eval.py

from __future__ import print_function

import numpy as np

import argparse
import torch
import torch.nn as nn
import pdb
import os
import pandas as pd
from utils.utils import *
from math import floor
import matplotlib.pyplot as plt
from dataset_modules.dataset_generic import Generic_WSI_Classification_Dataset, Generic_MIL_Dataset, save_splits
import h5py
from utils.eval_utils import *
from utils.task_utils import get_dataset_args, tasks

# Training settings
parser = argparse.ArgumentParser(description='CLAM Evaluation Script')
parser.add_argument('--data_root_dir', type=str, default='/scratch90/taghinia/pave_training',
                    help='data directory')
parser.add_argument('--results_dir', type=str, default='./results',
                    help='relative path to results folder, i.e. '+
                    'the directory containing models_exp_code relative to project root (default: ./results)')
parser.add_argument('--save_exp_code', type=str, default=None,
                    help='experiment code to save eval results')
parser.add_argument('--models_exp_code', type=str, default=None,
                    help='experiment code to load trained models (directory under results_dir containing model checkpoints')
parser.add_argument('--splits_dir', type=str, default=None,
                    help='splits directory, if using custom splits other than what matches the task (default: None)')
parser.add_argument('--model_size', type=str, choices=['small', 'big'], default='small', 
                    help='size of model (default: small)')
parser.add_argument('--model_type', type=str, choices=['clam_sb', 'clam_mb', 'mil', 'sao'], default='clam_sb', 
                    help='type of model (default: clam_sb)')
parser.add_argument('--use_pos_embed', action='store_true', default=False, help='Enable positional embeddings')
parser.add_argument('--k', type=int, default=1, help='number of folds (default: 1)')
parser.add_argument('--k_start', type=int, default=-1, help='start fold (default: -1, last fold)')
parser.add_argument('--k_end', type=int, default=-1, help='end fold (default: -1, first fold)')
parser.add_argument('--fold', type=int, default=-1, help='single fold to evaluate')
parser.add_argument('--micro_average', action='store_true', default=False, 
                    help='use micro_average instead of macro_avearge for multiclass AUC')
parser.add_argument('--split', type=str, choices=['train', 'val', 'test', 'all'], default='val')
parser.add_argument('--task', type=str, choices=list(tasks.keys()), required=True)
parser.add_argument('--seed', type=int, default=1, help='random seed for reproducible experiment (default: 1)')
parser.add_argument('--drop_out', type=float, default=0.25, help='dropout')
parser.add_argument('--embed_dim', type=int, default=1024)
parser.add_argument('--threshold', type=float, default=None, help='decision threshold for binary classification')
parser.add_argument('--temperature_optimize', action='store_true', default=False,
                    help='optimize temperature scaling on validation set for probability calibration (default: False)')
parser.add_argument('--temperature', type=float, default=1.0, help='temperature for scaling (default: 1.0)')

args = parser.parse_args()

device=torch.device("cuda" if torch.cuda.is_available() else "cpu")

args.save_dir = os.path.join('./eval_results', 'EVAL_' + str(args.save_exp_code))
args.models_dir = os.path.join(args.results_dir, str(args.models_exp_code))

os.makedirs(args.save_dir, exist_ok=True)

if args.splits_dir is None:
    args.splits_dir = args.models_dir
else:
    args.splits_dir = os.path.join('splits', str(args.splits_dir))

assert os.path.isdir(args.models_dir)
assert os.path.isdir(args.splits_dir)

if args.task:
    dataset_args = get_dataset_args(args, args.task)
    dataset = Generic_MIL_Dataset(**dataset_args)
else:
    raise NotImplementedError

if args.n_classes > 2 and args.model_type != 'sao':
    args.model_type = "clam_mb"

settings = {'task': args.task,
            'split': args.split,
            'save_dir': args.save_dir, 
            'models_dir': args.models_dir,
            'model_type': args.model_type,
            'drop_out': args.drop_out,
            'model_size': args.model_size,
            'threshold': args.threshold,
            'temperature': args.temperature,
            'temperature_optimize': args.temperature_optimize}

with open(args.save_dir + '/eval_experiment_{}.txt'.format(args.save_exp_code), 'w') as f:
    print(settings, file=f)
f.close()

print(settings)

if args.k_start == -1:
    start = 0
else:
    start = args.k_start
if args.k_end == -1:
    end = args.k
else:
    end = args.k_end

if args.fold == -1:
    folds = range(start, end)
else:
    folds = range(args.fold, args.fold+1)
ckpt_paths = [os.path.join(args.models_dir, 's_{}_checkpoint.pt'.format(fold)) for fold in folds]
datasets_id = {'train': 0, 'val': 1, 'test': 2, 'all': -1}

if __name__ == "__main__":
    all_results = []
    all_auc = []
    all_acc = []

    for ckpt_idx in range(len(ckpt_paths)):
        if datasets_id[args.split] < 0:
            split_dataset = dataset
        else:
            csv_path = '{}/splits_{}.csv'.format(args.splits_dir, folds[ckpt_idx])
            datasets = dataset.return_splits(from_id=False, csv_path=csv_path)
            split_dataset = datasets[datasets_id[args.split]]
        model, patient_results, test_error, auc, df, acc_logger  = eval(split_dataset, args, ckpt_paths[ckpt_idx])
        all_results.append(all_results)
        all_auc.append(auc)
        all_acc.append(1-test_error)
        df.to_csv(os.path.join(args.save_dir, 'fold_{}.csv'.format(folds[ckpt_idx])), index=False)

    final_df = pd.DataFrame({'folds': folds, 'test_auc': all_auc, 'test_acc': all_acc})
    if len(folds) != args.k:
        save_name = 'summary_partial_{}_{}.csv'.format(folds[0], folds[-1])
    else:
        save_name = 'summary.csv'
    final_df.to_csv(os.path.join(args.save_dir, save_name))