shaoshengsong / mobilenetv3-ssd Goto Github PK

或者说Windows平台兼容吗?

I found this code

Average Precision Per-class:
aeroplane: 0.5286851217064079
bicycle: 0.5684498038841389
bird: 0.2951959818198857
boat: 0.3374119115412078
bottle: 0.11274539710602145
bus: 0.5840837968782298
car: 0.5631029936677479
cat: 0.5605063414099851
chair: 0.24291365433556902
cow: 0.2908948526148929
diningtable: 0.4292561103683377
dog: 0.47845551032801703
horse: 0.6137629939581943
motorbike: 0.6033233235496501
person: 0.5106418290063527
pottedplant: 0.13170114489968404
sheep: 0.3551660800668141
sofa: 0.4682129143642543
train: 0.6344977720902926
tvmonitor: 0.4312501217962447

Average Precision Across All Classes:0.4370128827695964

请问您给的训练策略能把模型loss收敛到多小？我顶天了到3.7左右就再也降不下去了，不管是加epoch,调整学习率，还是调整batch_size都无法降低了？请问这是什么原因呢？

11.19更新：
训练集：VOC07+12，测试集：VOC07-test，
使用SGDR学习率调整策略，初始学习率为0.01，学习率调整周期为100个epoches，总共训练610个epoches，
预训练模型：mb3-ssd-lite-Epoch-149-Loss-5.782852862012213.pth，
batch_size：32，
能得到的最好的Loss是4.07，请问怎样进一步调整学习策略？

原提问：
你好，当我使用您提供的PASCAL的预训练模型（mb3-ssd-lite-Epoch-149-Loss-5.782852862012213.pth）对几张图片进行检测时，每张图片都检测不到任何一个物体。
然后，我在上述检查点的基础上进行finetune，使用VOC2007+VOC2012数据集训练了100个epoch，batch_size=32，lr=0.0005+multistep([40,80,])，使用新训练的模型再次检测，部分图片能检测到物体，但是准确度还是比较低。
对此，我有两个问题想请教您：

1. 您在使用上述PASCAL模型进行检测时效果如何？是我操作不当还是模型本身专准确度就不高？
2. 如果想进一步提高准确度，我需要如何训练？比如是否应该加大数据集或改进训练参数。

Baidu doesn't register users outside China, thanks.

Thank you a lot for your works

Thanks for the great work. I trained it on Titan V, but it is very slow. Do you have any idea of how to train it faster?

Traceback (most recent call last):
File "run_ssd_live_demo.py", line 73, in
boxes, labels, probs = predictor.predict(image, 10, 0.4)
File "/home/yajun/github/MobileNetV3-SSD/vision/ssd/predictor.py", line 37, in predict
scores, boxes = self.net.forward(images)
File "/home/yajun/github/MobileNetV3-SSD/vision/ssd/ssd.py", line 100, in forward
locations, self.priors, self.config.center_variance, self.config.size_variance
File "/home/yajun/github/MobileNetV3-SSD/vision/utils/box_utils.py", line 104, in convert_locations_to_boxes
locations[..., :2] * center_variance * priors[..., 2:] + priors[..., :2],
RuntimeError: expected backend CPU and dtype Float but got backend CUDA and dtype Float

为什么我在训练VOC2007+2012时候总是说找不到sub-train-annotations-bbox.csv

raise self.exc_type(msg)

