def __init__(self, dim = 64):
super(SRGAN_d,self).__init__()
self.conv1 = Conv2d(out_channels=dim, kernel_size=(4,4), stride=(2,2), act=tlx.LeakyReLU, padding='SAME', W_init=W_init)
self.conv2 = Conv2d(out_channels=dim * 2, kernel_size=(4,4), stride=(2,2), act=None, padding='SAME', W_init=W_init, b_init=None)
self.bn1 = BatchNorm2d(num_features=dim * 2, act=tlx.LeakyReLU, gamma_init=G_init)
self.conv3 = Conv2d(out_channels=dim * 4, kernel_size=(4,4), stride=(2,2), act=None, padding='SAME', W_init=W_init, b_init=None)
self.bn2 = BatchNorm2d(num_features=dim * 4,act=tlx.LeakyReLU, gamma_init=G_init)
self.conv4 = Conv2d(out_channels=dim * 8, kernel_size=(4, 4), stride=(2, 2), act=None, padding='SAME',W_init=W_init, b_init=None)
self.bn3 = BatchNorm2d(num_features=dim * 8, act=tlx.LeakyReLU, gamma_init=G_init)
self.conv5 = Conv2d(out_channels=dim * 16, kernel_size=(4, 4), stride=(2, 2), act=None, padding='SAME',
W_init=W_init, b_init=None)
self.bn4 = BatchNorm2d(num_features=dim * 16, act=tlx.LeakyReLU, gamma_init=G_init)
self.conv6 = Conv2d(out_channels=dim * 32, kernel_size=(4, 4), stride=(2, 2), act=None, padding='SAME',
W_init=W_init, b_init=None)
self.bn5 = BatchNorm2d(num_features=dim * 32,act=tlx.LeakyReLU, gamma_init=G_init)
self.conv7 = Conv2d(out_channels=dim * 16, kernel_size=(1, 1), stride=(1, 1), act=None, padding='SAME',
W_init=W_init, b_init=None)
self.bn6 = BatchNorm2d(num_features=dim * 16,act=tlx.LeakyReLU, gamma_init=G_init)
self.conv8 = Conv2d(out_channels=dim * 8, kernel_size=(1, 1), stride=(1, 1), act=None, padding='SAME',
W_init=W_init, b_init=None)
self.bn7 = BatchNorm2d(num_features=dim * 8,act=None, gamma_init=G_init)
self.conv9 = Conv2d(out_channels=dim * 2, kernel_size=(1, 1), stride=(1, 1), act=None, padding='SAME',
W_init=W_init, b_init=None)
self.bn8 = BatchNorm2d(num_features=dim * 2,act=tlx.LeakyReLU, gamma_init=G_init)
self.conv10 = Conv2d(out_channels=dim * 2, kernel_size=(3, 3), stride=(1, 1), act=None, padding='SAME',
W_init=W_init, b_init=None)
self.bn9 = BatchNorm2d(num_features=dim * 2,act=tlx.LeakyReLU, gamma_init=G_init)
self.conv11 = Conv2d(out_channels=dim * 8, kernel_size=(3, 3), stride=(1, 1), act=None, padding='SAME',
W_init=W_init, b_init=None)
self.bn10 = BatchNorm2d(num_features=dim * 8, gamma_init=G_init)
self.add = Elementwise(combine_fn=tlx.add, act=tlx.LeakyReLU)
self.flat = Flatten()
self.dense = Linear(out_features=1, W_init=W_init)
def __init__(self):
super(Vgg19_simple_api,self).__init__()
""" conv1 """
self.conv1 = Conv2d(out_channels=64, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv2 = Conv2d(out_channels=64, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.maxpool1 = MaxPool2d(kernel_size=(2,2), stride=(2,2), padding='SAME')
""" conv2 """
self.conv3 = Conv2d(out_channels=128, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv4 = Conv2d(out_channels=128, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.maxpool2 = MaxPool2d(kernel_size=(2,2), stride=(2,2), padding='SAME')
""" conv3 """
self.conv5 = Conv2d(out_channels=256, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv6 = Conv2d(out_channels=256, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv7 = Conv2d(out_channels=256, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv8 = Conv2d(out_channels=256, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.maxpool3 = MaxPool2d(kernel_size=(2,2), stride=(2,2), padding='SAME')
""" conv4 """
self.conv9 = Conv2d(out_channels=512, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv10 = Conv2d(out_channels=512, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv11 = Conv2d(out_channels=512, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv12 = Conv2d(out_channels=512, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.maxpool4 = MaxPool2d(kernel_size=(2,2), stride=(2,2), padding='SAME') # (batch_size, 14, 14, 512)
""" conv5 """
self.conv13 = Conv2d(out_channels=512, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv14 = Conv2d(out_channels=512, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv15 = Conv2d(out_channels=512, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.conv16 = Conv2d(out_channels=512, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME')
self.maxpool5 = MaxPool2d(kernel_size=(2,2), stride=(2,2), padding='SAME') # (batch_size, 7, 7, 512)
""" fc 6~8 """
self.flat = Flatten()
self.dense1 = Linear(out_features=4096, act=tlx.ReLU)
self.dense2 = Linear(out_features=4096, act=tlx.ReLU)
self.dense3 = Linear(out_features=1000, act=tlx.identity)
def make_layers(config, batch_norm=False, end_with='outputs'):
layer_list = []
is_end = False
for layer_group_idx, layer_group in enumerate(config):
if isinstance(layer_group, list):
for idx, layer in enumerate(layer_group):
layer_name = layer_names[layer_group_idx][idx]
n_filter = layer
if idx == 0:
if layer_group_idx > 0:
in_channels = config[layer_group_idx - 2][-1]
else:
in_channels = 3
else:
in_channels = layer_group[idx - 1]
layer_list.append(
Conv2d(
out_channels=n_filter, kernel_size=(3, 3), stride=(1, 1), act=tlx.ReLU, padding='SAME',
in_channels=in_channels, name=layer_name, data_format='channels_first'
)
)
if batch_norm:
layer_list.append(BatchNorm(num_features=n_filter, data_format='channels_first'))
if layer_name == end_with:
is_end = True
break
else:
layer_name = layer_names[layer_group_idx]
if layer_group == 'M':
layer_list.append(MaxPool2d(kernel_size=(2, 2), stride=(2, 2), padding='SAME', name=layer_name, data_format='channels_first'))
elif layer_group == 'O':
layer_list.append(Linear(out_features=1000, in_features=4096, name=layer_name))
elif layer_group == 'F':
layer_list.append(Flatten(name='flatten'))
elif layer_group == 'fc1':
layer_list.append(Linear(out_features=4096, act=tlx.ReLU, in_features=512 * 7 * 7, name=layer_name))
elif layer_group == 'fc2':
layer_list.append(Linear(out_features=4096, act=tlx.ReLU, in_features=4096, name=layer_name))
if layer_name == end_with:
is_end = True
if is_end:
break
return Sequential(layer_list)
def __init__(self, ):
super(SRGAN_d2, self).__init__()
self.conv1 = Conv2d(out_channels=64, kernel_size=(3,3), stride=(1,1), act=tlx.LeakyReLU(alpha=0.2), padding='SAME', W_init=W_init)
self.conv2 = Conv2d(out_channels=64, kernel_size=(3,3), stride=(2,2), act=tlx.LeakyReLU(alpha=0.2), padding='SAME', W_init=W_init, b_init=None)
self.bn1 = BatchNorm2d( gamma_init=G_init)
self.conv3 = Conv2d(out_channels=128, kernel_size=(3,3), stride=(1,1), act=tlx.LeakyReLU(alpha=0.2), padding='SAME', W_init=W_init, b_init=None)
self.bn2 = BatchNorm2d( gamma_init=G_init)
self.conv4 = Conv2d(out_channels=128, kernel_size=(3,3), stride=(2,2), act=tlx.LeakyReLU(alpha=0.2), padding='SAME', W_init=W_init, b_init=None)
self.bn3 = BatchNorm2d(gamma_init=G_init)
self.conv5 = Conv2d(out_channels=256, kernel_size=(3,3), stride=(1,1), act=tlx.LeakyReLU(alpha=0.2), padding='SAME', W_init=W_init, b_init=None)
self.bn4 = BatchNorm2d( gamma_init=G_init)
self.conv6 = Conv2d(out_channels=256, kernel_size=(3,3), stride=(2,2), act=tlx.LeakyReLU(alpha=0.2), padding='SAME', W_init=W_init, b_init=None)
self.bn5 = BatchNorm2d( gamma_init=G_init)
self.conv7 = Conv2d(out_channels=512, kernel_size=(3,3), stride=(1,1), act=tlx.LeakyReLU(alpha=0.2), padding='SAME', W_init=W_init, b_init=None)
self.bn6 = BatchNorm2d( gamma_init=G_init)
self.conv8 = Conv2d(out_channels=512, kernel_size=(3,3), stride=(2,2), act=tlx.LeakyReLU(alpha=0.2), padding='SAME', W_init=W_init, b_init=None)
self.bn7 = BatchNorm2d( gamma_init=G_init)
self.flat = Flatten()
self.dense1 = Linear(out_features=1024, act=tlx.LeakyReLU(alpha=0.2))
self.dense2 = Linear(out_features=1)