Weight matrix dimension in TensorFlow

Hey @siddhesh_mane,
It’s all about the convention followed. The lecture videos follow a particular convention, and Tensorflow follows another, that’s it. You will find that the core concepts still remain the same.

In the lecture videos, the input has the dimensions `(n_x, m)`, the weight matrix has the dimensions `(n_y, n_x)`, where `n_y` is the output number of neurons. The computation that happens is W \ x, and the output has the dimensions `(n_y, m)`.

In tensorflow, the input has the dimensions `(m, n_x)`, the weight matrix has the dimensions `(n_x, n_y)`, once again where `n_y` is the output number of neurons. The computation that happens is x \ W, and the output has the dimensions `(m, n_y)`.

So, you see the concept remains the same, just the difference in the shape, which includes the difference in shapes of the outputs as well.

For verifying the last part, feel free to run the below piece of code

``````model = tf.keras.Sequential([
tfl.Dense(2, input_shape = (4,))
])

print("Shape of the weights:")
print(np.array(model.weights[0]).shape)

x = np.random.randn(10, 4)
print("Shape of the input:")
print(x.shape)

y = model(x)
print("Shape of the output:")
print(y.shape)
``````

I hope this helps.

Cheers,
Elemento

1 Like