Triplet loss for facial recognition.

First attempt using Google Colab

It was quite hard and messy to make Python code work using git clone etc. In fact I had many problems

• in passing args to executed python scripts
• in unzipping or downloading the dataset
• in coding by the way, i hate jupyter notebook
• in managing to code in all that mess

So i'll try to do it an other and more simple way

Source of the new code (inceptionV3 and resnet!)

pytorch.org/tutorials

My attempt to have better accuracy will be based on this tutorial code, and will be executed locally

This code uses the default ImageFolder dataset class:

ImageFolderDocumentation

Other quite simple tutorial: tuto

In finetuning, we start with a pretrained model and update all of the model’s parameters for our new task, in essence retraining the whole model. In feature extraction, we start with a pretrained model and only update the final layer weights from which we derive predictions.

--> So we'll use feature extraction because we don't want to train model

to install requirements:

alias py3=/usr/bin/python3

py3 -m pip install -r requirements.txt

Description of script files:

• resnet18/train.py: Ylies script: resnet18 with triplet loss
• multi/train.py: original script from tutorial site (can choose model) but using criteria nn.CrossEntropyLoss()
• multi/train2.py: script adapted to have model class still using criteria nn.CrossEntropyLoss()

Trainings using hymenoptera_data datataset

## First test using resnet model

Training complete in 1m 56s

Best val Acc: 0.941176

## Second test using InceptionV3 model

We can see that the accuracy is slightly better for inception at the beginning of training But for all the rest, it's quite the same

Training complete in 6m 59s

Best val Acc: 0.928105

## Third test using squeezenet model

This one seems to be very stable, but has quite the same accuracy as the two others I would put a bit more trust in this one personally

Training complete in 1m 54s

Best val Acc: 0.928105

Fourth test comparison between Resnet18 LR 0.001 Momentum 0.9 and Resnet18 LR 0.0001 Momentum 0.95

Both are already trained and use feature extraction

Seem to be quite the sames

Fifth test comparison between Resnet18 and densenet

As i have a complete code ready to put into action, just for test, i'll compare resnet to densenet

Trainings using triplet face dataset:

The number of classes in our dataset: ls dataset/test/ -A | wc -l It returned 17

To fit to the code and gain some time, i just renamed the folders train and test to train and val

## First test using resnet model

I suppose that the accuracy is so good because we have mny images for each class (person)

This result isn't a good one to interpret, because transformations on images were way too easy for the algorithm to learn. So new tests have been done using resnet in google colab

Training complete in 37m 21s

Best val Acc: 0.999383

# Going back on Google Colab

Yes, I know, I don't have Nvidia GPU on my computer, so it's quite hard to execute triplet loss with so many images hehe.

I'll do the same type of tests as on my computer but trying to concentrate on our dataset this time.

So i made a comparison between between resnet and squeezenet (20 epochs):

• Best val Acc: 0.999383 for squeezenet
• Best val Acc: 0.988279 for resnet18

So, i'll try to focus on squeezenet

Triplet Face

The repository contains code for the application of triplet loss training to the task of facial recognition. This code has been produced for a lecture and is not going to be maintained in any sort.

Architecture

The proposed architecture is pretty simple and does not implement state of the art performances. The chosen architecture is a fine tuning example of the resnet18 CNN model. The model includes the freezed CNN part of resnet, and its FC part has been replaced to be trained to output latent variables for the facial image input.

The dataset needs to be formatted in the following form:

dataset/
| test/
| | 0/
| | | 00563.png
| | | 01567.png
| | | ...
| | 1/
| | | 00011.png
| | | 00153.png
| | | ...s
| | ...
| train/
| | 0/
| | | 00001.png
| | | 00002.png
| | | ...
| | 1/
| | | 00001.png
| | | 00002.png
| | | ...
| | ...
| labels.csv        # id;label

Install

Install all dependencies ( pip command may need sudo ):

cd TripletFace/
pip3 install -r requirements.txt

Usage

For training:

usage: train.py [-h] -s DATASET_PATH -m MODEL_PATH [-i INPUT_SIZE]
                [-z LATENT_SIZE] [-b BATCH_SIZE] [-e EPOCHS]
                [-l LEARNING_RATE] [-w N_WORKERS] [-r N_SAMPLES]

optional arguments:
  -h, --help            show this help message and exit
  -s DATASET_PATH, --dataset_path DATASET_PATH
  -m MODEL_PATH, --model_path MODEL_PATH
  -i INPUT_SIZE, --input_size INPUT_SIZE
  -z LATENT_SIZE, --latent_size LATENT_SIZE
  -b BATCH_SIZE, --batch_size BATCH_SIZE
  -e EPOCHS, --epochs EPOCHS
  -l LEARNING_RATE, --learning_rate LEARNING_RATE
  -w N_WORKERS, --n_workers N_WORKERS
  -r N_SAMPLES, --n_samples N_SAMPLES

References

• Resnet Paper: Arxiv
• Triplet Loss Paper: Arxiv
• TripletTorch Helper Module: Github

Todo ( For the students )

Deadline Decembre 14th 2019 at 12am

The students are asked to complete the following tasks:

• Fork the Project
• Improve the model by playing with Hyperparameters and by changing the Architecture ( may not use resnet )
• JIT compile the model ( see Documentation )
• Add script to generate Centroids and Thesholds using few face images from one person
• Generate those for each of the student included in the dataset
• Add inference script in order to use the final model
• Change README.md in order to include the student choices explained and a table containing the Centroids and Thesholds for each student of the dataset with a vizualisation ( See the one above )
• Send the github link by mail