The image classification CNN served with Flask

The model allows configuring a network architecture by:

• The number of filters at the first convolutional layer.
• The number of convolution blocks composed of single or double convolutional layers with a doubled number of filters with respect to the previous block.
• The number of neurons at the last fully-connected layer.
• The amount of dropout at the fully-connected layer.
• The activation, loss and accuracy functions.

The model allows applying data augmentation provided by imgaug module with customizable parameters of horizontal flipping, contrast normalization, and affine transformation. The input images can be pre-processed (by default) to select the central object of interest. The trained model can be served with Docker as a micro-service.

Build a container:

sudo docker build -t image-cnn:latest .

Run traning of a model wit data augmentation and save a model parameters to /app/model-output/params.ini:

 sudo docker run -v ~/images:/app/images -it -d image-cnn run.py train \
 --images-dir /app/images/*/*.jpeg --network-name net \
 --model-output-dir /app/model-output \
 --model-params-output-path /app/model-output/params.ini \
 --num-epochs 100 --depth 4 --num-base-filters 8 \
 --num-dense-neurons 128 --batch-size 8 --batch-normalization  \
 --learning-rate 0.001 --num-classes 2 --optimizer-name rmsprop \
 --do-augmentation --verbosity 1 --print-model

Save state of the container:

 sudo docker commit DOCKER_CONTAINER_ID image-cnn/trained

Run a web-server to perform prediction with a model, which configuration is stored in /app/model-output/params.ini:

 sudo docker run -p 6000:80 -d -t image-cnn/trained app.py \
 --port 80 --host=0.0.0.0 --model-params-path /app/model-output/params.ini

Send a GET request with an ID of shared image at Google Drive:

curl localhost:6000/predict_gd?image_id=GOOGLE_DRIVE_IMAGE_ID

or with URL:

curl localhost:6000/predict_url?image_url=URL