def call(self, inputs):
# Crop boxes [batch, num_boxes, (y1, x1, y2, x2)] in normalized coords
boxes = inputs[0]
# Image meta
# Holds details about the image. See compose_image_meta()
image_meta = inputs[1]
# Feature Maps. List of feature maps from different level of the
# feature pyramid. Each is [batch, height, width, channels]
feature_maps = inputs[2:]
# Assign each ROI to a level in the pyramid based on the ROI area.
y1, x1, y2, x2 = tf.split(boxes, 4, axis=2)
h = y2 - y1
w = x2 - x1
# Use shape of first image. Images in a batch must have the same size.
image_shape = parse_image_meta_graph(image_meta)['image_shape'][0]
# Equation 1 in the Feature Pyramid Networks paper. Account for
# the fact that our coordinates are normalized here.
# e.g. a 224x224 ROI (in pixels) maps to P4
image_area = tf.cast(image_shape[0] * image_shape[1], tf.float32)
roi_level = log2_graph(tf.sqrt(h * w) / (224.0 / tf.sqrt(image_area)))
roi_level = tf.minimum(
5, tf.maximum(2, 4 + tf.cast(tf.round(roi_level), tf.int32)))
roi_level = tf.squeeze(roi_level, 2)
# Loop through levels and apply ROI pooling to each. P2 to P5.
pooled = []
box_to_level = []
for i, level in enumerate(range(2, 6)):
ix = tf.where(tf.equal(roi_level, level))
level_boxes = tf.gather_nd(boxes, ix)
# Box indices for crop_and_resize.
box_indices = tf.cast(ix[:, 0], tf.int32)
# Keep track of which box is mapped to which level
box_to_level.append(ix)
# Stop gradient propogation to ROI proposals
level_boxes = tf.stop_gradient(level_boxes)
box_indices = tf.stop_gradient(box_indices)
# Crop and Resize
# From Mask R-CNN paper: "We sample four regular locations, so
# that we can evaluate either max or average pooling. In fact,
# interpolating only a single value at each bin center (without
# pooling) is nearly as effective."
#
# Here we use the simplified approach of a single value per bin,
# which is how it's done in tf.crop_and_resize()
# Result: [batch * num_boxes, pool_height, pool_width, channels]
pooled.append(
tf.image.crop_and_resize(feature_maps[i],
level_boxes,
box_indices,
self.pool_shape,
method="bilinear"))
# Pack pooled features into one tensor
pooled = tf.concat(pooled, axis=0)
# Pack box_to_level mapping into one array and add another
# column representing the order of pooled boxes
box_to_level = tf.concat(box_to_level, axis=0)
box_range = tf.expand_dims(tf.range(tf.shape(box_to_level)[0]), 1)
box_to_level = tf.concat([tf.cast(box_to_level, tf.int32), box_range],
axis=1)
# Rearrange pooled features to match the order of the original boxes
# Sort box_to_level by batch then box index
# TF doesn't have a way to sort by two columns, so merge them and sort.
sorting_tensor = box_to_level[:, 0] * 100000 + box_to_level[:, 1]
ix = tf.nn.top_k(sorting_tensor,
k=tf.shape(box_to_level)[0]).indices[::-1]
ix = tf.gather(box_to_level[:, 2], ix)
pooled = tf.gather(pooled, ix)
# Re-add the batch dimension
shape = tf.concat([tf.shape(boxes)[:2], tf.shape(pooled)[1:]], axis=0)
pooled = tf.reshape(pooled, shape)
return pooled
def build(self, mode: str, config: ShapesConfig):
"""Build Mask R-CNN architecture.
input_shape: The shape of the input image.
mode: Either "training" or "inference". The inputs and
outputs of the model differ accordingly.
"""
assert mode in ['training', 'inference']
# Image size must be dividable by 2 multiple times
h, w = config.IMAGE_SHAPE[:2]
if h / 2**6 != int(h / 2**6) or w / 2**6 != int(w / 2**6):
raise Exception(
"Image size must be dividable by 2 at least 6 times "
"to avoid fractions when downscaling and upscaling."
"For example, use 256, 320, 384, 448, 512, ... etc. ")
# Inputs
input_image = KL.Input(shape=[None, None, config.IMAGE_SHAPE[2]],
name="input_image")
input_image_meta = KL.Input(shape=[config.IMAGE_META_SIZE],
name="input_image_meta")
if mode == "training":
# RPN GT
input_rpn_match = KL.Input(shape=[None, 1],
name="input_rpn_match",
dtype=tf.int32)
input_rpn_bbox = KL.Input(shape=[None, 4],
name="input_rpn_bbox",
dtype=tf.float32)
# Detection GT (class IDs, bounding boxes, and masks)
# 1. GT Class IDs (zero padded)
input_gt_class_ids = KL.Input(shape=[None],
name="input_gt_class_ids",
dtype=tf.int32)
# 2. GT Boxes in pixels (zero padded)
# [batch, MAX_GT_INSTANCES, (y1, x1, y2, x2)] in image coordinates
input_gt_boxes = KL.Input(shape=[None, 4],
name="input_gt_boxes",
dtype=tf.float32)
# Normalize coordinates
gt_boxes = KL.Lambda(lambda x: norm_boxes_graph(
x,
K.shape(input_image)[1:3]))(input_gt_boxes)
# 3. GT Masks (zero padded)
# [batch, height, width, MAX_GT_INSTANCES]
if config.USE_MINI_MASK:
input_gt_masks = KL.Input(shape=[
config.MINI_MASK_SHAPE[0], config.MINI_MASK_SHAPE[1], None
],
name="input_gt_masks",
dtype=bool)
else:
input_gt_masks = KL.Input(
shape=[config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1], None],
name="input_gt_masks",
dtype=bool)
elif mode == "inference":
# Anchors in normalized coordinates
input_anchors = KL.Input(shape=[None, 4], name="input_anchors")
