pytorch-auto-drive
pytorch-auto-drive copied to clipboard
voldemortX
fine-tuning a llamas model with culanes data
I have a model trained on llamas dataset using erfnet_scnn_llamas configuration and would like to introduce culanes data set to this model as well. I use --checkpoint to load the model and used culanes config file for the second round training on culanes. However I get an error because flattened_size=4500, for erfnet_culanes.py config file and it is 4400 in the erfnet_llamas.py. This causes an error while loading the checkpoint. I tried to change the input size for culane config file and change the model parameters as well but still getting error. Could you please guide me how to solve this issue?
@MehdiAlilou I have never done this before. Could you share your latest culane fine-tuning config file?
I used the default config file available in /configs/lane_detection/scnn/erfnet_culane.py. The input size is different from the llamas' config file. I tried to make the same input size as the llamas' config file but it did not work. Previously I used erfnet_llamas.py to train on llamas.
@MehdiAlilou I can't use my testing env for now. But I think this config could do the finetuning, maybe try it out? culane_finetune.zip
@voldemortX thanks for your prompt response. I will test it tomorrow and will let you know the result. By the way I had another question while using --train mode how to involve the validation set so we can save improving checkpoints?
Unfortunately, we did not leave that functional for lane detection. You will have to do some hacking yourself.
As earlier discussed in #79 . However, in your finetuning case checkpointing may really be important. Perhaps you can add a bit code to save checkpoints every some iterations. And we will welcome a PR if this is implemented through command line option like --checkpointing-steps for example.
I actually did some changes to save every couple of epochs, just wanted to double-check with you. I can work on that.
This issue seems to be addressed. I'll close it for now. Feel free to continue commenting for reopen/open a new one.