State farm distracted driver detection

Overview

Implementation of Kaggle driver detection

CNN Models implemented

• simple model
• vgg16 - input shape 224×224
• vgg19 - input shape 224×224
• inception v3 - input shape 299×299
• xception - input shape 299×299

Train and evaluate

Downloading and splitting data

Using kaggle API:

./src/colab_config.sh

Train

python train.py [...options]

Example of train:

python train.py --model vgg16 --width 224 --optimizer adam --lr 1e-5 --batch 16 --epochs 50

Options

• --model <model>
  type of model used. <model> one of the following: simple, vgg16, vgg19, inception_v3, xception. Default us simple.
• --epochs <number>
<number> - number of training epochs. Default is 20.
• --width <width>
  Width of image, that will be input of model. Images from dataset are resized to this <width>. Default is 150.
• --optimizer <optimizer>
  Optimizer, used in train process. <optimizer> one of the following: adam, sgd, rmsprop
• --summary [True|False]
  Print summary of model. Default is False
• --lr <learning_rate>
  Set learning rate for <optimizer>. Default value is 5e-5
• --weight_path <path>
  Path to weights. If specified, weight from file is loaded, if not - weights initialized randomly.
• --fc <count_of_layers>
  Count of fully connected layers, used in fine-tuning. Default is 2.
• --fc_dim <dimension>
  Dimension of fully connected layers. Default is 4096.
• --dropout <dropout>
  Dropout after each fully-connected layer. If < 0, than no dropout layers added. Default is 0.5.
• --batch <batch>
  Batch size in train process. Default is 32.

Evaluation

Files predict.sh and predict-all.sh.

predict.sh

Second param: file name to saved weights in /content/drive/models folder:

./predict.sh --path_to_model vgg-04-acc-0.9722-loss-0.0997.hdf5 --model vgg16 --width 224

predict-all.sh

For all saved models in folder /content/drive/models make prediction and save to /content/drive/submissions:

./predict-all.sh --model vgg16 --width 224

Environment

Training was running on Colaboratory Google Platform on GPU environment.