anchors = input_anchors
# Build the shared convolutional layers.
# Bottom-up Layers
# Returns a list of the last layers of each stage, 5 in total.
# Don't create the thead (stage 5), so we pick the 4th item in the list.
if callable(config.BACKBONE):
_, C2, C3, C4, C5 = config.BACKBONE(input_image,
stage5=True,
train_bn=config.TRAIN_BN)
else:
_, C2, C3, C4, C5 = resnet_graph(input_image,
config.BACKBONE,
stage5=True,
train_bn=config.TRAIN_BN)
# Top-down Layers
# TODO: add assert to varify feature map sizes match what's in config
P5 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1),
name='fpn_c5p5')(C5)
P4 = KL.Add(name="fpn_p4add")([
KL.UpSampling2D(size=(2, 2), name="fpn_p5upsampled")(P5),
KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1),
name='fpn_c4p4')(C4)
])
P3 = KL.Add(name="fpn_p3add")([
KL.UpSampling2D(size=(2, 2), name="fpn_p4upsampled")(P4),
KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1),
name='fpn_c3p3')(C3)
])
P2 = KL.Add(name="fpn_p2add")([
KL.UpSampling2D(size=(2, 2), name="fpn_p3upsampled")(P3),
KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1),
name='fpn_c2p2')(C2)
])
# Attach 3x3 conv to all P layers to get the final feature maps.
P2 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3),
padding="SAME",
name="fpn_p2")(P2)
P3 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3),
padding="SAME",
name="fpn_p3")(P3)
P4 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3),
padding="SAME",
name="fpn_p4")(P4)
P5 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3),
padding="SAME",
name="fpn_p5")(P5)
# P6 is used for the 5th anchor scale in RPN. Generated by
# subsampling from P5 with stride of 2.
P6 = KL.MaxPooling2D(pool_size=(1, 1), strides=2, name="fpn_p6")(P5)
# Note that P6 is used in RPN, but not in the classifier heads.
rpn_feature_maps = [P2, P3, P4, P5, P6]
mrcnn_feature_maps = [P2, P3, P4, P5]
# Anchors
if mode == "training":
anchors = self.get_anchors(config.IMAGE_SHAPE)
# Duplicate across the batch dimension because Keras requires it
# TODO: can this be optimized to avoid duplicating the anchors?
anchors = np.broadcast_to(anchors,
(config.BATCH_SIZE, ) + anchors.shape)
# A hack to get around Keras's bad support for constants
anchors = KL.Lambda(lambda x: tf.Variable(anchors),
name="anchors")(input_image)
# RPN Model
rpn = build_rpn_model(config.RPN_ANCHOR_STRIDE,
len(config.RPN_ANCHOR_RATIOS),
config.TOP_DOWN_PYRAMID_SIZE)
# Loop through pyramid layers
layer_outputs = [] # list of lists
for p in rpn_feature_maps:
layer_outputs.append(rpn([p]))
# Concatenate layer outputs
# Convert from list of lists of level outputs to list of lists
# of outputs across levels.
# e.g. [[a1, b1, c1], [a2, b2, c2]] => [[a1, a2], [b1, b2], [c1, c2]]
output_names = ["rpn_class_logits", "rpn_class", "rpn_bbox"]
outputs = list(zip(*layer_outputs))
outputs = [
KL.Concatenate(axis=1, name=n)(list(o))
for o, n in zip(outputs, output_names)
]
rpn_class_logits, rpn_class, rpn_bbox = outputs
# Generate proposals
# Proposals are [batch, N, (y1, x1, y2, x2)] in normalized coordinates
# and zero padded.
proposal_count = config.POST_NMS_ROIS_TRAINING if mode == "training" \
else config.POST_NMS_ROIS_INFERENCE
rpn_rois = ProposalLayer(proposal_count=proposal_count,
nms_threshold=config.RPN_NMS_THRESHOLD,
name="ROI",
config=config)([rpn_class, rpn_bbox, anchors])
if mode == "training":
# Class ID mask to mark class IDs supported by the dataset the image
# came from.
active_class_ids = KL.Lambda(lambda x: parse_image_meta_graph(x)[
"active_class_ids"])(input_image_meta)
if not config.USE_RPN_ROIS:
# Ignore predicted ROIs and use ROIs provided as an input.
input_rois = KL.Input(shape=[config.POST_NMS_ROIS_TRAINING, 4],
name="input_roi",
dtype=np.int32)
# Normalize coordinates
target_rois = KL.Lambda(lambda x: norm_boxes_graph(
x,
K.shape(input_image)[1:3]))(input_rois)
else:
