.prod(self.output_sample_shape))) self.sources_["all2all/softmax"] = {"REDUCE_SIZE": self.reduce_size} retval = super(All2AllSoftmax, self).initialize(device=device, **kwargs) if retval: return retval if self.output.mem.size // self.output.mem.shape[0] <= 1: raise error.BadFormatError( "Output sample size should be greater than 1 for SoftMax.") if not self.max_idx: self.max_idx.reset( numpy.zeros(self.output.shape[0], dtype=numpy.int32)) self.max_idx.initialize(self.device) return retval def numpy_apply_exp(self): self.output.map_write() self.max_idx.map_invalidate() out = self.output.mem out = reshape(out, (out.shape[0], out.size // out.shape[0])) for i, sample in enumerate(out): im = sample.argmax() self.max_idx[i] = im m = sample[im] sample -= m numpy.exp(sample, sample) smm = sample.sum() sample /= smm def ocl_apply_exp(self): self.unmap_vectors(self.output, self.max_idx) global_size = (self.output.shape[0] * self.reduce_size, ) local_size = (self.reduce_size, ) self.execute_kernel(global_size, local_size, self.krn_sm_) def cuda_apply_exp(self): self.unmap_vectors(self.output, self.max_idx) global_size = (self.output.shape[0], 1, 1) local_size = (self.reduce_size, 1, 1) self.execute_kernel(global_size, local_size, self.krn_sm_) def numpy_run(self): """Forward propagation from batch on CPU only. """ super(All2AllSoftmax, self).numpy_run() if not self._force_gpu_apply_exp: self.numpy_apply_exp() def ocl_run(self): """Forward propagation from batch on GPU. """ self._force_gpu_apply_exp = True super(All2AllSoftmax, self).ocl_run() self.ocl_apply_exp() def cuda_run(self): """Forward propagation from batch on GPU. """ self._force_gpu_apply_exp = True super(All2AllSoftmax, self).cuda_run() self.cuda_apply_exp() def ocl_init(self): super(All2AllSoftmax, self).ocl_init() self.krn_sm_ = self.get_kernel("apply_exp") self.krn_sm_.set_args(self.output.devmem, self.max_idx.devmem) def cuda_init(self): super(All2AllSoftmax, self).cuda_init() self.krn_sm_ = self.get_kernel("apply_exp") self.krn_sm_.set_args(self.output.devmem, self.max_idx.devmem)
class OffsetPooling(Pooling): """Pooling by offset forward propagation. Must be assigned before initialize(): Updates after run(): input_offset Creates within initialize(): input_offset Attributes: input_offset: offsets in the input where elements are passed through. """ MAPPING = set() hide_from_registry = True def __init__(self, workflow, **kwargs): super(OffsetPooling, self).__init__(workflow, **kwargs) self.input_offset = Array() self.demand("input") def initialize(self, device, **kwargs): super(OffsetPooling, self).initialize(device=device, **kwargs) if self._no_output: return if not self.input_offset: self.input_offset.reset(numpy.zeros(self.output.shape, dtype=numpy.int32)) else: assert self.input_offset.shape == self.output.shape self.input_offset.initialize(self.device) def set_args(self, *args): super(OffsetPooling, self).set_args(self.input, self.output, self.input_offset, *args) def ocl_run(self): self.input_offset.unmap() super(OffsetPooling, self).ocl_run() def cuda_run(self): self.input_offset.unmap() super(OffsetPooling, self).cuda_run() def numpy_run(self): self.input_offset.map_invalidate() super(OffsetPooling, self).numpy_run() def numpy_run_cut(self, cut, coords): batch, y1, x1, ch, out_y, out_x = coords cut_index = self.numpy_run_cut_offset( cut, numpy.ravel_multi_index((batch, out_y, out_x, ch), self.output.shape)) i, j = numpy.unravel_index(cut_index, cut.shape) idx = numpy.ravel_multi_index((batch, y1 + i, x1 + j, ch), self.input.shape) val = numpy.ravel(self.input.mem)[idx] self.input_offset.mem[batch, out_y, out_x, ch] = idx return val
class KohonenForward(KohonenBase, AcceleratedUnit): """Kohonen forward layer. Must be assigned before initialize(): input weights minibatch_offset (if total == True) minibatch_size (if total == True) batch_size (if total == True) argmins speeds up run() if linked from KohonenTrainer Updates after run(): output Creates within initialize(): output Attributes: input: input as batch of samples. weights: the weights of the neurons in Kohonen layer. output: the list of winners. total: if total=True is passed in __init__(), the overall winners table """ def __init__(self, workflow, **kwargs): super(KohonenForward, self).__init__(workflow, **kwargs) self.demand("input", "weights") self.argmins = None self._distances = Array() self.output = Array() self._chunk_size_ = 0 self.weights_transposed = False self.total = Array() if kwargs.get("total", False) else None if self.total is not None: self.minibatch_offset = None self.minibatch_size = None self.batch_size = None def init_unpickled(self): super(KohonenForward, self).init_unpickled() self.sources_["kohonen"] = {"FORWARD": 1} @property def neurons_number(self): return self.weights.mem.shape[0] @property def sample_length(self): return self.weights.mem.shape[1] @property def chunk_size(self): return self._chunk_size_ def initialize(self, device, **kwargs): super(KohonenForward, self).initialize(device=device, **kwargs) assert self.input.mem.shape[1] == self.sample_length batch_size = self.input.mem.shape[0] self.output.reset(numpy.zeros(batch_size, dtype=numpy.int32)) if self.argmins is None: self._distances.reset(numpy.zeros( [batch_size, self.neurons_number], dtype=self.weights.mem.dtype)) if self.total is not None: self.total.reset(numpy.zeros(self.batch_size, dtype=numpy.int32)) self._minibatch_offset_ = numpy.zeros(1, dtype=numpy.int32) def ocl_init(self): batch_size = self.input.mem.shape[0] self.output.initialize(self.device) if self.argmins is None: self.input.initialize(self.device) self.weights.initialize(self.device) self._distances.initialize(self.device) elif self.total is None: return if self.total is not None: self.total.initialize(self.device) copy_chunk_size = int(numpy.ceil(batch_size / self.device.max_group_size)) chunk_size = self.neurons_number // self.device.max_group_size if chunk_size < 2: chunk_size = self.neurons_number // 2 + 1 self.argmin_group_size = \ int(numpy.ceil(self.neurons_number / chunk_size)) block_size, vector_opt = self.device.device_info.get_kernel_bs_vo( kernel="matrix_multiplication", dtype=self.input.dtype) defines = { 'BLOCK_SIZE': block_size, 'VECTOR_OPT': int(bool(vector_opt)), 'BATCH': batch_size, 'SAMPLE_LENGTH': self.sample_length, 'NEURONS_NUMBER': self.neurons_number, 'CHUNK_SIZE': chunk_size, 'COPY_CHUNK_SIZE': copy_chunk_size, } if self.weights_transposed: defines['WEIGHTS_TRANSPOSED'] = 1 self.build_program(defines, "%s_%d_%d_%d" % (self.__class__.__name__, batch_size, self.sample_length, self.neurons_number), dtype=self.weights.mem.dtype) if self.total is not None: self._set_total_global_size_ = \ [int(numpy.ceil(batch_size / copy_chunk_size))] self._krn_set_total_ = self.get_kernel("set_total") self._krn_set_total_.set_args(self.output.devmem, cl.skip, self.total.devmem) if self.argmins is not None: return self._krn_distances_ = self.get_kernel("calculate_distances") self._krn_distances_.set_args(self.input.devmem, self.weights.devmem, self._distances.devmem) self._krn_argmin_ = self.get_kernel("calculate_argmin") self._krn_argmin_.set_args(self._distances.devmem, self.output.devmem, None) self._gs_distance = [ roundup(self.neurons_number, block_size), roundup(batch_size, block_size)] self._ls_distance = [block_size, block_size] def ocl_run(self): self.output.unmap() if self.total is not None: self.total.unmap() if self.argmins is None: self.input.unmap() self.weights.unmap() self.execute_kernel(self._gs_distance, self._ls_distance, self._krn_distances_) self.execute_kernel([self.argmin_group_size], [self.argmin_group_size], self._krn_argmin_) else: self.argmins.unmap() self.argmins.map_read() self.output.map_write() self.output.mem[:] = self.argmins.mem self.output.unmap() self.argmins.unmap() if self.total is not None: self._minibatch_offset_[0] = \ self.minibatch_offset - self.minibatch_size self._krn_set_total_.set_arg(1, self._minibatch_offset_) self.execute_kernel(self._set_total_global_size_, None, self._krn_set_total_) def numpy_run(self): self.output.map_invalidate() if self.argmins is not None: self.argmins.map_read() self.output.mem[:] = self.argmins.mem else: self.input.map_read() self.weights.map_read() if self.total is not None: self.total.map_invalidate() length = self.minibatch_size if self.total is not None \ else self.input.mem.shape[0] for sindex in range(length): if self.argmins is None: dist = self.weights.mem - self.input[sindex] winner = numpy.argmin(self.numpy_linalg_norm(dist)) self.output[sindex] = winner else: winner = self.argmins[sindex] if self.total is not None: index = sindex + self.minibatch_offset - self.minibatch_size self.total[index] = winner
class Cutter1D(AcceleratedUnit): """Cuts the specified interval from each 1D sample of input batch into output. y = alpha * x + beta * y """ def __init__(self, workflow, **kwargs): super(Cutter1D, self).__init__(workflow, **kwargs) self.alpha = kwargs.get("alpha") self.beta = kwargs.get("beta") self.output_offset = kwargs.get("output_offset", 0) self.output = Array() self.demand("alpha", "beta", "input") # TODO: add input_offset and length to demand and not to crash lstm # TODO: unit test def init_unpickled(self): super(Cutter1D, self).init_unpickled() self.sources_["cutter"] = {} def initialize(self, device, **kwargs): super(Cutter1D, self).initialize(device, **kwargs) if not self.output or self.output.shape[0] != self.input.shape[0]: self.output.reset( numpy.zeros( (self.input.shape[0], self.output_offset + self.length), dtype=self.input.dtype)) else: assert self.output.sample_size >= self.output_offset + self.length self.init_vectors(self.input, self.output) def cuda_init(self): dtype = self.input.dtype itemsize = self.input.itemsize limit = self.input.shape[0] * self.length self.build_program({}, "%s" % self.__class__.__name__, dtype=dtype) self.assign_kernel("cutter_1d_forward") self.set_args( int(self.input.devmem) + self.input_offset * itemsize, numpy.array([self.alpha], dtype=dtype), numpy.array([self.input.sample_size], dtype=numpy.int32), int(self.output.devmem) + self.output_offset * itemsize, numpy.array([self.beta], dtype=dtype), numpy.array([self.output.sample_size], dtype=numpy.int32), numpy.array([self.length], dtype=numpy.int32), numpy.array([limit], dtype=numpy.int32)) block_size = self.device.suggest_block_size(self._kernel_) self._global_size = (int(numpy.ceil(limit / block_size)), 1, 1) self._local_size = (block_size, 1, 1) def ocl_init(self): dtype = self.input.dtype self.build_program({}, "%s" % self.__class__.__name__, dtype=dtype) self.assign_kernel("cutter_1d_forward") self.set_args( self.input.devmem, numpy.array([self.input_offset], dtype=numpy.int32), numpy.array([self.alpha], dtype=dtype), numpy.array([self.input.sample_size], dtype=numpy.int32), self.output.devmem, numpy.array([self.output_offset], dtype=numpy.int32), numpy.array([self.beta], dtype=dtype), numpy.array([self.output.sample_size], dtype=numpy.int32)) self._global_size = (self.input.shape[0], self.length) self._local_size = None def _gpu_run(self): self.unmap_vectors(self.input, self.output) self.execute_kernel(self._global_size, self._local_size) def cuda_run(self): return self._gpu_run() def ocl_run(self): return self._gpu_run() def numpy_run(self): self.input.map_read() self.output.map_write() out = self.output.matrix[:, self.output_offset:self.output_offset + self.length] if self.beta: out *= self.beta else: out[:] = 0 out += (self.input.matrix[:, self.input_offset:self.input_offset + self.length] * self.alpha)
class InputJoiner(AcceleratedUnit): """Joins several minibatch inputs into one continuous minibatch output. Must be assigned before initialize(): inputs Updates after run(): output Creates within initialize(): output Attributes: inputs: list of inputs of type memory.Array(). output: memory.Array(). minibatch_size: size of the minibatch (will be set to the minimum of the first shapes from the inputs if not provided prior to the initialize) """ def __init__(self, workflow, **kwargs): super(InputJoiner, self).__init__(workflow, **kwargs) self.inputs = kwargs["inputs"] self.output = Array() self.registered_inputs = {} def init_unpickled(self): super(InputJoiner, self).init_unpickled() self.sources_["join"] = {} @property def inputs(self): return self._inputs @inputs.setter def inputs(self, value): if not hasattr(value, "__iter__"): raise TypeError("inputs must be iterable") self._inputs = list(value) if len(self._inputs) == 0: raise ValueError("inputs may not be empty") def register_offset_length_attributes(self, inp): idx = len(self.registered_inputs) attrs = ("offset_%d" % idx, "length_%d" % idx) for attr in attrs: setattr(self, attr, -1) self.registered_inputs[inp] = attrs return attrs def _init_offset_length_attributes(self): offsets = [] lengths = [] offset = 0 for inp in self.inputs: offsets.append(offset) lengths.append(inp.sample_size) offset += lengths[-1] for inp, attrs in self.registered_inputs.items(): try: idx = self.inputs.index(inp) vals = (offsets[idx], lengths[idx]) except ValueError: vals = (-1, -1) for i, attr in enumerate(attrs): setattr(self, attr, vals[i]) def initialize(self, device, **kwargs): if any(i.mem is None for i in self.inputs): # Not yet ready to initialize return True self._init_offset_length_attributes() super(InputJoiner, self).initialize(device=device, **kwargs) minibatch_size = min(i.shape[0] for i in self.inputs) if any(i.shape[0] > minibatch_size for i in self.inputs): self.warning( "Detected inputs of different sizes. Sizes will be " "cut to the lowest value (%d)", minibatch_size) output_shape = (minibatch_size, sum(i.size // i.shape[0] for i in self.inputs)) if not self.output: self.output.reset(numpy.zeros(output_shape, self.inputs[0].dtype)) else: assert self.output.shape == output_shape self.init_vectors(self.output, *self.inputs) def _gpu_init(self): defines = { 'etype': opencl_types.numpy_dtype_to_opencl(self.output.dtype), } self.build_program( defines, "%s_%d_%s" % (type(self).__name__, self.output.shape[0], "_".join( map(str, self.output.shape[1:]))), inputs=self.inputs) self.assign_kernel("join") self.set_args(self.output, *self.inputs) def ocl_init(self): self._gpu_init() def cuda_init(self): self._gpu_init() def numpy_run(self): self.output.map_invalidate() # we will update output on CPU minibatch_size = self.output.shape[0] low = 0 for inp in self.inputs: inp.map_read() high = low + inp.size // inp.shape[0] if low >= high: break self.output.mem[:, low:high] = inp[:minibatch_size] low = high def ocl_run(self): for inp in self.inputs: inp.unmap() self.execute_kernel(*((self.output.shape[0], ), ) * 2) def cuda_run(self): for inp in self.inputs: inp.unmap() # TODO(a.kazantsev): rewrite CUDA kernel for proper grid size self.execute_kernel((1, 1, 1), (self.output.shape[0], 1, 1))
class EvaluatorMSE(EvaluatorBase): MAPPING = "evaluator_mse" LOSS = "mse" """Evaluator for nn softmax output from the batch labels. Must be assigned before initialize(): output target batch_size labels (may be None) class_targets (may be None) Updates after run(): err_output confusion_matrix max_err_output_sum n_err (only if labels and class_targets is not None) Creates within initialize(): err_output n_err (only if labels and class_targets is not None) max_err_output_sum Attributes: output: output of the network_common as Batch. target: target for the current Batch. err_output: backpropagation errors. batch_size: number of elements in output to evaluate. metrics: [0] - sum of sample's mse, [1] - max of sample's mse, [2] - min of sample's mse. mse: array of mse for each sample in minibatch. krn_constants_i_: numpy array for constant arguments to kernel. labels: labels for a batch (may be None). class_targets: target for each class (may be None). n_err: number of wrongly recognized samples (if labels and class_targets is not None). """ def __init__(self, workflow, **kwargs): super(EvaluatorMSE, self).__init__(workflow, **kwargs) self.metrics = Array() self.mse = Array() self.labels = None self.class_targets = None self.n_err = Array() self.root = kwargs.get("root", True) self.demand("target", "normalizer") @property def root(self): """ :return: True if error metric is RMSE, otherwise, MSE (mean sum of squares). Default is True. """ return self._root @root.setter def root(self, value): if not isinstance(value, bool): raise TypeError("root must be boolean (got %s)" % type(value)) self._root = value def initialize(self, device, **kwargs): super(EvaluatorMSE, self).initialize(device=device, **kwargs) if self.testing: return if self.target.size != self.output.size: raise error.BadFormatError( "target.size != output.size (%s != %s)" % (self.target.size, self.output.size)) self.sources_["evaluator_mse"] = {} self.sources_["denormalization"] = {} dtype = self.output.dtype self.metrics.reset(numpy.zeros(3, dtype=dtype)) self.metrics[2] = 1.0e30 # mse_min self.mse.reset(numpy.zeros(self.err_output.mem.shape[0], dtype)) self.n_err.reset(numpy.zeros(2, dtype=numpy.int32)) self.init_vectors(self.n_err, self.target, self.metrics, self.mse) if self.class_targets: self.class_targets.initialize(self.device) def _gpu_init(self): dtype = self.output.dtype block_size = min(self.err_output.shape[0], 128) if self.class_targets: self.sources_["mse_find_closest"] = { "target_dtype": numpy_dtype_to_opencl(self.class_targets.dtype) } self.build_program(cache_file_name="%s_%d_%d" % (self.__class__.