如何确保 con1D 的 output_shape 与 keras 自动编码器中时间序列的 input_shape 相同？

Question

Conv1D output shape incorrect in keras autoencoder model when running autoencoder fit.

I try to use keras autoencoder model to compress and decompress my time-series data. but when I change the layer with Conv1D , the output shape is incorrect.

I have some time series data with the shape of (4000, 689), where represent 4000 samples and each sample has 689 features. I want to use Conv1D to compress the data but the Upsampling layer's and last Conv1D layer's output shape(?, 688, 1) is not equal to input shape (, 689, 1).

How should I set those layers' parameters? Thanks in advance.

x_train = data[0:4000].values
x_test = data[4000:].values
print('x_train shape:', x_train.shape)
print('x_test shape:', x_test.shape)

x_train shape: (4000, 689)
x_test shape: (202, 689)

I reshaped the x_train, x_test to 3dim, like below.

x_tr = x_train.reshape(4000,689,1)
x_te = x_test.reshape(202,689,1)
print('x_tr shape:', x_tr.shape)
print('x_te shape:', x_te.shape)

x_tr shape: (4000, 689, 1)
x_te shape: (202, 689, 1)

input_img = Input(shape=(689,1))

x = Conv1D(16, 3, activation='relu', padding='same')(input_img)
print(x)
x = MaxPooling1D(2, padding='same')(x)
print(x)
x = Conv1D(8, 3, activation='relu', padding='same')(x)
print(x)
x = MaxPooling1D(2, padding='same')(x)
print(x)
x = Conv1D(8, 3, activation='relu', padding='same')(x)
print(x)
encoded = MaxPooling1D(2)(x)
print(encoded)
print('--------------')
    
    
x = Conv1D(8, 3, activation='relu', padding='same')(encoded)
print(x)
x = UpSampling1D(2)(x)
print(x)
x = Conv1D(8, 3, activation='relu', padding='same')(x)
print(x)
x = UpSampling1D(2)(x)
print(x)
x = Conv1D(16, 3, activation='relu', padding='same')(x)
print(x)
x = UpSampling1D(2)(x)
print(x)
decoded = Conv1D(1, 3, activation='sigmoid', padding='same')(x)
print(decoded)

autoencoder = Model(input_img, decoded)
autoencoder.compile(optimizer='adam', loss='mse')

When I imported those models and run cells above in Jupyter, It seems ok. Maybe. But I get the error in next code when running autoencoder.fit .

autoencoder.fit(x_tr, x_tr, epochs=50, batch_size=128, shuffle=True, validation_data=(x_te, x_te)) 

So I print each layer.

The each layers' print result below.

Tensor("conv1d_166/Relu:0", shape=(?, 689, 16), dtype=float32)
Tensor("max_pooling1d_71/Squeeze:0", shape=(?, 345, 16), dtype=float32)
Tensor("conv1d_167/Relu:0", shape=(?, 345, 8), dtype=float32)
Tensor("max_pooling1d_72/Squeeze:0", shape=(?, 173, 8), dtype=float32)
Tensor("conv1d_168/Relu:0", shape=(?, 173, 8), dtype=float32)
Tensor("max_pooling1d_73/Squeeze:0", shape=(?, 86, 8), dtype=float32)

Tensor("conv1d_169/Relu:0", shape=(?, 86, 8), dtype=float32)
Tensor("up_sampling1d_67/concat:0", shape=(?, 172, 8), dtype=float32)
Tensor("conv1d_170/Relu:0", shape=(?, 172, 8), dtype=float32)
Tensor("up_sampling1d_68/concat:0", shape=(?, 344, 8), dtype=float32)
Tensor("conv1d_171/Relu:0", shape=(?, 344, 16), dtype=float32)
Tensor("up_sampling1d_69/concat:0", shape=(?, 688, 16), dtype=float32)
Tensor("conv1d_172/Sigmoid:0", shape=(?, 688, 1), dtype=float32) 

ValueError bellow:

ValueError                                Traceback (most recent call last)
<ipython-input-74-56836006a800> in <module>
      3                 batch_size=128,
      4                 shuffle=True,
----> 5                 validation_data=(x_te, x_te)
      6                 )

~/anaconda3/envs/keras/lib/python3.6/site-packages/keras/engine/training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)
    950             sample_weight=sample_weight,
    951             class_weight=class_weight,
--> 952             batch_size=batch_size)
    953         # Prepare validation data.
    954         do_validation = False

~/anaconda3/envs/keras/lib/python3.6/site-packages/keras/engine/training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, check_array_lengths, batch_size)
    787                 feed_output_shapes,
    788                 check_batch_axis=False,  # Don't enforce the batch size.
--> 789                 exception_prefix='target')
    790 
    791             # Generate sample-wise weight values given the `sample_weight` and

~/anaconda3/envs/keras/lib/python3.6/site-packages/keras/engine/training_utils.py in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)
    136                             ': expected ' + names[i] + ' to have shape ' +
    137                             str(shape) + ' but got array with shape ' +
--> 138                             str(data_shape))
    139     return data
    140 

ValueError: Error when checking target: expected conv1d_172 to have shape (688, 1) but got array with shape (689, 1)

Is floor function make this happen?
How to fix the error and autoencoder.fit correctly?
Thanks in advance.

Answer 1

Using Convolutional layers, you need to infer your output size based on the input size, kernel size and other parameters. The simplest way to do so is to feed a data sample through the.network and see the final vector size after your last convolutional layer. Then, you can define further layers based on that size.

Here is an example from my pytorch project:

def _infer_flat_size(self):
        encoder_output = self.encoder(torch.ones(1, *self.input_size))
        return int(np.prod(encoder_output.size()[1:])), encoder_output.size()[1:]