def inference(image, output_shape, keep_prob, is_train):
wd = 0.0001
leaky_param = 0.01
input_shape = image.get_shape().as_list()
conv1 = mf.convolution_2d_layer(image, [5, 5, 3, 64], [1, 1], 'SAME', wd,
'conv1')
conv1_pool = mf.maxpool_2d_layer(conv1, [2, 2], [2, 2], 'maxpool1')
res1 = mf.copy_layer(conv1_pool, mf.res_layer, 3, 'res', [3, 3, 64, 64],
[1, 1], "SAME", wd, 'res_1', 2, leaky_param, is_train)
res1_pool = mf.maxpool_2d_layer(res1, [2, 2], [2, 2], 'maxpool2')
res1_pad = mf.res_pad(res1_pool, 64, 128, 'res_pad1')
res2 = mf.copy_layer(res1_pad, mf.res_layer, 4, 'res', [3, 3, 128, 128],
[1, 1], "SAME", wd, 'res_2', 2, leaky_param, is_train)
res2_pool = mf.maxpool_2d_layer(res2, [2, 2], [2, 2], 'maxpool2')
res2_pad = mf.res_pad(res2_pool, 128, 256, 'res_pad2')
res3 = mf.copy_layer(res2_pad, mf.res_layer, 6, 'res', [3, 3, 256, 256],
[1, 1], "SAME", wd, 'res_3', 2, leaky_param, is_train)
#res3_pool = mf.maxpool_2d_layer(res3, [2,2], [2,2], 'maxpool3')
#res3_pad = mf.res_pad(res3_pool, 256, 512, 'res_pad3')
#res4 = mf.copy_layer(res3_pad, mf.res_layer, 3, 'res', [3,3, 512, 512], [1,1], "SAME", wd, 'res_4', 2, leaky_param, is_train)
res1_unpool = mf.unpooling_layer(res1, [input_shape[1], input_shape[2]],
'res1_unpool')
res2_unpool = mf.unpooling_layer(res2, [input_shape[1], input_shape[2]],
'res2_unpool')
res3_unpool = mf.unpooling_layer(res3, [input_shape[1], input_shape[2]],
'res3_unpool')
#res4_unpool = mf.unpooling_layer(res4, [input_shape[1], input_shape[2]], 'res4_unpool')
hypercolumn = tf.concat(3, [res1_unpool, res2_unpool, res3_unpool],
name='hypercolumn')
output_shape = input_shape
output_shape[3] = 1
#heatmap = mf.deconvolution_2d_layer(hypercolumn, [1,1,1,hypercolumn.get_shape()[3]], [1,1], output_shape, 'SAME', wd, 'heatmap')
heatmap = mf.convolution_2d_layer(
hypercolumn, [1, 1, hypercolumn.get_shape()[3], 1], [1, 1], 'SAME', wd,
'heatmap')
return heatmap
def dense_transition_layer(input_tensor, is_pool, data_format, is_train,
leaky_param, wd, layer_name):
with tf.variable_scope(layer_name):
if data_format == "NCHW":
in_channle = input_tensor.get_shape()[1]
elif data_format == "NHWC":
in_channle = input_tensor.get_shape()[3]
else:
raise NotImplementedError
if not is_pool:
conv = mf.convolution_2d_layer(input_tensor, in_channle, [3, 3], [2, 2], "SAME",
data_format, True, is_train, leaky_param, wd, "conv")
else:
conv = mf.convolution_2d_layer(input_tensor, in_channle, [1, 1], [1, 1], "SAME",
data_format, True, is_train, leaky_param, wd, "conv")
conv = mf.maxpool_2d_layer(conv, [2, 2], [2, 2], data_format, 'maxpool')
return conv
def _add_adapt_pair_loss_single(self, feature, scale, leaky_param,
wd, layer_name):
with tf.variable_scope(layer_name):
feat = flip_gradient(feature, scale)
keep_prob_ph = self.keep_prob_ph
data_format = "NHWC"
conv1 = mf.convolution_2d_layer(
feat, 256, [3, 3], [1, 1],
"SAME", data_format=data_format, leaky_params=leaky_param,
wd=wd, layer_name="dann_conv1")
conv2 = mf.convolution_2d_layer(
conv1, 512, [3, 3], [1, 1],
"SAME", data_format=data_format, leaky_params=leaky_param,
wd=wd, layer_name="dann_conv2")
conv3 = mf.convolution_2d_layer(
conv2, 512, [3, 3], [1, 1],
"SAME", data_format=data_format, leaky_params=leaky_param,
wd=wd, layer_name="dann_conv3")
conv3_maxpool = mf.maxpool_2d_layer(
conv3, [2, 2], [2, 2], data_format=data_format,
layer_name="dann_maxpool3")
a_fc1 = mf.fully_connected_layer(
conv3_maxpool, 2048, leaky_param, wd, "a_fc1")
a_fc2 = mf.fully_connected_layer(
a_fc1, 2048, leaky_param, wd, "a_fc2")
a_fc3 = mf.fully_connected_layer(
a_fc2, 2, leaky_param, wd, "a_fc3")
return a_fc3
def model(self):
wd = 0.0004
leaky_param = 0.01
with tf.variable_scope("G"):
rf2 = mf.add_leaky_relu(
mf.fully_connected_layer(self.ran_code_ph, 8 * 8 * 64, wd,
"fc2"), leaky_param)
rf3 = tf.reshape(rf2, [self.bsize, 8, 8, 64], name="fc3")
rdeconv1 = mf.add_leaky_relu(
mf.deconvolution_2d_layer(rf3, [2, 2, 128, 64], [2, 2],
[self.bsize, 16, 16, 128], "VALID",
wd, "deconv1"), leaky_param)
rdeconv2 = mf.add_leaky_relu(
mf.deconvolution_2d_layer(rdeconv1, [2, 2, 256, 128], [2, 2],
[self.bsize, 32, 32, 256], "VALID",
wd, "deconv2"), leaky_param)
deconv1 = mf.add_leaky_relu(
mf.deconvolution_2d_layer(rdeconv2, [2, 2, 512, 256], [2, 2],
[self.bsize, 64, 64, 512], "VALID",
wd, "deconv3"), leaky_param)
deconv2 = mf.add_leaky_relu(
mf.deconvolution_2d_layer(deconv1, [2, 2, 512, 512], [2, 2],
[self.bsize, 128, 128, 512], "VALID",
wd, "deconv4"), leaky_param)
conv1 = mf.convolution_2d_layer(deconv2, [1, 1, 512, 1], [1, 1],
"SAME", wd, "conv1")
self.g_image = tf.sigmoid(conv1, "g_image")
tf.add_to_collection("image_to_write", self.g_image)
tf.add_to_collection("image_to_write", self.image_data_ph)
with tf.variable_scope("D"):
concat = tf.concat(0, [self.g_image, self.image_data_ph])
#conv1 = mf.convolution_2d_layer(self.image_data_ph, [5, 5, 2, 64], [2,2], "VALID", wd, "conv1")
conv1 = mf.add_leaky_relu(
mf.convolution_2d_layer(concat, [3, 3, 1, 32], [2, 2], "SAME",
wd, "conv1"), leaky_param)
conv1_maxpool = mf.maxpool_2d_layer(conv1, [2, 2], [2, 2],
"maxpool1")
conv2 = mf.add_leaky_relu(
mf.convolution_2d_layer(conv1_maxpool, [3, 3, 32, 64], [2, 2],
"SAME", wd, "conv2"), leaky_param)
conv2_maxpool = mf.maxpool_2d_layer(conv2, [2, 2], [2, 2],
"maxpool2")
conv3 = mf.add_leaky_relu(
mf.convolution_2d_layer(conv2, [3, 3, 64, 128], [2, 2], "SAME",
wd, "conv3"), leaky_param)
conv3_maxpool = mf.maxpool_2d_layer(conv3, [2, 2], [2, 2],
"maxpool3")
conv4 = mf.add_leaky_relu(
mf.convolution_2d_layer(conv3_maxpool, [3, 3, 128, 128],
[2, 2], "SAME", wd, "conv4"),
leaky_param)
conv4_maxpool = mf.maxpool_2d_layer(conv4, [2, 2], [2, 2],
"maxpool4")
self.fc = mf.fully_connected_layer(conv4_maxpool, 1, wd, "fc")
def model_infer(self, data_ph, model_params):
input_ph = data_ph.get_input()
gl_ph = data_ph.get_gl()
train_test_ph = data_ph.get_train_test()
input_ph = tf.cond(train_test_ph, lambda: self.add_data_arg(
input_ph), lambda: input_ph)
data_format = "NHWC"
bn = False
keep_prob_ph = data_ph.get_keep_prob()
self.keep_prob_ph = keep_prob_ph
b, _, _, _ = input_ph.get_shape()
self.b = b
leaky_param = model_params["leaky_param"]
wd = model_params["weight_decay"]
hyper_list = list()
print(input_ph)
conv11 = mf.convolution_2d_layer(
input_ph, 64, [3, 3], [1, 1], "SAME", data_format=data_format,
leaky_params=leaky_param, wd=wd, layer_name="conv11")
print(conv11)
conv1_maxpool = mf.maxpool_2d_layer(conv11, [2, 2],
[2, 2], data_format, "maxpool1")
print(conv1_maxpool)
conv21 = mf.convolution_2d_layer(
conv1_maxpool, 128, [3, 3], [1, 1], "SAME", data_format=data_format,
leaky_params=leaky_param, wd=wd, layer_name="conv21")
print(conv21)
conv2_maxpool = mf.maxpool_2d_layer(conv21, [2, 2],
[2, 2], data_format, "maxpool2")
print(conv2_maxpool)
conv31 = mf.convolution_2d_layer(
conv2_maxpool, 256, [3, 3], [1, 1], "SAME", data_format=data_format,
leaky_params=leaky_param, wd=wd, layer_name="conv31")
print(conv31)
if model_params['adapt']:
if self.model_params["adapt_loss_type"] == "MULTI":
self._add_adapt_multi_loss(conv31,
gl_ph, leaky_param, wd)
elif self.model_params["adapt_loss_type"] == "PAIR":
self._add_adapt_pair_loss(conv31,
gl_ph, leaky_param, wd)
else:
raise NotImplementedError
conv3_maxpool = mf.maxpool_2d_layer(
conv31, [2, 2], [2, 2], data_format, "maxpool3")
conv41 = mf.convolution_2d_layer(
conv3_maxpool, 256, [3, 3], [1, 1], "SAME", data_format=data_format,
leaky_params=leaky_param, wd=wd, layer_name="conv4")
fc1 = mf.fully_connected_layer(
conv41, 2048, leaky_param, wd, "fc1")
fc1_drop = tf.nn.dropout(fc1, keep_prob_ph, name="dropout1")
fc2 = mf.fully_connected_layer(
fc1_drop, 1024, leaky_param, wd, "fc2")
fc2_drop = tf.nn.dropout(fc2, keep_prob_ph, name="dropout2")
fc3 = mf.fully_connected_layer(fc2_drop, 10, 0.0, wd, "fc3")
self.fc = fc3