def batch_func(self, param ): pickled_samples, resolution, face_type, data_format = param samples = pickle.loads(pickled_samples) shuffle_idxs = [] idxs = [*range(len(samples))] random_flip = True rotation_range=[-10,10] scale_range=[-0.05, 0.05] tx_range=[-0.05, 0.05] ty_range=[-0.05, 0.05] random_bilinear_resize_chance, random_bilinear_resize_max_size_per = 25,75 motion_blur_chance, motion_blur_mb_max_size = 25, 5 gaussian_blur_chance, gaussian_blur_kernel_max_size = 25, 5 bs = self.batch_size while True: batches = [ [], [] ] n_batch = 0 while n_batch < bs: try: if len(shuffle_idxs) == 0: shuffle_idxs = idxs.copy() np.random.shuffle(shuffle_idxs) idx = shuffle_idxs.pop() sample = samples[idx] img = sample.load_bgr() h,w,c = img.shape mask = np.zeros ((h,w,1), dtype=np.float32) sample.seg_ie_polys.overlay_mask(mask) warp_params = imagelib.gen_warp_params(resolution, random_flip, rotation_range=rotation_range, scale_range=scale_range, tx_range=tx_range, ty_range=ty_range ) if face_type == sample.face_type: if w != resolution: img = cv2.resize( img, (resolution, resolution), cv2.INTER_LANCZOS4 ) mask = cv2.resize( mask, (resolution, resolution), cv2.INTER_LANCZOS4 ) else: mat = LandmarksProcessor.get_transform_mat (sample.landmarks, resolution, face_type) img = cv2.warpAffine( img, mat, (resolution,resolution), borderMode=cv2.BORDER_CONSTANT, flags=cv2.INTER_LANCZOS4 ) mask = cv2.warpAffine( mask, mat, (resolution,resolution), borderMode=cv2.BORDER_CONSTANT, flags=cv2.INTER_LANCZOS4 ) if len(mask.shape) == 2: mask = mask[...,None] img = imagelib.warp_by_params (warp_params, img, can_warp=True, can_transform=True, can_flip=True, border_replicate=False) mask = imagelib.warp_by_params (warp_params, mask, can_warp=True, can_transform=True, can_flip=True, border_replicate=False) img = np.clip(img.astype(np.float32), 0, 1) mask[mask < 0.5] = 0.0 mask[mask >= 0.5] = 1.0 mask = np.clip(mask, 0, 1) if np.random.randint(2) == 0: img = imagelib.apply_random_hsv_shift(img, mask=sd.random_circle_faded ([resolution,resolution])) else: img = imagelib.apply_random_rgb_levels(img, mask=sd.random_circle_faded ([resolution,resolution])) img = imagelib.apply_random_motion_blur( img, motion_blur_chance, motion_blur_mb_max_size, mask=sd.random_circle_faded ([resolution,resolution])) img = imagelib.apply_random_gaussian_blur( img, gaussian_blur_chance, gaussian_blur_kernel_max_size, mask=sd.random_circle_faded ([resolution,resolution])) img = imagelib.apply_random_bilinear_resize( img, random_bilinear_resize_chance, random_bilinear_resize_max_size_per, mask=sd.random_circle_faded ([resolution,resolution])) if data_format == "NCHW": img = np.transpose(img, (2,0,1) ) mask = np.transpose(mask, (2,0,1) ) batches[0].append ( img ) batches[1].append ( mask ) n_batch += 1 except: io.log_err ( traceback.format_exc() ) yield [ np.array(batch) for batch in batches]
def process(samples, sample_process_options, output_sample_types, debug, ct_sample=None): SPST = SampleProcessor.SampleType SPCT = SampleProcessor.ChannelType SPFMT = SampleProcessor.FaceMaskType sample_rnd_seed = np.random.randint(0x80000000) outputs = [] for sample in samples: sample_face_type = sample.face_type sample_bgr = sample.load_bgr() sample_landmarks = sample.landmarks ct_sample_bgr = None h, w, c = sample_bgr.shape def get_full_face_mask(): if sample.eyebrows_expand_mod is not None: full_face_mask = LandmarksProcessor.get_image_hull_mask( sample_bgr.shape, sample_landmarks, eyebrows_expand_mod=sample.eyebrows_expand_mod) else: full_face_mask = LandmarksProcessor.get_image_hull_mask( sample_bgr.shape, sample_landmarks) return np.clip(full_face_mask, 0, 1) def get_eyes_mask(): eyes_mask = LandmarksProcessor.get_image_eye_mask( sample_bgr.shape, sample_landmarks) return np.clip(eyes_mask, 0, 1) is_face_sample = sample_landmarks is not None if debug and is_face_sample: LandmarksProcessor.draw_landmarks(sample_bgr, sample_landmarks, (0, 