__name__, self.output.shape[0], self.output.sample_size), dtype=dtype, max_batch_size=self.err_output.shape[0], block_size=block_size, output_size=self.err_output.sample_size, root=self.root, normalization=self.normalizer.MAPPING, targets_number=self.class_targets.shape[0] if self.class_targets else None, coeffs=self.normalizer.coefficients) self.assign_kernel("evaluate_mse") self.set_args(self.output, self.target, self.skip_args(2), self.metrics, self.mse.devmem, self.err_output) if self.labels and self.class_targets: assert (self.labels.dtype == self.n_err.dtype == numpy.int32) self.krn_find_closest_ = self.get_kernel("mse_find_closest") self.krn_find_closest_.set_args(self.output.devmem, self.class_targets.devmem, self.labels.devmem, self.n_err.devmem) return block_size def ocl_init(self): if self.testing: return block_size = self._gpu_init() self._local_size = [block_size] self._global_size = self._local_size self._global_size_find_closest_ = lambda: (self.batch_size, ) self._local_size_find_closest = None def cuda_init(self): if self.testing: return block_size = self._gpu_init() self._local_size = (block_size, 1, 1) self._global_size = (1, 1, 1) self._global_size_find_closest_ = lambda: (self.batch_size, 1, 1) self._local_size_find_closest = (1, 1, 1) def _gpu_run(self): self.unmap_vectors(self.err_output, self.output, self.target, self.metrics, self.mse) batch_size = self.batch_size self.krn_constants_i_[0] = batch_size self.set_arg(2, self.krn_constants_i_[0:1]) self.krn_constants_f_[0] = 1.0 / self.batch_size if self.mean else 1.0 self.set_arg(3, self.krn_constants_f_[0:1]) self.execute_kernel(self._global_size, self._local_size) if self.labels and self.class_targets: self.unmap_vectors(self.class_targets, self.labels, self.n_err) self.execute_kernel(self._global_size_find_closest_(), self._local_size_find_closest, self.krn_find_closest_) self.n_err.map_write() self.n_err.mem[1] += batch_size def ocl_run(self): return self._gpu_run() def cuda_run(self): return self._gpu_run() def numpy_run(self): self.output.map_read() self.target.map_read() self.metrics.map_write() self.err_output.map_invalidate() self.mse.map_invalidate() assert (self.output.size == self.target.size == self.err_output.size) batch_size = self.batch_size err_output = self.err_output.matrix[:batch_size] assert_addr(err_output, self.err_output.mem) output = self.output.matrix[:batch_size] assert_addr(output, self.output.mem) target = self.target.matrix[:batch_size] assert_addr(target, self.target.mem) mse = self.mse.mem[:batch_size] assert_addr(mse, self.mse.mem) err_output[:] = output - target if not isinstance(self.normalizer, NoneNormalizer): output_copy = output.copy() target_copy = target.copy() self.normalizer.denormalize(output_copy) self.normalizer.denormalize(target_copy) denormed_err_output = output_copy - target_copy else: denormed_err_output = err_output self.err_output.mem[batch_size:] = 0 mse[:] = numpy.square(denormed_err_output).sum(axis=1) / \ denormed_err_output.shape[1] if self.mean: err_output /= batch_size if self.root: numpy.sqrt(mse, mse) self.mse.mem[batch_size:] = 0 self.metrics.mem[0] += mse.sum() self.metrics.mem[1] = max(self.metrics.mem[1], mse.max()) self.metrics.mem[2] = min(self.metrics.mem[2], mse.min()) if self.labels and self.class_targets: self.class_targets.map_read() self.labels.map_read() self.n_err.map_write() class_targets = self.class_targets.matrix labels = self.labels.mem for i, sample in enumerate(output): lbl = numpy.linalg.norm(class_targets - sample, axis=1).argmin() if lbl != labels[i]: self.n_err.mem[0] += 1 self.n_err.mem[1] += 1 def merge_output(self): if not isinstance(self.normalizer, NoneNormalizer): output = self.output[:self.batch_size].copy() self.normalizer.denormalize(output) else: output = self.output.mem self.merged_output[self.offset - self.batch_size:self.offset] = output
class ZeroFiller(ForwardBase, TriviallyDistributable): """Fills weights of given unit with zero on every step""" MAPPING = {"zero_filter"} def __init__(self, workflow, **kwargs): super(ZeroFiller, self).__init__(workflow, **kwargs) self.mask = Array() self.grouping = kwargs.get("grouping", 1) self.demand("weights") def init_unpickled(self): super(ZeroFiller, self).init_unpickled() self.sources_["weights_zerofilling"] = {} @property def effective_shape(self): return (self.weights.shape[0], self.weights.size // self.weights.shape[0]) @property def grouping(self): return self._grouping @grouping.setter def grouping(self, value): if not isinstance(value, int): raise TypeError( "grouping value must be an integer (got %s)" % type(value)) if value < 2: raise ValueError("grouping value %d is invalid" % value) self._grouping = value def initialize(self, device=None, **kwargs): super(ZeroFiller, self).initialize(device, **kwargs) if not self.weights: return True if not self.mask: if self.effective_shape[1] % self.grouping != 0: raise ValueError( "Non-multiple of grouping weights shape detected: " "%s, grouping=%d" % (self.weights.shape, self.grouping)) self.mask.reset(numpy.zeros(self.effective_shape, dtype=self.weights.dtype)) self.mask.map_invalidate() # TODO(a.kazantsev): add check for transposed weights. for kernel in range(self.effective_shape[0]): for chan in range(self.effective_shape[1]): self.mask[kernel, chan] = not ( kernel % self.grouping == chan % self.grouping) else: assert self.mask.shape == self.effective_shape for vec in self.mask, self.weights: vec.initialize(device) def _gpu_init(self): self.build_program(cache_file_name="zero_filling_%d" % self.grouping, dtype=self.weights.dtype) self.assign_kernel("multiply_by_mask") self.set_args(self.mask, self.weights) def ocl_init(self): self._gpu_init() self._global_size = [self.weights.size] self._local_size = None def cuda_init(self): self._gpu_init() self._global_size = (self.weights.size, 1, 1) self._local_size = (1, 1, 1) def numpy_run(self): self.mask.map_read() self.weights.map_write() self.weights.mem *= self.mask.mem def _gpu_run(self): self.weights.unmap() self.mask.unmap() self.execute_kernel(self._global_size, self._local_size) def ocl_run(self): self._gpu_run() def cuda_run(self): self._gpu_run()
class ZeroFiller(ForwardBase, TriviallyDistributable): """Fills weights of given unit with zero on every step""" MAPPING = {"zero_filter"} def __init__(self, workflow, **kwargs): super(ZeroFiller, self).__init__(workflow, **kwargs) self.mask = Array() self.grouping = kwargs.get("grouping", 1) self.demand("weights") def init_unpickled(self): super(ZeroFiller, self).init_unpickled() self.sources_["weights_zerofilling"] = {} @property def effective_shape(self): return (self.weights.shape[0], self.weights.size // self.weights.shape[0]) @property def grouping(self): return self._grouping @grouping.setter def grouping(self, value): if not isinstance(value, int): raise TypeError("grouping value must be an integer (got %s)" % type(value)) if value < 2: raise ValueError("grouping value %d is invalid" % value) self._grouping = value def initialize(self, device=None, **kwargs): super(ZeroFiller, self).initialize(device, **kwargs) if not self.weights: return True if not self.mask: if self.effective_shape[1] % self.grouping != 0: raise ValueError( "Non-multiple of grouping weights shape detected: " "%s, grouping=%d" % (self.weights.shape, self.grouping)) self.mask.reset( numpy.zeros(self.effective_shape, dtype=self.weights.dtype)) self.mask.map_invalidate() # TODO(a.kazantsev): add check for transposed weights. for kernel in range(self.effective_shape[0]): for chan in range(self.effective_shape[1]): self.mask[kernel, chan] = not (kernel % self.grouping == chan % self.grouping) else: assert self.mask.shape == self.effective_shape for vec in self.mask, self.weights: vec.initialize(device) def _gpu_init(self): self.build_program(cache_file_name="zero_filling_%d" % self.grouping, dtype=self.weights.dtype) self.assign_kernel("multiply_by_mask") self.set_args(self.mask, self.weights) def ocl_init(self): self._gpu_init() self._global_size = [self.weights.size] self._local_size = None def cuda_init(self): self._gpu_init() self._global_size = (self.weights.size, 1, 1) self._local_size = (1, 1, 1) def numpy_run(self): self.mask.map_read() self.weights.map_write() self.weights.mem *= self.mask.mem def _gpu_run(self): self.weights.unmap() self.mask.unmap() self.execute_kernel(self._global_size, self._local_size) def ocl_run(self): self._gpu_run() def cuda_run(self): self._gpu_run()
class BatchWeights(AcceleratedUnit, EmptyDeviceMethodsMixin): """Make weigths and biases from batch v and h. Must be assigned before initialize(): * v * h * batch_size Updates after run(): * hbias_batch * vbias_batch * W_batch Creates within initialize(): * hbias_batch * vbias_batch * W_batch Attributes: v: input data batch h: hidden states of input batch batch_size: size of batch hbias_batch: bias calculated from h vbias_batch: bias calculated from v W_batch: weigths calculated from batch v and h """ def __init__(self, workflow, **kwargs): super(BatchWeights, self).__init__(workflow, **kwargs) self.vbias_batch = Array() self.hbias_batch = Array() self.weights_batch = Array() self.demand("v", "h", "batch_size") def initialize(self, device, **kwargs): super(BatchWeights, self).initialize(device=device, **kwargs) vbias_size = self.v.size // self.v.shape[0] hbias_size = self.h.size // self.h.shape[0] W_size = vbias_size * hbias_size if not self.hbias_batch: self.hbias_batch.reset(numpy.zeros((1, hbias_size), dtype=self.h.mem.dtype)) else: assert self.hbias_batch.size == hbias_size if not self.vbias_batch: self.vbias_batch.reset(numpy.zeros((1, vbias_size), dtype=self.h.mem.dtype)) else: assert self.vbias_batch.size == vbias_size if not self.weights_batch: self.weights_batch.reset(numpy.zeros((vbias_size, hbias_size), dtype=self.h.mem.dtype)) else: assert self.weights_batch.size == W_size self.init_vectors(self.weights_batch, self.vbias_batch, self.hbias_batch, self.v, self.h) def run(self): self.v.map_read() self.h.map_read() for v in self.weights_batch, self.hbias_batch, self.vbias_batch: v.map_invalidate() self.weights_batch.mem[:] = numpy.dot( numpy.transpose(self.v.mem[0: self.batch_size, :]), self.h.mem[0: self.batch_size, :]) / \ self.batch_size for bv in (self.vbias_batch, self.v), (self.hbias_batch, self.h): bv[0].mem[:] = (numpy.sum(bv[1].mem[:self.batch_size, :], 0) / self.batch_size) bv[0].shape = (1, bv[0].size)
class GradientsCalculator(AcceleratedUnit, EmptyDeviceMethodsMixin): """ Making gradients for weights, hbias and vbias, using hbias0, vbias0 and vbias1, hbias1, which calculated with help BatchWeights. Must be assigned before initialize(): * hbias0 * vbias0 * hbias1 * vbias1 * weights1 * weights0 Updates after run(): * hbias_grad * vbias_grad * weights_grad Creates within initialize(): * hbias_grad * vbias_grad * weights_grad Attributes: vbias0: calculated with help BatchWeights from v0 hbias0: calculated with help BatchWeights from h0 vbias1: calculated with help BatchWeights from v1 hbias1: calculated with help BatchWeights from h1 weights1: calculated with help BatchWeights from v1. weights0: calculated with help BatchWeights from h1. hbias_grad: gradient for hbias vbias_grad: gradient for vbias weights_grad: gradient for weights """ def __init__(self, workflow, **kwargs): super(GradientsCalculator, self).__init__(workflow, **kwargs) self.vbias_grad = Array() self.hbias_grad = Array() self.weights_grad = Array() self.demand("hbias1", "vbias1", "hbias0", "vbias0", "weights0", "weights1") def initialize(self, device, **kwargs): super(GradientsCalculator, self).initialize(device=device, **kwargs) if not self.hbias_grad: self.hbias_grad.reset(numpy.zeros(self.hbias0.shape, dtype=self.hbias0.dtype)) else: assert self.hbias_grad.shape == self.hbias0.shape if not self.vbias_grad: self.vbias_grad.reset(numpy.zeros(self.vbias0.shape, dtype=self.vbias0.dtype)) else: assert self.vbias_grad.shape == self.vbias0.shape if not self.weights_grad: self.weights_grad.reset(numpy.zeros(self.weights0.shape, dtype=self.weights0.dtype)) else: assert self.weights_grad.shape == self.weights0.shape for v in (self.weights_grad, self.hbias_grad, self.vbias_grad, self.hbias0, self.vbias0, self.weights0, self.hbias1, self.vbias1, self.weights1): v.initialize(self.device) def run(self): for v in (self.hbias0, self.vbias0, self.weights0, self.hbias1, self.vbias1, self.weights1): v.map_read() for v in (self.weights_grad, self.vbias_grad, self.hbias_grad): v.map_invalidate() self.vbias_grad.mem[:] = self.vbias0.mem - self.vbias1.mem self.hbias_grad.mem[:] = self.hbias0.mem - self.hbias1.mem self.weights_grad.mem[:] = self.weights0.mem - self.weights1.mem
class MeanDispNormalizer(AcceleratedUnit, TriviallyDistributable): """Normalizes multichannel byte images according to dataset mean and dispersion. Attributes: input: minibatch of images (dtype=numpy.uint8, shape[0]=minibatch_size). mean: mean image over the dataset (dtype=numpy.uint8). rdisp: 1.0 / dispersion over the dataset (float datatype). output: normalized float images of the same dtype as rdisp. """ def __init__(self, workflow, **kwargs): kwargs["view_group"] = kwargs.get("view_group", "WORKER") super(MeanDispNormalizer, self).__init__(workflow, **kwargs) self.output = Array() self.global_size = None self.local_size = None self.demand("input", "mean", "rdisp") def init_unpickled(self): super(MeanDispNormalizer, self).init_unpickled() self.sources_["mean_disp_normalizer"] = {} def initialize(self, device, **kwargs): super(MeanDispNormalizer, self).initialize(device, **kwargs) for arr in self.input, self.mean, self.rdisp: if not isinstance(arr, Array): raise TypeError("veles.memory.Array type expected (got %s)" % type(arr)) if not arr: raise ValueError("Invalid Array state") if len(self.input.shape) < 2: raise ValueError("input should be at least 2D") sample_size = self.mean.size if (self.input.sample_size != sample_size or self.rdisp.size != sample_size): raise ValueError( "Sample size of input differs from mean-rdisp size") if not self.output: self.output.reset(numpy.zeros(self.input.shape, self.rdisp.dtype)) else: assert self.output.shape == self.input.shape self.init_vectors(self.input, self.mean, self.rdisp, self.output) def _gpu_init(self): dtype = self.rdisp.dtype sample_size = self.mean.size defines = { "input_type": numpy_dtype_to_opencl(self.input.dtype), "mean_type": numpy_dtype_to_opencl(self.mean.dtype), "SAMPLE_SIZE": sample_size } self.build_program(defines, self.__class__.__name__, dtype=dtype) self.assign_kernel("normalize_mean_disp") self.set_args(self.input, self.mean, self.rdisp, self.output) def ocl_init(self): self._gpu_init() self.global_size = [self.mean.size, self.input.shape[0]] def cuda_init(self): self._gpu_init() self.local_size = 1, 1, 1 self.global_size = self.mean.size, self.input.shape[0], 1 def _gpu_run(self): self.unmap_vectors(self.input, self.mean, self.rdisp, self.output) self.execute_kernel(self.global_size, self.local_size) def ocl_run(self): self._gpu_run() def cuda_run(self): self._gpu_run() def numpy_run(self): self.input.map_read() self.mean.map_read() self.rdisp.map_read() self.output.map_invalidate() dtype = self.output.dtype self.output.matrix[:] = ( self.input.matrix.astype(dtype)[:] - self.mean.plain.astype(dtype)) * self.rdisp.plain
class KohonenForward(KohonenBase, AcceleratedUnit): """Kohonen forward layer. Must be assigned before initialize(): input weights minibatch_offset (if total == True) minibatch_size (if total == True) batch_size (if total == True) argmins speeds up run() if linked from KohonenTrainer Updates after run(): output Creates within initialize(): output Attributes: input: input as batch of samples. weights: the weights of the neurons in Kohonen layer. output: the list of winners. total: if total=True is passed in __init__(), the overall winners table """ def __init__(self, workflow, **kwargs): super(KohonenForward, self).__init__(workflow, **kwargs) self.demand("input", "weights") self.argmins = None self._distances = Array() self.output = Array() self._chunk_size_ = 0 self.weights_transposed = False self.total = Array() if kwargs.get("total", False) else None if self.total is not None: self.minibatch_offset = None self.minibatch_size = None self.batch_size = None def init_unpickled(self): super(KohonenForward, self).init_unpickled() self.sources_["kohonen"] = {"FORWARD": 1} @property def neurons_number(self): return self.weights.mem.shape[0] @property def sample_length(self): return self.weights.mem.shape[1] @property def chunk_size(self): return self._chunk_size_ def initialize(self, device, **kwargs): super(KohonenForward, self).initialize(device=device, **kwargs) assert self.input.mem.shape[1] == self.sample_length batch_size = self.input.mem.shape[0] self.output.reset(numpy.zeros(batch_size, dtype=numpy.int32)) if self.argmins is None: self._distances.reset( numpy.zeros([batch_size, self.neurons_number], dtype=self.weights.mem.dtype)) if self.total is not None: self.total.reset(numpy.zeros(self.batch_size, dtype=numpy.int32)) self._minibatch_offset_ = numpy.zeros(1, dtype=numpy.int32) def ocl_init(self): batch_size = self.input.mem.shape[0] self.output.initialize(self.device) if self.argmins is None: self.input.initialize(self.device) self.weights.initialize(self.device) self._distances.initialize(self.device) elif self.total is None: return if self.total is not None: self.total.initialize(self.device) copy_chunk_size = int( numpy.ceil(batch_size / self.device.max_group_size)) chunk_size = self.neurons_number // self.device.max_group_size if chunk_size < 2: chunk_size = self.neurons_number // 2 + 1 self.argmin_group_size = \ int(numpy.ceil(self.neurons_number / chunk_size)) block_size, vector_opt = self.device.device_info.get_kernel_bs_vo( kernel="matrix_multiplication", dtype=self.input.dtype) defines = { 'BLOCK_SIZE': block_size, 'VECTOR_OPT': int(bool(vector_opt)), 'BATCH': batch_size, 'SAMPLE_LENGTH': self.sample_length, 'NEURONS_NUMBER': self.neurons_number, 'CHUNK_SIZE': chunk_size, 'COPY_CHUNK_SIZE': copy_chunk_size, } if self.weights_transposed: defines['WEIGHTS_TRANSPOSED'] = 1 self.build_program(defines, "%s_%d_%d_%d" % (self.__class__.__name__, batch_size, self.sample_length, self.neurons_number), dtype=self.weights.mem.dtype) if self.total is not None: self._set_total_global_size_ = \ [int(numpy.ceil(batch_size / copy_chunk_size))] self._krn_set_total_ = self.get_kernel("set_total") self._krn_set_total_.set_args(self.output.devmem, cl.skip, self.total.devmem) if self.argmins is not None: return self._krn_distances_ = self.get_kernel("calculate_distances") self._krn_distances_.set_args(self.input.devmem, self.weights.devmem, self._distances.devmem) self._krn_argmin_ = self.get_kernel("calculate_argmin") self._krn_argmin_.set_args(self._distances.devmem, self.output.devmem, None) self._gs_distance = [ roundup(self.neurons_number, block_size), roundup(batch_size, block_size) ] self._ls_distance = [block_size, block_size] def ocl_run(self): self.output.unmap() if self.total is not None: self.total.unmap() if self.argmins is None: self.input.unmap() self.weights.unmap() self.execute_kernel(self._gs_distance, self._ls_distance, self._krn_distances_) self.execute_kernel([self.argmin_group_size], [self.argmin_group_size], self._krn_argmin_) else: self.argmins.unmap() self.argmins.map_read() self.output.map_write() self.output.mem[:] = self.argmins.mem self.output.unmap() self.argmins.unmap() if self.total is not None: self._minibatch_offset_[0] = \ self.minibatch_offset - self.minibatch_size self._krn_set_total_.set_arg(1, self._minibatch_offset_) self.execute_kernel(self._set_total_global_size_, None, self._krn_set_total_) def numpy_run(self): self.output.map_invalidate() if self.argmins is not None: self.argmins.map_read() self.output.mem[:] = self.argmins.mem else: self.input.map_read() self.weights.map_read() if self.total is not None: self.total.map_invalidate() length = self.minibatch_size if self.total is not None \ else self.input.mem.shape[0] for sindex in range(length): if self.argmins is None: dist = self.weights.mem - self.input[sindex] winner = numpy.argmin(self.numpy_linalg_norm(dist)) self.output[sindex] = winner else: winner = self.argmins[sindex] if self.total is not None: index = sindex + self.minibatch_offset - self.minibatch_size self.total[index] = winner
