RuntimeError: The size of tensor a (32) must match the size of tensor b (64) at non-singleton dimension 2

[liuyjchj1]

Is the input image size of the SAM model fixed at 1024? It seems that an error occurs when inputting images of other sizes. I tried using a 512-sized image and it resulted in a RuntimeError: The size of tensor a (32) must match the size of tensor b (64) at non-singleton dimension 2.

[liuyjchj1]

我输入的图像尺寸设置为512，但是报错为RuntimeError: The size of tensor a (32) must match the size of tensor b (64) at non-singleton dimension 2，查了一下，发现SAM模型本身的输入就是1024，那怎么训练其他尺寸的数据呢？是硬padding为1024吗？

[LinhuiHu]

sam源码中的sam.py有一个process函数是这样直接padding的,感觉训练应该也是吧

[Issues-translate-bot]

Bot detected the issue body's language is not English, translate it automatically. 👯👭🏻🧑‍🤝‍🧑👫🧑🏿‍🤝‍🧑🏻👩🏾‍🤝‍👨🏿👬🏿

---

The sam.py in the sam source code has a process function that directly pads in this way. I feel that the training should be the same.

[WuJunde]

yes, you need to change some hyper-parameters to train image with other sizes. if you want to solve it in preprocessing, I will recommend resize, instead of padding.

[liuyjchj1]

yes, you need to change some hyper-parameters to train image with other sizes. if you want to solve it in preprocessing, I will recommend resize, instead of padding.

Thank you！ I will follow the method you suggested！

[gm1019]

So this still mean resizing the image to 1024?

[ShChen233]

yes, you need to change some hyper-parameters to train image with other sizes. if you want to solve it in preprocessing, I will recommend resize, instead of padding.

Thank you！ I will follow the method you suggested！

所以是重新训练的吗

[WuJunde]

image resolution now can be resized, just set -image_size parameter in the running. check my newest update.