process/eval.py

import json
from pathlib import Path

import numpy as np
import fire
import h5py
from sklearn.metrics import accuracy_score, mean_absolute_error, f1_score

from infer_eval_utils import read_time_series_data, concat_base_path, non_zero_rel_mae, DATASET_TO_TASK


class Runner(object):

    def multitask_classification(self,
                                 infer_path: str = "",
                                 gts: list = [],
                                 preds: list = []):
        tasks = gts[0].keys()
        output_fpath = Path(
            infer_path).parent / f"results/{Path(infer_path).stem}.json"
        output_fpath.parent.mkdir(parents=True, exist_ok=True)
        res_dict = {}
        success = 0
        fail = 0
        for task in tasks:
            correct_count = 0
            for gt, pred in zip(gts, preds):
                try:
                    if pred[task] == gt[task][0] or pred[task].lower(
                    ) == gt[task][0]:
                        correct_count += 1
                    success += 1
                except:
                    fail += 1
            acc = correct_count / len(gts)
            print(f"Accuracy for {task}: {acc}")
            class_f1s = []
            task_gts = [gt[task][0] for gt in gts]
            labels = list(set(task_gts))
            for label in labels:
                # Find all samples with this true label
                true_indices = [
                    i for i, gt in enumerate(task_gts) if gt == label
                ]
                if len(true_indices) == 0:
                    recall = 0.0
                else:
                    correct_predictions = 0
                    for idx in true_indices:
                        pred = preds[idx][task]
                        gt = task_gts[idx]
                        if pred == gt or pred.lower() == gt:
                            correct_predictions += 1

                    recall = correct_predictions / len(true_indices)

                pred_indices = [
                    i for i, pred in enumerate(preds)
                    if pred[task].lower() == label or pred[task] == label
                ]
                if len(pred_indices) == 0:
                    precision = 0.0
                else:
                    correct_predictions = 0
                    for idx in pred_indices:
                        pred = preds[idx][task]
                        gt = task_gts[idx]
                        if pred == gt or pred.lower() == gt:
                            correct_predictions += 1
                    precision = correct_predictions / len(pred_indices)

                f1 = (2 * recall * precision) / (recall + precision + 1e-6)
                class_f1s.append(f1)

            res_dict[task] = {
                "acc": acc,
                "f1": f1,
                "success": success,
                "fail": fail,
                "success_rate": success / (success + fail)
            }

        res_dict["overall"] = {
            "f1": np.mean(class_f1s),
            "acc": np.mean([r["acc"] for r in res_dict.values()]),
        }

        with open(output_fpath, "w") as writer:
            json.dump(res_dict, writer, indent=4)
            writer.write("\n")

    def multichoice_classification(self,
                                   infer_path: str = "",
                                   gts: list = [],
                                   preds: list = []):
        all_labels = set()
        for gt in gts:
            if isinstance(gt, list):
                all_labels.update(gt)
            else:
                all_labels.add(gt)

        all_labels = sorted(list(all_labels))

        # Convert ground truth and predictions to multi-label format
        y_true_multilabel = []
        y_pred_multilabel = []
        success = 0
        fail = 0
        for gt, pred in zip(gts, preds):
            # Process ground truth
            if isinstance(gt, list):
                gt_labels = gt
            else:
                gt_labels = [gt]

            # Process predictions
            while '\n\n' in pred:
                pred = pred.replace('\n\n', '\n')
            pred_labels = [x.strip() for x in pred.split("\n")]

            # Convert to binary vectors
            gt_binary = [
                1 if label in gt_labels else 0 for label in all_labels
            ]
            pred_binary = []
            for label in all_labels:
                is_found = False
                for pred_label in pred_labels:
                    if pred_label == label or pred_label.lower() == label:
                        is_found = True
                        break
                if is_found:
                    pred_binary.append(1)
                else:
                    pred_binary.append(0)

            y_true_multilabel.append(gt_binary)
            y_pred_multilabel.append(pred_binary)

        y_true_multilabel = np.array(y_true_multilabel)
        y_pred_multilabel = np.array(y_pred_multilabel)

        # Calculate F1 score for each class
        f1_scores = []
        for i, label in enumerate(all_labels):
            f1 = f1_score(y_true_multilabel[:, i],
                          y_pred_multilabel[:, i],
                          zero_division=0)
            f1_scores.append(f1)
            print(f"F1 score for class {label}: {f1:.4f}")

        # Calculate mean of F1 scores
        macro_f1 = np.mean(f1_scores)
        print(f"Macro F1 score (mean of all classes): {macro_f1:.4f}")