class GradientDescentBase(AcceleratedUnit): """Base class for gradient descent units. Attributes: input: input layer values. output: output layer values. err_output: error to backpropagate. err_input: backpropagated error. weights: weights. bias: bias. batch_size: current minibatch size. learning_rate: gradient descent speed (positive). learning_rate_bias weights_decay: regularization for weights (see l1_vs_l2). weights_decay_bias gradient_moment: moment coefficient for weights. gradient_moment_bias gradient_weights_with_moment: accumulated moment. gradient_bias_with_moment batch_size: effective batch size (if None, get it from y). weights_transposed: assume weights matrix as a transposed one. apply_gradient: will apply gradient. gradient_changed: when True, slave will send gradients to master (assigned to True just before the run call, so it can be set to False inside ocl_run, numpy_run if necessary). ocl_set_const_args: True when constant arguments for the kernel had been changed and need to be set again. """ hide_from_registry = True MAPPING = set() REDUCE_SIZE = 64 # used for updating bias def __init__(self, workflow, **kwargs): kwargs["view_group"] = kwargs.get("view_group", "TRAINER") super(GradientDescentBase, self).__init__(workflow, **kwargs) self.err_input = Array(shallow_pickle=True) self.ocl_set_const_args = True self.weights = None self.bias = None self.demand("input", "err_output") self.learning_rate = kwargs.get("learning_rate", 0.01) self.learning_rate_bias = kwargs.get("learning_rate_bias", self.learning_rate) self.weights_decay = kwargs.get("weights_decay", 0.00005) self.weights_decay_bias = kwargs.get("weights_decay_bias", 0.0) self.l1_vs_l2 = kwargs.get("l1_vs_l2", 0) self.l1_vs_l2_bias = kwargs.get("l1_vs_l2_bias", self.l1_vs_l2) self.gradient_moment = kwargs.get("gradient_moment", 0) self.gradient_moment_bias = kwargs.get("gradient_moment_bias", self.gradient_moment) self.weights_transposed = kwargs.get("weights_transposed", False) self.need_err_input = kwargs.get("need_err_input", True) self.include_bias = kwargs.get("include_bias", True) self.factor_ortho = kwargs.get("factor_ortho", 0) self.col_sums = Array() # for orthogonalization # Current gradient as it is without applying learning_rate etc. self.gradient_weights = Array() self.gradient_bias = Array() # Gradient with applied learning_rate etc. # optionally accumulated from the previous run self.accumulate_gradient = kwargs.get("accumulate_gradient", False) # When accumulate_gradient set to True: # 1. Calculate gd # 2. acc = acc_alpha * gd + acc_beta * acc # 3. gd = gd_alpha * acc + gd_beta * gd # 4. Apply moments to gd # 5. weights += gd if apply_gradient set to True self.acc_alpha = kwargs.get("acc_alpha", 0.0) self.acc_beta = kwargs.get("acc_beta", 0.0) self.gd_alpha = kwargs.get("gd_alpha", 0.0) self.gd_beta = kwargs.get("gd_beta", 1.0) self.accumulated_gradient_weights = Array() self.accumulated_gradient_bias = Array() # Gradient with accumulated moments self.gradient_weights_with_moment = Array() self.gradient_bias_with_moment = Array() # Sets to True when gradient changes self.gradient_changed = False # Gradient will be applied to weights immediately just after computing self.apply_gradient = kwargs.get("apply_gradient", not workflow.is_slave) @property def current_batch_size(self): batch_size = getattr(self, "batch_size", None) if batch_size is None: return self.err_output.mem.shape[0] return int(batch_size) def initialize(self, device, **kwargs): super(GradientDescentBase, self).initialize(device, **kwargs) if self.weights: assert len(self.weights.shape) == 2 self.weights_shape = (tuple(reversed(self.weights.shape)) if self.weights_transposed else self.weights.shape) else: self.weights_shape = None self.learning_rate = kwargs.get("learning_rate", self.learning_rate) self.weights_decay = kwargs.get("weights_decay", self.weights_decay) self.gradient_moment = kwargs.get("gradient_moment", self.gradient_moment) self.learning_rate_bias = kwargs.get("learning_rate_bias", self.learning_rate_bias) self.weights_decay_bias = kwargs.get("weights_decay_bias", self.weights_decay_bias) self.gradient_moment_bias = kwargs.get("gradient_moment_bias", self.gradient_moment_bias) if self.weights: if not self.gradient_weights: self.gradient_weights.reset(numpy.zeros_like(self.weights.mem)) else: assert self.gradient_weights.size == self.weights.size if self.weights and self.accumulate_gradient: if not self.accumulated_gradient_weights: self.accumulated_gradient_weights.reset( numpy.zeros_like(self.weights.mem)) else: assert (self.accumulated_gradient_weights.size == self.weights.size) if self.weights and (self.gradient_moment or not self.is_standalone): if not self.gradient_weights_with_moment: self.gradient_weights_with_moment.reset( numpy.zeros_like(self.weights.mem)) else: assert self.gradient_weights_with_moment.size == \ self.weights.size if (self.include_bias and self.bias and (not self.gradient_bias or self.gradient_bias.size != self.bias.size)): self.gradient_bias.reset(numpy.zeros_like(self.bias.mem)) if (self.include_bias and self.bias and self.accumulate_gradient and (not self.accumulated_gradient_bias or self.accumulated_gradient_bias.size != self.bias.size)): self.accumulated_gradient_bias.reset( numpy.zeros_like(self.bias.mem)) if (self.include_bias and self.bias and (self.gradient_moment_bias or not self.is_standalone)): if not self.gradient_bias_with_moment: self.gradient_bias_with_moment.reset( numpy.zeros_like(self.bias.mem)) else: assert self.gradient_bias_with_moment.size == self.bias.size dtype = self.err_output.dtype if self.need_err_input: if not self.err_input: self.err_input.reset(numpy.zeros(self.input.shape, dtype)) else: assert self.err_input.shape == self.input.shape if self.weights: side = self.weights_shape[0] other = self.weights.size // side if self.factor_ortho: if not self.col_sums: self.col_sums.reset(numpy.zeros(other, dtype=dtype)) else: assert self.col_sums.size == other self.col_sums.initialize(self.device) self.reduce_size = roundup(min(self.reduce_size, other), 32) self.weights.initialize(self.device) for vec in self.bias, self.input, self.err_input: if vec: vec.initialize(self.device) self.init_vectors(self.err_output, self.gradient_weights, self.gradient_bias, self.accumulated_gradient_weights, self.accumulated_gradient_bias, self.gradient_weights_with_moment, self.gradient_bias_with_moment) def gpu_weights_update(self): self.unmap_vectors(self.input, self.err_output, self.weights, self.gradient_weights, self.accumulated_gradient_weights, self.gradient_weights_with_moment) if self.factor_ortho: self.col_sums.unmap() self.execute_kernel(self._global_size_ortho, self._local_size_ortho, self.krn_compute_col_sums_) self._weights_const[12] = self.factor_ortho self.krn_weights_.set_arg(12, self._weights_const[12:13]) self._weights_const[4:12] = (self.learning_rate, self.weights_decay, self.l1_vs_l2, self.gradient_moment, self.acc_alpha, self.acc_beta, self.gd_alpha, self.gd_beta) self.krn_weights_.set_args( self.device.skip(4), self._weights_const[4:5], self._weights_const[5:6], self._weights_const[6:7], self._weights_const[7:8], self._weights_const[8:9], self._weights_const[9:10], self._weights_const[10:11], self._weights_const[11:12]) self.execute_kernel(self._global_size_weights, self._local_size_weights, self.krn_weights_) def gpu_bias_update(self): if not self.include_bias: return self.unmap_vectors(self.err_output, self.bias, self.gradient_bias, self.accumulated_gradient_bias, self.gradient_bias_with_moment) self._bias_const[5:13] = (self.learning_rate_bias, self.weights_decay_bias, self.l1_vs_l2_bias, self.gradient_moment_bias, self.acc_alpha, self.acc_beta, self.gd_alpha, self.gd_beta) self.krn_bias_.set_args(self.device.skip(5), self._bias_const[5:6], self._bias_const[6:7], self._bias_const[7:8], self._bias_const[8:9], self._bias_const[9:10], self._bias_const[10:11], self._bias_const[11:12], self._bias_const[12:13]) self.execute_kernel(self._global_size_bias, self._local_size_bias, self.krn_bias_) def gpu_err_output_update(self): """Multiply err_output by activation derivative by output. """ if self.krn_err_output_ is None: return self.err_output.unmap() self.output.unmap() self.execute_kernel(self._global_size_err_output, self._local_size_err_output, self.krn_err_output_) def numpy_err_output_update(self): """Multiply err_output by activation derivative by output. """ pass def print_debug_data(self): """ Show weights statistics """ if not self.logger.isEnabledFor(logging.DEBUG): return self.weights.map_read() self.bias.map_read() self.gradient_bias.map_read() self.gradient_weights.map_read() weights = self.weights.mem bias = self.bias.mem grad_weights = self.gradient_weights.mem grad_bias = self.gradient_bias.mem weight_table = PrettyTable("TYPE", "Mean", "StdDev", "Min", "Max") weight_table.float_format = ".10" for (w_name, w_array) in [("Weight", weights), ("Bias", bias), ("Grad Weight", grad_weights), ("Grad Bias", grad_bias)]: w_mean = w_stddev = w_min = w_max = None if w_array is not None and w_array.size > 0: w_mean = numpy.mean(w_array) w_stddev = numpy.std(w_array) w_min = numpy.min(w_array) w_max = numpy.max(w_array) weight_table.add_row(w_name, w_mean, w_stddev, w_min, w_max) self.debug("\n" + weight_table.get_string()) def generate_data_for_slave(self, slave): return (self.learning_rate, self.weights_decay, self.gradient_moment, self.learning_rate_bias, self.weights_decay_bias, self.gradient_moment_bias) @staticmethod def fill_zeros(vector): if not vector: return vector.map_invalidate() vector.mem[:] = 0 def apply_data_from_master(self, data): self.learning_rate = data[0] self.weights_decay = data[1] self.gradient_moment = data[2] self.learning_rate_bias = data[3] self.weights_decay_bias = data[4] self.gradient_moment_bias = data[5] self.fill_zeros(self.gradient_weights_with_moment) self.fill_zeros(self.gradient_bias_with_moment) self.fill_zeros(self.gradient_weights) self.fill_zeros(self.gradient_bias) self.fill_zeros(self.accumulated_gradient_weights) self.fill_zeros(self.accumulated_gradient_bias) def generate_data_for_master(self): if not self.gradient_changed: return None self.gradient_changed = False self.gradient_weights_with_moment.map_read() self.gradient_bias_with_moment.map_read() return (self.gradient_weights_with_moment.mem, self.gradient_bias_with_moment.mem) def apply_data_from_slave(self, data, slave): if self.weights: self.weights.map_write() self.gradient_weights_with_moment.map_write() self.gradient_weights_with_moment.mem *= self.gradient_moment self.gradient_weights_with_moment.mem += data[0] self.weights.mem += self.gradient_weights_with_moment.mem if self.bias: self.bias.map_write() self.gradient_bias_with_moment.map_write() self.gradient_bias_with_moment.mem *= self.gradient_moment_bias self.gradient_bias_with_moment.mem += data[1] self.bias.mem += self.gradient_bias_with_moment.mem def drop_slave(self, slave): pass def accumulate_gradient_f(self, accumulated_gradient, gradient): if accumulated_gradient and self.accumulate_gradient: accumulated_gradient[:] = ( gradient * self.acc_alpha + (self.acc_beta * accumulated_gradient if self.acc_beta else 0)) gradient *= self.gd_beta gradient += self.gd_alpha * accumulated_gradient return gradient @staticmethod def numpy_gradient_step(weight, gradient, lr, factor_l12, l1_vs_l2, factor_ortho=0, weights_transposed=False): gradient = gradient.copy() gradient += factor_l12 * ( (1.0 - l1_vs_l2) * weight + 0.5 * l1_vs_l2 * numpy.sign(weight)) if factor_ortho: col_sums = (reshape_transposed(weight).sum( axis=1) if weights_transposed else weight.sum(axis=0)) for i, row in enumerate(gradient): row += (col_sums - weight[i]) * factor_ortho / weight.shape[0] gradient *= lr return gradient def run(self): self.gradient_changed = True super(GradientDescentBase, self).run() self.ocl_set_const_args = False
class Deconv(TriviallyDistributable, ConvolutionalBase, nn_units.Forward): # TriviallyDistributable overrides nn_units.Forward IDistributable """Deconvolutional layer for simple convolutional layer with linear activation and without bias. Must be assigned before initialize(): input weights output_shape_source Updates after run(): output Creates within initialize(): output Attributes: input: input as batch of multichannel interleaved images. output: output as batch of multichannel interleaved images. weights: matrix of weights. output_shape_source: Array to get output shape from. n_kernels: number of convolutional kernels in the corresponding convolutional layer. kx: kernel width. ky: kernel height. sliding: tuple of kernel sliding (by x-axis, by y-axis), kx, ky MUST be a multiple of sliding to avoid irregularities. padding: tuple of virtual sample padding (left, top, right, bottom), will be computed automatically based on sliding. weights_transposed: assume weights matrix as a transposed one. unsafe_padding: flag to enable unsafe padding and/or sliding. """ MAPPING = {"deconv"} @staticmethod def compute_padding(sx, sy, kx, ky, sliding): """Computes required padding. """ return (kx - sliding[1], ky - sliding[0], kx - sx % sliding[1] if sx % sliding[1] != 0 else kx - sliding[1], ky - sy % sliding[0] if sy % sliding[0] != 0 else ky - sliding[0]) @staticmethod def check_padding_is_safe(kx, ky, sliding): if sliding[0] > (ky >> 1) or sliding[1] > (kx >> 1): raise ValueError( "sliding should not be greater than half of the kernel size") if kx % sliding[0] != 0 or kx % sliding[1] != 0: raise ValueError( "Kernel size should be multiple of sliding") def __init__(self, workflow, **kwargs): super(Deconv, self).__init__(workflow, **kwargs) self.unsafe_padding = kwargs.get("unsafe_padding", False) self.hits = Array() self.krn_clear_output_ = None self._global_size = None self._local_size = None del self.bias self.demand("n_kernels", "kx", "ky", "padding", "sliding", "input", "weights", "output_shape_source") def init_unpickled(self): super(Deconv, self).init_unpickled() self.sources_["deconv/forward"] = {} def initialize(self, device, **kwargs): super(Deconv, self).initialize(device, **kwargs) self._dtype = self.input.dtype self.weights_shape = (tuple(reversed(self.weights.shape)) if self.weights_transposed else self.weights.shape) if hasattr(self, "bias"): raise ValueError("bias should not be set") if (len(self.input.shape) != 4 or self.input.shape[3] != self.n_kernels): raise ValueError("Incorrectly shaped input encountered") if (len(self.weights_shape) != 2 or self.weights_shape[0] != self.n_kernels or self.weights_shape[1] % (self.kx * self.ky) != 0): raise ValueError("Incorrectly shaped weights encountered") output_shape = tuple(self.output_shape_source.shape) if len(output_shape) != 4: raise ValueError("Incorrect output_shape_source shape") if output_shape[0] != self.input.shape[0]: raise ValueError( "output_shape_source.shape[0] != input.shape[0]") try: self.check_padding_is_safe(self.kx, self.ky, self.sliding) except ValueError as e: if not self.unsafe_padding: raise from_none(e) self.warning("The padding will be unsafe") self._create_hits(output_shape) padding = Deconv.compute_padding( output_shape[2], output_shape[1], self.kx, self.ky, self.sliding) if self.padding is None: # pylint: disable=E0203 self.padding = padding elif self.padding != padding: if not self.unsafe_padding: raise ValueError( "Expected padding %s but got %s" % (padding, self.padding)) self._create_hits(output_shape) if self.output: assert self.output.shape[1:] == output_shape[1:] if not self.output or self.output.shape[0] != output_shape[0]: self.output.reset(numpy.zeros(output_shape, dtype=self._dtype)) self._output_shape = output_shape self._sy, self._sx, self._n_channels = self._output_shape[1:] self._kernel_size = self.kx * self.ky * self._n_channels self._kernel_app_per_image = self.input.sample_size // self.n_kernels self._kernel_app_total = (self._kernel_app_per_image * self.input.shape[0]) self.init_vectors(self.input, self.weights, self.output, self.hits) def _create_hits(self, output_shape): if not self.hits: self.hits.reset( numpy.zeros(output_shape, dtype=numpy.int32)) else: assert self.hits.size == int(numpy.prod(output_shape)) def _gpu_init(self, blas_class): defines = { "USE_ATOMICS": 1, "WEIGHTS_TRANSPOSED": int(self.weights_transposed), "BATCH": self._output_shape[0], "SX": self._sx, "SY": self._sy, "N_CHANNELS": self._n_channels, "KX": self.kx, "KY": self.ky, "N_KERNELS": self.n_kernels, "PAD_LEFT": self.padding[0], "PAD_TOP": self.padding[1], "PAD_RIGHT": self.padding[2], "PAD_BOTTOM": self.padding[3], "SLIDE_X": self.sliding[0], "SLIDE_Y": self.sliding[1], "USE_HITS": int(bool(self.hits)), "DECONV_MODE": int(bool(self.hits)) + 1, "OUTPUT_SIZE": self.output.size } self.build_program( defines, "%s/%s_%d_%dx%dx%d_%dx%d_%d" % ( root.common.dirs.cache, self.__class__.