在另一个模型中加载一个张量流模型，并连接两个模型

Question

The idea is to create a deconvolution model following my convolution model to see the importance of the learned pixels.

Overview of the model

I am having problems that I cannot explain. The first is that when I have created my 3 models, if I train the convolution model and test my auto model everything works fine. Whereas if I create my 3 models then I load my convolution model (trained previously) and if I test my auto model everything looks as if I had not loaded the weights of my convolution model. My first question is how to load the weights into my convolution model and that they are taken into account for my auto model ?

The second problem is perhaps related to the first one, it's that everything works well when I use predict on my auto model but if I decompose it, it doesn't work. By decomposing I mean take a x_test predict it with convolution model then use what we get to predict the deconv model. I'm getting an error when I give the result of the convolution model to the deconv model. "Invalid argument: You must feed a value for placeholder tensor 'input_CNN' with dtype float"

To create the auto model I'm doing :

inputs = layers.Input(shape(128,128,1),name='input_CNN')
model_auto = models.Model(inputs,model_deconv(model_conv(inputs)))

I can give you more details if needed.

Edit :

def CNN():
   inputs = layers.Input(shape=(128,128,1),name="input_CNN")
   layers_CNN = CNN_1_layers(inputs)
   model_conv = models.Model(inputs,layers_CNN,name='CNN_1')
   model_conv.compile(loss='categorical_crossentropy',optimizer=Adam(), metrics=[accuracy"])

   inputs_deconv = layers.Inputs(shape=(1,10),name="input_CNN_deconv")
   layers_CNN_deconv = CNN_1_deconv_layers(inputs_deconv)
   model_deconv = models.Model(inputs_deconv, layers_CNN_deconv,name="CNN_1_deconv")
   model_deconv.compile(loss='categorical_crossentropy',optimize=Adam(), metrics=["accuracy"])

   model_auto = models.Model(inputs,model_deconv(model_conv(inputs)))
   model_auto.compile(loss='categorical_crossentropy',optimize=Adam(), metrics=["accuracy"])

   return model_auto, model_conv, model_deconv

The last layer of my model_conv :

def CNN_1_layers(inputs):
   x = layers.Flattend(input_shape(1,1,10))(x)
   x = layers.Dense(10,activation='softmax')(x)

   return x

And the deconv layers :

def CNN_1_deconv_layers(inputs_deconv):
   x = layers.Reshape((1,1,10))(x)
   w1 = tf.get_variable("w1",shape=[4,4,128,10],dtype=tf.float32)
   x = layers.Lambda(lambda x: tf.nn.conv2d_transpose(x,w1,output_shape=[1,4,4,128],strides=(1,1,1,1),padding='VALID'),name='deconv_0')(x)
   ...
   return x


def apprentissage(model,nb_epoch=100):
   checkpoint = ModelCheckpoint(filepath="CNN_1.h5",monitor='vall_acc', save_best_only=True,save_weights_only=False,mode='auto')
   hist= model.fit(train_X, train_y, batch_size=128, nb_epoch=nb_epoch, validation_data=(test_X,test_y), callbacks=[checkpoint])
   return hist

I'm coding with another computer off the network so I have to rewrite everything by hand so I'm summarizing a bit.

Answer 1

I don't know if my comment helped to your first question or not (actually I did not expect your first problem to happen but anyway I gave you a suggestion).

For your second problem I've tested your code (man you had a lot of typo errors :D) and it had no problem. I will write the code below:

import numpy as np
import tensorflow as tf
from tensorflow import keras

# use below as you need
from tensorflow.keras.models import Sequential
from tensorflow.keras import layers
from tensorflow.keras.layers import *

# ----------------------

def CNN():
    inputs = layers.Input(shape=(128,128,1),name="input_CNN")
    layers_CNN = CNN_1_layers(inputs)
    model_conv = keras.Model(inputs,layers_CNN,name='CNN_1')
    model_conv.compile(loss='categorical_crossentropy',optimizer='adam', metrics=["accuracy"])

    inputs_deconv = layers.Input(shape=(1,10),name="input_CNN_deconv")
    layers_CNN_deconv = CNN_1_deconv_layers(inputs_deconv)
    model_deconv = keras.Model(inputs_deconv, layers_CNN_deconv,name="CNN_1_deconv")
    model_deconv.compile(loss='categorical_crossentropy',optimizer='adam', metrics=["accuracy"])

    model_auto = keras.Model(inputs,model_deconv(model_conv(inputs)))
    model_auto.compile(loss='categorical_crossentropy',optimizer='adam', metrics=["accuracy"])
    return model_auto, model_conv, model_deconv

# ----------------------

def CNN_1_layers(inputs):
   x = layers.Flatten(input_shape = (1,1,10))(inputs)
   x = layers.Dense(10,activation='softmax')(x)

   return x

# ----------------------

testX = np.random.rand(10, 128, 128, 1)

model_auto, model_conv, model_deconv = CNN()

print(model_auto(testX).shape)
conv_out = model_conv(testX)
deconv_out = model_deconv(conv_out)
print(deconv_out.shape)

Answer 2

Thanks Amin, You answered my question, the problem came from the type of my input variable. (uint8 instead of float32) The predict managed to work even if the type was wrong but by fixing this problem I can now decompose my model and everything works!

It only stays the problem to load the data. If you tell me you didn't have a problem doing something similar I'll do some more testing.

Answer 3

好吧，我只是放弃并将我所有的库上传到最后一个它似乎可以工作的库。