#!/usr/bin/env python
# -*- coding:utf-8-*-
# file: model_test1.py
# @author: jory.d
# @contact:
# @time: 2022/01/07 22:41
# @desc: 模型测试, 查看误检和漏检
"""
python tests/model_test1_qr_code.py
"""
import os
import os.path as osp
import copy
import json
import cv2
import numpy as np
import torch
from mmcv import Config
from mmcv.cnn import fuse_conv_bn
from mmcv.runner import (load_checkpoint)
from mmdet.core.bbox.iou_calculators import bbox_overlaps
from mmdet.datasets import build_dataset, build_dataloader
from mmdet.models import build_detector
BATCH_SIZE = 1
THRESHOLD = 0.6
CONFIG = "configs/xx/cfg_2_ssd_128x128.py"
WORK_DIR = "train_result/20221011_cfg_2_ssd_128x128-11"
EPOCH = 1340
CKPT = f"{WORK_DIR}/epoch_{EPOCH}.pth"
SAVE_DIR = f"{WORK_DIR}/test_2"
SAVE_IMG_FLAG = 1
TP_ASSIGN_IOU = 0.65
def build_model():
cfg = Config.fromfile(CONFIG)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
_ckpt = CKPT
if not osp.isfile(_ckpt) or not osp.exists(_ckpt):
raise FileNotFoundError(f'{_ckpt} is not existed.')
# build the dataloader
dataset = build_dataset(cfg.data.test)
data_loader = build_dataloader(
dataset,
samples_per_gpu=BATCH_SIZE,
workers_per_gpu=0,
dist=False,
shuffle=False)
# build the model and load checkpoint
cfg.model.train_cfg = None
_test_cfg = cfg.get('test_cfg')
if _test_cfg is None:
_test_cfg = cfg.model.get('test_cfg')
assert _test_cfg is not None
cfg.model['test_cfg']['score_thr'] = 0.2
cfg.model['test_cfg']['nms']['iou_threshold'] = 0.5
model = build_detector(cfg.model)
checkpoint = load_checkpoint(model, _ckpt, map_location='cpu')
model = fuse_conv_bn(model)
model.eval()
# old versions did not save class info in checkpoints, this walkaround is
# for backward compatibility
if 'CLASSES' in checkpoint.get('meta', {}):
model.CLASSES = checkpoint['meta']['CLASSES']
else:
model.CLASSES = dataset.CLASSES
return data_loader, model
def dummy_img():
INPUT_W, INPUT_H,INPUT_C = 128,128,3
MEAN = [128., ] * INPUT_C
STD = [128., ] * INPUT_C
# read pic
img0 = cv2.imread("./14.png", cv2.IMREAD_COLOR)
# img0 = np.random.randint(0, 255, [INPUT_H, INPUT_W, 3], dtype=np.uint8)
h, w = img0.shape[:2]
# ========================== preprocess img =============================
img1 = cv2.resize(img0, (INPUT_W, INPUT_H))
ori_image = cv2.cvtColor(img1, cv2.COLOR_BGR2RGB)
# ori_image = img1
if 1 == INPUT_C:
ori_image = cv2.cvtColor(ori_image, cv2.COLOR_RGB2GRAY)
input_image = ori_image.astype(np.float32) - np.asarray(MEAN)
input_image /= np.asarray(STD)
input_image = input_image[np.newaxis, ...]
else:
# [h,w,c]
input_image = ori_image.astype(np.float32) - np.asarray([[MEAN]])
input_image /= np.asarray([[STD]])
input_image = np.transpose(input_image, [2, 0, 1])
img_c, img_h, img_w = input_image.shape
img_data = input_image[np.newaxis, :, :, :]
return img_data
@torch.no_grad()
def model_test():
_dataloader, _model = build_model()
_num_class = len(_model.CLASSES)
miss_num = {i: 0 for i in range(_num_class)}
error_num, true_num = copy.deepcopy(miss_num), copy.deepcopy(miss_num)
total_num = copy.deepcopy(miss_num)
ignore_num = 0
for i, data in enumerate(_dataloader):
# if i >=1: break
img_metas = data['img_metas'].data[0]
img_tensors = data['img'].data[0]
b, c, h, w = img_tensors.shape
assert len(img_metas) == b
print(img_tensors.dtype)
a = np.asarray(dummy_img(), dtype=np.float32)
img_tensors = torch.from_numpy(a)
print(img_tensors)
results = _model(return_loss=False, rescale=False, img=[img_tensors], img_metas=[img_metas])
print(results)
exit(10)
for jj, meta in enumerate(img_metas):
# print(meta)
img_filepath = meta['filename']
fname = osp.basename(img_filepath)
ori_shape = meta['ori_shape'] # [h,w,c]
mean = meta['img_norm_cfg']['mean']
std = meta['img_norm_cfg']['std']
gt_boxes = meta['gt_bboxes'].data
gt_boxes_ignore = meta['gt_bboxes_ignore'].data
gt_labels = meta['gt_labels'].data
det_result = results[jj] #[class_0_res, class_1_res, ...]
