search

allenai / interactron

A Model for Embodied Adaptive Object Detection
Apache License 2.0
43 stars 7 forks source link

FileNotFoundError #3

Open Cccanvas opened 2 years ago

Cccanvas commented 2 years ago

Thanks very much for providing the codes! I tried to train the interactron model with the commond:

python train.py --config=configs/interactron.yaml

But some data is missing:

File "/disk1/chenze/code/interactron/interactron-main/datasets/sequence_dataset.py", line 60, in getitem frame = Image.open(img_path) File "/disk1/chenze/anaconda3/envs/interactron/lib/python3.8/site-packages/PIL/Image.py", line 3068, in open fp = builtins.open(filename, "rb") FileNotFoundError: [Errno 2] No such file or directory: 'data/interactron/train/FloorPlan6_00014/pos=[-1.12,0.90,-1.49]_rot=[240deg].jpg'

I have downloaded the data you provided, but this problem still exists. Directory FloorPlan6_00014 does not exist: FloorPlan6_00000/ FloorPlan6_00002/ FloorPlan6_00004/ FloorPlan6_00007/ FloorPlan6_00010/ FloorPlan6_00013/ FloorPlan6_00018/ FloorPlan6_00020/ FloorPlan6_00023/ FloorPlan6_00001/ FloorPlan6_00003/ FloorPlan6_00006/ FloorPlan6_00008/ FloorPlan6_00011/ FloorPlan6_00016/ FloorPlan6_00019/ FloorPlan6_00022/ FloorPlan6_00024/

Any suggestions for this problem? Thanks!

GilgameshD commented 2 years ago

I also found that some training data folders are missing. Any solutions here?

klemenkotar commented 2 years ago

Hey @GilgameshD and @Cccanvas. Thank you for your interest. I am currently working on a big update to address several issues including this one. I will let you know as soon as I push it.

GilgameshD commented 2 years ago

Thanks for your reply @klemenkotar. One follow up question, is it possible to release the scripts for generating this dataset in AI2-THOR? That would help a lot of active vision tasks!

klemenkotar commented 2 years ago

Hi @GilgameshD and @Cccanvas,

I have pushed a solution to this problem as well as a few other issues. After running some additional tests we discovered that there was quite a bit of instability in the training process. We also found a bug in the saving code which kept the best-performing checkpoint around. After fixing the code and re-running the experiments we found that the main trends still hold but the numbers are slightly different. You can find the updated numbers in our latest arxiv update.

You will need to re-download both the dataset and model checkpoint tar files to get things working. We also increased the training duration so the model should take about 5 days to train on a single high-power gpu machine.

@GilgameshD I have also added the script for collecting the AI2-THOR data. It contains a few settings that can be easily modified to adjust the data collection parameters.

If you have any questions about the work feel free to also reach out to me directly at klemenk@allenai.org.