zihaozhang9

Input_size=416 . So the mobilenet model may be larger than in the paper. > Hello, thank you very much for your work. After training according to your method, the result...

> @zihaozhang9 train_search.py里的model就是要搜索的网络 > nao包含的就是encoder-predictor-decoder结构 > model的valid accuracy作为评价该model的指标，encoder-predictor要学的就是从model到它的valid accuracy这的这种映射关系。 非常感谢

> @zihaozhang9 train_search.py里的model就是要搜索的网络 > nao包含的就是encoder-predictor-decoder结构 > model的valid accuracy作为评价该model的指标，encoder-predictor要学的就是从model到它的valid accuracy这的这种映射关系。 我貌似看懂了一些。 1. 先训练model，在[499行nao_train_dataset](https://github.com/renqianluo/NAO_pytorch/blob/master/train_search.py#L499)变量收集训练的acc，做为nao的训练数据。 2. nao将学到的结构在[549行child_arch_pool](https://github.com/renqianluo/NAO_pytorch/blob/master/train_search.py#L549) 变量反馈给model，改变网络结构。 3. model的最初始结构在[450行utils.generate_arch](https://github.com/renqianluo/NAO_pytorch/blob/master/train_search.py#L450)随机定义了一个结构。 4. 然后训练时model的forward中动态改变网络结构，在model_search.py文件中NASNetworkCIFAR的forward由arch控制[#L220](https://github.com/renqianluo/NAO_pytorch/blob/master/model_search.py#L220)

Hello! I can't download the model. Is there any other download address?

我已收到您发送的邮件。我会及时查看。