Toxicless texts with AI – how to measure text toxicity in the browser


In this article I will show how to measure comments toxicity using Machine Learning models.

Before you will continue reading please watch short introduction:

Hate, rude and toxic comments are common problem in the internet which affects many people. Today, we will prepare neural network, which detects comments toxicity, directly in the browser. The goal is to create solution which will detect toxicity in the realtime and warn the user during writing, which can discourage from writing toxic comments.

To do this, we will train the tensorflow lite model, which will run in the browser using WebAssembly backend. The WebAssembly (WASM) allows running C, C++ or RUST code at native speed. Thanks to this, prediction performance will be better than running it using javascript tensorflowjs version.
Moreover, we can serve the model, on the static page, with no additional backend servers required.

web assembly

To train the model, we will use the Kaggle Toxic Comment Classification Challenge training data,
which contains the labeled comments, with toxicity types:
* toxic
* severe_toxic
* obscene
* threat
* insult
* identity_hate

data set

Our model, will only classify, if the text is toxic, or not. Thus we need to start with preprocessing training data. Then we will use the tensorflow lite model maker library.
We will also use the Averaging Word Embedding specification which will create words embeddings and dictionary mappings using training data thus we can train the model in the different languages.
The Averaging Word Embedding specification based model will be small <1MB.
If we have small dataset we can use the pretrained embeddings. We can choose MobileBERT or BERT-Base specification.
In this case models will much more bigger 25MB w/ quantization 100MB w/o quantization for
MobileBERT and 300MB for BERT-Base (based on:


Using simple model architecture (Averaging Word Embedding), we can achieve about nighty five percent accuracy, and small model size, appropriate 
for the web browser, and web assembly. 

tensorflow lite

Now, let’s prepare the non-toxic forum web application, where we can write the comments.
When we write non-toxic comments, the model won’t block it.
On the other hand, the toxic comments will be blocked, and the user warned.

Of course, this is only client side validation, which can discourage users, from writing toxic comments.

web application

To run the example simply clone git repository and run simple server to serve the static page:

git clone
cd ai-toxicless-texts
python3 -m http.server

The code to for preparing data, training and exporting model is here:

How to extract music sources: bass, drums, vocals and other ? – music separation with AI


In this article I will show how we can extract music sources: bass, drums, vocals and other accompaniments using neural networks.

Before you will continue reading please watch short introduction:

Separation of individual instruments from arranged music is another area where machine learning
algorithms could help. Demucs solves this problem using neural networks.

The trained model ( use U-NET architecture which contains two parts encoder and decoder.
On the encoder input we put the original track and after processing we get bass, drums, vocals and other accompaniments at the decoder output.
The encoder, is connected to the decoder, through additional LSTM layer,
as well as residual connections between subsequent layers.

neural network architecture

Ok, we have neural network architecture but what about the training data ?
This is another difficulty which can be handled by the unlabeled data remixing pipeline.
We start with another classifier, which can find the parts of music, which do not contain the specific instruments, for example drums.
Then, we mix it with well known drums signal, and separate the tracks using the model. 

Now we can compare, the separation results, with known drums track and mixture of other instruments. 

According to this, we can calculate the loss (L1 loss), and use it during the training. 

Additionally, we set different loss weights, for known track and the other. 

training data

The whole UI is kept in the docker image thus you can simply try it:

#for CPU
docker run --name aiaudioseparation -it -p 8000:8000 -v $(pwd)/checkpoints:/root/.cache/torch/hub/checkpoints --rm qooba/aimusicseparation

#for GPU
docker run --name aiaudioseparation --gpus all -it -p 8000:8000 -v $(pwd)/checkpoints:/root/.cache/torch/hub/checkpoints --rm qooba/aimusicseparation

web UI

Bored with classical computers? – Quantum AI with OpenFermion


In this article I will show how we can prepare and perform calculations on quantum computers
using OpenFermion, Cirq and PySCF.

Before you will continue reading please watch short introduction:

Currently, there are many supercomputing centers, where we can run complicated simulations.
However, there are still problems that are beyond the capabilities of classical computers,
which can be addressed by quantum computers.

materials science

Quantum chemistry and materials science problems which that are described by the laws of
quantum mechanics can be mapped to the quantum computers and projected to qubits.

