def yolo_head(feats, anchors, num_classes, input_shape): """Convert final layer features to bounding box parameters.""" num_anchors = len(anchors) # Reshape to batch, height, width, num_anchors, box_params. anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2]) grid_shape = K.shape(feats)[1:3] # height, width grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]), [1, grid_shape[1], 1, 1]) grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]), [grid_shape[0], 1, 1, 1]) grid = K.concatenate([grid_x, grid_y]) grid = K.cast(grid, K.dtype(feats)) feats = K.reshape( feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5]) box_xy = K.sigmoid(feats[..., :2]) box_wh = K.exp(feats[..., 2:4]) box_confidence = K.sigmoid(feats[..., 4:5]) box_class_probs = K.sigmoid(feats[..., 5:]) # Adjust preditions to each spatial grid point and anchor size. box_xy = (box_xy + grid) / K.cast(grid_shape[::-1], K.dtype(feats)) box_wh = box_wh * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats)) return box_xy, box_wh, box_confidence, box_class_probs
def yolo_head(feats, anchors, num_classes): # convert anchors to shape 1 , 1 , 1 , len of anchors , 2 num_anchors = len(anchors) anchors_tensor = K.reshape(K.variable(anchors), [1, 1, 1, num_anchors, 2]) # conv_dims , width and hight of the grid _, conv_height, conv_width, _ = K.int_shape(feats) conv_dims = K.variable([conv_width, conv_height]) # reshape yolo network output to None , grid_width , grid_hight , num of amnchors , num of classes + 5 feats = K.reshape( feats, [-1, conv_dims[0], conv_dims[1], num_anchors, num_classes + 5]) # convert conv_dims after casting it to feats datatype to 1 , 1 , 1 , 2 conv_dims = K.cast(K.reshape(conv_dims, [1, 1, 1, 1, 2]), K.dtype(feats)) # create grid from (0 , 0 ) to (width , hight) conv_index = np.array([_ for _ in np.ndindex(conv_width, conv_height)]) conv_index = conv_index[:, [1, 0]] # swap columns for YOLO ordering. conv_index = K.variable( conv_index.reshape(1, conv_height, conv_width, 1, 2)) box_confidence = K.sigmoid(feats[..., 4:5]) box_xy = K.sigmoid(feats[..., :2]) box_wh = K.exp(feats[..., 2:4]) box_class_probs = K.softmax(feats[..., 5:]) box_xy = (box_xy + conv_index) / conv_dims box_wh = box_wh * anchors_tensor / conv_dims return box_confidence, box_xy, box_wh, box_class_probs
def loop_body(b, ignore_mask): true_box = tf.boolean_mask(y_true[layer][b, ..., 0:4], object_mask_bool[b, ..., 0]) iou = box_iou(pred_box[b], true_box) best_iou = K.max(iou, axis=-1) ignore_mask = ignore_mask.write( b, K.cast(best_iou < ignore_thresh, K.dtype(true_box))) return b + 1, ignore_mask
def _yolo_head(self, feats, anchors, num_classes, input_shape, calc_loss=False): """Convert final layer features to bounding box parameters""" num_anchors = len(anchors) anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2]) # height, width grid_shape = K.shape(feats)[1:3] grid_y = K.tile( K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]), [1, grid_shape[1], 1, 1], ) grid_x = K.tile( K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]), [grid_shape[0], 1, 1, 1], ) grid = K.concatenate([grid_x, grid_y]) grid = K.cast(grid, K.dtype(feats)) feats = K.reshape( feats, [-1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5]) # Adjust preditions to each spatial grid point and anchor size. box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast( grid_shape[::-1], K.dtype(feats)) box_wh = (K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))) box_confidence = K.sigmoid(feats[..., 4:5]) box_class_probs = K.sigmoid(feats[..., 5:]) if calc_loss == True: return grid, feats, box_xy, box_wh return box_xy, box_wh, box_confidence, box_class_probs
