diff --git a/README.md b/README.md index 9098fdc..b6af93c 100644 --- a/README.md +++ b/README.md @@ -157,6 +157,20 @@ export PYTHONPATH=$PWD python3 test.py --weights 权重路径 ``` +## Comparision with other methods in github +github上也有一些其他作者实现的mobile-yolo系列，其中paddle和keras都是用mobilev1作为backbone，无法直接计算Flops。 +darknet的Flops采用作者Readme中的BFlops。 + +|Model|resolution|COCO mAP|VOC mAP|Size| +|----|----|----|----|----| +|Ours|320|23.2|71.7|1x| +|Ours|640|27.5|74.4|4x| +|[caffe](https://github.com/eric612/MobileNet-YOLO)|352|-|71.5|x1.2| +|[paddle](https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.5/docs/MODEL_ZOO_cn.md)|608|29.3|76.2|-| +|[darknet](https://github.com/dog-qiuqiu/MobileNet-Yolo)|352|-|70.7|x2.2| +|[keras](https://github.com/Adamdad/keras-YOLOv3-mobilenet)|320|-|74.56|-| +|[tensorflow](https://github.com/fsx950223/mobilenetv2-yolov3)|416|-|66.9|x1.7| + ## Inference with TensorRT 1. python3 model/onnx_export.py --weights weight_path 2. 参考[https://github.com/Syencil/tensorRT](https://github.com/Syencil/tensorRT) diff --git a/script/coco2darknet.py b/script/coco2darknet.py index 87f64bd..d72fffc 100644 --- a/script/coco2darknet.py +++ b/script/coco2darknet.py @@ -469,13 +469,15 @@ def convert_ath_json(json_dir): # dir contains json annotations and images print('Done. Output saved to %s' % Path(dir).absolute()) -def convert_coco_json(json_dir='../coco/annotations/'): +def convert_coco_json(json_dir='/data/dataset/coco/annotations/'): dir = make_folders(path='out/') # output directory jsons = glob.glob(json_dir + '*.json') coco80 = coco91_to_coco80_class() # Import json for json_file in sorted(jsons): + if 'instances' not in json_file: + continue fn = 'out/labels/%s/' % Path(json_file).stem.replace('instances_', '') # folder name os.mkdir(fn) with open(json_file) as f: diff --git a/test.py b/test.py index 22c15e6..1690b65 100644 --- a/test.py +++ b/test.py @@ -87,10 +87,8 @@ def test(data, collate_fn=dataset.collate_fn) seen = 0 - names = {0: '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'} + names = {k: v for k, v in enumerate(model.names if hasattr(model, 'names') else model.module.names)} + coco91class = coco80_to_coco91_class() s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', 'mAP@.5', 'mAP@.5:.95') @@ -365,7 +363,8 @@ def test(data, opt.iou_thres, opt.save_json, opt.single_cls, - opt.augment) + opt.augment, + verbose=opt.verbose) elif opt.task == 'study': # run over a range of settings and save/plot for weights in ['yolov5s.pt', 'yolov5m.pt', diff --git a/train.py b/train.py index fcedf55..4f1ceb5 100644 --- a/train.py +++ b/train.py @@ -201,22 +201,22 @@ def get_hooks(): c1 = 256 c2 = 512 c3 = 1024 - # S_Converter_1 = Converter(32, c1, act=True) - # S_Converter_2 = Converter(96, c2, act=True) - # S_Converter_3 = Converter(320, c3, act=True) - # S_Converter_1.to(device) - # S_Converter_2.to(device) - # S_Converter_3.to(device) - # S_Converter_1.train() - # S_Converter_2.train() - # S_Converter_3.train() - - # T_Converter_1 = nn.ReLU6() - # T_Converter_2 = nn.ReLU6() - # T_Converter_3 = nn.ReLU6() - T_Converter_1 = Converter(c1, 32, act=True) - T_Converter_2 = Converter(c2, 96, act=True) - T_Converter_3 = Converter(c3, 320, act=True) + S_Converter_1 = Converter(32, c1, act=True) + S_Converter_2 = Converter(96, c2, act=True) + S_Converter_3 = Converter(320, c3, act=True) + S_Converter_1.to(device) + S_Converter_2.to(device) + S_Converter_3.to(device) + S_Converter_1.train() + S_Converter_2.train() + S_Converter_3.train() + + T_Converter_1 = nn.ReLU6() + T_Converter_2 = nn.ReLU6() + T_Converter_3 = nn.ReLU6() + # T_Converter_1 = Converter(c1, 32, act=True) + # T_Converter_2 = Converter(c2, 96, act=True) + # T_Converter_3 = Converter(c3, 320, act=True) T_Converter_1.to(device) T_Converter_2.to(device) T_Converter_3.to(device) @@ -360,14 +360,16 @@ def get_hooks(): if opt.dist: if opt.d_online: t_pred = t_model(imgs) + for p in t_pred: + p = p.detach() else: with torch.no_grad(): t_pred = t_model(imgs) if opt.d_feature: - # s_f1 = S_Converter_1(activation["s_f1"]) - # s_f2 = S_Converter_2(activation["s_f2"]) - # s_f3 = S_Converter_3(activation["s_f3"]) - # s_f = [s_f1, s_f2, s_f3] + s_f1 = S_Converter_1(activation["s_f1"]) + s_f2 = S_Converter_2(activation["s_f2"]) + s_f3 = S_Converter_3(activation["s_f3"]) + s_f = [s_f1, s_f2, s_f3] s_f = (activation["s_f1"], activation["s_f2"], activation["s_f3"]) t_f1 = T_Converter_1(activation["t_f1"]) t_f2 = T_Converter_2(activation["t_f2"]) @@ -375,7 +377,7 @@ def get_hooks(): t_f = [t_f1, t_f2, t_f3] # t_f = (activation["t_f1"], activation["t_f2"], activation["t_f3"]) # Loss - loss, loss_items = compute_loss(pred, targets.to(device), model) + loss, loss_items = compute_loss(pred, targets.to(device), model, None) # Sparse Learning if opt.sl > 0: @@ -384,7 +386,7 @@ def get_hooks(): # distillation if opt.dist: if opt.d_online: - loss, _ = compute_loss(t_pred, targets.to(device), model, loss) + loss, _ = compute_loss(t_pred, targets.to(device), t_model, loss) loss = compute_distillation_output_loss(pred, t_pred, model, loss) if opt.d_feature: loss = compute_distillation_feature_loss(s_f, t_f, model, loss) @@ -524,12 +526,12 @@ def get_hooks(): parser.add_argument('--nw', type=int, default=None, help='num of worker') # pruning parser.add_argument('--sl', default=0, type=float, help='sparse learning') - parser.add_argument('--ft', action='store_true', help='fine-tune') + parser.add_argument('--ft', action='store_true', default=False, help='fine-tune') # distillation parser.add_argument('--dist', action='store_true', help='distillation') - parser.add_argument('--t_weights', type=str, help='teacher model for distillation') - parser.add_argument('--d_feature', action='store_true', help='if true, distill both feature and output layers') - parser.add_argument('--d_online', action='store_true', help='if true, using online-distillation') + parser.add_argument('--t_weights', type=str, default="", help='teacher model for distillation') + parser.add_argument('--d_feature', action='store_true', default=False, help='if true, distill both feature and output layers') + parser.add_argument('--d_online', action='store_true', default=False, help='if true, using online-distillation') opt = parser.parse_args() if opt.type == "mcocos": @@ -546,7 +548,7 @@ def get_hooks(): opt.data = "data/coco.yaml" opt.name = opt.type opt.weights = "outputs/dmvocs/weights/best_dmvocs.pt" - opt.epochs = 50 + opt.epochs = 10 opt.batch_size = 24 opt.multi_scale = False opt.dist = True @@ -737,7 +739,7 @@ def get_hooks(): opt.multi_scale = False opt.dist = True opt.d_online = True - opt.t_weights = "outputs/voc/weights/best_voc.pt" + opt.t_weights = "/data/checkpoints/yolov5/yolov5s.pt" hyp["dist"] = 1 if opt.nw is None: