Classification using local dataset (images)

[ggous]

Hello, I am trying to accomplish a classification training. I have tif images in various folders and a csv which refers to every image and what kind of label it has.

After successfully have done the training using :

ClassificationModelConfig
SolverConfig
ClassificationLearnerConfig

and a CustomDataConfig, something like:

class CustomDataConfig(DataConfig):
    root_dir: Optional[str] = None
    
    def build(self, df, valid_split, test_split):
        # Shuffle the dataframe before splitting
        df = df.sample(frac=1).reset_index(drop=True)
       .....
      # Create datasets
        train_ds = CustomTIFDataset(
            df=df_train,
            root_dir=self.root_dir,
            transform=data_augmentation_transform,
        )

        valid_ds = CustomTIFDataset(
            df=df_valid,
            root_dir=self.root_dir,
            transform=None, 
        )

and

 learner = ClassificationLearner(

        cfg=learner_config,
        output_dir='data/output',
        train_ds=train_ds,
        valid_ds=valid_ds
    )

I want to proceed to predictions.

So, I am trying something like:

predictions = learner.predict_dataset(test_ds,
                                      raw_out=True,          
                                      numpy_out=True,   
                                      progress_bar=True)

pred_labels = ChipClassificationLabels.from_predictions(
    test_ds,
    predictions)

scores = pred_labels.get_scores()

but the scores is an empty list.

I am not sure if I have to use ChipClassificationLabels though. I want a Classifications Labels, probably not Chip but I can't find something else!

[AdeelH]

You should be passing test_ds.windows to ChipClassificationLabels.from_predictions() and not test_ds.

pred_labels = ChipClassificationLabels.from_predictions(
    test_ds.windows,
    predictions)

[ggous]

Hello,

the thing is that I have tif images in folders and I just load them with rasterio. So, train_ds, test_ds do not have a windows method.

I tried to write one like:

    def get_windows(self):
        windows = []
        for idx in range(len(self)):
            image, _ = self[idx]
            window_size = (32, 32)
            for y in range(0, image.shape[0] - window_size[0] + 1, window_size[0]):
                for x in range(0, image.shape[1] - window_size[1] + 1, window_size[1]):
                    box = Box(x, y, x + window_size[1], y + window_size[0]) 
                    windows.append(box)
        return windows

My images have size (64, 64). Whatever size I am trying with window_size = (32, 32), when i run train_ds.get_windows() it returns an empty list.

[AdeelH]

The prediction code above is designed to work with a SlidingWindowGeoDataset (as shown in this tutorial for example). The idea is to make individual predictions for all the windows in the full image and then stitch them together using Labels.from_predictions(). If your images are already small and you do not want to split them further into smaller chips, you can set a size and stride (e.g. set both to 64) such that only one window is generated.

You can also skip ChipClassificationLabels.from_predictions() and just access the predictions directly. predictions in your code above is just an iterator, so you can do list(predictions) or loop over it to access the individual predictions and then save them however you like.

Hope this helps. Please provide more details if I have misunderstood what you're trying to do.

[ggous]

Thanks! I thought I have to use ChipClassificationLabels.