evaluate.py

import torch
import argparse
from torch.utils.data import DataLoader
import os
import pandas as pd
import time

from dataset import AdvDataset
from model import get_model
from utils import BASE_ADV_PATH, accuracy, AverageMeter

def arg_parse():
    parser = argparse.ArgumentParser(description='')
    parser.add_argument('--adv_path', type=str, default='', help='the path of adversarial examples.')
    parser.add_argument('--gpu', type=str, default='0', help='gpu device.')
    parser.add_argument('--batch_size', type=int, default=20, metavar='N',
                    help='input batch size for reference (default: 16)')
    parser.add_argument('--model_name', type=str, default='', help='')
    args = parser.parse_args()
    args.adv_path = os.path.join(BASE_ADV_PATH, args.adv_path)
    return args

if __name__ == '__main__':
    args = arg_parse()
    
    os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu

    # Loading dataset
    dataset = AdvDataset(args.model_name, args.adv_path)
    data_loader = DataLoader(dataset, batch_size=args.batch_size, shuffle=False, num_workers=4)
    print (len(dataset))
    # Loading model
    model = get_model(args.model_name)
    model.cuda()
    model.eval()

    # main 
    top1 = AverageMeter()
    top5 = AverageMeter()
    batch_time = AverageMeter()

    prediction = []
    gts = []
    with torch.no_grad():
        end = time.time()
        for batch_idx, batch_data in enumerate(data_loader):
            if batch_idx%10 == 0:
                print ('Ruing batch_idx', batch_idx)
            batch_x = batch_data[0].cuda()
            batch_y = batch_data[1].cuda()
            batch_name = batch_data[2]

            output = model(batch_x)
            acc1, acc5 = accuracy(output.detach(), batch_y, topk=(1, 5))
            top1.update(acc1.item(), batch_x.size(0))
            top5.update(acc5.item(), batch_x.size(0))

            batch_time.update(time.time() - end)
            end = time.time()

            _, pred = output.detach().topk(1, 1, True, True)
            pred = pred.t()
            prediction += list(torch.squeeze(pred.cpu()).numpy())
            gts += list(batch_y.cpu().numpy())

    df = pd.DataFrame(columns = ['path', 'pre', 'gt'])
    df['path'] = dataset.paths[:len(prediction)]
    df['pre'] = prediction
    df['gt'] = gts
    df.to_csv(os.path.join(args.adv_path, 'prediction-model_{}-top1_{:.3f}.csv'.format(args.model_name, top1.avg)), index=False)