Understanding input_shape parameter in LSTM with Keras

I'm trying to use the example described in the Keras documentation named "Stacked LSTM for sequence classification" (see code below) and can't figure out the input_shape parameter in the context of my data.

I have as input a matrix of sequences of 25 possible characters encoded in integers to a padded sequence of maximum length 31. As a result, my x_train has the shape (1085420, 31) meaning (n_observations, sequence_length).

from keras.models import Sequential
from keras.layers import LSTM, Dense
import numpy as np

data_dim = 16
timesteps = 8
num_classes = 10

# expected input data shape: (batch_size, timesteps, data_dim)
model = Sequential()
input_shape=(timesteps, data_dim)))  # returns a sequence of vectors of dimension 32
model.add(LSTM(32, return_sequences=True))  # returns a sequence of vectors of dimension 32
model.add(LSTM(32))  # return a single vector of dimension 32

model.compile(loss='categorical_crossentropy',
optimizer='rmsprop',
metrics=['accuracy'])

# Generate dummy training data
x_train = np.random.random((1000, timesteps, data_dim))
y_train = np.random.random((1000, num_classes))

# Generate dummy validation data
x_val = np.random.random((100, timesteps, data_dim))
y_val = np.random.random((100, num_classes))

model.fit(x_train, y_train,
batch_size=64, epochs=5,
validation_data=(x_val, y_val))


In this code x_train has the shape (1000, 8, 16), as for an array of 1000 arrays of 8 arrays of 16 elements. There I get completely lost on what is what and how my data can reach this shape.

Looking at Keras doc and various tutorials and Q&A, it seems I'm missing something obvious. Can someone give me a hint of what to look for ?

LSTM shapes are tough so don't feel bad, I had to spend a couple days battling them myself:

If you will be feeding data 1 character at a time your input shape should be (31,1) since your input has 31 timesteps, 1 character each. You will need to reshape your x_train from (1085420, 31) to (1085420, 31,1) which is easily done with this command :

 x_train=x_train.reshape(x_train.shape[0],x_train.shape[1],1))


Check this git repository LSTM Keras summary diagram and i believe you should get everything crystal clear.

This git repo includes a Keras LSTM summary diagram that shows:

• the use of parameters like return_sequences, batch_size, time_step...
• the real structure of lstm layers
• the concept of these layers in keras
• how to manipulate your input and output data to match your model requirements how to stack LSTM's layers

And more

• Thank you for that, @MohammadFneish. This looks like it would be more helpful now. However, it isn't clear that this is quite an answer to the question, as opposed to a helpful suggestion. Be aware that Cross Validated is strictly a Q&A site, not a forum. Can you add [still more] information to explain the input shape parameter? – gung May 19 at 1:31
• @gung i really appreciate the way you are managing to review these answers to keep on standards, but i think that i can't elaborate even more about these parameters where there is many technical details concerning it. I just think that my answer could be helpful for developers facing similar issues with keras inputs and not necessarily this particular issue. Thanks – Mohammad Fneish May 19 at 1:43

I know it is not direct answer to your question. This is a simplified example with just one LSTM cell, helping me understand the reshape operation for the input data.

from keras.models import Model
from keras.layers import Input
from keras.layers import LSTM
import numpy as np

# define model
inputs1 = Input(shape=(2, 3))
lstm1, state_h, state_c = LSTM(1, return_sequences=True, return_state=True)(inputs1)
model = Model(inputs=inputs1, outputs=[lstm1, state_h, state_c])

# define input data
data = np.random.rand(2, 3)
data = data.reshape((1,2,3))

# make and show prediction
print(model.predict(data))


This would be an example of the LSTM network with just a single LSTM cell and with the input data of specific shape.

As it turns out, we are just predicting in here, training is not present for simplicity, but look how we needed to reshape the data (to add additional dimension) before the predict method.