dslighting 1.7.1__py3-none-any.whl → 1.7.6__py3-none-any.whl

mlebench/competitions/herbarium-2022-fgvc9/prepare_val.py

@@ -0,0 +1,213 @@
+import json
+import random
+import shutil
+from pathlib import Path
+
+import pandas as pd
+from sklearn.model_selection import train_test_split
+from tqdm import tqdm
+
+from mlebench.utils import get_logger
+
+logger = get_logger(__name__)
+
+
+def _organize_data_by_category(metadata: dict) -> dict:
+    """Organizes image and annotation data by category ID."""
+    annotations_images_by_category = {}
+    for annotation, image in list(zip(metadata["annotations"], metadata["images"])):
+        assert annotation["image_id"] == image["image_id"]
+        category_id = annotation["category_id"]
+        if category_id not in annotations_images_by_category:
+            annotations_images_by_category[category_id] = []
+        annotations_images_by_category[category_id].append(
+            {
+                "annotation": annotation,
+                "image": image,
+            }
+        )
+    return annotations_images_by_category
+
+
+def _split_data(data_by_category: dict, test_size: float, random_state: int) -> tuple[dict, dict]:
+    """Splits data for each category into train and test sets."""
+    train_annotations_images_by_category = {}
+    test_annotations_images_by_category = {}
+    for category_id, annotations_images in data_by_category.items():
+        # Each category needs to be in both train and test
+        train_annotations_images, test_annotations_images = train_test_split(
+            annotations_images, test_size=test_size, random_state=random_state
+        )
+        assert len(train_annotations_images) > 0 and len(test_annotations_images) > 0
+        train_annotations_images_by_category[category_id] = train_annotations_images
+        test_annotations_images_by_category[category_id] = test_annotations_images
+    return train_annotations_images_by_category, test_annotations_images_by_category
+
+
+def _process_train_set(
+    train_data: dict, base_metadata: dict, raw_path: Path, output_public_path: Path
+):
+    """Processes and writes the training set data, images, and metadata."""
+    logger.info(f"Processing train set for output: {output_public_path}")
+    new_train_metadata = base_metadata.copy()
+    new_train_metadata.update({"annotations": [], "images": []})
+    train_sample_count = sum(len(v) for v in train_data.values())
+
+    output_train_images_path = output_public_path / "train_images"
+
+    with tqdm(
+        desc=f"Creating train dataset in {output_public_path.name}",
+        total=train_sample_count,
+    ) as pbar:
+        for category_id, annotations_images in train_data.items():
+            category_subdir = f"{category_id // 100:03d}/{category_id % 100:02d}"
+            (output_train_images_path / category_subdir).mkdir(exist_ok=True, parents=True)
+            for idx, annotation_image in enumerate(annotations_images):
+                new_image_id = f"{category_id:05d}__{(idx + 1):03d}"
+                new_file_name = f"{category_subdir}/{new_image_id}.jpg"
+
+                new_annotation = annotation_image["annotation"].copy()
+                new_annotation["image_id"] = new_image_id
+                new_train_metadata["annotations"].append(new_annotation)
+
+                new_image = annotation_image["image"].copy()
+                new_image["image_id"] = new_image_id
+                new_image["file_name"] = new_file_name
+                new_train_metadata["images"].append(new_image)
+
+                src_path = raw_path / "train_images" / annotation_image["image"]["file_name"]
+                dst_path = output_train_images_path / new_file_name
+                shutil.copyfile(src=src_path, dst=dst_path)
+
+                pbar.update(1)
+
+    with open(output_public_path / "train_metadata.json", "w") as f:
+        json.dump(new_train_metadata, f, indent=4, sort_keys=True)
+
+    assert len(list(output_train_images_path.glob("**/*.jpg"))) == len(
+        new_train_metadata["images"]
+    )
+    assert len(new_train_metadata["annotations"]) == len(new_train_metadata["images"])
+
+
+def _process_test_set(
+    test_data: dict, raw_path: Path, output_public_path: Path, output_private_path: Path
+):
+    """Processes and writes the test set data, images, metadata, and private answers."""
+    logger.info(
+        f"Processing test set for outputs: {output_public_path} and {output_private_path}"
+    )
+    new_test_metadata = {"annotations": [], "images": []}
+    test_annotations_images = [item for sublist in test_data.values() for item in sublist]
+    random.Random(0).shuffle(test_annotations_images)
+
+    output_test_images_path = output_public_path / "test_images"
+
+    for idx, annotation_image in tqdm(
+        enumerate(test_annotations_images),
+        desc=f"Creating test dataset in {output_public_path.name}",
+        total=len(test_annotations_images),
+    ):
+        new_image_id = str(idx)
+        new_file_name = f"{idx // 1000:03d}/test-{idx:06d}.jpg"
+
+        new_annotation = annotation_image["annotation"].copy()
