Open ChuhuaW opened 1 week ago
Hi ChuhuaW,
thank you for your interest in our work. The missing files have been added. Please let me know whether the code works now.
@algvr Thanks! I’m able to get the model running. However, when I load the checkpoint "translated_ego4dv2.pth", I encountered aKeyError
:
File "/nfs/tynamo/home/data/cw234/transfusion/runner/nao/egov2_naf_trainer.py", line 65, in load_state_dict
if ckpt_pos_embedding.shape[1] < self.state_dict()[k].shape[1]:
KeyError: 'model.cross_fusion_layers.0.pos_embedding_layer.pos_embedding'
Upon further investigation, it seems there might be multiple keys in the state_dict with unexpected/missing/size mismatch. I’ve uploaded the full log here. Could you please take a look? Thank you.
Hi ChuhuaW,
thank you for providing the log file. I cannot reproduce the issue using our provided codebase. Are you trying to load translated_ego4dv2.pth
using the Trainer.fit
function of PyTorch Lightning, or perhaps using the --resume_from
argument? This file does not contain a checkpoint of our entire model, but rather the visual encoder weights from the original Ego4D baseline, which is the initialization point for our visual encoder. These weights are loaded automatically in the get_rcnn_model
function of modeling/obj_detection/rcnn_factory.py
when using the default configuration (pretrained
set to ${CODE}/checkpoints/translated_ego4dv2.pth
in the configuration YAML), so no further action is needed to load them.
Hi @algvr,
Thanks for the clarification! I’ll start training on my end and get back to you once I have some results. By any chances are you going to release the checkpoint our your entire model?
Hi,
Thanks for sharing the work. Would you mind also releasing the
EgoNaoDataset
class? Thedata_preproceessing/datasets/
folder is missing. And it would be helpful if you can also release the code formodeling/narration_embeds/datasets
.Thank you!
Best,