My 121st place solution to the Lyft Motion Prediction for Autonomous vehicles competition hosted on Kaggle by Lyft.
I joined rather late to this competition and then had to battle with memory leaks and errors but even still I feel like I did pretty well. This competition was interesting because you needed a long compute time which is not possible with online cloud environments like Colab. So I had to use my personal GPU and train for hours on end.
I would also like to shoutout @louis925 whose discussion posts about bug fixes[1][2] made it possible for me to compete!
My final solution was basically just the baseline model trained for 3.5 days. The final single model was simply a ResNet18 with a linear head dropout rate of 0.5. I trained it on 5% of the total training data randomly sampled.
Since training took so long and because I started experimenting late I couldn't try very many models. I only tried ResNet18 and ResNet34 variations of the baseline model. The final change that got my leaderboard score was adding a dropout layer in the linear head.
I didn't do anything fancy with the dataset. I just used the vanilla l5kit AgentDatasets and validated with the full validation.zarr chopped.
Since there was such a large amount of data and I didn't dive deep into the data, so I had limited knowledge of what the data was, I didn't use any augmentation.
I used a simple pipeline for training:
- Adam optimizer with One cycle LR schedule
- Trained for 15 epochs with a batch size of 64
- Used the given neg multi log loss likelihood loss as the loss function
- Trained for 5% of the length of full_train.zarr each epoch and randomly sampled the entire train AgentDataset for every item.
Training the final model took 85.25 hours or 3.5 days on my local RTX2080 gpu.
Did not ensemble or blend at all, just submitted one model once.
Since everyone said the validation scheme matched so well with the leaderboard I only submitted my best model validation score wise.
Validation: 20.39
Public LB: 19.89
Private LB: 20.54Nothing super important from this competition, just had fun and learned a lot as usual! I am content with my final placement as I just wanted to beat the baseline kernals, I was surprised I even got close to a bronze.
Previous Competition: OSIC Pulmonary Fibrosis Progression
Next Competition: [placeholder]