Mmdetection框架使用
太久没用有点忘了，记录一下
环境篇
环境篇没啥好说的主要是有几个包要注意一下
在原先的环境上多了这几个环境包
powershell
conda create -n mmdet python=3.9 anaconda
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"
https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -v -e .
pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 -i https://pypi.mirrors.ustc.edu.cn/simple
pip install numpy==1.26.4 -i https://pypi.mirrors.ustc.edu.cn/simple
pip install opencv-python==4.8.1.78 opencv-python-headless==4.8.1.78 -i https://pypi.tuna.tsinghua.edu.cn/simple要注意的是有几个包会冲突实测下来这几个能完美兼容（numpy和opencv-python）
脚本转换
yolo2coco
python
# --coding:utf-8--
# 声明文件编码为utf-8
# --coding:utf-8--
# 声明文件编码为utf-8
import os # 导入os模块，用于操作文件和目录
import cv2 # 导入OpenCV库，用于图像处理
import json # 导入json模块，用于处理JSON格式数据
from tqdm import tqdm # 导入tqdm模块，用于显示进度条
import argparse # 导入argparse模块，用于解析命令行参数
classes = ['ship'] # 定义类别列表，当前只有一个类别：'ship'
# 创建ArgumentParser对象，用于解析命令行参数
parser = argparse.ArgumentParser()
parser.add_argument('--image_path', default=r'/root/dataset/images/train', type=str, help="path of images") # 图片路径
parser.add_argument('--label_path', default=r'/root/dataset/labels/train', type=str, help="path of labels .txt") # 标签路径
parser.add_argument('--save_path', type=str, default='/root/dataset/coco-labels/train/train.json', help="if not split the dataset, give a path to a json file") # 保存路径
arg = parser.parse_args() # 解析命令行参数
def yolo2coco(arg):
print("Loading data from ", arg.image_path, arg.label_path) # 打印加载数据的路径信息
# 确保图片路径和标签路径存在
assert os.path.exists(arg.image_path), f"Image path {arg.image_path} does not exist"
assert os.path.exists(arg.label_path), f"Label path {arg.label_path} does not exist"
originImagesDir = arg.image_path # 图片路径
originLabelsDir = arg.label_path # 标签路径
# 获取图片目录下的所有文件名
indexes = os.listdir(originImagesDir)
# 初始化COCO格式的数据结构
dataset = {'categories': [], 'annotations': [], 'images': []}
# 添加类别信息到COCO格式中
for i, cls in enumerate(classes, 0):
dataset['categories'].append({'id': i, 'name': cls, 'supercategory': 'mark'})
ann_id_cnt = 0 # 初始化标注ID计数器
# 遍历图片目录中的每个文件
for k, index in enumerate(tqdm(indexes)):
txtFile = f'{index[:index.rfind(".")]}.txt' # 获取对应的标签文件名
stem = index[:index.rfind(".")] # 获取文件名（不带扩展名）
try:
im = cv2.imread(os.path.join(originImagesDir, index)) # 读取图片
height, width, _ = im.shape # 获取图片的高度、宽度
except Exception as e:
print(f'{os.path.join(originImagesDir, index)} read error.\nerror:{e}') # 如果读取失败，打印错误信息
continue # 跳过该图片
# 如果没有对应的标签文件，则跳过
if not os.path.exists(os.path.join(originLabelsDir, txtFile)):
continue
# 添加图片信息到COCO格式中
dataset['images'].append({
'file_name': index,
'id': stem,
'width': width,
'height': height
})
# 读取标签文件内容
with open(os.path.join(originLabelsDir, txtFile), 'r') as fr:
labelList = fr.readlines() # 按行读取标签内容
# 遍历标签文件中的每一行
for label in labelList:
label = label.strip().split() # 去除空格并分割字段
# 跳过空行（负样本图片的标签文件可能包含空行）
if len(label) == 0:
continue
# 验证标签格式是否正确（YOLO格式应该有5个值：class x y w h）
if len(label) != 5:
print(f"Warning: Invalid label format in {txtFile}: {label}")
continue
try:
x = float(label[1]) # 归一化中心点x坐标
y = float(label[2]) # 归一化中心点y坐标
w = float(label[3]) # 归一化宽度
h = float(label[4]) # 归一化高度
except ValueError as e:
print(f"Warning: Invalid label values in {txtFile}: {label}, error: {e}")
continue
# 将YOLO格式的(x, y, w, h)转换为COCO格式的(x1, y1, x2, y2)
H, W, _ = im.shape
x1 = (x - w / 2) * W
y1 = (y - h / 2) * H
x2 = (x + w / 2) * W
y2 = (y + h / 2) * H
# 计算实际的宽度和高度
width = max(0, x2 - x1)
height = max(0, y2 - y1)
# 添加标注信息到COCO格式中
dataset['annotations'].append({
'area': width * height, # 目标区域面积
'bbox': [x1, y1, width, height], # 边界框坐标
'category_id': int(label[0]), # 类别ID
'id': ann_id_cnt, # 标注ID
'image_id': stem, # 图片ID
'iscrowd': 0, # 是否为人群目标
'segmentation': [[x1, y1, x2, y1, x2, y2, x1, y2]] # 分割信息（矩形顶点）
})
ann_id_cnt += 1 # 更新标注ID计数器
