ValueError: Input 0 of layer "sequential" is incompatible with the layer

Question

I am currently implementing my first project using keras/tensorflow, and so far it works pretty good, but now I ran into an error I don't know how to solve. I googled a lot and found lots of relating questions, but none solved my issue.

Here is what I want to achieve:

I have a model for a chatbot that I trained with a dataset to provide "standard" conversations like hello, how are u and stuff. Now I want to "extend" the existing Model with a dataset that provides answers to questions related to shipping, what's in stock etc.

Here is my working/already trained model:

# create Sequential model
    model = Sequential()

    # add first layer with input shape dependent on size of input and "relu" activation function
    model.add(Dense(256, input_shape=(len(training_data_x[0]),), activation=activations.relu))

    # add Dropout to prevent overfitting
    model.add(Dropout(0.6))
    # additional layer with 64 neurons
    model.add(Dense(128, activation=activations.relu))
    model.add(Dropout(0.2))
    # Additional dense layer with num of neurons of classes & softmax activation function
    # -> adds results in output layer to "1" to get %
    model.add(Dense(len(training_data_y[0]), activation=activations.softmax))
    # print(len(training_data_y[0])) = 71
    sgd = SGD(learning_rate=0.01, decay=1e-6, momentum=0.9, nesterov=True)
    # compile the model
    model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])

    output = model.fit(np.array(training_data_x), np.array(training_data_y), epochs=200, batch_size=5, verbose=1)
    plot_model_output(output)
    model.summary()
    model.save('./MyModel_tf', save_format='tf')

The training data is prepared in a separate class and takes a certain json file as input.

Now I just swapped the JSON file for the one with data related to the stuff I want to add to the model and tried fitting it like this:

json_data = json.loads(open('data.json').read())

model = load_model('MyModel_tf')

model.fit(np.array(training_data_x), np.array(training_data_y), epochs=200, batch_size=5, verbose=1)

However when I run it I get this error:

ValueError: Input 0 of layer "sequential" is incompatible with the layer: expected shape=(None, 652), found shape=(None, 71)

I am assuming the data is the problem.. however it is structured exactly the same just way shorter.

Now my question(s):

1. does it make sense the way I try to implement it?
2. should I try adding the additional data in a different way?
3. Does the second dataset have to be the same length as the first one?

Any help appreciated!

Answer 1

The problem is probably coming from this line

model.add(Dense(256, input_shape=(len(training_data_x[0]),), activation=activations.relu))

where you define the input shape of your model based on the size of the feature dimension of training_data_x . Now that you have defined this very specific input shape, all data fed into your model must have the same feature dimension size. That is the reason for your error.