+        new_annotation["image_id"] = new_image_id
+        new_test_metadata["annotations"].append(new_annotation)
+
+        new_image = annotation_image["image"].copy()
+        new_image["image_id"] = new_image_id
+        new_image["file_name"] = new_file_name
+        new_test_metadata["images"].append(new_image)
+
+        src_path = raw_path / "train_images" / annotation_image["image"]["file_name"]
+        dst_path = output_test_images_path / new_file_name
+        dst_path.parent.mkdir(exist_ok=True, parents=True)
+        shutil.copyfile(src=src_path, dst=dst_path)
+
+    with open(output_public_path / "test_metadata.json", "w") as f:
+        json.dump(new_test_metadata["images"], f, indent=4, sort_keys=True)
+
+    answers_rows = [
+        {"Id": image["image_id"], "Predicted": annotation["category_id"]}
+        for image, annotation in zip(new_test_metadata["images"], new_test_metadata["annotations"])
+    ]
+    answers_df = pd.DataFrame(answers_rows)
+    answers_df.to_csv(output_private_path / "answers.csv", index=False)
+
+    sample_rows = [{"Id": image["image_id"], "Predicted": 42} for image in new_test_metadata["images"]]
+    sample_df = pd.DataFrame(sample_rows)
+    sample_df.to_csv(output_public_path / "sample_submission.csv", index=False)
+
+    assert len(list(output_test_images_path.glob("**/*.jpg"))) == len(new_test_metadata["images"])
+    assert len(new_test_metadata["annotations"]) == len(new_test_metadata["images"])
+    assert len(answers_df) == len(new_test_metadata["images"])
+    assert len(sample_df) == len(answers_df)
+
+
+def prepare(raw: Path, public: Path, private: Path):
+    """
+    Splits the raw data into public and private datasets with appropriate test/train splits.
+
+    `train_metadata.json` is the "table of contents" for our data, with the following structure:
+    (More details at https://www.kaggle.com/competitions/herbarium-2022-fgvc9/data)
+    ```
+    {
+        "annotations" : [annotation],
+        "categories" : [category],
+        "genera" : [genus]
+        "images" : [image],
+        "distances" : [distance],
+        "licenses" : [license],
+        "institutions" : [institution]
+    }
+    ```
+    - `images` and `annotations` are both N-length lists corresponding to the N samples.
+        We'll need to split each of these lists into train and test.
+    - The other fields are dataset-wide metadata that we don't need to touch.
+
+    Other notes:
+    - train/test splits need to occur per category (each category should be in both train and test).
+    - The `test_images` and `train_images` folders have nested subdirs to make it easier to browse
+        - `train_images` is structured as `{category_id[:3]}/{category_id[3:]}/{image_id}.jpg`
+        - `test_images` is structured as `{image_idx[:3]}/test-{image_idx}.jpg` (to not reveal the category)
+    - When we create the new splits, we re-assign image indices so that we don't give away labels based on the index
+        - train images are indexed within their own category
+        - test images follow a flat index after shuffling the categories
+    """
+    # Load raw data and organize it by category
+    with open(raw / "train_metadata.json") as f:
+        old_train_metadata = json.load(f)
+    annotations_images_by_category = _organize_data_by_category(old_train_metadata)
+
+    # --- 1. Create the original public/private datasets ---
+    # This first split creates the main train and test sets.
+    # The outputs in `public` and `private` will be identical to the original script.
+    logger.info("--- Creating original train/test split for 'public' and 'private' directories ---")
+    original_train_split, original_test_split = _split_data(
+        annotations_images_by_category, test_size=0.2, random_state=0
+    )
+
+    _process_train_set(original_train_split, old_train_metadata, raw, public)
+    _process_test_set(original_test_split, raw, public, private)
+    logger.info("Finished creating original 'public' and 'private' datasets.")
+
+    # --- 2. Create the validation datasets from the original training set ---
+    # This second split takes the `original_train_split` and splits it *again*
+    # to create a new, smaller training set and a validation set.
+    # The outputs are saved to new `public_val` and `private_val` directories.
+    logger.info("--- Creating validation split for 'public_val' and 'private_val' directories ---")
+    public_val = public.parent / "public_val"
+    private_val = private.parent / "private_val"
+    public_val.mkdir(exist_ok=True, parents=True)
+    private_val.mkdir(exist_ok=True, parents=True)
+
+    # The input for this split is the *train set* from the first split.
+    # We use the exact same splitting logic to ensure consistency.
+    train_val_split, test_val_split = _split_data(
+        original_train_split, test_size=0.2, random_state=0
+    )
+
+    _process_train_set(train_val_split, old_train_metadata, raw, public_val)
+    _process_test_set(test_val_split, raw, public_val, private_val)
+    logger.info("Finished creating validation 'public_val' and 'private_val' datasets.")

