Skip to content

Latest commit

 

History

History
82 lines (54 loc) · 3.42 KB

kaggle_to_production.md

File metadata and controls

82 lines (54 loc) · 3.42 KB

Kaggle to Production steps/packaging

1. Data augmentation pipeline

Data augmentation proofed to be the single most important step to boost competition results:

  • Create simple modular Python pipeline that can apply all data augmentation steps automatically and that allows for quick changes of steps and treatments.
  • Apply augmentations that proofed successful during the Kaggle competition
  • Deliverable: simple Python module that can be called like this:
$ augment_data --input_folder {foldername} --output_folder {foldername) --label_file {filename} --new_label_and_weights_file {filename}

2. Package all model dependencies

  • Create an easily reproducable route to create an environment in which a) models can be trained, b) deployed. Options would be 1. provisioning tools like (Ansible, Chef, Puppet, and the like), 2. Docker containers, or 3. prepackaged server images with all the relevant products installed (most likely a combination of 1 and 3), or readily available cloud servers.
  • We could start with Faster-RCNN which provided best during the competition (one model for performance), Keras, Tensorflow?

3. Create and train fish detection model

During the competition explicit fish detection models proofed to be more successful then implicite object classification algorithms as build into some of the applied model products. If we decide to implement a two step recognition process (1. fish, 2. species). Retrain model and deploy that it can be run as:

$ create_fish_images --input_folder --output_folder

Label files should be unaffected in this step.

4. Setup easy workflow for fast retraining of models

  • Create simple command line wrapper for retraining models.
  • Allow for plug and play of different models
  • Goal should be to find the single best performaing model (with potential trade-offs for speed)
  • Create the workflow in a way that it lends itself to automation

5. Create command line interface for speedy model application

  • A command line interface allows for the implementation of the algorithm in a wide range of contexts (such as desktop apps, RESTful APIs, or code in virtually any coding language, etc.):
$ classify --image {imagepath}

['ALB']
  • Would be nice if a binary could be piped to the command, which would it make easier to use in a network context:
$ cat {imagepath} | classify
  • A verbose version could provide additional information, e.g.:
$ classify --image {imagepath} --verbose

[['ALB', .33443], ['YFT', .0001], ...]

where the numbers refer to the likelyhood of the image being in that category.

  • Explore potential steps on the incomming image that could improve outcome, performance (e.g. subtraction of background if applied to a single camera).

6. Create simple web application that provides an online interface to 5

Something like:

curl -X POST codefornature.org/classifyfish --data-binary=/path/to/image
  • Interface should match Camio's requirements.

7. Create enhanced training data

  • Concomitantly work to enhance training data after model is put into production environment
  • Apply lessons learned from Kaggle competition to improve training/test data sets for production model training
  • Balanced species distribution
  • More images of sensitive conservation relevant species e.g. sharks, turtles, etc.
  • Better distribution of data across vessels
  • Third party data
  • Other lessons learned from competition results and forums