OpenFermion is the library which can help to perform such calculations on a quantum computer.

Additionally we will use the PySCF package which will help to perform initial structure optimization (if you are interested in PySCF package I have shared the example DFT based band structure calculation of the single layer graphene structure pyscf_graphene.ipynb).

materials science

In our example we will investigate H_2 molecule for simplicity. We will use the PySCF package to find optimal bond length of the molecule.

Thanks to the OpenFermion-PySCF plugin we can smoothly use the molecule initial state obtained from PySCF package run in OpenFermion library (openfermionpyscf_h2.ipynb).

from openfermion.chem import MolecularData
from openfermionpyscf import run_pyscf

geometry = create_molecule(bond_length)
basis = 'sto-3g'
multiplicity = 1

run_scf = 1
run_mp2 = 1
run_cisd = 0
run_ccsd = 0
run_fci = 1

molecule = MolecularData(geometry, basis, multiplicity)

# Run pyscf.
molecule = run_pyscf(molecule,

materials science

Now it is time to compile the molecule to the representation readable by the quantum computer
using OpenFermion and Cirq library.
Currently you can use several methods to achieve this:
* Constructing a Basis Change Circuits
* Constructing a Diagonal Coulomb Trotter Step
* Constructing Trotter Steps With Low-Rank Decomposition

Using one of this methods we get optimized quantum circuit.
In our case the quantum cirquit for H_2 system will be represented by 4 qubits and operations that act on them (moment is collection of operations that act at the same abstract time slice).

materials science

Finally we can use quantum circuit to run the calculations on the cirq simulator or on the real quantum computer.

TinyMLOps with Arduino


In this article I will show how to build MLOps process TinyML on Arduino Nano 33 BLE Sense.

Before you will continue reading please watch short introduction:

In the last article (TinyML with Arduino) I have shown the example TinyML model which will classify
jelly bears using RGB sensor.
The next step, will be to build a process that will simplify, the model versions management, and the deployment.


The MLflow project is prepared in the Jupyter Notebook. Then we can convert the Notebook to the python code using nbdev library and version it in the Git repository.

Now we are ready to run the MLflow project using command:

mlflow run https://git_repository.git#path --no-conda --experiment-name="arduino"

The model is saved in the MLflow registry and the model version is associated with
the git commit version.

mlops git

The MLflow model contains additional artifacts:
* artifacts.ino – the arduino code which loads and uses the model
* model.h – the Tensorflow Lite model encoded to hex
* reduirements.ino.txt – the list of Arduino dependencies required by the arduino code

Example requirements.ino.txt file:


mlops arduino

Finally we can run the command:

docker run -it --network app_default --device=/dev/ttyACM0:/dev/ttyACM0 -e AWS_ACCESS_KEY_ID=minio -e AWS_SECRET_ACCESS_KEY=minio123 -e MLFLOW_S3_ENDPOINT_URL=http://minio:9000 -e MLFLOW_TRACKING_URI=http://mlflow:5000 qooba/tinyml-arduino:mlops ./ -r ${RUN_ID}

* –device=/dev/ttyACM0 – is arduino device connected using USB
* AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY – are minio credentials
* MLFLOW_S3_ENDPOINT_URL – is minio url
* MLFLOW_TRACKING_URI – is mlflow url
* ${RUN_ID} – is run id of model saved in MLflow registry

Additionally we have several command line options:


Syntax: docker run -it qooba/tinyml-arduino:mlops -h [-r MLFLOW_RUN_ID] [-s ARDUINO_SERIAL] [-c ARDUINO_CORE] [-m ARDUINO_MODEL]
-h|--help     Print help
-r|--run      MLflow run id
-s|--serial   Arduino device serial (default: /dev/ttyACM0)
-c|--core     Arduino core (default: arduino:mbed_nano)
-m|--model    Arduino model (default: arduino:mbed_nano:nano33ble)

arduino docker

After running the code the docker image qooba/tinyml-arduino:mlops
will fetch the model for indicated RUN_ID from MLFlow.
Then it will install required dependencies using the file requirements.ino.txt.

It will compile the model and the Arduino code.
And finally upload it to the device.