1, 0)) params_per_resolution = {} warp_rnd_state = np.random.RandomState(sample_rnd_seed - 1) for opts in output_sample_types: resolution = opts.get('resolution', None) if resolution is None: continue params_per_resolution[resolution] = imagelib.gen_warp_params( resolution, sample_process_options.random_flip, rotation_range=sample_process_options.rotation_range, scale_range=sample_process_options.scale_range, tx_range=sample_process_options.tx_range, ty_range=sample_process_options.ty_range, rnd_state=warp_rnd_state) outputs_sample = [] for opts in output_sample_types: sample_type = opts.get('sample_type', SPST.NONE) channel_type = opts.get('channel_type', SPCT.NONE) resolution = opts.get('resolution', 0) warp = opts.get('warp', False) transform = opts.get('transform', False) motion_blur = opts.get('motion_blur', None) gaussian_blur = opts.get('gaussian_blur', None) random_bilinear_resize = opts.get('random_bilinear_resize', None) random_rgb_levels = opts.get('random_rgb_levels', False) random_hsv_shift = opts.get('random_hsv_shift', False) random_circle_mask = opts.get('random_circle_mask', False) normalize_tanh = opts.get('normalize_tanh', False) ct_mode = opts.get('ct_mode', None) data_format = opts.get('data_format', 'NHWC') if sample_type == SPST.FACE_MASK or sample_type == SPST.IMAGE: border_replicate = False elif sample_type == SPST.FACE_IMAGE: border_replicate = True border_replicate = opts.get('border_replicate', border_replicate) borderMode = cv2.BORDER_REPLICATE if border_replicate else cv2.BORDER_CONSTANT if sample_type == SPST.FACE_IMAGE or sample_type == SPST.FACE_MASK: if not is_face_sample: raise ValueError( "face_samples should be provided for sample_type FACE_*" ) if sample_type == SPST.FACE_IMAGE or sample_type == SPST.FACE_MASK: face_type = opts.get('face_type', None) face_mask_type = opts.get('face_mask_type', SPFMT.NONE) if face_type is None: raise ValueError( "face_type must be defined for face samples") if face_type > sample.face_type: raise Exception( 'sample %s type %s does not match model requirement %s. Consider extract necessary type of faces.' % (sample.filename, sample.face_type, face_type)) if sample_type == SPST.FACE_MASK: if face_mask_type == SPFMT.FULL_FACE: img = get_full_face_mask() elif face_mask_type == SPFMT.EYES: img = get_eyes_mask() elif face_mask_type == SPFMT.FULL_FACE_EYES: img = get_full_face_mask() + get_eyes_mask() else: img = np.zeros(sample_bgr.shape[0:2] + (1, ), dtype=np.float32) if sample.ie_polys is not None: sample.ie_polys.overlay_mask(img) if sample_face_type == FaceType.MARK_ONLY: mat = LandmarksProcessor.get_transform_mat( sample_landmarks, warp_resolution, face_type) img = cv2.warpAffine( img, mat, (warp_resolution, warp_resolution), flags=cv2.INTER_LINEAR) img = imagelib.warp_by_params( params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=border_replicate, cv2_inter=cv2.INTER_LINEAR) img = cv2.resize(img, (resolution, resolution), cv2.INTER_LINEAR) else: if face_type != sample_face_type: mat = LandmarksProcessor.get_transform_mat( sample_landmarks, resolution, face_type) img = cv2.warpAffine(img, mat, (resolution, resolution), borderMode=borderMode, flags=cv2.INTER_LINEAR) else: if w != resolution: img = cv2.resize(img, (resolution, resolution), cv2.INTER_CUBIC) img = imagelib.warp_by_params( params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=border_replicate, cv2_inter=cv2.INTER_LINEAR) if len(img.shape) == 2: img = img[..., None] if channel_type == SPCT.G: out_sample = img.astype(np.float32) else: raise ValueError( "only channel_type.G supported for the mask") elif sample_type == SPST.FACE_IMAGE: img = sample_bgr if random_rgb_levels: random_mask = sd.random_circle_faded( [w, w], rnd_state=np.random.RandomState( sample_rnd_seed) ) if random_circle_mask else None img = imagelib.apply_random_rgb_levels( img, mask=random_mask, rnd_state=np.random.RandomState( sample_rnd_seed)) if