def build_graph(self, image, label):
xys = np.array([(y, x, 1) for y in range(WARP_TARGET_SIZE)
for x in range(WARP_TARGET_SIZE)], dtype='float32')
xys = tf.constant(xys, dtype=tf.float32, name='xys') # p x 3
image = image / 255.0 - 0.5 # bhw2
def get_stn(image):
stn = (LinearWrap(image)
.AvgPooling('downsample', 2)
.Conv2D('conv0', 20, 5, padding='VALID')
.MaxPooling('pool0', 2)
.Conv2D('conv1', 20, 5, padding='VALID')
.FullyConnected('fc1', 32)
.FullyConnected('fct', 6, activation=tf.identity,
kernel_initializer=tf.constant_initializer(),
bias_initializer=tf.constant_initializer([1, 0, HALF_DIFF, 0, 1, HALF_DIFF]))())
# output 6 parameters for affine transformation
stn = tf.reshape(stn, [-1, 2, 3], name='affine') # bx2x3
stn = tf.reshape(tf.transpose(stn, [2, 0, 1]), [3, -1]) # 3 x (bx2)
coor = tf.reshape(tf.matmul(xys, stn),
[WARP_TARGET_SIZE, WARP_TARGET_SIZE, -1, 2])
coor = tf.transpose(coor, [2, 0, 1, 3], 'sampled_coords') # b h w 2
sampled = GridSample('warp', [image, coor], borderMode='constant')
return sampled
with argscope([Conv2D, FullyConnected], activation=tf.nn.relu):
with tf.variable_scope('STN1'):
sampled1 = get_stn(image)
with tf.variable_scope('STN2'):
sampled2 = get_stn(image)
# For visualization in tensorboard
with tf.name_scope('visualization'):
padded1 = tf.pad(sampled1, [[0, 0], [HALF_DIFF, HALF_DIFF], [HALF_DIFF, HALF_DIFF], [0, 0]])
padded2 = tf.pad(sampled2, [[0, 0], [HALF_DIFF, HALF_DIFF], [HALF_DIFF, HALF_DIFF], [0, 0]])
img_orig = tf.concat([image[:, :, :, 0], image[:, :, :, 1]], 1) # b x 2h x w
transform1 = tf.concat([padded1[:, :, :, 0], padded1[:, :, :, 1]], 1)
transform2 = tf.concat([padded2[:, :, :, 0], padded2[:, :, :, 1]], 1)
stacked = tf.concat([img_orig, transform1, transform2], 2, 'viz')
tf.summary.image('visualize',
tf.expand_dims(stacked, -1), max_outputs=30)
sampled = tf.concat([sampled1, sampled2], 3, 'sampled_concat')
logits = (LinearWrap(sampled)
.FullyConnected('fc1', 256, activation=tf.nn.relu)
.FullyConnected('fc2', 128, activation=tf.nn.relu)
.FullyConnected('fct', 19, activation=tf.identity)())
tf.nn.softmax(logits, name='prob')
cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=label)
cost = tf.reduce_mean(cost, name='cross_entropy_loss')
wrong = tf.cast(tf.logical_not(tf.nn.in_top_k(logits, label, 1)), tf.float32, name='incorrect_vector')
summary.add_moving_summary(tf.reduce_mean(wrong, name='train_error'))
wd_cost = tf.multiply(1e-5, regularize_cost('fc.*/W', tf.nn.l2_loss),
name='regularize_loss')
summary.add_moving_summary(cost, wd_cost)
return tf.add_n([wd_cost, cost], name='cost')
def visualize_conv_activations(activation, name):
"""Visualize activations for convolution layers.
Remarks:
This tries to place all activations into a square.
Args:
activation: tensor with the activation [B,H,W,C]
name: label for tensorboard
Returns:
image of almost all activations
"""
import math
with tf.name_scope('visualize_act_' + name):
_, h, w, c = activation.get_shape().as_list()
rows = []
c_per_row = int(math.sqrt(c))
for y in range(0, c - c_per_row, c_per_row):
row = activation[:, :, :,
y:y + c_per_row] # [?, H, W, 32] --> [?, H, W, 5]
cols = tf.unstack(row, axis=3) # [?, H, W, 5] --> 5 * [?, H, W]
row = tf.concat(cols, 1)
rows.append(row)
viz = tf.concat(rows, 2)
tf.summary.image('visualize_act_' + name, tf.expand_dims(viz, -1))
def GridSample(inputs, borderMode='repeat'):
"""
Sample the images using the given coordinates, by bilinear interpolation.
