CODEWITHSUNDEEP
pythontensorflowmachine-learningkeraslstm
Eghbal
Eghbal

Reputation: 3783

Reset all weights and biases of the model in Keras (restore the model after training)

Suppose that I have something like this.

model = Sequential()
model.add(LSTM(units = 10 input_shape = (x1, x2)))
model.add(Activation('tanh'))
model.compile(optimizer = 'adam', loss = 'mse')

## Step 1.
model.fit(X_train, Y_train, epochs = 10)

After training the model, I want to reset everything (weights and biases) in the model. So I want to restore the model after compile function (Step 1). What is the fastest way to that in Keras?

Upvotes: 1

Views: 2652

Answers (1)

fuglede
fuglede

Reputation: 18201

Whether it's the fastest is probably up in the air, but it's certainly straightforward and might be good enough for your case: Serialize the initial weights, then deserialize when necessary, and use something like io.BytesIO to avoid the disk I/O hit (and having to clean up afterwards):

from io import BytesIO

model = Sequential()
model.add(LSTM(units = 10, input_shape = (x1, x2)))
model.add(Activation('tanh'))
model.compile(optimizer = 'adam', loss = 'mse')
f = BytesIO()
model.save_weights(f)  # Stores the weights
model.fit(X_train, Y_train, epochs = 10)
# [Do whatever you want with your trained model here]
model.load_weights(f)  # Resets the weights

Upvotes: 4

Related Questions