target_rois = rpn_rois
# Generate detection targets
# Subsamples proposals and generates target outputs for training
# Note that proposal class IDs, gt_boxes, and gt_masks are zero
# padded. Equally, returned rois and targets are zero padded.
rois, target_class_ids, target_bbox, target_mask = \
DetectionTargetLayer(config, name="proposal_targets")([
target_rois, input_gt_class_ids, gt_boxes, input_gt_masks])
# Network Heads
# TODO: verify that this handles zero padded ROIs
mrcnn_class_logits, mrcnn_class, mrcnn_bbox = \
fpn_classifier_graph(rois, mrcnn_feature_maps, input_image_meta,
config.POOL_SIZE, config.NUM_CLASSES,
train_bn=config.TRAIN_BN,
fc_layers_size=config.FPN_CLASSIF_FC_LAYERS_SIZE)
mrcnn_mask = build_fpn_mask_graph(rois,
mrcnn_feature_maps,
input_image_meta,
config.MASK_POOL_SIZE,
config.NUM_CLASSES,
train_bn=config.TRAIN_BN)
# TODO: clean up (use tf.identify if necessary)
output_rois = KL.Lambda(lambda x: x * 1, name="output_rois")(rois)
# Losses
rpn_class_loss = KL.Lambda(lambda x: rpn_class_loss_graph(*x),
name="rpn_class_loss")(
[input_rpn_match, rpn_class_logits])
rpn_bbox_loss = KL.Lambda(
lambda x: rpn_bbox_loss_graph(config, *x),
name="rpn_bbox_loss")(
[input_rpn_bbox, input_rpn_match, rpn_bbox])
class_loss = KL.Lambda(lambda x: mrcnn_class_loss_graph(*x),
name="mrcnn_class_loss")([
target_class_ids, mrcnn_class_logits,
active_class_ids
])
bbox_loss = KL.Lambda(lambda x: mrcnn_bbox_loss_graph(*x),
name="mrcnn_bbox_loss")([
target_bbox, target_class_ids, mrcnn_bbox
])
mask_loss = KL.Lambda(lambda x: mrcnn_mask_loss_graph(*x),
name="mrcnn_mask_loss")([
target_mask, target_class_ids, mrcnn_mask
])
# Model
inputs = [
input_image, input_image_meta, input_rpn_match, input_rpn_bbox,
input_gt_class_ids, input_gt_boxes, input_gt_masks
]
if not config.USE_RPN_ROIS:
inputs.append(input_rois)
outputs = [
rpn_class_logits, rpn_class, rpn_bbox, mrcnn_class_logits,
mrcnn_class, mrcnn_bbox, mrcnn_mask, rpn_rois, output_rois,
rpn_class_loss, rpn_bbox_loss, class_loss, bbox_loss, mask_loss
]
model = KM.Model(inputs, outputs, name='mask_rcnn')
else:
# Network Heads
# Proposal classifier and BBox regressor heads
mrcnn_class_logits, mrcnn_class, mrcnn_bbox = \
fpn_classifier_graph(rpn_rois, mrcnn_feature_maps, input_image_meta,
config.POOL_SIZE, config.NUM_CLASSES,
train_bn=config.TRAIN_BN,
fc_layers_size=config.FPN_CLASSIF_FC_LAYERS_SIZE)
# Detections
# output is [batch, num_detections, (y1, x1, y2, x2, class_id, score)] in
# normalized coordinates
detections = DetectionLayer(config, name="mrcnn_detection")(
[rpn_rois, mrcnn_class, mrcnn_bbox, input_image_meta])
# Create masks for detections
detection_boxes = KL.Lambda(lambda x: x[..., :4])(detections)
mrcnn_mask = build_fpn_mask_graph(detection_boxes,
mrcnn_feature_maps,
input_image_meta,
config.MASK_POOL_SIZE,
config.NUM_CLASSES,
train_bn=config.TRAIN_BN)
model = KM.Model([input_image, input_image_meta, input_anchors], [
detections, mrcnn_class, mrcnn_bbox, mrcnn_mask, rpn_rois,
rpn_class, rpn_bbox
],
name='mask_rcnn')
# Add multi-GPU support.
if config.GPU_COUNT > 1:
from mrcnn.parallel_model import ParallelModel
model = ParallelModel(model, config.GPU_COUNT)
return model
def build(self, mode, config):
"""Build the Mask R-CNN teacher-student architecture for mimic training.
"""
assert mode in ['training', 'inference']
# Image size must be dividable by 2 multiple times
h, w = config.IMAGE_SHAPE[:2]
if h / 2 ** 6 != int(h / 2 ** 6) or w / 2 ** 6 != int(w / 2 ** 6):
raise Exception("Image size must be dividable by 2 at least 6 times "
"to avoid fractions when downscaling and upscaling."