# 将COCO格式的数据保存为JSON文件
folder = os.path.dirname(arg.save_path)
if folder and not os.path.exists(folder):
os.makedirs(folder)
with open(arg.save_path, 'w') as f:
json.dump(dataset, f)
print('Save annotation to {}'.format(arg.save_path))
if __name__ == "__main__":
yolo2coco(arg) # 主程序入口，调用yolo2coco函数改路径就行
训练
powershell
python tools/train.py <your-config-file>验证
powershell
python tools/test.py <your-config-file> <your-model-weights-file> --out <save-pickle-path>mmdet2yolo
python
import os
import torch
import cv2
import math
import tqdm
import argparse
import json
import pickle
import numpy as np
from prettytable import PrettyTable
def clip_boxes(boxes, shape):
# Clip boxes (xyxy) to image shape (height, width)
if isinstance(boxes, torch.Tensor): # faster individually
boxes[..., 0].clamp_(0, shape[1]) # x1
boxes[..., 1].clamp_(0, shape[0]) # y1
boxes[..., 2].clamp_(0, shape[1]) # x2
boxes[..., 3].clamp_(0, shape[0]) # y2
else: # np.array (faster grouped)
boxes[..., [0, 2]] = boxes[..., [0, 2]].clip(0, shape[1]) # x1, x2
boxes[..., [1, 3]] = boxes[..., [1, 3]].clip(0, shape[0]) # y1, y2
def scale_boxes(img1_shape, boxes, img0_shape, ratio_pad=None):
# Rescale boxes (xyxy) from img1_shape to img0_shape
if ratio_pad is None: # calculate from img0_shape
gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1]) # gain = old / new
pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2 # wh padding
else:
gain = ratio_pad[0][0]
pad = ratio_pad[1]
boxes[..., [0, 2]] -= pad[0] # x padding
boxes[..., [1, 3]] -= pad[1] # y padding
boxes[..., :4] /= gain
clip_boxes(boxes, img0_shape)
return boxes
def box_iou(box1, box2, eps=1e-7):
"""
Calculate intersection-over-union (IoU) of boxes. Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
Based on https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
"""
(a1, a2), (b1, b2) = box1.float().unsqueeze(1).chunk(2, 2), box2.float().unsqueeze(0).chunk(2, 2)
inter = (torch.min(a2, b2) - torch.max(a1, b1)).clamp_(0).prod(2)
return inter / ((a2 - a1).prod(2) + (b2 - b1).prod(2) - inter + eps)
def process_batch(detections, labels, iouv):
"""
Return correct prediction matrix
Arguments:
detections (array[N, 6]), x1, y1, x2, y2, conf, class
labels (array[M, 5]), class, x1, y1, x2, y2
Returns:
correct (array[N, 10]), for 10 IoU levels
"""
correct = np.zeros((detections.shape[0], iouv.shape[0])).astype(bool)
iou = box_iou(labels[:, 1:], detections[:, :4])
correct_class = labels[:, 0:1] == detections[:, 5]
for i in range(len(iouv)):
x = torch.where((iou >= iouv[i]) & correct_class) # IoU > threshold and classes match
if x[0].shape[0]:
matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy() # [label, detect, iou]
if x[0].shape[0] > 1:
matches = matches[matches[:, 2].argsort()[::-1]]
matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
correct[matches[:, 1].astype(int), i] = True
return torch.tensor(correct, dtype=torch.bool, device=iouv.device)
def smooth(y, f=0.05):
# Box filter of fraction f
nf = round(len(y) * f * 2) // 2 + 1 # number of filter elements (must be odd)
p = np.ones(nf // 2) # ones padding
yp = np.concatenate((p * y[0], y, p * y[-1]), 0) # y padded
return np.convolve(yp, np.ones(nf) / nf, mode='valid') # y-smoothed
def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=(), eps=1e-16, prefix=''):
""" Compute the average precision, given the recall and precision curves. """
# Sort by objectness
i = np.argsort(-conf)
tp, conf, pred_cls = tp[i], conf[i], pred_cls[i]
# Find unique classes
unique_classes, nt = np.unique(target_cls, return_counts=True)
nc = unique_classes.shape[0] # number of classes, number of detections
# Create Precision-Recall curve and compute AP for each class
px, py = np.linspace(0, 1, 1000), [] # for plotting
ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000))