# print('det_result: ', det_result)
# print('det_result size: ', det_result)
# print('gt_boxes: ', gt_boxes.numpy())
for k, gt_label in enumerate(gt_labels.numpy()):
total_num[gt_label] += 1
ignore_num += gt_boxes_ignore.shape[0]
miss_labels, error_labels = [], []
miss_boxes, error_boxes = [], []
true_labels, true_boxes = [], []
if len(det_result) == 0:
for k, gt_label in enumerate(gt_labels.numpy()):
miss_num[gt_label] += 1
miss_labels.append(gt_label)
miss_boxes.append(gt_boxes[k].numpy())
continue
for class_idx, det_bboxes in enumerate(det_result):
det_bboxes = np.asarray(det_bboxes) # [n,5]
det_bboxes_src = det_bboxes[det_bboxes[:, -1] >= THRESHOLD]
det_bboxes = det_bboxes_src[:, :4]
_inds = (gt_labels == class_idx)
_gt_boxes = gt_boxes[_inds]
_gt_labels = gt_labels[_inds]
# print('_gt_labels: ', _gt_labels)
# print('_gt_boxes: ', _gt_boxes)
# print(f'cls: {class_idx}, det_bboxes: {det_bboxes}')
# continue
det_num = det_bboxes.shape[0]
gt_num = _gt_boxes.shape[0]
gt_num2 = gt_boxes_ignore.shape[0]
if det_num == 0 and gt_num > 0:
miss_num[class_idx] += 1
miss_labels.append(_gt_labels.numpy())
miss_boxes.append(_gt_boxes.numpy())
elif gt_num == 0 and gt_num2 == 0 and det_num > 0:
error_num[class_idx] += 1
error_labels.append([class_idx for _ in range(det_num)])
error_boxes.append(det_bboxes_src)
else:
_ious = bbox_overlaps(torch.from_numpy(det_bboxes), _gt_boxes, mode='iou') # [n,m]
# print('_ious: ', _ious)
assign_inds = []
_tp_labels, _fp_labels = [], []
_tp_boxes, _fp_boxes = [], []
for i, iou_d2gts in enumerate(_ious):
is_assign = False
# foreach ious of det_box and gt_boxes
for j, iou in enumerate(iou_d2gts):
if iou >= TP_ASSIGN_IOU and class_idx == _gt_labels[j]:
is_assign = True
assign_inds.append(j)
_tp_labels.append(class_idx)
_tp_boxes.append(det_bboxes_src[i])
break
if not is_assign:
_fp_labels.append(class_idx)
_fp_boxes.append(det_bboxes_src[i])
# print(_tp_labels, _tp_boxes)
if len(_tp_labels)>0:
true_num[class_idx] += len(_tp_labels)
true_labels.append(_tp_labels)
true_boxes.append(_tp_boxes)
if len(_fp_labels)>0:
error_num[class_idx] += len(_fp_labels)
error_labels.append(_fp_labels)
error_boxes.append(_fp_boxes)
all_gt_inds = [i for i in range(_gt_boxes.size(0))]
if len(assign_inds) < len(all_gt_inds):
miss_inds = list(set(all_gt_inds).difference(set(assign_inds)))
if len(miss_inds) > 0:
miss_num[class_idx] += len(miss_inds)
miss_labels.append(_gt_labels[miss_inds].numpy())
miss_boxes.append(_gt_boxes[miss_inds].numpy())
print(f'miss_num: {miss_num}', f'error_num: {error_num}', f'true_num: {true_num}',
f'total_num: {total_num}')
_error_flag = len(error_labels) > 0
_miss_flag = len(miss_labels) > 0
_true_flag = len(true_labels) > 0
# if not _error_flag and not _miss_flag:
# continue
if _error_flag:
# print('error_labels: ', error_labels)
# print('error_boxes: ', error_boxes)
error_labels = np.asarray(error_labels).reshape(-1)
error_boxes = np.asarray(error_boxes).reshape(-1, 5)
if _miss_flag:
# print('miss_labels: ', miss_labels)
# print('miss_boxes: ', miss_boxes)
miss_labels = np.asarray(miss_labels).reshape(-1)
miss_boxes = np.asarray(miss_boxes).reshape(-1, 4)
if len(true_labels) > 0:
true_labels = np.asarray(true_labels).reshape(-1)
true_boxes = np.asarray(true_boxes).reshape(-1, 5)
if SAVE_IMG_FLAG:
img = img_tensors[jj].permute(1, 2, 0) # [c,h,w]
img = img * std + mean
_ch = img.size(-1)
img = np.asarray(img.numpy(), dtype=np.uint8)
if _ch == 1:
img = cv2.cvtColor(np.squeeze(img), cv2.COLOR_GRAY2BGR)
else:
img = img[:, :, ::-1]
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
h, w = img.shape[:2]
for idx, (cls, box) in enumerate(zip(gt_labels, gt_boxes)):
x1, y1, x2, y2 = list(map(int, box))
cls = int(cls)
cv2.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 1)
cv2.putText(img, f"cls:{cls}", (x1, max(5, y1 - 5)),
cv2.FONT_HERSHEY_COMPLEX, 0.5, (0, 255, 0), 1)
for idx, (cls, box) in enumerate(zip(true_labels, true_boxes)):
x1, y1, x2, y2 = list(map(int, box[:4]))
p = box[-1]
cls = int(cls)
cv2.rectangle(img, (x1, y1), (x2, y2), (255, 0, 0), 1)
# print(f'============= {x1,y1, x2,y2, p, cls}')
cv2.putText(img, "cls:{}({:.2f})".format(cls, p), (x1, min(h - 10, y2 + 5)),
cv2.FONT_HERSHEY_COMPLEX, 0.5, (255, 0, 0), 1)
for idx, (cls, box) in enumerate(zip(miss_labels, miss_boxes)):
x1, y1, x2, y2 = list(map(int, box))
cv2.rectangle(img, (x1, y1), (x2, y2), (255, 255, 0), 3)
for idx, (cls, box) in enumerate(zip(error_labels, error_boxes)):
x1, y1, x2, y2 = list(map(int, box[:4]))
p = box[-1]
cls = int(cls)
cv2.rectangle(img, (x1, y1), (x2, y2), (0, 0, 255), 1)
cv2.putText(img, "cls:{}({:.2f})".format(cls, p), (x1, min(h - 10, y2 + 5)),
cv2.FONT_HERSHEY_COMPLEX, 0.5, (0, 0, 255), 1)
img = cv2.resize(img, dsize=(0, 0), fx=2, fy=2)
if _miss_flag:
_savefilepath = f'{SAVE_DIR}/miss/{fname}'
os.makedirs(osp.dirname(_savefilepath), exist_ok=True)
cv2.imwrite(_savefilepath, img)
if _error_flag:
_savefilepath = f'{SAVE_DIR}/error/{fname}'
os.makedirs(osp.dirname(_savefilepath), exist_ok=True)
cv2.imwrite(_savefilepath, img)
if _true_flag:
_savefilepath = f'{SAVE_DIR}/true/{fname}'
os.makedirs(osp.dirname(_savefilepath), exist_ok=True)
cv2.imwrite(_savefilepath, img)
#####################################################################
_report_json_path = f'{SAVE_DIR}/report.json'
_data = {i: 0 for i in range(_num_class)}
for i in range(_num_class):
_data[i] = {
"total": total_num[i],
"ignore": ignore_num,
"miss": miss_num[i],
"miss_p": miss_num[i] / total_num[i],
"error": error_num[i],
"error_p": error_num[i] / total_num[i],
"true": true_num[i],
"true_p": true_num[i] / total_num[i],
}
from pprint import pprint
pprint(_data)
os.makedirs(osp.dirname(_report_json_path), exist_ok=True)
with open(_report_json_path, 'w', encoding='utf-8') as f:
json.dump(_data, f, indent=4)
print('done.')
if __name__ == '__main__':
model_test()
标签：误检率,num,img,labels,漏检,正确率,boxes,gt,miss From： https://www.cnblogs.com/dxscode/p/16802672.html