Error training model - band regex?

[graceebc9]

Hello!

I've got a custom datamodule for Landcover / Modis / Sentinel data. The data module works fine when called directly (I can plot the 1 mask 4 channels) by sampling with a dataloader.

Issue comes when trying to run this datamodule with a semantic segmentation binary task - there's an issue with the bands within the geo.py raster dataset - ' no such group' . I've looked at the source code in datasets/geo.py but I'm not clear how to solve.

It seems to be some kind of issue with the band and the regex - we seem to match the date ok, but potentially fail with the band. However I copied the form of the band regex from the torchgeo sentinel2class.

from torchgeo.datasets import Sentinel2

class Sentinel2(Sentinel2):
    filename_glob = '*B03.tif'
    filename_regex = '^(?P<date>\d{6})\S{4}(?P<band>B[018][\dA]).tif$'
    date_format = '%Y%m'
    all_bands = ['B03', 'B08', 'B11']

def main():

    datamodule = MODISJDLandcoverSimpleDataModule(
      modis_root_dir="MODIS/",
      landcover_root_dir="landcover/Classified/",
      sentinel_root_dir ='sentinel/',
      patch_size=250,
      batch_size=10,
      length=10,
      num_workers=0,
      one_hot_encode=False,
      balance_samples=False,
      burn_prop = 0, 
      grid_sampler = False,
      units = Units.PIXELS
)

    # ignore_zeros=True corresponds to ignoring the background class
    # in metrics evaluation
    model = BinarySemanticSegmentationTask(
        segmentation_model="unet",
        encoder_name="resnet18",
        encoder_weights=None, #"imagenet",
        in_channels=4,
        num_filters=64,
        num_classes=2,
        loss="jaccard",
        # tversky_alpha=0.7,
        # tversky_beta=0.3,
        # tversky_gamma=1.0,
        learning_rate=0.1,
        ignore_zeros=False,
        learning_rate_schedule_patience=5,
    )

    trainer = Trainer(gpus=1, fast_dev_run=True)


    # this is used when automatically finding the learning rate
    trainer.tune(
        model, datamodule
    )  
    trainer.fit(model, datamodule)


if __name__ == "__main__":

    # set random seed for reproducibility
    pl.seed_everything(0)

    # TRAIN
    main()

Global seed set to 0
GPU available: True, used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
Running in fast_dev_run mode: will run a full train, val, test and prediction loop using 1 batch(es).
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]

  | Name          | Type             | Params
---------------------------------------------------
0 | model         | Unet             | 14.3 M
1 | loss          | JaccardLoss      | 0     
2 | train_metrics | MetricCollection | 0     
3 | val_metrics   | MetricCollection | 0     
4 | test_metrics  | MetricCollection | 0     
---------------------------------------------------
14.3 M    Trainable params
0         Non-trainable params
14.3 M    Total params
57.325    Total estimated model params size (MB)
/usr/local/lib/python3.7/dist-packages/pytorch_lightning/trainer/data_loading.py:433: UserWarning: The number of training samples (1) is smaller than the logging interval Trainer(log_every_n_steps=50). Set a lower value for log_every_n_steps if you want to see logs for the training epoch.
  f"The number of training samples ({self.num_training_batches}) is smaller than the logging interval"
Epoch 0: 0%
0/2 [00:00<?, ?it/s]
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
[<ipython-input-33-74529b3a24c6>](https://localhost:8080/#) in <module>()
     72 
     73     # TRAIN
---> 74     main()

26 frames
[/usr/local/lib/python3.7/dist-packages/torchgeo/datasets/geo.py](https://localhost:8080/#) in __getitem__(self, query)
    415                     if match:
    416                         if "date" in match.groupdict():
--> 417                             start = match.start("band")
    418                             end = match.end("band")
    419                             filename = filename[:start] + band + filename[end:]

IndexError: no such group

[graceebc9]

format of name of files is '201808_21_B03.tif'

[adamjstewart]

This one took me a long time to figure out. You need to use a raw string (filename_regex = r'...') or replace all \ with \\.

[graceebc9]

Thank you very much!!!!!! That solved that but now I'm getting the error:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
[<ipython-input-62-d2a0032e8c00>](https://localhost:8080/#) in <module>()
     70 
     71     # TRAIN
---> 72     main()

39 frames
[/usr/local/lib/python3.7/dist-packages/segmentation_models_pytorch/unet/decoder.py](https://localhost:8080/#) in forward(self, x, skip)
     36         x = F.interpolate(x, scale_factor=2, mode="nearest")
     37         if skip is not None:
---> 38             x = torch.cat([x, skip], dim=1)
     39             x = self.attention1(x)
     40         x = self.conv1(x)

RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 64 but got size 63 for tensor number 1 in the list.

[adamjstewart]

This seems related to the following bug reports. Basically, the UNet that comes with SMP requires images with patch_size divisible by 32. Can you try switching from 250 to 256 and see if that solves your issue?

[graceebc9]

yes that worked, thanks so much!!!!