script.py

import pandas as pd
import numpy as np
import os
import sys
from tqdm import tqdm
import timm
import torchvision.transforms as T
from PIL import Image
import torch
from multiprocessing import Pool

from mmpretrain.apis import ImageClassificationInferencer, FeatureExtractor

import mmpretrain.utils.progress as progress
progress.disable_progress_bar = True


sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.realpath(__file__))))


def load_image(path : str, images_root_path="/tmp/data/private_testset"):
    return np.array(Image.open(os.path.join(images_root_path, path)))[:, :, ::-1]

def rerank_poison(posison_status_list : pd.DataFrame, pred_scores : np.array) -> tuple[int, float]:
    class_id = np.argmax(pred_scores)
    class_score = np.max(pred_scores)

    poisonous = posison_status_list.copy()
    poisonous['score'] = pred_scores
    poisonous.sort_values(by=['score'], ascending=False, inplace=True)
    first_poisonous = poisonous[poisonous['poisonous'] == 1].iloc[0]

    if 13 * first_poisonous['score'] > class_score:
        class_id = first_poisonous['class_id']
        class_score = first_poisonous['score']

    return class_id, class_score


def make_submission(test_metadata, model_path, model_name, output_csv_path="./submission.csv", images_root_path="/tmp/data/private_testset"):
    """Make submission with given """

    #inferencer = ImageClassificationInferencer(model=model_name, pretrained=model_path, device="cuda:0")
    feature_extractor = FeatureExtractor(model=model_name, pretrained=model_path, device="cuda:0")

    predictions = []
    prediction_scores = []
    prediction_scores_dict = {}
    prediction_feats_dict = {}
    obs_imgs_dict = {}

    BATCH_SIZE = 4
    p = Pool(BATCH_SIZE)
    # image_paths_next_batch = test_metadata['image_path'][0:BACTH_SIZE]
    # next_batch = p.map_async(load_image, image_paths_next_batch)
    for i in tqdm(range(int(np.ceil(test_metadata.shape[0] / BATCH_SIZE)))):
        # batch_imgs = next_batch.get()
        # image_paths_next_batch = test_metadata['image_path'][(i+1) * BACTH_SIZE:(i+2) * BATCH_SIZE]
        # next_batch = p.map_async(load_image, image_paths_next_batch)
        img_paths_batch = test_metadata['image_path'][(i) * BATCH_SIZE:(i+1) * BATCH_SIZE]
        batch_imgs = p.map(load_image, img_paths_batch)
        #batch_imgs = [np.array(Image.open(os.path.join(images_root_path, x)))[:, :, ::-1] for x in test_metadata['image_path'][(i) * BATCH_SIZE:(i+1) * BATCH_SIZE]]
        #results = inferencer(batch_imgs, batch_size=BATCH_SIZE)
        feats = feature_extractor(batch_imgs, batch_size=BATCH_SIZE)
        feats = (torch.stack([x[0] for x in feats], dim=0),)
        results = feature_extractor.model.head.task_heads['species'].predict(feats, img_paths=img_paths_batch)
        for res, f, obs_id, img_path in zip(results, feats[0], test_metadata['observation_id'][(i) * BATCH_SIZE:(i+1) * BATCH_SIZE], img_paths_batch):
            #pred_scores = res.species.pred_score.detach().cpu().numpy()
            pred_scores = res.pred_score.detach().cpu().numpy()
            #pred_scores = res['pred_scores']
            predictions.append(np.argmax(pred_scores))
            prediction_scores.append(pred_scores)
            prediction_scores_dict.setdefault(obs_id, []).append(pred_scores)
            prediction_feats_dict.setdefault(obs_id, []).append(f)
            obs_imgs_dict[obs_id] = img_path

    print('finished inference')

    test_metadata["class_id"] = predictions
    test_metadata["max_score"] = prediction_scores

    poison_status_list = pd.read_csv('poison_status_list.csv')
    poison_status_list = poison_status_list.sort_values(by=['class_id'])

    poison_classes = set(poison_status_list[poison_status_list['poisonous'] == 1]['class_id'])

    for obs_id, pred_feats in tqdm(prediction_feats_dict.items()):
        #fusion_scores = np.prod(np.array(pred_scores), axis=0)
        #fusion_scores = np.mean(np.array(pred_scores), axis=0)
        #fusion_scores = np.max(np.array(pred_scores), axis=0)
        fusion_feats = torch.mean(torch.stack(pred_feats, dim=0), dim=0, keepdim=True)
        results = feature_extractor.model.head.task_heads['species'].predict((fusion_feats,), img_paths=[obs_imgs_dict[obs_id]])
        fusion_scores = results[0].pred_score.detach().cpu().numpy()
        class_score = np.max(fusion_scores)
        class_id = np.argmax(fusion_scores)
        class_id, class_score = rerank_poison(poison_status_list, fusion_scores)
        entropy = -np.sum(fusion_scores * np.log(fusion_scores))
        if entropy > 7 or (class_id not in poison_classes and entropy > 2.5):
            class_id = -1
        #if class_score < 0.1:
        #    class_id = -1
        test_metadata.loc[test_metadata["observation_id"] == obs_id, "class_id"] = class_id
        test_metadata.loc[test_metadata["observation_id"] == obs_id, "max_score"] = class_score

    user_pred_df = test_metadata.drop_duplicates("observation_id", keep="first")
    user_pred_df[["observation_id", "class_id"]].to_csv(output_csv_path, index=None)


if __name__ == "__main__":

    import zipfile

    with zipfile.ZipFile("/tmp/data/private_testset.zip", 'r') as zip_ref:
        zip_ref.extractall("/tmp/data")

    MODEL_PATH = "swinv2_base_w24_b16x8-fp16_fungi+val2_res_384_genus-loss_no-unknown-target-zero_metadata_epochs_4_epoch_2_20240524-a429ecac.pth"
    MODEL_NAME = "swinv2_base_w24_b16x8-fp16_fungi+val2_res_384_genus-loss_no-unknown-target-zero_metadata_epochs_4.py"

    metadata_file_path = "./FungiCLEF2024_TestMetadata.csv"
    test_metadata = pd.read_csv(metadata_file_path)

    make_submission(
        test_metadata=test_metadata,
        model_path=MODEL_PATH,
        model_name=MODEL_NAME
    )