bits-bytes-nn / state-farm-detection

State Farm Distracted Driver Detection via Image Classification

State Farm Distracted Driver Detection

Goals

• Experiment with various modern deep learning techniques from the paper, Bag of Tricks for Image Classification with Convolutional Neural Networks through an image classification task.
• Build a high-speed and efficient data pipeline using the TFRecord file format, Dataset API, and the albumentations library.
• Explain the model via Grad-CAM and host it using TF Serving or SageMaker to enable real-time inference.
• (Optional) Effectively manage experiments through the wandb library.

Requirements

• The dataset can be downloaded from this Kaggle competition.
• In addition to the Anaconda libraries, you need to install tensorflow, tensorflow-addons, tensorflow-hub, albumentations and wandb.

Experimental Setup

• EfficientNet-B0 was used as the base model. On top of that, fully connected layers and dropout layers were added.
• The batch size was set to 32, the number of epochs was set to 500, and an early stopping option was applied.
• A Cosine Decay Schedule with Restarts was used. In this case, the initial learning rate was set to 0.001 and the first decay step was set to 1,000.
• The images were resized to 224 x 224 and image data augmentation through rotation, scaling, and shifting was applied. Below are examples of data augmentation.

Experiment Result

• The evaluation criterion for this Kaggle competition is multi-class logarithmic loss.
• As the validation set, 25% of the images were randomly assigned. However, in the case of the 5-fold CV ensemble, the dataset was divided into 5 equal parts.

Treatment	Public Score	Private Score
RAdam	1.0260	0.6792
AdamW	0.9117	0.7140
RAdam + SWA	1.1547	0.7527
RAdam + Mixup	0.8331	0.6423
RAdam + Label Smoothing (0.1)	0.9047	0.7891
RAdam + Mixup + TTA (3 times)	0.7434	0.5777
RAdam + Mixup + TTA + 5-fold CV Ensemble (OOF)	0.6877	0.5419
Pseudo Labeling (> 0.9) + RAdam + Mixup + TTA + 5-fold CV Ensemble	0.6504	0.5169

Model Explainability with Grad-CAM

• 20 samples were randomly selected from the test set and visualized using the Grad-CAM technique. Labels shown are predicted.

Model Serving

TF Serving

• You need to download and run the Docker image via scripts/run.sh file. Then, you can test model inference through a locally hosted TF Serving.

SageMaker

• SageMaker allows you to train TensorFlow models and deploy endpoints for serving. You can also use the SageMaker Estimator's Pipe mode to train a model without downloading a dataset directly.