Tensoflow Model.fit adds extra dimension when training with custom model

Question

I made this colab so that you can play around with the code.

tldr; The shape of my x_data is (# of examples, 12, 1, 100, 1 ) and when the data is passed to the model, the shape of each example becomes (None, 12, 1, 100, 1 ). How do I get rid of the None?

I'm trying to train a series of examples with the following data structure:

import numpy as np

x_train = np.reshape(np.random.rand(200, 12, 100),(200, 12, 1, 100, 1)) #X training data (see dimensions in output)
y_train =  np.random.rand(200, 20) #Y training data (see dimensions in output)
n_class = np.shape(y_train)[1] #Number of classes in the classification task

print(np.shape(x_train))
print(np.shape(y_train))
____________
(200, 12, 1, 100, 1)
(200, 20)

x_train dimensions explained:

There are 200 examples. Each x_train example has a corresponding y_train example with a one-hot encoded 20 dim vector representing the class it belongs to.
Each example contains 12-time series. Each time series is 100 units long and we want to convolve over them separately.
We have shaped each time series as (1, 100, 1) because of the Conv1D keras documentation.
As per the link above, the input dimensions should be: (batch_size, (steps, input_dim))

batch_size = 1, steps = 100, input_dim = 1

The model is as follows (it can also be found in the colab):

def custom_model(x_data, n_class):
    """
    This is the custom model used to convolve over the data
    We want to convolve over each dimension independently
    """
    conv_inputs = []
    conv_outputs = []
    amount_of_filters = np.shape(x_data[0])[0] 
    # Let's create a separate convolution channel for each one of the dims in x_data
    for i in range(amount_of_filters):
        x0 = keras.Input(batch_shape=(1, 100, 1))
        # conv_inputs will therefore have a size of (12, 1, 100, 1)
        conv_inputs.append(x0)
        x0 = Conv1D(filters=20,kernel_size=10,strides=2,padding="same",data_format="channels_last")(x0)
        x0 = MaxPool1D(pool_size=6, data_format="channels_last")(x0)
        x0 = Conv1D(filters=20, kernel_size=4, padding="same", data_format="channels_last")(x0)
        x0 = MaxPool1D(pool_size=3, data_format="channels_last", padding="same")(x0)
        x0 = Flatten()(x0)
        x0 = Dense(100, activation="relu")(x0)
        x0 = Dropout(0.2)(x0)
        x0 = Dense(10, activation="relu")(x0)
        conv_outputs.append(x0)
    # Convolving all outputs from the above convolutions. 
    x1 = concatenate(conv_outputs)
    x1 = Dense(100, activation="relu")(x1)
    x1 = Dense(100, activation="relu")(x1)
    out = Dense(n_class)(x1)

    modelConv = keras.Model(inputs=conv_inputs, outputs=out)

    return modelConv

When running the model:

model = custom_model(x_train, n_class)
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
history = model.fit(x_train, y_train, epochs=250, verbose=1, shuffle=True)

The following error comes up:

ValueError: Layer model expects 12 input(s), but it received 1 input tensors. Inputs received: []

You can see that the input shape is (None, 12, 1, 100, 1 ). How do I get rid of the None? As shown above, when printing the shape of x_train, the None dimension does not appear. My question is how do I feed in the x_data with each example being (12, 1, 100, 1) dimensional?

Thank you!

Tensoflow Model.fit adds extra dimension when training with custom model

0 Answers