Keras Conv1d输入形状问题，conv1d层的输入0与层不兼容：：预期min_ndim=3，发现ndim=2

Question

I am trying to use a conv1d to predict time series, but I have trouble with the conv1d input shape. My data 406 samples of 10 values in temporal order. The goal is to predict sample N+1 using sample N as input.

Here is a sample of my code:

print(data_x.shape)
# (406, 10)
print(data_y.shape)
# (406, 10)


inputs = Input(10, 1)
x = Conv1D(64, 2, input_shape=(10,1))(inputs)
x = Dense(64, "relu")(x)
x = Dense(64, "relu")(x)
x = Dense(10, "sigmoid")(x)

model = Model(inputs, x)
model.compile(loss='mse', metrics=['accuracy'], optimizer='adam')
history = model.fit(data_x, data_y,
                    batch_size=10, epochs=EPOCHS)

But I get this error ValueError: Input 0 of layer conv1d is incompatible with the layer: : expected min_ndim=3, found ndim=2. Full shape received: (1, 10) ValueError: Input 0 of layer conv1d is incompatible with the layer: : expected min_ndim=3, found ndim=2. Full shape received: (1, 10) .

I don't know what I am missing, I even tried to do data_x = data_x.reshape(-1,10,1) but with the same results.

Answer 1

Use tf.expand_dims(x, axis=0) instead of reshape.

Working sample code

import tensorflow as tf

inputs = (10, 1)
x = tf.random.normal(inputs)
inp = tf.expand_dims(x, axis=0)
x = tf.keras.layers.Conv1D(64, 2, input_shape=(10,1))(inp)

Output

(1, 9, 64)