"For example, use 256, 320, 384, 448, 512, ... etc. ")
# Input
input_image = KL.Input(shape=[None, None, config.IMAGE_SHAPE[2]], name="input_image")
input_image_meta = KL.Input(shape=[config.IMAGE_META_SIZE], name="input_image_meta")
if mode == "training":
# RPN GT
input_rpn_match = KL.Input(shape=[None, 1], name="input_rpn_match", dtype=tf.int32)
input_rpn_bbox = KL.Input(shape=[None, 4], name="input_rpn_bbox", dtype=tf.float32)
# Detection GT (class IDs, bounding boxes, and masks)
# 1. GT Class IDs (zero padded)
input_gt_class_ids = KL.Input(shape=[None], name="input_gt_class_ids", dtype=tf.int32)
# 2. GT Boxes in pixels (zero padded)
# [batch, MAX_GT_INSTANCES, (y1, x1, y2, x2)] in image coordinates
input_gt_boxes = KL.Input(shape=[None, 4], name="input_gt_boxes", dtype=tf.float32)
# Normalize coordinates
gt_boxes = KL.Lambda(lambda x: modellib.norm_boxes_graph(x, K.shape(input_image)[1:3]))(input_gt_boxes)
# 3. GT Masks (zero padded)
# [batch, height, width, MAX_GT_INSTANCES]
if config.USE_MINI_MASK:
input_gt_masks = KL.Input(
shape=[config.MINI_MASK_SHAPE[0],
config.MINI_MASK_SHAPE[1], None],
name="input_gt_masks", dtype=bool)
else:
input_gt_masks = KL.Input(
shape=[config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1], None],
name="input_gt_masks", dtype=bool)
# Class ID mask to mark class IDs supported by the dataset the image came from.
active_class_ids = KL.Lambda(
lambda x: modellib.parse_image_meta_graph(x)["active_class_ids"]
)(input_image_meta)
elif mode == "inference":
pass
# Build the architecture of the teacher model
# Backbone
_, C2, C3, C4, C5 = modellib.resnet_graph(input_image, config.TEACHER_BACKBONE, stage5=True, train_bn=False)
# Top-down Layers
P5 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c5p5')(C5)
P4 = KL.Add(name="fpn_p4add")([
KL.UpSampling2D(size=(2, 2), name="fpn_p5upsampled")(P5),
KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c4p4')(C4)])
P3 = KL.Add(name="fpn_p3add")([
KL.UpSampling2D(size=(2, 2), name="fpn_p4upsampled")(P4),
KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c3p3')(C3)])
P2 = KL.Add(name="fpn_p2add")([
KL.UpSampling2D(size=(2, 2), name="fpn_p3upsampled")(P3),
KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name='fpn_c2p2')(C2)])
# Attach 3x3 conv to all P layers to get the final feature maps.
P2 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p2")(P2) # N x 256 x 256 x 256
P3 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p3")(P3) # N x 128 x 128 x 256
P4 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p4")(P4) # N x 64 x 64 x 256
P5 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name="fpn_p5")(P5) # N x 32 x 32 x 256
# Note that P6 is used in RPN, but not in the classifier heads.
t_mrcnn_feature_maps = [P2, P3, P4, P5]
# Build the architecture of the student model
s_prefix = 's_'
# Backbone
_, S2, S3, S4, S5 = s_resnet_graph(input_image, config.STUDENT_BACKBONE, prefix=s_prefix, train_bn=None)
# Top-down Layers
Q5 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name=s_prefix + 'fpn_s5q5')(S5)
Q4 = KL.Add(name=s_prefix + "fpn_q4add")([
KL.UpSampling2D(size=(2, 2), name=s_prefix + "fpn_q5upsampled")(Q5),
KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name=s_prefix + 'fpn_s4q4')(S4)])
Q3 = KL.Add(name=s_prefix + "fpn_q3add")([
KL.UpSampling2D(size=(2, 2), name=s_prefix + "fpn_q4upsampled")(Q4),
KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name=s_prefix + 'fpn_s3q3')(S3)])
Q2 = KL.Add(name=s_prefix + "fpn_q2add")([
KL.UpSampling2D(size=(2, 2), name=s_prefix + "fpn_q3upsampled")(Q3),
KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (1, 1), name=s_prefix + 'fpn_s2q2')(S2)])
# Attach 3x3 conv to all Q layers to get the final feature maps.
Q2 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name=s_prefix + "fpn_q2")(
Q2) # N x 256 x 256 x 256
Q3 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name=s_prefix + "fpn_q3")(
Q3) # N x 128 x 128 x 256
Q4 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name=s_prefix + "fpn_q4")(
Q4) # N x 64 x 64 x 256
Q5 = KL.Conv2D(config.TOP_DOWN_PYRAMID_SIZE, (3, 3), padding="SAME", name=s_prefix + "fpn_q5")(
Q5) # N x 32 x 32 x 256
# Q6 is used for the 5th anchor scale in RPN. Generated by
# subsampling from P5 with stride of 2.
Q6 = KL.MaxPooling2D(pool_size=(1, 1), strides=2, name=s_prefix + "fpn_q6")(Q5)
# Note that P6 is used in RPN, but not in the classifier heads.
s_rpn_feature_maps = [Q2, Q3, Q4, Q5, Q6]
s_mrcnn_feature_maps = [Q2, Q3, Q4, Q5]
# RPN Model
s_rpn = s_build_rpn_model(config.RPN_ANCHOR_STRIDE,
len(config.RPN_ANCHOR_RATIOS), config.TOP_DOWN_PYRAMID_SIZE, prefix=s_prefix)
# Loop through pyramid layers
s_layer_outputs = [] # list of lists
for p in s_rpn_feature_maps:
s_layer_outputs.append(s_rpn([p]))
