训练DETR
一、数据准备二、配置DETR三、绘图四、推理五、一些小bugReferences
一、数据准备
DETR用的是COCO格式的数据集。
如果要用DETR训练自己的数据集，直接利用Labelimg标注成COCO格式。
如果是VOC数据集的话，要做一个格式转换。网上一大堆格式转换的代码都很乱，所以自己写了一个针对VOC数据集的转换。
COCO数据集的格式类似这样，annotations文件夹里面有对应的train、val数据集的json文件。train2017则是训练集图片，其他同理。
VOC数据集的存放方式是这样的，转换格式就是找出Main文件夹下用于目标检测的图片。
Main文件夹下有train.txt文件，记录了训练集的图片。val.txt记录了验证集的图片
只需要修改注释中的两个路径即可（创建文件夹时没有加判断语句严谨一点应该加上）。
import osimport shutilimport sysimport jsonimport globimport xml.etree.ElementTree as ETSTART_BOUNDING_BOX_ID = 1# PRE_DEFINE_CATEGORIES = None# If necessary, pre-define category and its idPRE_DEFINE_CATEGORIES = {"aeroplane": 1, "bicycle": 2, "bird": 3, "boat": 4, "bottle": 5, "bus": 6, "car": 7, "cat": 8, "chair": 9, "cow": 10, "diningtable": 11, "dog": 12, "horse": 13, "motorbike": 14, "person": 15, "pottedplant": 16, "sheep": 17, "sofa": 18, "train": 19, "tvmonitor": 20}def get(root, name): vars = root.findall(name) return varsdef get_and_check(root, name, length): vars = root.findall(name) if len(vars) == 0: raise ValueError("Can not find %s in %s." % (name, root.tag)) if length > 0 and len(vars) != length: raise ValueError( "The size of %s is supposed to be %d, but is %d." % (name, length, len(vars)) ) if length == 1: vars = vars[0] return varsdef get_filename_as_int(filename): try: filename = filename.replace("\\", "/") filename = os.path.splitext(os.path.basename(filename))[0] return int(filename) except: raise ValueError( "Filename %s is supposed to be an integer." % (filename))def get_categories(xml_files): """Generate category name to id mapping from a list of xml files. Arguments: xml_files {list} -- A list of xml file paths. Returns: dict -- category name to id mapping. """ classes_names = [] for xml_file in xml_files: tree = ET.parse(xml_file) root = tree.getroot() for member in root.findall("object"): classes_names.append(member[0].text) classes_names = list(set(classes_names)) classes_names.sort() return {name: i for i, name in enumerate(classes_names)}def convert(xml_files, json_file): json_dict = {"images": [], "type": "instances", "annotations": [], "categories": []} if PRE_DEFINE_CATEGORIES is not None: categories = PRE_DEFINE_CATEGORIES else: categories = get_categories(xml_files) bnd_id = START_BOUNDING_BOX_ID for xml_file in xml_files: tree = ET.parse(xml_file) root = tree.getroot() path = get(root, "path") if len(path) == 1: filename = os.path.basename(path[0].text) elif len(path) == 0: filename = get_and_check(root, "filename", 1).text else: raise ValueError("%d paths found in %s" % (len(path), xml_file)) # The filename must be a number image_id = get_filename_as_int(filename) size = get_and_check(root, "size", 1) width = int(get_and_check(size, "width", 1).text) height = int(get_and_check(size, "height", 1).text) image = { "file_name": filename, "height": height, "width": width, "id": image_id, } json_dict["images"].append(image) # Currently we do not support segmentation. # segmented = get_and_check(root, 'segmented', 1).text # assert segmented == '0' for obj in get(root, "object"): category = get_and_check(obj, "name", 1).text if category not in categories: new_id = len(categories) categories[category] = new_id category_id = categories[category] bndbox = get_and_check(obj, "bndbox", 1) xmin = int(get_and_check(bndbox, "xmin", 1).text) - 1 ymin = int(get_and_check(bndbox, "ymin", 1).text) - 1 xmax = int(get_and_check(bndbox, "xmax", 