Thanks to this, we can more easily manage subsequent versions of models, and automate the deployment process.

TinyML with Arduino


In this article I will show how to build Tensorflow Lite based jelly bears classifier using Arduino Nano 33 BLE Sense.

Before you will continue reading please watch short introduction:

Currently a machine learning solution can be deployed not only on very powerful machines containing GPU cards but also on a really small devices. Of course such a devices has a some limitation eg. memory etc. To deploy ML model we need to prepare it. The Tensorflow framework allows you to convert neural networks to Tensorflow Lite which can be installed on the edge devices eg. Arduino Nano.

Arduino Nano 33 BLE Sense is equipped with many sensors that allow for the implementation of many projects eg.:
* Digital microphone
* Digital proximity, ambient light, RGB and gesture sensor
* 3D magnetometer, 3D accelerometer, 3D gyroscope
* Capacitive digital sensor for relative humidity and temperature

Examples which I have used in this project can be found here.

Arduino sensors

To simplify device usage I have build Arduino Lab project where you can test and investigate listed sensors directly on the web browser.

The project dependencies are packed into docker image to simplify usage.

Before you start the project you will need to connect Arduino through USB (the Arduino will communicate with docker container through /dev/ttyACM0)

git clone
cd tinyml-arduino
# in another terminal tab
# go inside server container 
docker exec -it arduino /bin/bash

For each sensor type you can click Prepare button which will build and deploy appropriate Arduino code.

Sometimes you will have to deploy to arduino manually to do this you will need to
go to arduino container

docker exec -it arduino /bin/bash
cd /arduino
make rgb

Here you have complete Makefile with all types of implemented sensors.

You can start observations using Watch button.
Arduino pdm
Arduino temperature
Arduino rgb

Now we will build TinyML solution.
In the first step we will capture training data:
Arduino capture

The training data will be saved in the csv format. You will need to repeat the proces for each class you will detect.

Captured data will be uploaded to the Colab Notebook.
Here I fully base on the project Fruit identification using Arduino and TensorFlow.
In the notebook we train the model using Tensorflow then convert it to Tensorflow Lite and finally encode to hex format (model.h header file) which is readable by Arduino.

Now we compile and upload model.h header file using drag and drop mechanism.

Arduino upload

Finally we can classify the jelly bears by the color:

Arduino classify

Feast with AI – feed your MLflow models with feature store


In this article I will show how to prepare complete MLOPS solution based on the Feast feature store and MLflow platform.

Before you will continue reading please watch short introduction:

The whole solution will be deployed on the kubernetes (mlflow_feast.yaml).


We will use:
* Feast – as a Feature Store
* MLflow – as model repository
* Minio – as a S3 storage
* Jupyter notebook – as a workspace
* Redis – for a online features store

propensity to buy

To better visualize the whole process we will use the Propensity to buy example where I base on the Kaggle examples and data.


We start in Jupyter Notebook where we prepare Feast feature store schema which is kept in S3.

We can simply inspect the Feast schema in Jupyter Notebook:

from feast import FeatureStore
from IPython.core.display import display, HTML
import json
from json2html import *
import warnings

class FeastSchema:
    def __init__(self, repo_path: str): = FeatureStore(repo_path=repo_path)

    def show_schema(self, skip_meta: bool= False):
        display(HTML(json2html.convert(json = feast_schema)))

    def show_table_schema(self, table: str, skip_meta: bool= False):
        display(HTML(json2html.convert(json = {table:feasture_tables_dictionary[table]})))

    def __project_show_schema(self, skip_meta: bool= False):
        for entity in feast_entities:

        for feature_table in feast_feature_tables:
            if 'entities' in feature_table_spec:
                for entity in feature_table_spec['entities']:

            if not skip_meta:


        return feasture_tables_dictionary


In our case we store the data in Apache Parquet files in S3 bucket.
Using the Feast we can fetch the historical features and train the model using Scikit-learn library


store = FeatureStore(repo_path=".")

s3 = fs.S3FileSystem(endpoint_override=os.environ.get("FEAST_S3_ENDPOINT_URL"))
entity_df=pd.read_parquet(f'{bucket_name}/{filename}_entities.parquet', filesystem=s3)

training_df = store.get_historical_features(
    feature_refs = [

predictors = training_df.drop(['propensity_data__ordered','UserID','event_timestamp'], axis=1)
targets = training_df['propensity_data__ordered']