random_hsv_shift: random_mask = sd.random_circle_faded( [w, w], rnd_state=np.random.RandomState( sample_rnd_seed + 1)) if random_circle_mask else None img = imagelib.apply_random_hsv_shift( img, mask=random_mask, rnd_state=np.random.RandomState( sample_rnd_seed + 1)) if face_type != sample_face_type: mat = LandmarksProcessor.get_transform_mat( sample_landmarks, resolution, face_type) img = cv2.warpAffine(img, mat, (resolution, resolution), borderMode=borderMode, flags=cv2.INTER_CUBIC) else: if w != resolution: img = cv2.resize(img, (resolution, resolution), cv2.INTER_CUBIC) # Apply random color transfer if ct_mode is not None and ct_sample is not None: if ct_sample_bgr is None: ct_sample_bgr = ct_sample.load_bgr() img = imagelib.color_transfer( ct_mode, img, cv2.resize(ct_sample_bgr, (resolution, resolution), cv2.INTER_LINEAR)) img = imagelib.warp_by_params( params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=border_replicate) img = np.clip(img.astype(np.float32), 0, 1) if motion_blur is not None: random_mask = sd.random_circle_faded( [resolution, resolution], rnd_state=np.random.RandomState( sample_rnd_seed + 2)) if random_circle_mask else None img = imagelib.apply_random_motion_blur( img, *motion_blur, mask=random_mask, rnd_state=np.random.RandomState( sample_rnd_seed + 2)) if gaussian_blur is not None: random_mask = sd.random_circle_faded( [resolution, resolution], rnd_state=np.random.RandomState( sample_rnd_seed + 3)) if random_circle_mask else None img = imagelib.apply_random_gaussian_blur( img, *gaussian_blur, mask=random_mask, rnd_state=np.random.RandomState( sample_rnd_seed + 3)) if random_bilinear_resize is not None: random_mask = sd.random_circle_faded( [resolution, resolution], rnd_state=np.random.RandomState( sample_rnd_seed + 4)) if random_circle_mask else None img = imagelib.apply_random_bilinear_resize( img, *random_bilinear_resize, mask=random_mask, rnd_state=np.random.RandomState( sample_rnd_seed + 4)) # Transform from BGR to desired channel_type if channel_type == SPCT.BGR: out_sample = img elif channel_type == SPCT.G: out_sample = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)[..., None] elif channel_type == SPCT.GGG: out_sample = np.repeat( np.expand_dims( cv2.cvtColor(img, cv2.COLOR_BGR2GRAY), -1), (3, ), -1) # Final transformations if not debug: if normalize_tanh: out_sample = np.clip(out_sample * 2.0 - 1.0, -1.0, 1.0) if data_format == "NCHW": out_sample = np.transpose(out_sample, (2, 0, 1)) elif sample_type == SPST.IMAGE: img = sample_bgr img = imagelib.warp_by_params( params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=True) img = cv2.resize(img, (resolution, resolution), cv2.INTER_CUBIC) out_sample = img if data_format == "NCHW": out_sample = np.transpose(out_sample, (2, 0, 1)) elif sample_type == SPST.LANDMARKS_ARRAY: l = sample_landmarks l = np.concatenate([ np.expand_dims(l[:, 0] / w, -1), np.expand_dims(l[:, 1] / h, -1) ], -1) l = np.clip(l, 0.0, 1.0) out_sample = l elif sample_type == SPST.PITCH_YAW_ROLL or sample_type == SPST.PITCH_YAW_ROLL_SIGMOID: pitch, yaw, roll = sample.get_pitch_yaw_roll() if params_per_resolution[resolution]['flip']: yaw = -yaw if sample_type == SPST.PITCH_YAW_ROLL_SIGMOID: pitch = np.clip((pitch / math.pi) / 2.0 + 0.5, 0, 1) yaw = np.clip((yaw / math.pi) / 2.0 + 0.5, 0, 1) roll = np.clip((roll / math.pi) / 2.0 + 0.5, 0, 1) out_sample = (pitch, yaw) else: raise ValueError('expected sample_type') outputs_sample.append(out_sample) outputs += [outputs_sample] return outputs
