RandCropByPosNegLabeld AttributeError: 'list' object has no attribute 'keys'

[OeslleLucena]

Running a minimal example for a patch-based segmentation based on the spleen Lightning segmentation tutorial from here.
However, the returned training batch is a list instead of a dict. The validation bit works with no error.
Here is a print of the error:

| Name      | Type              | Params
------------------------------------------------
0 | net       | Generic_UNet      | 13.9 M
1 | criterion | BCEWithLogitsLoss | 0     
------------------------------------------------
13.9 M    Trainable params
0         Non-trainable params
13.9 M    Total params
/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/utilities/distributed.py:49: UserWarning: The dataloader, val dataloader 0, does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` (try 8 which is the number of cpus on this machine) in the `DataLoader` init to improve performance.
  warnings.warn(*args, **kwargs)
Validation sanity check:   0%|          | 0/2 [00:00<?, ?it/s]/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/monai/data/utils.py:300: UserWarning: Modifying image pixdim from [1.25 1.25 1.25 1.  ] to [  1.25         1.25         1.25       170.76592166]
  warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}")
/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/monai/data/utils.py:300: UserWarning: Modifying image pixdim from [1.25 1.25 1.25 1.  ] to [  1.25         1.25         1.25       170.76592166]
  warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}")
outputs torch.Size([1, 73, 145, 174, 145])
Validation sanity check:  50%|█████     | 1/2 [00:12<00:12, 12.25s/it]outputs torch.Size([1, 73, 145, 174, 145])
Epoch 0:   0%|          | 0/84 [00:00<?, ?it/s] /home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/utilities/distributed.py:49: UserWarning: The dataloader, train dataloader, does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` (try 8 which is the number of cpus on this machine) in the `DataLoader` init to improve performance.
  warnings.warn(*args, **kwargs)
/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/monai/data/utils.py:300: UserWarning: Modifying image pixdim from [1.25 1.25 1.25 1.  ] to [  1.25         1.25         1.25       170.76592166]
  warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}")
/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/monai/data/utils.py:300: UserWarning: Modifying image pixdim from [1.25 1.25 1.25 1.  ] to [  1.25         1.25         1.25       170.76592166]
  warnings.warn(f"Modifying image pixdim from {pixdim} to {norm}")
Epoch 0:   0%|          | 0/84 [00:04<?, ?it/s]
Traceback (most recent call last):
  File "/home/ol18/Codes/TractUncertainty/segmentation_sh_torchIO_PIL MONAI.py", line 378, in <module>
    trainer.fit(model) #tune(model)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/trainer/trainer.py", line 509, in fit
    results = self.accelerator_backend.train()
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/accelerators/accelerator.py", line 57, in train
    return self.train_or_test()
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/accelerators/accelerator.py", line 74, in train_or_test
    results = self.trainer.train()
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/trainer/trainer.py", line 560, in train
    self.train_loop.run_training_epoch()
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/trainer/training_loop.py", line 534, in run_training_epoch
    batch_output = self.run_training_batch(batch, batch_idx, dataloader_idx)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/trainer/training_loop.py", line 692, in run_training_batch
    self.optimizer_step(optimizer, opt_idx, batch_idx, train_step_and_backward_closure)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/trainer/training_loop.py", line 475, in optimizer_step
    using_lbfgs=is_lbfgs,
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/core/lightning.py", line 1264, in optimizer_step
    optimizer.step(closure=optimizer_closure)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/core/optimizer.py", line 286, in step
    self.__optimizer_step(*args, closure=closure, profiler_name=profiler_name, **kwargs)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/core/optimizer.py", line 144, in __optimizer_step
    optimizer.step(closure=closure, *args, **kwargs)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/torch/optim/lr_scheduler.py", line 67, in wrapper
    return wrapped(*args, **kwargs)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/torch/autograd/grad_mode.py", line 15, in decorate_context
    return func(*args, **kwargs)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/torch/optim/rmsprop.py", line 66, in step
    loss = closure()
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/trainer/training_loop.py", line 687, in train_step_and_backward_closure
    self.trainer.hiddens
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/trainer/training_loop.py", line 780, in training_step_and_backward
    result = self.training_step(split_batch, batch_idx, opt_idx, hiddens)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/trainer/training_loop.py", line 301, in training_step
    training_step_output = self.trainer.accelerator_backend.training_step(args)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py", line 71, in training_step
    return self._step(self.trainer.model.training_step, args)
  File "/home/ol18/miniconda3/envs/tract_uncertainty/lib/python3.6/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py", line 66, in _step
    output = model_step(*args)`` 
  File "/home/ol18/Codes/TractUncertainty/segmentation_sh_torchIO_PIL MONAI.py", line 224, in training_step
    print('train',batch.keys())
AttributeError: 'list' object has no attribute 'keys'

Below is shown how the transforms, training/validation steps, and loaders are coded.

    def prepare_data(self):
        data_path = Path(self.hparams.data_path)
        csv_path = Path(self.hparams.csv_path)

        if self.hparams.norm:
            input_data = 'input_norm'
        else:
            input_data = 'input'

        subjects_folders = [data_path / input_data,
                            data_path / 'label',
                            data_path / 'mask']
        subjects_prefixes = ['dwish', 'tractsmask', 'dwimask']

        fold_file = csv_path / f'fold{self.hparams.fold}.csv'

        df = pd.read_csv(fold_file, names=['SUBJECTS', 'SET'])
        subjects_names = {x: list(df[df.SET == x].SUBJECTS)
                          for x in [TRAINING, VALIDATION, INFERENCE]
                          }

        subjects_list = {x: monai_data_dicts(subjects_folders,
                                             subjects_prefixes,
                                             subjects_names[x])
                         for x in [TRAINING, VALIDATION, INFERENCE]
                         }

        # set deterministic training for reproducibility
        set_determinism(seed=0)

        # define the data transforms
        train_transforms = Compose(
            [
                LoadImaged(keys=["image", "label"]),
                Orientationd(keys=["image", "label"], axcodes="RAS"),
                # randomly crop out patch samples from big image based on pos / neg ratio
                # the image centers of negative samples must be in valid image area
                RandCropByPosNegLabeld(
                    keys=["image", "label"],
                    label_key="label",
                    spatial_size=(32, 32, 32),
                    num_samples=4,
                ),
                # DataStatsd(keys=['image', 'label'], data_value=False),
                ToTensord(keys=["image", "label"])
            ]
        )
        val_transforms = Compose(
            [
                LoadImaged(keys=["image", "label"]),
                Orientationd(keys=["image", "label"], axcodes="RAS"),
                ToTensord(keys=["image", "label"])
            ]
        )

        # we use cached datasets - these are 10x faster than regular datasets
        self.train_ds = Dataset(
            data=subjects_list[TRAINING], transform=train_transforms,
        )
        self.val_ds = Dataset(
            data=subjects_list[VALIDATION], transform=val_transforms,
        )

     def training_step(self, batch, batch_idx):
        print('train',batch.keys())
        inputs = batch['image']
        labels = batch['label']
        outputs = self(inputs)

        loss = self.criterion(outputs, labels)
        dice = dice_score(torch.sigmoid(outputs), labels)

        # Calling self.log will surface up scalars for you in TensorBoard
        self.log('train_loss', loss, on_step=True, on_epoch=True, prog_bar=True, logger=True)
        self.log('train_dice', dice, on_step=True, on_epoch=True, prog_bar=True, logger=True)
        return loss

    def validation_step(self, batch, batch_idx):
        inputs = batch['image']
        labels = batch['label']
        roi_size = (32, 32, 32)
        sw_batch_size = 4
        outputs = sliding_window_inference(inputs, roi_size, sw_batch_size, self.forward)
        print('outputs', outputs.shape)
        loss = self.criterion(outputs, labels)
        dice = dice_score(torch.sigmoid(outputs), labels)

        self.log_image_tensorboard(labels, torch.sigmoid(outputs) >= 0.5, 'val_images')

        # Calling self.log will surface up scalars for you in TensorBoard
        self.log('val_loss', loss, on_epoch=True, prog_bar=True, logger=True)
        self.log('val_dice', dice, on_epoch=True, prog_bar=True, logger=True)
        return loss

    def train_dataloader(self):
        train_loader = DataLoader(self.train_ds,
                                  batch_size=1, #self.hparams.batch_size,
                                  num_workers=2,
                                  pin_memory=True
                                  )
        return train_loader

    def val_dataloader(self):
        val_loader = DataLoader(self.val_ds,
                                  batch_size=1,#self.hparams.batch_size,
                                  num_workers=2,
                                  pin_memory=True
                                )

[rijobro]

Hi, what makes you think this is related to RandCropByPosNegLabeld?

This part of the error:

output = model_step(*args)`` 
  File "/home/ol18/Codes/TractUncertainty/segmentation_sh_torchIO_PIL MONAI.py", line 224, in training_step
    print('train',batch.keys())
AttributeError: 'list' object has no attribute 'keys'

implies to me that it's trying to use a dictionary-based transform on a list. I.e., subjects_list[TRAINING] looks to be a list as opposed to a dictionary. is that correct?

[OeslleLucena]

Hi @rijobr! You're right, I just don't understand why this is happening. I swap subjects_list[TRAINING] for subjects_list[VALIDATION] to see if the problem was with subjects_list[TRAINING] but the error is the same.
Any idea what could potentially be?

[rijobro]

In MONAI, there are two ways of doing transforms. You can either do them via arrays or dictionaries. An example of the array transforms is given in this notebook, whereas an example of dictionary transforms is given in this one.

You are opting for the latter, which means that subjects_list[TRAINING] should have the form:

subjects_list[TRAINING] = []
for i, s in zip(image_filenames, segementation_filenames):
    subjects_list[TRAINING].append({
        "image": i,
        "label": s,
    })

I can't see the definition of monai_data_dicts, so I can't tell what the structure of your subjects_list[TRAINING] is. At any rate, you need to make sure that subjects_list[TRAINING] is a list of dictionaries, each one containing the keys image and label.

[OeslleLucena]

I am based on that notebook, My subject lists are in the following format, so I do not see the error:

[{'image': '/media/ol18/Seagate Expansion Drive/Datasets/TractSeg_Data/HCP_105_monai/input_norm/dwish622236.nii.gz', ...]

[OeslleLucena]

And they contain both image and label.

[rijobro]

Ok thanks for verifying!
In that case, I can't see anything obviously wrong in your code snippet. However, your code snippet isn't the whole script, so it's hard to say for sure. I think you'll want to check to see where your code diverges from the spleen segmentation example. Clearly, we expect batch to be of type dict as it arrives into training_step, but in your case it seems to be list for some reason.

I just double-checked with the tutorial, by putting print(type(batch)) at the top of training_step and it returned dict as expected. This makes me think (I could be wrong!) that it's more likely to be a small error in your code as opposed to a MONAI bug.

If you still can't see the different between your code and the tutorial, then perhaps start again with the tutorial and gradually change it so that it matches your needs?

[OeslleLucena]

@rijobro I had a bug in the subjects_list[TRAINING] dict creation. Once I sorted that out, the code worked fine.
Thank you very much for helping me out.