extract_train.py

## This code is based on the bioptimus/H-optimus-0 model of Hugging Face Hub. 
## Source: https://huggingface.co/bioptimus/H-optimus-0 
## This code has been partially modified from the original.

## hdf5 usage under HDF5 License. For details see:
## See https://docs.h5py.org/en/stable/licenses.html
## Redistribution and use in source and binary forms, with or without modification, are permitted for any purpose.

## h5py usage under the terms specified by the HDF5 License. 
## Copyright (c) 2008 Andrew Collette and contributors.
## All rights reserved.
## Redistribution and use in source and binary forms, with or without modification, are permitted provided the following conditions are met:
## 1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
## 2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions, and the following disclaimer in the documentation and/or other materials provided with the distribution.
## See https://docs.h5py.org/en/stable/licenses.html

# This software includes components from webdataset provided by NVIDIA CORPORATION.
# Copyright 2020 NVIDIA CORPORATION. All rights reserved.


# import package
import yaml
import os
import torch
from glob import glob
import timm
from torchvision import transforms
import pandas as pd
from sklearn.preprocessing import LabelEncoder
import webdataset as wds
from torch.utils.data import DataLoader, Dataset
import math
import h5py
import numpy as np
from tqdm import tqdm
from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from sklearn.metrics import accuracy_score, balanced_accuracy_score
from huggingface_hub import login
import braceexpand

# set directories
work_dir = "."

# load config
config_path = os.path.join(work_dir, "config.yaml")
with open(config_path, 'r') as file:
    configs = yaml.safe_load(file)
        
# model name and path
## The following features were extended from the original bioptimus/H-optimus-0
model_dic = {
    "h_optimus": "hf-hub:bioptimus/H-optimus-0",
    # if you want to use other model, please check the path
}
configs["model_path"] = model_dic[configs["model_name"]]
configs["eval_name"] = configs.get("eval_name", "logreg")  # ["logreg", "knn", "proto"]
configs["max_iter"] = configs.get("max_iter", 1000)
configs["cost"] = configs.get("cost", 0.0001)
configs["k"] = configs.get("k", 10)

# load meta data
metadata_path = os.path.join(work_dir, "train_val_test_split.csv")
df = pd.read_csv(metadata_path)

# get huggingface token
split = configs["split_type"]
file_range = 9 if split == "internal" else 8
patterns = {
    'train': [os.path.join(work_dir, f"data/dataset_{split}_train_part{str(i).zfill(3)}.tar") for i in range(39)],
    'valid': [os.path.join(work_dir, f"data/dataset_{split}_valid_part{str(i).zfill(3)}.tar") for i in range(file_range)],
    'test': [os.path.join(work_dir, f"data/dataset_{split}_test_part{str(i).zfill(3)}.tar") for i in range(file_range)],
}

# define hdf5 dataset class
class HDF5Dataset(Dataset):
    def __init__(self, hdf5_file_path):
        self.hdf5_file = h5py.File(hdf5_file_path, 'r')
        self.features = self.hdf5_file['features']
        self.labels = self.hdf5_file['labels']
    
    def __len__(self):
        return len(self.features)
    
    def __getitem__(self, idx):
        feature = torch.tensor(self.features[idx], dtype=torch.float32)
        label = torch.tensor(self.labels[idx], dtype=torch.long)
        return feature, label

    def __del__(self):
        self.hdf5_file.close()


def main():
    # check config
    global configs
    print(configs)

    # huggingface login
    # login()

    # get model and transform
    model_name = configs["model_name"]
    model, transform = get_model_transform(model_name)

    # make dataloader
    label_encoder = LabelEncoder()
    label_encoder.fit(df['case'].unique())
    
    train_loader = make_dataloader(batch_size=8, split=split, transform=transform, label_encoder=label_encoder, mode="train")
    valid_loader = make_dataloader(batch_size=8, split=split, transform=transform, label_encoder=label_encoder, mode="valid")
    test_loader = make_dataloader(batch_size=8, split=split, transform=transform, label_encoder=label_encoder, mode="test")

    # set output file (train)
    train_hdf5 = f"features/{model_name}_{split}_train.h5"
    valid_hdf5 = f"features/{model_name}_{split}_valid.h5"
    test_hdf5 = f"features/{model_name}_{split}_test.h5"

    features_dir = os.path.join(work_dir, "features")
    if not os.path.exists(features_dir):
        os.makedirs(features_dir)
        print(f"Created directory: {features_dir}")
    else:
        print(f"Directory already exists: {features_dir}")
        
    # Save Features in HDF5
    model.to(configs["device"])
    if not configs["feature_exist"]:
        for loader, hdf5 in zip([train_loader, valid_loader, test_loader], [train_hdf5, valid_hdf5, test_hdf5]):
            output_file = os.path.join(work_dir, hdf5)
            save_features_to_hdf5_in_batches(model, loader, output_file)
    else:
        for hdf5 in [train_hdf5, valid_hdf5, test_hdf5]:
            output_file = os.path.join(work_dir, hdf5)
            assert os.path.isfile(output_file), f"There is not {output_file}"

    # load feats and labels
    train_feats, train_labels = get_feats_labels(train_hdf5)
    valid_feats, valid_labels = get_feats_labels(valid_hdf5, mode="valid") 
    test_feats, test_labels = get_feats_labels(test_hdf5, mode="test")