        # Save results
        output_fpath = Path(
            infer_path).parent / f"results/{Path(infer_path).stem}.json"
        output_fpath.parent.mkdir(parents=True, exist_ok=True)

        results = {
            "macro_f1": macro_f1,
            "per_class_f1": dict(zip(all_labels, f1_scores))
        }

        with open(output_fpath, "w") as writer:
            json.dump(results, writer, indent=4)
            writer.write("\n")

    def classification(self, infer_path: str = ""):
        gts, preds = [], []
        with open(infer_path, "r") as f:
            for line in f:
                item = json.loads(line)
                if "id" not in item:
                    continue
                gts.append(item["ground_truth"])
                preds.append(item["output"])

        if any(isinstance(gt, list) for gt in gts):
            return self.multichoice_classification(infer_path, gts, preds)

        if isinstance(preds[0], dict):
            return self.multitask_classification(infer_path, gts, preds)
        # Custom comparison function: consider both exact match and case-insensitive match
        # This is because LLM outputs sometimes capitalize the first letter to follow English grammar
        correct_count = 0
        for gt, pred in zip(gts, preds):
            if pred == gt or pred.lower() == gt:
                correct_count += 1

        acc = correct_count / len(gts)
        print(f"Accuracy: {acc}")

        # Using the same case-insensitive matching as accuracy calculation
        labels = list(set(gts))

        # Manually calculate recall for each class
        class_recalls = []
        class_precisions = []
        class_f1s = []
        for label in labels:
            # Find all samples with this true label
            true_indices = [i for i, gt in enumerate(gts) if gt == label]
            if len(true_indices) == 0:
                recall = 0.0
                class_recalls.append(0.0)
            else:
                # Calculate recall for this label
                correct_predictions = 0
                for idx in true_indices:
                    pred = preds[idx]
                    gt = gts[idx]
                    if pred == gt or pred.lower() == gt:
                        correct_predictions += 1

                recall = correct_predictions / len(true_indices)
                class_recalls.append(recall)

            pred_indices = [
                i for i, pred in enumerate(preds)
                if pred.lower() == label or pred == label
            ]
            if len(pred_indices) == 0:
                precision = 0.0
                class_precisions.append(0.0)
            else:
                correct_predictions = 0
                for idx in pred_indices:
                    pred = preds[idx]
                    gt = gts[idx]
                    if pred == gt or pred.lower() == gt:
                        correct_predictions += 1

                precision = correct_predictions / len(pred_indices)
                class_precisions.append(precision)

            f1 = (2 * recall * precision) / (recall + precision + 1e-6)
            class_f1s.append(f1)

        output_fpath = Path(
            infer_path).parent / f"results/{Path(infer_path).stem}.json"
        output_fpath.parent.mkdir(parents=True, exist_ok=True)
        with open(output_fpath, "w") as writer:
            json.dump({
                "acc": acc,
                "f1": np.mean(class_f1s),
            },
                      writer,
                      indent=4)
            writer.write("\n")

    def mcq(self, infer_path: str = ""):
        gts, preds = [], []
        with open(infer_path, "r") as f:
            for line in f:
                item = json.loads(line)
                if "id" not in item:
                    continue
                gts.append(item["ground_truth"])
                preds.append(item["output"])

        # Custom comparison function: consider both exact match and case-insensitive match
        # This is because LLM outputs sometimes capitalize the first letter to follow English grammar
        correct_count = 0
        for gt, pred in zip(gts, preds):
            if pred == gt or pred.lower() == gt:
                correct_count += 1

        acc = correct_count / len(gts)
        print(f"Accuracy: {acc}")

        # Calculate recall with custom matching logic
        # Using the same case-insensitive matching as accuracy calculation
        try:
            labels = list(set(gts))

            # Manually calculate recall for each class
            class_recalls = []
            for label in labels:
                # Find all samples with this true label
                true_indices = [i for i, gt in enumerate(gts) if gt == label]
                if len(true_indices) == 0:
                    class_recalls.append(0.0)
                    continue

                # Calculate recall for this label (using lenient matching)
                correct_predictions = 0
                for idx in true_indices:
                    pred = preds[idx]
                    gt = gts[idx]
                    if pred == gt or pred.lower() == gt:
                        correct_predictions += 1

                recall = correct_predictions / len(true_indices)
                class_recalls.append(recall)

