def forward_prop():
for l, l_info in enumerate(ConvModel.layer_info):
func_name, func_args = l_info[0], l_info[1:]
act_in = ConvModel.act_tensor[l]
act_out, dz_out, maxpool_mask = None, None, None
# perform layer-wise convolution operation
if func_name == 'convolve3d':
z_out_3d = Layers.convolve3d(act_in, ConvModel.w_tensor[l], ConvModel.b_tensor[l], func_args[0],
func_args[1])
dz_out = Layers.drelu(np.array(z_out_3d))
act_out = Layers.relu(np.array(z_out_3d))
# perform mlp layer-wise forward pass operation
if func_name == 'fc_layer':
act_in = np.ravel(act_in)
z_out = np.dot(act_in, ConvModel.w_tensor[l]) + ConvModel.b_tensor[l]
act_out = getattr(Layers, func_args[0])(np.array(z_out))
dz_out = getattr(Layers, 'd' + func_args[0])(np.array(z_out))
# perform layer-wise pooling operation
if func_name == 'maxpool3d':
pool_outputs = Layers.maxpool3d(input=act_in, filter_dim=func_args[0], stride=func_args[1])
act_out = pool_outputs[0]
maxpool_mask = pool_outputs[1]
dz_out = Layers.drelu(np.array(act_out))
ConvModel.act_tensor[l + 1] = np.array(act_out)
ConvModel.dz_tensor[l] = np.array(dz_out)
ConvModel.maxpool_masks[l] = np.array(maxpool_mask)
def subsampled(inputs, reuse=False):
# Less border effect
inputs = Layers.pad(inputs)
with tf.variable_scope('subsampled', reuse=reuse):
conv1 = Layers.conv2d(inputs, 3, 32, 9, 1, 'SAME', 'conv1')
norm1 = Layers.instance_norm(conv1)
relu1 = Layers.relu(norm1)
conv2 = Layers.conv2d(relu1, 32, 64, 3, 2, 'SAME', 'conv2')
norm2 = Layers.instance_norm(conv2)
relu2 = Layers.relu(norm2)
conv3 = Layers.conv2d(relu2, 64, 128, 3, 2, 'SAME', 'conv3')
norm3 = Layers.instance_norm(conv3)
relu3 = Layers.relu(norm3)
return relu3
def upsampling(inputs, reuse=False):
with tf.variable_scope('upsampling', reuse=reuse):
deconv1 = Layers.resize_conv2d(inputs, 128, 64, 3, 2, 'SAME', 'deconv1')
denorm1 = Layers.instance_norm(deconv1)
derelu1 = Layers.relu(denorm1)
deconv2 = Layers.resize_conv2d(derelu1, 64, 32, 3, 2, 'SAME', 'deconv2')
denorm2 = Layers.instance_norm(deconv2)
derelu2 = Layers.relu(denorm2)
deconv3 = Layers.resize_conv2d(derelu2, 32, 3, 9, 1, 'SAME', 'deconv3')
denorm3 = Layers.instance_norm(deconv3)
detanh3 = tf.nn.tanh(denorm3)
y = (detanh3 + 1) * 127.5
# Remove the border effect
y = Layers.remove_pad(y)
return y