ValueError: Caught ValueError in DataLoader worker process 7.
Original Traceback (most recent call last):
File "/home/yww/anaconda3/envs/pytorch1.3/lib/python3.6/site-packages/torch/utils/data/_utils/worker.py", line 178, in _worker_loop
data = fetcher.fetch(index)
File "/home/yww/anaconda3/envs/pytorch1.3/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py", line 44, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/home/yww/anaconda3/envs/pytorch1.3/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py", line 44, in
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/home/yww/anaconda3/envs/pytorch1.3/lib/python3.6/site-packages/torch/utils/data/dataset.py", line 207, in getitem
return self.datasets[dataset_idx][sample_idx]
File "/home/yww/MobileNetV3-SSD/vision/datasets/voc_dataset.py", line 65, in getitem
image, boxes, labels = self.transform(image, boxes, labels)
File "/home/yww/MobileNetV3-SSD/vision/ssd/data_preprocessing.py", line 34, in call
return self.augment(img, boxes, labels)
File "/home/yww/MobileNetV3-SSD/vision/transforms/transforms.py", line 55, in call
img, boxes, labels = t(img, boxes, labels)
File "/home/yww/MobileNetV3-SSD/vision/transforms/transforms.py", line 278, in call
if overlap.min() < min_iou and max_iou < overlap.max():
File "/home/yww/anaconda3/envs/pytorch1.3/lib/python3.6/site-packages/numpy/core/_methods.py", line 29, in _amin
return umr_minimum(a, axis, None, out, keepdims)
ValueError: zero-size array to reduction operation minimum which has no identity

怀疑是数据集的处理过程中出现0-size 的batch

Hello,
Thank you for your repo.
I have used your pretrained model, trained it one more time with VOC dataset for 100 epochs but Loss does not go less than 4.8, and the model accuracy is not good at all. I have used VOC 2012 and 2007 dataset for training and 2007 for testing. Do you have any idea what it could be happened?
Regards

如果可以，是否有demo可以给出，万分感谢~

你好，请问要在嵌入式部署这两种OD网络，哪个更好一些啊~他们的速度和精度能有多大差距呢

Can somebody help me?