# Concatenate layer outputs
# Convert from list of lists of level outputs to list of lists of outputs across levels.
# e.g. [[a1, b1, c1], [a2, b2, c2]] => [[a1, a2], [b1, b2], [c1, c2]]
s_output_names = ["s_rpn_class_logits", "s_rpn_class", "s_rpn_bbox"]
s_outputs = list(zip(*s_layer_outputs))
s_outputs = [KL.Concatenate(axis=1, name=n)(list(o))
for o, n in zip(s_outputs, s_output_names)]
s_rpn_class_logits, s_rpn_class, s_rpn_bbox = s_outputs
# Proposals
anchors = self.get_anchors(config.IMAGE_SHAPE)
# Duplicate across the batch dimension because Keras requires it
# TODO: can this be optimized to avoid duplicating the anchors?
anchors = np.broadcast_to(anchors, (config.BATCH_SIZE,) + anchors.shape)
# A hack to get around Keras's bad support for constants
anchors = KL.Lambda(lambda x: tf.Variable(anchors), name="anchors")(input_image)
# Generate proposals
# Proposals are [batch, N, (y1, x1, y2, x2)] in normalized coordinates
# and zero padded.
proposal_count = config.POST_NMS_ROIS_TRAINING if mode == "training" \
else config.POST_NMS_ROIS_INFERENCE
s_rpn_rois = s_ProposalLayer(
proposal_count=proposal_count,
nms_threshold=config.RPN_NMS_THRESHOLD,
name="s_ROI",
config=config)([s_rpn_class, s_rpn_bbox, anchors])
# Generate detection targets
# Subsamples proposals and generates target outputs for training
# Note that proposal class IDs, gt_boxes, and gt_masks are zero
# padded. Equally, returned rois and targets are zero padded.
s_rois, target_class_ids, target_bbox, target_mask = \
modellib.DetectionTargetLayer(config, name="proposal_targets")([
s_rpn_rois, input_gt_class_ids, gt_boxes, input_gt_masks])
# s_rois: [batch, TRAIN_ROIS_PER_IMAGE, (y1, x1, y2, x2)] in normalized coordinates
# Add a transformer layer to the feature maps from the student model before comparison with the teacher model
transformer = build_transformer_layer(config.TOP_DOWN_PYRAMID_SIZE)
s_transformed_feature_maps = [transformer([p]) for p in s_mrcnn_feature_maps]
# Losses
rpn_class_loss = KL.Lambda(lambda x: modellib.rpn_class_loss_graph(*x), name="rpn_class_loss")(
[input_rpn_match, s_rpn_class_logits])
rpn_bbox_loss = KL.Lambda(lambda x: modellib.rpn_bbox_loss_graph(config, *x), name="rpn_bbox_loss")(
[input_rpn_bbox, input_rpn_match, s_rpn_bbox])
rpn_mimic_loss = KL.Lambda(lambda x: rpn_mimic_loss_graph(config, *x), name="rpn_mimic_loss")(
[s_transformed_feature_maps[0], s_transformed_feature_maps[1],
s_transformed_feature_maps[2], s_transformed_feature_maps[3],
t_mrcnn_feature_maps[0], t_mrcnn_feature_maps[1],
t_mrcnn_feature_maps[2], t_mrcnn_feature_maps[3],
s_rois, input_image_meta])
# Model
inputs = [input_image, input_image_meta, input_rpn_match, input_rpn_bbox,
input_gt_class_ids, input_gt_boxes, input_gt_masks]
outputs = [s_rpn_class_logits, s_rpn_class, s_rpn_bbox,
s_rpn_rois, rpn_class_loss, rpn_bbox_loss, rpn_mimic_loss]
model = KM.Model(inputs, outputs, name='mimic_mask_rcnn')
# Add multi-GPU support.
if config.GPU_COUNT > 1:
from mrcnn.parallel_model import ParallelModel
model = ParallelModel(model, config.GPU_COUNT)
return model
def build(self, mode, config):
"""Build Mask R-CNN architecture.
input_shape: The shape of the input image.
mode: Either "training" or "inference". The inputs and
outputs of the model differ accordingly.
"""
assert mode in ['training', 'inference']
# Image size must be dividable by 2 multiple times
h, w = config.IMAGE_SHAPE[:2]
if h / 2**6 != int(h / 2**6) or w / 2**6 != int(w / 2**6):
raise Exception("Image size must be dividable by 2 at least 6 times "
"to avoid fractions when downscaling and upscaling."
