Error in lstm_backward

Course Q&A Deep Learning Specialization Sequence Models
1

{moderator edit - solution code removed}

ERROR is :
ValueError Traceback (most recent call last)
in
18
19 da_tmp = np.random.randn(5, 10, 4)
—> 20 gradients_tmp = lstm_backward(da_tmp, caches_tmp)
21
22 print(“gradients["dx"][1][2] =”, gradients_tmp[“dx”][1][2])

in lstm_backward(da, caches)
56
57 # Fix the dimensions of dx[:,:,t] to match gradients["dxt"]
—> 58 dx[:, : ,t] = gradients[“dxt”]
59
60 dWf += gradients[“dWf”]

ValueError: could not broadcast input array from shape (5,3) into shape (3,10)

2

Hi @AkibButt

In your code, gradients["dxt"] has a shape of (5, 3) and dx[:, :, t] expects a shape of (3, 10). This is causing a dimension mismatch.

Make sure that the input and the forward/backward pass dimensions match, especially in the initialization of dx to match gradients["dxt"]. You can also print the shapes and other intermediate values before assignment to find the issue.

Hope it helps!

3

The code you show looks correct. Maybe the problem is actually in lstm_cell_backward. It looks like the gradient for dxt is the wrong shape. So how could that happen and why didn’t the test case for lstm_cell_backward catch it?

I added some print statements in both functions and here’s what I see when I run that test case:

dxt.shape (3, 10)
shape of dx[:,:,t] (3, 10)
dxt.shape (3, 10)
shape of dx[:,:,t] (3, 10)
dxt.shape (3, 10)
shape of dx[:,:,t] (3, 10)
dxt.shape (3, 10)
shape of dx[:,:,t] (3, 10)