def _correct_boxes(self, box_xy, box_wh, input_shape, image_shape): """Get corrected boxes""" box_yx = box_xy[..., ::-1] box_hw = box_wh[..., ::-1] input_shape = K.cast(input_shape, K.dtype(box_yx)) image_shape = K.cast(image_shape, K.dtype(box_yx)) new_shape = K.round(image_shape * K.min(input_shape / image_shape)) offset = (input_shape - new_shape) / 2.0 / input_shape scale = input_shape / new_shape box_yx = (box_yx - offset) * scale box_hw *= scale box_mins = box_yx - (box_hw / 2.0) box_maxes = box_yx + (box_hw / 2.0) boxes = K.concatenate([ box_mins[..., 0:1], # y_min box_mins[..., 1:2], # x_min box_maxes[..., 0:1], # y_max box_maxes[..., 1:2], # x_max ]) # Scale boxes back to original image shape. boxes *= K.concatenate([image_shape, image_shape]) return boxes
def yolo_head(feats, anchors, num_classes, input_shape, calc_loss=False): """Convert final layer features to bounding box parameters.""" num_anchors = anchors_per_level # Reshape to batch, height, width, num_anchors, box_params. anchors_tensor = K.reshape(K.constant(anchors), [1, 1, 1, num_anchors, 2]) grid_shape = K.shape(feats)[1:3] # height, width grid_y = K.tile( tf.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1], name='yolo_head/tile/reshape/grid_y'), [1, grid_shape[1], 1, 1]) grid_x = K.tile( tf.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1], name='yolo_head/tile/reshape/grid_x'), [grid_shape[0], 1, 1, 1]) grid = tf.concat([grid_x, grid_y], axis=-1, name='yolo_head/concatenate/grid') grid = K.cast(grid, K.dtype(feats)) feats = tf.reshape(feats, [ -1, grid_shape[0], grid_shape[1], num_anchors, num_classes + 5 + NUM_ANGLES3 ], name='yolo_head/reshape/feats') # Adjust predictions to each spatial grid point and anchor size. box_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast( grid_shape[..., ::-1], K.dtype(feats)) box_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast( input_shape[..., ::-1], K.dtype(feats)) box_confidence = K.sigmoid(feats[..., 4:5]) box_class_probs = K.sigmoid(feats[..., 5:5 + num_classes]) polygons_confidence = K.sigmoid(feats[..., 5 + num_classes + 2:5 + num_classes + NUM_ANGLES3:3]) polygons_x = K.exp(feats[..., 5 + num_classes:num_classes + 5 + NUM_ANGLES3:3]) dx = K.square(anchors_tensor[..., 0:1] / 2) dy = K.square(anchors_tensor[..., 1:2] / 2) d = K.cast(K.sqrt(dx + dy), K.dtype(polygons_x)) a = K.pow(input_shape[..., ::-1], 2) a = K.cast(a, K.dtype(feats)) b = K.sum(a) diagonal = K.cast(K.sqrt(b), K.dtype(feats)) polygons_x = polygons_x * d / diagonal polygons_y = feats[..., 5 + num_classes + 1:num_classes + 5 + NUM_ANGLES3:3] polygons_y = K.sigmoid(polygons_y) if calc_loss == True: return grid, feats, box_xy, box_wh, polygons_confidence return box_xy, box_wh, box_confidence, box_class_probs, polygons_x, polygons_y, polygons_confidence
def yolo_head(feats, anchors, num_classes): """Convert final layer features to bounding box parameters. Parameters ---------- feats : tensor Final convolutional layer features. anchors : array-like Anchor box widths and heights. num_classes : int Number of target classes. Returns ------- box_xy : tensor x, y box predictions adjusted by spatial location in conv layer. box_wh : tensor w, h box predictions adjusted by anchors and conv spatial resolution. box_conf : tensor Probability estimate for whether each box contains any object. box_class_pred : tensor Probability distribution estimate for each box over class labels. """ num_anchors = len(anchors) # Reshape to batch, height, width, num_anchors, box_params. anchors_tensor = K.reshape(K.variable(anchors), [1, 1, 1, num_anchors, 2]) # Static implementation for fixed models. # TODO: Remove or add option for static implementation. # _, conv_height, conv_width, _ = K.int_shape(feats) # conv_dims = K.variable([conv_width, conv_height]) # Dynamic implementation of conv dims for fully convolutional model. conv_dims = K.shape(feats)[1:3] # assuming channels last # In YOLO the height index is the inner most iteration. conv_height_index = K.arange(0, stop=conv_dims[0]) conv_width_index = K.arange(0, stop=conv_dims[1]) conv_height_index = K.tile(conv_height_index, [conv_dims[1]]) # TODO: Repeat_elements and tf.split doesn't support dynamic splits. # conv_width_index = K.repeat_elements(conv_width_index, conv_dims[1], axis=0) conv_width_index = K.tile( K.expand_dims(conv_width_index, 0), [conv_dims[0], 1]) conv_width_index = K.flatten(K.transpose(conv_width_index)) conv_index = K.transpose(K.stack([conv_height_index, conv_width_index])) conv_index = K.reshape(conv_index, [1, conv_dims[0], conv_dims[1], 1, 2]) conv_index = K.cast(conv_index, K.dtype(feats)) feats = K.reshape( feats, [-1, conv_dims[0], conv_dims[1], num_anchors, num_classes + 5]) conv_dims = K.cast(K.reshape(conv_dims, [1, 1, 1, 1, 2]), K.dtype(feats)) # Static generation of conv_index: # conv_index = np.array([_ for _ in np.ndindex(conv_width, conv_height)]) # conv_index = conv_index[:, [1, 0]] # swap columns for YOLO ordering. # conv_index = K.variable( # conv_index.reshape(1, conv_height, conv_width, 1, 2)) # feats = Reshape( # (conv_dims[0], conv_dims[1], num_anchors, num_classes + 5))(feats) box_xy = K.sigmoid(feats[..., :2]) box_wh = K.exp(feats[..., 2:4]) box_confidence = K.sigmoid(feats[..., 4:5]) box_class_probs = K.softmax(feats[..., 5:]) # Adjust preditions to each spatial grid point and anchor size. # Note: YOLO iterates over height index before width index. box_xy = (box_xy + conv_index) / conv_dims box_wh = box_wh * anchors_tensor / conv_dims return box_xy, box_wh, box_confidence, box_class_probs
def yolo_loss(args, anchors, num_classes, rescore_confidence=False, print_loss=False): """YOLO localization loss function. Parameters ---------- yolo_output : tensor Final convolutional layer features. true_boxes : tensor Ground truth boxes tensor with shape [batch, num_true_boxes, 5] containing box x_center, y_center, width, height, and class. detectors_mask : array 0/1 mask for detector positions where there is a matching ground truth. matching_true_boxes : array Corresponding ground truth boxes for positive detector positions. Already adjusted for conv height and width. anchors : tensor Anchor boxes for model. num_classes : int Number of object classes. rescore_confidence : bool, default=False If true then set confidence target to IOU of best predicted box with the closest matching ground truth box. print_loss : bool, default=False If True then use a tf.Print() to print the loss components. Returns ------- mean_loss : float mean localization loss across minibatch """ (yolo_output, true_boxes, detectors_mask, matching_true_boxes) = args num_anchors = len(anchors) object_scale = 5 no_object_scale = 1 class_scale = 1 coordinates_scale = 1 pred_xy, pred_wh, pred_confidence, pred_class_prob = yolo_head( yolo_output, anchors, num_classes) # Unadjusted box predictions for loss. # TODO: Remove extra computation shared with yolo_head. yolo_output_shape = K.shape(yolo_output) feats = K.reshape(yolo_output, [ -1, yolo_output_shape[1], yolo_output_shape[2], num_anchors, num_classes + 5 ]) pred_boxes = K.concatenate( (K.sigmoid(feats[..., 0:2]), feats[..., 2:4]), axis=-1) # TODO: Adjust predictions by image width/height for non-square images? # IOUs may be off due to different aspect ratio. # Expand pred x,y,w,h to allow comparison with ground truth. # batch, conv_height, conv_width, num_anchors, num_true_boxes, box_params pred_xy = K.expand_dims(pred_xy, 4) pred_wh = K.expand_dims(pred_wh, 4) pred_wh_half = pred_wh / 2. pred_mins = pred_xy - pred_wh_half pred_maxes = pred_xy + pred_wh_half true_boxes_shape = K.shape(true_boxes) # batch, conv_height, conv_width, num_anchors, num_true_boxes, box_params true_boxes = K.reshape(true_boxes, [ true_boxes_shape[0], 1, 1, 1, true_boxes_shape[1], true_boxes_shape[2] ]) true_xy = true_boxes[..., 0:2] true_wh = true_boxes[..., 2:4] # Find IOU of each predicted box with each ground truth box. true_wh_half = true_wh / 2. true_mins = true_xy - true_wh_half true_maxes = true_xy + true_wh_half intersect_mins = K.maximum(pred_mins, true_mins) intersect_maxes = K.minimum(pred_maxes, true_maxes) intersect_wh = K.maximum(intersect_maxes - intersect_mins, 0.) intersect_areas = intersect_wh[..., 0] * intersect_wh[..., 1] pred_areas = pred_wh[..., 0] * pred_wh[..., 1] true_areas = true_wh[..., 0] * true_wh[..., 1] union_areas = pred_areas + true_areas - intersect_areas iou_scores = intersect_areas / union_areas # Best IOUs for each location. best_ious = K.max(iou_scores, axis=4) # Best IOU scores. best_ious = K.expand_dims(best_ious) # A detector has found an object if IOU > thresh for some true box. object_detections = K.cast(best_ious > 0.6, K.dtype(best_ious)) # TODO: Darknet region training includes extra coordinate loss for early # training steps to encourage predictions to match anchor priors. # Determine confidence weights from object and no_object weights. # NOTE: YOLO does not use binary cross-entropy here. no_object_weights = (no_object_scale * (1 - object_detections) * (1 - detectors_mask)) no_objects_loss = no_object_weights * K.square(-pred_confidence) if rescore_confidence: objects_loss = (object_scale * detectors_mask * K.square(best_ious - pred_confidence)) else: objects_loss = (object_scale * detectors_mask * K.square(1 - pred_confidence)) confidence_loss = objects_loss + no_objects_loss # Classification loss for matching detections. # NOTE: YOLO does not use categorical cross-entropy loss here. matching_classes = K.cast(matching_true_boxes[..., 4], 'int32') matching_classes = K.one_hot(matching_classes, num_classes) classification_loss = (class_scale * detectors_mask * K.square(matching_classes - pred_class_prob)) # Coordinate loss for matching detection boxes. matching_boxes = matching_true_boxes[..., 0:4] coordinates_loss = (coordinates_scale * detectors_mask * K.square(matching_boxes - pred_boxes)) confidence_loss_sum = K.sum(confidence_loss) classification_loss_sum = K.sum(classification_loss) coordinates_loss_sum = K.sum(coordinates_loss) total_loss = 0.5 * ( confidence_loss_sum + classification_loss_sum + coordinates_loss_sum) if print_loss: total_loss = tf.Print( total_loss, [ total_loss, confidence_loss_sum, classification_loss_sum, coordinates_loss_sum ], message='yolo_loss, conf_loss, class_loss, box_coord_loss:') return total_loss
