Kaggle-Featured Code Competition
-
In this competition, you’ll apply your data science skills to build motion prediction models for self-driving vehicles.
-
You'll have access to the largest Prediction Dataset ever released to train and test your models.
-
Your knowledge of machine learning will then be required to predict how cars, cyclists,and pedestrians move in the AV's environment.
scenes, aerial_map, semantic_map
sample.zarr, test.zarr, train.zarr, validate.zarr, mask.npz *
scenes, frames, agents, traffic_light_faces, agents_mask**
*: In test, the mask (provided in files as mask.npz) masks out any test object for which predictions are NOT required.
**: A mask that (for train and validation) masks out objects that aren't useful for training.
- aerial_map - an aerial map used when rasterisation is performed with mode "py_satellite"
- semantic_map - a high definition semantic map used when rasterisation is performed with mode "py_semantic"
- sample.zarr - a small sample set, designed for exploration
- train.zarr - the training set, in .zarr format
- validate.zarr - a validation set (roughly the size of train)
- test.csv - the test set, in .zarr format
- mask.npz - a boolean mask for the test set. All and only the agents included in the mask should be submitted
- *sample_submission.csv - two sample submissions, one in multi-mode format, the other in single-mode
In order for the "Submit to Competition" button to be active after a commit, the following conditions must be met:
- CPU Notebook <= 9 hours run-time
- GPU Notebook <= 9 hours run-time
- TPU Notebook <= 3 hours run-time
Freely & publicly available external data is allowed, including pre-trained models.
Submission file must be named submission.csv
https://github.com/lyft/l5kit/blob/master/competition.md
https://github.com/lyft/l5kit/blob/master/data_format.md
https://numpy.org/doc/stable/user/basics.rec.html
https://pytorch.org/tutorials/beginner/saving_loading_models.html
https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py
https://ai.googleblog.com/2019/05/efficientnet-improving-accuracy-and.html
https://github.com/lyft/l5kit/blob/master/README.md
https://github.com/lyft/l5kit/blob/master/competition.md
https://github.com/lyft/l5kit/blob/master/coords_systems.md
https://github.com/lyft/l5kit/blob/master/data_format.md
https://github.com/lyft/l5kit/blob/master/how_to_contribute.md
- pytorch --- improve the baseline of Peter’s -> primary process
- keras --- improve the baseline of GoSabres’s
- tensorflow --- improve the baseline of Ashi’s
- include our best model if possible
- focus on emsamble last few weeks of End date
-
Pytorch process
-
Dependency of parameters/models on LB score
-
Current Issue
- Model
- Parameters
- Save the model
- load the model
- Test the model
- raster size (384, 384) : LB 3593.289
- raster size (300, 300) : LB 2354.393
- raster size (224, 224) : LB 1584.348
- batch size 12 : LB 1584.348 #raster size (224, 224), num_steps 10,000
- batch size 32 :
- batch size 64 : evaluating --> Took much time to compute (>9 hours)
- num_steps 10,000 : LB 1584.348 #raster size (224, 224), batch size 12
- num_steps 15,000 :
- num_steps 20,000 :
- lr 1e-2 :
- lr 1e-3 : LB 1584.348
- lr 1e-4 :
- optimizer Adam : LB 1584.348
- optimizer SGD :
- optimizer ASGD :
- Resnet 18 :
- Resnet 50 : LB 1584.348
- EfficientNet :
-> W/ kaggle notebook is required in Testing/Prediction
-> Separate Training from Testing/Prediction
-> W/O kaggle notebook in Training
model_resnet34_output_1000-chop.pth 'lr': 1e-4 (--- ver.57)
#nn.Dropout(0.2) LB: 2840.352 --- ver.65
nn.Dropout(0.2) LB: 84.224 --- ver.64
nn.Dropout(0.2) LB: 84.211 --- ver.60
0918-predictor-full/0918_predictor_full.pt
#nn.Dropout(0.2) LB: 451990.306 --- ver.62
nn.Dropout(0.2) LB: 7469.120 --- ver.61
model_multi_update_lyft_public.pth
nn.Dropout(0.2) LB: --- ver.67
- Combine solutions [1, 2, 3, 4]
- weights = [0.3, 0.7] LB:46.603
- weights = [0.35, 0.65] LB:46.370
- weights = [0.4, 0.6] LB:46.189 ---Original Score
- weights = [0.420, 0.580] LB:46.084
- weights = [0.425, 0.575] LB:46.069
- weights = [0.430, 0.570] LB:46.061
- weights = [0.435, 0.565] LB:46.059 ---Got Best Score
- weights = [0.440, 0.560] LB:46.062
- weights = [0.45, 0.55] LB:46.082
- weights = [0.475, 0.525] LB:46.179
- weights = [0.5, 0.5] LB:46.331
- weights = [0.5, 0.5] LB:46.641
- weights = [0.4, 0.6] LB:45.064
- weights = [0.3, 0.7] LB:44.346
- weights = [0.15, 0.85] LB:42.173
- weights = [0.105, 0.895] LB:41.855
- weights = [0.1, 0.9] LB:41.840
- weights = [0.095, 0.905] LB:41.830
- weights = [0.090, 0.910] LB:41.825 ---Got Best Score
- weights = [0.085, 0.915] LB:41.825
- weights = [0.080, 0.920] LB:41.829
- weights = [0.05, 0.985] LB:41.947
- weights = [0.5, 0.5] LB:132.087
- weights = [0.105, 0.895] LB:34.058
- weights = [0.1, 0.9] LB:33.756
- weights = [0.095, 0.905] LB:33.745
- weights = [0.085, 0.915] LB:33.726
- weights = [0.075, 0.925] LB:33.718 ---Got Best Score
- weights = [0.070, 0.930] LB:33.719
- weights = [0.065, 0.935] LB:33.723
- weights = [0.05, 0.95] LB:33.748