Vincent Duan
Vincent Duan
I am trying to reproduce the Highest Score on mini-Imagenet (80.51%). Resnet18 is mentioned to use as the Backbone; however, I wonder if Relation Model, Reshaper, Concentrator, and Projector both...
Besides, I found the code in the `code`, it seems reshaper, concentrator, and projector both build with `Bottleneck`, so I wonder if this means that even backbone are 4 Conv,...
具体是哪一个链接啊
Hi, I have the same issue too. I solved it by download it manually. 1. Download `LanguageTool-stable.zip` from [here](https://www.languagetool.org/download/). 2. unzip the `zip` file, then copy the whole unzipped file...
@WonderSeven I reproduce the MatchingNet for 5way-5shot on mini-Imagenet with train-aug reach acc 62.12% +- 0.69% (close to the report), PN with no train-aug reach acc 66.54% +- 0.69% (close...
@WonderSeven The code I used to train the model is exactly as same as the original one. For Matching Network, I didn't change anything. BTW, I also train on K100...
闭源的情况下甚至你很难判断这个是一个算法系统，还是是一个e2e的模型。
> Dear all, > > The query shuffle is enabled in the updated version. The CML-BGNN-3 can achieve 87.05% on the 5-way 1-shot classification task of the tiered-imagenet dataset, which...
一样的问题，在我的 case 中通过修改 `ppocr/modeling/backbones/vqa_layoutlm.py` 中修改 70行，在 from_pretrained 参数中增加 `ignore_mismatched_sizes=True` 可以解决这个问题。 从后续的实验来看，训练能够收敛，准确率和预期的结果相若。