How to set weights in Keras with a numpy array?

Using NumPy arrays to set model weights in Keras is a common practice, particularly when you have pre-trained weights or weights trained in different environments. Below, I'll provide a detailed example to explain how to set weights in Keras using NumPy arrays.

Step 1: Import Necessary Libraries

First, we need to import the required libraries, including Keras and NumPy, since we'll use NumPy arrays to manipulate weights.

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

Step 2: Create the Model

Next, we create a simple model. Here, I'll construct a model with a single Dense layer, which has an input dimension of 10 and an output dimension of 10.

python
model = Sequential()
model.add(Dense(10, input_dim=10, activation='relu'))

Step 3: Initialize Weights

Before setting the weights, we must ensure their dimensions match those of the model. For a Dense layer, weights are stored as a (input_dim, output_dim) array, and biases as a (output_dim,) array. Let's initialize some random weights and biases.

python
weights = np.random.rand(10, 10)  # Corresponding to input_dim and output_dim
biases = np.random.rand(10)       # Corresponding to output_dim

Step 4: Set Weights

Now, we can use the initialized weights and biases to set the layer's weights. In Keras, this is achieved using the set_weights method, which accepts a list containing the weights and biases as NumPy arrays.

python
model.layers[0].set_weights([weights, biases])

Step 5: Verify Weights

To confirm the weights are correctly set, we can use the get_weights method to retrieve the current layer's weights and verify they match what we set.

python
current_weights, current_biases = model.layers[0].get_weights()
print("Are weights the same? ", np.array_equal(weights, current_weights))
print("Are biases the same? ", np.array_equal(biases, current_biases))

This completes the process of setting model weights in Keras using NumPy arrays. With this approach, you can easily import pre-trained weights or fine-tune the model.