PangziZhang523
PangziZhang523
Will there be water drop artifacts when replacing ModulatedConv2d with Conv + AdaIN? I happened to encounter this problem recently. When I used the encoder-decoder to generate the feature map...
I tried BN and LN, but there are still water drop artifacts. My artifact is in the form of half purple and half white. Is there any other way? Thanks...
I used IN at the beginning, and my task is to regress a feature map. I don’t want any water drop artifacts on the feature map.I don't know what to...
I'm trying to replace traditional convolution and IN with modulation convolution. However, it is not known whether to replace all or only the decoder part.
官方能给一下多轮对话的数据格式嘛？
> https://github.com/THUDM/VisualGLM-6B/blob/f4429a009ee533b76e8757dce6917fbf0b0408f9/finetune_visualglm.py#L118-L120 > > ``` > xxx问：xxx > 答：xxx > 问：xxx > 答：xxx > ``` 这样还是有前面老哥提到的问题，轮数越多越复杂了
with open('img_path', 'rb') as image_file: image_data = image_file.read() image_encoded = base64.b64encode(image_data).decode('utf-8') 这样可以正常调用
 一直卡在这里，正常嘛？
对Slice Encoding的Positional Embedding Interpolation 这部分代码是在哪里实现的