Open afiff2 opened 5 days ago
Hello, you can use this config file if you want to train a faster-rcnn for the VG150 dataset.
By the way, please be kind when posting issues on Github, I am working on this codebase alone in my free time and I am providing it entirely for free for the community. I also mentioned in the readme that this codebase is highly experimental. Right now I don't have a lot of time to maintain it properly, however Github is also a place where people can contribute, so if you find mistakes you can create a pull request and I would be happy to merge it.
I am not a native English speaker. I sincerely apologize if my previous message came across as unkind. I truly appreciate the hard work you have put on SGG..
Yes the last part came out a bit rude to be honest ah ah But it's ok we're cool now :)
I have encountered several issues while attempting to use Faster R-CNN pre-training. The problems include, but are not limited to, the following:
Configuration File Adjustments: In the config file, the setting INPUT:
MIN_SIZE_TRAIN:(600,)
should be modified toMIN_SIZE_TRAIN: 600
. In theMODEL: BACKBONE: section, CONV_BODY: "R-101-FPN"
should be changed toTYPE: "R-101-FPN"
.Modifications in detector_pretrain_net.py:
The following code block:
if args.use_wandb: run_name = cfg.OUTPUT_DIR.split('/')[-1] if args.distributed: wandb.init(project="scene-graph-benchmark", entity="maelic", group="DDP", name=run_name, config=cfg) wandb.init(project="scene-graph-benchmark", entity="maelic", name=run_name, config=cfg) use_wandb = True
should be updated to include:else: use_wandb = False
Changes in paths_catalog.py: Comment out:
for k, v in args.items(): args[k] = os.path.join(data_dir, v)
Due to a lack of code management, I am uncertain whether these are all the modifications I made, and whether the training process will function correctly after these changes.