(py3.6) D:\winfred\mobilenetv3_ssd>python train_ssd.py --dataset_type voc --datasets ./data/VOC2007 ./data/VOC2012 --validation_dataset ./data/test/VOC2007 --net mb3-ssd-lite --pretrained_ssd ./models/mb3-ssd-lite-Epoch-149-Loss-5.782852862012213.pth
2019-10-25 17:22:45,411 - root - INFO - Namespace(balance_data=False, base_net=None, base_net_lr=None, batch_size=32, checkpoint_folder='models/', dataset_type='voc', datasets=['./data/VOC2007', './data/VOC2012'], debug_steps=100, extra_layers_lr=None, freeze_base_net=False, freeze_net=False, gamma=0.1, lr=0.001, mb2_width_mult=1.0, milestones='80,100', momentum=0.9, net='mb3-ssd-lite', num_epochs=120, num_workers=4, pretrained_ssd='./models/mb3-ssd-lite-Epoch-149-Loss-5.782852862012213.pth', resume=None, scheduler='multi-step', t_max=120, use_cuda=True, validation_dataset='./data/test/VOC2007', validation_epochs=5, weight_decay=0.0005)
2019-10-25 17:22:45,413 - root - INFO - Prepare training datasets.
2019-10-25 17:22:45,415 - root - INFO - No labels file, using default VOC classes.
2019-10-25 17:22:45,418 - root - INFO - No labels file, using default VOC classes.
2019-10-25 17:22:45,419 - root - INFO - Stored labels into file models/voc-model-labels.txt.
2019-10-25 17:22:45,420 - root - INFO - Train dataset size: 16551
2019-10-25 17:22:45,420 - root - INFO - Prepare Validation datasets.
2019-10-25 17:22:45,422 - root - INFO - No labels file, using default VOC classes.
2019-10-25 17:22:45,424 - root - INFO - validation dataset size: 4952
2019-10-25 17:22:45,428 - root - INFO - Build network.
SSD(
(base_net): Sequential(
(0): Conv2d(3, 16, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
(1): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
(3): MobileBlock(
(conv): Sequential(
(0): Conv2d(16, 16, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
(depth_conv): Sequential(
(0): Conv2d(16, 16, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=16)
(1): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(squeeze_block): SqueezeBlock(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(dense): Sequential(
(0): Linear(in_features=16, out_features=4, bias=True)
(1): ReLU(inplace)
(2): Linear(in_features=4, out_features=16, bias=True)
(3): h_sigmoid()
)
)
(point_conv): Sequential(
(0): Conv2d(16, 16, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
)
(4): MobileBlock(
(conv): Sequential(
(0): Conv2d(16, 72, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(72, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
(depth_conv): Sequential(
(0): Conv2d(72, 72, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=72)
(1): BatchNorm2d(72, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(point_conv): Sequential(
(0): Conv2d(72, 24, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
)
(5): MobileBlock(
(conv): Sequential(
(0): Conv2d(24, 88, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(88, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
(depth_conv): Sequential(
(0): Conv2d(88, 88, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=88)
(1): BatchNorm2d(88, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(point_conv): Sequential(
(0): Conv2d(88, 24, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
)
(6): MobileBlock(
(conv): Sequential(
(0): Conv2d(24, 96, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
(depth_conv): Sequential(
(0): Conv2d(96, 96, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), groups=96)
(1): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(squeeze_block): SqueezeBlock(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(dense): Sequential(
(0): Linear(in_features=96, out_features=24, bias=True)
(1): ReLU(inplace)
(2): Linear(in_features=24, out_features=96, bias=True)
(3): h_sigmoid()
)
)
(point_conv): Sequential(
(0): Conv2d(96, 40, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
)
(7): MobileBlock(
(conv): Sequential(
(0): Conv2d(40, 240, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
(depth_conv): Sequential(
(0): Conv2d(240, 240, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=240)
(1): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(squeeze_block): SqueezeBlock(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(dense): Sequential(
(0): Linear(in_features=240, out_features=60, bias=True)
(1): ReLU(inplace)
(2): Linear(in_features=60, out_features=240, bias=True)
(3): h_sigmoid()
)
)
(point_conv): Sequential(
(0): Conv2d(240, 40, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
)
(8): MobileBlock(
(conv): Sequential(
(0): Conv2d(40, 240, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
(depth_conv): Sequential(
(0): Conv2d(240, 240, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=240)
(1): BatchNorm2d(240, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(squeeze_block): SqueezeBlock(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(dense): Sequential(