for ci, c in enumerate(unique_classes):
i = pred_cls == c
n_l = nt[ci] # number of labels
n_p = i.sum() # number of predictions
if n_p == 0 or n_l == 0:
continue
# Accumulate FPs and TPs
fpc = (1 - tp[i]).cumsum(0)
tpc = tp[i].cumsum(0)
# Recall
recall = tpc / (n_l + eps) # recall curve
r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0) # negative x, xp because xp decreases
# Precision
precision = tpc / (tpc + fpc) # precision curve
p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1) # p at pr_score
# AP from recall-precision curve
for j in range(tp.shape[1]):
ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j])
if plot and j == 0:
py.append(np.interp(px, mrec, mpre)) # precision at mAP@0.5
# Compute F1 (harmonic mean of precision and recall)
f1 = 2 * p * r / (p + r + eps)
i = smooth(f1.mean(0), 0.1).argmax() # max F1 index
p, r, f1 = p[:, i], r[:, i], f1[:, i]
tp = (r * nt).round() # true positives
fp = (tp / (p + eps) - tp).round() # false positives
return tp, fp, p, r, f1, ap, unique_classes.astype(int)
def compute_ap(recall, precision):
""" Compute the average precision, given the recall and precision curves """
# Append sentinel values to beginning and end
mrec = np.concatenate(([0.0], recall, [1.0]))
mpre = np.concatenate(([1.0], precision, [0.0]))
# Compute the precision envelope
mpre = np.flip(np.maximum.accumulate(np.flip(mpre)))
# Integrate area under curve
method = 'interp' # methods: 'continuous', 'interp'
if method == 'interp':
x = np.linspace(0, 1, 101) # 101-point interp (COCO)
ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate
else: # 'continuous'
i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes
ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve
return ap, mpre, mrec
def parse_opt():
parser = argparse.ArgumentParser()
parser.add_argument('--label_coco', type=str, default='/root/dataset/coco-labels/test/test.json', help='label coco path')
parser.add_argument('--pred_coco', type=str, default='/root/program/mmdetection-main/tood.pkl', help='pred coco path')
parser.add_argument('--iou', type=float, default=0.7, help='iou threshold')
parser.add_argument('--conf', type=float, default=0.001, help='conf threshold')
parser.add_argument('--img_size', type=int, default=640, help='default image size for unlabeled images') # 新增参数
opt = parser.parse_known_args()[0]
return opt
if __name__ == '__main__':
opt = parse_opt()
iouv = torch.linspace(0.5, 0.95, 10) # iou vector for mAP@0.5:0.95
niou = iouv.numel()
stats = []
# 加载标注数据
label_coco_json_path, pred_coco_json_path = opt.label_coco, opt.pred_coco
with open(label_coco_json_path) as f:
label = json.load(f)
# 获取类别名称
classes = []
for data in label['categories']:
classes.append(data['name'])
# ========== 修正1：构建映射关系 + 收集标注图片ID ==========
image_id_hw_dict = {} # 标注image_id -> (height, width)
image_name_to_id_dict = {} # 图片文件名（无后缀）-> 标注image_id
annotated_image_ids = set() # 所有有标注的image_id集合
for data in label['images']:
image_id = data['id']
image_id_hw_dict[image_id] = [data['height'], data['width']]
# 提取图片文件名（无后缀）
image_file_name = os.path.splitext(os.path.basename(data['file_name']))[0]
image_name_to_id_dict[image_file_name] = image_id
annotated_image_ids.add(image_id)
# ========== 修正2：初始化label_id_dict，包含所有标注image_id（负样本为空列表） ==========
label_id_dict = {img_id: [] for img_id in image_id_hw_dict.keys()}
for data in tqdm.tqdm(label['annotations'], desc='Process label...'):
img_id = data['image_id']
category_id = data['category_id']
x_min, y_min, w, h = data['bbox'][0], data['bbox'][1], data['bbox'][2], data['bbox'][3]
x_max, y_max = x_min + w, y_min + h
label_id_dict[img_id].append(np.array([int(category_id), x_min, y_min, x_max, y_max]))
# ========== 修正3：处理预测数据（改进版） ==========
pred_id_dict = {}
if pred_coco_json_path.endswith('json'):
with open(pred_coco_json_path) as f:
