Hi there, I’m trying to reimplement SCAN.
My code is here.
The loss didn’t change after each iteration. When I called backward to calculate the gradients, it happened to be NaN.
Do you guys have any suggestions?
Hi,
I’d try Automatic differentiation package - torch.autograd — PyTorch 1.8.0 documentation to which backward part emit nan first.
1 Like
It showed Function 'DivBackward0' returned nan values in its 0th output.
I’m trying to figure out but found nothing :<
@ptrblck Sir please give me a hand!
Oh, it’s a little bit hard to identify which layer.
nan can occur for some reasons but mainly it’s oftentimes 0/inf related maths.
For example, in SCAN code (SCAN/model.py at master · kuanghuei/SCAN · GitHub), nan and inf can happen in forward of l1norm and l2norm. So, I think it’s better to investigate where those bad values are generated, for example, by using forward/backward hooks.
I think one implementation of a backward hook with module name is as follows.
Also, forward hook with similar functionalities will be helpful.
# backward hook with module name
def get_backward_hook(module_name: str):
class BackwardHook:
name: str
def __init__(self, name):
self.name = name
def __call__(self, module, grad_input, grad_output):
for i, g_in in enumerate(grad_input):
print(module_name, torch.any(torch.isnan(g_in)))
if torch.any(torch.isnan(g_in)):
print(f"{module_name}'s {i}th input gradient is nan")
for i, g_out in enumerate(grad_output):
if torch.any(torch.isnan(g_out)):
print(f"{module_name}'s {i}th output gradient is nan")
return BackwardHook(module_name)
...
model = SomeCoolModel(...)
for name, module in model.named_modules():
module.register_full_backward_hook(get_backward_hook(name))
Hope this
2 Likes
Thank you for your generous response. I found that there is something wrong in this line of code but I can’t figure out. Could you please have a look at it?!
def forward(self, images: Tensor, captions: Tensor, cap_lens: Tensor) -> Tensor:
"""
Parameters
----------
images: (batch, regions, hidden_dim)
captions: (batch, cap_lenghts, hidden_dim)
cap_lengths: (batch)
Returns
-------
sim_scores: (batch cap, batch image)
"""
# Projecting the image/caption to embedding dimensions
img_embs = self.img_enc(images)
cap_embs = self.txt_enc(captions, cap_lens)
# -> img_embs: (batch, regions, hidden)
# -> cap_embs: (batch, max_seq_len, hidden)
# Compute similarity between each region and each token
# cos(u, v) = (u @ v) / (||u|| * ||v||)
# = (u / ||u||) @ (v / ||v||)
img_embs = F.normalize(img_embs, dim=-1)
cap_embs = F.normalize(cap_embs, dim=-1)
img_embs.unsqueeze_(0)
cap_embs.unsqueeze_(1)
# -> img_embs: (1, batch, regions, hidden)
# -> cap_embs: (batch, 1, max_seq_len, hidden)
# After normalizing: cos(u, v) = u @ v
sim_token_region = torch.matmul(cap_embs, img_embs.transpose(-1, -2))
# -> sim_token_region: (batch, batch, max_seq_len, regions)
att_score = self.lambda_softmax * sim_token_region
# -> att_score: (batch, batch, max_seq_len, regions)
# Create mask from caption length
# torch.arange(max_seq_len_SIZE).expand(batch_SIZE, max_seq_len_SIZE)
padding_mask = torch.arange(cap_embs.size(2)).expand(
cap_embs.size(0), cap_embs.size(2)
) >= cap_lens.unsqueeze(1)
# padding_mask: (batch, max_seq_len)
padding_mask = padding_mask.unsqueeze(-1).unsqueeze(1).to(cap_embs.device)
# -> padding_mask: (batch, 1, max_seq_len, 1)
# mask score of padding tokens
att_score.data.masked_fill_(padding_mask, -float('inf'))
# softmax along regions axis
attention_weights = F.softmax(att_score, dim=-1)
# -> attention_weights: (batch, batch, max_seq_len, regions)
# Calculate weighted sum of regions -> attended image vectors
attention = torch.matmul(attention_weights, img_embs)
# -> attention: (batch, batch, max_seq_len, hidden_dim)
# Calculate the importance of each attended image vector
# w.r.t each token of sentence
r = F.cosine_similarity(cap_embs, attention, dim=-1)
# -> r: (batch, batch, max_seq_len)
# Zeros similarity of padding tokens
r[r.isnan()] = 0
# Calculate similarity of each caption and each image by averaging all tokens in a caption.
sim_scores = r.sum(dim=-1) / cap_lens.view(-1, 1).to(cap_embs.device)
# -> sim_score: (batch, batch)
return sim_scores
Did you find that in forward computation? or backward?
Either way, I’m not sure but I’d check the values of img_embs and backward input values.
I met the ‘nan’ because of empty input tensor.
A simple example is :
impot torch.nn as nn
loss = nn.TripletMarginLoss(margin=1.0, p=2)
a = torch.randn(0, 100)
p = torch.randn(0, 100)
n = torch.randn(0, 100)
ret = loss(a, p, n) # the return value is `tensor(nan)`