问题
I'm using Python 3.7.7. and Tensorflow 2.1.0.
I have a pre-trained U-Net network, and I want to get its encoder and its decoder.
In the following picture:
You can see a convolutional encoder-decoder architecture. I want to get the encoder part, that is, the layers that appears on the left of the image:
And the decoder part:
I get the U-Net model from this function:
def get_unet_uncompiled(img_shape = (200,200,1)):
inputs = Input(shape=img_shape)
conv1 = Conv2D(64, (5, 5), activation='relu', padding='same', data_format="channels_last", name='conv1_1')(inputs)
conv1 = Conv2D(64, (5, 5), activation='relu', padding='same', data_format="channels_last", name='conv1_2')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool1')(conv1)
conv2 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv2_1')(pool1)
conv2 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv2_2')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool2')(conv2)
conv3 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv3_1')(pool2)
conv3 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv3_2')(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool3')(conv3)
conv4 = Conv2D(256, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv4_1')(pool3)
conv4 = Conv2D(256, (4, 4), activation='relu', padding='same', data_format="channels_last", name='conv4_2')(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool4')(conv4)
conv5 = Conv2D(512, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv5_1')(pool4)
conv5 = Conv2D(512, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv5_2')(conv5)
up_conv5 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv5')(conv5)
ch, cw = get_crop_shape(conv4, up_conv5)
crop_conv4 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv4')(conv4)
up6 = concatenate([up_conv5, crop_conv4])
conv6 = Conv2D(256, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv6_1')(up6)
conv6 = Conv2D(256, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv6_2')(conv6)
up_conv6 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv6')(conv6)
ch, cw = get_crop_shape(conv3, up_conv6)
crop_conv3 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv3')(conv3)
up7 = concatenate([up_conv6, crop_conv3])
conv7 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv7_1')(up7)
conv7 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv7_2')(conv7)
up_conv7 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv7')(conv7)
ch, cw = get_crop_shape(conv2, up_conv7)
crop_conv2 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv2')(conv2)
up8 = concatenate([up_conv7, crop_conv2])
conv8 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv8_1')(up8)
conv8 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv8_2')(conv8)
up_conv8 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv8')(conv8)
ch, cw = get_crop_shape(conv1, up_conv8)
crop_conv1 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv1')(conv1)
up9 = concatenate([up_conv8, crop_conv1])
conv9 = Conv2D(64, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv9_1')(up9)
conv9 = Conv2D(64, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv9_2')(conv9)
ch, cw = get_crop_shape(inputs, conv9)
conv9 = ZeroPadding2D(padding=(ch, cw), data_format="channels_last", name='conv9_3')(conv9)
conv10 = Conv2D(1, (1, 1), activation='sigmoid', data_format="channels_last", name='conv10_1')(conv9)
model = Model(inputs=inputs, outputs=conv10)
return model
The auxiliary function is:
def get_crop_shape(target, refer):
# width, the 3rd dimension
cw = (target.get_shape()[2] - refer.get_shape()[2])
assert (cw >= 0)
if cw % 2 != 0:
cw1, cw2 = cw // 2, cw // 2 + 1
else:
cw1, cw2 = cw // 2, cw // 2
# height, the 2nd dimension
ch = (target.get_shape()[1] - refer.get_shape()[1])
assert (ch >= 0)
if ch % 2 != 0:
ch1, ch2 = ch // 2, ch // 2 + 1
else:
ch1, ch2 = ch // 2, ch // 2
return (ch1, ch2), (cw1, cw2)
The graph for this model is:
I train the network, and after train it I get the encoder doing this:
first_encoder_layer = 0
last_encoder_layer = 14
old_model = get_unet_uncompiled()
old_model.compile(tf.keras.optimizers.Adam(lr=(1e-4) * 2),
loss='binary_crossentropy',
metrics=['accuracy'])
encoder: Model = Model(inputs=old_model.layers[first_encoder_layer].input,
outputs=old_model.layers[last_encoder_layer].output,
name='encoder')
And it works perfectly.
But, when I try to extract the decoder with:
decoder: Model = Model(inputs=old_model.layers[last_encoder_layer + 1].input,
outputs=old_model.layers[-1].output,
name='decoder')
I get these errors which I don't understand:
WARNING:tensorflow:Model inputs must come from `tf.keras.Input` (thus holding past layer metadata), they cannot be the output of a previous non-Input layer. Here, a tensor specified as input to "decoder" was not an Input tensor, it was generated by layer up_conv5.
Note that input tensors are instantiated via `tensor = tf.keras.Input(shape)`.
The tensor that caused the issue was: up_conv5/Identity:0
WARNING:tensorflow:Model inputs must come from `tf.keras.Input` (thus holding past layer metadata), they cannot be the output of a previous non-Input layer. Here, a tensor specified as input to "decoder" was not an Input tensor, it was generated by layer crop_conv4.
Note that input tensors are instantiated via `tensor = tf.keras.Input(shape)`.
The tensor that caused the issue was: crop_conv4/Identity:0
A TraceBack, and then another error:
ValueError: Graph disconnected: cannot obtain value for tensor Tensor("input_1:0", shape=(None, 200, 200, 1), dtype=float32) at layer "input_1". The following previous layers were accessed without issue: []
I have tried this code to get the decoder:
decoder_input = Input(shape=(12, 12, 512), name='dec_input')
z = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_dec_conv5')(decoder_input)
decoder: Model = Model(inputs=z,
outputs=old_model.layers[-1].output,
name='decoder')
But I get mostly the same error:
WARNING:tensorflow:Model inputs must come from `tf.keras.Input` (thus holding past layer metadata), they cannot be the output of a previous non-Input layer. Here, a tensor specified as input to "decoder" was not an Input tensor, it was generated by layer up_dec_conv5.
