def _deconv2_wrapper(self, input_tensor, sample_tensor, output_channel, wd,
layer_name):
[b, h, w, _] = sample_tensor.get_shape().as_list()
[_, _, _, c] = input_tensor.get_shape().as_list()
deconv = mf.deconvolution_2d_layer(input_tensor,
[3, 3, output_channel, c], [2, 2],
[b, h, w, output_channel], 'VALID',
wd, layer_name)
return deconv
def _deconv2_wrapper(self, input_tensor, sample_tensor, output_channel, wd,
layer_name):
[b, h, w, _] = sample_tensor.get_shape().as_list()
[_, _, _, c] = input_tensor.get_shape().as_list()
b = tf.shape(sample_tensor)[0]
#deconv_shape = [16, h, w, output_channel]
deconv_shape = tf.pack([b, h, w, output_channel])
deconv = mf.deconvolution_2d_layer(input_tensor,
[3, 3, output_channel, c], [2, 2],
deconv_shape, 'VALID', wd,
layer_name)
deconv.set_shape([None, h, w, output_channel])
return deconv
def _resize_deconv(self, input_tensor, desire_shape, output_channel, wd,
layer_name):
b, w, h, c = input_tensor.get_shape().as_list()
with tf.variable_scope(layer_name):
if w != desire_shape[0] or h != desire_shape[1]:
input_tensor = tf.image.resize_images(
input_tensor, [desire_shape[0], desire_shape[1]])
deconv = mf.deconvolution_2d_layer(
input_tensor, [3, 3, output_channel, c], [1, 1],
[b, desire_shape[0], desire_shape[1], output_channel], "SAME",
wd, 'deconv')
return deconv
def _deconv2_wrapper(self, input_tensor, sample_tensor, output_channel, wd,
lp, layer_name):
[b, h, w, _] = sample_tensor.get_shape().as_list()
[_, _, _, c] = input_tensor.get_shape().as_list()
deconv = mf.deconvolution_2d_layer(input_tensor, output_channel,
[3, 3], [2, 2], "VALID", [h, w],
"NHWC", False, False, lp, wd,
layer_name)
# deconv = mf.deconvolution_2d_layer(input_tensor,
# [3, 3, output_channel, c],
# [2, 2], [b, h, w, output_channel], 'VALID',
# wd, layer_name)
return deconv