def process(samples, sample_process_options, output_sample_types, debug, ct_sample=None): SPST = SampleProcessor.SampleType SPCT = SampleProcessor.ChannelType SPFMT = SampleProcessor.FaceMaskType sample_rnd_seed = np.random.randint(0x80000000) outputs = [] for sample in samples: sample_face_type = sample.face_type sample_bgr = sample.load_bgr() sample_landmarks = sample.landmarks ct_sample_bgr = None h, w, c = sample_bgr.shape def get_full_face_mask(): xseg_mask = sample.get_xseg_mask() if xseg_mask is not None: if xseg_mask.shape[0] != h or xseg_mask.shape[1] != w: xseg_mask = cv2.resize(xseg_mask, (w, h), interpolation=cv2.INTER_CUBIC) xseg_mask = imagelib.normalize_channels(xseg_mask, 1) return np.clip(xseg_mask, 0, 1) else: full_face_mask = LandmarksProcessor.get_image_hull_mask( sample_bgr.shape, sample_landmarks, eyebrows_expand_mod=sample.eyebrows_expand_mod) return np.clip(full_face_mask, 0, 1) def get_eyes_mask(): eyes_mask = LandmarksProcessor.get_image_eye_mask( sample_bgr.shape, sample_landmarks) # set eye masks to 1-2 clip = np.clip(eyes_mask, 0, 1) clip[clip > 0.1] += 1 return clip def get_mouth_mask(): mouth_mask = LandmarksProcessor.get_image_mouth_mask( sample_bgr.shape, sample_landmarks) # set eye masks to 2-3 clip = np.clip(mouth_mask, 0, 1) clip[clip > 0.1] += 2 return clip is_face_sample = sample_landmarks is not None if debug and is_face_sample: LandmarksProcessor.draw_landmarks(sample_bgr, sample_landmarks, (0, 1, 0)) params_per_resolution = {} warp_rnd_state = np.random.RandomState(sample_rnd_seed - 1) for opts in output_sample_types: resolution = opts.get('resolution', None) if resolution is None: continue params_per_resolution[resolution] = imagelib.gen_warp_params( resolution, sample_process_options.random_flip, rotation_range=sample_process_options.rotation_range, scale_range=sample_process_options.scale_range, tx_range=sample_process_options.tx_range, ty_range=sample_process_options.ty_range, rnd_state=warp_rnd_state) outputs_sample = [] for opts in output_sample_types: sample_type = opts.get('sample_type', SPST.NONE) channel_type = opts.get('channel_type', SPCT.NONE) resolution = opts.get('resolution', 0) nearest_resize_to = opts.get('nearest_resize_to', None) warp = opts.get('warp', False) transform = opts.get('transform', False) random_downsample = opts.get('random_downsample', False) random_noise = opts.get('random_noise', False) random_blur = opts.get('random_blur', False) random_jpeg = opts.get('random_jpeg', False) motion_blur = opts.get('motion_blur', None) gaussian_blur = opts.get('gaussian_blur', None) random_bilinear_resize = opts.get('random_bilinear_resize', None) random_rgb_levels = opts.get('random_rgb_levels', False) random_hsv_shift = opts.get('random_hsv_shift', False) random_circle_mask = opts.get('random_circle_mask', False) normalize_tanh = opts.get('normalize_tanh', False) ct_mode = opts.get('ct_mode', None) data_format = opts.get('data_format', 'NHWC') if sample_type == SPST.FACE_MASK or sample_type == SPST.IMAGE: border_replicate = False elif sample_type == SPST.FACE_IMAGE: border_replicate = True border_replicate = opts.get('border_replicate', border_replicate) borderMode = cv2.BORDER_REPLICATE if border_replicate else cv2.BORDER_CONSTANT if sample_type == SPST.FACE_IMAGE or sample_type == SPST.FACE_MASK: if not is_face_sample: raise ValueError( "face_samples should be provided for sample_type FACE_*" ) if sample_type == SPST.FACE_IMAGE or sample_type == SPST.FACE_MASK: face_type = opts.get('face_type', None) face_mask_type = opts.get('face_mask_type', SPFMT.NONE) if face_type is None: raise ValueError( "face_type must be defined for face samples") if sample_type == SPST.FACE_MASK: if face_mask_type == SPFMT.FULL_FACE: img = get_full_face_mask() elif face_mask_type == SPFMT.EYES: img = get_eyes_mask() elif face_mask_type == SPFMT.FULL_FACE_EYES: # sets both eyes and mouth mask parts img = get_full_face_mask() mask = img.copy() mask[mask != 0.0] = 1.0 eye_mask = get_eyes_mask() * mask img = np.where(eye_mask > 1, eye_mask, img) mouth_mask = get_mouth_mask() * mask img = np.where(mouth_mask > 2, mouth_mask, img) else: img = np.zeros(sample_bgr.shape[0:2] + (1, ), dtype=np.float32) if sample_face_type == FaceType.MARK_ONLY: mat = LandmarksProcessor.get_transform_mat( sample_landmarks, warp_resolution, face_type) img = cv2.warpAffine( img, mat, (warp_resolution, warp_resolution), flags=cv2.INTER_LINEAR) img = imagelib.warp_by_params( params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=border_replicate, cv2_inter=cv2.INTER_LINEAR) img = cv2.resize(img, (resolution, resolution), interpolation=cv2.INTER_LINEAR) else: if face_type != sample_face_type: mat = LandmarksProcessor.get_transform_mat( sample_landmarks, resolution, face_type) img = cv2.warpAffine(img, mat, (resolution, resolution), borderMode=borderMode, flags=cv2.INTER_LINEAR) else: if w != resolution: img = cv2.resize( img, (resolution, resolution), interpolation=cv2.INTER_LINEAR) img = imagelib.warp_by_params( params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=border_replicate, cv2_inter=cv2.INTER_LINEAR) if len(img.shape) == 2: img = img[..., None] if channel_type == SPCT.G: out_sample = img.astype(np.float32) else: raise ValueError( "only channel_type.G supported for the mask") elif sample_type == SPST.FACE_IMAGE: img = sample_bgr if random_rgb_levels: random_mask = sd.random_circle_faded( [w, w], rnd_state=np.random.RandomState( sample_rnd_seed) ) if random_circle_mask else None img = imagelib.apply_random_rgb_levels( img, mask=random_mask, rnd_state=np.random.RandomState( sample_rnd_seed)) if random_hsv_shift: random_mask = sd.random_circle_faded( [w, w], rnd_state=np.random.RandomState( sample_rnd_seed + 1)) if random_circle_mask else None img = imagelib.apply_random_hsv_shift( img, mask=random_mask, rnd_state=np.random.RandomState( sample_rnd_seed + 1)) if face_type != sample_face_type: mat = LandmarksProcessor.get_transform_mat( sample_landmarks, resolution, face_type) img = cv2.warpAffine(img, mat, (resolution, resolution), borderMode=borderMode, flags=cv2.INTER_CUBIC) else: if w != resolution: img = cv2.resize(img, (resolution, resolution), interpolation=cv2.INTER_CUBIC) # Apply random color transfer if ct_mode is not None and ct_sample is not None or ct_mode == 'fs-aug': if ct_mode == 'fs-aug': img = imagelib.color_augmentation( img, sample_rnd_seed) else: if ct_sample_bgr is None: ct_sample_bgr = ct_sample.load_bgr() img = imagelib.color_transfer( ct_mode, img, cv2.resize(ct_sample_bgr, (resolution, resolution), interpolation=cv2.INTER_LINEAR)) randomization_order = ['blur', 'noise', 'jpeg', 'down'] np.random.shuffle(randomization_order) for random_distortion in randomization_order: # Apply random blur if random_distortion == 'blur' and random_blur: blur_type = np.random.choice( ['motion', 'gaussian']) if blur_type == 'motion': blur_k = np.random.randint(10, 20) blur_angle = 360 * np.random.random() img = LinearMotionBlur( img, blur_k, blur_angle) elif blur_type == 'gaussian': blur_sigma = 5 * np.random.random() + 3 if blur_sigma < 5.0: kernel_size = 2.9 * blur_sigma # 97% of weight else: kernel_size = 2.6 * blur_sigma # 95% of weight kernel_size = int(kernel_size) kernel_size = kernel_size + 1 if kernel_size % 2 == 0 else kernel_size img = cv2.GaussianBlur( img, (kernel_size, kernel_size), blur_sigma) # Apply random noise if random_distortion == 'noise' and random_noise: noise_type = np.random.choice( ['gaussian', 'laplace', 'poisson']) noise_scale = (20 * np.random.random() + 20) if noise_type == 'gaussian': noise = np.random.normal(scale=noise_scale, size=img.shape) img += noise / 255.0 elif noise_type == 'laplace': noise = np.random.laplace( scale=noise_scale, size=img.shape) img += noise / 255.0 elif noise_type == 'poisson': noise_lam = (15 * np.random.random() + 15) noise = np.random.poisson(lam=noise_lam, size=img.shape) img += noise / 255.0 # Apply random jpeg compression if random_distortion == 'jpeg' and random_jpeg: img = np.clip(img * 255, 0, 255).astype(np.uint8) jpeg_compression_level = np.random.randint( 50, 85) encode_param = [ int(cv2.IMWRITE_JPEG_QUALITY), jpeg_compression_level ] _, enc_img = cv2.imencode( '.jpg', img, encode_param) img = cv2.imdecode( enc_img, cv2.IMREAD_UNCHANGED).astype( np.float32) / 255.0 # Apply random downsampling if random_distortion == 'down' and random_downsample: down_res = np.random.randint( int(0.125 * resolution), int(0.25 * resolution)) img = cv2.resize(img, (down_res, down_res), interpolation=cv2.INTER_CUBIC) img = cv2.resize(img, (resolution, resolution), interpolation=cv2.INTER_CUBIC) img = imagelib.warp_by_params( params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=border_replicate) img = np.clip(img.astype(np.float32), 0, 1) if motion_blur is not None: random_mask = sd.random_circle_faded( [resolution, resolution], rnd_state=np.random.RandomState( sample_rnd_seed + 2)) if random_circle_mask else None img = imagelib.apply_random_motion_blur( img, *motion_blur, mask=random_mask, rnd_state=np.random.RandomState( sample_rnd_seed + 2)) if gaussian_blur is not None: random_mask = sd.random_circle_faded( [resolution, resolution], rnd_state=np.random.RandomState( sample_rnd_seed + 3)) if random_circle_mask else None img = imagelib.apply_random_gaussian_blur( img, *gaussian_blur, mask=random_mask, rnd_state=np.random.RandomState( sample_rnd_seed + 3)) if random_bilinear_resize is not None: random_mask = sd.random_circle_faded( [resolution, resolution], rnd_state=np.random.RandomState( sample_rnd_seed + 4)) if random_circle_mask else None img = imagelib.apply_random_bilinear_resize( img, *random_bilinear_resize, mask=random_mask, rnd_state=np.random.RandomState( sample_rnd_seed + 4)) # Transform from BGR to desired channel_type if channel_type == SPCT.BGR: out_sample = img elif channel_type == SPCT.LAB_RAND_TRANSFORM: out_sample = random_lab_rotation( img, sample_rnd_seed) elif channel_type == SPCT.G: out_sample = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)[..., None] elif channel_type == SPCT.GGG: out_sample = np.repeat( np.expand_dims( cv2.cvtColor(img, cv2.COLOR_BGR2GRAY), -1), (3, ), -1) # Final transformations if nearest_resize_to is not None: out_sample = cv2_resize( out_sample, (nearest_resize_to, nearest_resize_to), interpolation=cv2.INTER_NEAREST) if not debug: if normalize_tanh: out_sample = np.clip(out_sample * 2.0 - 1.0, -1.0, 1.0) if data_format == "NCHW": out_sample = np.transpose(out_sample, (2, 0, 1)) elif sample_type == SPST.IMAGE: img = sample_bgr img = imagelib.warp_by_params( params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=True) img = cv2.resize(img, (resolution, resolution), interpolation=cv2.INTER_CUBIC) out_sample = img if data_format == "NCHW": out_sample = np.transpose(out_sample, (2, 0, 1)) elif sample_type == SPST.LANDMARKS_ARRAY: l = sample_landmarks l = np.concatenate([ np.expand_dims(l[:, 0] / w, -1), np.expand_dims(l[:, 1] / h, -1) ], -1) l = np.clip(l, 0.0, 1.0) out_sample = l elif sample_type == SPST.PITCH_YAW_ROLL or sample_type == SPST.PITCH_YAW_ROLL_SIGMOID: pitch, yaw, roll = sample.get_pitch_yaw_roll() if params_per_resolution[resolution]['flip']: yaw = -yaw if sample_type == SPST.PITCH_YAW_ROLL_SIGMOID: pitch = np.clip((pitch / math.pi) / 2.0 + 0.5, 0, 1) yaw = np.clip((yaw / math.pi) / 2.0 + 0.5, 0, 1) roll = np.clip((roll / math.pi) / 2.0 + 0.5, 0, 1) out_sample = (pitch, yaw) else: raise ValueError('expected sample_type') outputs_sample.append(out_sample) outputs += [outputs_sample] return outputs