X_train, X_test, y_train, y_test = train_test_split(predictors, targets, test_size=.3)



ac_score=sklearn.metrics.accuracy_score(y_test, predictions)

propensity_model_path = 'propensity.joblib'
joblib.dump(classifier, propensity_model_path)

artifacts = {
    "propensity_model": propensity_model_path,
    "feature_store": "feature_store.yaml"

The model will use online Feast redis features as well as additional features from the request thus we need to wrap the MLflow model and define it:

import mlflow.pyfunc
class PropensityWrapper(mlflow.pyfunc.PythonModel):

    def load_context(self, context):
        import joblib
        from feast import FeatureStore
        import pandas as pd 
        import os

        self.model = joblib.load(context.artifacts["propensity_model"]) = FeatureStore(repo_path=os.environ.get("FEAST_REPO_PATH"))

    def predict(self, context, model_input):

        feature_vector =
            entity_rows=[{"UserID": uid} for uid in users]

        merged_data = pd.merge(model_input,data, how="inner", on=["UserID"], suffixes=('_x', '')).drop(['UserID'], axis=1)
        return self.model.predict(merged_data)

Now we can log the MLflow model to the repository:

import warnings
import sys

import pandas as pd
import numpy as np
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
from sklearn.model_selection import train_test_split
from sklearn.linear_model import ElasticNet
from urllib.parse import urlparse
import mlflow
import mlflow.sklearn
import mlflow.pyfunc


with mlflow.start_run():

    #mlflow.log_param("var_smoothing", input_params['var_smoothing'])
    mlflow.log_metric("accuracy_score", ac_score)

    tracking_url_type_store = urlparse(mlflow.get_tracking_uri()).scheme

    if tracking_url_type_store != "file":

We can export the code and run is using MLflow cli:

mlflow run . --no-conda --experiment-name="propensity" -P var_smoothing=1e-9

Now we need to materialize features to Redis:

feast materialize 2021-03-22T23:42:00 2021-06-22T23:42:00

Using MLflow we can simply deploy model as a microservice in k8s.
In our case we want to deploy the model models:/propensity_model/Production
which is currently assigned for Production. During start the MLflow will automatically fetch the proper model from S3:

apiVersion: apps/v1
kind: Deployment
  name: mlflow-serving
  namespace: qooba
  replicas: 1
      app: mlflow-serving
      version: v1
        app: mlflow-serving
        version: v1
      - image: qooba/mlflow:serving
        imagePullPolicy: IfNotPresent
        name: mlflow-serving
        - name: MLFLOW_TRACKING_URI
          value: http://mlflow.qooba.svc.cluster.local:5000
        - name: AWS_ACCESS_KEY_ID
              name: minio-auth
              key: username
        - name: AWS_SECRET_ACCESS_KEY
              name: minio-auth
              key: password
        - name: MLFLOW_S3_ENDPOINT_URL
          value: http://minio.qooba.svc.cluster.local:9000
        - name: FEAST_S3_ENDPOINT_URL
          value: http://minio.qooba.svc.cluster.local:9000
        - name: REDIS_TYPE
          value: REDIS
          value: redis.qooba.svc.cluster.local:6379,db=0
        - name: FEAST_TELEMETRY
          value: "false"
        - name: FEAST_REPO_PATH
          value: /feast_repository
        - name: PORT
          value: "5000"
        - name: MODEL
          value: models:/propensity_model/Production
        - containerPort: 5000
          - mountPath: /feast_repository
            name: config
        - name: config
            name: mlflow-serving
            - key: feature_store
              path: feature_store.yaml

On each HTTP request:

import requests
import json


    'Content-Type': 'application/json; format=pandas-records'

    {"UserID": "a720-6b732349-a720-4862-bd21-644732",
     'propensity_data:device_mobile': 1.0,
     'propensity_data:device_computer': 0.0,
     'propensity_data:device_tablet': 0.0
], data=json.dumps(data), headers=headers)

The model will fetch the client features (based on UserID) from Redis and HTTP request and generate prediction.