Note that input tensors are instantiated via `tensor = tf.keras.Input(shape)`.
The tensor that caused the issue was: up_dec_conv5/Identity:0
How can I get the decoder from a pretrained U-Net network?
回答1:
what I suggest is to define a structure for encoder and decoder (get_encoder, get_decoder). After the training of the whole model, the idea is to create a new decoder architecture (through get_decoder) that we can fill with the decoder trained weights
pythonic speaking you can do that in this way...
def get_crop_shape(target, refer):
# width, the 3rd dimension
cw = (target.get_shape()[2] - refer.get_shape()[2])
assert (cw >= 0)
if cw % 2 != 0:
cw1, cw2 = cw // 2, cw // 2 + 1
else:
cw1, cw2 = cw // 2, cw // 2
# height, the 2nd dimension
ch = (target.get_shape()[1] - refer.get_shape()[1])
assert (ch >= 0)
if ch % 2 != 0:
ch1, ch2 = ch // 2, ch // 2 + 1
else:
ch1, ch2 = ch // 2, ch // 2
return (ch1, ch2), (cw1, cw2)
def get_encoder(img_shape):
inp = Input(shape=img_shape)
conv1 = Conv2D(64, (5, 5), activation='relu', padding='same', data_format="channels_last", name='conv1_1')(inp)
conv1 = Conv2D(64, (5, 5), activation='relu', padding='same', data_format="channels_last", name='conv1_2')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool1')(conv1)
conv2 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv2_1')(pool1)
conv2 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv2_2')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool2')(conv2)
conv3 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv3_1')(pool2)
conv3 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv3_2')(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool3')(conv3)
conv4 = Conv2D(256, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv4_1')(pool3)
conv4 = Conv2D(256, (4, 4), activation='relu', padding='same', data_format="channels_last", name='conv4_2')(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2), data_format="channels_last", name='pool4')(conv4)
conv5 = Conv2D(512, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv5_1')(pool4)
conv5 = Conv2D(512, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv5_2')(conv5)
return conv5,conv4,conv3,conv2,conv1,inp
def get_decoder(convs):
conv5,conv4,conv3,conv2,conv1,inputs = convs
up_conv5 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv5')(conv5)
ch, cw = get_crop_shape(conv4, up_conv5)
crop_conv4 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv4')(conv4)
up6 = concatenate([up_conv5, crop_conv4])
conv6 = Conv2D(256, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv6_1')(up6)
conv6 = Conv2D(256, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv6_2')(conv6)
up_conv6 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv6')(conv6)
ch, cw = get_crop_shape(conv3, up_conv6)
crop_conv3 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv3')(conv3)
up7 = concatenate([up_conv6, crop_conv3])
conv7 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv7_1')(up7)
conv7 = Conv2D(128, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv7_2')(conv7)
up_conv7 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv7')(conv7)
ch, cw = get_crop_shape(conv2, up_conv7)
crop_conv2 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv2')(conv2)
up8 = concatenate([up_conv7, crop_conv2])
conv8 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv8_1')(up8)
conv8 = Conv2D(96, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv8_2')(conv8)
up_conv8 = UpSampling2D(size=(2, 2), data_format="channels_last", name='up_conv8')(conv8)
ch, cw = get_crop_shape(conv1, up_conv8)
crop_conv1 = Cropping2D(cropping=(ch, cw), data_format="channels_last", name='crop_conv1')(conv1)
up9 = concatenate([up_conv8, crop_conv1])
conv9 = Conv2D(64, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv9_1')(up9)
conv9 = Conv2D(64, (3, 3), activation='relu', padding='same', data_format="channels_last", name='conv9_2')(conv9)
ch, cw = get_crop_shape(inputs, conv9)
conv9 = ZeroPadding2D(padding=(ch, cw), data_format="channels_last", name='conv9_3')(conv9)
conv10 = Conv2D(1, (1, 1), activation='sigmoid', data_format="channels_last", name='conv10_1')(conv9)
return conv10
def get_unet(img_shape = (200,200,1)):
enc = get_encoder(img_shape)
dec = get_decoder(enc)
model = Model(inputs=enc[-1], outputs=dec)
return model
create the whole model and fit
img_shape = (200,200,1)
old_model = get_unet(img_shape)
# old_model.compile(...)
# old_model.fit(...)
extract encoder as always
# extract encoder
first_encoder_layer = 0
last_encoder_layer = 14
encoder_output_layer = [14, 11, 8, 5, 2, 0]
encoder = Model(inputs=old_model.layers[first_encoder_layer].input,
outputs=[old_model.layers[l].output for l in encoder_output_layer],
name='encoder')
encoder.summary()
create the decoder structure and assign the trained weights
# extract decoder fitted weights
restored_w = []
for w in old_model.layers[last_encoder_layer + 1:]:
restored_w.extend(w.get_weights())
# reconstruct decoder architecture setting the fitted weights
new_inp = [Input(l.shape[1:]) for l in get_encoder(img_shape)]
new_dec = get_decoder(new_inp)
decoder = Model(new_inp, new_dec)
decoder.set_weights(restored_w)
decoder.summary()
return predictions
# generate random images
n_images = 20
X = np.random.uniform(0,1, (n_images,200,200,1)).astype('float32')
# get encoder predictions
pred_encoder = encoder.predict(X)
print([p.shape for p in pred_encoder])
# get decoder predictions
pred_decoder = decoder.predict(pred_encoder)
print(pred_decoder.shape)
here the running notebook
来源：https://stackoverflow.com/questions/63756756/model-inputs-must-come-from-tf-keras-input-they-cannot-be-the-output-of-a