(0): Linear(in_features=240, out_features=60, bias=True)
(1): ReLU(inplace)
(2): Linear(in_features=60, out_features=240, bias=True)
(3): h_sigmoid()
)
)
(point_conv): Sequential(
(0): Conv2d(240, 40, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU(inplace)
)
)
(9): MobileBlock(
(conv): Sequential(
(0): Conv2d(40, 120, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(120, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
)
(depth_conv): Sequential(
(0): Conv2d(120, 120, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=120)
(1): BatchNorm2d(120, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(squeeze_block): SqueezeBlock(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(dense): Sequential(
(0): Linear(in_features=120, out_features=30, bias=True)
(1): ReLU(inplace)
(2): Linear(in_features=30, out_features=120, bias=True)
(3): h_sigmoid()
)
)
(point_conv): Sequential(
(0): Conv2d(120, 48, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
)
)
(10): MobileBlock(
(conv): Sequential(
(0): Conv2d(48, 144, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
)
(depth_conv): Sequential(
(0): Conv2d(144, 144, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=144)
(1): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(squeeze_block): SqueezeBlock(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(dense): Sequential(
(0): Linear(in_features=144, out_features=36, bias=True)
(1): ReLU(inplace)
(2): Linear(in_features=36, out_features=144, bias=True)
(3): h_sigmoid()
)
)
(point_conv): Sequential(
(0): Conv2d(144, 48, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
)
)
(11): MobileBlock(
(conv): Sequential(
(0): Conv2d(48, 288, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(288, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
)
(depth_conv): Sequential(
(0): Conv2d(288, 288, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), groups=288)
(1): BatchNorm2d(288, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(squeeze_block): SqueezeBlock(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(dense): Sequential(
(0): Linear(in_features=288, out_features=72, bias=True)
(1): ReLU(inplace)
(2): Linear(in_features=72, out_features=288, bias=True)
(3): h_sigmoid()
)
)
(point_conv): Sequential(
(0): Conv2d(288, 96, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
)
)
(12): MobileBlock(
(conv): Sequential(
(0): Conv2d(96, 576, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(576, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
)
(depth_conv): Sequential(
(0): Conv2d(576, 576, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=576)
(1): BatchNorm2d(576, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(squeeze_block): SqueezeBlock(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(dense): Sequential(
(0): Linear(in_features=576, out_features=144, bias=True)
(1): ReLU(inplace)
(2): Linear(in_features=144, out_features=576, bias=True)
(3): h_sigmoid()
)
)
(point_conv): Sequential(
(0): Conv2d(576, 96, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
)
)
(13): MobileBlock(
(conv): Sequential(
(0): Conv2d(96, 576, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(576, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
)
(depth_conv): Sequential(
(0): Conv2d(576, 576, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), groups=576)
(1): BatchNorm2d(576, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
(squeeze_block): SqueezeBlock(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(dense): Sequential(
(0): Linear(in_features=576, out_features=144, bias=True)
(1): ReLU(inplace)
(2): Linear(in_features=144, out_features=576, bias=True)
(3): h_sigmoid()
)
)
(point_conv): Sequential(
(0): Conv2d(576, 96, kernel_size=(1, 1), stride=(1, 1))
(1): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): h_swish()
)
)
(14): Conv2d(96, 576, kernel_size=(1, 1), stride=(1, 1))
(15): SqueezeBlock(
(avg_pool): AdaptiveAvgPool2d(output_size=1)
(dense): Sequential(
(0): Linear(in_features=576, out_features=144, bias=True)
(1): ReLU(inplace)
(2): Linear(in_features=144, out_features=576, bias=True)
(3): h_sigmoid()
)
)
(16): BatchNorm2d(576, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(17): h_swish()
(18): Conv2d(576, 1280, kernel_size=(1, 1), stride=(1, 1))
(19): h_swish()
)
(extras): ModuleList(
(0): InvertedResidual(
(conv): Sequential(
(0): Conv2d(1280, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6(inplace)
(3): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=256, bias=False)
(4): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): ReLU6(inplace)
(6): Conv2d(256, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
(7): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(1): InvertedResidual(
(conv): Sequential(
