How to load trained autoencoder weights for decoder?

问题

I have a CNN 1d autoencoder which has a dense central layer. I would like to train this Autoencoder and save its model. I would also like to save the decoder part, with this goal: feed some central features (calculated independently) to the trained and loaded decoder, to see what are the images of these independently calculated features through the decoder.

## ENCODER
encoder_input = Input(batch_shape=(None,501,1))
x  = Conv1D(256,3, activation='tanh', padding='valid')(encoder_input)
x  = MaxPooling1D(2)(x)
x  = Conv1D(32,3, activation='tanh', padding='valid')(x)
x  = MaxPooling1D(2)(x)
_x = Flatten()(x)
encoded = Dense(32,activation = 'tanh')(_x)

## DECODER (autoencoder)
y = Conv1D(32, 3, activation='tanh', padding='valid')(x)
y = UpSampling1D(2)(y)
y = Conv1D(256, 3, activation='tanh', padding='valid')(y)
y = UpSampling1D(2)(y)
y = Flatten()(y)
y = Dense(501)(y)
decoded = Reshape((501,1))(y)

autoencoder = Model(encoder_input, decoded)
autoencoder.save('autoencoder.hdf5')

## DECODER (independent)
decoder_input = Input(batch_shape=K.int_shape(x))  # import keras.backend as K
y = Conv1D(32, 3, activation='tanh', padding='valid')(decoder_input)
y = UpSampling1D(2)(y)
y = Conv1D(256, 3, activation='tanh', padding='valid')(y)
y = UpSampling1D(2)(y)
y = Flatten()(y)
y = Dense(501)(y)
decoded = Reshape((501,1))(y)

decoder = Model(decoder_input, decoded)
decoder.save('decoder.hdf5')

EDIT:

Just to make sure that it is clear, I first need to JOIN encoded and the first y, in the sense that y has to take encoded as input. Once this is done, I need a way to load a trained decoder and replace encoded with some new central features, which I will feed my decoder with.

EDIT following answer:

I implemented the suggestion, see code below

## ENCODER
encoder_input = Input(batch_shape=(None,501,1))
x  = Conv1D(256,3, activation='tanh', padding='valid')(encoder_input)
x  = MaxPooling1D(2)(x)
x  = Conv1D(32,3, activation='tanh', padding='valid')(x)
x  = MaxPooling1D(2)(x)
_x = Flatten()(x)
encoded = Dense(32,activation = 'tanh')(_x)

## DECODER (autoencoder)
encoded = Reshape((32,1))(encoded)
y = Conv1D(32, 3, activation='tanh', padding='valid')(encoded)
y = UpSampling1D(2)(y)
y = Conv1D(256, 3, activation='tanh', padding='valid')(y)
y = UpSampling1D(2)(y)
y = Flatten()(y)
y = Dense(501)(y)
decoded = Reshape((501,1))(y)

autoencoder = Model(encoder_input, decoded)
autoencoder.compile(optimizer='adam', loss='mse')
epochs = 10
batch_size = 100
validation_split = 0.2
# train the model
history = autoencoder.fit(x = training, y = training,
                    epochs=epochs,
                    batch_size=batch_size,
                    validation_split=validation_split)
autoencoder.save_weights('autoencoder_weights.h5')


## DECODER (independent)
decoder_input = Input(batch_shape=K.int_shape(encoded))  # import keras.backend as K
y = Conv1D(32, 3, activation='tanh', padding='valid', name='decod_conv1d_1')(decoder_input)
y = UpSampling1D(2, name='decod_upsampling1d_1')(y)
y = Conv1D(256, 3, activation='tanh', padding='valid', name='decod_conv1d_2')(y)
y = UpSampling1D(2, name='decod_upsampling1d_2')(y)
y = Flatten(name='decod_flatten')(y)
y = Dense(501, name='decod_dense1')(y)
decoded = Reshape((501,1), name='decod_reshape')(y)

decoder = Model(decoder_input, decoded)
decoder.save_weights('decoder_weights.h5')


encoder = Model(inputs=encoder_input, outputs=encoded, name='encoder')
features = encoder.predict(training) # features
np.savetxt('features.txt', np.squeeze(features))

predictions = autoencoder.predict(training)
predictions = np.squeeze(predictions)
np.savetxt('predictions.txt', predictions)

Then I open another file and I do

import h5py
import keras.backend as K

def load_weights(model, filepath):
    with h5py.File(filepath, mode='r') as f:
        file_layer_names = [n.decode('utf8') for n in f.attrs['layer_names']]
        model_layer_names = [layer.name for layer in model.layers]

        weight_values_to_load = []
        for name in file_layer_names:
            if name not in model_layer_names:
                print(name, "is ignored; skipping")
                continue
            g = f[name]
            weight_names = [n.decode('utf8') for n in g.attrs['weight_names']]

            weight_values = []
            if len(weight_names) != 0:
                weight_values = [g[weight_name] for weight_name in weight_names]
            try:
                layer = model.get_layer(name=name)
            except:
                layer = None
            if layer is not None:
                symbolic_weights = (layer.trainable_weights + 
                                    layer.non_trainable_weights)
                if len(symbolic_weights) != len(weight_values):
                    print('Model & file weights shapes mismatch')
                else:
                    weight_values_to_load += zip(symbolic_weights, weight_values)

