DtYXs

> > https://openaipublic.azureedge.net/clip/models/3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02/ViT-L-14-336px.pt > > https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt > > how to download other model? the link of ViT-B/32 in `clip/clip.py` _MODELS = { "RN50": "https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt", "RN101": "https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt", "RN50x4": "https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt", "RN50x16": "https://openaipublic.azureedge.net/clip/models/52378b407f34354e150460fe41077663dd5b39c54cd0bfd2b27167a4a06ec9aa/RN50x16.pt",...

when START(i) < START(j)

> I agree that this function is wrong, but I don't think the spans "si.i1 < sj.i1

> Yes, these spans should be accepted. For example: > "Mister Smith and his wife Missus Smith..." > We have that "Mister Smith", "Missus Smith" and "Mister Smith and his...

可以参考[预测及评估](https://github.com/OFA-Sys/Chinese-CLIP?tab=readme-ov-file#%E5%9B%BE%E6%96%87%E7%89%B9%E5%BE%81%E6%8F%90%E5%8F%96)部分的流程，将`resume`替换为保存的模型路径

您好，微调时没有用两阶段的方式，可以参考我们的[技术报告](https://arxiv.org/pdf/2211.01335.pdf)中给出的超参数配置。给出的脚本默认超参应该不适合1*V100训练，需要自行调整一下超参设置。

您好，错误看起来是`args`没有解析到`use_flash_attention`参数，但这个错误看起来比较奇怪，因为`use_flash_attention`默认是`False`，您可以检查一下代码和训练脚本。

Hello, RN50 is pretrained at 224 input resolution, and you may need to resize the input image to 224. If you want to train with your custom ResNet, you may...

您好，请问能提供下您的pytorch和flash_attention版本吗。支持flash_attention时的官方版本是0.2.8，已经比较旧了，可能需要适配新版本。

您好，可能是由于Windows系统不支持NCCL分布式后端，可以参考 https://github.com/OFA-Sys/Chinese-CLIP/issues/57#issuecomment-1439787723 尝试将分布式后端设置为gloo看看，更建议在Linux系统训练