__name__, self.input.shape[0], self._output_shape[2], self._output_shape[1], self._output_shape[3], self.kx, self.ky, self.n_kernels), dtype=self._dtype) self.krn_pack_ = self.get_kernel("DirectPack") unpack_bytes = (self._kernel_app_per_image * self.unpack_size * self._kernel_size * self.input.itemsize) self.device.request_temp_buffer(unpack_bytes) if self.hits: self.krn_pack_.set_arg(3, self.hits.devmem) self.krn_apply_hits_ = self.get_kernel("apply_hits") self.krn_apply_hits_.set_args(self.output.devmem, self.hits.devmem) self.gemm_ = blas_class.gemm(self._dtype) self.np_one = numpy.ones(1, dtype=self._dtype) self.np_zero = numpy.zeros(1, dtype=self._dtype) self._const_i = numpy.zeros(1, dtype=numpy.int64) def ocl_init(self): ocl_blas.OCLBLAS.attach_to_device(self.device) self._gpu_init(ocl_blas.OCLBLAS) self._global_size_pack = lambda size: (size,) self._local_size_pack = None if self.hits: self.krn_clear_hits_ = self.get_kernel("clear_hits") self.krn_clear_hits_.set_arg(0, self.hits.devmem) self._global_size_hits = (self.output.size,) self._local_size_hits = None self.krn_clear_output_ = self.get_kernel("clear_output") self.krn_clear_output_.set_arg(0, self.output.devmem) self._clear_output = lambda: ( self.execute_kernel((self.output.size,), None, self.krn_clear_output_)) self._clear_hits = lambda: ( self.execute_kernel((self.hits.size,), None, self.krn_clear_hits_)) self._process_subblock = self._ocl_process_subblock self.krn_pack_.set_arg(1, self.output.devmem) def cuda_init(self): self._gpu_init(cublas.CUBLAS) block_size = self.device.suggest_block_size(self.krn_pack_) self._global_size_pack = ( lambda size: (int(numpy.ceil(size / block_size)), 1, 1)) self._local_size_pack = (block_size, 1, 1) if self.hits: block_size = self.device.suggest_block_size(self.krn_apply_hits_) self._global_size_hits = ( int(numpy.ceil(self.output.size / block_size)), 1, 1) self._local_size_hits = (block_size, 1, 1) self._clear_output = lambda: self.output.devmem.memset32_async() self._clear_hits = lambda: self.hits.devmem.memset32_async() self._process_subblock = self._cuda_process_subblock def ocl_run(self): self.gpu_run() def cuda_run(self): self.gpu_run() def gpu_run(self): self.unmap_vectors(self.output, self.input, self.weights) unpack_data = self.device.get_temp_buffer() self._clear_output() if self.hits: self.hits.unmap() self._clear_hits() batch_size = self.output.shape[0] for i in range(0, batch_size, self.unpack_size): self._process_subblock(i, min(batch_size - i, self.unpack_size), unpack_data) if self.hits: self.execute_kernel(self._global_size_hits, self._local_size_hits, self.krn_apply_hits_) def _cuda_process_subblock(self, start_image, image_count, unpack_data): output_offs = (start_image * self.input.sample_size * self.input.itemsize) unpack_side = self._kernel_app_per_image * image_count self.gemm_( self.device.blas, cublas.CUBLAS_OP_T if self.weights_transposed else cublas.CUBLAS_OP_N, cublas.CUBLAS_OP_N, self._kernel_size, unpack_side, self.weights_shape[0], self.np_one, self.weights.devmem, int(self.input.devmem) + output_offs, self.np_zero, unpack_data) self.krn_pack_.set_arg(0, unpack_data) self.krn_pack_.set_arg( 1, int(self.output.devmem) + start_image * self.output.sample_size * self.output.itemsize) limit = unpack_side * self._kernel_size self._const_i[0] = limit self.krn_pack_.set_arg(2, self._const_i) self.execute_kernel(self._global_size_pack(limit), self._local_size_pack, self.krn_pack_) def _ocl_process_subblock(self, start_image, image_count, unpack_data): output_offs = start_image * self.input.sample_size unpack_side = self._kernel_app_per_image * image_count self.gemm_( self.device.blas, cublas.CUBLAS_OP_T if self.weights_transposed else cublas.CUBLAS_OP_N, cublas.CUBLAS_OP_N, self._kernel_size, unpack_side, self.weights_shape[0], self.np_one, self.weights.devmem, self.input.devmem, self.np_zero, unpack_data, offsetB=output_offs) self.krn_pack_.set_arg(0, unpack_data) self._const_i[0] = start_image * self.output.sample_size self.krn_pack_.set_arg(2, self._const_i) limit = unpack_side * self._kernel_size self.execute_kernel(self._global_size_pack(limit), self._local_size_pack, self.krn_pack_) def numpy_run(self): raise NotImplementedError()
class EvaluatorMSE(EvaluatorBase): MAPPING = "evaluator_mse" LOSS = "mse" """Evaluator for nn softmax output from the batch labels. Must be assigned before initialize(): output target batch_size labels (may be None) class_targets (may be None) Updates after run(): err_output confusion_matrix max_err_output_sum n_err (only if labels and class_targets is not None) Creates within initialize(): err_output n_err (only if labels and class_targets is not None) max_err_output_sum Attributes: output: output of the network_common as Batch. target: target for the current Batch. err_output: backpropagation errors. batch_size: number of elements in output to evaluate. metrics: [0] - sum of sample's mse, [1] - max of sample's mse, [2] - min of sample's mse. mse: array of mse for each sample in minibatch. krn_constants_i_: numpy array for constant arguments to kernel. labels: labels for a batch (may be None). class_targets: target for each class (may be None). n_err: number of wrongly recognized samples (if labels and class_targets is not None). """ def __init__(self, workflow, **kwargs): super(EvaluatorMSE, self).__init__(workflow, **kwargs) self.metrics = Array() self.mse = Array() self.labels = None self.class_targets = None self.n_err = Array() self.root = kwargs.get("root", True) self.demand("target", "normalizer") @property def root(self): """ :return: True if error metric is RMSE, otherwise, MSE (mean sum of squares). Default is True. """ return self._root @root.setter def root(self, value): if not isinstance(value, bool): raise TypeError("root must be boolean (got %s)" % type(value)) self._root = value def initialize(self, device, **kwargs): super(EvaluatorMSE, self).initialize(device=device, **kwargs) if self.testing: return if self.target.size != self.output.size: raise error.BadFormatError( "target.size != output.size (%s != %s)" % (self.target.size, self.output.size)) self.sources_["evaluator_mse"] = {} self.sources_["denormalization"] = {} dtype = self.output.dtype self.metrics.reset(numpy.zeros(3, dtype=dtype)) self.metrics[2] = 1.0e30 # mse_min self.mse.reset(numpy.zeros(self.err_output.mem.shape[0], dtype)) self.n_err.reset(numpy.zeros(2, dtype=numpy.int32)) self.init_vectors(self.n_err, self.target, self.metrics, self.mse) if self.class_targets: self.class_targets.initialize(self.device) def _gpu_init(self): dtype = self.output.dtype block_size = min(self.err_output.shape[0], 128) if self.class_targets: self.sources_["mse_find_closest"] = { "target_dtype": numpy_dtype_to_opencl(self.class_targets.dtype) } self.build_program( cache_file_name="%s_%d_%d" % (self.__class__.__name__, self.output.shape[0], self.output.sample_size), dtype=dtype, max_batch_size=self.err_output.shape[0], block_size=block_size, output_size=self.err_output.sample_size, root=self.root, normalization=self.normalizer.MAPPING, targets_number=self.class_targets.shape[0] if self.class_targets else None, coeffs=self.normalizer.coefficients) self.assign_kernel("evaluate_mse") self.set_args(self.output, self.target, self.skip_args(2), self.metrics, self.mse.devmem, self.err_output) if self.labels and self.class_targets: assert(self.labels.dtype == self.n_err.dtype == numpy.int32) self.krn_find_closest_ = self.get_kernel("mse_find_closest") self.krn_find_closest_.set_args( self.output.devmem, self.class_targets.devmem, self.labels.devmem, self.n_err.devmem) return block_size def ocl_init(self): if self.testing: return block_size = self._gpu_init() self._local_size = [block_size] self._global_size = self._local_size self._global_size_find_closest_ = lambda: (self.batch_size,) self._local_size_find_closest = None def cuda_init(self): if self.testing: return block_size = self._gpu_init() self._local_size = (block_size, 1, 1) self._global_size = (1, 1, 1) self._global_size_find_closest_ = lambda: (self.batch_size, 1, 1) self._local_size_find_closest = (1, 1, 1) def _gpu_run(self): self.unmap_vectors(self.err_output, self.output, self.target, self.metrics, self.mse) batch_size = self.batch_size self.krn_constants_i_[0] = batch_size self.set_arg(2, self.krn_constants_i_[0:1]) self.krn_constants_f_[0] = 1.0 / self.batch_size if self.mean else 1.0 self.set_arg(3, self.krn_constants_f_[0:1]) self.execute_kernel(self._global_size, self._local_size) if self.labels and self.class_targets: self.unmap_vectors(self.class_targets, self.labels, self.n_err) self.execute_kernel(self._global_size_find_closest_(), self._local_size_find_closest, self.krn_find_closest_) self.n_err.map_write() self.n_err.mem[1] += batch_size def ocl_run(self): return self._gpu_run() def cuda_run(self): return self._gpu_run() def numpy_run(self): self.output.map_read() self.target.map_read() self.metrics.map_write() self.err_output.map_invalidate() self.mse.map_invalidate() assert(self.output.size == self.target.size == self.err_output.size) batch_size = self.batch_size err_output = self.err_output.matrix[:batch_size] assert_addr(err_output, self.err_output.mem) output = self.output.matrix[:batch_size] assert_addr(output, self.output.mem) target = self.target.matrix[:batch_size] assert_addr(target, self.target.mem) mse = self.mse.mem[:batch_size] assert_addr(mse, self.mse.mem) err_output[:] = output - target if not isinstance(self.normalizer, NoneNormalizer): output_copy = output.copy() target_copy = target.copy() self.normalizer.denormalize(output_copy) self.normalizer.denormalize(target_copy) denormed_err_output = output_copy - target_copy else: denormed_err_output = err_output self.err_output.mem[batch_size:] = 0 mse[:] = numpy.square(denormed_err_output).sum(axis=1) / \ denormed_err_output.shape[1] if self.mean: err_output /= batch_size if self.root: numpy.sqrt(mse, mse) self.mse.mem[batch_size:] = 0 self.metrics.mem[0] += mse.sum() self.metrics.mem[1] = max(self.metrics.mem[1], mse.max()) self.metrics.mem[2] = min(self.metrics.mem[2], mse.min()) if self.labels and self.class_targets: self.class_targets.map_read() self.labels.map_read() self.n_err.map_write() class_targets = self.class_targets.matrix labels = self.labels.mem for i, sample in enumerate(output): lbl = numpy.linalg.norm(class_targets - sample, axis=1).argmin() if lbl != labels[i]: self.n_err.mem[0] += 1 self.n_err.mem[1] += 1 def merge_output(self): if not isinstance(self.normalizer, NoneNormalizer): output = self.output[:self.batch_size].copy() self.normalizer.denormalize(output) else: output = self.output.mem self.merged_output[self.offset - self.batch_size:self.offset] = output
class KohonenTrainer(KohonenBase, AcceleratedUnit): """KohonenForward train pass. Must be assigned before initialize(): input shape Creates within initialize(): weights winners argmins _distances _coords Updates after run(): weights Attributes: weights: weights of the current layer. input: input of the current layer as batch of 1D samples. krn_dist_: computes distances between input and neuron weights. _krn_argmin_: finds indexes of minimal computed distances. krn_gravity_: computes gravity to the winner neuron. krn_apply_gradients_: applies gradient to weights. """ def __init__(self, workflow, **kwargs): super(KohonenTrainer, self).__init__(workflow, **kwargs) self._distances = Array() self.argmins = Array() self._coords = Array() self.weights = Array() self.winners = Array() self.weights_filling = kwargs.get("weights_filling", "uniform") self.weights_stddev = kwargs.get("weights_stddev", None) self.weights_transposed = kwargs.get("weights_transposed", False) self.time = 0 self._sigma = 0 self.gradient_decay = kwargs.get("gradient_decay", lambda t: 0.1 / (1.0 + t * 0.05)) self.radius_decay = kwargs.get("radius_decay", lambda t: 1.0 / (1.0 + t * 0.05)) self.demand("input", "shape") self._shape = kwargs.get("shape") def init_unpickled(self): super(KohonenTrainer, self).init_unpickled() self.sources_["kohonen"] = {"TRAIN": 1} self._krn_distances_ = None self._krn_argmin_ = None self._krn_gravity_ = None self._krn_compute_gradients_ = None self._krn_apply_gradients_ = None @property def gravity_radius(self): return self.radius_decay(self.time) * self._sigma @property def gradient_multiplier(self): return self.gradient_decay(self.time) @property def shape(self): return self._shape @shape.setter def shape(self, value): self._shape = value def initialize(self, device, **kwargs): super(KohonenTrainer, self).initialize(device=device, **kwargs) self._neurons_number = self.shape[0] * self.shape[1] self._sample_length = self.input.mem.size // self.input.mem.shape[0] # Initialize weights if self.weights_stddev is None: # Get weights magnitude and cap it to 0.05 self.weights_stddev = min(self._get_weights_magnitude(), 0.05) weights_size = (self._sample_length * self._neurons_number) if not self.weights: self.weights.reset(numpy.zeros(weights_size, dtype=self.input.mem.dtype)) filling = { "uniform": lambda rand: rand.fill( self.weights.mem, -self.weights_stddev, self.weights_stddev), "gaussian": lambda rand: rand.fill_normal_real( self.weights.mem, 0, self.weights_stddev) } filling[self.weights_filling](prng.get()) self.weights.mem = self.weights.mem.reshape(( self._neurons_number, self._sample_length)) else: assert self.weights.shape == (self._neurons_number, self._sample_length) if self.weights_transposed: # Reshape weights as a matrix: wtrncopy = self.weights.mem.transpose().copy() self.weights.mem.shape = wtrncopy.shape self.weights.mem[:] = wtrncopy[:] self._sample_length = \ self.weights.mem.shape[0 if self.weights_transposed else 1] # Initialize winners self.winners.reset(numpy.zeros(self._neurons_number, numpy.int32)) # Initialize distances batch_size = self.input.mem.shape[0] self._distances.reset(numpy.zeros( [batch_size, self._neurons_number], dtype=self.weights.mem.dtype)) self.argmins.reset(numpy.zeros(batch_size, dtype=numpy.int32)) self._coords.reset(numpy.zeros([self._neurons_number, 2], dtype=self.weights.mem.dtype)) sz = self._neurons_number rows = int(numpy.round(numpy.sqrt(sz))) cols = sz // rows if sz % rows != 0: cols += 1 x_min = -1.0 x_max = 1.0 y_min = -1.0 y_max = 1.0 x_step = (x_max - x_min) / (cols - 1) if cols > 1 else 0 y = y_min y_step = (y_max - y_min) / (rows - 1) if rows > 1 else 0 offs = 0 mem = self._coords.mem for _row in range(rows): x = x_min + (x_step * 0.5 if _row & 1 else 0) for _col in range(cols): mem[offs, 0] = x mem[offs, 1] = y offs += 1 x += x_step y += y_step self._sigma = (self._coords.mem.ravel().max() - self._coords.mem.ravel().min()) * 1.42 def ocl_init(self): self.input.initialize(self.device) self.weights.initialize(self.device) self.winners.initialize(self.device) self.argmins.initialize(self.device) self._distances.initialize(self.device) self._coords.initialize(self.device) batch_size = self.input.mem.shape[0] chunk_size = self._neurons_number // self.device.max_group_size if chunk_size < 2: chunk_size = self._neurons_number // 2 + 1 self.argmin_group_size = int(numpy.ceil(float(self._neurons_number) / chunk_size)) block_size, vector_opt = self.device.device_info.get_kernel_bs_vo( kernel="matrix_multiplication", dtype=self.input.dtype) defines = { 'BLOCK_SIZE': block_size, 'VECTOR_OPT': int(bool(vector_opt)), 'BATCH': batch_size, 'SAMPLE_LENGTH': self._sample_length, 'NEURONS_NUMBER': self._neurons_number, 'CHUNK_SIZE': chunk_size, 'GRADIENT_CHUNK_SIZE': self.device.max_group_size, 'coord_type': "%s%d" % (opencl_types.numpy_dtype_to_opencl(self._coords.mem.dtype), self._coords.mem.shape[-1]) } if self.weights_transposed: defines['WEIGHTS_TRANSPOSED'] = 1 self.build_program(defines, "%s_%d_%d_%d" % (self.__class__.__name__, batch_size, self._sample_length, self._neurons_number), dtype=self.weights.mem.dtype) self.ocl_consts_ = numpy.zeros(1, dtype=self.weights.mem.dtype) self._krn_distances_ = self.get_kernel("calculate_distances") self._krn_distances_.set_args(self.input.devmem, self.weights.devmem, self._distances.devmem) self._krn_argmin_ = self.get_kernel("calculate_argmin") self._krn_argmin_.set_args(self._distances.devmem, self.argmins.devmem, self.winners.devmem) self._krn_gravity_ = self.get_kernel("compute_gravity") self._krn_gravity_.set_args(self.argmins.devmem, self._coords.devmem) self._krn_gravity_.set_arg(3, self._distances.devmem) self._krn_apply_gradient_ = self.get_kernel("apply_gradient") self._krn_apply_gradient_.set_args(self.input.devmem, self._distances.devmem) self._krn_apply_gradient_.set_arg(3, self.weights.devmem) self._gs_distance = [ roundup(self._neurons_number, block_size), roundup(batch_size, block_size)] self._ls_distance = [block_size, block_size] def iteration(fn): def wrapped(self, *args, **kwargs): result = fn(self, *args, **kwargs) self.time += 1 return result name = getattr(fn, '__name__', getattr(fn, 'func', wrapped).__name__) wrapped.__name__ = name + '_iteration' return wrapped @iteration def numpy_run(self): batch_size = self.input.mem.shape[0] neurons_number = self._neurons_number dists = numpy.empty(neurons_number) gradients = numpy.zeros(self.weights.mem.shape) sigma = self.gravity_radius gmult = self.gradient_multiplier self.input.map_read() self.weights.map_invalidate() self.winners.map_invalidate() for sindex in range(batch_size): dist = self.weights.mem - self.input[sindex] winner = numpy.argmin(self.numpy_linalg_norm(dist)) self.winners[winner] += 1 winner_coords = self._coords.mem[winner] for nindex in range(neurons_number): dist = self._coords.mem[nindex] - winner_coords dists[nindex] = numpy.sum(dist * dist) gravity = numpy.exp(dists / (-2 * sigma * sigma)) gradients += gravity.reshape((1, neurons_number)).transpose() * \ (self.input[sindex] - self.weights.mem) * gmult self.weights.mem += gradients @iteration def ocl_run(self): self.unmap_vectors(self.input, self.weights, self.winners, self._distances, self.argmins, self._coords) batch_size = self.input.mem.shape[0] self.execute_kernel(self._gs_distance, self._ls_distance, self._krn_distances_) self.execute_kernel([self.argmin_group_size], [self.argmin_group_size], self._krn_argmin_) self.ocl_consts_[0] = self.gravity_radius self._krn_gravity_.set_arg(2, self.ocl_consts_[0:1]) self.execute_kernel([batch_size, self._neurons_number], None, self._krn_gravity_) self.ocl_consts_[0] = self.gradient_multiplier self._krn_apply_gradient_.set_arg(2, self.ocl_consts_[0:1]) self.execute_kernel( [int(numpy.ceil(self._sample_length / self.device.max_group_size)), self.device.max_group_size], None, self._krn_apply_gradient_) iteration = staticmethod(iteration) def _get_weights_magnitude(self): """ Returns: weights magnitude for initial random distribution, such that activation function will be near maximum if all input values are at their supposed max value. Doesn't matter for classic Kohonen networks, get values as in All2AllTanh. """ d = self.input.max_supposed * self._sample_length if self.input.mem.dtype in (numpy.complex64, numpy.complex128): return 1.0 / d return 9.0 / d
