My program was broken by AdamW and then my console became red!

This happened when I trained resnet18 with CIFAR10. If I use SGD, then the program can run normally. This is how I get AdamW:

parameters = {'params':model.parameters(), 'lr':learning_rate}
parameters['weight_decay'] = 1e-4
optimizer = optim.AdamW(**parameters)

Could you post a minimal and executable code snippet to reproduce the issue, please?

Thank you! I found the crux. I set the learning rate in AdamW with a tensor. Here is the code:

import torch
from torch.optim import SGD, AdamW
from torch.cuda import amp
from torch.cuda.amp.grad_scaler import GradScaler
from torchvision import models
from torch.nn import functional as F

model = models.resnet18().cuda()

scaler = GradScaler()
L = torch.nn.CrossEntropyLoss(reduction = 'mean').cuda()

parameters = {'params':model.parameters(), 'lr':0.002}
parameters['weight_decay'] = 2e-5
optimizer = AdamW(**parameters)                 #SGD works fine
for param_group in optimizer.param_groups:
    param_group['lr'] = torch.tensor(0.)

images = torch.randn(128, 3, 32, 32).cuda(non_blocking = True)
labels = F.one_hot(torch.randint(0, 9, (128,)), 10).type(torch.float32).cuda(non_blocking = True)
with amp.autocast():
    output = model(images)
    loss = L(output, labels)

Hi @jack_S,

it looks a bit surprising to me that SGD works with Tensor learning rate considering the type hint of pytorch/ at fc63d710fee323eee8b135fd193ee37e9f06ed55 · pytorch/pytorch · GitHub where lr’s type hint is float.

One way to let AdamW with Tensor lr would be to pass fused=False, foreach=False to AdamW constructor.

Would you mind telling me why you want to have lr in Tensor?

Well, it’s not very important for me to assign lr with a tensor. I can turn to float in other experiments if necessary. I did it just because I want to see what will happen if I set the lr casually. For example, if I want to set a random learning schedule for my training, I may code like below:

total_epochs = 200
lr_schedule = torch.normal(mean = 0.05, std = 0.01, size = (total_epochs, )).abs()
optimizer = AdamW(model.parameters(), 0.001)
for i in range(0, total_epochs):
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr_schedule[i]   

So this introduces tensor in lr.