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 cut_out(self, sample):
if self.minibatch_class == 2:
rand = prng.get()
h_off = rand.randint(sample.shape[0] - self.crop_size_sy + 1)
w_off = rand.randint(sample.shape[1] - self.crop_size_sx + 1)
else:
h_off = (sample.shape[0] - self.crop_size_sy) / 2
w_off = (sample.shape[1] - self.crop_size_sx) / 2
sample = sample[h_off:h_off + self.crop_size_sy,
w_off:w_off + self.crop_size_sx, :self.channels]
return sample
def cut_out(self, sample):
if self.minibatch_class == 2:
rand = prng.get()
h_off = rand.randint(
sample.shape[0] - self.crop_size_sy + 1)
w_off = rand.randint(
sample.shape[1] - self.crop_size_sx + 1)
else:
h_off = (sample.shape[0] - self.crop_size_sy) / 2
w_off = (sample.shape[1] - self.crop_size_sx) / 2
sample = sample[
h_off:h_off + self.crop_size_sy,
w_off:w_off + self.crop_size_sx, :self.channels]
return sample
def fill_indices(self, start_offset, count):
self.minibatch_indices.map_invalidate()
idxs = self.minibatch_indices.mem
self.shuffled_indices.map_read()
idxs[:count] = self.shuffled_indices[start_offset:start_offset + count]
if self.is_master:
return True
if self.matrixes_filename is not None:
self.has_mean_file = True
self.minibatch_data.map_invalidate()
self.minibatch_labels.map_invalidate()
sample = numpy.zeros(
[self.sy, self.sx, self.channels], dtype=numpy.uint8)
sample_bytes = sample.nbytes
for index, index_sample in enumerate(idxs[:count]):
self.file_samples.seek(int(index_sample) * sample_bytes)
self.file_samples.readinto(sample)
rand = prng.get()
self.do_mirror = self.mirror and bool(rand.randint((2)))
image = self.transform_sample(sample)
self.minibatch_data.mem[index] = image
self.minibatch_labels.mem[
index] = self.original_labels[int(index_sample)]
if count < len(idxs):
idxs[count:] = self.class_lengths[1] # no data sample is there
self.croped_mean = self.cut_out(self.mean)
self.minibatch_data.mem[count:] = self.croped_mean
self.minibatch_labels.mem[count:] = 0 # 0 is no data
return True
def fill_indices(self, start_offset, count):
self.minibatch_indices.map_invalidate()
idxs = self.minibatch_indices.mem
self.shuffled_indices.map_read()
idxs[:count] = self.shuffled_indices[start_offset:start_offset + count]
if self.is_master:
return True
if self.matrixes_filename is not None:
self.has_mean_file = True
self.minibatch_data.map_invalidate()
self.minibatch_labels.map_invalidate()
sample = numpy.zeros([self.sy, self.sx, self.channels],
dtype=numpy.uint8)
sample_bytes = sample.nbytes
for index, index_sample in enumerate(idxs[:count]):
self.file_samples.seek(int(index_sample) * sample_bytes)
self.file_samples.readinto(sample)
rand = prng.get()
self.do_mirror = self.mirror and bool(rand.randint((2)))
image = self.transform_sample(sample)
self.minibatch_data.mem[index] = image
self.minibatch_labels.mem[index] = self.original_labels[int(
index_sample)]
if count < len(idxs):
idxs[count:] = self.class_lengths[1] # no data sample is there
self.croped_mean = self.cut_out(self.mean)
self.minibatch_data.mem[count:] = self.croped_mean
self.minibatch_labels.mem[count:] = 0 # 0 is no data
return True
