用GhostModule替换Conv2d，loss降的很慢？

[yc-cui]

我直接将efficientnet里面的MBConvBlock中的Conv2d替换为GhostModule：
Conv2d(in_channels=inp, out_channels=oup, kernel_size=1, bias=False)
替换为
GhostModule(inp, oup)，
其他参数不变，为什么损失比以前收敛的更慢了，一直降不下来？请问需要修改其他什么参数吗？

[iamhankai]

试试GhostModule( inp, oup, kernel_size=1, ratio=2, dw_size=3, stride=1, relu=False)

另外，EfficientNet里PWConv换成GhostModule，DWConv就不要再换成GhostModule了

[yc-cui]

试试GhostModule( inp, oup, kernel_size=1, ratio=2, dw_size=3, stride=1, relu=False)

另外，EfficientNet里PWConv换成GhostModule，DWConv就不要再换成GhostModule了

是的，我只把PWConv替换成GhostModule。按您给的参数还是不行诶，和Conv2d差别很大

[iamhankai]

网络首尾层的Conv不要换，影响较大。另外，log能发来看看吗？

[yc-cui]

网络首尾层的Conv不要换，影响较大。另外，log能发来看看吗？

首尾的1x1卷积比较耗时吧，我想用GhostModule替换掉。我想要牺牲精度提高EfficientNet的效率，如果不替换首尾，那该替换哪呢？

[yc-cui]

这是替换之前的：

loading annotations into memory...
Done (t=0.01s)
creating index...
index created!
loading annotations into memory...
Done (t=0.00s)
creating index...
index created!
[Warning] Ignoring Error(s) in loading state_dict for EfficientDetBackbone:
	size mismatch for classifier.header.pointwise_conv.conv.weight: copying a param with shape torch.Size([810, 64, 1, 1]) from checkpoint, the shape in current model is torch.Size([63, 64, 1, 1]).
	size mismatch for classifier.header.pointwise_conv.conv.bias: copying a param with shape torch.Size([810]) from checkpoint, the shape in current model is torch.Size([63]).
[Warning] Don't panic if you see this, this might be because you load a pretrained weights with different number of classes. The rest of the weights should be loaded already.
[Info] loaded weights: efficientdet-d0.pth, resuming checkpoint from step: 0
[Info] freezed backbone
Step: 55. Epoch: 0/10. Iteration: 56/56. Cls loss: 1.52058. Reg loss: 2.57542. T
Val. Epoch: 0/10. Classification loss: 2.12950. Regression loss: 3.11589. Total loss: 5.24539
Step: 99. Epoch: 1/10. Iteration: 44/56. Cls loss: 1.35428. Reg loss: 3.01937. Tcheckpoint...
Step: 111. Epoch: 1/10. Iteration: 56/56. Cls loss: 1.15300. Reg loss: 3.10199.
Val. Epoch: 1/10. Classification loss: 1.50583. Regression loss: 2.84335. Total loss: 4.34918
Step: 167. Epoch: 2/10. Iteration: 56/56. Cls loss: 0.96853. Reg loss: 2.74884.
Val. Epoch: 2/10. Classification loss: 1.06936. Regression loss: 2.83312. Total loss: 3.90248
Step: 199. Epoch: 3/10. Iteration: 32/56. Cls loss: 0.98485. Reg loss: 2.56581. checkpoint...
Step: 223. Epoch: 3/10. Iteration: 56/56. Cls loss: 0.86978. Reg loss: 2.86461.
Val. Epoch: 3/10. Classification loss: 0.87215. Regression loss: 2.57996. Total loss: 3.45211
Step: 279. Epoch: 4/10. Iteration: 56/56. Cls loss: 0.88644. Reg loss: 2.45166.
Val. Epoch: 4/10. Classification loss: 0.81143. Regression loss: 2.66739. Total loss: 3.47882
Step: 299. Epoch: 5/10. Iteration: 20/56. Cls loss: 0.77803. Reg loss: 2.22479. checkpoint...
Step: 335. Epoch: 5/10. Iteration: 56/56. Cls loss: 0.77692. Reg loss: 2.67725.
Val. Epoch: 5/10. Classification loss: 0.77923. Regression loss: 2.72656. Total loss: 3.50579
Step: 391. Epoch: 6/10. Iteration: 56/56. Cls loss: 0.77824. Reg loss: 1.80291.
Val. Epoch: 6/10. Classification loss: 0.75823. Regression loss: 2.68391. Total loss: 3.44214
Step: 399. Epoch: 7/10. Iteration: 8/56. Cls loss: 0.83059. Reg loss: 2.75574. Tcheckpoint...
Step: 447. Epoch: 7/10. Iteration: 56/56. Cls loss: 0.72287. Reg loss: 2.84142.
Val. Epoch: 7/10. Classification loss: 0.72159. Regression loss: 2.79952. Total loss: 3.52111
Step: 499. Epoch: 8/10. Iteration: 52/56. Cls loss: 0.76601. Reg loss: 2.46491. checkpoint...
Step: 503. Epoch: 8/10. Iteration: 56/56. Cls loss: 0.73658. Reg loss: 2.11717.
Val. Epoch: 8/10. Classification loss: 0.70479. Regression loss: 2.62323. Total loss: 3.32802
Step: 559. Epoch: 9/10. Iteration: 56/56. Cls loss: 0.65757. Reg loss: 2.30679.
Val. Epoch: 9/10. Classification loss: 0.69079. Regression loss: 2.78529. Total loss: 3.47609

