frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix LBFGS warning convert a tensor with requires_grad=True to a scalar (#160389)

Browse Source 
Fixes #160197

## Test Result

```python
In [1]: import warnings
   ...: warnings.simplefilter('error')
   ...: import torch
   ...: print(torch.__version__)
   ...: a, b = torch.rand((2, 32, 32))
   ...: a.requires_grad_()
   ...: optimizer = torch.optim.LBFGS([a])
   ...: loss_fn = lambda x, y: (x-y).pow(2).mean()
   ...:
   ...: def closure():
   ...:     optimizer.zero_grad()
   ...:     loss = loss_fn(a, b)
   ...:     loss.backward()
   ...:     return loss
   ...:
   ...: for i in range(100):
   ...:     optimizer.step(closure)
   ...:     print(i, loss_fn(a, b))
   ...:
2.9.0a0+gitf33f3f8
0 tensor(5.8066e-11, grad_fn=<MeanBackward0>)
1 tensor(5.8066e-11, grad_fn=<MeanBackward0>)
2 tensor(5.8066e-11, grad_fn=<MeanBackward0>)
3 tensor(5.8066e-11, grad_fn=<MeanBackward0>)
4 tensor(5.8066e-11, grad_fn=<MeanBackward0>)
5 tensor(5.8066e-11, grad_fn=<MeanBackward0>)
6 tensor(5.8066e-11, grad_fn=<MeanBackward0>)
7 tensor(5.8066e-11, grad_fn=<MeanBackward0>)
8 tensor(5.8066e-11, grad_fn=<MeanBackward0>)
9 tensor(5.8066e-11, grad_fn=<MeanBackward0>)
10 tensor(5.8066e-11, grad_fn=<MeanBackward0>)

...

```

```bash
pytest test/test_optim.py -vv

...
test/test_optim.py::TestOptimRenewedCUDA::test_tensor_lr_num_dim_2_NAdam_cuda_float32 PASSED [2.7192s]                                                                                                                                           [ 99%]
test/test_optim.py::TestOptimRenewedCUDA::test_tensor_lr_num_dim_2_RAdam_cuda_float32 PASSED [2.5370s]                                                                                                                                           [ 99%]
test/test_optim.py::TestOptimRenewedCUDA::test_tensor_lr_num_dim_2_RMSprop_cuda_float32 PASSED [2.0190s]                                                                                                                                         [ 99%]
test/test_optim.py::TestOptimRenewedCUDA::test_tensor_lr_num_dim_2_Rprop_cuda_float32 PASSED [1.8554s]                                                                                                                                           [ 99%]
test/test_optim.py::TestOptimRenewedCUDA::test_tensor_lr_num_dim_2_SGD_cuda_float32 PASSED [2.0433s]                                                                                                                                             [ 99%]
test/test_optim.py::TestOptimRenewedCUDA::test_tensor_lr_num_dim_2_SparseAdam_cuda_float32 PASSED [1.1788s]                                                                                                                                      [100%]

================== 1471 passed, 242 skipped in 2440.52s (0:40:40) ============
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160389
Approved by: https://github.com/janeyx99

Co-authored-by: albanD <desmaison.alban@gmail.com>
This commit is contained in:
Zesheng Zong
2025-08-26 03:07:47 +00:00
committed by PyTorch MergeBot
parent e34b6a0103
commit 8c442e4fd3
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
torch/optim/lbfgs.py
View File

@ -454,7 +454,8 @@ class LBFGS(Optimizer):
# the reason we do this: in a stochastic setting,
# no use to re-evaluate that function here
with torch.enable_grad():
loss = float(closure())
loss = closure()
loss = float(loss)
flat_grad = self._gather_flat_grad()
opt_cond = flat_grad.abs().max() <= tolerance_grad
ls_func_evals = 1