umbertogriffo / Kaggle_Invasive-Species-Monitoring

My solutions that uses Keras (81st place LB 0.99046).

Kaggle-Invasive Species Monitoring

My solutions that uses Keras (81st/513 place LB 0.99046)

Colab

You can try the code directly on Colab. Save a copy in your drive and enjoy It!

Environment

* 2 Intel Xeon E5-2630 v4 2.2GHz, 25M Cache, 8.0 GT/s QPI, Turbo, HT, 10C/20T (85W) Max Mem 2133MHz
* 128 GB Ram
* 1 TB Disk

My approach to achieve 0.99046

Image Sizing: The images in the dataset were size 1154px x 866px. I used a small range of image sizes 128X128, 190X256 and 256X356. If I used a greater size I cloud have a better final score, but my environment doesn't have GPU.

Cross Validation: 8-fold cross-validation. So each fold I ran predictions on the test set and then took the average of 8 folds.

Data Augmentation: Data augmentation is very important to avoid overfitting. I used some random augmentations like horizontal/vertical flips and rotations and I tried other augmentations like shift and zoom.

Ensembling: It is very important to average predictions to help the models generalize and to avoid any bias because each trained model will potentially pick up on different information.

FineTuning: At the end I used finetuning with the same base model Inception-v3 pre-trained on Imagenet.

Code

Version 1: cnn_v1.py - Image size 128X128 - Naive Bagging (8 ANN) - Random rotation, shift and horizontal and vertical flip of images (LB ~0.95)

Version 2: cnn_v2.py - Increased the Image size to 190X256 (LB ~0.96)

Version 3: cnn_v3.py - Increased the Image size to 256X356 - Increased the momentum in SGD (LB ~0.97600)

Version 4: cnn_v4.py - New CNN structure. (LB ~0.97889)

Version 5: cnn_v5.py - Using pre-trained Inception-V3 CNN. Allows to get 0.99046 on LB. (LB ~0.99046)

References

Kaggle Invasive Species Monitoring