1).text) ymax = int(get_and_check(bndbox, "ymax", 1).text) assert xmax > xmin assert ymax > ymin o_width = abs(xmax - xmin) o_height = abs(ymax - ymin) ann = { "area": o_width * o_height, "iscrowd": 0, "image_id": image_id, "bbox": [xmin, ymin, o_width, o_height], "category_id": category_id, "id": bnd_id, "ignore": 0, "segmentation": [], } json_dict["annotations"].append(ann) bnd_id = bnd_id + 1 for cate, cid in categories.items(): cat = {"supercategory": "none", "id": cid, "name": cate} json_dict["categories"].append(cat) os.makedirs(os.path.dirname(json_file), exist_ok=True) json_fp = open(json_file, "w") json_str = json.dumps(json_dict) json_fp.write(json_str) json_fp.close()if __name__ == "__main__": # 只需修改以下两个路径 # VOC数据集根目录 voc_path = "VOC2012" # 保存coco格式数据集根目录 save_coco_path = "VOC2COCO" # VOC只分了训练集和验证集即train.txt和val.txt data_type_list = ["train", "val"] for data_type in data_type_list: os.makedirs(os.path.join(save_coco_path, data_type+"2017")) os.makedirs(os.path.join(save_coco_path, data_type+"_xml")) with open(os.path.join(voc_path, "ImageSets\Main", data_type+".txt"), "r") as f: txt_ls = f.readlines() txt_ls = [i.strip() for i in txt_ls] for i in os.listdir(os.path.join(voc_path, "JPEGImages")): if os.path.splitext(i)[0] in txt_ls: shutil.copy(os.path.join(voc_path, "JPEGImages", i), os.path.join(save_coco_path, data_type+"2017", i)) shutil.copy(os.path.join(voc_path, "Annotations", i[:-4]+".xml"), os.path.join( save_coco_path, data_type+"_xml", i[:-4]+".xml")) xml_path = os.path.join(save_coco_path, data_type+"_xml") xml_files = glob.glob(os.path.join(xml_path, "*.xml")) convert(xml_files, os.path.join(save_coco_path, "annotations", "instances_"+data_type+"2017.json")) shutil.rmtree(xml_path)结果如图所示，在voc2coco文件夹下有三个文件：
二、配置DETR
修改main.py文件中的参数、超参数：
这个最好不改，就设为coco。去修改models/detr.py 文件的num_classes(大概在三百多行)。这里作者也解释了num_classes其实并不是类别数，因为coco只有80类，因为coco的id是不连续的，coco数据集最大的ID是90，所以原论文时写的MAX ID +1 即91。对于我们自定义的和转化的VOC数据集num_classes就是类别数。
coco_path改成自己的coco路径。
其中预训练权重需要修改一下，coco是80类，不能直接加载官方的模型。voc是20类。把num_classes改成21。传入得到的detr_r50_21.pth新的权重文件。
import torchpretrained_weights=torch.load('detr-r50-e632da11.pth')num_classes=21pretrained_weights["model"]["class_embed.weight"].resize_(num_classes+1,256)pretrained_weights["model"]["class_embed.bias"].resize_(num_classes+1)torch.save(pretrained_weights,"detr_r50_%d.path"%num_classes)运行日志（特别难训练）：
三、绘图
在util文件夹下有plot_utils.py文件，可以绘制损失和mAP曲线。
在plot_utils.py文件中加入代码运行即可：
if __name__ == "__main__":# 路径更换为保存输出的eval路径# mAP曲线 files=list(Path("./outputs/eval").glob("*.pth")) plot_precision_recall(files) plt.show() # 路径更换为保存输出的路径 # 损失曲线 plot_logs(Path("./output")) plt.show()四、推理
训练完毕后我们会得到一个checkpoint.pth的文件，可以用自己训练得到的模型来推理图片，代码如下：
import argparseimport numpy as npfrom models.detr import DETRfrom models.backbone import Backbone, build_backbonefrom models.transformer import build_transformerfrom PIL import Imageimport cv2import matplotlib.pyplot as pltimport torchimport torchvision.transforms as Ttorch.set_grad_enabled(False)def get_args_parser(): parser = argparse.ArgumentParser('Set transformer detector', add_help=False) parser.add_argument('--lr', default=1e-4, type=float) parser.add_argument('--lr_backbone', default=1e-5, type=float) parser.add_argument('--batch_size', default=2, type=int) parser.add_argument('--weight_decay', default=1e-4, type=float) parser.add_argument('--epochs', default=300, type=int) parser.add_argument('--lr_drop', default=200, type=int) parser.add_argument('--clip_max_norm', default=0.1, type=float, help='gradient