search

automl / NASLib

NASLib is a Neural Architecture Search (NAS) library for facilitating NAS research for the community by providing interfaces to several state-of-the-art NAS search spaces and optimizers.
Apache License 2.0
528 stars 117 forks source link

Different taskonomy n. classes #190

Open casarinsof opened 2 months ago

casarinsof commented 2 months ago

Hi, I am using the code, and I noticed that in the graph creation of TransBench101 both Micro and Macro the number of classes is different from those reported on the original work

See line 44 https://github.com/yawen-d/TransNASBench/blob/training_code_with_config/configs/train_from_scratch/class_scene/config.py, and all the config files, and the original paper https://arxiv.org/pdf/2105.11871 Table 1.

While in this work in naslib>search_spaces>transbench101>graph.py the classes are so defined

    if dataset == "jigsaw":
        self.num_classes = 1000
    elif dataset == "class_object":
        self.num_classes = 100 # why not 75?
    elif dataset == "class_scene":
        self.num_classes = 63 # why not 47??

Is ther a reason for doing so? How does this affect the results?