class ImageLoader(LoaderWithValidationRatio): """Base class for all image loaders. It is generally used for loading large datasets. Attributes: color_space: the color space to which to convert images. Can be any of the values supported by OpenCV, e.g., GRAY or HSV. source_dtype: dtype to work with during various image operations. shape: image shape (tuple) - set after initialize(). Must be overriden in child classes: get_image_label() get_image_info() get_image_data() get_keys() """ def __init__(self, workflow, **kwargs): super(ImageLoader, self).__init__(workflow, **kwargs) self.color_space = kwargs.get("color_space", "RGB") self._source_dtype = numpy.float32 self._original_shape = tuple() self.class_keys = [[], [], []] self.verify_interface(IImageLoader) self.path_to_mean = kwargs.get("path_to_mean", None) self.add_sobel = kwargs.get("add_sobel", False) self.mirror = kwargs.get("mirror", False) # True, False, "random" self.scale = kwargs.get("scale", 1.0) self.scale_maintain_aspect_ratio = kwargs.get( "scale_maintain_aspect_ratio", True) self.rotations = kwargs.get("rotations", (0.0,)) # radians self.crop = kwargs.get("crop", None) self.crop_number = kwargs.get("crop_number", 1) self._background = None self.background_image = kwargs.get("background_image", None) self.background_color = kwargs.get( "background_color", (0xff, 0x14, 0x93)) self.smart_crop = kwargs.get("smart_crop", True) self.minibatch_label_values = Array() @property def source_dtype(self): return self._source_dtype @property def color_space(self): return self._color_space @color_space.setter def color_space(self, value): self._validate_color_space(value) self._color_space = value @Loader.shape.getter def shape(self): """ :return: Final cropped image shape. """ if self.crop is not None: shape = self.crop else: shape = self.uncropped_shape if self.channels_number > 1: shape += (self.channels_number,) return shape @property def uncropped_shape(self): """ :return: Uncropped (but scaled) image shape. """ if not isinstance(self.scale, tuple): if self._original_shape == tuple(): return tuple() return self._scale_shape(self._original_shape)[:2] else: return self.scale @property def original_shape(self): return self._original_shape @original_shape.setter def original_shape(self, value): if value is None: raise ValueError("shape must not be None") if not isinstance(value, tuple): raise TypeError("shape must be a tuple (got %s)" % (value,)) if len(value) not in (2, 3): raise ValueError("len(shape) must be equal to 2 or 3 (got %s)" % (value,)) for i, d in enumerate(value): if not isinstance(d, int): raise TypeError("shape[%d] is not an integer (= %s)" % (i, d)) if d < 1: raise ValueError("shape[%d] < 1 (= %s)" % (i, d)) self._original_shape = value @property def scale(self): return self._scale @scale.setter def scale(self, value): if not isinstance(value, (float, tuple)): raise TypeError("scale must be either float or tuple of two ints" " (got %s of type %s)" % (value, value.__class__)) if isinstance(value, tuple): if len(value) != 2: raise ValueError("scale must have length 2 (not %d in %s)" % (len(value), value)) if not isinstance(value[0], int) or not isinstance(value[1], int): raise ValueError("scale must consist of integers (got %s)" % value) self._scale = value @property def crop(self): return self._crop @crop.setter def crop(self, value): if value is None: self._crop = None return if not isinstance(value, tuple): raise TypeError( "crop must be a tuple of 2 integers or floats (got %s)" % value) if len(value) != 2: raise ValueError("invalid crop length (got %d for %s), must be 2" % (len(value), value)) for i, val in enumerate(value): if not isinstance(val, (int, float)): raise TypeError( "crop[%d] = %s is neither an integer nor a float" % (i, val[i])) if isinstance(val, int) and val < 1: raise ValueError( "crop[%d] = %s is out of range" % (i, val)) if isinstance(val, float): if val <= 0 or val > 1: raise ValueError( "Out of range crop %s: %s" % (("height", "width")[i], val)) self._crop = value @property def crop_number(self): return self._crop_number @crop_number.setter def crop_number(self, value): if not isinstance(value, int): raise TypeError("crop_number must be an integer (got %s)" % value) if value < 1: raise ValueError( "crop_number must be greater than zero (got %d)" % value) if value > 1 and self.crop is None: raise ValueError( "crop parameter is None, refusing to set crop_number") self._crop_number = value @property def smart_crop(self): """ :return: Value indicating whether to crop only around bboxes. """ return self._smart_crop @smart_crop.setter def smart_crop(self, value): if not isinstance(value, bool): raise TypeError("smart_crop must be a boolean value") self._smart_crop = value @property def mirror(self): return self._mirror @mirror.setter def mirror(self, value): if value not in (False, True, "random"): raise ValueError( "mirror must be any of the following: False, True, \"random\"") self._mirror = value @property def rotations(self): return self._rotations @rotations.setter def rotations(self, value): if not isinstance(value, tuple): raise TypeError("rotations must be a tuple (got %s)" % value) for i, rot in enumerate(value): if not isinstance(rot, float): raise TypeError( "rotations[%d] = %s is not a float" % (i, rot)) if rot >= numpy.pi * 2: raise ValueError( "rotations[%d] = %s is greater than 2π" % (i, rot)) self._rotations = tuple(sorted(value)) @property def samples_inflation(self): return (1 if self.mirror is not True else 2) * len(self.rotations) * \ self.crop_number @property def background_image(self): return self._background_image @background_image.setter def background_image(self, value): if isinstance(value, str): with open(value, "rb") as fin: self.background_image = fin elif hasattr(value, "read") and hasattr(value, "seek"): self.background_image = numpy.array(Image.open(value)) elif isinstance(value, numpy.ndarray): if value.shape != self.shape: raise error.BadFormatError( "background_image's shape %s != sample's shape " "%s" % (value.shape, self.shape)) self._background_image = value if getattr(self, "background_color", None) is not None: self.warning( "background_color = %s is ignored in favor of " "background_image", self.background_color) elif value is None: self._background_image = None else: raise ValueError( "background_image must be any of the following: " "file name, file object, numpy array or None") @property def background_color(self): return self._background_color @background_color.setter def background_color(self, value): if value is None: self._background_color = None return if not isinstance(value, tuple): raise TypeError( "background_color must be a tuple (got %s)" % value) if len(value) != self.channels_number: raise ValueError( "background_color must have the same length as the number of " "channels = %d (got length %d for %s)" % (self.channels_number, len(value), value)) for i, col in enumerate(value): if not isinstance(col, int): raise TypeError( "background_color[%d] = %s is not an integer" % (i, col)) if getattr(self, "background_image", None) is not None: self.warning( "background_color = %s is ignored in favor of " "background_image", value) self._background_color = value @property def background(self): if self._background is None: if self.background_image is not None: self._background = self.background_image else: self._background = numpy.zeros(self.shape) self._background[:] = self.background_color return self._background.copy() @property def channels_number(self): channels = COLOR_CHANNELS_MAP[self.color_space] if self.add_sobel: channels += 1 return channels def get_effective_image_info(self, key): info = self.get_image_info(key) if self.scale == 1.0: return info if isinstance(self.scale, tuple): return self.scale, info[1] else: return self._scale_shape(info[0]), info[1] def get_image_bbox(self, key, size): """ Override this method for custom label <-> bbox mapping. :param key: The image key. :param size: The image size (for optimization purposes). :return: (ymin, ymax, xmin, xmax). """ return 0, size[0], 0, size[1] def preprocess_image(self, data, color, crop, bbox): """ Transforms images before serving. :param data: the loaded image data. :param color: The loaded image color space. :param crop: True if must crop the scaled image; otherwise, False. :param bbox: The bounding box of the labeled object. Tuple (ymin, ymax, xmin, xmax). :return: The transformed image data, the label value (from 0 to 1). """ if color != self.color_space: method = getattr( cv2, "COLOR_%s2%s" % (color, self.color_space), None) if method is None: aux_method = getattr(cv2, "COLOR_%s2BGR" % color) try: data = cv2.cvtColor(data, aux_method) except cv2.error as e: self.error("Failed to perform '%s' conversion", aux_method) raise from_none(e) method = getattr(cv2, "COLOR_BGR2%s" % self.color_space) try: data = cv2.cvtColor(data, method) except cv2.error as e: self.error("Failed to perform '%s' conversion", method) raise from_none(e) if self.add_sobel: data = self.add_sobel_channel(data) if self.scale != 1.0: data, bbox = self.scale_image(data, bbox) if crop and self.crop is not None: data, label_value = self.crop_image(data, bbox) else: label_value = 1 return data, label_value, bbox def scale_image(self, data, bbox): bbox = numpy.array(bbox, float) if self.scale_maintain_aspect_ratio: if data.shape[1] >= data.shape[0]: dst_width = self.uncropped_shape[:2][1] dst_height = int(numpy.round( float(dst_width) * data.shape[0] / data.shape[1])) else: dst_height = self.uncropped_shape[:2][0] dst_width = int(numpy.round( float(dst_height) * data.shape[1] / data.shape[0])) dst_x_min = int( numpy.round( 0.5 * (self.uncropped_shape[:2][1] - dst_width))) dst_y_min = int( numpy.round( 0.5 * (self.uncropped_shape[:2][0] - dst_height))) data = cv2.resize( data, (dst_width, dst_height), interpolation=cv2.INTER_CUBIC) dst_x_max = dst_x_min + data.shape[1] dst_y_max = dst_y_min + data.shape[0] sample = self.background sample[dst_y_min:dst_y_max, dst_x_min:dst_x_max] = data data = sample.copy() bbox[:2] *= (dst_y_max - dst_y_min) / (bbox[1] - bbox[0]) bbox[:2] += dst_y_min bbox[2:] *= (dst_x_max - dst_x_min) / (bbox[3] - bbox[2]) bbox[2:] += dst_x_min else: data = cv2.resize( data, tuple(reversed(self.uncropped_shape[:2])), interpolation=cv2.INTER_CUBIC) bbox[:2] *= self.uncropped_shape[0] / (bbox[1] - bbox[0]) bbox[2:] *= self.uncropped_shape[1] / (bbox[3] - bbox[2]) return data, tuple(bbox.astype(numpy.int32)) def add_sobel_channel(self, data): original_data = data if self.channels_number == 1 + 1: original_data = original_data.reshape( original_data.shape[:2] + (1,)) elif self.color_space in ("RGB", "BGR", "RGBA", "BGRA"): data = cv2.cvtColor( data, getattr(cv2, "COLOR_%s2GRAY" % self.color_space)) elif self.color_space == "HSV": data = data[:, :, 2] elif self.color_space == "YCR_CB": data = data[:, :, 0] else: raise NotImplementedError( "Conversion from %s to GRAY is not ready" % self.color_space) sobel_xy = tuple(cv2.Sobel(data, cv2.CV_32F, *d, ksize=3) for d in ((1, 0), (0, 1))) sobel_data = numpy.zeros( shape=data.shape + (original_data.shape[2] + 1,), dtype=original_data.dtype) sobel_data[:, :, -1] = numpy.linalg.norm(sobel_xy) sobel_data[:, :, :-1] = original_data return sobel_data def crop_image(self, data, bbox): """ Cuts a rectangular part of an image. :param data: The source image to crop. :param bbox: (ymin, ymax, xmin, xmax) :return: tuple (image part randomly cropped around the bbox,\ intersection ratio) """ crop_hw_yx = [[0, 0], [0, 0]] for i in 0, 1: crop_hw_yx[0][i] = self.crop[i] if isinstance(self.crop[i], int) \ else int(self.crop[i] * data.shape[i]) crop_size = crop_hw_yx[0][i] crop_hw_yx[1][i] = self.prng.randint( max(bbox[i * 2] - crop_size, 0), min(data.shape[i] - crop_size + 1, bbox[i * 2 + 1] + crop_size)) crop_first = crop_hw_yx[1] crop_last = tuple(crop_hw_yx[1][i] + crop_hw_yx[0][i] for i in (0, 1)) crop_bbox = crop_first[0], crop_last[0], crop_first[1], crop_last[1] return data[crop_bbox[0]:crop_bbox[1], crop_bbox[2]:crop_bbox[3]], \ self._intersection(bbox, crop_bbox) def distort(self, data, mirror, rot): if mirror: data = cv2.flip(data, 1) data = numpy.resize(data, data.shape[:2] + (data.shape[-1] + 1,)) data[:, :, -1] = 1 center = tuple(reversed(tuple(data.shape[i] // 2 for i in (0, 1)))) rot_matrix = cv2.getRotationMatrix2D( center, rot * 180 / numpy.pi, 1.0) data = cv2.warpAffine(data, rot_matrix, tuple(reversed(data.shape[:2]))) real = data[:, :, :-1] imag = data[:, :, -1] real *= imag[..., None] real += self.background * (1 - imag)[..., None] return real def get_distortion_by_index(self, index): index //= self.crop_number if self.mirror is True: return index % 2 == 1, self.rotations[index // 2] elif self.mirror == "random": mirror = bool(self.prng.randint(2)) else: mirror = False return mirror, self.rotations[index] def load_keys(self, keys, pbar, data, labels, label_values, crop=True): """Loads data from the specified keys. """ index = 0 has_labels = False for key in keys: obj, label_value, _ = self._load_image(key) label, has_labels = self._load_label(key, has_labels) if (self.crop is None or not crop) and \ obj.shape[:2] != self.uncropped_shape: self.warning( "Ignored %s (label %s): shape %s", key, label, obj.shape[:2]) continue if data is not None: data[index] = obj if labels is not None: labels[index] = label if label_values is not None: label_values[index] = label_value index += 1 if pbar is not None: pbar.inc() return has_labels def load_labels(self): if not self.has_labels: return self.info("Reading labels...") different_labels = defaultdict(int), defaultdict(int), defaultdict(int) label_key_map = defaultdict(list), defaultdict(list), defaultdict(list) pb = ProgressBar(maxval=self.total_samples, term_width=40) pb.start() for class_index in range(3): for key in self.class_keys[class_index]: label, has_labels = self._load_label(key, True) assert has_labels different_labels[class_index][label] += 1 label_key_map[class_index][label].append(key) self._samples_mapping[label].add(key) pb.inc() pb.finish() return different_labels, label_key_map def initialize(self, **kwargs): self._restored_from_pickle_ = kwargs["snapshot"] super(ImageLoader, self).initialize(**kwargs) del self._restored_from_pickle_ def load_data(self): try: super(ImageLoader, self).load_data() except AttributeError: pass if self._restored_from_pickle_: self.info("Scanning for changes...") progress = ProgressBar(maxval=self.total_samples, term_width=40) progress.start() for keys in self.class_keys: for key in keys: progress.inc() size, _ = self.get_effective_image_info(key) if size != self.uncropped_shape: raise error.BadFormatError( "%s changed the effective size (now %s, was %s)" % (key, size, self.uncropped_shape)) progress.finish() return for keys in self.class_keys: del keys[:] for index, class_name in enumerate(CLASS_NAME): keys = set(self.get_keys(index)) self.class_keys[index].extend(keys) self.class_lengths[index] = len(keys) * self.samples_inflation self.class_keys[index].sort() if self.uncropped_shape == tuple(): raise error.BadFormatError( "original_shape was not initialized in get_keys()") self.info( "Found %d samples of shape %s (%d TEST, %d VALIDATION, %d TRAIN)", self.total_samples, self.shape, *self.class_lengths) # Perform a quick (unreliable) test to determine if we have labels keys = next(k for k in self.class_keys if len(k) > 0) self._has_labels = self.load_keys( (keys[RandomGenerator(None).randint(len(keys))],), None, None, None, None) self._resize_validation_keys(self.load_labels()) def create_minibatch_data(self): self.minibatch_data.reset(numpy.zeros( (self.max_minibatch_size,) + self.shape, dtype=self.dtype)) self.minibatch_label_values.reset(numpy.zeros( self.max_minibatch_size, numpy.float32)) def keys_from_indices(self, indices): for index in indices: class_index, origin_index, _ = \ self._get_class_origin_distortion_from_index(index) yield self.class_keys[class_index][origin_index] def fill_minibatch(self): indices = self.minibatch_indices.mem[:self.minibatch_size] assert self.has_labels == self.load_keys( self.keys_from_indices(indices), None, self.minibatch_data.mem, self.raw_minibatch_labels, self.minibatch_label_values) if self.samples_inflation == 1: return for pos, index in enumerate(indices): _, _, dist_index = \ self._get_class_origin_distortion_from_index(index) self.minibatch_data[pos] = self.distort( self.minibatch_data[pos], *self.get_distortion_by_index(dist_index)) def _resize_validation_keys(self, label_analysis): if label_analysis is None: return different_labels, label_key_map = label_analysis if self.validation_ratio is None: self._setup_labels_mapping(different_labels) return if self.validation_ratio < 0: self.class_keys[TRAIN] += self.class_keys[VALID] self.class_lengths[TRAIN] += self.class_lengths[VALID] del self.class_keys[VALID][:] self.class_lengths[VALID] = 0 merged = {k: (different_labels[VALID][k] + different_labels)[TRAIN][k] for k in label_key_map[TRAIN]} self._setup_labels_mapping((different_labels[TEST], {}, merged)) return overall = sum(len(ck) for ck in self.class_keys[VALID:]) target_validation_length = int(overall * self.validation_ratio) if not self.has_labels: keys = list(chain.from_iterable(self.class_keys[VALID:])) keys.sort() self.prng.shuffle(keys) del self.class_keys[VALID][:] self.class_keys[VALID].extend(keys[:target_validation_length]) del self.class_keys[TRAIN][:] self.class_keys[TRAIN].extend(keys[target_validation_length:]) self._finalize_resizing_validation(different_labels, label_key_map) return # We must ensure that each set has the same labels # The first step is to pick two keys for each label and distribute them # into VALID and TRAIN evenly if len(label_key_map[TRAIN]) > target_validation_length: raise LoaderError( "Unable to set the new size of the validation set to %d (%.3f)" " since the number of labels is %d" % (target_validation_length * self.samples_inflation, self.validation_ratio, len(label_key_map[TRAIN]))) if overall - target_validation_length < len(label_key_map[TRAIN]): raise LoaderError( "Unable to set the new size of the training set to %d (%.3f) " "since the number of labels is %d" % ((overall - target_validation_length) * self.samples_inflation, 1.0 - self.validation_ratio, len(label_key_map[TRAIN]))) vt_label_key_map = {l: (label_key_map[VALID].get(l, []) + label_key_map[TRAIN].get(l, [])) for l in label_key_map[TRAIN]} for i in VALID, TRAIN: del self.class_keys[i][:] for label, keys in sorted(vt_label_key_map.items()): if len(keys) < 2: raise LoaderError("Label %s has less than 2 keys" % label) choice = self.prng.choice(len(keys), 2, replace=False) assert choice[0] != choice[1] for i in VALID, TRAIN: self.class_keys[i].append(keys[choice[i - 1]]) for c in sorted(choice, reverse=True): del keys[c] # Distribute the left keys randomly left_keys = list(sorted(chain.from_iterable( vt_label_key_map.values()))) self.prng.shuffle(left_keys) offset_val_length = \ target_validation_length - len(vt_label_key_map) self.class_keys[VALID].extend(left_keys[:offset_val_length]) self.class_keys[TRAIN].extend(left_keys[offset_val_length:]) self._finalize_resizing_validation(different_labels, label_key_map) def _finalize_resizing_validation(self, different_labels, label_key_map): for ck in self.class_keys[VALID:]: ck.sort() for i in VALID, TRAIN: self.class_lengths[i] = len(self.class_keys[i]) * \ self.samples_inflation new_diff = defaultdict(int), defaultdict(int) key_label_map = {} for ci in VALID, TRAIN: key_label_map.update({k: l for l, keys in label_key_map[ci].items() for k in keys}) for ci in VALID, TRAIN: for key in self.class_keys[ci]: new_diff[ci - 1][key_label_map[key]] += 1 self._setup_labels_mapping((different_labels[TEST],) + new_diff) def _get_class_origin_distortion_from_index(self, index): class_index, key_remainder = self.class_index_by_sample_index(index) key_index = self.class_lengths[class_index] - key_remainder return (class_index,) + divmod(key_index, self.samples_inflation) def _load_image(self, key, crop=True): """Returns the data to serve corresponding to the given image key and the label value (from 0 to 1). """ data = self.get_image_data(key) size, color = self.get_image_info(key) bbox = self.get_image_bbox(key, size) return self.preprocess_image(data, color, crop, bbox) def _load_label(self, key, has_labels): label = self.get_image_label(key) if label is not None: has_labels = True if has_labels and label is None: raise error.BadFormatError( "%s does not have a label, but others do" % key) return label, has_labels def _intersection(self, bbox_a, bbox_b): ymin_a, ymax_a, xmin_a, xmax_a = bbox_a ymin_b, ymax_b, xmin_b, xmax_b = bbox_b x_intersection = min(xmax_a, xmax_b) - max(xmin_a, xmin_b) y_intersection = min(ymax_a, ymax_b) - max(ymin_a, ymin_b) if int(x_intersection) | int(y_intersection) <= 0: return 0 else: return x_intersection * y_intersection def _scale_shape(self, shape): return tuple(int(shape[i] * self.scale) for i in (0, 1)) + shape[2:] def _validate_color_space(self, value): if not isinstance(value, str): raise TypeError( "db_colorpsace must be a string (got %s)" % type(value)) if value != "RGB" and not hasattr(cv2, "COLOR_%s2RGB" % value): raise ValueError("Unsupported color space: %s" % value)