(0): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6(inplace)
(3): Conv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=128, bias=False)
(4): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): ReLU6(inplace)
(6): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(7): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(2): InvertedResidual(
(conv): Sequential(
(0): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6(inplace)
(3): Conv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=128, bias=False)
(4): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): ReLU6(inplace)
(6): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(7): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
(3): InvertedResidual(
(conv): Sequential(
(0): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6(inplace)
(3): Conv2d(64, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=64, bias=False)
(4): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(5): ReLU6(inplace)
(6): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(7): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
)
)
)
(classification_headers): ModuleList(
(0): Sequential(
(0): Conv2d(288, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=288)
(1): BatchNorm2d(288, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6()
(3): Conv2d(288, 126, kernel_size=(1, 1), stride=(1, 1))
)
(1): Sequential(
(0): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=1280)
(1): BatchNorm2d(1280, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6()
(3): Conv2d(1280, 126, kernel_size=(1, 1), stride=(1, 1))
)
(2): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=512)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6()
(3): Conv2d(512, 126, kernel_size=(1, 1), stride=(1, 1))
)
(3): Sequential(
(0): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6()
(3): Conv2d(256, 126, kernel_size=(1, 1), stride=(1, 1))
)
(4): Sequential(
(0): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6()
(3): Conv2d(256, 126, kernel_size=(1, 1), stride=(1, 1))
)
(5): Conv2d(64, 126, kernel_size=(1, 1), stride=(1, 1))
)
(regression_headers): ModuleList(
(0): Sequential(
(0): Conv2d(288, 288, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=288)
(1): BatchNorm2d(288, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6()
(3): Conv2d(288, 24, kernel_size=(1, 1), stride=(1, 1))
)
(1): Sequential(
(0): Conv2d(1280, 1280, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=1280)
(1): BatchNorm2d(1280, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6()
(3): Conv2d(1280, 24, kernel_size=(1, 1), stride=(1, 1))
)
(2): Sequential(
(0): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=512)
(1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6()
(3): Conv2d(512, 24, kernel_size=(1, 1), stride=(1, 1))
)
(3): Sequential(
(0): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6()
(3): Conv2d(256, 24, kernel_size=(1, 1), stride=(1, 1))
)
(4): Sequential(
(0): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256)
(1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(2): ReLU6()
(3): Conv2d(256, 24, kernel_size=(1, 1), stride=(1, 1))
)
(5): Conv2d(64, 24, kernel_size=(1, 1), stride=(1, 1))
)
(source_layer_add_ons): ModuleList()
)
2019-10-25 17:22:48,105 - root - INFO - Init from pretrained ssd ./models/mb3-ssd-lite-Epoch-149-Loss-5.782852862012213.pth
2019-10-25 17:22:49,881 - root - INFO - Took 1.78 seconds to load the model.
2019-10-25 17:22:50,180 - root - INFO - Learning rate: 0.001, Base net learning rate: 0.001, Extra Layers learning rate: 0.001.
2019-10-25 17:22:50,180 - root - INFO - Uses MultiStepLR scheduler.
2019-10-25 17:22:50,181 - root - INFO - Start training from epoch 0.
Traceback (most recent call last):
File "train_ssd.py", line 338, in
device=DEVICE, debug_steps=args.debug_steps, epoch=epoch)
File "train_ssd.py", line 119, in train
for i, data in enumerate(loader):
File "D:\Anaconda3\envs\py3.6\lib\site-packages\torch\utils\data\dataloader.py", line 819, in iter
return _DataLoaderIter(self)
File "D:\Anaconda3\envs\py3.6\lib\site-packages\torch\utils\data\dataloader.py", line 560, in init
w.start()
File "D:\Anaconda3\envs\py3.6\lib\multiprocessing\process.py", line 105, in start
self._popen = self._Popen(self)
File "D:\Anaconda3\envs\py3.6\lib\multiprocessing\context.py", line 223, in _Popen
return _default_context.get_context().Process._Popen(process_obj)
File "D:\Anaconda3\envs\py3.6\lib\multiprocessing\context.py", line 322, in _Popen
return Popen(process_obj)
File "D:\Anaconda3\envs\py3.6\lib\multiprocessing\popen_spawn_win32.py", line 65, in init
reduction.dump(process_obj, to_child)
File "D:\Anaconda3\envs\py3.6\lib\multiprocessing\reduction.py", line 60, in dump
ForkingPickler(file, protocol).dump(obj)
AttributeError: Can't pickle local object 'TrainAugmentation.init..'

