ValueError：检查目标时出错：预期activation_5 的形状为（1，），但数组的形状为（100，）

Question

I'm currently trying to train a LSTM network with sequential video data. The problem that keeps occurring now though is an error with the output shape of the data itself. From the video, I have generated 686 sample clips of 100 time steps (frames). Then with another CNN, I've created an embedding of shape 2048 of each image. In other words, the shape of X_train in my case is (686,100,2048) and the shape of Y_train is (686,100). Now when I pass my dataset through my network, I get this shape error.

My model:

from keras.layers import Activation, Input, Dense, Lambda, LSTM, Flatten
from keras.models import Model



def model_builder(input_shape):
    base_input = Input(shape = input_shape)
    x = LSTM(units=50, name='LSTM1', return_sequences=True)(base_input)
    x = Flatten()(x)
    x = Dense(units = 3)(x)
    x = Activation('softmax')(x)
    classification_model = Model(base_input, x,name='classifier')
    classification_model.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['accuracy'])

    return classification_model

And I run it like this:

batch_size = 64
epochs = 20
timesteps = 100
embedding_shape=2048

classification_model = model_builder((timesteps,embedding_shape))

try:
    Y_train=Y_train.reshape((686,timesteps))
    X_train = np.reshape(X_train,(686, timesteps,embedding_shape))

    outcome = classification_model.fit(x=X_train, y=Y_train, batch_size=batch_size, epochs=epochs, verbose=1, callbacks=None, validation_split=(6200/68600), validation_data=None, shuffle=False, class_weight=None, sample_weight=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None, validation_freq=1)

except KeyboardInterrupt:
    pass

Resulting in this error message:

Any ideas as to what I might be doing wrong?

Model Summary:

Answer 1

Without checking your code in depth, 99% of this Errors occur if you have the false loss function.

Please modify your model compilation with the loss='categorical_crossentropy' instead of loss='sparse_categorical_crossentropy':

classification_model.compile(loss='categorical_crossentropy',........)

The difference is the encodings of you targets. If your targets are one-hot encoded, use categorical_crossentropy.