            # Calculate macro average recall
            weighted_recall = np.mean(class_recalls)
            print(f"Weighted Recall: {weighted_recall}")

            output_fpath = Path(
                infer_path).parent / f"results/{Path(infer_path).stem}.json"
            output_fpath.parent.mkdir(parents=True, exist_ok=True)
            with open(output_fpath, "w") as writer:
                json.dump({
                    "acc": acc,
                    "uar": weighted_recall
                },
                          writer,
                          indent=4)
                writer.write("\n")

        except Exception as e:
            print(f"Error calculating Weighted Recall: {e}")
            print(
                "Possible reasons: labels are not numeric or contain non-numeric labels"
            )

    def anomaly_detection(self, infer_path: str = ""):
        gts, preds = [], []
        success, fail = 0, 0
        with open(infer_path, "r") as f:
            for line in f:
                item = json.loads(line)
                if "id" not in item:
                    continue

                if item["output"].lower() == "yes":
                    preds.append(True)
                elif item["output"].lower() == "no":
                    preds.append(False)
                else:
                    fail += 1
                    continue
                success += 1
                gts.append(item["ground_truth"])

        correct_count = 0
        for gt, pred in zip(gts, preds):
            if pred == gt:
                correct_count += 1
        print(f"Success: {success}, Fail: {fail}")
        acc = correct_count / len(gts)
        print(f"Accuracy: {acc}")

        f1 = f1_score(gts, preds)
        print(f"F1 Score: {f1}")
        output_fpath = Path(
            infer_path).parent / f"results/{Path(infer_path).stem}.json"
        output_fpath.parent.mkdir(parents=True, exist_ok=True)
        with open(output_fpath, "w") as writer:
            json.dump({
                "acc": acc,
                "f1": f1,
                "success": success,
                "fail": fail
            },
                      writer,
                      indent=4)
            writer.write("\n")

    def forecasting(self, infer_path: str = ""):
        gt_arrs = []
        pred_arrs = []
        success = 0
        fail = 0
        with h5py.File(infer_path, "r") as f:
            base_path = Path(f["base_path"][()].decode("utf-8"))
            for id in f.keys():
                try:
                    if id not in [
                            "base_path", "dataset_name", "domain", "task",
                            "scene"
                    ]:
                        gt_path = concat_base_path(
                            base_path,
                            f[id]["gt_path"][()].decode("utf-8").strip("/"))
                        gt_data = read_time_series_data(gt_path)
                        gt_data = np.array(gt_data, dtype=np.float32)
                        pred = f[id]["pred_result"][()]

                        if pred.shape != gt_data.shape:
                            raise ValueError(
                                f"Pred shape {pred.shape} does not match gt shape {gt_data.shape}"
                            )
                        gt_arrs.append(gt_data.reshape(-1))
                        pred_arrs.append(pred.reshape(-1))
                        success += 1
                except Exception as e:
                    print(f"Error processing {id}: {e}")
                    fail += 1

        if len(gt_arrs) == 0:
            mae = "N/A"
            rel_mae = "N/A"
        else:
            gt_arrs = np.concatenate(gt_arrs)
            pred_arrs = np.concatenate(pred_arrs)

            # mse = mean_squared_error(gt_arrs, pred_arrs)
            mae = mean_absolute_error(gt_arrs, pred_arrs)
            rel_mae = non_zero_rel_mae(gt_arrs, pred_arrs)
        print(
            f"MAE: {mae}, REL_MAE: {rel_mae}, Success: {success}, Fail: {fail}"
        )
        output_fpath = Path(
            infer_path).parent / f"results/{Path(infer_path).stem}.json"
        output_fpath.parent.mkdir(parents=True, exist_ok=True)
        with open(output_fpath, "w") as writer:
            json.dump(
                {
                    "rel_mae": rel_mae,
                    "mae": mae,
                    "success": success,
                    "fail": fail,
                    "success_rate": success / (success + fail)
                },
                writer,
                indent=4)
            writer.write("\n")

    def synthesize(self, infer_path: str = ""):
        return self.forecasting(infer_path)

    def imputation(self, infer_path: str = ""):
        gt_arrs = []
        pred_arrs = []
        success = 0
        fail = 0
        with h5py.File(infer_path, "r") as f:
            base_path = Path(f["base_path"][()].decode("utf-8"))
            for id in f.keys():
                try:
                    if id not in [
                            "base_path", "dataset_name", "domain", "task",
                            "scene"
                    ]:
                        # gt_path = base_path / f[id]["gt_path"][
                        #     ()].decode("utf-8")
                        gt_path = concat_base_path(
                            base_path,
                            f[id]["gt_path"][()].decode("utf-8").strip("/"))
                        gt_data = read_time_series_data(gt_path)