    # train and test
    train_eval(train_feats, train_labels, test_feats, test_labels) # if you want to tune parameter, you can use validation data
     
    pass


# define function that load model and transform
def get_model_transform(model_name):
    global configs

    ## The following features were extended from the original bioptimus/H-optimus-0
    if model_name == "h_optimus":  
        model = timm.create_model(
            configs["model_path"], pretrained=True, init_values=1e-5, dynamic_img_size=False
        ) # h-optimus license
        transform = transforms.Compose([
            transforms.Resize(size=(224, 224)), # resize for model
            transforms.ToTensor(),
            transforms.Normalize(
                mean=(0.707223, 0.578729, 0.703617), 
                std=(0.211883, 0.230117, 0.177517)
            ),
        ]) # h-optimus license
    # elif ***:
    # if you want to use other model, please write here
    else: 
        assert False, "This model name cannot be used."
    return model, transform

# define function that make dataloader
def encode_labels(labels, label_encoder):
    return label_encoder.transform(labels).item() # change scalar format


def make_dataloader(batch_size,
                    split,
                    transform,
                    label_encoder,
                    mode="train", 
                    is_all_data_shuffle=True):
    global df, patterns

    if split=="internal":
        buffer_size = len(df[df.split_internal == mode]) if is_all_data_shuffle else 1000 # if OOM, make flag False
    else:
        buffer_size = len(df[df.split_external == mode]) if is_all_data_shuffle else 1000

    def func_transform(image):
        return transform(image)

    dataset = wds.WebDataset(patterns[mode], shardshuffle=False) \
            .shuffle(buffer_size, seed=42) \
            .decode("pil").to_tuple("jpg", "json") \
            .map_tuple(func_transform, lambda x: encode_labels([x["label"]], label_encoder))

    dataloader = DataLoader(dataset, batch_size=batch_size)

    return dataloader

# define function that save feature in hdf5
## The following features were extended from the original h5py
def save_features_to_hdf5_in_batches(model, dataloader, output_file, chunk_size=100, mode="train"):
    global configs, df
    if "internal" in output_file:
        total_iterations = math.ceil(len(df[df.split_internal == mode]) / dataloader.batch_size)
    elif "external" in output_file:
        total_iterations = math.ceil(len(df[df.split_external == mode]) / dataloader.batch_size)
    
    # change model to evaluation mode
    model.eval()

    # Open HDF5 file
    with h5py.File(output_file, 'w') as hdf5_file:
        # Make Feature Dataset (batch_size)
        first_batch = next(iter(dataloader))
        sample_images, _ = first_batch
        with torch.no_grad():
            num_features = model(sample_images.to(configs["device"])).shape[1]
        
        dset_features = hdf5_file.create_dataset('features', shape=(0, num_features), maxshape=(None, num_features), chunks=True)
        dset_labels = hdf5_file.create_dataset('labels', shape=(0,), maxshape=(None,), chunks=True)

        with torch.no_grad():
            for images, labels in tqdm(dataloader, total=total_iterations):
                images = images.to(configs["device"])
                
                # inference feature
                with torch.autocast(device_type=configs["device"], dtype=torch.float16):
                    features = model(images).cpu().numpy() 

                labels = labels.cpu().numpy() 
                
                # add data
                dset_features.resize(dset_features.shape[0] + features.shape[0], axis=0)
                dset_features[-features.shape[0]:] = features 

                dset_labels.resize(dset_labels.shape[0] + labels.shape[0], axis=0)
                dset_labels[-labels.shape[0]:] = labels

                torch.cuda.empty_cache()

    print(f"Features and labels have been saved to {output_file}.")

# define fuction that load feats and labels
def get_feats_labels(hdf5_file_path, mode="train", batch_size=32):
    dataset = HDF5Dataset(hdf5_file_path)
    shuffle = mode=="train"
    dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=shuffle)

    feats_list = []
    labels_list = []

    for feats, labels in dataloader:
        feats_list.append(feats.numpy())
        labels_list.append(labels.numpy())

    all_feats = np.concatenate(feats_list, axis=0)
    all_labels = np.concatenate(labels_list, axis=0)

    all_feats = torch.tensor(all_feats, dtype=torch.float32)
    all_labels = torch.tensor(all_labels, dtype=torch.long)

    return all_feats, all_labels

# training and evaluation
def train_eval(train_feats, train_labels, test_feats, test_labels):
    global configs
    
    # define model, train, evaluation
    if configs["eval_name"] == "logreg":
        model = LogisticRegression(C=configs["cost"], max_iter=configs["max_iter"])
        model.fit(train_feats, train_labels)
        pred = model.predict(test_feats)
        
    if configs["eval_name"] == "knn":
        model = KNeighborsClassifier(n_neighbors=configs["k"])
        model.fit(train_feats.numpy(), train_labels.numpy())
        pred = model.predict(test_feats.numpy())
        test_labels = test_labels.numpy()
        
    if configs["eval_name"] == "proto":
        unique_labels = sorted(np.unique(train_labels.numpy()))
        feats_proto = torch.vstack([
            train_feats[train_labels == c].mean(dim=0) for c in unique_labels
        ])
        labels_proto = torch.tensor(unique_labels)
        pw_dist = (test_feats[:, None] - feats_proto[None, :]).norm(dim=-1, p=2)
        pred = labels_proto[pw_dist.argmin(dim=1)]     

    # result
    acc = accuracy_score(test_labels, pred)
    balanced_acc = balanced_accuracy_score(test_labels, pred)
    print(f"Accuracy = {acc:.3f}, Balanced Accuracy = {balanced_acc:.3f}")

if __name__ == "__main__":
    main()