This was described in the paper:
`Spatial Transformer Networks <http://arxiv.org/abs/1506.02025>`_.
This is equivalent to `torch.nn.functional.grid_sample`,
up to some non-trivial coordinate transformation.
This implementation returns pixel value at pixel (1, 1) for a floating point coordinate (1.0, 1.0).
Note that this may not be what you need.
Args:
inputs (list): [images, coords]. images has shape NHWC.
coords has shape (N, H', W', 2), where each pair of the last dimension is a (y, x) real-value
coordinate.
borderMode: either "repeat" or "constant" (zero-filled)
Returns:
tf.Tensor: a tensor named ``output`` of shape (N, H', W', C).
"""
image, mapping = inputs
assert image.get_shape().ndims == 4 and mapping.get_shape().ndims == 4
input_shape = image.get_shape().as_list()[1:]
assert None not in input_shape, \
"Images in GridSample layer must have fully-defined shape"
assert borderMode in ['repeat', 'constant']
orig_mapping = mapping
mapping = tf.maximum(mapping, 0.0)
lcoor = tf.floor(mapping)
ucoor = lcoor + 1
diff = mapping - lcoor
neg_diff = 1.0 - diff # bxh2xw2x2
lcoory, lcoorx = tf.split(lcoor, 2, 3)
ucoory, ucoorx = tf.split(ucoor, 2, 3)
lyux = tf.concat([lcoory, ucoorx], 3)
uylx = tf.concat([ucoory, lcoorx], 3)
diffy, diffx = tf.split(diff, 2, 3)
neg_diffy, neg_diffx = tf.split(neg_diff, 2, 3)
ret = tf.add_n([sample(image, lcoor) * neg_diffx * neg_diffy,
sample(image, ucoor) * diffx * diffy,
sample(image, lyux) * neg_diffy * diffx,
sample(image, uylx) * diffy * neg_diffx], name='sampled')
if borderMode == 'constant':
max_coor = tf.constant([input_shape[0] - 1, input_shape[1] - 1], dtype=tf.float32)
mask = tf.greater_equal(orig_mapping, 0.0)
mask2 = tf.less_equal(orig_mapping, max_coor)
mask = tf.logical_and(mask, mask2) # bxh2xw2x2
mask = tf.reduce_all(mask, [3]) # bxh2xw2 boolean
mask = tf.expand_dims(mask, 3)
ret = ret * tf.cast(mask, tf.float32)
return tf.identity(ret, name='output')
def build_graph(self, image, label):
"""This function should build the model which takes the input variables (defined above)
and return cost at the end."""
# In tensorflow, inputs to convolution function are assumed to be
# NHWC. Add a single channel here.
image = tf.expand_dims(image, 3)
image = image * 2 - 1 # center the pixels values at zero
# The context manager `argscope` sets the default option for all the layers under
# this context. Here we use 32 channel convolution with shape 3x3
# See tutorial at https://tensorpack.readthedocs.io/tutorial/symbolic.html
with argscope(Conv2D, kernel_size=3, activation=tf.nn.relu, filters=32):
# LinearWrap is just a syntax sugar.
# See tutorial at https://tensorpack.readthedocs.io/tutorial/symbolic.html
logits = (LinearWrap(image)
.Conv2D('conv0')
.MaxPooling('pool0', 2)
.Conv2D('conv1')
.Conv2D('conv2')
.MaxPooling('pool1', 2)
.Conv2D('conv3')
.FullyConnected('fc0', 512, activation=tf.nn.relu)
.Dropout('dropout', rate=0.5)
.FullyConnected('fc1', 10, activation=tf.identity)())
# a vector of length B with loss of each sample
cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=label)
cost = tf.reduce_mean(cost, name='cross_entropy_loss') # the average cross-entropy loss
correct = tf.cast(tf.nn.in_top_k(predictions=logits, targets=label, k=1), tf.float32, name='correct')
accuracy = tf.reduce_mean(correct, name='accuracy')