class All2AllSoftmax(All2All): """All2All with linear activation and softmax normalization. Must be assigned before initialize(): Updates after run(): max_idx Creates within initialize(): max_idx Attributes: krn_sm_: kernel for softmax activation calculation. max_idx: indexes of element with maximum value for each sample. """ __id__ = "420219fc-3e1a-45b1-87f8-aaa0c1540de4" MAPPING = {"softmax"} def __init__(self, workflow, **kwargs): super(All2AllSoftmax, self).__init__(workflow, **kwargs) self.max_idx = Array() self.reduce_size = 256 def init_unpickled(self): super(All2AllSoftmax, self).init_unpickled() self.krn_sm_ = None self._force_gpu_apply_exp = False def initialize(self, device, **kwargs): self.reduce_size = min(self.reduce_size, int(numpy.prod(self.output_sample_shape))) self.sources_["all2all/softmax"] = { "REDUCE_SIZE": self.reduce_size } retval = super(All2AllSoftmax, self).initialize( device=device, **kwargs) if retval: return retval if self.output.mem.size // self.output.mem.shape[0] <= 1: raise error.BadFormatError( "Output sample size should be greater than 1 for SoftMax.") if not self.max_idx: self.max_idx.reset(numpy.zeros(self.output.shape[0], dtype=numpy.int32)) self.max_idx.initialize(self.device) return retval def numpy_apply_exp(self): self.output.map_write() self.max_idx.map_invalidate() out = self.output.mem out = reshape(out, (out.shape[0], out.size // out.shape[0])) for i, sample in enumerate(out): im = sample.argmax() self.max_idx[i] = im m = sample[im] sample -= m numpy.exp(sample, sample) smm = sample.sum() sample /= smm def ocl_apply_exp(self): self.unmap_vectors(self.output, self.max_idx) global_size = (self.output.shape[0] * self.reduce_size,) local_size = (self.reduce_size,) self.execute_kernel(global_size, local_size, self.krn_sm_) def cuda_apply_exp(self): self.unmap_vectors(self.output, self.max_idx) global_size = (self.output.shape[0], 1, 1) local_size = (self.reduce_size, 1, 1) self.execute_kernel(global_size, local_size, self.krn_sm_) def numpy_run(self): """Forward propagation from batch on CPU only. """ super(All2AllSoftmax, self).numpy_run() if not self._force_gpu_apply_exp: self.numpy_apply_exp() def ocl_run(self): """Forward propagation from batch on GPU. """ self._force_gpu_apply_exp = True super(All2AllSoftmax, self).ocl_run() self.ocl_apply_exp() def cuda_run(self): """Forward propagation from batch on GPU. """ self._force_gpu_apply_exp = True super(All2AllSoftmax, self).cuda_run() self.cuda_apply_exp() def ocl_init(self): super(All2AllSoftmax, self).ocl_init() self.krn_sm_ = self.get_kernel("apply_exp") self.krn_sm_.set_args(self.output.devmem, self.max_idx.devmem) def cuda_init(self): super(All2AllSoftmax, self).cuda_init() self.krn_sm_ = self.get_kernel("apply_exp") self.krn_sm_.set_args(self.output.devmem, self.max_idx.devmem)
class KohonenTrainer(KohonenBase, AcceleratedUnit): """KohonenForward train pass. Must be assigned before initialize(): input shape Creates within initialize(): weights winners argmins _distances _coords Updates after run(): weights Attributes: weights: weights of the current layer. input: input of the current layer as batch of 1D samples. krn_dist_: computes distances between input and neuron weights. _krn_argmin_: finds indexes of minimal computed distances. krn_gravity_: computes gravity to the winner neuron. krn_apply_gradients_: applies gradient to weights. """ def __init__(self, workflow, **kwargs): super(KohonenTrainer, self).__init__(workflow, **kwargs) self._distances = Array() self.argmins = Array() self._coords = Array() self.weights = Array() self.winners = Array() self.weights_filling = kwargs.get("weights_filling", "uniform") self.weights_stddev = kwargs.get("weights_stddev", None) self.weights_transposed = kwargs.get("weights_transposed", False) self.time = 0 self._sigma = 0 self.gradient_decay = kwargs.get("gradient_decay", lambda t: 0.1 / (1.0 + t * 0.05)) self.radius_decay = kwargs.get("radius_decay", lambda t: 1.0 / (1.0 + t * 0.05)) self.demand("input", "shape") self._shape = kwargs.get("shape") def init_unpickled(self): super(KohonenTrainer, self).init_unpickled() self.sources_["kohonen"] = {"TRAIN": 1} self._krn_distances_ = None self._krn_argmin_ = None self._krn_gravity_ = None self._krn_compute_gradients_ = None self._krn_apply_gradients_ = None @property def gravity_radius(self): return self.radius_decay(self.time) * self._sigma @property def gradient_multiplier(self): return self.gradient_decay(self.time) @property def shape(self): return self._shape @shape.setter def shape(self, value): self._shape = value def initialize(self, device, **kwargs): super(KohonenTrainer, self).initialize(device=device, **kwargs) self._neurons_number = self.shape[0] * self.shape[1] self._sample_length = self.input.mem.size // self.input.mem.shape[0] # Initialize weights if self.weights_stddev is None: # Get weights magnitude and cap it to 0.05 self.weights_stddev = min(self._get_weights_magnitude(), 0.05) weights_size = (self._sample_length * self._neurons_number) if not self.weights: self.weights.reset( numpy.zeros(weights_size, dtype=self.input.mem.dtype)) filling = { "uniform": lambda rand: rand.fill(self.weights.mem, -self.weights_stddev, self.weights_stddev), "gaussian": lambda rand: rand.fill_normal_real(self.weights.mem, 0, self. weights_stddev) } filling[self.weights_filling](prng.get()) self.weights.mem = self.weights.mem.reshape( (self._neurons_number, self._sample_length)) else: assert self.weights.shape == (self._neurons_number, self._sample_length) if self.weights_transposed: # Reshape weights as a matrix: wtrncopy = self.weights.mem.transpose().copy() self.weights.mem.shape = wtrncopy.shape self.weights.mem[:] = wtrncopy[:] self._sample_length = \ self.weights.mem.shape[0 if self.weights_transposed else 1] # Initialize winners self.winners.reset(numpy.zeros(self._neurons_number, numpy.int32)) # Initialize distances batch_size = self.input.mem.shape[0] self._distances.reset( numpy.zeros([batch_size, self._neurons_number], dtype=self.weights.mem.dtype)) self.argmins.reset(numpy.zeros(batch_size, dtype=numpy.int32)) self._coords.reset( numpy.zeros([self._neurons_number, 2], dtype=self.weights.mem.dtype)) sz = self._neurons_number rows = int(numpy.round(numpy.sqrt(sz))) cols = sz // rows if sz % rows != 0: cols += 1 x_min = -1.0 x_max = 1.0 y_min = -1.0 y_max = 1.0 x_step = (x_max - x_min) / (cols - 1) if cols > 1 else 0 y = y_min y_step = (y_max - y_min) / (rows - 1) if rows > 1 else 0 offs = 0 mem = self._coords.mem for _row in range(rows): x = x_min + (x_step * 0.5 if _row & 1 else 0) for _col in range(cols): mem[offs, 0] = x mem[offs, 1] = y offs += 1 x += x_step y += y_step self._sigma = (self._coords.mem.ravel().max() - self._coords.mem.ravel().min()) * 1.42 def ocl_init(self): self.input.initialize(self.device) self.weights.initialize(self.device) self.winners.initialize(self.device) self.argmins.initialize(self.device) self._distances.initialize(self.device) self._coords.initialize(self.device) batch_size = self.input.mem.shape[0] chunk_size = self._neurons_number // self.device.max_group_size if chunk_size < 2: chunk_size = self._neurons_number // 2 + 1 self.argmin_group_size = int( numpy.ceil(float(self._neurons_number) / chunk_size)) block_size, vector_opt = self.device.device_info.get_kernel_bs_vo( kernel="matrix_multiplication", dtype=self.input.dtype) defines = { 'BLOCK_SIZE': block_size, 'VECTOR_OPT': int(bool(vector_opt)), 'BATCH': batch_size, 'SAMPLE_LENGTH': self._sample_length, 'NEURONS_NUMBER': self._neurons_number, 'CHUNK_SIZE': chunk_size, 'GRADIENT_CHUNK_SIZE': self.device.max_group_size, 'coord_type': "%s%d" % (opencl_types.numpy_dtype_to_opencl( self._coords.mem.dtype), self._coords.mem.shape[-1]) } if self.weights_transposed: defines['WEIGHTS_TRANSPOSED'] = 1 self.build_program(defines, "%s_%d_%d_%d" % (self.__class__.__name__, batch_size, self._sample_length, self._neurons_number), dtype=self.weights.mem.dtype) self.ocl_consts_ = numpy.zeros(1, dtype=self.weights.mem.dtype) self._krn_distances_ = self.get_kernel("calculate_distances") self._krn_distances_.set_args(self.input.devmem, self.weights.devmem, self._distances.devmem) self._krn_argmin_ = self.get_kernel("calculate_argmin") self._krn_argmin_.set_args(self._distances.devmem, self.argmins.devmem, self.winners.devmem) self._krn_gravity_ = self.get_kernel("compute_gravity") self._krn_gravity_.set_args(self.argmins.devmem, self._coords.devmem) self._krn_gravity_.set_arg(3, self._distances.devmem) self._krn_apply_gradient_ = self.get_kernel("apply_gradient") self._krn_apply_gradient_.set_args(self.input.devmem, self._distances.devmem) self._krn_apply_gradient_.set_arg(3, self.weights.devmem) self._gs_distance = [ roundup(self._neurons_number, block_size), roundup(batch_size, block_size) ] self._ls_distance = [block_size, block_size] def iteration(fn): def wrapped(self, *args, **kwargs): result = fn(self, *args, **kwargs) self.time += 1 return result name = getattr(fn, '__name__', getattr(fn, 'func', wrapped).__name__) wrapped.__name__ = name + '_iteration' return wrapped @iteration def numpy_run(self): batch_size = self.input.mem.shape[0] neurons_number = self._neurons_number dists = numpy.empty(neurons_number) gradients = numpy.zeros(self.weights.mem.shape) sigma = self.gravity_radius gmult = self.gradient_multiplier self.input.map_read() self.weights.map_invalidate() self.winners.map_invalidate() for sindex in range(batch_size): dist = self.weights.mem - self.input[sindex] winner = numpy.argmin(self.numpy_linalg_norm(dist)) self.winners[winner] += 1 winner_coords = self._coords.mem[winner] for nindex in range(neurons_number): dist = self._coords.mem[nindex] - winner_coords dists[nindex] = numpy.sum(dist * dist) gravity = numpy.exp(dists / (-2 * sigma * sigma)) gradients += gravity.reshape((1, neurons_number)).transpose() * \ (self.input[sindex] - self.weights.mem) * gmult self.weights.mem += gradients @iteration def ocl_run(self): self.unmap_vectors(self.input, self.weights, self.winners, self._distances, self.argmins, self._coords) batch_size = self.input.mem.shape[0] self.execute_kernel(self._gs_distance, self._ls_distance, self._krn_distances_) self.execute_kernel([self.argmin_group_size], [self.argmin_group_size], self._krn_argmin_) self.ocl_consts_[0] = self.gravity_radius self._krn_gravity_.set_arg(2, self.ocl_consts_[0:1]) self.execute_kernel([batch_size, self._neurons_number], None, self._krn_gravity_) self.ocl_consts_[0] = self.gradient_multiplier self._krn_apply_gradient_.set_arg(2, self.ocl_consts_[0:1]) self.execute_kernel([ int(numpy.ceil(self._sample_length / self.device.max_group_size)), self.device.max_group_size ], None, self._krn_apply_gradient_) iteration = staticmethod(iteration) def _get_weights_magnitude(self): """ Returns: weights magnitude for initial random distribution, such that activation function will be near maximum if all input values are at their supposed max value. Doesn't matter for classic Kohonen networks, get values as in All2AllTanh. """ d = self.input.max_supposed * self._sample_length if self.input.mem.dtype in (numpy.complex64, numpy.complex128): return 1.0 / d return 9.0 / d
class GDMultiplier(AcceleratedUnit): """Gradient descent for Multiplier. """ def __init__(self, workflow, **kwargs): super(GDMultiplier, self).__init__(workflow, **kwargs) self.err_x = Array() self.err_y = Array() self.demand("x", "y", "err_output") def initialize(self, device, **kwargs): super(GDMultiplier, self).initialize(device, **kwargs) if not self.err_x: self.err_x.reset(numpy.zeros_like(self.x.mem)) else: assert self.err_x.shape == self.x.shape if not self.err_y: self.err_y.reset(numpy.zeros_like(self.y.mem)) else: assert self.err_y.shape == self.y.shape self.init_vectors(self.err_x, self.err_y, self.x, self.y, self.err_output) def init_unpickled(self): super(GDMultiplier, self).init_unpickled() self.sources_["multiplier"] = {} def _gpu_init(self): self.build_program({"OUTPUT_SIZE": self.err_output.size}, "%s_%d" % (self.__class__.__name__, self.err_output.size), dtype=self.x.dtype) self.assign_kernel("multiply_backward") self.set_args(self.x, self.y, self.err_output, self.err_x, self.err_y) def cuda_init(self): self._gpu_init() block_size = self.device.suggest_block_size(self._kernel_) self._global_size = ( int(numpy.ceil(self.err_output.size / block_size)), 1, 1) self._local_size = (block_size, 1, 1) def ocl_init(self): self._gpu_init() self._global_size = (self.err_output.size, 1, 1) self._local_size = None def numpy_init(self): pass # nothing to init def _gpu_run(self): self.unmap_vectors(self.x, self.y, self.err_output, self.err_x, self.err_y) self.execute_kernel(self._global_size, self._local_size) def cuda_run(self): self._gpu_run() def ocl_run(self): self._gpu_run() def numpy_run(self): self.x.map_read() self.y.map_read() self.err_output.map_read() self.err_x.map_invalidate() self.err_y.map_invalidate() numpy.multiply(self.err_output.mem, self.y.mem, self.err_x.mem) numpy.multiply(self.err_output.mem, self.x.mem, self.err_y.mem)
class Deconv(TriviallyDistributable, ConvolutionalBase, nn_units.Forward): # TriviallyDistributable overrides nn_units.Forward IDistributable """Deconvolutional layer for simple convolutional layer with linear activation and without bias. Must be assigned before initialize(): input weights output_shape_source Updates after run(): output Creates within initialize(): output Attributes: input: input as batch of multichannel interleaved images. output: output as batch of multichannel interleaved images. weights: matrix of weights. output_shape_source: Array to get output shape from. n_kernels: number of convolutional kernels in the corresponding convolutional layer. kx: kernel width. ky: kernel height. sliding: tuple of kernel sliding (by x-axis, by y-axis), kx, ky MUST be a multiple of sliding to avoid irregularities. padding: tuple of virtual sample padding (left, top, right, bottom), will be computed automatically based on sliding. weights_transposed: assume weights matrix as a transposed one. unsafe_padding: flag to enable unsafe padding and/or sliding. """ MAPPING = {"deconv"} @staticmethod def compute_padding(sx, sy, kx, ky, sliding): """Computes required padding. """ return (kx - sliding[1], ky - sliding[0], kx - sx % sliding[1] if sx % sliding[1] != 0 else kx - sliding[1], ky - sy % sliding[0] if sy % sliding[0] != 0 else ky - sliding[0]) @staticmethod def check_padding_is_safe(kx, ky, sliding): if sliding[0] > (ky >> 1) or sliding[1] > (kx >> 1): raise ValueError( "sliding should not be greater than half of the kernel size") if kx % sliding[0] != 0 or kx % sliding[1] != 0: raise ValueError("Kernel size should be multiple of sliding") def __init__(self, workflow, **kwargs): super(Deconv, self).