mlebench/competitions/histopathologic-cancer-detection/grade.py

@@ -0,0 +1,12 @@
+import numpy as np
+import pandas as pd
+from sklearn.metrics import roc_auc_score
+
+from mlebench.competitions.utils import prepare_for_auroc_metric
+
+
+def grade(submission: pd.DataFrame, answers: pd.DataFrame) -> float:
+    roc_auc_inputs = prepare_for_auroc_metric(
+        submission=submission, answers=answers, id_col="id", target_col="label"
+    )
+    return roc_auc_score(y_true=roc_auc_inputs["y_true"], y_score=roc_auc_inputs["y_score"])

mlebench/competitions/histopathologic-cancer-detection/prepare.py

@@ -0,0 +1,59 @@
+import json
+import shutil
+from pathlib import Path
+
+import pandas as pd
+from sklearn.model_selection import train_test_split
+from tqdm import tqdm
+
+
+def prepare(raw: Path, public: Path, private: Path):
+    old_train = pd.read_csv(raw / "train_labels.csv")
+
+    num_test = len(list((raw / "test").glob("*.tif")))
+    test_ratio = num_test / (len(old_train) + num_test)
+
+    new_train_ids, new_test_ids = train_test_split(
+        old_train["id"], test_size=test_ratio, random_state=0
+    )
+    new_train = old_train[old_train["id"].isin(new_train_ids)]
+    new_test = old_train[old_train["id"].isin(new_test_ids)]
+
+    assert set(new_train["id"]).isdisjoint(
+        set(new_test["id"])
+    ), "Train should not contain id's of test images"
+    assert len(new_train_ids) + len(new_test_ids) == len(
+        old_train
+    ), "The combined length of new_train_ids and new_test_ids should equal the length of old_train"
+
+    # Copy over files
+    (public / "train").mkdir(exist_ok=True)
+    (public / "test").mkdir(exist_ok=True)
+
+    for file_id in tqdm(new_train_ids, desc="Copying train images"):
+        shutil.copyfile(
+            src=raw / "train" / f"{file_id}.tif",
+            dst=public / "train" / f"{file_id}.tif",
+        )
+    for file_id in tqdm(new_test_ids, desc="Copying test images"):
+        shutil.copyfile(
+            src=raw / "train" / f"{file_id}.tif",
+            dst=public / "test" / f"{file_id}.tif",
+        )
+
+    # Create sample submission
+    sample_submission = new_test.copy()
+    sample_submission["label"] = 0
+
+    # Copy over files
+    new_train.to_csv(public / "train_labels.csv", index=False)
+    new_test.to_csv(private / "answers.csv", index=False)
+    sample_submission.to_csv(public / "sample_submission.csv", index=False)
+
+    # Check integrity of files copied
+    assert len(list(public.glob("train/*.tif"))) == len(
+        new_train_ids
+    ), "Number of train images should be equal to the number of unique id's in the train set"
+    assert len(list(public.glob("test/*.tif"))) == len(
+        new_test_ids
+    ), "Number of test images should be equal to the number of unique id's in the test set"