"For example, use 256, 320, 384, 448, 512, ... etc. ")
# Inputs
input_image = KL.Input(
shape=config.IMAGE_SHAPE.tolist(), name="input_image")
# CHANGE: add target input
if not config.NUM_TARGETS:
config.NUM_TARGETS = 1
input_target = KL.Input(
shape=[config.NUM_TARGETS] + config.TARGET_SHAPE.tolist(), name="input_target")
input_image_meta = KL.Input(shape=[config.IMAGE_META_SIZE],
name="input_image_meta")
if mode == "training":
# RPN GT
input_rpn_match = KL.Input(
shape=[None, 1], name="input_rpn_match", dtype=tf.int32)
input_rpn_bbox = KL.Input(
shape=[None, 4], name="input_rpn_bbox", dtype=tf.float32)
# Detection GT (class IDs, bounding boxes, and masks)
# 1. GT Class IDs (zero padded)
input_gt_class_ids = KL.Input(
shape=[None], name="input_gt_class_ids", dtype=tf.int32)
# 2. GT Boxes in pixels (zero padded)
# [batch, MAX_GT_INSTANCES, (y1, x1, y2, x2)] in image coordinates
input_gt_boxes = KL.Input(
shape=[None, 4], name="input_gt_boxes", dtype=tf.float32)
# Normalize coordinates
gt_boxes = KL.Lambda(lambda x: modellib.norm_boxes_graph(
x, K.shape(input_image)[1:3]))(input_gt_boxes)
# 3. GT Masks (zero padded)
# [batch, height, width, MAX_GT_INSTANCES]
if config.USE_MINI_MASK:
input_gt_masks = KL.Input(
shape=[config.MINI_MASK_SHAPE[0],
config.MINI_MASK_SHAPE[1], None],
name="input_gt_masks", dtype=bool)
else:
input_gt_masks = KL.Input(
shape=[config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1], None],
name="input_gt_masks", dtype=bool)
elif mode == "inference":
# Anchors in normalized coordinates
input_anchors = KL.Input(shape=[None, 4], name="input_anchors")
# Build the shared convolutional layers.
# CHANGE: Use weightshared FPN model for image and target
# Create FPN Model
resnet = build_resnet_model(self.config)
fpn = build_fpn_model(feature_maps=self.config.FPN_FEATUREMAPS)
# Create Image FP
_, IC2, IC3, IC4, IC5 = resnet(input_image)
IP2, IP3, IP4, IP5, IP6 = fpn([IC2, IC3, IC4, IC5])
# Create Target FR
input_targets = [KL.Lambda(lambda x: x[:,idx,...])(input_target) for idx in range(input_target.shape[1])]
for k, one_target in enumerate(input_targets):
_, TC2, TC3, TC4, TC5 = resnet(one_target)
out = fpn([TC2, TC3, TC4, TC5])
if k == 0:
target_pyramid = out
else:
target_pyramid = [KL.Add(name="target_adding_{}_{}".format(k, i))(
[target_pyramid[i], out[i]]) for i in range(len(out))]
TP2, TP3, TP4, TP5, TP6 = [KL.Lambda(lambda x: x / config.NUM_TARGETS)(
target_pyramid[i]) for i in range(len(target_pyramid))]
# one_target = KL.Lambda(lambda x: x[:,0,...])(input_target)
# one_target = input_target[:,0,...]
# _, TC2, TC3, TC4, TC5 = resnet(one_target)
# TP2, TP3, TP4, TP5, TP6 = fpn([TC2, TC3, TC4, TC5])
# CHANGE: add siamese distance copmputation
# Combine FPs using L1 distance
P2 = l1_distance_graph(IP2, TP2, feature_maps = 3*self.config.FPN_FEATUREMAPS//2, name='P2')
P3 = l1_distance_graph(IP3, TP3, feature_maps = 3*self.config.FPN_FEATUREMAPS//2, name='P3')
P4 = l1_distance_graph(IP4, TP4, feature_maps = 3*self.config.FPN_FEATUREMAPS//2, name='P4')
P5 = l1_distance_graph(IP5, TP5, feature_maps = 3*self.config.FPN_FEATUREMAPS//2, name='P5')
P6 = l1_distance_graph(IP6, TP6, feature_maps = 3*self.config.FPN_FEATUREMAPS//2, name='P6')
# Note that P6 is used in RPN, but not in the classifier heads.
rpn_feature_maps = [P2, P3, P4, P5, P6]
mrcnn_feature_maps = [P2, P3, P4, P5]
# Anchors
if mode == "training":
anchors = self.get_anchors(config.IMAGE_SHAPE)
# Duplicate across the batch dimension because Keras requires it
# TODO: can this be optimized to avoid duplicating the anchors?
anchors = np.broadcast_to(anchors, (config.BATCH_SIZE,) + anchors.shape)
# A hack to get around Keras's bad support for constants
anchors = KL.Lambda(lambda x: tf.Variable(anchors), name="anchors")(input_image)
else:
anchors = input_anchors
# RPN Model
# CHANGE: Set number of filters to [3*self.config.FPN_FEATUREMAPS//2]
rpn = modellib.build_rpn_model(config.RPN_ANCHOR_STRIDE,
len(config.RPN_ANCHOR_RATIOS), 3*self.config.FPN_FEATUREMAPS//2)
# Loop through pyramid layers
layer_outputs = [] # list of lists
for p in rpn_feature_maps:
layer_outputs.append(rpn([p]))