Flink with AI – how to use Flink with MLflow model in Jupyter Notebook


In this article I will show how to process streams with Apache Flink and MLflow model

Before you will continue reading please watch short introduction:

Apache Flink allows for an efficient and scalable way of processing streams. It is a distributed processing engine which supports multiple sources like: Kafka, NiFi and many others
(if we need custom, we can create them ourselves).

Apache Flink also provides the framework for defining streams operations in languages like:
Java, Scala, Python and SQL.

To simplify the such definitions we can use Jupyter Notebook as a interface. Of course we can write in Python using PyFlink library but we can make it even easier using writing jupyter notebook extension (“magic words”).

Using Flink extension (magic.ipynb) we can simply use Flink SQL sql syntax directly in Jupyter Notebook.

To use the extesnions we need to load it:

%reload_ext flinkmagic

Then we need to initialize the Flink StreamEnvironment:


Now we can use the SQL code for example:

FileSystem connector:

    word varchar,
    cnt bigint) WITH (
        'connector.type' = 'filesystem',
        'format.type' = 'csv',
        'connector.path' = '/opt/flink/notebooks/data/word_count_output1')

MySQL connector:

    smstext varchar,
    smstype varchar) WITH (
        'connector.type' = 'jdbc',
        'connector.url' = 'jdbc:mysql://mysql:3306/test',
        'connector.table' = 'sms',
        'connector.driver' = 'com.mysql.jdbc.Driver',
        'connector.write.flush.interval' = '10',
        'connector.username' = 'root',
        'connector.password' = 'my-secret-pw')

Kafka connector:

CREATE TABLE MySourceKafkaTable (word varchar) WITH (
    'connector.type' = 'kafka',
    'connector.version' = 'universal',
    'connector.topic' = 'test',
    'connector.startup-mode' = 'latest-offset',
    '' = 'kafka:9092',
    '' = 'test',
    'format.type' = 'csv'

The magic keyword will automatically execute SQL in existing StreamingEnvironment.

Now we can apply the Machine Learning model. In plain Flink we can use UDF function defined
in python but we will use MLflow model which wraps the ML frameworks (like PyTorch, Tensorflow, Scikit-learn etc.). Because MLflow expose homogeneous interface we can
create another “jupyter magic” which will automatically load MLflow model as a Flink function.

%flink_mlflow "SPAM_CLASSIFIER" "/mlflow/mlruns/2/64a89b0a6b7346498316bfae4c298535/artifacts/model" "[DataTypes.STRING()]" "DataTypes.STRING()"

Now we can simply write Flink SQL query:

SELECT word as smstext, SPAM_CLASSIFIER(word) as smstype FROM MySourceKafkaTable

which in our case will fetch kafka events and classify it using MLflow spam classifier. The
results will be displayed in the realtime in the Jupyter Notebook as a events DataFrame.

If we want we can simply use other python libraries (like matplotlib and others) to create
graphical representation of the results eg. pie chart.

You can find the whole code including: Flink examples, extension and Dockerfiles here:

You can also use docker image: qooba/flink:dev to test and run notebooks inside.
Please check the
where you have all components (Kafka, MySQL, Jupyter with Flink, MLflow repository).

Animated Art with AI – face reeanactment in action


In this article I will show how to use artificial intelligence to add motion to the images and photos.

Before you will continue reading please watch short introduction:

Face reenactment

To bring photos to life we can use the face reenactment algorithm designed to transfer the facial movements in the video to another image.

face reenactment diagram

In this project I have used github implementation: Where the extensive description of the neural network architecture can be found in this paper. The solution contains of two parts: motion module and generation module.
The motion module at the first stage extracts the key points from the source and target image. In fact in the solution we assume that reference image which we can to the source and target image exists and at the first stage the transformations from reference image to source (T_{S \leftarrow R} (p_k)) and target (T_{T \leftarrow R} (p_k)) image is calculated respectively. Then the first order Taylor expansions \frac{d}{dp}T_{S \leftarrow R} (p)| {p=p_k} and \frac{d}{dp}T_{T \leftarrow R} (p)| {p=p_k} is used to calculate dense motion field.
The generation module use calculated dense motion field and source image to generate new image that will resemble target image.

face reenactment diagram

The whole solution is packed into docker image thus we can simply reproduce the results using command:

docker run -it --rm --gpus all -v $(pwd)/torch_models:/root/.torch/models -v $(pwd)/checkpoints:/ai/checkpoints -v $(pwd)/test:/ai/test qooba/aifacereeanactment python3 ./ --source_image /ai/test/test.jpg --driving_video /ai/test/test.mp4 --output /ai/test/test_generated.mp4

NOTE: additional volumes (torch_models and checkpoints) are mount because during first run the trained neural networks are downloaded.

To reproduce the results we need to provide two files motion video and source image. In above example I put them into test directory and mount it into docker container (-v $(pwd)/test:/ai/test) to use them into it.

Below you have all command line options:

usage: [-h] [--config CONFIG] [--checkpoint CHECKPOINT]
                  [--source_image SOURCE_IMAGE]
                  [--driving_video DRIVING_VIDEO] [--crop_image]
                  [--crop_image_padding CROP_IMAGE_PADDING [CROP_IMAGE_PADDING ...]]
                  [--crop_video] [--output OUTPUT] [--relative]
                  [--no-relative] [--adapt_scale] [--no-adapt_scale]
                  [--find_best_frame] [--best_frame BEST_FRAME] [--cpu]


optional arguments:
  -h, --help            show this help message and exit
  --config CONFIG       path to config
  --checkpoint CHECKPOINT
                        path to checkpoint to restore
  --source_image SOURCE_IMAGE
                        source image
  --driving_video DRIVING_VIDEO
                        driving video
  --crop_image, -ci     autocrop image
                        autocrop image paddings left, upper, right, lower
  --crop_video, -cv     autocrop video
  --output OUTPUT       output video
  --relative            use relative or absolute keypoint coordinates
  --no-relative         don't use relative or absolute keypoint coordinates
  --adapt_scale         adapt movement scale based on convex hull of keypoints
  --no-adapt_scale      no adapt movement scale based on convex hull of
  --find_best_frame     Generate from the frame that is the most alligned with
                        source. (Only for faces, requires face_aligment lib)
  --best_frame BEST_FRAME
                        Set frame to start from.
  --cpu                 cpu mode.

New Face with AI


In this article I will show how to use artificial intelligence to generate human faces.

Before you will continue reading please watch short introduction:

Generative adversarial network

To generate realistic human faces, we can use neural networks with GAN (Generative adversarial network) architecture.

neural network architecture

The GaN network consists of two parts of the Generator whose task is to generate the image from random input and a discriminator that checks if the generated image is realistic.

training progress

During training, the networks compete with each other, the generator tries to generate better and better images
and thereby deceive the Discriminator. On the other hand, the Discriminator learns to distinguish between real and generated photos.

To train the discriminator, we use both real photos and those generated by the generator.

Finally, we can achieve the following results using DCGAN network.
As you can see some faces look realistic while some are distorted, additionally the network can only generate low resolution images.

training results

We can achieve much better results using the StyleGaN (arxiv article) network, which, among other things, differs in that the next layers of the network are progressively added during training.

I generated the images using pretrained networks and the effect is really amazing.

results stylegan

Unblur low quality face images with AI


In this article I will show how to improve the quality of blurred face images using
artificial intelligence. For this purpose I will use neural networks and FastAI library (ver. 1)

The project code is available on my github:
You can also use ready docker image:

Before you will continue reading please watch short introduction:

I have based o lot on the fastai course thus I definitely recommend to go through it.


To train neural network how to rebuild the face images we need to provide the
faces dataset which will show how low quality and blurred images should be reconstructed.
Thus we need pairs of low and high quality images.

To prepare the data set we can use available fases dataset eg. FFHQ, Tufts Face Database, CelebA

We will treat the original images as a high resolution data and rescale them
to prepare low resolution input:

import fastai
from import *
from fastai.callbacks import *
from fastai.utils.mem import *
from torchvision.models import vgg16_bn
from pathlib import Path

path = Path('/opt/notebooks/faces')
path_hr = path/'high_resolution'
path_lr = path/'small-96'

il = ImageList.from_folder(path_hr)

def resize_one(fn, i, path, size):
    dest = path/fn.relative_to(path_hr)
    dest.parent.mkdir(parents=True, exist_ok=True)
    img =
    targ_sz = resize_to(img, size, use_min=True)
    img = img.resize(targ_sz, resample=PIL.Image.BILINEAR).convert('RGB'), quality=60)

sets = [(path_lr, 96)]
for p,size in sets:
    if not p.exists(): 
        print(f"resizing to {size} into {p}")
        parallel(partial(resize_one, path=p, size=size), il.items)

Now we can create data bunch for training:

arch = models.resnet34
src = ImageImageList.from_folder(path_lr).split_by_rand_pct(0.1, seed=42)

def get_data(bs,size):
    data = (src.label_from_func(lambda x: path_hr/
           .transform(get_transforms(max_zoom=2.), size=size, tfm_y=True)
           .databunch(bs=bs,num_workers=0).normalize(imagenet_stats, do_y=True))

    data.c = 3
    return data

data = get_data(bs,size)


In this solution we will use a neural network with UNET architecture.

neural network architecture

The UNET neural network contains two parts Encoder and Decoder which are used to reconstruct the face image.
During the first stage Encoder fetch the input, extracts and aggregates the image features. At each stage the features maps are donwsampled.
Then Decoder uses extracted features and tries to rebuild the image upsampling it at each decoding stage. Finally we get regenerated images.

Additionally we need to define the Loss Function which will tell the model if the image was rebuilt correctly and allow to train the model.

To do this we will use additional neural network VGG-16. We will put Generated image and Original image (which is our target) to the network input. Then will compare the features extracted for both images at selected layers and according to this calculated the loss.

Finally we will use Adam optmizer to minimize the loss and achieve better result.

def gram_matrix(x):
    n,c,h,w = x.size()
    x = x.view(n, c, -1)
    return (x @ x.transpose(1,2))/(c*h*w)

base_loss = F.l1_loss

vgg_m = vgg16_bn(True).features.cuda().eval()
requires_grad(vgg_m, False)

blocks = [i-1 for i,o in enumerate(children(vgg_m)) if isinstance(o,nn.MaxPool2d)]

class FeatureLoss(nn.Module):
    def __init__(self, m_feat, layer_ids, layer_wgts):
        self.m_feat = m_feat
        self.loss_features = [self.m_feat[i] for i in layer_ids]
        self.hooks = hook_outputs(self.loss_features, detach=False)
        self.wgts = layer_wgts
        self.metric_names = ['pixel',] + [f'feat_{i}' for i in range(len(layer_ids))
              ] + [f'gram_{i}' for i in range(len(layer_ids))]

    def make_features(self, x, clone=False):
        return [(o.clone() if clone else o) for o in self.hooks.stored]

    def forward(self, input, target):
        out_feat = self.make_features(target, clone=True)
        in_feat = self.make_features(input)
        self.feat_losses = [base_loss(input,target)]
        self.feat_losses += [base_loss(f_in, f_out)*w
                             for f_in, f_out, w in zip(in_feat, out_feat, self.wgts)]
        self.feat_losses += [base_loss(gram_matrix(f_in), gram_matrix(f_out))*w**2 * 5e3
                             for f_in, f_out, w in zip(in_feat, out_feat, self.wgts)]
        self.metrics = dict(zip(self.metric_names, self.feat_losses))
        return sum(self.feat_losses)

    def __del__(self): self.hooks.remove()

feat_loss = FeatureLoss(vgg_m, blocks[2:5], [5,15,2])

learn = unet_learner(data, arch, wd=wd, loss_func=feat_loss, callback_fns=LossMetrics,
                     blur=True, norm_type=NormType.Weight)


After training we can use the model to regenerate the images:



Finally we can export the model and create the drag and drop application which fix the face images in web application.


The whole solution is packed into docker images thus you can simply start it using commands:

# with GPU
docker run -d --gpus all --rm -p 8000:8000 --name aiunblur qooba/aiunblur

# without GPU
docker run -d --rm -p 8000:8000 --name aiunblur qooba/aiunblur

To use GPU additional nvidia drivers (included in the NVIDIA CUDA Toolkit) are needed.