(py3.6) D:\winfred\mobilenetv3_ssd>Traceback (most recent call last):
File "", line 1, in
File "D:\Anaconda3\envs\py3.6\lib\multiprocessing\spawn.py", line 105, in spawn_main
exitcode = _main(fd)
File "D:\Anaconda3\envs\py3.6\lib\multiprocessing\spawn.py", line 115, in _main
self = reduction.pickle.load(from_parent)
EOFError: Ran out of input

I have trained the pytorch model.pth on my dataset, then, I use the convert_to_caffe2_models.py to convert ONXX model, but it raise an error like this:

Traceback (most recent call last):
  File "convert_to_caffe2_models.py", line 50, in <module>
    torch.onnx.export(net, dummy_input, model_path, verbose=False, output_names=['scores', 'boxes'])
  File "/home/tsd/.local/lib/python3.6/site-packages/torch/onnx/__init__.py", line 25, in export
    return utils.export(*args, **kwargs)
  File "/home/tsd/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 131, in export
    strip_doc_string=strip_doc_string)
  File "/home/tsd/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 363, in _export
    _retain_param_name, do_constant_folding)
  File "/home/tsd/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 278, in _model_to_graph
    _disable_torch_constant_prop=_disable_torch_constant_prop)
  File "/home/tsd/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 188, in _optimize_graph
    graph = torch._C._jit_pass_onnx(graph, operator_export_type)
  File "/home/tsd/.local/lib/python3.6/site-packages/torch/onnx/__init__.py", line 50, in _run_symbolic_function
    return utils._run_symbolic_function(*args, **kwargs)
  File "/home/tsd/.local/lib/python3.6/site-packages/torch/onnx/utils.py", line 589, in _run_symbolic_function
    return fn(g, *inputs, **attrs)
  File "/home/tsd/.local/lib/python3.6/site-packages/torch/onnx/symbolic.py", line 130, in wrapper
    args = [_parse_arg(arg, arg_desc) for arg, arg_desc in zip(args, arg_descriptors)]
  File "/home/tsd/.local/lib/python3.6/site-packages/torch/onnx/symbolic.py", line 130, in <listcomp>
    args = [_parse_arg(arg, arg_desc) for arg, arg_desc in zip(args, arg_descriptors)]
  File "/home/tsd/.local/lib/python3.6/site-packages/torch/onnx/symbolic.py", line 90, in _parse_arg
    raise RuntimeError("Failed to export an ONNX attribute, "
RuntimeError: Failed to export an ONNX attribute, since it's not constant, please try to make things (e.g., kernel size) static if possible

Traceback (most recent call last):
File "train_ssd.py", line 335, in
device=DEVICE, debug_steps=args.debug_steps, epoch=epoch)
File "train_ssd.py", line 117, in train
for i, data in enumerate(loader):
File "/home/ai/anaconda3/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 801, in next
return self._process_data(data)
File "/home/ai/anaconda3/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 846, in _process_data
data.reraise()
File "/home/ai/anaconda3/lib/python3.6/site-packages/torch/_utils.py", line 385, in reraise
raise self.exc_type(msg)
IndexError: Caught IndexError in DataLoader worker process 3.

IndexError: too many indices for array

代码质量太差了。。。

没有使用预训练模型，我在coco提取了5类70K图片进行训练，但是loss降到3.6左右就降不下来了，请问有什么策略嘛？

Hello,

Whenever I try to run demo, I am having an issue. Could you help me with this? I would be appreciated.

The Error message is:

RuntimeError: Error(s) in loading state_dict for SSD:
size mismatch for classification_headers.0.3.weight: copying a param with shape torch.Size([126, 288, 1, 1]) from checkpoint, the shape in current model is torch.Size([18, 288, 1, 1]).
size mismatch for classification_headers.0.3.bias: copying a param with shape torch.Size([126]) from checkpoint, the shape in current model is torch.Size([18]).
size mismatch for classification_headers.1.3.weight: copying a param with shape torch.Size([126, 1280, 1, 1]) from checkpoint, the shape in current model is torch.Size([18, 1280, 1, 1]).
size mismatch for classification_headers.1.3.bias: copying a param with shape torch.Size([126]) from checkpoint, the shape in current model is torch.Size([18]).
size mismatch for classification_headers.2.3.weight: copying a param with shape torch.Size([126, 512, 1, 1]) from checkpoint, the shape in current model is torch.Size([18, 512, 1, 1]).
size mismatch for classification_headers.2.3.bias: copying a param with shape torch.Size([126]) from checkpoint, the shape in current model is torch.Size([18]).
size mismatch for classification_headers.3.3.weight: copying a param with shape torch.Size([126, 256, 1, 1]) from checkpoint, the shape in current model is torch.Size([18, 256, 1, 1]).
size mismatch for classification_headers.3.3.bias: copying a param with shape torch.Size([126]) from checkpoint, the shape in current model is torch.Size([18]).
size mismatch for classification_headers.4.3.weight: copying a param with shape torch.Size([126, 256, 1, 1]) from checkpoint, the shape in current model is torch.Size([18, 256, 1, 1]).
size mismatch for classification_headers.4.3.bias: copying a param with shape torch.Size([126]) from checkpoint, the shape in current model is torch.Size([18]).
size mismatch for classification_headers.5.weight: copying a param with shape torch.Size([126, 64, 1, 1]) from checkpoint, the shape in current model is torch.Size([18, 64, 1, 1]).
size mismatch for classification_headers.5.bias: copying a param with shape torch.Size([126]) from checkpoint, the shape in current model is torch.Size([18]).

