WRL-Agriculture-Vision

This repository is the Team WRL's solution to the Agriculture-Vision Prize Challenge.

Generate result with trained models

1.1 Download all checkpoints in models/README.md and put all checkpoints in the model folder

1.2 Modify path in all configs. Check config/README.md for details. The default path for dataset is ./supervised/Agriculture-Vision-2021/ and ./raw/ (for semi-supervised learning only).

1.3 Run python test.py and python test-semi.py. We can also specify the input and output dir as well as the number of GPUs that are used. Check Reproduction of training models sections for more details.

1.4 By default, a directory called sumbission will store the submission for supervised track and a directory called `semi-submission will store the submission for the semi-supervised track.

1.5 Single model submission can be generated by python test-single.py -c <CONFIG>

Reproduction of solution

1. For supervised learning track

1.1 Modify path in all configs. Check config/README.md for details.

1.2 Train those two models which are required 4 X 2080 Ti GPUs:

python main.py -c config/DeepLabV3Plus-efficientnet-b3.yaml
python main.py -c config/FPN-efficientnet-b4.yaml

1.3 Train those two models, which are required 8 X 2080 Ti GPUs of two nodes, check here to set how to train models on more than one node:

python main.py -c config/DeepLabV3Plus-efficientnet-b5.yaml
python main.py -c config/FPN-efficientnet-b5.yaml

1.4 Run test script:

python test.py -o <OUT-PATH>

<OUT-PATH> represents the out path of submission, default value is submission.

If 4 GPUs are used, could run code:

  python test.py -f0 -t4 \
& python test.py -f1 -t4 \
& python test.py -f2 -t4 \
& python test.py -f3 -t4

2. For semi-supervised learning track

2.1 Train all models in supervised section 1 (supervised learning track)

2.2 Run python gen-semi-data.py, check its CLI for more details.

2.3 Run python gen-semi-label.py, check its CLI for more details.

2.4 Train all four models again and use *-semi.yaml to train models again use similar command e.g:

python main.py -c config/DeepLabV3Plus-efficientnet-b3-semi.yaml

2.5 Run test script, the arguments are the same as test.py while and default value of <OUT-PATH> is set to semi-submission:

python test-semi.py

Touble Shooting

Train models with more than one node will need install lastest version of torchmetrics from github. The old version is wrongly set-up. check requirements.txt for details or use: