def pre_trained_net(weights, image):
layers = (
'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1',
'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2',
'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3', 'relu3_3', 'pool3',
'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3', 'relu4_3', 'pool4',
'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3', 'relu5_3',
)
net = {} # net stores the output tensor of each layer
cur_data = image
for i, name in enumerate(layers):
layer_type = name[:4]
if layer_type == 'conv':
kernels = weights[name]["weights"][0][0]
bias = weights[name]["bias"][0][0]
# obtain kernel and bias from vgg net (note the difference in layout)
# matconvnet: weights are [out_channels, in_channels, height, width]
# tensorflow: weights are [height, width, in_channels, out_channels]
kernels = net_utils.get_variable(np.transpose(kernels, (2, 3, 1, 0)), name=name + "_w")
bias = net_utils.get_variable(bias.reshape(-1), name=name + "_b")
cur_data = net_utils.conv2d_basic(cur_data, kernels, bias)
elif layer_type == 'relu':
cur_data = tf.nn.relu(cur_data, name=name)
elif layer_type == 'pool':
cur_data = net_utils.max_pool_2x2(cur_data)
net[name] = cur_data
return net
def inference(image, weights):
# start a new variable scope "inference"
with tf.variable_scope("inference"):
##################################################
# obtain the forward result of each layer in vgg
##################################################
# as well as the tensor after last conv layer in each stage
# NOTE: we do not make use of the result of conv layers in stage 1
image_net = pre_trained_net(weights, image)
output_stage_2 = image_net['conv2_2']
output_stage_3 = image_net['conv3_3']
output_stage_4 = image_net['conv4_3']
output_stage_5 = image_net['conv5_3']
##############################
# Preperation for upsampling
##############################
# Prep 2
output_stage_2_shape = output_stage_2.get_shape()
nChannels_in = output_stage_2_shape[3].value
kernels = weights['conv2_2_16']['weights'][0][0]
bias = weights["conv2_2_16"]["bias"][0][0]
prep2_w = net_utils.get_variable(np.transpose(kernels, (2, 3, 1, 0)), name="prep2_w")
prep2_b = net_utils.get_variable(bias.reshape(-1), name="prep2_b")
prep_2 = net_utils.conv2d_basic(output_stage_2, prep2_w, prep2_b)
# Prep 3
output_stage_3_shape = output_stage_3.get_shape()
nChannels_in = output_stage_3_shape[3].value
kernels = weights['conv3_3_16']['weights'][0][0]
bias = weights['conv3_3_16']['bias'][0][0]
prep3_w = net_utils.get_variable(np.transpose(kernels, (2, 3, 1, 0)), name="prep3_w")
prep3_b = net_utils.get_variable(bias.reshape(-1), name="prep3_b")
prep_3 = net_utils.conv2d_basic(output_stage_3, prep3_w, prep3_b)
# Prep 4
output_stage_4_shape = output_stage_4.get_shape()
nChannels_in = output_stage_4_shape[3].value
kernels = weights['conv4_3_16']['weights'][0][0]
bias = weights['conv4_3_16']['bias'][0][0]
prep4_w = net_utils.get_variable(np.transpose(kernels, (2, 3, 1, 0)), name="prep4_w")
prep4_b = net_utils.get_variable(bias.reshape(-1), name="prep4_b")
prep_4 = net_utils.conv2d_basic(output_stage_4, prep4_w, prep4_b)
# Prep 5
output_stage_5_shape = output_stage_5.get_shape()
nChannels_in = output_stage_5_shape[3].value
kernels = weights['conv5_3_16']['weights'][0][0]
bias = weights['conv5_3_16']['bias'][0][0]
prep5_w = net_utils.get_variable(np.transpose(kernels, (2, 3, 1, 0)), name="prep5_w")
prep5_b = net_utils.get_variable(bias.reshape(-1), name="prep5_b")
prep_5 = net_utils.conv2d_basic(output_stage_5, prep5_w, prep5_b)
#############################
# Upsampling for each stage
#############################
# matconvnet: weights are [out_channels, in_channels, height, width]
# tensorflow: weights are [height, width, out_channels, in_channels]
image_shape = tf.shape(image)
upsample_shape = tf.stack([FLAGS.batch_size, 480, 854, 16])
# upsample output_stage_2: upsample by ratio of 2
# t2_w = net_utils.weight_variable([4, 4, 16, 16], name="t2_w")
# Directly assign a bilinear kernel filter to the weight
t2_w = weights['upsample2_']['weights'][0][0]
t2_w = np.transpose(t2_w, (2, 3, 0, 1))
upsample_2 = net_utils.conv2d_transpose_strided(prep_2, t2_w, output_shape=upsample_shape, stride=2)
# upsample output_stage_3: upsample by ratio of 4
# t3_w = net_utils.weight_variable([8, 8, 16, 16], name="t3_w")
# Directly assign a bilinear kernel filter to the weight
t3_w = weights['upsample4_']['weights'][0][0]
t3_w = np.transpose(t3_w, (2, 3, 0, 1))
upsample_3 = net_utils.conv2d_transpose_strided(prep_3, t3_w, output_shape=upsample_shape, stride=4)
# upsample output_stage_4: upsample by ratio of 8
# t4_w = net_utils.weight_variable([16, 16, 16, 16], name="t4_w")
# Directly assign a bilinear kernel filter to the weight
t4_w = weights['upsample8_']['weights'][0][0]
t4_w = np.transpose(t4_w, (2, 3, 0, 1))
upsample_4 = net_utils.conv2d_transpose_strided(prep_4, t4_w, output_shape=upsample_shape, stride=8)
# upsample output_stage_5: upsample by ratio of 16
# t5_w = net_utils.weight_variable([32, 32, 16, 16], name="t5_w")
# Directly assign a bilinear kernel filter to the weight
t5_w = weights['upsample16_']['weights'][0][0]
t5_w = np.transpose(t5_w, (2, 3, 0, 1))
upsample_5 = net_utils.conv2d_transpose_strided(prep_5, t5_w, output_shape=upsample_shape, stride=16)
########################################
# Concatenation and Weighted Summation
########################################
fuse = tf.concat([upsample_2, upsample_3, upsample_4, upsample_5], 3)
fuse_shape = fuse.get_shape()
kernels = weights['new_score_weighting']['weights'][0][0]
bias = weights['new_score_weighting']['bias'][0][0]
fuse_w = net_utils.get_variable(np.transpose(kernels, (2, 3, 1, 0)), name="fuse_w")
fuse_b = net_utils.get_variable(bias.reshape(-1), name="fuse_b")
output_fuse = net_utils.conv2d_basic(fuse, fuse_w, fuse_b)
return output_fuse
