A CapsNet network implementing a facial recognition for a LFW and similar data sets.
The provided capsnet package can be easily installed via PIP:
pip install git+https://github.com/tumido/capsnet-faceThe package provides binding to a Kaggle data set, if you desire use it as your data source please install kaggle package as well and setup accordingly.
If you want to create a developer setup for this package, you can either use Conda or Pipenv to do so:
$ git clone https://github.com/tumido/capsnet-face
Cloning into 'capsnet-face'...
remote: Enumerating objects: 95, done.
remote: Counting objects: 100% (95/95), done.
remote: Compressing objects: 100% (67/67), done.
remote: Total 330 (delta 52), reused 69 (delta 28), pack-reused 235
Receiving objects: 100% (330/330), 2.02 MiB | 4.12 MiB/s, done.
Resolving deltas: 100% (200/200), done.
$ cd capsnet-faceThen sync via Pipenv:
$ pipenv sync --dev
Creating a virtualenv for this project…
...
⠏ Creating virtual environment...
...
Installing dependencies from Pipfile.lock (2dc3d7)…
🐍 ▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉▉ 96/96 — 00:01:15
To activate this project's virtualenv, run pipenv shell.
Alternatively, run a command inside the virtualenv with pipenv run.
All dependencies are now up-to-date!
$ pipenv shell
Launching subshell in virtual environment…
...Or use Anaconda/Conda:
$ conda env create -f environment.yml
Collecting package metadata: done
Solving environment: done
Preparing transaction: done
...
$ conda activate keras_cpu
...Python 3.6 supported.
The package provides a CLI interface:
$ capsnet --help
Using TensorFlow backend.
Usage: cli.py [OPTIONS] COMMAND [ARGS]...
CapsNet for facial recognition. CLI interface.
Options:
--help Show this message and exit.
Commands:
predict Predict a label for input image
train Train your model.To invoke training please do:
$ capsnet train -d lfw
Using TensorFlow backend.
Loading data set: LFW
Creating CapsNet network...
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_image (InputLayer) (None, 32, 32, 3) 0
__________________________________________________________________________________________________
encoder_conv2d (Conv2D) (None, 24, 24, 256) 62464 input_image[0][0]
__________________________________________________________________________________________________
encoder_dropout (Dropout) (None, 24, 24, 256) 0 encoder_conv2d[0][0]
__________________________________________________________________________________________________
encoder_feature_caps_conv2d (Co (None, 10, 10, 256) 1638656 encoder_dropout[0][0]
__________________________________________________________________________________________________
encoder_feature_caps_reshape (R (None, 1600, 16) 0 encoder_feature_caps_conv2d[0][0]
__________________________________________________________________________________________________
encoder_feature_caps_squash (La (None, 1600, 16) 0 encoder_feature_caps_reshape[0][0
__________________________________________________________________________________________________
encoder_pred_caps (PredictionCa (None, 42, 32) 34406400 encoder_feature_caps_squash[0][0]
__________________________________________________________________________________________________
input_label (InputLayer) (None, 42) 0
__________________________________________________________________________________________________
mask (Mask) (None, 1344) 0 encoder_pred_caps[0][0]
input_label[0][0]
__________________________________________________________________________________________________
capsnet (Lambda) (None, 42) 0 encoder_pred_caps[0][0]
__________________________________________________________________________________________________
decoder (Sequential) (None, 32, 32, 3) 540479 mask[0][0]
==================================================================================================
Total params: 36,647,999
Trainable params: 36,647,999
Non-trainable params: 0
__________________________________________________________________________________________________
Starting training...
Epoch 1/200
...This launches the standard training routine over LFW dataset. Kaggle PINS dataset is also available
Predictions can be run directly using a saved model:
$ predict -m saved_models/2019-05-14_11-caps_75-acc.tar.gz Serena_Williams_0002.jpg
Loading CapsNet...
Loading model from saved_models/2019-05-14_11-caps_75-acc.tar.gz...
Loading "train" architecture... Done
Loading "test" architecture... Done
Loading weights... Done
Extracting labels... Done
Loading image...
Predicting...
Guessed likelihood per label:
Label Probability
Ariel Sharon 2.83%
Colin Powell 8.25%
Donald Rumsfeld 3.25%
George W Bush 0.79%
Gerhard Schroeder 0.25%
Hugo Chavez 0.08%
Jacques Chirac 0.08%
Jean Chretien 0.11%
John Ashcroft 0.65%
Junichiro Koizumi 0.16%
Serena Williams 91.07% <-- Best match
Tony Blair 0.23%When you desire to experiment with this implementation, please feel free to do so by importing the CapsNet class from the installed package. We encourage to explore help(CapsNet) to list all methods and arguments possible.
This package provides multiple objects:
CapsNet, the network classpreprocessfunction for input data preprocessingfetch_pins_peoplea Kaggle PINS data set collector with API consistent to thefetch_lfw_people
There is a notebooks folder available in this repository. If you would like to know more about the process how the models were trained or evaluated, please feel free to explore these.
| Identities | Data set (images) | Routing iterations | Acc. Train | Acc. Validation | Acc. Test | Loss |
|---|---|---|---|---|---|---|
| 42 | LFW (2588) | 1 | 46.2% | 42.5% | 42.5% | 0.5002 |
| 42 | LFW (2588) | 3 | 56.4% | 53.7% | 42.5% | 0.3915 |
| 11 | LFW (1560) | 1 | 52.6% | 63.2% | 61.5% | 0.2952 |
| 11 | LFW (1560) | 3 | 69.3% | 75.0% | 73.7% | 0.2013 |
The trained models are available at Google Drive.
