def naive()

in fast_grad/gradient_funcs.py [0:0]

• 13 lines of code
• 3 McCabe index (conditional complexity)

def naive(model, X, y):
	"""
	Computes the predictions in a full-batch fasion,
	then call backward on the individual losses
	"""
	grad_list = []
	logits, _, _ = model.forward(X)
	N = X.shape[0]
	for n in range(N):
		model.zero_grad()
		loss = F.binary_cross_entropy_with_logits(logits[n], y[n].view(-1,))
		loss.backward(retain_graph=True)

		grad_list.append(list([p.grad.clone() for p in model.parameters()]))

	grads = []
	for p_id in range(len(list(model.parameters()))):
		grads.append(torch.cat([grad_list[n][p_id].unsqueeze(0) for n in range(N)]))

	return grads