clipping max norm') # Model parameters parser.add_argument('--frozen_weights', type=str, default=None, help="Path to the pretrained model. If set, only the mask head will be trained") # * Backbone parser.add_argument('--backbone', default='resnet50', type=str, help="Name of the convolutional backbone to use") parser.add_argument('--dilation', action='store_true', help="If true, we replace stride with dilation in the last convolutional block (DC5)") parser.add_argument('--position_embedding', default='sine', type=str, choices=('sine', 'learned'), help="Type of positional embedding to use on top of the image features") # * Transformer parser.add_argument('--enc_layers', default=6, type=int, help="Number of encoding layers in the transformer") parser.add_argument('--dec_layers', default=6, type=int, help="Number of decoding layers in the transformer") parser.add_argument('--dim_feedforward', default=2048, type=int, help="Intermediate size of the feedforward layers in the transformer blocks") parser.add_argument('--hidden_dim', default=256, type=int, help="Size of the embeddings (dimension of the transformer)") parser.add_argument('--dropout', default=0.1, type=float, help="Dropout applied in the transformer") parser.add_argument('--nheads', default=8, type=int, help="Number of attention heads inside the transformer's attentions") parser.add_argument('--num_queries', default=100, type=int, help="Number of query slots") parser.add_argument('--pre_norm', action='store_true') # * Segmentation parser.add_argument('--masks', action='store_true', help="Train segmentation head if the flag is provided") # Loss parser.add_argument('--no_aux_loss', dest='aux_loss', action='store_false', help="Disables auxiliary decoding losses (loss at each layer)") # * Matcher parser.add_argument('--set_cost_class', default=1, type=float, help="Class coefficient in the matching cost") parser.add_argument('--set_cost_bbox', default=5, type=float, help="L1 box coefficient in the matching cost") parser.add_argument('--set_cost_giou', default=2, type=float, help="giou box coefficient in the matching cost") # * Loss coefficients parser.add_argument('--mask_loss_coef', default=1, type=float) parser.add_argument('--dice_loss_coef', default=1, type=float) parser.add_argument('--bbox_loss_coef', default=5, type=float) parser.add_argument('--giou_loss_coef', default=2, type=float) parser.add_argument('--eos_coef', default=0.1, type=float, help="Relative classification weight of the no-object class") # dataset parameters parser.add_argument('--dataset_file', default='coco') parser.add_argument('--coco_path', type=str, default=r"F:\DLdata\VOC2COCO") parser.add_argument('--coco_panoptic_path', type=str) parser.add_argument('--remove_difficult', action='store_true') parser.add_argument('--output_dir', default='./output', help='path where to save, empty for no saving') parser.add_argument('--device', default='cuda', help='device to use for training / testing') parser.add_argument('--seed', default=42, type=int) parser.add_argument('--resume', default='detr_r50_21.path', help='resume from checkpoint') # parser.add_argument('--resume', default='detr-r50-e632da11.pth', help='resume from