4D input in LSTM layer in Keras

Question

I have data with a shape of (10000, 20, 15, 4) where num samples = 10000 , num series in time = 20 , height = 15 , weight = 4 . So I have table 15x4 which is distributed over time. Here is the model I want to train it over this data:

...
model.add((LSTM(nums-1,return_sequences=True,input_shape=(20,15,4), activation='relu')))
model.add((LSTM(nums-1,return_sequences=False,input_shape=(20,15,4), activation='tanh')))
model.add(Dense(15,activation='relu'))
...

However, I get the following error:

ValueError: Input 0 is incompatible with layer lstm_1: expected ndim=3, 
found ndim=4

How do I define a LSTM layer with 4D input shape?

Answer 1

LSTM layer accepts a 3D array as input which has a shape of (n_sample, n_timesteps, n_features) . Since the features of each timestep in your data is a (15,4) array, you need to first flatten them to a feature vector of length 60 and then pass it to your model:

X_train = X_train.reshape(10000, 20, -1)

# ...
model.add(LSTM(...,input_shape=(20,15*4), ...)) # modify input_shape accordingly

Alternatively, you can use a Flatten layer wrapped in a TimeDistributed layer as the first layer of your model to flatten each timestep:

model.add(TimeDistributed(Flatten(input_shape=(15,4))))

Further, note that if each timestep (ie array (15, 4) ) is a feature map where there is a local spatial relationship between its elements, say like an image patch, you can also use ConvLSTM2D instead of LSTM layer. Otherwise, flattening the timesteps and using LSTM would be fine.

---

As a side note: you only need to specify input_shape argument on the first layer of the model. Specifying it on other layers would be redundant and will be ignored since their input shape is automatically inferred by Keras.