05-Regression

[47482-pawansingh]

1. Quick clarification: In cell where we are plotting prediction from untrained model, we are generating new set of X which is between [-2, 2], whereas original X is between [-1, 1]. Not sure why we are plotting y_pretrained which is prediction of untrained model on original X, with respect to new X. As per me, correct method would be either plot with respect to original X or while creating y_pretrained, we should predict on new X. This suggestions does not change the answer and conclusion per se, but seems right thing to do. Correct code would be

for y_pretrain_idx in y_pretrain:
    # New X that ranges from -2 to 2 instead of -1 to 1
    X_new = torch.unsqueeze(torch.linspace(-2, 2, 100), dim=1)
        
#     plt.plot(X_new.numpy(), y_pretrain_idx.cpu().numpy(), 'r-', lw=1)
    plt.plot(X.cpu().numpy(), y_pretrain_idx.cpu().numpy(), 'r-', lw=1)

plt.scatter(X.cpu().numpy(), y.cpu().numpy(), label='data')
plt.axis('square')
# plt.axis((-1.1, 1.1, -1.1, 1.1));
y_combo = torch.stack(y_pretrain)
plt.plot(X_new.numpy(), y_combo.var(dim=0).cpu().numpy(), 'g', label='variance');
plt.legend()

2. In cell where we are plotting prediction distribution post plotting, variance multiplier is 10, when it should be 30.

Incorrect code:

plt.plot(X_new.cpu().numpy(), 10 * y_combo.var(dim=0).cpu().numpy(), 'g', label='30 × variance')

Correct code:

plt.plot(X_new.cpu().numpy(), 30 * y_combo.var(dim=0).cpu().numpy(), 'g', label='30 × variance')