Some problems occurred in predict.py.

[DL-Alva]

Hi, sorry for troubling you. But when I run predict.py, it remind AttributeError: 'numpy.ndarray' object has no attribute 'unsqueeze':

Traceback (most recent call last):
File "origin_predict.py", line 135, in
device=device)
File "origin_predict.py", line 32, in predict_img
img = img.unsqueeze(0)
AttributeError: 'numpy.ndarray' object has no attribute 'unsqueeze'

Do you know how to deal with it? Thank you so much.

[milesial]

Hi, can you try after pulling the last commit on master ?

[DL-Alva]

Thank you for your reply. I try to use torch.from_numpy(img).unsqeeze(0），to transfer the numpy type of img to that of torch.

[glesx]

I'm getting the following error:

python predict.py -i validation.jpg --viz --no-save

Traceback (most recent call last):
  File "predict.py", line 132, in <module>
    device=device)
  File "predict.py", line 25, in predict_img
    ds = BasicDataset('', '', scale=scale_factor)
  File "C:\Users\glesx\Documents\GitHub\Pytorch-UNet\utils\dataset.py", line 18, in __init__
    self.ids = [splitext(file)[0] for file in listdir(imgs_dir)
FileNotFoundError: [WinError 3] The system cannot find the path specified: ''

When I add a path, I get this error instead:

Traceback (most recent call last):
  File "predict.py", line 132, in <module>
    device=device)
  File "predict.py", line 44, in predict_img
    transforms.Resize(full_img.shape[1]),
AttributeError: 'JpegImageFile' object has no attribute 'shape'

[cmlyldz]

I am also getting similar error
python predict.py -i austin.png --viz --no-save
`Reloaded modules: unet, unet.unet_model, unet.unet_parts, utils.data_vis, utils.dataset
Traceback (most recent call last):

File "D:\PW\courses\neuralNetworks\repo\Pytorch-UNet\predict.py", line 141, in
device=device)

File "D:\PW\courses\neuralNetworks\repo\Pytorch-UNet\predict.py", line 26, in predict_img
ds = BasicDataset('', '', scale=scale_factor)

File "D:\PW\courses\neuralNetworks\repo\Pytorch-UNet\utils\dataset.py", line 18, in init
self.ids = [splitext(file)[0] for file in listdir(imgs_dir)

FileNotFoundError: [WinError 3] The system cannot find the path specified: ''`

[milesial]

Hi, can you try again after pulling the last commit ?

[cmlyldz]

Commit is works for me. But there is another problem.

I passing this error with - transforms.Resize(np.array(full_img).shape[1]), - but when I use this my predicted image is totally white.

I want to add my traning log, maybe there is something wrong with traning I am not sure about it.

[milesial]

You validation DICE is > 1 and your loss is negative so there is a problem with your masks. If you have 2 classes, make sure the mask values are 0s or 1s.

[cmlyldz]

I have only one class. As an example my train image;

And my mask;

[milesial]

Verify that after loading your mask, the tensor contains only 0s and 1s

[cmlyldz]

It has only true and falses, I think it means 0 and 1;

[milesial]

can you print the resulting true mask tensor and the network output values ?

[cmlyldz]

Can you be more spesific? Where are they?

[milesial]

In the train loop

Pytorch-UNet/train.py

Line 81 in ebe3bbb

masks_pred = net(imgs)

print masks_pred and true_masks and make sure they are as expected

[cmlyldz]

True mask;

mask_pred;

How can I understand they are okay or not?

[milesial]

Please show some of the values of these tensors : min, max, type, ...

[cmlyldz]

I hope you mention about following images;

[milesial]

Yes that what I meant. The values of the true mask should be between [0, 1], as they are probabilities. Divide your true masks by 255 in your loader.

( I'm doing that in my loader, did you bypass that ?

Pytorch-UNet/utils/dataset.py

Line 40 in ebe3bbb

img_trans = img_trans / 255

)

[cmlyldz]

no its same. its not bypassed.

[cmlyldz]

I solve the problem, it is about .tif files.

I can reproduce that on Windows. Downgrading libtiff to 4.0.10 (conda install libtiff=4.0.10) works. Investigating it...

PS: it works fine on macOS and Linux, only Windows is failing.
[https://github.com/conda-forge/pillow-feedstock/issues/69]

After changing libtiff version it works properly.

[Janene-Li]

I'm getting the following error:

python predict.py -i validation.jpg --viz --no-save

Traceback (most recent call last):
  File "predict.py", line 132, in <module>
    device=device)
  File "predict.py", line 25, in predict_img
    ds = BasicDataset('', '', scale=scale_factor)
  File "C:\Users\glesx\Documents\GitHub\Pytorch-UNet\utils\dataset.py", line 18, in __init__
    self.ids = [splitext(file)[0] for file in listdir(imgs_dir)
FileNotFoundError: [WinError 3] The system cannot find the path specified: ''

When I add a path, I get this error instead:

Traceback (most recent call last):
  File "predict.py", line 132, in <module>
    device=device)
  File "predict.py", line 44, in predict_img
    transforms.Resize(full_img.shape[1]),
AttributeError: 'JpegImageFile' object has no attribute 'shape'

I have come to similar problem like this:

Traceback (most recent call last):
  File "predict.py", line 131, in <module>
    device=device)
  File "predict.py", line 43, in predict_img
    transforms.Resize(full_img.shape[2]),
AttributeError: 'PngImageFile' object has no attribute 'shape'

It seems like png picture cannot use '.shape[2]'.... I change it into '.width' and finally succesfully run the project~

[milesial]

@Janene-Li make sure you are using the latest version of Pillow