JasonQSY / vehicle-recognition

ROB 535 Perception Project

Vehicle Recognition from a Single Image

Code for reproducing the results in our ROB 535 Self-Driving Cars Perception Project, 2018 Fall.

Set up

• pytorch
• opencv
• numpy
• scipy
• imageio
• tqdm

Training

To set up,

mkdir exp

To start training a new model,

python train.py [-c checkpoint] -e experiment -t task

# train a model for task 1 from scratch
python train.py -e resnet50_bs16_2e-5_aug -t t1

# continue training from checkpoint resnet50_bs16_2e-5_aug_5 for task 1
python train.py -c resnet50_bs16_2e-5_aug_5 -e resnet50_bs16_2e-5_aug -t t1

# train a model for task 2 from scratch
python train.py -e resnet50_t2 -t t2

The training code would automatically save ${model}_${epoch} under exp. For example, if we train a model resnet for 10 epochs, there would be resnet_1, resnet_2, etc. under exp. These snapshots are used for validation.

Evaluation

To generate the prediction, check generate.py. It should store results into a result.csv or task2.csv file. You need specify checkpoint and task.

python generate.py -c resnet50_bs16_2e-5_aug_5 -t t1

Kaggle Submission

See https://github.com/Kaggle/kaggle-api

In general,

kaggle competitions submit -c fall2018-rob535-task1 -f result.csv -m "msg"