pred = json.load(f)
for data in tqdm.tqdm(pred, desc='Process pred...'):
img_id = data['image_id']
if img_id not in pred_id_dict:
pred_id_dict[img_id] = []
score = data['score']
category_id = data['category_id']
x_min, y_min, w, h = data['bbox'][0], data['bbox'][1], data['bbox'][2], data['bbox'][3]
x_max, y_max = x_min + w, y_min + h
pred_id_dict[img_id].append(np.array([x_min, y_min, x_max, y_max, float(score), int(category_id)]))
else:
with open(pred_coco_json_path, 'rb') as f:
pred = pickle.load(f)
for data in tqdm.tqdm(pred, desc='Process pred...'):
# 提取预测图片的文件名（无后缀）
image_file_name = os.path.splitext(os.path.basename(data['img_path']))[0]
# 对无标注的图片，生成临时唯一ID
if image_file_name in image_name_to_id_dict:
img_id = image_name_to_id_dict[image_file_name]
else:
img_id = f"unlabeled_{image_file_name}"
# 新增：为无标注图片添加默认尺寸信息
if img_id not in image_id_hw_dict:
image_id_hw_dict[img_id] = [opt.img_size, opt.img_size]
# 新增：为无标注图片初始化空的label列表
if img_id not in label_id_dict:
label_id_dict[img_id] = []
if img_id not in pred_id_dict:
pred_id_dict[img_id] = []
# 遍历预测实例
for i in range(data['pred_instances']['labels'].size(0)):
score = data['pred_instances']['scores'][i]
category_id = data['pred_instances']['labels'][i]
bboxes = data['pred_instances']['bboxes'][i]
x_min, y_min, x_max, y_max = bboxes.cpu().detach().numpy()
pred_id_dict[img_id].append(np.array([x_min, y_min, x_max, y_max, float(score), int(category_id)]))
# ========== 修正4：遍历所有图片ID（改进版） ==========
all_image_ids = set(image_id_hw_dict.keys()).union(set(pred_id_dict.keys()))
# 统计信息
total_images = len(all_image_ids)
negative_samples = 0
positive_samples = 0
for image_id in tqdm.tqdm(all_image_ids, desc="Cal mAP..."):
# 1. 获取标注数据
label_data = np.array(label_id_dict.get(image_id, []))
nl = label_data.shape[0]
# 统计负样本和正样本数量
if nl == 0:
negative_samples += 1
else:
positive_samples += 1
# 2. 获取预测数据
if image_id in pred_id_dict:
pred_data = torch.from_numpy(np.array(pred_id_dict[image_id]))
else:
pred_data = torch.empty((0, 6))
npr = pred_data.shape[0]
correct = torch.zeros(npr, niou, dtype=torch.bool)
# 3. 处理不同情况
if npr == 0:
# 无预测的情况
if nl:
# 有标注但无预测，记录为漏检（FN）
stats.append((correct, *torch.zeros((2, 0)), torch.from_numpy(label_data[:, 0])))
# 负样本无预测的情况（TN），跳过（不影响mAP计算）
continue
# 4. 核心逻辑：根据是否有标注来处理
if nl == 0:
# 负样本（无标注）：所有预测都是FP（False Positive）
correct = torch.zeros(npr, niou, dtype=torch.bool)
else:
# 正样本（有标注）：通过IoU匹配来判断TP/FP
correct = process_batch(pred_data, torch.from_numpy(label_data), iouv)
# 5. 处理target_cls
target_cls = torch.from_numpy(label_data[:, 0]) if nl > 0 else torch.tensor([], dtype=torch.int64)
stats.append((correct, pred_data[:, 4], pred_data[:, 5], target_cls))
# 打印统计信息
print(f"\n========== 数据集统计 ==========")
print(f"总图片数: {total_images}")
print(f"正样本数（有标注）: {positive_samples}")
print(f"负样本数（无标注）: {negative_samples}")
print(f"负样本比例: {negative_samples/total_images*100:.2f}%")
print(f"================================\n")
# 计算mAP
if len(stats) == 0:
print("警告：无有效数据用于计算mAP")
else:
# 合并统计数据
stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)]
if len(stats) >= 4 and stats[0].size > 0:
tp, fp, p, r, f1, ap, ap_class = ap_per_class(*stats)
print(f'precision: {p}')
print(f'recall: {r}')
print(f'mAP@0.5: {ap[:, 0]}')
# 生成结果表格
table = PrettyTable()
table.title = "Metrics"
table.field_names = ["Classes", 'Precision', 'Recall', 'mAP50', 'mAP50-95']
table.add_row(['all', f'{np.mean(p):.3f}', f'{np.mean(r):.3f}', f'{np.mean(ap[:, 0]):.3f}', f'{np.mean(ap):.3f}'])
for cls_idx, cls_name in enumerate(classes):
if cls_idx < len(p):
table.add_row([cls_name, f'{p[cls_idx]:.3f}', f'{r[cls_idx]:.3f}', f'{ap[cls_idx, 0]:.3f}', f'{ap[cls_idx, :].mean():.3f}'])
else:
table.add_row([cls_name, '0.000', '0.000', '0.000', '0.000'])
print(table)
else:
print("警告：统计数据不足，无法计算mAP")