尝试跟踪Keras 2中完全卷积密集网络的this implementation。我更新了某些层的函数调用并更新了连接层
def transition_up(added, wd=0):
x = concat(added)
_, r, c, ch = x.get_shape().as_list()
return Deconvolution2D(ch, 3, 3, (None,r*2,c*2,ch), init='he_uniform',
border_mode='same', subsample=(2,2), W_regularizer=l2(wd))(x)
到
def transition_up(added, wd=0):
x = concat(added)
_, r, c, ch = x.get_shape().as_list()
return Deconv2D(ch, (3, 3), input_shape=(r*2, c*2, ch), padding='same',
kernel_initializer='he_uniform', kernel_regularizer=l2(wd)) (x)
但输入并不像在向下路径之后那样加倍。这是使用Keras 2创建模型的完整代码。
def conv(x, nf, sz, wd, p, stride=1):
x = Conv2D(nf, (sz, sz), strides=(stride, stride), padding='same',
kernel_initializer='he_uniform', kernel_regularizer=l2(wd)) (x)
return dropout(x, p)
def conv_relu_bn(x, nf, sz=3, wd=0, p=0, stride=1):
return conv(relu_bn(x), nf, sz, wd=wd, p=p, stride=stride)
def dense_block(n, x, growth_rate, p, wd):
added = []
for i in range(n):
b = conv_relu_bn(x, growth_rate, p=p, wd=wd)
print('x:', x, 'b:', b)
x = concat([x, b])
added.append(b)
return x, added
def transition_dn(x, p, wd):
return conv_relu_bn(x, x.get_shape().as_list()[-1], sz=1, p=p, wd=wd, stride=2)
def down_path(x, nb_layers, growth_rate, p, wd):
skips = []
for i, n in enumerate(nb_layers):
x, added = dense_block(n, x, growth_rate, p, wd)
skips.append(x)
x = transition_dn(x, p=p, wd=wd)
return skips, added
def transition_up(added, wd=0):
x = concat(added)
_, r, c, ch = x.get_shape().as_list()
return Deconv2D(ch, (3, 3), input_shape=(r*2, c*2, ch), padding='same',
kernel_initializer='he_uniform', kernel_regularizer=l2(wd)) (x)
def up_path(added, skips, nb_layers, growth_rate, p, wd):
for i, n in enumerate(nb_layers):
x = transition_up(added, wd)
x = concat([x, skips[i]])
x, added = dense_block(n, x, growth_rate, p, wd)
return x
def reverse(x):
return list(reversed(x))
def create_tiramisu(nb_classes, img_input, nb_dense_block=6, growth_rate=16, nb_filter=48,
nb_layers_per_block=5, p=None, wd=0):
if type(nb_layers_per_block) is list or type(nb_layers_per_block) is tuple:
nb_layers = list(nb_layers_per_block)
else:
nb_layers = [nb_layers_per_block] * nb_dense_block
x = conv(img_input, nb_filter, 3, wd, 0)
skips, added = down_path(x, nb_layers, growth_rate, p, wd)
x = up_path(added, reverse(skips[:-1]), reverse(nb_layers[:-1]), growth_rate, p, wd)
x = conv(x, nb_classes, 1, wd, 0)
_, r, c, f = x.get_shape().as_list()
x = Reshape((-1, nb_classes)) (x)
return Activation('softmax') (x)
input_shape = (224, 224, 3)
img_input = Input(shape=input_shape)
x = create_tiramisu(12, img_input, nb_layers_per_block=[4, 5, 7, 10, 12, 15], p=0.2, wd=1e-4)
当它尝试使用Keras 2语法创建模型时，这是堆栈跟踪。当在上采样路径的开头它应该是（None，14,14,656）时，连接层的一个输入不会从（None，7,7,656）加倍。使用Keras 1语法不会发生这种情况，我也可以发布。
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-133-fb96a61666a1> in <module>()
----> 1 x = create_tiramisu(12, img_input, nb_layers_per_block=[4, 5, 7, 10, 12, 15], p=0.2, wd=1e-4)
<ipython-input-131-f58faf58b799> in create_tiramisu(nb_classes, img_input, nb_dense_block, growth_rate, nb_filter, nb_layers_per_block, p, wd)
62 print('p', p)
63 print('wd', wd)
---> 64 x = up_path(added, reverse(skips[:-1]), reverse(nb_layers[:-1]), growth_rate, p, wd)
65 x = conv(x, nb_classes, 1, wd, 0)
66 _, r, c, f = x.get_shape().as_list()
<ipython-input-131-f58faf58b799> in up_path(added, skips, nb_layers, growth_rate, p, wd)
38 print('up path', 'i:', i, 'n:', n)
39 x = transition_up(added, wd)
---> 40 x = concat([x, skips[i]])
41 x, added = dense_block(n, x, growth_rate, p, wd)
42 return x
<ipython-input-130-c6d2942be960> in concat(xs)
5 def bn(x): return BatchNormalization(axis=-1)(x)
6 def relu_bn(x): return relu(bn(x))
----> 7 def concat(xs): return concatenate(xs, axis=-1)
/home/clu/anaconda3/lib/python3.6/site-packages/keras/layers/merge.py in concatenate(inputs, axis, **kwargs)
389 A tensor, the concatenation of the inputs alongside axis `axis`.
390 """
--> 391 return Concatenate(axis=axis, **kwargs)(inputs)
392
393
/home/clu/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py in __call__(self, inputs, **kwargs)
528 self.build(input_shapes[0])
529 else:
--> 530 self.build(input_shapes)
531 self.built = True
532
/home/clu/anaconda3/lib/python3.6/site-packages/keras/layers/merge.py in build(self, input_shape)
158 'inputs with matching shapes '
159 'except for the concat axis. '
--> 160 'Got inputs shapes: %s' % (input_shape))
161
162 def call(self, inputs):
ValueError: `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 7, 7, 240), (None, 14, 14, 656)]
答案 0 :(得分：1)
在反卷积层中使用2的步幅。