如何用合适的输入形状训练 CNN？

Question

I am trying to train a CNN and LSTM.network with S&P 500 dataset. This is the shape of my train dataset:

xtrain shape: (6445, 16) ytrain shape: (6445,)

this is the input shape I have gave to CNN:

model.add(TimeDistributed(Conv1D(filters=64, kernel_size=1, activation='relu'), input_shape=(None,16)))

withe input shape parameter shown in the code I get this error:

ValueError: Input 0 of layer conv1d_8 is incompatible with the layer: : expected min_ndim=3, found ndim=2. Full shape received: [None, 16]

Answer 1

expected min_ndim=3, found ndim=2.

Keras expects three-dimensional arrays when working with Conv1D: the expected shape is [batch_size, sequence_length, feature_dimension] . In your case, you only have one feature dimension, I suspect the price, but imagine you also wanted to pass the trading volumes data, you would have xtrain.shape == (6445,16,2) . The last dimension would contain information about the price and the volume.

To solve your issue, you need to reshape your xtrain to

(batch_size, sequence_length, feature_dimension=(6445,16,1)

To do so, you can use tensorflow:

xtrain = tf.expand_dims(xtrain, axis=-1) # -1 means expand the LAST axis

or with numpy:

xtrain = np.expand_dims(xtrain, axis=-1) # -1 means expand the LAST axis

This function just does what the name implies: it adds a new axis in the position specified by axis . This results in xtrain having the shape we wanted, now you can just proceed with your model, for example:

model = keras.models.Sequential()
model.add(keras.layers.TimeDistributed(keras.layers.Conv1D(filters=64, kernel_size=1, activation='relu'), input_shape=(None,16,1)))