                        # input_path = base_path / f[id]["input_ts_path"][
                        # ()].decode("utf-8")
                        input_path = concat_base_path(
                            base_path, f[id]["input_ts_path"][(
                            )].decode("utf-8").strip("/"))
                        input_data = read_time_series_data(input_path)

                        pred_indices = np.where(input_data == "X")[0]
                        pred = f[id]["pred_result"][()]

                        pred = pred[pred_indices]
                        gt_data = gt_data[pred_indices]
                        if len(pred) != len(gt_data):
                            length_mismatch += 1
                        else:
                            success += 1
                        if len(pred) < len(gt_data):
                            pred = pred[:len(gt_data)]
                        if len(pred) > len(gt_data):
                            gt_data = gt_data[:len(pred)]
                        gt_arrs.append(gt_data)
                        pred_arrs.append(pred)
                        # success += 1
                except Exception as e:
                    print(f"Error processing {id}: {e}")
                    fail += 1

        gt_arrs = np.concatenate(gt_arrs)
        pred_arrs = np.concatenate(pred_arrs)
        # mse = mean_squared_error(gt_arrs, pred_arrs)
        rel_mae = non_zero_rel_mae(gt_arrs, pred_arrs)
        mae = mean_absolute_error(gt_arrs, pred_arrs)

        print(
            f"REL_MAE: {rel_mae}, MAE: {mae}, Success: {success}, Fail: {fail}"
        )
        output_fpath = Path(
            infer_path).parent / f"results/{Path(infer_path).stem}.json"
        output_fpath.parent.mkdir(parents=True, exist_ok=True)
        with open(output_fpath, "w") as writer:
            json.dump(
                {
                    "rel_mae": rel_mae,
                    "mae": mae,
                    "success": success,
                    "fail": fail,
                    "success_rate": success / (success + fail)
                },
                writer,
                indent=4)
            writer.write("\n")

    def event_detection(self, infer_path: str = ""):
        event_gts, event_preds = [], []
        seq_length = None
        success = 0
        total = 0
        timestamp_gts, timestamp_preds = [], []
        with open(infer_path, "r") as f:
            for line in f:
                item = json.loads(line)
                if "id" not in item:
                    seq_length = item["seq_length"]
                    continue
                event_gt = item["ground_truth"]["contain"]
                event_gts.append(1 if event_gt else 0)

                if "\n" in item["output"]:
                    while '\n\n' in item["output"]:
                        item["output"] = item["output"].replace('\n\n', '\n')
                    event_pred, *timestamps = item["output"].split("\n")
                else:
                    event_pred = item["output"]
                    timestamps = None
                event_preds.append(1 if event_pred.lower() == "yes" else 0)

                if event_gt:
                    if "start_time" in item["ground_truth"]:
                        gt_timestamps = [item["ground_truth"]["start_time"]]
                    elif "start_time_p" in item["ground_truth"]:
                        gt_timestamps = [
                            item["ground_truth"]["start_time_p"],
                            item["ground_truth"]["start_time_s"]
                        ]

                    if timestamps is None:
                        pass
                    else:
                        try:
                            assert len(timestamps) == len(gt_timestamps)
                            for pred_timestamp, gt_timestamp in zip(
                                    timestamps, gt_timestamps):
                                pred_timestamp = eval(pred_timestamp)
                                timestamp_preds.append(pred_timestamp)
                                timestamp_gts.append(gt_timestamp)
                            success += 1
                        except Exception as e:
                            print(str(e))
                    total += 1

        event_acc = accuracy_score(event_gts, event_preds)
        event_f1 = f1_score(event_gts, event_preds)
        timestamp_gts = np.array(timestamp_gts)
        timestamp_preds = np.array(timestamp_preds)
        mape = non_zero_rel_mae(timestamp_gts, timestamp_preds)
        output_fpath = Path(
            infer_path).parent / f"results/{Path(infer_path).stem}.json"
        output_fpath.parent.mkdir(parents=True, exist_ok=True)
        with open(output_fpath, "w") as writer:
            json.dump(
                {
                    "acc": event_acc,
                    "f1": event_f1,
                    "mape": mape,
                    "success_rate": success / total
                },
                writer,
                indent=4)
            writer.write("\n")
        print({
            "acc": event_acc,
            "f1": event_f1,
            "mape": mape,
            "success_rate": success / total
        })

    def evaluate(self, infer_dir: str):
        for infer_path in Path(infer_dir).glob("*"):
            if infer_path.is_dir():
                continue
            dataset_id = infer_path.stem
            task = DATASET_TO_TASK[dataset_id]
            print(f"evaluating {dataset_id} ...")
            getattr(self, task)(infer_path)


if __name__ == "__main__":
    fire.Fire(Runner)