loss : nan when train custom data set

[rezha130]

Hi @BelBES

I tried several batch-size from 8,16,32,64,128,256..but always end with loss : nan in every epoch when training my custom data set.

python train.py --data-path datatrain --test-init True --test-epoch 10 --output-dir snapshot --abc 0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz:/. --batch-size 8

Test phase
acc: 0.0000; avg_ed: 0.0000: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 18.10it/s]
acc: 0.0	acc_best: 0; avg_ed: 18.428571428571427
epoch: 0; iter: 1998; lr: 1.0000000000000002e-06; loss_mean: nan; loss: nan: 100%|██████████████████████████████████████████████████████████████████████████████████████| 2000/2000 [00:42<00:00, 46.69it/s]
epoch: 1; iter: 3998; lr: 1.0000000000000004e-10; loss_mean: nan; loss: nan: 100%|██████████████████████████████████████████████████████████████████████████████████████| 2000/2000 [00:42<00:00, 46.74it/s]
epoch: 2; iter: 5998; lr: 1.0000000000000006e-14; loss_mean: nan; loss: nan: 100%|██████████████████████████████████████████████████████████████████████████████████████| 2000/2000 [00:43<00:00, 45.84it/s]

I am using PyTorch 0.4, Python 3.6, GTX 1080 Ti and Ubuntu 16.04

Can you help me how to solve this problem?

Kindly Regards

[bes-dev]

Hi,

Can you provide small reproducer for this bug?

[rezha130]

Sorry @BelBES , would you please explain about "small reproducer"?

FYI, this is structure of my custom data set:

datatrain
---- data
-------- folderA/img_filename_0.jpg
...
-------- folderB/img_filename_1.jpg
---- desc.json

And, this is structure of my custom desc.json:

{
"abc": "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz:/.",
"train": [
{
"text": "text_on_image0"
"name": "folderA/img_filename_0.jpg"
},
...
{
"text": "text_on_image1"
"name": "folderB/img_filename_1.jpg"
}
],
"test": [
{
"text": "text_on_image3"
"name": "folderC/img_filename_3.jpg"
},
...
{
"text": "text_on_image4"
"name": "folderD/img_filename_4.jpg"
}
]
}

In text_data.py, i used this syntax in line 32:
img = cv2.imread(os.path.join(self.data_path, "data", name))

But still have same loss : nan issue. Please help.

[rezha130]

When i tried to debug using cuda = false (in CPU) on my dev laptop, this is the result of loss.data[0] that cause loss : nan

[0]:<Tensor>
_backward_hooks:None
_base:<Tensor, len() = 1>
_cdata:140460563260592
_grad:None
_grad_fn:None
_version:0
data:<Tensor>
device:device(type='cpu')
dtype:torch.float32
grad:None
grad_fn:None
is_cuda:False
is_leaf:True
is_sparse:False
layout:torch.strided
name:None
output_nr:0

Note: i set cuda = False in my CPU dev-laptop, but set cuda = True on my GPU server above.