        K.batch_set_value(weight_values_to_load)

## DECODER (independent)
decoder_input = Input(batch_shape=(None,32,1))
y = Conv1D(32, 3, activation='tanh',padding='valid',name='decod_conv1d_1')(decoder_input)
y = UpSampling1D(2, name='decod_upsampling1d_1')(y)
y = Conv1D(256, 3, activation='tanh', padding='valid', name='decod_conv1d_2')(y)
y = UpSampling1D(2, name='decod_upsampling1d_2')(y)
y = Flatten(name='decod_flatten')(y)
y = Dense(501, name='decod_dense1')(y)
decoded = Reshape((501,1), name='decod_reshape')(y)

decoder = Model(decoder_input, decoded)
#decoder.save_weights('decoder_weights.h5')

load_weights(decoder, 'autoencoder_weights.h5')

# Read autoencoder
decoder.summary()

# read encoded features
features = np.loadtxt('features.txt'.format(batch_size, epochs))
features = np.reshape(features, [1500,32,1])

# evaluate loaded model on features
prediction = decoder.predict(features)



autoencoderpredictions = np.loadtxt('predictions.txt'.format(batch_size, epochs))

fig, ax = plt.subplots(5, figsize=(10,20))
for i in range(5):
        ax[i].plot(prediction[100*i], color='blue', label='Decoder')
        ax[i].plot(autoencoderpredictions[100*i], color='red', label='AE')
        ax[i].set_xlabel('Time components', fontsize='x-large')
        ax[i].set_ylabel('Amplitude', fontsize='x-large')
        ax[i].set_title('Seismogram n. {:}'.format(1500+100*i+1), fontsize='x-large')
        ax[i].legend(fontsize='x-large')
plt.subplots_adjust(hspace=1)
plt.close()

prediction and autoencoderpredictions do not agree. It seems as if prediction is just small noise, whereas autoencoder predictions has reasonable values.

回答1:

You'll need to: (1) save weights of AE (autoencoder); (2) load weights file; (3) deserialize the file and assign only those weights that are compatible with the new model (decoder).

(1): .save does include the weights, but with an extra deserialization step that's spared by using .save_weights instead. Also, .save saves optimizer state and model architecture, latter which is irrelevant for your new decoder
(2): load_weights by default attempts to assign all saved weights, which won't work

Code below accomplishes (3) (and remedies (2)) as follows:

Load all weights
Retrieve loaded weight names and store them in file_layer_names (list)
Retrieve current model weight names and store them in model_layer_names (list)
Iterate over file_layer_names as name; if name is in model_layer_names, append loaded weight with that name to weight_values_to_load
Assign weights in weight_values_to_load to model using K.batch_set_value

Note that this requires you to name every layer in both AE and decoder models and make them match. It's possible to rewrite this code to brute-force assign sequentially in a try-except loop, but that's both inefficient and bug-prone.

Usage:

## omitted; use code as in question but name all ## DECODER layers as below
autoencoder.save_weights('autoencoder_weights.h5')

## DECODER (independent)
decoder_input = Input(batch_shape=K.int_shape(x))
y = Conv1D(32, 3, activation='tanh',padding='valid',name='decod_conv1d_1')(decoder_input)
y = UpSampling1D(2, name='decod_upsampling1d_1')(y)
y = Conv1D(256, 3, activation='tanh', padding='valid', name='decod_conv1d_2')(y)
y = UpSampling1D(2, name='decod_upsampling1d_2')(y)
y = Flatten(name='decod_flatten')(y)
y = Dense(501, name='decod_dense1')(y)
decoded = Reshape((501,1), name='decod_reshape')(y)

decoder = Model(decoder_input, decoded)
decoder.save_weights('decoder_weights.h5')

load_weights(decoder, 'autoencoder_weights.h5')

Function:

import h5py
import keras.backend as K

def load_weights(model, filepath):
    with h5py.File(filepath, mode='r') as f:
        file_layer_names = [n.decode('utf8') for n in f.attrs['layer_names']]
        model_layer_names = [layer.name for layer in model.layers]

        weight_values_to_load = []
        for name in file_layer_names:
            if name not in model_layer_names:
                print(name, "is ignored; skipping")
                continue
            g = f[name]
            weight_names = [n.decode('utf8') for n in g.attrs['weight_names']]

            weight_values = []
            if len(weight_names) != 0:
                weight_values = [g[weight_name] for weight_name in weight_names]
            try:
                layer = model.get_layer(name=name)
            except:
                layer = None
            if layer is not None:
                symbolic_weights = (layer.trainable_weights + 
                                    layer.non_trainable_weights)
                if len(symbolic_weights) != len(weight_values):
                    print('Model & file weights shapes mismatch')
                else:
                    weight_values_to_load += zip(symbolic_weights, weight_values)

        K.batch_set_value(weight_values_to_load)

来源：https://stackoverflow.com/questions/58364974/how-to-load-trained-autoencoder-weights-for-decoder

标签

python

tensorflow

keras

autoencoder