Autoencoder shaping issue

Question

I am having an issue with my autoencoder as I am shaping the ouput incorrectly. Currently the autoencoder is coded lke this.

I Got This Error :

ValueError: Dimensions must be equal, but are 2000 and 3750 for '{{node mean_absolute_error/sub}} = Sub[T=DT_FLOAT](sequential_8/sequential_7/conv1d_transpose_14/BiasAdd, IteratorGetNext:1)' with input shapes: [?,2000,3], [?,3750,3].

Can someone help with adjusting the architecture if this is possible. I seem to be forgetting the original modifications I made to adjust for this initially.

import tensorflow as tf
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Conv1D, MaxPooling1D, UpSampling1D, concatenate
from tensorflow.keras.callbacks import EarlyStopping

# Provided encoder
encoder = tf.keras.models.Sequential([
    tf.keras.layers.Reshape([3750, 3], input_shape=[3750, 3]),
    tf.keras.layers.Conv1D(32, kernel_size=5, padding="same", activation="relu"),
    tf.keras.layers.MaxPool1D(pool_size=2),
    tf.keras.layers.Conv1D(64, kernel_size=5, padding="same", activation="relu"),
    tf.keras.layers.MaxPool1D(pool_size=2),
    tf.keras.layers.Conv1D(128, kernel_size=5, padding="same", activation="relu"),
    tf.keras.layers.MaxPool1D(pool_size=2),
    tf.keras.layers.Conv1D(256, kernel_size=5, padding="same", activation="relu"),
    tf.keras.layers.MaxPool1D(pool_size=2),
    tf.keras.layers.Conv1D(512, kernel_size=5, padding="same", activation="relu"),
    tf.keras.layers.MaxPool1D(pool_size=2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(512)
])

#latent space

decoder = tf.keras.models.Sequential([
    tf.keras.layers.Dense(512 * 125, input_shape=[512]),
    tf.keras.layers.Reshape([125, 512]),
    tf.keras.layers.Conv1DTranspose(512, kernel_size=5, strides=1, padding="same", activation="relu"),
    tf.keras.layers.UpSampling1D(size=2),
    tf.keras.layers.Conv1DTranspose(256, kernel_size=5, strides=1, padding="same", activation="relu"),
    tf.keras.layers.UpSampling1D(size=2),
    tf.keras.layers.Conv1DTranspose(128, kernel_size=5, strides=1, padding="same", activation="relu"),
    tf.keras.layers.UpSampling1D(size=2),
    tf.keras.layers.Conv1DTranspose(64, kernel_size=5, strides=1, padding="same", activation="relu"),
    tf.keras.layers.UpSampling1D(size=2),
    # Adjust the kernel size and padding to match the input shape
    tf.keras.layers.Conv1DTranspose(3, kernel_size=5, strides=1, padding="same", activation="linear")
])

# Add more layers with larger kernel sizes to both encoder and decoder.
ae = tf.keras.models.Sequential([encoder, decoder])

ae.compile(
    loss="mean_squared_error", 
    optimizer=tf.keras.optimizers.Adam(learning_rate=0.00001)
)
# Define the early stopping criteria
early_stopping = EarlyStopping(monitor='val_loss', patience=30, mode='min')

history = ae.fit(X_train, X_train, batch_size=8, epochs=150, validation_data=(X_val, X_val), callbacks=[early_stopping])    ```

Taha Akbari · Accepted Answer · 2024-03-18 00:17:16Z

0

As the error indicates there is a dimension mismatch in your autoencoder which causes error when computing the mean-squared-error for the input and output image. You can check the output shape of your model using model.summary() which results in the following result:

A quick fix would be to add padding layers for example you can use the following architecture for decoder:

decoder = tf.keras.models.Sequential([
    tf.keras.layers.Dense(512 * 125, input_shape=[512]),
    tf.keras.layers.Reshape([125, 512]),
    tf.keras.layers.Conv1DTranspose(512, kernel_size=5, strides=1, padding="same", activation="relu"),
    tf.keras.layers.UpSampling1D(size=2),
    tf.keras.layers.Conv1DTranspose(256, kernel_size=5, strides=1, padding="same", activation="relu"),
    tf.keras.layers.UpSampling1D(size=2),
    tf.keras.layers.Conv1DTranspose(128, kernel_size=5, strides=1, padding="same", activation="relu"),
    tf.keras.layers.UpSampling1D(size=2),
    tf.keras.layers.Conv1DTranspose(64, kernel_size=5, strides=1, padding="same", activation="relu"),
    tf.keras.layers.UpSampling1D(size=2),
    tf.keras.layers.ZeroPadding1D(padding=875),
    # Adjust the kernel size and padding to match the input shape
    tf.keras.layers.Conv1DTranspose(3, kernel_size=5, strides=1, padding="same", activation="linear")
])

Remark: It is better to add these embeddings between layers instead so you don't overwhelm you model with a lot of zeros in one layer.

answered Mar 18, 2024 at 0:17

Taha Akbari

2222 silver badges9 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

Omar Eweis Over a year ago

Thank you I’ve decided to adjust some of the upsampling layers for a cleaner look .

Collectives™ on Stack Overflow

Autoencoder shaping issue

1 Answer 1

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related