Poor results with pre-trained model. #140
Answered
by
rwightman
darleybarreto
asked this question in
Q&A
-
|
Hi, I've tried to run a basic coco validation pre-trained model (as in here) and got very bad results: Do you have any idea what could it be? |
Beta Was this translation helpful? Give feedback.
Answered by
rwightman
Dec 3, 2020
Replies: 4 comments 3 replies
-
|
@darleybarreto are you using a model that has pretrained weights defined? I don't think I emit a warning for this codebase if there are no weights. |
Beta Was this translation helpful? Give feedback.
1 reply
-
|
I tested with that one too, this is the output: And training one epoch with this command: python train.py $DIRPATH --model tf_efficientdet_d2 -b 10 --amp --lr .06 --sync-bn --opt fusedmomentum --warmup-epochs 5 --lr-noise 0.4 0.9 --model-ema --model-ema-decay 0.99995 --output $DIRPATH --pretrainedgives this: I can't see what's wrong :( |
Beta Was this translation helpful? Give feedback.
-
|
It's best to look at the source of truth, the model config definitions, and look at the presence of URL for pretrained weights https://github.com/rwightman/efficientdet-pytorch/blob/master/effdet/config/model_config.py |
Beta Was this translation helpful? Give feedback.
0 replies
-
|
@darleybarreto I don't know what the problem is... the current master branch has been verified on multiple machines, ubuntu 16.04, 18.04, 20.04 with volta, turing, pascal GPU and also one Windows 10 system. Just running on the system I'm on right now gives below... Make sure pycocotools==2.0.2 as mentioned in README. PyTorch 1.6 or 1.7 recommended now. |
Beta Was this translation helpful? Give feedback.
1 reply
-
|
I figured it out, I tweaked this init here doing this: def __init__(self, data_dir, parser=None, parser_kwargs=None, transform=None):
super(DetectionDatset, self).__init__()
parser_kwargs = parser_kwargs or {}
self.data_dir = data_dir
if isinstance(parser, str):
self._parser = create_parser(parser, **parser_kwargs)
else:
assert parser is not None and len(parser.img_ids)
self._parser = parser
self._transform = transform
self.img_ids = self._parser.img_ids
shuffle(self.img_ids) # <--- this is the problemAnd then I'm using a faction of the |
Beta Was this translation helpful? Give feedback.
-
|
@darleybarreto of course that's a problem, I use image index to reference dataset gt for evaluation, not the id because the id isn't always an integer for some datasets and I can't use strings reasonably in tensors. Any modifications should be mentioned before asking me others for help. If you want to remove a subset you need to do it like SkipSubset class I wrote that's already here, remap the indices. |
Beta Was this translation helpful? Give feedback.
1 reply
-
|
Oh, I'm so sorry about that, I thought those were simply indices I could play around with. Thank you so much for your attention! |
Beta Was this translation helpful? Give feedback.
Answer selected by
darleybarreto
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
@darleybarreto of course that's a problem, I use image index to reference dataset gt for evaluation, not the id because the id isn't always an integer for some datasets and I can't use strings reasonably in tensors. Any modifications should be mentioned before asking me others for help. If you want to remove a subset you need to do it like SkipSubset class I wrote that's already here, remap the indices.