mlebench/competitions/histopathologic-cancer-detection/prepare_val.py

@@ -0,0 +1,131 @@
+import shutil
+from pathlib import Path
+
+import pandas as pd
+from sklearn.model_selection import train_test_split
+from tqdm import tqdm
+
+
+def _create_split(
+    df_to_split: pd.DataFrame,
+    image_source_dir: Path,
+    test_split_size: float,
+    public_dest: Path,
+    private_dest: Path,
+    random_state: int,
+) -> pd.DataFrame:
+    """
+    Splits a dataframe of labels, copies corresponding images, and saves the results.
+
+    Args:
+        df_to_split: DataFrame containing image IDs and labels to be split.
+        image_source_dir: Path to the directory containing the source images.
+        test_split_size: The proportion of the dataset to allocate to the test split.
+        public_dest: The destination directory for public-facing files (train set, test images, sample submission).
+        private_dest: The destination directory for private files (test set answers).
+        random_state: The seed used by the random number generator for the split.
+
+    Returns:
+        The training portion of the split as a pandas DataFrame.
+    """
+    # Create destination directories
+    public_dest.mkdir(exist_ok=True, parents=True)
+    private_dest.mkdir(exist_ok=True, parents=True)
+    (public_dest / "train").mkdir(exist_ok=True)
+    (public_dest / "test").mkdir(exist_ok=True)
+
+    # Perform the split
+    train_ids, test_ids = train_test_split(
+        df_to_split["id"], test_size=test_split_size, random_state=random_state
+    )
+    train_df = df_to_split[df_to_split["id"].isin(train_ids)]
+    test_df = df_to_split[df_to_split["id"].isin(test_ids)]
+
+    assert set(train_df["id"]).isdisjoint(
+        set(test_df["id"])
+    ), "Train should not contain id's of test images"
+    assert len(train_ids) + len(test_ids) == len(
+        df_to_split
+    ), "The combined length of train_ids and test_ids should equal the length of the source df"
+
+    # Copy over image files
+    for file_id in tqdm(train_ids, desc=f"Copying train images to {public_dest}"):
+        shutil.copyfile(
+            src=image_source_dir / f"{file_id}.tif",
+            dst=public_dest / "train" / f"{file_id}.tif",
+        )
+    for file_id in tqdm(test_ids, desc=f"Copying test images to {public_dest}"):
+        shutil.copyfile(
+            src=image_source_dir / f"{file_id}.tif",
+            dst=public_dest / "test" / f"{file_id}.tif",
+        )
+
+    # Create and save label/submission files
+    sample_submission = test_df.copy()
+    sample_submission["label"] = 0
+
+    train_df.to_csv(public_dest / "train_labels.csv", index=False)
+    test_df.to_csv(private_dest / "answers.csv", index=False)
+    sample_submission.to_csv(public_dest / "sample_submission.csv", index=False)
+
+    # Check integrity of files copied
+    assert len(list(public_dest.glob("train/*.tif"))) == len(
+        train_ids
+    ), "Number of train images should be equal to the number of unique id's in the train set"
+    assert len(list(public_dest.glob("test/*.tif"))) == len(
+        test_ids
+    ), "Number of test images should be equal to the number of unique id's in the test set"
+
+    return train_df
+
+
+def prepare(raw: Path, public: Path, private: Path):
+    """
+    Prepares the data for the competition by performing two splits.
+    1. A main split of the raw data into a primary train/test set for the competition.
+       Outputs are saved to `public/` and `private/`.
+    2. A secondary split of the primary train set into a smaller train/validation set for user convenience.
+       Outputs are saved to `public_val/` and `private_val/`, mirroring the main structure.
+    """
+    # Common setup
+    all_train_labels = pd.read_csv(raw / "train_labels.csv")
+    image_source_dir = raw / "train"
+    RANDOM_STATE = 0
+
+    # --- 1. Main Split: Create the primary train/test sets ---
+    # This logic is identical to the original script to ensure outputs do not change.
+    num_test_from_pool = len(list((raw / "test").glob("*.tif")))
+    test_ratio_main = num_test_from_pool / (len(all_train_labels) + num_test_from_pool)
+
+    # The returned `primary_train_df` is the larger portion of the first split.
+    primary_train_df = _create_split(
+        df_to_split=all_train_labels,
+        image_source_dir=image_source_dir,
+        test_split_size=test_ratio_main,
+        public_dest=public,
+        private_dest=private,
+        random_state=RANDOM_STATE,
+    )
+
+    # --- 2. Validation Split: Create a secondary train/validation set ---
+    # Define new output paths for the validation split
+    public_val = public.parent / "public_val"
+    private_val = private.parent / "private_val"
+
+    # To get the same number of test samples as the main split, we must adjust the test_size
+    # for this second split, which operates on a smaller dataset (the primary_train_df).
+    if (1 - test_ratio_main) <= 0:
+        # Avoid division by zero in the unlikely case the main split uses all data for test
+        test_ratio_val = 0
+    else:
+        test_ratio_val = test_ratio_main / (1 - test_ratio_main)
+
+    # Perform the second split on the primary training data
+    _create_split(
+        df_to_split=primary_train_df,
+        image_source_dir=image_source_dir,
+        test_split_size=test_ratio_val,
+        public_dest=public_val,
+        private_dest=private_val,
+        random_state=RANDOM_STATE,
+    )