__init__(workflow, **kwargs) self.unsafe_padding = kwargs.get("unsafe_padding", False) self.hits = Array() self.krn_clear_output_ = None self._global_size = None self._local_size = None del self.bias self.demand("n_kernels", "kx", "ky", "padding", "sliding", "input", "weights", "output_shape_source") def init_unpickled(self): super(Deconv, self).init_unpickled() self.sources_["deconv/forward"] = {} def initialize(self, device, **kwargs): super(Deconv, self).initialize(device, **kwargs) self._dtype = self.input.dtype self.weights_shape = (tuple(reversed(self.weights.shape)) if self.weights_transposed else self.weights.shape) if hasattr(self, "bias"): raise ValueError("bias should not be set") if (len(self.input.shape) != 4 or self.input.shape[3] != self.n_kernels): raise ValueError("Incorrectly shaped input encountered") if (len(self.weights_shape) != 2 or self.weights_shape[0] != self.n_kernels or self.weights_shape[1] % (self.kx * self.ky) != 0): raise ValueError("Incorrectly shaped weights encountered") output_shape = tuple(self.output_shape_source.shape) if len(output_shape) != 4: raise ValueError("Incorrect output_shape_source shape") if output_shape[0] != self.input.shape[0]: raise ValueError("output_shape_source.shape[0] != input.shape[0]") try: self.check_padding_is_safe(self.kx, self.ky, self.sliding) except ValueError as e: if not self.unsafe_padding: raise from_none(e) self.warning("The padding will be unsafe") self._create_hits(output_shape) padding = Deconv.compute_padding(output_shape[2], output_shape[1], self.kx, self.ky, self.sliding) if self.padding is None: # pylint: disable=E0203 self.padding = padding elif self.padding != padding: if not self.unsafe_padding: raise ValueError("Expected padding %s but got %s" % (padding, self.padding)) self._create_hits(output_shape) if not self.output: self.output.reset(numpy.zeros(output_shape, dtype=self._dtype)) else: assert self.output.shape == output_shape self._output_shape = output_shape self._sy, self._sx, self._n_channels = self._output_shape[1:] self._kernel_size = self.kx * self.ky * self._n_channels self._kernel_app_per_image = self.input.sample_size // self.n_kernels self._kernel_app_total = (self._kernel_app_per_image * self.input.shape[0]) self.init_vectors(self.input, self.weights, self.output, self.hits) def _create_hits(self, output_shape): if not self.hits: self.hits.reset(numpy.zeros(output_shape, dtype=numpy.int32)) else: assert self.hits.size == int(numpy.prod(output_shape)) def _gpu_init(self, blas_class): defines = { "USE_ATOMICS": 1, "WEIGHTS_TRANSPOSED": int(self.weights_transposed), "BATCH": self._output_shape[0], "SX": self._sx, "SY": self._sy, "N_CHANNELS": self._n_channels, "KX": self.kx, "KY": self.ky, "N_KERNELS": self.n_kernels, "PAD_LEFT": self.padding[0], "PAD_TOP": self.padding[1], "PAD_RIGHT": self.padding[2], "PAD_BOTTOM": self.padding[3], "SLIDE_X": self.sliding[0], "SLIDE_Y": self.sliding[1], "USE_HITS": int(bool(self.hits)), "DECONV_MODE": int(bool(self.hits)) + 1, "OUTPUT_SIZE": self.output.size } self.build_program( defines, "%s/%s_%d_%dx%dx%d_%dx%d_%d" % (root.common.dirs.cache, self.__class__.__name__, self.input.shape[0], self._output_shape[2], self._output_shape[1], self._output_shape[3], self.kx, self.ky, self.n_kernels), dtype=self._dtype) self.krn_pack_ = self.get_kernel("DirectPack") unpack_bytes = (self._kernel_app_per_image * self.unpack_size * self._kernel_size * self.input.itemsize) self.device.request_temp_buffer(unpack_bytes) if self.hits: self.krn_pack_.set_arg(3, self.hits.devmem) self.krn_apply_hits_ = self.get_kernel("apply_hits") self.krn_apply_hits_.set_args(self.output.devmem, self.hits.devmem) self.gemm_ = blas_class.gemm(self._dtype) self.np_one = numpy.ones(1, dtype=self._dtype) self.np_zero = numpy.zeros(1, dtype=self._dtype) self._const_i = numpy.zeros(1, dtype=numpy.int64) def ocl_init(self): ocl_blas.OCLBLAS.attach_to_device(self.device) self._gpu_init(ocl_blas.OCLBLAS) self._global_size_pack = lambda size: (size, ) self._local_size_pack = None if self.hits: self.krn_clear_hits_ = self.get_kernel("clear_hits") self.krn_clear_hits_.set_arg(0, self.hits.devmem) self._global_size_hits = (self.output.size, ) self._local_size_hits = None self.krn_clear_output_ = self.get_kernel("clear_output") self.krn_clear_output_.set_arg(0, self.output.devmem) self._clear_output = lambda: (self.execute_kernel( (self.output.size, ), None, self.krn_clear_output_)) self._clear_hits = lambda: (self.execute_kernel( (self.hits.size, ), None, self.krn_clear_hits_)) self._process_subblock = self._ocl_process_subblock self.krn_pack_.set_arg(1, self.output.devmem) def cuda_init(self): self._gpu_init(cublas.CUBLAS) block_size = self.device.suggest_block_size(self.krn_pack_) self._global_size_pack = (lambda size: (int(numpy.ceil(size / block_size)), 1, 1)) self._local_size_pack = (block_size, 1, 1) if self.hits: block_size = self.device.suggest_block_size(self.krn_apply_hits_) self._global_size_hits = (int( numpy.ceil(self.output.size / block_size)), 1, 1) self._local_size_hits = (block_size, 1, 1) self._clear_output = lambda: self.output.devmem.memset32_async() self._clear_hits = lambda: self.hits.devmem.memset32_async() self._process_subblock = self._cuda_process_subblock def ocl_run(self): self.gpu_run() def cuda_run(self): self.gpu_run() def gpu_run(self): self.unmap_vectors(self.output, self.input, self.weights) unpack_data = self.device.get_temp_buffer() self._clear_output() if self.hits: self.hits.unmap() self._clear_hits() batch_size = self.output.shape[0] for i in range(0, batch_size, self.unpack_size): self._process_subblock(i, min(batch_size - i, self.unpack_size), unpack_data) if self.hits: self.execute_kernel(self._global_size_hits, self._local_size_hits, self.krn_apply_hits_) def _cuda_process_subblock(self, start_image, image_count, unpack_data): output_offs = (start_image * self.input.sample_size * self.input.itemsize) unpack_side = self._kernel_app_per_image * image_count self.gemm_( self.device.blas, cublas.CUBLAS_OP_T if self.weights_transposed else cublas.CUBLAS_OP_N, cublas.CUBLAS_OP_N, self._kernel_size, unpack_side, self.weights_shape[0], self.np_one, self.weights.devmem, int(self.input.devmem) + output_offs, self.np_zero, unpack_data) self.krn_pack_.set_arg(0, unpack_data) self.krn_pack_.set_arg( 1, int(self.output.devmem) + start_image * self.output.sample_size * self.output.itemsize) limit = unpack_side * self._kernel_size self._const_i[0] = limit self.krn_pack_.set_arg(2, self._const_i) self.execute_kernel(self._global_size_pack(limit), self._local_size_pack, self.krn_pack_) def _ocl_process_subblock(self, start_image, image_count, unpack_data): output_offs = start_image * self.input.sample_size unpack_side = self._kernel_app_per_image * image_count self.gemm_(self.device.blas, cublas.CUBLAS_OP_T if self.weights_transposed else cublas.CUBLAS_OP_N, cublas.CUBLAS_OP_N, self._kernel_size, unpack_side, self.weights_shape[0], self.np_one, self.weights.devmem, self.input.devmem, self.np_zero, unpack_data, offsetB=output_offs) self.krn_pack_.set_arg(0, unpack_data) self._const_i[0] = start_image * self.output.sample_size self.krn_pack_.set_arg(2, self._const_i) limit = unpack_side * self._kernel_size self.execute_kernel(self._global_size_pack(limit), self._local_size_pack, self.krn_pack_) def numpy_run(self): raise NotImplementedError()
class MeanDispNormalizer(AcceleratedUnit, TriviallyDistributable): """Normalizes multichannel byte images according to dataset mean and dispersion. Attributes: input: minibatch of images (dtype=numpy.uint8, shape[0]=minibatch_size). mean: mean image over the dataset (dtype=numpy.uint8). rdisp: 1.0 / dispersion over the dataset (float datatype). output: normalized float images of the same dtype as rdisp. """ def __init__(self, workflow, **kwargs): kwargs["view_group"] = kwargs.get("view_group", "WORKER") super(MeanDispNormalizer, self).__init__(workflow, **kwargs) self.output = Array() self.global_size = None self.local_size = None self.demand("input", "mean", "rdisp") def init_unpickled(self): super(MeanDispNormalizer, self).init_unpickled() self.sources_["mean_disp_normalizer"] = {} def initialize(self, device, **kwargs): super(MeanDispNormalizer, self).initialize(device, **kwargs) for arr in self.input, self.mean, self.rdisp: if not isinstance(arr, Array): raise TypeError( "veles.memory.Array type expected (got %s)" % type(arr)) if not arr: raise ValueError("Invalid Array state") if len(self.input.shape) < 2: raise ValueError("input should be at least 2D") sample_size = self.mean.size if (self.input.sample_size != sample_size or self.rdisp.size != sample_size): raise ValueError( "Sample size of input differs from mean-rdisp size") if not self.output: self.output.reset(numpy.zeros(self.input.shape, self.rdisp.dtype)) else: assert self.output.shape == self.input.shape self.init_vectors(self.input, self.mean, self.rdisp, self.output) def _gpu_init(self): dtype = self.rdisp.dtype sample_size = self.mean.size defines = { "input_type": numpy_dtype_to_opencl(self.input.dtype), "mean_type": numpy_dtype_to_opencl(self.mean.dtype), "SAMPLE_SIZE": sample_size } self.build_program(defines, self.__class__.__name__, dtype=dtype) self.assign_kernel("normalize_mean_disp") self.set_args(self.input, self.mean, self.rdisp, self.output) def ocl_init(self): self._gpu_init() self.global_size = [self.mean.size, self.input.shape[0]] def cuda_init(self): self._gpu_init() self.local_size = 1, 1, 1 self.global_size = self.mean.size, self.input.shape[0], 1 def _gpu_run(self): self.unmap_vectors(self.input, self.mean, self.rdisp, self.output) self.execute_kernel(self.global_size, self.local_size) def ocl_run(self): self._gpu_run() def cuda_run(self): self._gpu_run() def numpy_run(self): self.input.map_read() self.mean.map_read() self.rdisp.map_read() self.output.map_invalidate() dtype = self.output.dtype self.output.matrix[:] = ( self.input.matrix.astype(dtype)[:] - self.mean.plain.astype(dtype)) * self.rdisp.plain
class Forward(ForwardBase): """Class for forward propagation units. Attributes: input: input layer values. output: output layer values. weights: weights. bias: bias. weights_stddev: magnitude of the random distribution for weights. bias_stddev: magnitude of the random distribution for bias. rand: prng.Rand() object for initial weights generation. """ hide_from_registry = True MAPPING = set() def __init__(self, workflow, **kwargs): kwargs["view_group"] = kwargs.get("view_group", "WORKER") super(Forward, self).__init__(workflow, **kwargs) self.weights_stddev = kwargs.get("weights_stddev") self.bias_stddev = kwargs.get("bias_stddev", self.weights_stddev) self.weights_filling = kwargs.get("weights_filling", "uniform") self.bias_filling = kwargs.get("bias_filling", "uniform") self.rand = kwargs.get("rand", prng.get()) self.weights_transposed = kwargs.get("weights_transposed", False) self.include_bias = kwargs.get("include_bias", True) self.demand("input") self.output = Array(shallow_pickle=True) self.weights = Array() self.bias = Array() self.forward_mode = False self.exports = ["weights", "bias", "include_bias", "weights_transposed"] def package_export(self): data = {} for attr in self.exports: value = getattr(self, attr) if value is not None: if isinstance(value, Array): value.map_read() value = value.mem data[attr] = value return data @property def forward_mode(self): return self._forward_mode @forward_mode.setter def forward_mode(self, value): if not isinstance(value, bool): raise TypeError("forward_mode must be boolean (got %s)" % type(value)) self._forward_mode = value def initialize(self, device, **kwargs): self.forward_mode = kwargs.get("forward_mode", False) super(Forward, self).initialize(device=device, **kwargs) def generate_data_for_slave(self, slave): if self.forward_mode: return None data = [None, None] if self.weights: self.weights.map_read() data[0] = self.weights.mem if self.bias: self.bias.map_read() data[1] = self.bias.mem return data def generate_data_for_master(self): return None def apply_data_from_master(self, data): if self.forward_mode: return if self.weights: self.weights.map_invalidate() numpy.copyto(self.weights.mem, data[0]) else: self.weights.reset(data[0]) if self.bias: self.bias.map_invalidate() numpy.copyto(self.bias.mem, data[1]) else: self.bias.reset(data[1]) def apply_data_from_slave(self, data, slave): pass def drop_slave(self, slave): pass
class OffsetPooling(Pooling): """Pooling by offset forward propagation. Must be assigned before initialize(): Updates after run(): input_offset Creates within initialize(): input_offset Attributes: input_offset: offsets in the input where elements are passed through. """ MAPPING = set() hide_from_registry = True def __init__(self, workflow, **kwargs): super(OffsetPooling, self).__init__(workflow, **kwargs) self.input_offset = Array() self.demand("input") def initialize(self, device, **kwargs): super(OffsetPooling, self).initialize(device=device, **kwargs) if self._no_output: return if self.input_offset: assert self.input_offset.shape[1:] == self.output.shape[1:] if (not self.input_offset or self.input_offset.shape[0] != self.output.shape[0]): self.input_offset.reset(numpy.zeros(self.output.shape, dtype=numpy.int32)) self.input_offset.initialize(self.device) def set_args(self, *args): super(OffsetPooling, self).set_args(self.input, self.output, self.input_offset, *args) def ocl_run(self): self.input_offset.unmap() super(OffsetPooling, self).ocl_run() def cuda_run(self): self.input_offset.unmap() super(OffsetPooling, self).cuda_run() def numpy_run(self): self.input_offset.map_invalidate() super(OffsetPooling, self).numpy_run() def numpy_run_cut(self, cut, coords): batch, y1, x1, ch, out_y, out_x = coords cut_index = self.numpy_run_cut_offset( cut, numpy.ravel_multi_index((batch, out_y, out_x, ch), self.output.shape)) i, j = numpy.unravel_index(cut_index, cut.shape) idx = numpy.ravel_multi_index((batch, y1 + i, x1 + j, ch), self.input.shape) val = numpy.ravel(self.input.mem)[idx] self.input_offset.mem[batch, out_y, out_x, ch] = idx return val
class GradientDescentBase(AcceleratedUnit): """Base class for gradient descent units. Attributes: input: input layer values. output: output layer values. err_output: error to backpropagate. err_input: backpropagated error. weights: weights. bias: bias. batch_size: current minibatch size. learning_rate: gradient descent speed (positive). learning_rate_bias weights_decay: regularization for weights (see l1_vs_l2). weights_decay_bias gradient_moment: moment coefficient for weights. gradient_moment_bias gradient_weights_with_moment: accumulated moment. gradient_bias_with_moment batch_size: effective batch size (if None, get it from y). weights_transposed: assume weights matrix as a transposed one. apply_gradient: will apply gradient. gradient_changed: when True, slave will send gradients to master (assigned to True just before the run call, so it can be set to False inside ocl_run, numpy_run if necessary). ocl_set_const_args: True when constant arguments for the kernel had been changed and need to be set again. """ hide_from_registry = True MAPPING = set() REDUCE_SIZE = 64 # used for updating bias def __init__(self, workflow, **kwargs): kwargs["view_group"] = kwargs.get("view_group", "TRAINER") super(GradientDescentBase, self).__init__(workflow, **kwargs) self.err_input = Array(shallow_pickle=True) self.ocl_set_const_args = True self.weights = None self.bias = None self.demand("input", "err_output") self.learning_rate = kwargs.get("learning_rate", 0.01) self.learning_rate_bias = kwargs.get("learning_rate_bias", self.learning_rate) self.weights_decay = kwargs.get("weights_decay", 0.00005) self.weights_decay_bias = kwargs.get("weights_decay_bias", 0.0) self.l1_vs_l2 = kwargs.get("l1_vs_l2", 0) self.l1_vs_l2_bias = kwargs.get("l1_vs_l2_bias", self.l1_vs_l2) self.gradient_moment = kwargs.get("gradient_moment", 0) self.gradient_moment_bias = kwargs.get("gradient_moment_bias", self.gradient_moment) self.weights_transposed = kwargs.get("weights_transposed", False) self.need_err_input = kwargs.get("need_err_input", True) self.include_bias = kwargs.get("include_bias", True) self.factor_ortho = kwargs.get("factor_ortho", 0) self.col_sums = Array() # for orthogonalization # Current gradient as it is without applying learning_rate etc. self.gradient_weights = Array() self.gradient_bias = Array() # Gradient with applied learning_rate etc. # optionally accumulated from the previous run self.accumulate_gradient = kwargs.get("accumulate_gradient", False) # When accumulate_gradient set to True: # 1. Calculate gd # 2. acc = acc_alpha * gd + acc_beta * acc # 3. gd = gd_alpha * acc + gd_beta * gd # 4. Apply moments to gd # 5. weights += gd if apply_gradient set to True self.acc_alpha = kwargs.get("acc_alpha", 0.0) self.acc_beta = kwargs.get("acc_beta", 0.0) self.gd_alpha = kwargs.get("gd_alpha", 0.0) self.gd_beta = kwargs.get("gd_beta", 1.0) self.accumulated_gradient_weights = Array() self.accumulated_gradient_bias = Array() # Gradient with accumulated moments self.gradient_weights_with_moment = Array() self.gradient_bias_with_moment = Array() # Sets to True when gradient changes self.gradient_changed = False # Gradient will be applied to weights immediately just after computing self.apply_gradient = kwargs.get("apply_gradient", not workflow.is_slave) @property def current_batch_size(self): batch_size = getattr(self, "batch_size", None) if batch_size is None: return self.err_output.mem.shape[0] return int(batch_size) def initialize(self, device, **kwargs): super(GradientDescentBase, self).initialize(device, **kwargs) if self.weights: assert len(self.weights.shape) == 2 self.weights_shape = tuple(reversed(self.weights.shape)) if self.weights_transposed else self.weights.shape else: self.weights_shape = None self.learning_rate = kwargs.get("learning_rate", self.learning_rate) self.weights_decay = kwargs.get("weights_decay", self.weights_decay) self.gradient_moment = kwargs.get("gradient_moment", self.gradient_moment) self.learning_rate_bias = kwargs.get("learning_rate_bias", self.learning_rate_bias) self.weights_decay_bias = kwargs.get("weights_decay_bias", self.weights_decay_bias) self.gradient_moment_bias = kwargs.get("gradient_moment_bias", self.gradient_moment_bias) if self.weights: if not self.gradient_weights: self.gradient_weights.reset(numpy.zeros_like(self.weights.mem)) else: assert self.gradient_weights.size == self.weights.size if self.weights and self.accumulate_gradient: if not self.accumulated_gradient_weights: self.accumulated_gradient_weights.reset(numpy.zeros_like(self.weights.mem)) else: assert self.accumulated_gradient_weights.size == self.weights.size if self.weights and (self.gradient_moment or not self.is_standalone): if not self.gradient_weights_with_moment: self.gradient_weights_with_moment.reset(numpy.zeros_like(self.weights.mem)) else: assert self.gradient_weights_with_moment.size == self.weights.size if self.include_bias and self.bias and (not self.gradient_bias or self.gradient_bias.size != self.bias.size): self.gradient_bias.reset(numpy.zeros_like(self.bias.mem)) if ( self.include_bias and self.bias and self.accumulate_gradient and (not self.accumulated_gradient_bias or self.accumulated_gradient_bias.size != self.bias.size) ): self.accumulated_gradient_bias.reset(numpy.zeros_like(self.bias.mem)) if self.include_bias and self.bias and (self.gradient_moment_bias or not self.is_standalone): if not self.gradient_bias_with_moment: self.gradient_bias_with_moment.reset(numpy.zeros_like(self.bias.mem)) else: assert self.gradient_bias_with_moment.size == self.bias.size dtype = self.err_output.dtype if self.need_err_input: if not self.err_input: self.err_input.reset(numpy.zeros(self.input.shape, dtype)) else: assert self.err_input.shape == self.input.shape if self.weights: side = self.weights_shape[0] other = self.weights.size // side if self.factor_ortho: if not self.col_sums: self.col_sums.reset(numpy.zeros(other, dtype=dtype)) else: assert self.col_sums.size == other self.col_sums.initialize(self.device) self.reduce_size = roundup(min(self.reduce_size, other), 32) self.weights.initialize(self.device) for vec in self.bias, self.input, self.err_input: if vec: vec.initialize(self.device) self.init_vectors( self.err_output, self.gradient_weights, self.gradient_bias, self.accumulated_gradient_weights, self.accumulated_gradient_bias, self.gradient_weights_with_moment, self.gradient_bias_with_moment, ) def gpu_weights_update(self): self.unmap_vectors( self.input, self.err_output, self.weights, self.gradient_weights, self.accumulated_gradient_weights, self.gradient_weights_with_moment, ) if self.factor_ortho: self.col_sums.unmap() self.execute_kernel(self._global_size_ortho, self._local_size_ortho, self.krn_compute_col_sums_) self._weights_const[12] = self.factor_ortho self.krn_weights_.set_arg(12, self._weights_const[12:13]) self._weights_const[4:12] = ( self.learning_rate, self.weights_decay, self.l1_vs_l2, self.gradient_moment, self.acc_alpha, self.acc_beta, self.gd_alpha, self.gd_beta, ) self.krn_weights_.set_args( self.device.skip(4), self._weights_const[4:5], self._weights_const[5:6], self._weights_const[6:7], self._weights_const[7:8], self._weights_const[8:9], self._weights_const[9:10], self._weights_const[10:11], self._weights_const[11:12], ) self.execute_kernel(self._global_size_weights, self._local_size_weights, self.krn_weights_) def gpu_bias_update(self): if not self.include_bias: return self.unmap_vectors( self.err_output, self.bias, self.gradient_bias, self.accumulated_gradient_bias, self.gradient_bias_with_moment, ) self._bias_const[5:13] = ( self.learning_rate_bias, self.weights_decay_bias, self.l1_vs_l2_bias, self.gradient_moment_bias, self.acc_alpha, self.acc_beta, self.gd_alpha, self.gd_beta, ) self.krn_bias_.set_args( self.device.skip(5), self._bias_const[5:6], self._bias_const[6:7], self._bias_const[7:8], self._bias_const[8:9], self._bias_const[9:10], self._bias_const[10:11], self._bias_const[11:12], self._bias_const[12:13], ) self.execute_kernel(self._global_size_bias, self._local_size_bias, self.krn_bias_) def gpu_err_output_update(self): """Multiply err_output by activation derivative by output. """ if self.krn_err_output_ is None: return self.err_output.unmap() self.output.unmap() self.execute_kernel(self._global_size_err_output, self._local_size_err_output, self.krn_err_output_) def numpy_err_output_update(self): """Multiply err_output by activation derivative by output. """ pass def print_debug_data(self): """ Show weights statistics """ if not self.logger.isEnabledFor(logging.DEBUG): return self.weights.map_read() self.bias.map_read() self.gradient_bias.map_read() self.gradient_weights.map_read() weights = self.weights.mem bias = self.bias.mem grad_weights = self.gradient_weights.mem grad_bias = self.gradient_bias.mem weight_table = PrettyTable("TYPE", "Mean", "StdDev", "Min", "Max") weight_table.float_format = ".10" for (w_name, w_array) in [ ("Weight", weights), ("Bias", bias), ("Grad Weight", grad_weights), ("Grad Bias", grad_bias), ]: w_mean = w_stddev = w_min = w_max = None if w_array is not None and w_array.size > 0: w_mean = numpy.mean(w_array) w_stddev = numpy.std(w_array) w_min = numpy.min(w_array) w_max = numpy.max(w_array) weight_table.add_row(w_name, w_mean, w_stddev, w_min, w_max) self.debug("\n" + weight_table.get_string()) def generate_data_for_slave(self, slave): return ( self.learning_rate, self.weights_decay, self.gradient_moment, self.learning_rate_bias, self.weights_decay_bias, self.gradient_moment_bias, ) @staticmethod def fill_zeros(vector): if not vector: return vector.map_invalidate() vector.mem[:] = 0 def apply_data_from_master(self, data): self.learning_rate = data[0] self.weights_decay = data[1] self.gradient_moment = data[2] self.learning_rate_bias = data[3] self.weights_decay_bias = data[4] self.gradient_moment_bias = data[5] self.fill_zeros(self.gradient_weights_with_moment) self.fill_zeros(self.gradient_bias_with_moment) self.fill_zeros(self.gradient_weights) self.fill_zeros(self.gradient_bias) self.fill_zeros(self.accumulated_gradient_weights) self.fill_zeros(self.accumulated_gradient_bias) def generate_data_for_master(self): if not self.gradient_changed: return None self.gradient_changed = False self.gradient_weights_with_moment.map_read() self.gradient_bias_with_moment.map_read() return (self.gradient_weights_with_moment.mem, self.gradient_bias_with_moment.mem) def apply_data_from_slave(self, data, slave): if self.weights: self.weights.map_write() self.gradient_weights_with_moment.map_write() self.gradient_weights_with_moment.mem *= self.gradient_moment self.gradient_weights_with_moment.mem += data[0] self.weights.mem += self.gradient_weights_with_moment.mem if self.bias: self.bias.map_write() self.gradient_bias_with_moment.map_write() self.gradient_bias_with_moment.mem *= self.gradient_moment_bias self.gradient_bias_with_moment.mem += data[1] self.bias.mem += self.gradient_bias_with_moment.mem def drop_slave(self, slave): pass def accumulate_gradient_f(self, accumulated_gradient, gradient): if accumulated_gradient and self.accumulate_gradient: accumulated_gradient[:] = gradient * self.acc_alpha + ( self.acc_beta * accumulated_gradient if self.acc_beta else 0 ) gradient *= self.gd_beta gradient += self.gd_alpha * accumulated_gradient return gradient @staticmethod def numpy_gradient_step(weight, gradient, lr, factor_l12, l1_vs_l2, factor_ortho=0, weights_transposed=False): gradient = gradient.copy() gradient += factor_l12 * ((1.0 - l1_vs_l2) * weight + 0.5 * l1_vs_l2 * numpy.sign(weight)) if factor_ortho: col_sums = reshape_transposed(weight).sum(axis=1) if weights_transposed else weight.sum(axis=0) for i, row in enumerate(gradient): row += (col_sums - weight[i]) * factor_ortho / weight.shape[0] gradient *= lr return gradient def run(self): self.gradient_changed = True super(GradientDescentBase, self).run() self.ocl_set_const_args = False
class EvaluatorSoftmax(EvaluatorBase): MAPPING = "evaluator_softmax" LOSS = "softmax" """Evaluator for nn softmax output from the batch labels. Must be assigned before initialize(): output labels batch_size max_idx Updates after run(): err_output n_err confusion_matrix max_err_output_sum Creates within initialize(): err_output n_err confusion_matrix max_err_output_sum Attributes: labels: labels for Batch. output: output of the network_common as Batch. err_output: backpropagation errors based on labels. batch_size: number of elements in output to evaluate. confusion_matrix: confusion matrix for the output. compute_confusion_matrix: compute confusion matrix or not. max_idx: indexes of element with maximum real value for each sample. max_err_output_sum: maximum of backpropagated error sum by sample. """ def __init__(self, workflow, **kwargs): super(EvaluatorSoftmax, self).__init__(workflow, **kwargs) self.compute_confusion_matrix = kwargs.get("compute_confusion_matrix", True) self.confusion_matrix = Array() self.n_err = Array() self.max_err_output_sum = Array() self.class_keys = None self.demand("labels", "max_idx") if self.testing: self.demand("labels_mapping") def initialize(self, device, **kwargs): super(EvaluatorSoftmax, self).initialize(device=device, **kwargs) if self.testing: return self.sources_["evaluator"] = {} dtype = self.output.dtype if not self.n_err: self.n_err.reset(numpy.zeros(2, dtype=numpy.int32)) else: assert self.n_err.size == 2 out_size = self.output.sample_size if self.compute_confusion_matrix: if not self.confusion_matrix: self.confusion_matrix.reset( numpy.zeros([out_size, out_size], numpy.int32)) else: assert self.confusion_matrix.size == out_size * out_size else: self.confusion_matrix.reset() if not self.max_err_output_sum: self.max_err_output_sum.reset(numpy.zeros(1, dtype)) else: assert self.max_err_output_sum.size == 1 self.init_vectors(self.confusion_matrix, self.n_err, self.max_idx, self.labels, self.max_err_output_sum) def _gpu_init(self): dtype = self.output.dtype block_size = min(self.err_output.shape[0], 256) self.build_program(cache_file_name="%s_%d_%d" % (self.__class__.__name__, self.output.shape[0], self.output.sample_size), dtype=dtype, block_size=block_size, max_batch_size=self.err_output.shape[0], output_size=self.err_output.sample_size) self.assign_kernel("evaluate_softmax") self.set_args(self.output, self.max_idx, self.labels, self.skip_args(2), self.n_err, self.confusion_matrix, self.max_err_output_sum, self.err_output) return block_size def ocl_init(self): if self.testing: return block_size = self._gpu_init() self._global_size = [block_size] self._local_size = [block_size] def cuda_init(self): if self.testing: return block_size = self._gpu_init() self._global_size = (1, 1, 1) self._local_size = (block_size, 1, 1) def _gpu_run(self): self.unmap_vectors(self.err_output, self.output, self.max_idx, self.labels, self.n_err, self.confusion_matrix, self.max_err_output_sum) self.krn_constants_i_[0] = self.batch_size self.set_arg(3, self.krn_constants_i_[0:1]) self.krn_constants_f_[0] = 1.0 / self.batch_size if self.mean else 1.0 self.set_arg(4, self.krn_constants_f_[0:1]) self.execute_kernel(self._global_size, self._local_size) def ocl_run(self): return self._gpu_run() def cuda_run(self): return self._gpu_run() def numpy_run(self): self.err_output.map_invalidate() for vec in self.output, self.max_idx, self.labels: vec.map_read() for vec in self.n_err, self.confusion_matrix, self.max_err_output_sum: vec.map_write() batch_size = self.batch_size labels = self.labels.mem confusion_matrix = self.confusion_matrix.mem n_ok = 0 n_total = 0 multiplier = 1.0 / batch_size if self.mean else 1.0 for i in range(batch_size): # loop by batch if labels[i] < 0: self.err_output.mem[i] = 0.0 continue output = ravel(self.output[i]) err_output = ravel(self.err_output[i]) max_idx = self.max_idx[i] confusion_matrix[max_idx, labels[i]] += 1 if max_idx == labels[i]: n_ok += 1 n_total += 1 # Compute softmax output error gradient err_output[:] = output[:] err_output[labels[i]] -= 1.0 err_output *= multiplier if err_output.dtype in (numpy.complex64, numpy.complex128): self.max_err_output_sum[0] = max(self.max_err_output_sum[0], numpy.linalg.norm(err_output)) else: self.max_err_output_sum[0] = max( self.max_err_output_sum[0], (numpy.fabs(err_output)).sum()) # Set errors for excessive samples to zero if batch_size < self.err_output.mem.shape[0]: self.err_output.mem[batch_size:] = 0.0 self.n_err[0] += batch_size - n_ok self.n_err[1] += n_total def get_metric_values(self): if self.testing: output_labels = {} class_keys = getattr(self, "class_keys", None) for index, labels in enumerate(self.merged_output[:]): max_value = 0 for label_index, value in enumerate(labels): if value >= max_value: max_value = value max_index = label_index if class_keys is not None: output_labels[self.class_keys[TEST] [index]] = self.labels_mapping[max_index] else: output_labels[index] = self.labels_mapping[max_index] return {"Output": output_labels} return {}
class BatchWeights(AcceleratedUnit, EmptyDeviceMethodsMixin): """Make weigths and biases from batch v and h. Must be assigned before initialize(): * v * h * batch_size Updates after run(): * hbias_batch * vbias_batch * W_batch Creates within initialize(): * hbias_batch * vbias_batch * W_batch Attributes: v: input data batch h: hidden states of input batch batch_size: size of batch hbias_batch: bias calculated from h vbias_batch: bias calculated from v W_batch: weigths calculated from batch v and h """ def __init__(self, workflow, **kwargs): super(BatchWeights, self).__init__(workflow, **kwargs) self.vbias_batch = Array() self.hbias_batch = Array() self.weights_batch = Array() self.demand("v", "h", "batch_size") def initialize(self, device, **kwargs): super(BatchWeights, self).initialize(device=device, **kwargs) vbias_size = self.v.size // self.v.shape[0] hbias_size = self.h.size // self.h.shape[0] W_size = vbias_size * hbias_size if not self.hbias_batch: self.hbias_batch.reset( numpy.zeros((1, hbias_size), dtype=self.h.mem.dtype)) else: assert self.hbias_batch.size == hbias_size if not self.vbias_batch: self.vbias_batch.reset( numpy.zeros((1, vbias_size), dtype=self.h.mem.dtype)) else: assert self.vbias_batch.size == vbias_size if not self.weights_batch: self.weights_batch.reset( numpy.zeros((vbias_size, hbias_size), dtype=self.h.mem.dtype)) else: assert self.weights_batch.size == W_size self.init_vectors(self.weights_batch, self.vbias_batch, self.hbias_batch, self.v, self.h) def run(self): self.v.map_read() self.h.map_read() for v in self.weights_batch, self.hbias_batch, self.vbias_batch: v.map_invalidate() self.weights_batch.mem[:] = numpy.dot( numpy.transpose(self.v.mem[0: self.batch_size, :]), self.h.mem[0: self.batch_size, :]) / \ self.batch_size for bv in (self.vbias_batch, self.v), (self.hbias_batch, self.h): bv[0].mem[:] = (numpy.sum(bv[1].mem[:self.batch_size, :], 0) / self.batch_size) bv[0].shape = (1, bv[0].size)
class EvaluatorBase(AcceleratedUnit, TriviallyDistributable): hide_from_registry = True """Base class for evaluators. """ def __init__(self, workflow, **kwargs): kwargs["view_group"] = kwargs.get("view_group", "EVALUATOR") super(EvaluatorBase, self).__init__(workflow, **kwargs) self.mean = kwargs.get("mean", True) self.err_output = Array() self._merged_output = Array() self.krn_constants_i_ = None self.krn_constants_f_ = None self.demand("output", "batch_size") if self.testing: self.demand("class_lengths", "offset") @property def mean(self): """ :return: True if the error function averages values. Default is True. """ return self._mean @mean.setter def mean(self, value): if not isinstance(value, bool): raise TypeError("mean must be boolean (got %s)" % type(value)) self._mean = value @property def merged_output(self): assert self.testing return self._merged_output.mem def initialize(self, device, **kwargs): super(EvaluatorBase, self).initialize(device, **kwargs) dtype = self.output.dtype if self.testing: self._merged_output.reset( numpy.zeros( (self.class_lengths[TEST], ) + self.output.shape[1:], dtype)) return self.krn_constants_i_ = numpy.zeros(1, numpy.int32) self.krn_constants_f_ = numpy.zeros(1, dtype) self.err_output.reset(numpy.zeros_like(self.output.mem, dtype)) for vec in self.output, self.err_output: vec.initialize(self.device) def run(self): if self.testing: self.output.map_read() self.merge_output() return return super(EvaluatorBase, self).run() def merge_output(self): self.merged_output[self.offset - self.batch_size:self.offset] = \ self.output[:self.batch_size] def get_metric_names(self): if self.testing: return {"Output"} return set() def get_metric_values(self): if self.testing: return {"Output": self.merged_output} return {}
class Uniform(AcceleratedUnit): """Generates random numbers from uniform distribution. Attributes: num_states: number of random states for parallel generation. states: Array of random states. prng: veles.prng.RandomGenerator for initial states generation. output_bytes: number of output bytes to generate. """ backend_methods = AcceleratedUnit.backend_methods + ("fill",) def __init__(self, workflow, **kwargs): super(Uniform, self).__init__(workflow, **kwargs) self.num_states = kwargs.get("num_states", 256) self.states = Array() self.prng = kwargs.get("prng", get()) self.output_bytes = kwargs.get("output_bytes", 0) self.output = Array() self.cl_const = numpy.zeros(1, dtype=numpy.int32) def init_unpickled(self): super(Uniform, self).init_unpickled() self.sources_["random"] = {} def initialize(self, device, **kwargs): super(Uniform, self).initialize(device, **kwargs) if not self.states or self.states.size != self.num_states * 16: self.states.reset(numpy.empty(self.num_states * 16 * 2, dtype=numpy.uint32)) self.states.mem[:] = self.prng.randint(0, (1 << 32) + 1, self.states.size) if not self.output or self.output.nbytes < self.output_bytes: self.output_bytes = roundup(self.output_bytes, self.num_states * 16 * 8) self.output.reset(numpy.zeros(self.output_bytes, numpy.uint8)) else: self.output_bytes = self.output.nbytes self.init_vectors(self.states, self.output) def _gpu_init(self): self.build_program({}, "uniform_%d" % self.num_states) self.assign_kernel("random_xorshift1024star") self.set_args(self.states, self.cl_const, self.output) def ocl_init(self): self._gpu_init() self._global_size = [self.num_states] self._local_size = None def cuda_init(self): self._gpu_init() n = self.num_states l = 1 while not (n & 1) and l < 32: n >>= 1 l <<= 1 self._global_size = (n, 1, 1) self._local_size = (l, 1, 1) def _gpu_fill(self, nbytes): bytes_per_round = self.num_states * 16 * 8 nbytes = roundup(nbytes, bytes_per_round) if nbytes > self.output.nbytes: raise error.Bug("nbytes > self.output.nbytes") self.unmap_vectors(self.states, self.output) self.cl_const[0] = nbytes // bytes_per_round self.set_arg(1, self.cl_const) self.execute_kernel(self._global_size, self._local_size) def ocl_fill(self, nbytes): self._gpu_fill(nbytes) def cuda_fill(self, nbytes): self._gpu_fill(nbytes) def numpy_fill(self, nbytes): bytes_per_round = self.num_states * 16 * 8 nbytes = roundup(nbytes, bytes_per_round) if nbytes > self.output.nbytes: raise error.Bug("nbytes > self.output.nbytes") self.states.map_write() self.output.map_invalidate() n_rounds = nbytes // bytes_per_round u64 = numpy.array([1181783497276652981], dtype=numpy.uint64) s0 = numpy.zeros(1, dtype=numpy.uint64) s1 = numpy.zeros(1, dtype=numpy.uint64) states = self.states.mem.view(dtype=numpy.uint64) states = states.reshape(states.size // 16, 16) output = self.output.mem.view(dtype=numpy.uint64) for i in range(self.num_states): offs = i s = states[i] self.p = 0 for _round in range(n_rounds): for _iter in range(16): output[offs] = self._next_rand(s, s0, s1, u64) offs += self.num_states def _next_rand(self, s, s0, s1, u64): s0[0] = s[self.p] self.p = (self.p + 1) & 15 s1[0] = s[self.p] s1 ^= s1 << 31 s1 ^= s1 >> 11 s0 ^= s0 >> 30 s0 ^= s1 s[self.p] = s0[0] return (s0 * u64)[0] def fill(self, nbytes): self._backend_fill_(nbytes) def ocl_run(self): self.ocl_fill(self.output.nbytes) def cuda_run(self): self.cuda_fill(self.output.nbytes) def numpy_run(self): self.numpy_fill(self.output.nbytes)