# This will monitor training error & accuracy (in a moving average fashion). The value will be automatically
# 1. written to tensosrboard
# 2. written to stat.json
# 3. printed after each epoch
# You can also just call `tf.summary.scalar`. But moving summary has some other benefits.
# See tutorial at https://tensorpack.readthedocs.io/tutorial/summary.html
train_error = tf.reduce_mean(1 - correct, name='train_error')
summary.add_moving_summary(train_error, accuracy)
# Use a regex to find parameters to apply weight decay.
# Here we apply a weight decay on all W (weight matrix) of all fc layers
# If you don't like regex, you can certainly define the cost in any other methods.
wd_cost = tf.multiply(1e-5,
regularize_cost('fc.*/W', tf.nn.l2_loss),
name='regularize_loss')
total_cost = tf.add_n([wd_cost, cost], name='total_cost')
summary.add_moving_summary(cost, wd_cost, total_cost)
# monitor histogram of all weight (of conv and fc layers) in tensorboard
summary.add_param_summary(('.*/W', ['histogram', 'rms']))
# the function should return the total cost to be optimized
return total_cost
def visualize_conv_weights(filters, name):
"""Visualize use weights in convolution filters.
Args:
filters: tensor containing the weights [H,W,Cin,Cout]
name: label for tensorboard
Returns:
image of all weight
"""
with tf.name_scope('visualize_w_' + name):
filters = tf.transpose(
filters, (3, 2, 0, 1)) # [h, w, cin, cout] -> [cout, cin, h, w]
filters = tf.unstack(filters) # --> cout * [cin, h, w]
filters = tf.concat(filters, 1) # --> [cin, cout * h, w]
filters = tf.unstack(filters) # --> cin * [cout * h, w]
filters = tf.concat(filters, 1) # --> [cout * h, cin * w]
filters = tf.expand_dims(filters, 0)
filters = tf.expand_dims(filters, -1)
tf.summary.image('visualize_w_' + name, filters)
def build_graph(self, image, label):
"""This function should build the model which takes the input variables
and return cost at the end"""
# In tensorflow, inputs to convolution function are assumed to be
# NHWC. Add a single channel here.
image = tf.expand_dims(image, 3)
image = image * 2 - 1 # center the pixels values at zero
# The context manager `argscope` sets the default option for all the layers under
# this context. Here we use 32 channel convolution with shape 3x3
with argscope([tf.layers.conv2d], padding='same', activation=tf.nn.relu):
l = tf.layers.conv2d(image, 32, 3, name='conv0')
l = tf.layers.max_pooling2d(l, 2, 2, padding='valid')
l = tf.layers.conv2d(l, 32, 3, name='conv1')
l = tf.layers.conv2d(l, 32, 3, name='conv2')
l = tf.layers.max_pooling2d(l, 2, 2, padding='valid')
l = tf.layers.conv2d(l, 32, 3, name='conv3')
l = tf.layers.flatten(l)
l = tf.layers.dense(l, 512, activation=tf.nn.relu, name='fc0')
l = tf.layers.dropout(l, rate=0.5, training=self.training)
logits = tf.layers.dense(l, 10, activation=tf.identity, name='fc1')
# a vector of length B with loss of each sample
cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=label)
cost = tf.reduce_mean(cost, name='cross_entropy_loss') # the average cross-entropy loss
correct = tf.cast(tf.nn.in_top_k(logits, label, 1), tf.float32, name='correct')
accuracy = tf.reduce_mean(correct, name='accuracy')
# This will monitor training error & accuracy (in a moving average fashion). The value will be automatically
# 1. written to tensosrboard
# 2. written to stat.json
# 3. printed after each epoch
train_error = tf.reduce_mean(1 - correct, name='train_error')
summary.add_moving_summary(train_error, accuracy)
