Runist

Results 51 comments of Runist

> In addtion, did you try to use a large scale neural network training by a first-order approximation. Can it get a better result on dataset like miniImageNet? I think...

> > > In addtion, did you try to use a large scale neural network training by a first-order approximation. Can it get a better result on dataset like miniImageNet?...

> > > > > In addtion, did you try to use a large scale neural network training by a first-order approximation. Can it get a better result on dataset...

> 好吧,所以我才希望看到她本人来回答一下。 根据李宏毅教授的讲解,他是把二阶导近似等效为0或1。但是这样对结果的精度不好。

> > > 好吧,所以我才希望看到她本人来回答一下。 > > > > > > 根据李宏毅教授的讲解,他是把二阶导近似等效为0或1。但是这样对结果的精度不好。 > > 这个有讲义资料吗?求地址啊 B站上搜“李宏毅”第一个2020的你拉到下面有个Meta-Learning章节的,就是了

同问,请问你解决了嘛? > 想问一下如何训练可以得到您给出的在voc上77mAP的权重?可以问一下训练策略吗?我在你的代码中注释了对预训练文件的引入,直接使用原始ResNet网络进行train.py文件中100epoch的训练,得到的训练效果很差?想问怎样可以训练出和您相同的voc的结果呢?

> 呃,主要是我试了你用pretrain=True也是这样30%的map

# Train ![image](https://user-images.githubusercontent.com/38599727/168017648-6f7fb716-6941-472d-a7ef-60346981b442.png) ![image](https://user-images.githubusercontent.com/38599727/168017693-3811fd1e-a3eb-4b91-9e1b-7e7b704c06b1.png) # Train one epochs ![image](https://user-images.githubusercontent.com/38599727/168017858-b96e6fcf-1909-4494-8031-fdcfdc70f069.png) ![image](https://user-images.githubusercontent.com/38599727/168017897-eb220b4c-ac53-4451-b814-63e3e6b46afb.png) ![image](https://user-images.githubusercontent.com/38599727/168017952-c70e5201-f7ab-4840-87b4-3620428ee951.png) # Loss ![image](https://user-images.githubusercontent.com/38599727/168196031-7c9b2feb-0795-4ad0-b4fd-46926178d35c.png) ![image](https://user-images.githubusercontent.com/38599727/168196064-2ca88682-728a-4810-b440-cdb65c108606.png) ![image](https://user-images.githubusercontent.com/38599727/168196076-686545b4-78c8-448d-8d9b-fa9402e080ca.png) ![image](https://user-images.githubusercontent.com/38599727/168196096-1aeb9460-1ba4-4e48-b35e-5e30435341b1.png) # Dataloader ![image](https://user-images.githubusercontent.com/38599727/168196194-ce1d1e5c-b7d4-488c-a86d-01a403824834.png) ![image](https://user-images.githubusercontent.com/38599727/168196213-40c3df39-8b89-4da6-a2f6-5c4d2a249804.png) ![image](https://user-images.githubusercontent.com/38599727/168196230-31f26f96-70b3-4c90-b2a7-b88a6c7f90be.png) DataLoader 和 Loss基本按照你的方式写的。训练策略稍有改动,但我感觉Loss和data都一样的情况,应该没啥问题。

是呀,跟着你的讲解自己又重新加了一点。

> 找个我是在外面进行了permute,维度顺序是一致的,不然也没法计算。不过我想和你确认的是,你是基于resnet50 backbone的预训练权重,然后fine-tune就可以在VOC上达到77的MAP吗?那此时的val loss是多少呢?我也是采用resnet的backbone预训练权重,也没法达到这个77,而且val loss达到2.5的时候,就会反向上升(虽然我看了下原作者的repo说hm loss是正常的)。模型确实是学到东西了,直接测训练集Map是89,但验证集只有40多。很难受。。。