def batch_func(self, param ): samples, seg_sample_idxs, resolution, face_type, data_format = param shuffle_idxs = [] bg_shuffle_idxs = [] random_flip = True rotation_range=[-10,10] scale_range=[-0.05, 0.05] tx_range=[-0.05, 0.05] ty_range=[-0.05, 0.05] random_bilinear_resize_chance, random_bilinear_resize_max_size_per = 25,75 motion_blur_chance, motion_blur_mb_max_size = 25, 5 gaussian_blur_chance, gaussian_blur_kernel_max_size = 25, 5 def gen_img_mask(sample): img = sample.load_bgr() h,w,c = img.shape if sample.seg_ie_polys.has_polys(): mask = np.zeros ((h,w,1), dtype=np.float32) sample.seg_ie_polys.overlay_mask(mask) elif sample.has_xseg_mask(): mask = sample.get_xseg_mask() mask[mask < 0.5] = 0.0 mask[mask >= 0.5] = 1.0 else: raise Exception(f'no mask in sample {sample.filename}') if face_type == sample.face_type: if w != resolution: img = cv2.resize( img, (resolution, resolution), interpolation=cv2.INTER_LANCZOS4 ) mask = cv2.resize( mask, (resolution, resolution), interpolation=cv2.INTER_LANCZOS4 ) else: mat = LandmarksProcessor.get_transform_mat (sample.landmarks, resolution, face_type) img = cv2.warpAffine( img, mat, (resolution,resolution), borderMode=cv2.BORDER_CONSTANT, flags=cv2.INTER_LANCZOS4 ) mask = cv2.warpAffine( mask, mat, (resolution,resolution), borderMode=cv2.BORDER_CONSTANT, flags=cv2.INTER_LANCZOS4 ) if len(mask.shape) == 2: mask = mask[...,None] return img, mask bs = self.batch_size while True: batches = [ [], [] ] n_batch = 0 while n_batch < bs: try: if len(shuffle_idxs) == 0: shuffle_idxs = seg_sample_idxs.copy() np.random.shuffle(shuffle_idxs) sample = samples[shuffle_idxs.pop()] img, mask = gen_img_mask(sample) if np.random.randint(2) == 0: if len(bg_shuffle_idxs) == 0: bg_shuffle_idxs = seg_sample_idxs.copy() np.random.shuffle(bg_shuffle_idxs) bg_sample = samples[bg_shuffle_idxs.pop()] bg_img, bg_mask = gen_img_mask(bg_sample) bg_wp = imagelib.gen_warp_params(resolution, True, rotation_range=[-180,180], scale_range=[-0.10, 0.10], tx_range=[-0.10, 0.10], ty_range=[-0.10, 0.10] ) bg_img = imagelib.warp_by_params (bg_wp, bg_img, can_warp=False, can_transform=True, can_flip=True, border_replicate=False) bg_mask = imagelib.warp_by_params (bg_wp, bg_mask, can_warp=False, can_transform=True, can_flip=True, border_replicate=False) c_mask = (1-bg_mask) * (1-mask) img = img*(1-c_mask) + bg_img * c_mask warp_params = imagelib.gen_warp_params(resolution, random_flip, rotation_range=rotation_range, scale_range=scale_range, tx_range=tx_range, ty_range=ty_range ) img = imagelib.warp_by_params (warp_params, img, can_warp=True, can_transform=True, can_flip=True, border_replicate=False) mask = imagelib.warp_by_params (warp_params, mask, can_warp=True, can_transform=True, can_flip=True, border_replicate=False) img = np.clip(img.astype(np.float32), 0, 1) mask[mask < 0.5] = 0.0 mask[mask >= 0.5] = 1.0 mask = np.clip(mask, 0, 1) if np.random.randint(2) == 0: img = imagelib.apply_random_hsv_shift(img, mask=sd.random_circle_faded ([resolution,resolution])) else: img = imagelib.apply_random_rgb_levels(img, mask=sd.random_circle_faded ([resolution,resolution])) img = imagelib.apply_random_motion_blur( img, motion_blur_chance, motion_blur_mb_max_size, mask=sd.random_circle_faded ([resolution,resolution])) img = imagelib.apply_random_gaussian_blur( img, gaussian_blur_chance, gaussian_blur_kernel_max_size, mask=sd.random_circle_faded ([resolution,resolution])) img = imagelib.apply_random_bilinear_resize( img, random_bilinear_resize_chance, random_bilinear_resize_max_size_per, mask=sd.random_circle_faded ([resolution,resolution])) if data_format == "NCHW": img = np.transpose(img, (2,0,1) ) mask = np.transpose(mask, (2,0,1) ) batches[0].append ( img ) batches[1].append ( mask ) n_batch += 1 except: io.log_err ( traceback.format_exc() ) yield [ np.array(batch) for batch in batches]
def batch_func(self, param): pickled_samples, obstructions_images_paths, resolution, face_type, data_format = param samples = pickle.loads(pickled_samples) obstructions_images_paths_len = len(obstructions_images_paths) shuffle_o_idxs = [] o_idxs = [*range(obstructions_images_paths_len)] shuffle_idxs = [] idxs = [*range(len(samples))] random_flip = True rotation_range = [-10, 10] scale_range = [-0.05, 0.05] tx_range = [-0.05, 0.05] ty_range = [-0.05, 0.05] o_random_flip = True o_rotation_range = [-180, 180] o_scale_range = [-0.5, 0.05] o_tx_range = [-0.5, 0.5] o_ty_range = [-0.5, 