替换之后的：

loading annotations into memory...
Done (t=0.01s)
creating index...
index created!
loading annotations into memory...
Done (t=0.00s)
creating index...
index created!
[Warning] Ignoring Error(s) in loading state_dict for EfficientDetBackbone:
	size mismatch for classifier.header.pointwise_conv.conv.weight: copying a param with shape torch.Size([810, 64, 1, 1]) from checkpoint, the shape in current model is torch.Size([63, 64, 1, 1]).
	size mismatch for classifier.header.pointwise_conv.conv.bias: copying a param with shape torch.Size([810]) from checkpoint, the shape in current model is torch.Size([63]).
[Warning] Don't panic if you see this, this might be because you load a pretrained weights with different number of classes. The rest of the weights should be loaded already.
[Info] loaded weights: efficientdet-d0.pth, resuming checkpoint from step: 0
[Info] freezed backbone
Step: 55. Epoch: 0/500. Iteration: 56/56. Cls loss: 1.55586. Reg loss: 2.86230.
Val. Epoch: 0/500. Classification loss: 2.05592. Regression loss: 56.78297. Total loss: 58.83890
Step: 99. Epoch: 1/500. Iteration: 44/56. Cls loss: 1.39613. Reg loss: 2.59735. checkpoint...
Step: 111. Epoch: 1/500. Iteration: 56/56. Cls loss: 1.35521. Reg loss: 2.11275.
Val. Epoch: 1/500. Classification loss: 1.70027. Regression loss: 12.24190. Total loss: 13.94217
Step: 167. Epoch: 2/500. Iteration: 56/56. Cls loss: 1.39917. Reg loss: 3.25899.
Val. Epoch: 2/500. Classification loss: 1.40038. Regression loss: 5.59421. Total loss: 6.99460
Step: 199. Epoch: 3/500. Iteration: 32/56. Cls loss: 1.14606. Reg loss: 2.25805.checkpoint...
Step: 223. Epoch: 3/500. Iteration: 56/56. Cls loss: 0.98865. Reg loss: 2.36145.
Val. Epoch: 3/500. Classification loss: 1.18235. Regression loss: 3.75098. Total loss: 4.93333
Step: 279. Epoch: 4/500. Iteration: 56/56. Cls loss: 1.20481. Reg loss: 2.65963.
Val. Epoch: 4/500. Classification loss: 1.11369. Regression loss: 3.21690. Total loss: 4.33059
Step: 299. Epoch: 5/500. Iteration: 20/56. Cls loss: 1.13203. Reg loss: 2.45407.checkpoint...
Step: 335. Epoch: 5/500. Iteration: 56/56. Cls loss: 1.07970. Reg loss: 2.34846.
Val. Epoch: 5/500. Classification loss: 1.06054. Regression loss: 2.98431. Total loss: 4.04484
Step: 391. Epoch: 6/500. Iteration: 56/56. Cls loss: 1.38627. Reg loss: 2.23130.
Val. Epoch: 6/500. Classification loss: 1.00038. Regression loss: 2.97780. Total loss: 3.97818
Step: 399. Epoch: 7/500. Iteration: 8/56. Cls loss: 1.18486. Reg loss: 3.14593. checkpoint...
Step: 447. Epoch: 7/500. Iteration: 56/56. Cls loss: 0.94545. Reg loss: 2.42121.
Val. Epoch: 7/500. Classification loss: 0.96258. Regression loss: 3.05248. Total loss: 4.01506
Step: 499. Epoch: 8/500. Iteration: 52/56. Cls loss: 0.97633. Reg loss: 2.60206.checkpoint...
Step: 503. Epoch: 8/500. Iteration: 56/56. Cls loss: 0.91663. Reg loss: 2.54093.
Val. Epoch: 8/500. Classification loss: 0.93287. Regression loss: 3.02960. Total loss: 3.96247
Step: 559. Epoch: 9/500. Iteration: 56/56. Cls loss: 0.92465. Reg loss: 2.67289.
Val. Epoch: 9/500. Classification loss: 0.93854. Regression loss: 2.97227. Total loss: 3.91081
Step: 599. Epoch: 10/500. Iteration: 40/56. Cls loss: 0.99782. Reg loss: 2.43815checkpoint...

替换前后我都跑了500个epoch，替换之前效果挺好的，替换之后的loss几乎就降不下来了。

[iamhankai]

替换后的backbone也得先imagenet预训练，然后再在detection上finetune吧

[yc-cui]

替换后的backbone也得先imagenet预训练，然后再在detection上finetune吧

哦哦，好 那么这个将原来的1x1卷积替换为GhostModule应该是可行的吧

[iamhankai]

我觉得可行，不会掉点那么多的