search

Event-AHU / OpenPAR

[OpenPAR] An open-source framework for Pedestrian Attribute Recognition, based on PyTorch
MIT License
63 stars 6 forks source link

Issues with PA100k.py and AttrDataset.py #18

Closed erictan23 closed 4 months ago

erictan23 commented 6 months ago

Hi! I am trying out SequencePAR in this repository. I followed the steps for preprocessing of PA100k data, however i encountered some problems when I run the training script.

In the dataset_info OrderedDict there is no .attribute feature, .neg_attr_word, .expand_pos_attr_words, .expand_neg_attr_words, may I know what should I do to replace them for PA100k? image

So i tried to replace it with .attr_vector and got this error. May I know what is .attribute so that I can replace it correctly? image

1125178969 commented 5 months ago

[neg_attr_word, .expand_pos_attr_words, .expand_neg_attr_words] is useless, you can just comment them out

1125178969 commented 5 months ago

.attribute refers to a list of attributes, i.e. attr_words

erictan23 commented 5 months ago

Hello! Thanks for the help. I have tried replacing .attribute with .attr_words. It works, however, I am unable to train the model due to cuda out of memory issues.

Is there any other way that I am able to try sequencePAR via inference??

1125178969 commented 4 months ago

One can try to use ViT-B to replace the backbone or fp16, which are not implemented in our submission and which will somewhat impair the model performance.