checkpoint') parser.add_argument('--start_epoch', default=0, type=int, metavar='N', help='start epoch') parser.add_argument('--eval', action='store_true') parser.add_argument('--num_workers', default=0, type=int) # distributed training parameters parser.add_argument('--world_size', default=1, type=int, help='number of distributed processes') parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training') return parserCOLORS = [[0.000, 0.447, 0.741], [0.850, 0.325, 0.098], [0.929, 0.694, 0.125], [0.494, 0.184, 0.556], [0.466, 0.674, 0.188], [0.301, 0.745, 0.933]]transform_input = T.Compose([T.Resize(800), T.ToTensor(), T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])def box_cxcywh_to_xyxy(x): x_c, y_c, w, h = x.unbind(1) b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (x_c + 0.5 * w), (y_c + 0.5 * h)] return torch.stack(b, dim=1)def rescale_bboxes(out_bbox, size): img_w, img_h = size b = box_cxcywh_to_xyxy(out_bbox) b = b * torch.tensor([img_w, img_h, img_w, img_h], dtype=torch.float32, device="cuda") return bdef plot_results(pil_img, prob, boxes, img_save_path): plt.figure(figsize=(16, 10)) plt.imshow(pil_img) ax = plt.gca() colors = COLORS * 100 for p, (xmin, ymin, xmax, ymax), c in zip(prob, boxes.tolist(), colors): ax.add_patch(plt.Rectangle((xmin, ymin), xmax - xmin, ymax - ymin, fill=False, color=c, linewidth=3)) cl = p.argmax() text = f'{CLASSES[cl]}: {p[cl]:0.2f}' ax.text(xmin, ymin, text, fontsize=9, bbox=dict(facecolor='yellow', alpha=0.5)) plt.savefig(img_save_path) plt.axis('off') plt.show()def main(num_classes, chenkpoint_path, img_path, img_save_path, num_queries=100): parser = argparse.ArgumentParser('DETR training and evaluation script', parents=[get_args_parser()]) args = parser.parse_args() backbone = build_backbone(args) transform = build_transformer(args) model = DETR(backbone=backbone, transformer=transform, num_classes=num_classes, num_queries=100) device = "cuda" if torch.cuda.is_available() else "cpu" model.to(device) model_path = chenkpoint_path model_data = torch.load(model_path)['model'] model.load_state_dict(model_data) model.eval() path = img_path im = cv2.imread(path) im = Image.fromarray(cv2.cvtColor(im, cv2.COLOR_BGR2RGB)) img = transform_input(im).unsqueeze(0) outputs = model(img.to(device)) probs = outputs['pred_logits'].softmax(-1)[0, :, :-1] # 可修改阈值,只输出概率大于0.7的物体 keep = probs.max(-1).values > 0.7 # print(probs[keep]) bboxes_scaled = rescale_bboxes(outputs['pred_boxes'][0, keep], im.size) ori_img = np.array(im) plot_results(ori_img, probs[keep], bboxes_scaled, img_save_path)if __name__ == "__main__": CLASSES = ['N/A', "aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"] main(num_classes=21, chenkpoint_path="checkpoint.pth", img_path="test.png", img_save_path="result2.png")几点说明：
1.CLASSES是我们数据集对应的类别名，注意自己标注的顺序一定写对。第一个类别是背景类，这个是固定的，所有数据集都要有。
2.
num_classes：类别数+1
chenkpoint_path：保存的权重文件
img_path：测试的图片路径
img_save_path：保存结果路径
3.可修改阈值，论文中默认只输出概率大于0.7的物体。
用VOC数据集训练的模型推理效果：
（VOC数据集中没有自行车一类所以识别不出来）
五、一些小bug
UserWarning: floordiv is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the ‘trunc’ function NOT ‘floor’). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode=‘trunc’), or for actual floor division, use torch.div(a, b, rounding_mode=‘floor’).
这时一个torch版本原因导致的一个函数问题，报了一个警告。
将models/position_encoding.py文件中的第44行改成如下形式即可。
References
VOC2COCO代码参考Github
DETR预训练模型