Are config files required for efficientdet models ?

#1
by Manish1903 - opened

I am trying to load the model in a controlled environment (firewall blocked) by downloading the model weights manually from this repository.

import layoutparser as lp
model = lp.AutoLayoutModel(config_path = "models/PubLayNet/tf_efficientdet_d1/publaynet-tf_efficientdet_d1.pth.tar")

gives "ValueError: Invalid model config_path. Please suggest how to resolve this."

Thanks

I believe it should be

import layoutparser as lp
model = lp.models.effdet.layoutmodel.EfficientDetLayoutModel('tf_efficientdet_d1', model_path='path_to_model_file', label_map={0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"})

There also seems to be a weird bug with either the layoutparser package or the effdet package where the head doesn't reset when loading the model, resulting in the following error

size mismatch for class_net.predict.conv_pw.weight: copying a param with shape torch.Size([810, 88, 1, 1]) from checkpoint, the shape in current model is torch.Size([45, 88, 1, 1]).

This stems from the effdet library not resetting the head before loading the pretrained model. You can fix this in your effdet library by editing the factory.py file.

Sign up or log in to comment