def yolo_loss(args, anchors, num_classes, ignore_thresh=.5): """Return yolo_loss tensor Parameters ---------- yolo_outputs: list of tensor, the output of yolo_body or tiny_yolo_body y_true: list of array, the output of preprocess_true_boxes anchors: array, shape=(N, 2), wh num_classes: integer ignore_thresh: float, the iou threshold whether to ignore object confidence loss Returns ------- loss: tensor, shape=(1,) """ num_layers = 1 yolo_outputs = args[:num_layers] y_true = args[num_layers:] g_y_true = y_true input_shape = K.cast( K.shape(yolo_outputs[0])[1:3] * grid_size_multiplier, K.dtype(y_true[0])) grid_shapes = [ K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0])) for l in range(num_layers) ] loss = 0 m = K.shape(yolo_outputs[0])[0] # batch size, tensor mf = K.cast(m, K.dtype(yolo_outputs[0])) for layer in range(num_layers): object_mask = y_true[layer][..., 4:5] vertices_mask = y_true[layer][..., 5 + num_classes + 2:5 + num_classes + NUM_ANGLES3:3] true_class_probs = y_true[layer][..., 5:5 + num_classes] grid, raw_pred, pred_xy, pred_wh, pol_cnf = yolo_head( yolo_outputs[layer], anchors[anchor_mask[layer]], num_classes, input_shape, calc_loss=True) pred_box = K.concatenate([pred_xy, pred_wh]) raw_true_xy = y_true[layer][..., :2] * grid_shapes[layer][ ..., ::-1] - grid raw_true_polygon0 = y_true[layer][..., 5 + num_classes:5 + num_classes + NUM_ANGLES3] raw_true_wh = K.log(y_true[layer][..., 2:4] / anchors[anchor_mask[layer]] * input_shape[..., ::-1]) raw_true_wh = K.switch(object_mask, raw_true_wh, K.zeros_like(raw_true_wh)) # avoid log(0)=-inf raw_true_polygon_x = raw_true_polygon0[..., ::3] raw_true_polygon_y = raw_true_polygon0[..., 1::3] dx = K.square(anchors[anchor_mask[layer]][..., 0:1] / 2) dy = K.square(anchors[anchor_mask[layer]][..., 1:2] / 2) d = K.cast(K.sqrt(dx + dy), K.dtype(raw_true_polygon_x)) diagonal = K.sqrt( K.pow(input_shape[..., ::-1][0], 2) + K.pow(input_shape[..., ::-1][1], 2)) raw_true_polygon_x = K.log(raw_true_polygon_x / d * diagonal) raw_true_polygon_x = K.switch(vertices_mask, raw_true_polygon_x, K.zeros_like(raw_true_polygon_x)) box_loss_scale = 2 - y_true[layer][..., 2:3] * y_true[layer][..., 3:4] # Find ignore mask, iterate over each of batch. ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True) object_mask_bool = K.cast(object_mask, 'bool') def loop_body(b, ignore_mask): true_box = tf.boolean_mask(y_true[layer][b, ..., 0:4], object_mask_bool[b, ..., 0]) iou = box_iou(pred_box[b], true_box) best_iou = K.max(iou, axis=-1) ignore_mask = ignore_mask.write( b, K.cast(best_iou < ignore_thresh, K.dtype(true_box))) return b + 1, ignore_mask _, ignore_mask = tf.while_loop(lambda b, *args: b < m, loop_body, [0, ignore_mask]) ignore_mask = ignore_mask.stack() ignore_mask = K.expand_dims(ignore_mask, -1) # K.binary_crossentropy is helpful to avoid exp overflow. xy_loss = object_mask * box_loss_scale * K.binary_crossentropy( raw_true_xy, raw_pred[..., 0:2], from_logits=True) wh_loss = object_mask * box_loss_scale * 0.5 * K.square( raw_true_wh - raw_pred[..., 2:4]) confidence_loss = object_mask * K.binary_crossentropy( object_mask, raw_pred[..., 4:5], from_logits=True ) + (1 - object_mask) * K.binary_crossentropy( object_mask, raw_pred[..., 4:5], from_logits=True) * ignore_mask class_loss = object_mask * K.binary_crossentropy( true_class_probs, raw_pred[..., 5:5 + num_classes], from_logits=True) polygon_loss_x = object_mask * vertices_mask * box_loss_scale * 0.5 * K.square( raw_true_polygon_x - raw_pred[..., 5 + num_classes:5 + num_classes + NUM_ANGLES3:3]) polygon_loss_y = object_mask * vertices_mask * box_loss_scale * K.binary_crossentropy( raw_true_polygon_y, raw_pred[..., 5 + num_classes + 1:5 + num_classes + NUM_ANGLES3:3], from_logits=True) vertices_confidence_loss = object_mask * K.binary_crossentropy( vertices_mask, raw_pred[..., 5 + num_classes + 2:5 + num_classes + NUM_ANGLES3:3], from_logits=True) xy_loss = K.sum(xy_loss) / mf wh_loss = K.sum(wh_loss) / mf class_loss = K.sum(class_loss) / mf confidence_loss = K.sum(confidence_loss) / mf vertices_confidence_loss = K.sum(vertices_confidence_loss) / mf polygon_loss = K.sum(polygon_loss_x) / mf + K.sum(polygon_loss_y) / mf diou_loss = K.sum( object_mask * box_loss_scale * (1 - box_diou(pred_box, y_true[layer][..., 0:4]))) / mf loss += (xy_loss + wh_loss + confidence_loss + class_loss + 0.2 * polygon_loss + 0.2 * vertices_confidence_loss) / ( K.sum(object_mask) + 1) * mf return loss
def yolo_loss(args, anchors, num_classes, ignore_thresh=.5): '''Return yolo_loss tensor Parameters ---------- yolo_outputs: list of tensor, the output of yolo_body y_true: list of array, the output of preprocess_true_boxes anchors: array, shape=(T, 2), wh num_classes: integer ignore_thresh: float, the iou threshold whether to ignore object confidence loss Returns ------- loss: tensor, shape=(1,) ''' yolo_outputs = args[:3] y_true = args[3:] anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] input_shape = K.cast( K.shape(yolo_outputs[0])[1:3] * 32, K.dtype(y_true[0])) grid_shapes = [ K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0])) for l in range(3) ] loss = 0 m = K.shape(yolo_outputs[0])[0] for l in range(3): object_mask = y_true[l][..., 4:5] true_class_probs = y_true[l][..., 5:] pred_xy, pred_wh, pred_confidence, pred_class_probs = yolo_head( yolo_outputs[l], anchors[anchor_mask[l]], num_classes, input_shape) pred_box = K.concatenate([pred_xy, pred_wh]) # Darknet box loss. xy_delta = (y_true[l][..., :2] - pred_xy) * grid_shapes[l][::-1] wh_delta = K.log(y_true[l][..., 2:4]) - K.log(pred_wh) # Avoid log(0)=-inf. wh_delta = K.switch(object_mask, wh_delta, K.zeros_like(wh_delta)) box_delta = K.concatenate([xy_delta, wh_delta], axis=-1) box_delta_scale = 2 - y_true[l][..., 2:3] * y_true[l][..., 3:4] # Find ignore mask, iterate over each of batch. ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True) object_mask_bool = K.cast(object_mask, 'bool') def loop_body(b, ignore_mask): true_box = tf.boolean_mask(y_true[l][b, ..., 0:4], object_mask_bool[b, ..., 0]) iou = box_iou(pred_box[b], true_box) best_iou = K.max(iou, axis=-1) ignore_mask = ignore_mask.write( b, K.cast(best_iou < ignore_thresh, K.dtype(true_box))) return b + 1, ignore_mask _, ignore_mask = K.control_flow_ops.while_loop(lambda b, *args: b < m, loop_body, [0, ignore_mask]) ignore_mask = ignore_mask.stack() ignore_mask = K.expand_dims(ignore_mask, -1) box_loss = object_mask * K.square(box_delta * box_delta_scale) confidence_loss = object_mask * K.square(1-pred_confidence) + \ (1-object_mask) * K.square(0-pred_confidence) * ignore_mask class_loss = object_mask * K.square(true_class_probs - pred_class_probs) loss += K.sum(box_loss) + K.sum(confidence_loss) + K.sum(class_loss) return loss / K.cast(m, K.dtype(loss))