I'm having a problem when training ssd mobilenet v3 lite, I have lost inf at some steps. Does that affect my results?
---2019-08-22 13:54:37,701 - root - INFO - Epoch: 6, Step: 20, Average Loss: inf, Average Regression Loss inf, Average Classification Loss: 3.2975
2019-08-22 13:56:20,223 - root - INFO - Epoch: 6, Step: 30, Average Loss: inf, Average Regression Loss inf, Average Classification Loss: 3.3092
2019-08-22 13:58:55,497 - root - INFO - Epoch: 6, Step: 40, Average Loss: 6.5821, Average Regression Loss 3.2743, Average Classification Loss: 3.3079
2019-08-22 14:00:47,632 - root - INFO - Epoch: 6, Step: 50, Average Loss: 6.5586, Average Regression Loss 3.2818, Average Classification Loss: 3.2768
+I would like to ask if I download pretrain model voc but when I tested it, I did not return the results of the boxes
I thank

请问一下，执行run_ssd_example.py脚本报这个错该如何解决？ Traceback (most recent call last):
File "run_ssd_example.py", line 57, in
boxes, labels, probs = predictor.predict(image, 10, 0.4)
File "D:\Projects\DeepLearning\MobileNetV3-SSD-master\vision\ssd\predictor.py", line 37, in predict
scores, boxes = self.net.forward(images)
File "D:\Projects\DeepLearning\MobileNetV3-SSD-master\vision\ssd\ssd.py", line 100, in forward
locations, self.priors, self.config.center_variance, self.config.size_variance
File "D:\Projects\DeepLearning\MobileNetV3-SSD-master\vision\utils\box_utils.py", line 104, in convert_locations_to_boxes
locations[..., :2] * center_variance * priors[..., 2:] + priors[..., :2],
RuntimeError: expected backend CPU and dtype Float but got backend CUDA and dtype Float

can you provide a link to download mb3-ssd-lite-Epoch-99-Loss-2.5194434596402613.pth ?

你好 !
我按着步骤来 进行 预加载训练.
但是 训练时没有从Epoch150开始 而是直接又从0开始训练了
请问这是怎么回事? 我应该怎么调.

