search

hab-spc / hab-ml

Harmful algae bloom CNN detection model, developed in PyTorch
6 stars 1 forks source link

readme

HAB detection with HAB-ML

The code here is for training, validating, and deploying HAB classification models, using convolutional neural networks in PyTorch.

What is HAB-ML?

HAB-ML (Harmful Algae Bloom Machine Learning) is a machine learning library for detecting the algae species from images collected by the Scripps Plankton Camera (SPC). There are a total of 46 species of interest, that need to be detected from the SPC database. For all HAB-ML models, they are pretrained off of ImageNet and fine tuned for this task.

Our task

The specific task here is to identify a harmful algae species at the species taxonomy level, i.e. given an image, return the predicted species classification.

Getting Started

The approach here is to add a classifier on top of the pre-trained CNN model and fine tune the model parameters by training on our domain specific data, i.e., phytoplankton images. This is done in main.py.

System Requirement

  1. Python 3.5 or higher
  2. PyTorch 1.0.1 or higher
  3. Python libraries: pandas, scikit-learn, scikit-image, matplotlib, numpy, lxml
  4. Example: Create a python environment named hab_env and install required libraries using pip:
    • virtualenv hab_env
    • source hab_env/bin/activate
    • pip install --user -r requirements.txt
  5. 12GB GPU (recommended)

It is recommended to run the program on GPU with 12 GB memory. Training (fine tuning) the model on a GPU, with ~10k images in training data, takes ~1 hour for 5 or 6 epochs depending on the hyperparameters used. The scoring can be done on CPU, but running on GPU is ~10-30x faster than on CPU

Download required files:

  1. Copy the files and directories in this repo into your work directory.
  2. Download the pre-trained ResNet models into the model directory of your work directory.
    • Run python download_pretrained.py from the model directory.

Execute the job

Run a saved model

To run a saved model in main.py, include deploy type arguments with the script call. This will run the fine-tuned model saved in the specified model directory to evaluate a batch of data. An example is given below, where we are running the 20_class_run model version on a dataset collected in 2018.

deploy_data=/data6/phytoplankton-db/csv/field_2018.csv
model_dir=/data6/yuanzhouyuan/hab/hab-ml/experiments/20_class_run/
python main.py --mode deploy --batch_size 16 --deploy_data $deploy_data --model_dir $model_dir
Train and validate the model
1. Run the main.py

You can run it with different arguments, such as --mode, --model_dir, --epochs, --lr.

2. Answer the input questions:
  1. Enter training set date (ex.20190708) :
    • To answer this question, type the date. Then, the code will set 1data_dir1 to /data6/plankton_test_db_new/data/$(date)/ directory to search for train.log and val.log. [The current workshop data folder is stored to date 0]
  2. Do you want to load existed checkpoint ? (y/n)
    • If you type y, then, it will set resume to True. And when it was loading checkpoint in Trainer.py, you will be asked to query the sqlite database in /data6/plankton_test_db_new/model/model.db.
    • If you type n, it will just load the ImageNet pre-trained weights.
  3. Do you want to save model to sql database? (y/n)
    • If you type y, It will ask you to enter today's date. Then, it will set model_dir into /data6/plankton_test_db_new/model/$(date)/$(time)/, and save the model info into it.
    • If you type n, it will save model to your model_dir you specified in argument. If you did not specified model_dir in argument, it will save to the default directory.

Sample 1: In this sample we will i) select a new training set ii) save the model to the sql database

(hab) zhy185@gpu2:/data6/yuanzhouyuan/hab/hab-ml$ python main.py --mode TRAIN --model_dir ./experiments/proro_run/ --epochs 0

Enter training set date (ex.20190708) :
>>> 0
# now the data_dir is /data6/plankton_test_db_new/data/0

Do you want to load existed checkpoint ? (y/n)
>>> n
Do you want to save model to sql database? (y/n)
>>> y
Enter today date (ex.20190708) :
>>> 20190802
#now model_dir will be saved to /data6/plankton_test_db_new/model/20190802/$(current_time)/

#Inserting current model to sql database
Any Addtional Comment to this model?
>>> just for testing, better not to use it
#the additional comment of this specific model is added to sqlite

Sample 2: In this example we will i) load an existing checkpoint (6) and ii) save the model to the sql database

(hab) zhy185@gpu2:/data6/yuanzhouyuan/hab/hab-ml$ python main.py --mode TRAIN --model_dir ./experiments/proro_run/ --epochs 0

