Manually perform predictions using the model weights and biases (Keras)

Question

I'd like to manually predict data (regression task) using the model weights and biases in Keras (i.e., Z = f(WX+B), where Z is the output of a layer, W is the weight matrix, X is the input vector, and B is the bias vector).

My model architecture is as follows:

model = Sequential()
model.add(Dense(207, input_dim = 173, activation = 'relu'))
model.add(Dense(207, activation = 'relu'))
model.add(Dense(207, activation = 'relu'))
model.add(Dense(240))

I'm also done with training the model.

I have already tried extracting the weights and biases using the following:

layer_num = 1
layer_weights = model.layers[layer_num-1].get_weights()[0]
layer_biases  = model.layers[layer_num-1].get_weights()[1]

where layer_num is the layer number (layer_num runs from 1 to 4 for my model). However, I can't figure out how to do the Z = f(WX+B) manually for my test set x_test.

Using this manual method, I'd like to compare its results and model.predict, as well as get the prediction computation time per instance/row of x_test (but I suppose I can figure out this last part on my own).

Edit 1: I mentioned that I can figure out the computation time per instance/row on my own, but if you could help me with this also, I'd really appreciate it.

Answer 1

For a dense layer, when you call the get_weights method, you will get a list, whose first element is the weight matrix W and second element is the bias b:

W, b = dense_layer.get_weights()

Suppose you have a test data matrix X with shape [batch, 173]. Then you can write a code snippet like below to get all the weights and biases:

Ws, bs = [], []
for layer in model.layers:
    W, b = layer.get_weights()
    Ws.append(W)
    bs.append(b)

With them at hand, we can define our function:

def relu(x):
    return np.where(x >= 0, x, 0)


def identity(x):
    return x


def predict(Ws, bs, activations, X):
    Z = X
    for W, b, activation in zip(Ws, bs, activations):
        Z = Z @ W + b
        Z = activation(Z)
    return Z

Then you can use the predict function to manually predict for test X:

Y = predict(Ws, bs, (relu, relu, relu, identity), x_test)