Add Few-Shot Training Subset Creation

[RuslanAntjuschin]

This PR adds a python method to create few-shot training subsets for BirdSet datasets. There is a small usage example in the added notebook.

[lurauch]

thank you!

a few remarks:

• the object orientation is IMO not necessary, you could just have 2 different options ("lenient" and "strict") and then define these parameters in the code for example
• i tried your code, there a few steps missing that would be nice to have:
  • defining the path from where to load the dataset
  • defining the path and automatically saving the dataset directly to the respective folder
  • removing some columns
  • one-hot-encode the labels

e.g., you could add this to the code:

 def one_hot_encode_batch(batch, num_classes):
      """
      Converts integer class labels in a batch to one-hot encoded tensors.
      """
      label_list = batch["labels"]
      batch_size = len(label_list)
      one_hot = torch.zeros((batch_size, num_classes), dtype=torch.float32)
      for i, label in enumerate(label_list):
          one_hot[i, label] = 1
      return {"labels": one_hot}
      dataset = DatasetDict({"train": Dataset.from_list(selected_samples), "test": dataset["test_5s"]})

print("Selecting relevant columns and renaming...", flush=True)
columns_to_keep = ["filepath", "ebird_code_multilabel", "detected_events", "start_time", "end_time"]

dataset = DatasetDict({
split: dataset[split].select_columns(columns_to_keep).rename_column("ebird_code_multilabel", "labels")
for split in dataset.keys()
})

print("Applying one-hot encoding to labels...", flush=True)
dataset = dataset.map(lambda batch: one_hot_encode_batch(batch, NUM_CLASSES), batched=True)

print("Saving processed dataset to disk...", flush=True)
dataset.save_to_disk(SAVE_DIR)
print("Dataset saved to", SAVE_DIR, flush=True)

return dataset

[RuslanAntjuschin]

@lurauch I've added the requested changes.