# Concatenate layer outputs
# Convert from list of lists of level outputs to list of lists
# of outputs across levels.
# e.g. [[a1, b1, c1], [a2, b2, c2]] => [[a1, a2], [b1, b2], [c1, c2]]
output_names = ["rpn_class_logits", "rpn_class", "rpn_bbox"]
outputs = list(zip(*layer_outputs))
outputs = [KL.Concatenate(axis=1, name=n)(list(o))
for o, n in zip(outputs, output_names)]
rpn_class_logits, rpn_class, rpn_bbox = outputs
# Generate proposals
# Proposals are [batch, N, (y1, x1, y2, x2)] in normalized coordinates
# and zero padded.
proposal_count = config.POST_NMS_ROIS_TRAINING if mode == "training"\
else config.POST_NMS_ROIS_INFERENCE
rpn_rois = modellib.ProposalLayer(
proposal_count=proposal_count,
nms_threshold=config.RPN_NMS_THRESHOLD,
name="ROI",
config=config)([rpn_class, rpn_bbox, anchors])
if mode == "training":
# Class ID mask to mark class IDs supported by the dataset the image
# came from.
active_class_ids = KL.Lambda(
lambda x: modellib.parse_image_meta_graph(x)["active_class_ids"]
)(input_image_meta)
if not config.USE_RPN_ROIS:
# Ignore predicted ROIs and use ROIs provided as an input.
input_rois = KL.Input(shape=[config.POST_NMS_ROIS_TRAINING, 4],
name="input_roi", dtype=np.int32)
# Normalize coordinates
target_rois = KL.Lambda(lambda x: modellig.norm_boxes_graph(
x, K.shape(input_image)[1:3]))(input_rois)
else:
target_rois = rpn_rois
# Generate detection targets
# Subsamples proposals and generates target outputs for training
# Note that proposal class IDs, gt_boxes, and gt_masks are zero
# padded. Equally, returned rois and targets are zero padded.
rois, target_class_ids, target_bbox, target_mask =\
modellib.DetectionTargetLayer(config, name="proposal_targets")([
target_rois, input_gt_class_ids, gt_boxes, input_gt_masks])
# Network Heads
# TODO: verify that this handles zero padded ROIs
# CHANGE: reduce number of classes to 2
# CHANGE: replaced with custom 2 class function
mrcnn_class_logits, mrcnn_class, mrcnn_bbox =\
fpn_classifier_graph(rois, mrcnn_feature_maps, input_image_meta,
config.POOL_SIZE, num_classes=2,
train_bn=config.TRAIN_BN,
fc_layers_size=config.FPN_CLASSIF_FC_LAYERS_SIZE)
# CHANGE: reduce number of classes to 2
# CHANGE: replaced with custom 2 class function
if config.MODEL == 'mrcnn':
mrcnn_mask = fpn_mask_graph(rois, mrcnn_feature_maps,
input_image_meta,
config.MASK_POOL_SIZE,
num_classes=2,
train_bn=config.TRAIN_BN)
# TODO: clean up (use tf.identify if necessary)
output_rois = KL.Lambda(lambda x: x * 1, name="output_rois")(rois)
# Losses
rpn_class_loss = KL.Lambda(lambda x: modellib.rpn_class_loss_graph(*x), name="rpn_class_loss")(
[input_rpn_match, rpn_class_logits])
rpn_bbox_loss = KL.Lambda(lambda x: modellib.rpn_bbox_loss_graph(config, *x), name="rpn_bbox_loss")(
[input_rpn_bbox, input_rpn_match, rpn_bbox])
# CHANGE: use custom class loss without using active_class_ids
class_loss = KL.Lambda(lambda x: mrcnn_class_loss_graph(*x), name="mrcnn_class_loss")(
[target_class_ids, mrcnn_class_logits, active_class_ids])
bbox_loss = KL.Lambda(lambda x: modellib.mrcnn_bbox_loss_graph(*x), name="mrcnn_bbox_loss")(
[target_bbox, target_class_ids, mrcnn_bbox])
if config.MODEL == 'mrcnn':
mask_loss = KL.Lambda(lambda x: modellib.mrcnn_mask_loss_graph(*x), name="mrcnn_mask_loss")(
[target_mask, target_class_ids, mrcnn_mask])
# Model
# CHANGE: Added target to inputs
inputs = [input_image, input_image_meta, input_target,
input_rpn_match, input_rpn_bbox, input_gt_class_ids, input_gt_boxes, input_gt_masks]
if not config.USE_RPN_ROIS:
inputs.append(input_rois)
if config.MODEL == 'mrcnn':
outputs = [rpn_class_logits, rpn_class, rpn_bbox,
mrcnn_class_logits, mrcnn_class, mrcnn_bbox, mrcnn_mask,
rpn_rois, output_rois,
rpn_class_loss, rpn_bbox_loss, class_loss, bbox_loss, mask_loss]
elif config.MODEL =='frcnn':
outputs = [rpn_class_logits, rpn_class, rpn_bbox,
mrcnn_class_logits, mrcnn_class, mrcnn_bbox,
rpn_rois, output_rois,
rpn_class_loss, rpn_bbox_loss, class_loss, bbox_loss]
model = KM.Model(inputs, outputs, name='mask_rcnn')
else:
# Network Heads
# Proposal classifier and BBox regressor heads
# CHANGE: reduce number of classes to 2
# CHANGE: replaced with custom 2 class function
mrcnn_class_logits, mrcnn_class, mrcnn_bbox =\
fpn_classifier_graph(rpn_rois, mrcnn_feature_maps, input_image_meta,
config.POOL_SIZE, num_classes=2,
train_bn=config.TRAIN_BN,
fc_layers_size=config.FPN_CLASSIF_FC_LAYERS_SIZE)
# Detections
# output is [batch, num_detections, (y1, x1, y2, x2, class_id, score)] in
# normalized coordinates
detections = modellib.DetectionLayer(config, name="mrcnn_detection")(
[rpn_rois, mrcnn_class, mrcnn_bbox, input_image_meta])
# Create masks for detections
detection_boxes = KL.Lambda(lambda x: x[..., :4])(detections)
# CHANGE: reduce number of classes to 2
# CHANGE: replaced with custom 2 class function
if config.MODEL == 'mrcnn':
mrcnn_mask = fpn_mask_graph(detection_boxes, mrcnn_feature_maps,
input_image_meta,
config.MASK_POOL_SIZE,
num_classes=2,
train_bn=config.TRAIN_BN)
# CHANGE: Added target to the input
inputs = [input_image, input_image_meta, input_target, input_anchors]
if config.MODEL == 'mrcnn':
outputs = [detections, mrcnn_class, mrcnn_bbox,
mrcnn_mask, rpn_rois, rpn_class, rpn_bbox]
elif config.MODEL =='frcnn':
outputs = [detections, mrcnn_class, mrcnn_bbox,
rpn_rois, rpn_class, rpn_bbox]
model = KM.Model(inputs, outputs, name='mask_rcnn')
