search

mseg-dataset / mseg-semantic

An Official Repo of CVPR '20 "MSeg: A Composite Dataset for Multi-Domain Segmentation"
MIT License
460 stars 78 forks source link

Format of weights #8

Closed soumik12345 closed 4 years ago

soumik12345 commented 4 years ago

Hi, can you please tell me if the weights include the model structure or it's just the weights? If the last is true then to which model these weights refer to since under the models dir there are several model implementations?

johnwlambert commented 4 years ago

Hi @soumik12345 , thanks for your interest in our work.

The pre-trained models follow the HRNet-W48 architecture. The model structure is defined in the code here.

The saved weights provide a dictionary between keys (unique IDs for each weight identifying the corresponding layer) and values (the floating point weights).

AvivSham commented 4 years ago

Hi @soumik12345 and @johnwlambert , I have another question please, what was the image shape HRNet-W48 model was trained on?

Thank you for helping!

johnwlambert commented 4 years ago

Hi @AvivSham, please refer to https://github.com/mseg-dataset/mseg-semantic#experiment-settings.

At training time, images are resized to 1080p, then a 713x713 crop is taken for almost all models. There are only two exceptions: MSeg-3m-480p and MSeg-3m-720p (resized to 480p or 720p, respectively, before a crop is taken).

AvivSham commented 4 years ago

@johnwlambert thank you for the information!