Skip to content

alt0xFF/hongkong_flowers

Repository files navigation

Hong Kong Flower Identitfication APP

An APP for identify flowers in Hong Kong with Deep Learning technology.

screen.jpg

Prerequisite

Keras or pyTorch installed.

Docker

It is very easy to set up a docker container for pyTorch and Keras using the following command.

pyTorch:

$sudo docker pull floydhub/pytorch:0.2.0-gpu-py2.11

(Note: remove -gpu if you want CPU only. Change to py3.11 if using python3.)

$sudo nvidia-docker run -ti -v YOURDIRECTORY:/workspace/ -p 8889:8888 -p 8097:8097 floydhub/pytorch:0.2.0-gpu-py2.11

For CPU use normal docker. You can also add /bin/bash at the end of the command to use bash instead.

Go to your localhost:8889 to access pyTorch Jupyter notebook!

Keras

$sudo docker pull floydhub/tensorflow:1.3.0-gpu-py2_aws.12 (settings likewise as above.)

$sudo nvidia-docker run -ti -v YOURDIRECTORY:/workspace/ -p 8888:8888 -p 6006:6006 floydhub/tensorflow:1.3.0-gpu-py2_aws.12

Go to your localhost:8888 to access Keras and Tensorflow Jupyter notebook!

Fine-tuning

Fine-tuning the pre-trained ResNet50 with Oxford 102 flowers dataset

$./finetuning/boostrap.sh to download oxford102 dataset

$python resnet50.py to start fine-tuning

Training

Go to options.py and change data_dir to your own dataset ABSOLUTE path.

Choose which library, model, optimizer and loss to run in options.py by changing self.configs.

$python train.py to start training.

Adding your own model

To add your model, simply do the following:

  1. create your model class in core/YOURLIBRARYCHOICE/models, note that it must take two arguments (args, num_classes).
  2. add your model class to ModelsDict in core/YOURLIBRARYCHOICE/parser.py
  3. add your model, optimizer and loss function of your choice to CONFIGS in options.py
  4. change self.configs to your model in options.py

Deploy to GCloud

  1. First at all, install Google Cloud SDK
    $gcloud init # set up gcloud compute on your computer,
    $gcloud auth application-default login # or if you already initialized before
  1. Compile your trained model and upload to Google Cloud Storage
    $export MODEL_BUCKET=gs://dlhk-flower.appspot.com
    $python ./deploy \
        --trained_model=model_best_weights \
        --model_dir=./checkpoints/2017-10-04_experiment_0/\
        --bucket_dir=$MODEL_BUCKET            
  1. Create model on ml-engine
    $export MODEL_BINARIES=gs://dlhk-flower.appspot.com/export
    $export MODEL_NAME=baseline
    $gcloud ml-engine models create $MODEL_NAME --regions=asia-east1 # if you are first time to create model,
    $gcloud ml-engine versions create v1 --model $MODEL_NAME --origin $MODEL_BINARIES --runtime-version 1.3 # otherwise just create a new version

TODO

pyTorch:

  • Visdom.
  • logger for saving results.
  • model.test()

Keras:

  • Model.
  • dataloading.
  • model.train()
  • model.validate()
  • model.test()
  • tensorboard.

Citations:

Fine-tuning Deep Convolutional Networks for Plant Recognition

About

HK flowers project initial draft

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published