# Add multi-GPU support.
if config.GPU_COUNT > 1:
from mrcnn.parallel_model import ParallelModel
model = ParallelModel(model, config.GPU_COUNT)
return model
def rpn_mimic_loss_graph(config, Q2, Q3, Q4, Q5, P2, P3, P4, P5, boxes, image_meta):
student_maps = [Q2, Q3, Q4, Q5]
teacher_maps = [P2, P3, P4, P5]
num_boxes_batch = config.BATCH_SIZE * config.TRAIN_ROIS_PER_IMAGE
# Assign each ROI to a level in the pyramid based on the ROI area.
y1, x1, y2, x2 = tf.split(boxes, 4, axis=2) # [BATCH_SIZE, TRAIN_ROIS_PER_IMAGE, 1]
h = y2 - y1
w = x2 - x1
# Use shape of first image. Images in a batch must have the same size.
image_shape = modellib.parse_image_meta_graph(image_meta)['image_shape'][0]
# Equation 1 in the Feature Pyramid Networks paper. Account for
# the fact that our coordinates are normalized here.
# e.g. a 224x224 ROI (in pixels) maps to P4
image_area = tf.cast(image_shape[0] * image_shape[1], tf.float32)
roi_level = modellib.log2_graph(tf.sqrt(h * w) / (224.0 / tf.sqrt(image_area)))
roi_level = tf.minimum(5, tf.maximum(2, 4 + tf.cast(tf.round(roi_level), tf.int32))) # 4 + log2(roi_h * roi_w)
roi_level = tf.squeeze(roi_level, 2) # [BATCH_SIZE, TRAIN_ROIS_PER_IMAGE]
# init the loss for later accumulation of every pyramid level
loss = 0.
for i, level in enumerate(range(2, 6)):
ix = tf.where(tf.equal(roi_level, level)) # [BATCH_SIZE, TRAIN_ROIS_PER_IMAGE]
level_boxes = tf.gather_nd(boxes, ix) # [num_boxes_level, 4]
# Box indices for crop_and_resize (specify which image the bbox refers to)
box_indices = tf.cast(ix[:, 0], tf.int32) # [num_boxes_level]
P = num_boxes_batch - tf.shape(level_boxes)[0]
level_boxes = tf.pad(level_boxes, [(0, P), (0, 0)]) # padded with zero
box_indices = tf.pad(box_indices, [(0, P)]) # padded with zero
level_y1, level_x1, level_y2, level_x2 = tf.split(level_boxes, 4, axis=-1)
map_h = image_shape[0] / config.BACKBONE_STRIDES[i]
map_w = image_shape[1] / config.BACKBONE_STRIDES[i]
# calculates the coordinate of bounding boxes at the pyramid level
level_y1_p = tf.cast(tf.squeeze(tf.floor(level_y1 * map_h), axis=-1), tf.int32) # [num_boxes_batch]
level_x1_p = tf.cast(tf.squeeze(tf.floor(level_x1 * map_w), axis=-1), tf.int32)
level_y2_p = tf.cast(tf.squeeze(tf.ceil(level_y2 * map_h), axis=-1), tf.int32)
level_x2_p = tf.cast(tf.squeeze(tf.ceil(level_x2 * map_w), axis=-1), tf.int32)
level_h_p = level_y2_p - level_y1_p
level_w_p = level_x2_p - level_x1_p
# iterate over bounding boxes and accumulate the l2-distance loss
def cond(idx, _):
# return tf.less(idx, level_boxes.shape[0])
return tf.less(idx, num_boxes_batch)
def body(idx, loss):
s_cropped = tf.slice(student_maps[i], [box_indices[idx], level_y1_p[idx], level_x1_p[idx], 0],
[1, level_h_p[idx], level_w_p[idx], -1])
t_cropped = tf.slice(teacher_maps[i], [box_indices[idx], level_y1_p[idx], level_x1_p[idx], 0],
[1, level_h_p[idx], level_w_p[idx], -1])
loss += tf.reduce_mean(tf.square(s_cropped - t_cropped))
idx = tf.add(idx, 1)
return [idx, loss]
loss_temp = tf.while_loop(cond, body, [0, 0.])
loss += loss_temp[1]
loss = loss / (2.0 * num_boxes_batch)
return loss