# Use a regex to find parameters to apply weight decay.
# Here we apply a weight decay on all W (weight matrix) of all fc layers
# If you don't like regex, you can certainly define the cost in any other methods.
wd_cost = tf.multiply(1e-5,
regularize_cost('fc.*/kernel', tf.nn.l2_loss),
name='regularize_loss')
total_cost = tf.add_n([wd_cost, cost], name='total_cost')
summary.add_moving_summary(cost, wd_cost, total_cost)
# monitor histogram of all weight (of conv and fc layers) in tensorboard
summary.add_param_summary(('.*/kernel', ['histogram', 'rms']))
# the function should return the total cost to be optimized
return total_cost
def build_graph(self, image, label):
image = tf.expand_dims(image * 2 - 1, 3)
with argscope(Conv2D, kernel_shape=3, nl=tf.nn.relu, out_channel=32):
c0 = Conv2D('conv0', image)
p0 = MaxPooling('pool0', c0, 2)
c1 = Conv2D('conv1', p0)
c2 = Conv2D('conv2', c1)
p1 = MaxPooling('pool1', c2, 2)
c3 = Conv2D('conv3', p1)
fc1 = FullyConnected('fc0', c3, 512, nl=tf.nn.relu)
fc1 = Dropout('dropout', fc1, 0.5)
logits = FullyConnected('fc1', fc1, out_dim=10, nl=tf.identity)
with tf.name_scope('visualizations'):
visualize_conv_weights(c0.variables.W, 'conv0')
visualize_conv_activations(c0, 'conv0')
visualize_conv_weights(c1.variables.W, 'conv1')
visualize_conv_activations(c1, 'conv1')
visualize_conv_weights(c2.variables.W, 'conv2')
visualize_conv_activations(c2, 'conv2')
visualize_conv_weights(c3.variables.W, 'conv3')
visualize_conv_activations(c3, 'conv3')
tf.summary.image('input', (image + 1.0) * 128., 3)
cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,
labels=label)
cost = tf.reduce_mean(cost, name='cross_entropy_loss')
tf.reduce_mean(tf.cast(tf.nn.in_top_k(logits, label, 1), tf.float32),
name='accuracy')
wd_cost = tf.multiply(1e-5,
regularize_cost('fc.*/W', tf.nn.l2_loss),
name='regularize_loss')
return tf.add_n([wd_cost, cost], name='total_cost')
def roi_heads(self, image, features, proposals, targets):
image_shape2d = tf.shape(image)[2:] # h,w
assert len(features) == 5, "Features have to be P23456!"
gt_boxes, gt_labels, *_ = targets
if self.training:
proposals = sample_fast_rcnn_targets(proposals.boxes, gt_boxes,
gt_labels)
fastrcnn_head_func = getattr(model_frcnn, cfg.FPN.FRCNN_HEAD_FUNC)
if not cfg.FPN.CASCADE:
roi_feature_fastrcnn = multilevel_roi_align(
features[:4], proposals.boxes, 7)
head_feature = fastrcnn_head_func('fastrcnn', roi_feature_fastrcnn)
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_outputs(
'fastrcnn/outputs', head_feature, cfg.DATA.NUM_CATEGORY)
fastrcnn_head = FastRCNNHead(
proposals, fastrcnn_box_logits, fastrcnn_label_logits,
gt_boxes,
tf.constant(cfg.FRCNN.BBOX_REG_WEIGHTS, dtype=tf.float32))
else:
def roi_func(boxes):
return multilevel_roi_align(features[:4], boxes, 7)
fastrcnn_head = CascadeRCNNHead(proposals, roi_func,
fastrcnn_head_func,
(gt_boxes, gt_labels),
image_shape2d,
cfg.DATA.NUM_CATEGORY)
if self.training:
all_losses = fastrcnn_head.losses()
if cfg.MODE_MASK:
gt_masks = targets[2]
# maskrcnn loss
roi_feature_maskrcnn = multilevel_roi_align(
features[:4],
proposals.fg_boxes(),
14,
name_scope='multilevel_roi_align_mask')
maskrcnn_head_func = getattr(model_mrcnn,
cfg.FPN.MRCNN_HEAD_FUNC)
mask_logits = maskrcnn_head_func(
'maskrcnn', roi_feature_maskrcnn,
cfg.DATA.NUM_CATEGORY) # #fg x #cat x 28 x 28
target_masks_for_fg = crop_and_resize(
tf.expand_dims(gt_masks, 1),
proposals.fg_boxes(),
proposals.fg_inds_wrt_gt,
28,
pad_border=False) # fg x 1x28x28
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1,
'sampled_fg_mask_targets')
all_losses.append(
maskrcnn_loss(mask_logits, proposals.fg_labels(),
target_masks_for_fg))
return all_losses
else:
decoded_boxes = fastrcnn_head.decoded_output_boxes()
decoded_boxes = clip_boxes(decoded_boxes,
image_shape2d,
name='fastrcnn_all_boxes')
label_scores = fastrcnn_head.output_scores(
name='fastrcnn_all_scores')
final_boxes, final_scores, final_labels = fastrcnn_predictions(
decoded_boxes, label_scores, name_scope='output')
if cfg.MODE_MASK:
# Cascade inference needs roi transform with refined boxes.
roi_feature_maskrcnn = multilevel_roi_align(
features[:4], final_boxes, 14)
maskrcnn_head_func = getattr(model_mrcnn,
cfg.FPN.MRCNN_HEAD_FUNC)
mask_logits = maskrcnn_head_func(
'maskrcnn', roi_feature_maskrcnn,
cfg.DATA.NUM_CATEGORY) # #fg x #cat x 28 x 28
indices = tf.stack([
tf.range(tf.size(final_labels)),
tf.cast(final_labels, tf.int32) - 1
],
axis=1)
final_mask_logits = tf.gather_nd(mask_logits,
indices) # #resultx28x28
tf.sigmoid(final_mask_logits, name='output/masks')
return []
def preprocess(self, image):
image = tf.expand_dims(image, 0)
image = image_preprocess(image, bgr=True)
return tf.transpose(image, [0, 3, 1, 2])
def roi_heads(self, image, features, proposals, targets):
image_shape2d = tf.shape(image)[2:] # h,w
featuremap = features[0]
gt_boxes, gt_labels, *_ = targets
if self.training:
# sample proposal boxes in training
proposals = sample_fast_rcnn_targets(proposals.boxes, gt_boxes,
gt_labels)
# The boxes to be used to crop RoIs.
# Use all proposal boxes in inference
boxes_on_featuremap = proposals.boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE)
roi_resized = roi_align(featuremap, boxes_on_featuremap, 14)
feature_fastrcnn = resnet_conv5(
roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCKS[-1]) # nxcx7x7
# Keep C5 feature to be shared with mask branch
feature_gap = GlobalAvgPooling('gap',
feature_fastrcnn,
data_format='channels_first')
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_outputs(
'fastrcnn', feature_gap, cfg.DATA.NUM_CATEGORY)
fastrcnn_head = FastRCNNHead(
proposals, fastrcnn_box_logits, fastrcnn_label_logits, gt_boxes,
tf.constant(cfg.FRCNN.BBOX_REG_WEIGHTS, dtype=tf.float32))
if self.training:
all_losses = fastrcnn_head.losses()
if cfg.MODE_MASK:
gt_masks = targets[2]
# maskrcnn loss
# In training, mask branch shares the same C5 feature.
fg_feature = tf.gather(feature_fastrcnn, proposals.fg_inds())
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', fg_feature, cfg.DATA.NUM_CATEGORY,
num_convs=0) # #fg x #cat x 14x14
target_masks_for_fg = crop_and_resize(
tf.expand_dims(gt_masks, 1),
proposals.fg_boxes(),
proposals.fg_inds_wrt_gt,
14,
pad_border=False) # nfg x 1x14x14
target_masks_for_fg = tf.squeeze(target_masks_for_fg, 1,
'sampled_fg_mask_targets')
all_losses.append(
maskrcnn_loss(mask_logits, proposals.fg_labels(),
target_masks_for_fg))
return all_losses
else:
decoded_boxes = fastrcnn_head.decoded_output_boxes()
decoded_boxes = clip_boxes(decoded_boxes,
image_shape2d,
name='fastrcnn_all_boxes')
label_scores = fastrcnn_head.output_scores(
name='fastrcnn_all_scores')
final_boxes, final_scores, final_labels = fastrcnn_predictions(
decoded_boxes, label_scores, name_scope='output')
if cfg.MODE_MASK:
roi_resized = roi_align(
featuremap, final_boxes * (1.0 / cfg.RPN.ANCHOR_STRIDE),
14)
feature_maskrcnn = resnet_conv5(
roi_resized, cfg.BACKBONE.RESNET_NUM_BLOCKS[-1])
mask_logits = maskrcnn_upXconv_head(
'maskrcnn', feature_maskrcnn, cfg.DATA.NUM_CATEGORY,
0) # #result x #cat x 14x14
indices = tf.stack([
tf.range(tf.size(final_labels)),
tf.cast(final_labels, tf.int32) - 1
],
axis=1)
final_mask_logits = tf.gather_nd(mask_logits,
indices) # #resultx14x14
tf.sigmoid(final_mask_logits, name='output/masks')
return []