0.5] random_bilinear_resize_chance, random_bilinear_resize_max_size_per = 25, 75 motion_blur_chance, motion_blur_mb_max_size = 25, 5 gaussian_blur_chance, gaussian_blur_kernel_max_size = 25, 5 bs = self.batch_size while True: batches = [[], []] n_batch = 0 while n_batch < bs: try: if len(shuffle_idxs) == 0: shuffle_idxs = idxs.copy() np.random.shuffle(shuffle_idxs) idx = shuffle_idxs.pop() sample = samples[idx] img = sample.load_bgr() h, w, c = img.shape mask = np.zeros((h, w, 1), dtype=np.float32) sample.ie_polys.overlay_mask(mask) warp_params = imagelib.gen_warp_params( resolution, random_flip, rotation_range=rotation_range, scale_range=scale_range, tx_range=tx_range, ty_range=ty_range) if face_type == sample.face_type: if w != resolution: img = cv2.resize(img, (resolution, resolution), cv2.INTER_LANCZOS4) mask = cv2.resize(mask, (resolution, resolution), cv2.INTER_LANCZOS4) else: mat = LandmarksProcessor.get_transform_mat( sample.landmarks, resolution, face_type) img = cv2.warpAffine(img, mat, (resolution, resolution), borderMode=cv2.BORDER_CONSTANT, flags=cv2.INTER_LANCZOS4) mask = cv2.warpAffine(mask, mat, (resolution, resolution), borderMode=cv2.BORDER_CONSTANT, flags=cv2.INTER_LANCZOS4) if len(mask.shape) == 2: mask = mask[..., None] if obstructions_images_paths_len != 0: # apply obstruction if len(shuffle_o_idxs) == 0: shuffle_o_idxs = o_idxs.copy() np.random.shuffle(shuffle_o_idxs) o_idx = shuffle_o_idxs.pop() o_img = cv2_imread( obstructions_images_paths[o_idx]).astype( np.float32) / 255.0 oh, ow, oc = o_img.shape if oc == 4: ohw = max(oh, ow) scale = resolution / ohw #o_img = cv2.resize (o_img, ( int(ow*rate), int(oh*rate), ), cv2.INTER_CUBIC) mat = cv2.getRotationMatrix2D( (ow / 2, oh / 2), np.random.uniform(o_rotation_range[0], o_rotation_range[1]), 1.0) mat += np.float32([[0, 0, -ow / 2], [0, 0, -oh / 2]]) mat *= scale * np.random.uniform( 1 + o_scale_range[0], 1 + o_scale_range[1]) mat += np.float32( [[ 0, 0, resolution / 2 + resolution * np.random.uniform( o_tx_range[0], o_tx_range[1]) ], [ 0, 0, resolution / 2 + resolution * np.random.uniform( o_ty_range[0], o_ty_range[1]) ]]) o_img = cv2.warpAffine( o_img, mat, (resolution, resolution), borderMode=cv2.BORDER_CONSTANT, flags=cv2.INTER_LANCZOS4) if o_random_flip and np.random.randint(10) < 4: o_img = o_img[:, ::-1, ...] o_mask = o_img[..., 3:4] o_mask[o_mask > 0] = 1.0 o_mask = cv2.erode(o_mask, cv2.getStructuringElement( cv2.MORPH_ELLIPSE, (5, 5)), iterations=1) o_mask = cv2.GaussianBlur(o_mask, (5, 5), 0)[..., None] img = img * (1 - o_mask) + o_img[..., 0:3] * o_mask o_mask[o_mask < 0.5] = 0.0 #import code #code.interact(local=dict(globals(), **locals())) mask *= (1 - o_mask) #cv2.imshow ("", np.clip(o_img*255, 0,255).astype(np.uint8) ) #cv2.waitKey(0) img = imagelib.warp_by_params(warp_params, img, can_warp=True, can_transform=True, can_flip=True, border_replicate=False) mask = imagelib.warp_by_params(warp_params, mask, can_warp=True, can_transform=True, can_flip=True, border_replicate=False) img = np.clip(img.astype(np.float32), 0, 1) mask[mask < 0.5] = 0.0 mask[mask >= 0.5] = 1.0 mask = np.clip(mask, 0, 1) img = imagelib.apply_random_hsv_shift( img, mask=sd.random_circle_faded([resolution, resolution])) img = imagelib.apply_random_motion_blur( img, motion_blur_chance, motion_blur_mb_max_size, mask=sd.random_circle_faded([resolution, resolution])) img = imagelib.apply_random_gaussian_blur( img, gaussian_blur_chance, gaussian_blur_kernel_max_size, mask=sd.random_circle_faded([resolution, resolution])) img = imagelib.apply_random_bilinear_resize( img, random_bilinear_resize_chance, random_bilinear_resize_max_size_per, mask=sd.random_circle_faded([resolution, resolution])) if data_format == "NCHW": img = np.transpose(img, (2, 0, 1)) mask = np.transpose(mask, (2, 0, 1)) batches[0].append(img) batches[1].append(mask) n_batch += 1 except: io.log_err(traceback.format_exc()) yield [np.array(batch) for batch in batches]