XIANG Shuai comments

Repositories
Issues
Comments

Results 4 comments of


                                            XIANG Shuai

datasets for the first-stage autoencoders

according to the paper(Table 8.), the autoencoders are trained on OpenImages, evaluated on ImageNet-Val

Wired loss and reconstruction results when training the autoencoder

I think it might be that kl_weight is too small, you can try increasing kl_weight in the config

Allow user change photo labels (and update Container.item_label) at the same time.

You are changing the "ms.photos.label". However, when using the `roi.back2raw()` function, the parameter passed is "ms.photos.item_label", which is a dictionary, different from "ms.photos.label" So to change the labels correctly, you...

Crop the ROI along image boundarys

Suppose the size of the image is 1600*900, and the coordinates of a vertex in the ROI are (1200, 1000). The existing method seems to change the vertex coordinates to...