search

hellozhuo / pidinet

Code for the ICCV 2021 paper "Pixel Difference Networks for Efficient Edge Detection" (Oral).
Other
458 stars 70 forks source link

Question about the trained models #15

Open zhuoinoulu opened 2 years ago

zhuoinoulu commented 2 years ago

(1) In your github repository, you have provided the pre-trained models in Google Driver, but, I fail to find any .pth or .pt file. There are only one file named "data.pkl", and some digitally named files without any suffix. So I wonder where to find your pretrained pytorch models? A: Hi, to prevent confusing, I have changed the suffix ".tar" to ".pth" and loaded the models in the directory trained_models. They can be downloaded directly now.

(2) Your paper shows that the size of params. of HED and RCF are about 14.7 and 14.8 Mb in Table 1, respectively. However, their released pre-trained models are about 50Mb, which can be directly downloaded from the corresponding github repository. So I wonder what does the reported size of params. in the first row of Table 1 mean, does it mean the size of saved .pth file, or something else? A: Actually, the downloaded model file is usually larger than the model itself, due to the existence of other things in the file. For example, the optimizer, which can also take much storage space, and others. Specifically, I used the following functions to calculate the model size: https://github.com/zhuoinoulu/pidinet/blob/781924fe30469cdc64f63ce6666a3e1f5b4e576f/utils.py#L33-L36

Hope my answers can help~

ForawardStar commented 2 years ago

Hello, one more questions. Can your also provide the codes for calculating the multiply accumulates (MACs) illustrated in Table 1 of your paper? Thanks