Application Challenges MachineLearning TUM

Johannes Keim, Johannes H.,Dushyant D., Katarina G.

Setup

• Create conda environment and install dependencies from environment.yml file: conda env create -f environment.yml
• For the web-app, navigate to cd web-app and install dependencies with npm install

Adding dependencies

• Check if the new dependency is supported on the IBM Power platform https://docs.anaconda.com/anaconda/packages/py3.8_linux-ppc64le/ https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/#/
• Install, preferably via conda
• Update environment.yml by running: conda env export > environment.yml

Training

To start a training run from the project root: python -m training.train

Note: Please commit any changes before starting, so that W&B can associate the correct repository state with the run.

Evaluation

To calculate evaluation scores run from the project root: python -m classifier.faiss_evaluate.

The evaluation depends on the dataset, a pre-computed FAISS index and a model checkpoint. Please check in the script if paths are set accordingly.

Prerequisites

Evaluation depends on a pre-calculated FAISS index, which in turn requires a set of pre-computed embeddings for it's creation.

• To calculate embeddings from the dataset run python -m classifier.pre_compute_embeddings
• To pre-compute the index see python classifier/faiss_create.py

Application

Starting the API Backend

From project root run: python -m api.main

Starting the Web-App

Navigate to web-app folder: cd web-app

With node and npm installed, start locally: npm run dev

For more information have a look at the readme.

Folder Structure

• api: FastAPI backend service
• checkpoints: Put model state checkpoints from W&B here
• classifier: FAISS and L2 classifiers and scripts to pre-calculate embeddings, to create FAISS index, and for evaluation
• data: Batch sampler, two face alignment variants, dataset, dataloader and script to create pre-aligned dataset.
• datasets: The dataset folders, and pre-computed embeddings and FAISS indices go here.
• faiss_pcc: FAISS python library compiled for IBM PPC,
• models: PyTorch implementation of inception resnet V1
• training: training script, depending on triplet generation and triplet loss function.
• utils: Utils for visualization and embedding handling
• web-app: Svelte app as interface for face recognition system and embeddings visualization