mlebench/competitions/hms-harmful-brain-activity-classification/constants.py

@@ -0,0 +1,9 @@
+ID_COL = "eeg_id"
+TARGET_COLS = [
+    "seizure_vote",
+    "lpd_vote",
+    "gpd_vote",
+    "lrda_vote",
+    "grda_vote",
+    "other_vote",
+]

mlebench/competitions/hms-harmful-brain-activity-classification/grade.py

@@ -0,0 +1,43 @@
+import pandas as pd
+
+from mlebench.grade_helpers import InvalidSubmissionError
+
+from . import kullback_leibler_divergence as kl_divergence
+from .constants import ID_COL, TARGET_COLS
+
+
+def prepare_for_metric(submission: pd.DataFrame, answers: pd.DataFrame) -> tuple:
+
+    if ID_COL not in submission.columns:
+        raise InvalidSubmissionError(f"Submission must contain {ID_COL} column")
+    if not all(col in submission.columns for col in TARGET_COLS):
+        raise InvalidSubmissionError(f"Submission must contain all target columns: {TARGET_COLS}")
+    if len(submission) != len(answers):
+        raise InvalidSubmissionError("Submission and answers must have the same length")
+    if not (submission[TARGET_COLS].sum(axis=1).apply(lambda x: round(x, 5) == 1).all()):
+        raise InvalidSubmissionError("Submission probabilities must add to 1 for each row")
+    if not set(answers[ID_COL]) == set(submission[ID_COL]):
+        raise InvalidSubmissionError("Submission and answers must have the same IDs")
+
+    assert ID_COL in answers.columns, f"Answers must contain {ID_COL} column"
+    assert all(
+        col in answers.columns for col in TARGET_COLS
+    ), f"Answers must contain all target columns: {TARGET_COLS}"
+
+    submission = submission.sort_values(ID_COL).reset_index(drop=True)
+    answers = answers.sort_values(ID_COL).reset_index(drop=True)
+
+    answers = answers.copy()[[ID_COL] + TARGET_COLS]
+    # normalize answers to be max 1, by taking vote / sum(votes)
+    # https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/468705#2606605
+    answers[TARGET_COLS] = answers[TARGET_COLS].div(answers[TARGET_COLS].sum(axis=1), axis=0)
+
+    return submission, answers
+
+
+def grade(submission: pd.DataFrame, answers: pd.DataFrame) -> float:
+
+    submission, answers = prepare_for_metric(submission, answers)
+    score = kl_divergence.score(answers, submission, ID_COL)
+
+    return score

mlebench/competitions/hms-harmful-brain-activity-classification/kaggle_metric_utilities.py

