CODEWITHSUNDEEP
pythontensorflowkerasneural-network
Peter Smith
Peter Smith

Reputation: 23

Specify List of Parameters in Keras

I have defined two Variables biases for weights and biases. How do I use those variables in Keras? Basically, what I am trying to do is as follows:

w = tf.get_variable("weight", shape=[784, 512], trainable=True)
b = tf.get_variable("bias", shape=[512], trainable=True)

model = Sequential()
model.add(Dense(512, activation='relu', input_shape=(784,), weights=w, biases=b))

Does anyone know how to do this with Keras?

Upvotes: 2

Views: 913

Answers (1)

OverLordGoldDragon
OverLordGoldDragon

Reputation: 19836

Pass in a Numpy array directly, Keras will handle tensor conversion for you; also, weights handles both the 'regular' weights, and biases. Full example below:

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

input_shape = (784,)
dense_dim = 512

W = np.random.randn(input_shape[0], dense_dim)
b = np.random.randn(dense_dim)

model = Sequential()
model.add(Dense(dense_dim, activation='relu', input_shape=input_shape, weights=[W, b]))

Be sure to pass in the weights in order which the layer expects them - which can be inspected directly:

print(model.layers[0].weights)

[<tf.Variable 'dense_1/kernel:0' shape=(784, 512) dtype=float32_ref>,
 <tf.Variable 'dense_1/bias:0' shape=(512,) dtype=float32_ref>]

Set weights after building model: use layer.set_weights():

model.layers[0].set_weights([W, b]) # again, mind the order

Using tf.get_variable: can't do; from set_weights() source code, K.batch_set_value is used, which operates on raw array values rather than Tensors. If your goal is to track a layer's weight variables, you can simply fetch directly, and use K.eval() to get their values (or .numpy() for TF2):

import keras.backend as K
dense1_weights, dense1_biases = model.layers[0].weights

if tf.__version__[0] == '2':
    print(dense1_weights.numpy())
else:
    print(K.eval(dense1_weights))

Upvotes: 1

Related Questions