def yolo_loss(args, anchors, num_classes, ignore_thresh=.5, print_loss=False): '''Return yolo_loss tensor Parameters ---------- yolo_outputs: list of tensor, the output of yolo_body or tiny_yolo_body y_true: list of array, the output of preprocess_true_boxes anchors: array, shape=(N, 2), wh num_classes: integer ignore_thresh: float, the iou threshold whether to ignore object confidence loss Returns ------- loss: tensor, shape=(1,) ''' num_layers = len(anchors) // 3 # default setting yolo_outputs = args[:num_layers] y_true = args[num_layers:] anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2] ] if num_layers == 3 else [[3, 4, 5], [1, 2, 3]] input_shape = K.cast( K.shape(yolo_outputs[0])[1:3] * 32, K.dtype(y_true[0])) grid_shapes = [ K.cast(K.shape(yolo_outputs[l])[1:3], K.dtype(y_true[0])) for l in range(num_layers) ] loss = 0 m = K.shape(yolo_outputs[0])[0] # batch size, tensor mf = K.cast(m, K.dtype(yolo_outputs[0])) for l in range(num_layers): object_mask = y_true[l][..., 4:5] true_class_probs = y_true[l][..., 5:] grid, raw_pred, pred_xy, pred_wh = yolo_head(yolo_outputs[l], anchors[anchor_mask[l]], num_classes, input_shape, calc_loss=True) pred_box = K.concatenate([pred_xy, pred_wh]) # Darknet raw box to calculate loss. raw_true_xy = y_true[l][..., :2] * grid_shapes[l][::-1] - grid raw_true_wh = K.log(y_true[l][..., 2:4] / anchors[anchor_mask[l]] * input_shape[::-1]) raw_true_wh = K.switch(object_mask, raw_true_wh, K.zeros_like(raw_true_wh)) # avoid log(0)=-inf box_loss_scale = 2 - y_true[l][..., 2:3] * y_true[l][..., 3:4] # Find ignore mask, iterate over each of batch. ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True) object_mask_bool = K.cast(object_mask, 'bool') def loop_body(b, ignore_mask): true_box = tf.boolean_mask(y_true[l][b, ..., 0:4], object_mask_bool[b, ..., 0]) iou = box_iou(pred_box[b], true_box) best_iou = K.max(iou, axis=-1) ignore_mask = ignore_mask.write( b, K.cast(best_iou < ignore_thresh, K.dtype(true_box))) return b + 1, ignore_mask _, ignore_mask = K.control_flow_ops.while_loop(lambda b, *args: b < m, loop_body, [0, ignore_mask]) ignore_mask = ignore_mask.stack() ignore_mask = K.expand_dims(ignore_mask, -1) # K.binary_crossentropy is helpful to avoid exp overflow. xy_loss = object_mask * box_loss_scale * K.binary_crossentropy( raw_true_xy, raw_pred[..., 0:2], from_logits=True) wh_loss = object_mask * box_loss_scale * 0.5 * K.square( raw_true_wh - raw_pred[..., 2:4]) confidence_loss = object_mask * K.binary_crossentropy(object_mask, raw_pred[...,4:5], from_logits=True)+ \ (1-object_mask) * K.binary_crossentropy(object_mask, raw_pred[...,4:5], from_logits=True) * ignore_mask class_loss = object_mask * K.binary_crossentropy( true_class_probs, raw_pred[..., 5:], from_logits=True) xy_loss = K.sum(xy_loss) / mf wh_loss = K.sum(wh_loss) / mf confidence_loss = K.sum(confidence_loss) / mf class_loss = K.sum(class_loss) / mf loss += xy_loss + wh_loss + confidence_loss + class_loss if print_loss: loss = tf.Print(loss, [ loss, xy_loss, wh_loss, confidence_loss, class_loss, K.sum(ignore_mask) ], message='loss: ') return loss