search

autonomousvision / transfuser

[PAMI'23] TransFuser: Imitation with Transformer-Based Sensor Fusion for Autonomous Driving; [CVPR'21] Multi-Modal Fusion Transformer for End-to-End Autonomous Driving
MIT License
1.04k stars 175 forks source link

Finetune without `optimizer.pth` #196

Closed MaiRajborirug closed 7 months ago

MaiRajborirug commented 7 months ago

As a followup from ISSUE 194 The given pre-train weights don't have optimizer.pth along with it. However, when we load weight states. We are supposed to load an optimizer as shown in train.py line 183. Since we didn't have optimizer.pth. from the pre-trained folder What should I do to make the training go smoothly?

Kait0 commented 7 months ago

Just comment that line. I would expect that supervised finetuning works if you start with fresh adam momentums. Similar than starting a new training with pre-trained image net weights (where you also don't load the optimizer weights).

MaiRajborirug commented 7 months ago

Got it thank you so much!

MaiRajborirug commented 7 months ago

Then what should be the learning rate? Should I keep learning rate 10e-4 or should I take into account that the model already trained for 41 epoch and reduce the start learning rate in the model?

Kait0 commented 7 months ago

Well its not possible to predict the appropriate learning rate for a task in advance. Start with the default in the repository and test different values.