mindspore/model_zoo/official/cv/deeplabv3/postprocess.py

# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""post process for 310 inference"""
import os
import argparse
import numpy as np
import cv2

from eval import cal_hist, pre_process

def parse_args():
    parser = argparse.ArgumentParser(description="deeplabv3 accuracy calculation")
    parser.add_argument('--data_root', type=str, default='', help='root path of val data')
    parser.add_argument('--data_lst', type=str, default='', help='list of val data')
    parser.add_argument('--batch_size', type=int, default=1, help='batch size')
    parser.add_argument('--crop_size', type=int, default=513, help='crop size')
    parser.add_argument('--scales', type=float, action='append', help='scales of evaluation')
    parser.add_argument('--flip', action='store_true', help='perform left-right flip')
    parser.add_argument('--ignore_label', type=int, default=255, help='ignore label')
    parser.add_argument('--num_classes', type=int, default=21, help='number of classes')
    parser.add_argument('--result_path', type=str, default='./result_Files', help='result Files path')
    args, _ = parser.parse_known_args()
    return args

def eval_batch(args, result_file, img_lst, crop_size=513, flip=True):
    result_lst = []
    batch_size = len(img_lst)
    batch_img = np.zeros((args.batch_size, 3, crop_size, crop_size), dtype=np.float32)
    resize_hw = []
    for l in range(batch_size):
        img_ = img_lst[l]
        img_, resize_h, resize_w = pre_process(args, img_, crop_size)
        batch_img[l] = img_
        resize_hw.append([resize_h, resize_w])

    batch_img = np.ascontiguousarray(batch_img)
    net_out = np.fromfile(result_file, np.float32).reshape(args.batch_size, args.num_classes, crop_size, crop_size)

    for bs in range(batch_size):
        probs_ = net_out[bs][:, :resize_hw[bs][0], :resize_hw[bs][1]].transpose((1, 2, 0))
        ori_h, ori_w = img_lst[bs].shape[0], img_lst[bs].shape[1]
        probs_ = cv2.resize(probs_, (ori_w, ori_h))
        result_lst.append(probs_)

    return result_lst


def eval_batch_scales(args, eval_net, img_lst, scales,
                      base_crop_size=513, flip=True):
    sizes_ = [int((base_crop_size - 1) * sc) + 1 for sc in scales]
    probs_lst = eval_batch(args, eval_net, img_lst, crop_size=sizes_[0], flip=flip)
    print(sizes_)
    for crop_size_ in sizes_[1:]:
        probs_lst_tmp = eval_batch(args, eval_net, img_lst, crop_size=crop_size_, flip=flip)
        for pl, _ in enumerate(probs_lst):
            probs_lst[pl] += probs_lst_tmp[pl]

    result_msk = []
    for i in probs_lst:
        result_msk.append(i.argmax(axis=2))
    return result_msk


def acc_cal():
    args = parse_args()
    args.image_mean = [103.53, 116.28, 123.675]
    args.image_std = [57.375, 57.120, 58.395]
    # data list
    with open(args.data_lst) as f:
        img_lst = f.readlines()
    # evaluate
    hist = np.zeros((args.num_classes, args.num_classes))
    batch_img_lst = []
    batch_msk_lst = []
    bi = 0
    image_num = 0
    for i, line in enumerate(img_lst):
        img_path, msk_path = line.strip().split(' ')
        result_file = os.path.join(args.result_path, os.path.basename(img_path).split('.jpg')[0] + '_0.bin')
        img_path = os.path.join(args.data_root, img_path)
        msk_path = os.path.join(args.data_root, msk_path)
        img_ = cv2.imread(img_path)
        msk_ = cv2.imread(msk_path, cv2.IMREAD_GRAYSCALE)
        batch_img_lst.append(img_)
        batch_msk_lst.append(msk_)
        bi += 1
        if bi == args.batch_size:
            batch_res = eval_batch_scales(args, result_file, batch_img_lst, scales=args.scales,
                                          base_crop_size=args.crop_size, flip=args.flip)
            for mi in range(args.batch_size):
                hist += cal_hist(batch_msk_lst[mi].flatten(), batch_res[mi].flatten(), args.num_classes)

            bi = 0
            batch_img_lst = []
            batch_msk_lst = []
            print('processed {} images'.format(i+1))
        image_num = i

    if bi > 0:
        batch_res = eval_batch_scales(args, result_file, batch_img_lst, scales=args.scales,
                                      base_crop_size=args.crop_size, flip=args.flip)
        for mi in range(bi):
            hist += cal_hist(batch_msk_lst[mi].flatten(), batch_res[mi].flatten(), args.num_classes)
        print('processed {} images'.format(image_num + 1))

    print(hist)
    iu = np.diag(hist) / (hist.sum(1) + hist.sum(0) - np.diag(hist))
    print('per-class IoU', iu)
    print('mean IoU', np.nanmean(iu))

if __name__ == '__main__':
    acc_cal()