HPA-Special-Prize

The special prize winner in the kaggle HPA competition

SOFTWARE:

Python 2.7 (should work fine with 3.x) CUDA 9.0 cuddn 7.0.5 nvidia drivers v.384

Installation

The solution is using the OpenVINO by intel. please install it before attempting to do inference

apt-get install -y --no-install-recommends \
        build-essential \
        cpio \
        curl \
        git \
        lsb-release \
        pciutils \
        python3.5 \
        python3-pip \
        python3-dev \
        python3-setuptools \
        sudo

pip3 install tensorflow numpy pandas networkx tqdm

git clone https://github.com/opconty/keras-shufflenetV2.git

DATA SETUP (assumes the Kaggle API is installed)

below are the shell commands used in each step, as run from the top level directory

mkdir -p data/
cd data
kaggle competitions download -c human-protein-atlas-image-classificatin
mkdir test
mkdir train
cd test && unzip ../test.zip
#get external data
wget https://storage.googleapis.com/kaggle-forum-message-attachments/430860/10774/HPAv18RGBY_WithoutUncertain_wodpl.csv
python download_hpa.py
python conv_512.py

MODEL BUILD: There are three options to produce the solution.

python train train/shufflenet_test_enhanced.py

Freezing to pb

git clone https://github.com/amir-abdi/keras_to_tensorflow
python3 keras_to_tensorflow/keras_to_tensorflow.py --input_model model.model --output_model frozen_model.pb
# converting to openvino
python3 /opt/intel/computer_vision_sdk/deployment_tools/model_optimizer/mo_tf.py --input_model frozen_model.pb --input_shape [1,512,512,3] --data_type FP32

References

References

(paper)ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

(repo)Trents iterative stratification

(repo)Keras implementation of shufflenet