Possible bug when loading GTSRB dataset about trojanzoo

[ain-soph]

Thanks for your bug report!
I admit there are potentially quite some bugs after so many times code re-organization. And we have no unit-test scripts yet.
I'll figure it out in one day, (hopefully).

A suggestion about your test.py is that you should always enable environ like the examples, even though it doesn't matter in this small scenario. It allows you to set config_path, random_seed, default cuda device and some other basic environment settings.

Here's the modified test.py: (run with python test.py --verbose 1 --color --tqdm --dataset gtsrb --download)

# -*- coding: utf-8 -*-

import trojanvision.environ
import trojanvision.datasets
from trojanvision.utils import summary
import argparse


parser = argparse.ArgumentParser()
trojanvision.environ.add_argument(parser)
trojanvision.datasets.add_argument(parser)
args = parser.parse_args()

env = trojanvision.environ.create(**args.__dict__)
dataset = trojanvision.datasets.create(**args.__dict__)
if env['verbose']:
    summary(env=env, dataset=dataset)

Btw, data_format is a very new and unstable argument that haven't been tested well. I'm currently working on it. The basic idea is to load data into memory to avoid latency for data loading on HDD disk. Keep it as default for now.

And another question is, are you using the pip to install trojanzoo? Although it's convenient, sometimes it'll be a little out-of-date. Try to git clone the repo and put test.py in the root dir.

---

And I guess it's possibly already fixed with the recent commits :)

[liu00222]

Many thanks for the response! I think my problem is gone now.