cjd@jordan:~/car_FPN-G-F-MobileNetV3-SSD$ python3.6 train_ssd.py --datasets /home/cjd/ssd.pytorch/data/VOCdevkit/VOC2007 --validation_dataset /home/cjd/ssd.pytorch/data/VOCdevkit/VOC2007 --net mb3-ssd-lite --pretrained_ssd models_new/mb3-ssd-lite-Epoch-149-Loss-2.2059922739140037.pth --scheduler cosine --lr 0.01 --t_max 100 --validation_epochs 5 --num_epochs 300 --base_net_lr 0.0012 --batch_size 5
2020-07-15 17:08:40,220 - root - INFO - Use Cuda.
2020-07-15 17:08:40,220 - root - INFO - Namespace(balance_data=False, base_net=None, base_net_lr=0.0012, batch_size=5, checkpoint_folder='models_new/', dataset_type='voc', datasets=['/home/cjd/ssd.pytorch/data/VOCdevkit/VOC2007'], debug_steps=100, extra_layers_lr=None, freeze_base_net=False, freeze_net=False, gamma=0.1, lr=0.01, mb2_width_mult=1.0, milestones='80,100', momentum=0.9, net='mb3-ssd-lite', num_epochs=300, num_workers=0, pretrained_ssd='models_new/mb3-ssd-lite-Epoch-149-Loss-2.2059922739140037.pth', resume=None, scheduler='cosine', t_max=100.0, use_cuda=True, validation_dataset='/home/cjd/ssd.pytorch/data/VOCdevkit/VOC2007', validation_epochs=5, weight_decay=0.0005)
2020-07-15 17:08:40,233 - root - INFO - Prepare training datasets.
2020-07-15 17:08:40,240 - root - INFO - No labels file, using default VOC classes.
2020-07-15 17:08:40,246 - root - INFO - Stored labels into file models_new/voc-model-labels.txt.
2020-07-15 17:08:40,246 - root - INFO - Train dataset size: 3300
2020-07-15 17:08:40,246 - root - INFO - Prepare Validation datasets.
2020-07-15 17:08:40,248 - root - INFO - No labels file, using default VOC classes.
2020-07-15 17:08:40,248 - root - INFO - validation dataset size: 826
2020-07-15 17:08:40,248 - root - INFO - Build network.
2020-07-15 17:08:40,397 - root - INFO - Init from pretrained ssd models_new/mb3-ssd-lite-Epoch-149-Loss-2.2059922739140037.pth
2020-07-15 17:08:40,519 - root - INFO - Took 0.12 seconds to load the model.
2
2020-07-15 17:08:43,911 - root - INFO - Learning rate: 0.01, Base net learning rate: 0.0012, Extra Layers learning rate: 0.01.
2020-07-15 17:08:43,911 - root - INFO - Uses CosineAnnealingLR scheduler.
2020-07-15 17:08:43,911 - root - INFO - Start training from epoch 0.

when i use voc0712 training，it appears the following error，please help me:
CUDA_VISIBLE_DEVICES=4 python2 train_ssd.py --dataset_type voc --datasets /usr/local/zjc/ActionRecog/VOCdevkit --net mb3-ssd-lite --scheduler cosine --lr 0.01 --t_max 100 --validation_epochs 5 --num_epochs 100 --base_net_lr 0.001 --batch_size 5
File "train_ssd.py", line 223
raise ValueError(f"Dataset tpye {args.dataset_type} is not supported.")
^
SyntaxError: invalid syntax

Hello, first, thank you for providing pytorch code for MobilenetV3-SSD. I want to ask about video detection. Please answer how to run_ssd_live_demo.py. Thanks.

When I compare the implementation of yours and official paper, I found some differences probably due to the paper update in last November.

Changes are in Mobilenet_v3.py:
line 159
[112, 160, 5, 1, "HS", True, 672] needs to be [112, 160, 5, 2, "HS", True, 672]

line 159
[160, 160, 5, 2, "HS", True, 672] needs to be [160, 160, 5, 1, "HS", True, 960]

If I am not mistaken,
Line 246
out_conv2_out = _make_divisible(1280 * multiplier) needs to be

out_conv2_out = _make_divisible(1024 * multiplier), since this is MobilenetV3_Small.
Mobilenetv3_Large uses 1280.

out_conv2_out = _make_divisible(1280 * multiplier)

Hello,
thank you for this repo. Could you provide a Google Drive link or something else than Baidu?
Regards

I tried to load the weights of https://github.com/kuan-wang/pytorch-mobilenet-v3,
it gave the error that size mismatch. Could you tell me where the pretrained model can be downloaded?

我使用mbv3_small.pth 加载到这个程序当中为什么会出现乱码 下面是我加载的命令

python3.6 train_ssd.py --datasets /home/cjd/ssd.pytorch/data/VOCdevkit/VOC2007/ --validation_dataset /home/cjd/ssd.pytorch/data/VOCdevkit/VOC2007/ --net mb2-ssd-lite --base_net models/mbv3_small.pth --batch_size 8 --scheduler cosine --lr 0.01 --t_max 200 --validation_epochs 5 --num_epochs 200

就一会出现乱码 报错 不知道是v3的预训练不行还是 此程序不行

请指点下 谢谢
@shaoshengsong

我怎么改，用我的数据，比如 20 类的权重，我只有 5类