@@ -0,0 +1,96 @@
+"""
+This script exists to reduce code duplication across metrics.
+Source: https://www.kaggle.com/code/metric/kaggle-metric-utilities
+Linked from: https://www.kaggle.com/code/metric/kullback-leibler-divergence
+Linked from: https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification
+"""
+
+from typing import Union
+
+import numpy as np
+import pandas as pd
+import pandas.api.types
+
+
+class ParticipantVisibleError(Exception):
+    pass
+
+
+class HostVisibleError(Exception):
+    pass
+
+
+def treat_as_participant_error(
+    error_message: str, solution: Union[pd.DataFrame, np.ndarray]
+) -> bool:
+    """Many metrics can raise more errors than can be handled manually. This function attempts
+    to identify errors that can be treated as ParticipantVisibleError without leaking any competition data.
+
+    If the solution is purely numeric, and there are no numbers in the error message,
+    then the error message is sufficiently unlikely to leak usable data and can be shown to participants.
+
+    We expect this filter to reject many safe messages. It's intended only to reduce the number of errors we need to manage manually.
+    """
+    # This check treats bools as numeric
+    if isinstance(solution, pd.DataFrame):
+        solution_is_all_numeric = all(
+            [pandas.api.types.is_numeric_dtype(x) for x in solution.dtypes.values]
+        )
+        solution_has_bools = any(
+            [pandas.api.types.is_bool_dtype(x) for x in solution.dtypes.values]
+        )
+    elif isinstance(solution, np.ndarray):
+        solution_is_all_numeric = pandas.api.types.is_numeric_dtype(solution)
+        solution_has_bools = pandas.api.types.is_bool_dtype(solution)
+
+    if not solution_is_all_numeric:
+        return False
+
+    for char in error_message:
+        if char.isnumeric():
+            return False
+    if solution_has_bools:
+        if "true" in error_message.lower() or "false" in error_message.lower():
+            return False
+    return True
+
+
+def safe_call_score(metric_function, solution, submission, **metric_func_kwargs):
+    """
+    Call score. If that raises an error and that already been specifically handled, just raise it.
+    Otherwise make a conservative attempt to identify potential participant visible errors.
+    """
+    try:
+        score_result = metric_function(solution, submission, **metric_func_kwargs)
+    except Exception as err:
+        error_message = str(err)
+        if err.__class__.__name__ == "ParticipantVisibleError":
+            raise ParticipantVisibleError(error_message)
+        elif err.__class__.__name__ == "HostVisibleError":
+            raise HostVisibleError(error_message)
+        else:
+            if treat_as_participant_error(error_message, solution):
+                raise ParticipantVisibleError(error_message)
+            else:
+                raise err
+    return score_result
+
+
+def verify_valid_probabilities(df: pd.DataFrame, df_name: str):
+    """Verify that the dataframe contains valid probabilities.
+
+    The dataframe must be limited to the target columns; do not pass in any ID columns.
+    """
+    if not pandas.api.types.is_numeric_dtype(df.values):
+        raise ParticipantVisibleError(f"All target values in {df_name} must be numeric")
+
+    if df.min().min() < 0:
+        raise ParticipantVisibleError(f"All target values in {df_name} must be at least zero")
+
+    if df.max().max() > 1:
+        raise ParticipantVisibleError(f"All target values in {df_name} must be no greater than one")
+
+    if not np.allclose(df.sum(axis=1), 1):
+        raise ParticipantVisibleError(
+            f"Target values in {df_name} do not add to one within all rows"
+        )