Enter training set date (ex.20190708) :
>>> 0
# data_dir will be /data6/plankton_test_db_new/data/0
Do you want to load existed checkpoint ? (y/n)
>>> y
Do you want to save model to sql database? (y/n)
>>> y
Enter today date (ex.20190708) :
>>> 20190802
#model_dir will be /data6/plankton_test_db_new/model/20190802/$(current_time)/

# search for checkpoint
Enter query command (ex. SELECT * FROM models WHERE train_acc > 90) :
>>> SELECT * FROM models

#below are checkpoints info
#[ID, model name, model path, frezzing layers number, uselss variable, learning rate, batch size, best train acc, best test acc, epochs, date, class number, additional comment]
(1, 'resnet50', '/data6/plankton_test_db_new/model/20190730/1564555883/', 0, 'n', 0.001, 16, 38.833333333333336, 21.15506329113924, 1, 20190730, 24, 'just for testing, better not to use it')
(2, 'resnet50', '/data6/plankton_test_db_new/model/20190730/1564556191/', 0, 'n', 0.001, 16, 57.06944444444444, 30.799050632911392, 1, 20190730, 24, 'used for testing resume, no need to use it')
(3, 'resnet50', '/data6/plankton_test_db_new/model/20190730/1564556466/', 0, 'n', 0.001, 16, 0.0, 0.0, 15, 20190730, 24, 'used for testing VAL mode, no need to use it')
(4, 'resnet50', '/data6/plankton_test_db_new/model/20190730/1564557013/', 0, 'n', 0.001, 16, 89.80555555555556, 68.44145569620254, 15, 20190730, 24, 'testing without summary_state report')
(5, 'resnet50', '/data6/plankton_test_db_new/model/20190731/1564629304/', 0, 'n', 0.001, 16, 94.18402777777777, 70.57773109243698, 15, 20190731, 23, 'first try with 23 classes, performance is not good')
(6, 'resnet50', '/data6/plankton_test_db_new/model/20190802/1564786478/', 0, 'n', 0.001, 16, 0.0, 0.0, 0, 20190802, 14, 'just for testing, better not to use it')

Enter the ID number of the model:
>>> 6

The model you find is:
[(6, 'resnet50', '/data6/plankton_test_db_new/model/20190802/1564786478/', 0, 'n', 0.001, 16, 0.0, 0.0, 0, 20190802, 14, 'just for testing, better not to use it')]

# checkpoint '/data6/plankton_test_db_new/model/20190802/1564786478/model_best.pth.tar' will be loaded

#After training, i can enter some additional comment to the newly created model
Any Addtional Comment to this model?
>>> just for testing, don't use

Sample 3: In this example, we will not be loading an existing checkpoint or saving the the model

(hab) zhy185@gpu2:/data6/yuanzhouyuan/hab/hab-ml$ python main.py --mode TRAIN --model_dir ./experiments/proro_run/ --epochs 0

Enter training set date (ex.20190708) :
>>> 0
# data_dir will be /data6/plankton_test_db_new/data/0
Do you want to load existed checkpoint ? (y/n)
>>> n
Do you want to save model to sql database? (y/n)
>>> n
#model_dir will be the one specified in the arguments. The model_dir will be ./experiments/proro_run/

Benchmark

We fine tuned the ResNet50 model with different model size and hyperparameters and tested on one of our validation sets. The saved model, model, uses the highlighted parameters below.

Model input_size batch_size learning_rate epochs training_time Accuracy (val)

Directory Structure

The directory structure of the HAB-ML project looks like this: 

├── README.md          <- The top-level README for developers using this project.
├── data               <- Scripts to download or generate data
│   ├── dataloader.py
│   ├── d_utils.py
│   └── prepare_db.py
│
├── models             <- Trained and serialized models, model predictions, or model summaries
│   ├── download_pretrained.py
│   ├── layers.py
│   └── resnet.py
├── spc                <- SPICI module to download and upload SPC images
│
├── sql database       <- Scripts to get and set data in sql database
│
├── utils              <- Miscellaneous scripts
│   ├── config.py
│   ├── constants.py
│   ├── eval_utils.py
│   ├── logger.py
│   └── model_sql.py
├── requirements.txt   <- The requirements file for reproducing the analysis environment, e.g.
│                         generated with `pip freeze > requirements.txt`
│
├── main.py            <- Script to train models and then use trained models to make
│                         predictions
│
└── trainer.py         <- Module to handle model training