default_collate error while trying to quantize my pytorch model

[CostatoMarco]

Hi,
I'm trying to compress my model using this framework's implementation of post training quantization. Here is the code I'm using:

val_dataset = datasets.ImageFolder(path1)
dataset_loader = torch.utils.data.DataLoader(val_dataset)

def transform_fn(data_item):
    images,_=data_item
    print(type(images))
    
    return(images)

calibration_dataset = nncf.Dataset(dataset_loader, transform_fn)
quantized_model = nncf.quantize(model, calibration_dataset)
torch.save(model, path2)

I keep running in this error:

default_collate: batch must contain tensors, numpy arrays, numbers, dicts or lists; found <class 'PIL.Image.Image'>

on line:

quantized_model = nncf.quantize(model, calibration_dataset)

I'm new to pytorch and I have no idea on how to fix this. Any help?

I'm running on pytorch version 2.0.0

[alexsu52]

Hello @CostatoMarco,

I guess that images weren’t transformed to Tensors. The same error was discussed here link

I would recommend to check that your transform_fn is correct. The transformation function is a function that takes a sample from the dataset and returns data that can be passed to the model for inference (doc). Thus, you can use the following code to check it:

data_item = next(iter(dataset_loader))
output = model(transform_fn(data_item))

[CostatoMarco]

Hi! Thank you for your time.
I tried to implement my transformation function and to check it with the code you wrote:

val_dataset = datasets.ImageFolder(path1)
dataset_loader = torch.utils.data.DataLoader(val_dataset)

def transform_fn(data_item):
    images,_=data_item
    #print(type(images))
    numpy_image=np.array(images)
    cv2_image = cv2.cvtColor(numpy_image, cv2.COLOR_RGB2BGR)
    transformed_image = cv2.resize(cv2_image.astype(np.float32), (224, 224),
                                        interpolation=cv2.INTER_LINEAR)
    transformed_image = transformed_image.transpose(2, 0, 1)
    images = torch.from_numpy(transformed_image).unsqueeze(0)
    
    return(images)

data_item = next(iter(dataset_loader))
output = model(transform_fn(data_item))


calibration_dataset = nncf.Dataset(dataset_loader, transform_fn)
quantized_model = nncf.quantize(model, calibration_dataset)
torch.save(model, path2)

but I still get the same error, this time on line:

data_item = next(iter(dataset_loader))

TypeError: default_collate: batch must contain tensors, numpy arrays, numbers, dicts or lists; found <class 'PIL.Image.Image'>

Here is the full traceback :

Traceback (most recent call last):
  File "/home/marco/Desktop/Stage/Pruning/Pruning.py", line 68, in <module>
    data_item = next(iter(dataset_loader))
  File "/home/marco/.local/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 634, in __next__
    data = self._next_data()
  File "/home/marco/.local/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 678, in _next_data
    data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
  File "/home/marco/.local/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 54, in fetch
    return self.collate_fn(data)
  File "/home/marco/.local/lib/python3.8/site-packages/torch/utils/data/_utils/collate.py", line 264, in default_collate
    return collate(batch, collate_fn_map=default_collate_fn_map)
  File "/home/marco/.local/lib/python3.8/site-packages/torch/utils/data/_utils/collate.py", line 142, in collate
    return [collate(samples, collate_fn_map=collate_fn_map) for samples in transposed]  # Backwards compatibility.
  File "/home/marco/.local/lib/python3.8/site-packages/torch/utils/data/_utils/collate.py", line 142, in <listcomp>
    return [collate(samples, collate_fn_map=collate_fn_map) for samples in transposed]  # Backwards compatibility.
  File "/home/marco/.local/lib/python3.8/site-packages/torch/utils/data/_utils/collate.py", line 150, in collate
    raise TypeError(default_collate_err_msg_format.format(elem_type))
TypeError: default_collate: batch must contain tensors, numpy arrays, numbers, dicts or lists; found <class 'PIL.Image.Image'>

[vshampor]

This time it seems that the error is unrelated to NNCF. Looks like you have to specify a transform=... parameter when instantiating the ImageFolder as specified in the docs to maybe transform the PIL image to a torch.Tensor or something.

[CostatoMarco]

Yep, just implemented this and it seems to be working. Thank you so much for the help