def block_forward(self,
inp,
weights,
ss_weights,
block,
reuse,
scope,
block_last_layer=False):
"""The function to forward a resnet block during meta-train phase
Args:
inp: input feature maps.
weights: input resnet weights.
ss_weights: input scaling weights.
block: the string to indicate which block we are processing.
reuse: reuse the batch norm weights or not.
scope: the label to indicate which layer we are processing.
block_last_layer: whether it is the last layer of this block.
Return:
The processed feature maps.
"""
net = resnet_conv_block(inp, self.process_ss_weights(weights, ss_weights, block + '_conv1'), \
ss_weights[block + '_bias1'], reuse, scope+block+'0')
net = resnet_conv_block(net, self.process_ss_weights(weights, ss_weights, block + '_conv2'), \
ss_weights[block + '_bias2'], reuse, scope+block+'1')
res = resnet_nob_conv_block(inp, weights[block + '_conv_res'], reuse,
scope + block + 'res')
net = net + res
if block_last_layer:
net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], 'SAME')
net = tf.nn.dropout(net, keep_prob=1)
return net
def pretrain_block_forward(self, inp, weights, block, reuse, scope):
"""The function to forward a resnet block during pre-train phase
Args:
inp: input feature maps.
weights: input resnet weights.
block: the string to indicate which block we are processing.
reuse: reuse the batch norm weights or not.
scope: the label to indicate which layer we are processing.
Return:
The processed feature maps.
"""
net = resnet_conv_block(inp, weights[block + '_conv1'],
weights[block + '_bias1'], reuse,
scope + block + '0')
net = resnet_conv_block(net, weights[block + '_conv2'],
weights[block + '_bias2'], reuse,
scope + block + '1')
net = resnet_conv_block(net, weights[block + '_conv3'],
weights[block + '_bias3'], reuse,
scope + block + '2')
res = resnet_nob_conv_block(inp, weights[block + '_conv_res'], reuse,
scope + block + 'res')
net = net + res
net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], 'VALID')
net = tf.nn.dropout(net, keep_prob=FLAGS.pretrain_dropout_keep)
return net
def block_forward(self, inp, weights, ss_weights, block, reuse, scope):
net = resnet_conv_block(inp, self.process_ss_weights(weights, ss_weights, block + '_conv1'), ss_weights[block + '_bias1'], reuse, scope+block+'0')
net = resnet_conv_block(net, self.process_ss_weights(weights, ss_weights, block + '_conv2'), ss_weights[block + '_bias2'], reuse, scope+block+'1')
net = resnet_conv_block(net, self.process_ss_weights(weights, ss_weights, block + '_conv3'), ss_weights[block + '_bias3'], reuse, scope+block+'2')
res = resnet_nob_conv_block(inp, weights[block + '_conv_res'], reuse, scope+block+'res')
net = net + res
net = tf.nn.max_pool(net, [1,2,2,1], [1,2,2,1], 'VALID')
net = tf.nn.dropout(net, keep_prob=1)
return net
def pretrain_block_forward(self, inp, weights, block, reuse, scope):
net = resnet_conv_block(inp, weights[block + '_conv1'], weights[block + '_bias1'], reuse, scope+block+'0')
net = resnet_conv_block(net, weights[block + '_conv2'], weights[block + '_bias2'], reuse, scope+block+'1')
net = resnet_conv_block(net, weights[block + '_conv3'], weights[block + '_bias3'], reuse, scope+block+'2')
res = resnet_nob_conv_block(inp, weights[block + '_conv_res'], reuse, scope+block+'res')
net = net + res
net = tf.nn.max_pool(net, [1,2,2,1], [1,2,2,1], 'VALID')
net = tf.nn.dropout(net, keep_prob=FLAGS.pretrain_dropout_keep)
return net
def forward_swn(self, inp, weights, reuse):
net = resnet_conv_block(inp, weights['RN_conv1'], weights['RN_bias1'], reuse, 'RN_0')
net = resnet_conv_block(net, weights['RN_conv2'], weights['RN_bias2'], reuse, 'RN_1')
net = tf.reshape(net, [-1, np.prod([int(dim) for dim in net.get_shape()[1:]])])
net = tf.matmul(net, weights['RN_fc_w1']) + weights['RN_fc_b1']
net = leaky_relu(net)
net = tf.matmul(net, weights['RN_fc_w2']) + weights['RN_fc_b2']
return net