def construct_conv_bn_block(current, net, weights, start_weight_index,
param_names, layer_names, stride, keep_prob,
is_training):
# conv
kernel = weights[start_weight_index][1]
kernel = utils.get_variable(kernel, name=param_names[0])
if stride == 1:
current = tf.nn.conv2d(current,
kernel,
strides=[1, 1, 1, 1],
padding='SAME',
name=layer_names[0])
else:
kernel_size = kernel.shape[0]
padding = int((int(kernel_size) - 1) / 2)
current = tf.nn.conv2d(tf.pad(
current,
tf.constant([[0, 0], [padding, padding], [padding, padding],
[0, 0]])),
kernel,
strides=[1, stride, stride, 1],
padding='VALID',
name=layer_names[0])
net[layer_names[0]] = tf.nn.dropout(current, keep_prob=keep_prob)
# bn
# print(layer_names)
# print(param_names)
# print()
current, net = construct_batch_normalisation_block(current, net, weights,
start_weight_index,
param_names,
layer_names,
is_training)
return current, net
def construct_batch_normalisation_block(current, net, weights,
start_weight_index, param_names,
layer_names, is_training):
scale = weights[start_weight_index + 1][1].reshape(-1)
scale = utils.get_variable(scale, name=param_names[1])
offset = weights[start_weight_index + 2][1].reshape(-1)
offset = utils.get_variable(offset, name=param_names[2])
mean = weights[start_weight_index + 3][1][:, 0].reshape(-1)
mean = utils.get_variable(mean, name=param_names[3] + '_mean')
variance = weights[start_weight_index + 3][1][:, 1].reshape(-1)
variance = utils.get_variable(variance * variance,
name=param_names[3] + '_variance')
batch_mean, batch_var = tf.nn.moments(current, [0, 1, 2],
name='batch_moments')
decay = 1 - weights[start_weight_index + 3][2][0][0]
ema = tf.train.ExponentialMovingAverage(decay=decay)
# decay = 0.9999
def mean_var_with_update():
ema_apply_op = ema.apply([batch_mean, batch_var])
with tf.control_dependencies([ema_apply_op]):
return tf.identity(batch_mean), tf.identity(batch_var)
mean, variance = tf.cond(
is_training, mean_var_with_update, lambda:
(ema.average(batch_mean), ema.average(batch_var)))
current = tf.nn.batch_normalization(current,
mean,
variance,
offset,
scale,
1e-5,
name=layer_names[1])
net[layer_names[1]] = current
return current, net
def resnet101_net(image, weights):
net = {}
current = image
start_param_index = 0
# conv1 block
conv1_layer_names = ['conv1', 'bn_conv1', 'conv1_relu', 'pool1']
conv1_param_names = create_param_names_from_layers(conv1_layer_names)
current, net = construct_conv_bn_relu_block(current, net, weights, start_param_index, conv1_param_names, conv1_layer_names, 2)
current = tf.nn.max_pool(current, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME', name=conv1_layer_names[3])
net[conv1_layer_names[3]] = current
start_param_index += 4
current, net, start_param_index = construct_res_xa_block(2, current, net, weights, start_param_index, False)
current, net, start_param_index = construct_res_xxx_block(2, 'b', current, net, weights, start_param_index)
current, net, start_param_index = construct_res_xxx_block(2, 'c', current, net, weights, start_param_index)
current, net, start_param_index = construct_res_xa_block(3, current, net, weights, start_param_index, True)
for i in range(1, 4):
current, net, start_param_index = construct_res_xxx_block(3, 'b' + str(i), current, net, weights, start_param_index)
current, net, start_param_index = construct_res_xa_block(4, current, net, weights, start_param_index, True)
for i in range(1, 23):
current, net, start_param_index = construct_res_xxx_block(4, 'b' + str(i), current, net, weights, start_param_index)
current, net, start_param_index = construct_res_xa_block(5, current, net, weights, start_param_index, True)
current, net, start_param_index = construct_res_xxx_block(5, 'b', current, net, weights, start_param_index)
current, net, start_param_index = construct_res_xxx_block(5, 'c', current, net, weights, start_param_index)
current = tf.nn.avg_pool(current, ksize=[1, 7, 7, 1], strides=[1, 1, 1, 1], padding='VALID', name='pool5')
net['pool5'] = current
fc1000_kernel = utils.get_variable(weights[start_param_index][1], name='fc1000_filter')
fc1000_bias = utils.get_variable(weights[start_param_index + 1][1].reshape(-1), name='fc1000_bias')
current = tf.nn.bias_add(tf.nn.conv2d(current, fc1000_kernel, strides=[1, 1, 1, 1], padding="VALID"), fc1000_bias, name='fc1000')
net['fc1000'] = current
current = tf.nn.softmax(current, name='prob')
net['prob'] = current
return net
def vgg_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', 'conv3_4', 'relu3_4', 'pool3',
'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3',
'relu4_3', 'conv4_4', 'relu4_4', 'pool4',
'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3',
'relu5_3', 'conv5_4', 'relu5_4', 'pool5',
'fc6', 'relu6', 'fc7', 'relu7', 'fc8', 'prob'
)
net = {}
current = image
for i, name in enumerate(layers):
if len(name) >= 4:
kind = name[:4]
else:
kind = name[:2]
if kind == 'conv' or kind == 'fc':
kernels, bias = weights[i][0][0][0][0]
print('Kernel size:', kernels.shape)
print('Bias size:', bias.shape)
# matconvnet: weights are [width, height, in_channels, out_channels]
# tensorflow: weights are [height, width, in_channels, out_channels]
# kernels = utils.get_variable(np.transpose(kernels, (0, 1, 2, 3)), name=name + "_w")
kernels = utils.get_variable(kernels, name=name + "_W")
bias = utils.get_variable(bias.reshape(-1), name=name + "_b")
if kind == 'conv':
current = utils.conv2d_basic(current, kernels, bias)
elif kind == 'fc':
current = tf.nn.bias_add(tf.nn.conv2d(current, kernels, strides=[1, 1, 1, 1], padding="VALID"), bias)
elif kind == 'relu':
current = tf.nn.relu(current, name=name)
elif kind == 'pool':
current = utils.max_pool_2x2(current)
elif kind == 'prob':
current = tf.nn.softmax(current, name=name)
net[name] = current
return net