doem97 / slc_chen

Image segmentation and classification framework, based on KERAS, features in controlling of large-scale experiments.

🔬SLC Chen

Skin Lesion Classification framework based on Keras.

皮肤镜图像病理分类框架。

Introduction 💡

SLC is an KERAS framework for large-scale deep learning experiments. It feautres in simple usage and convinient experienments tracking.

One-click Usage 🧯

The main idea of the project is that the deep-learning experiment include two modules: process control and model control. You can train model as simple as:

exp = ProcessControl('config_file')
model = ModelControl('config_file')
exp.train(mc)

All-in-one config file. As in above pseudo-code, config_file is an all-in-one config of the training pipeline, model design, data processing and evaluation.

Detail Usage 🕹

A simple classification exp to show the power of SLC.

Pre-Settings

1. Run ./environment_set.sh in the path where you want the SLC_chen project be placed.

   E.g. in path home/user/, the script will:

   1. create file structure into home/user/SLC_chen/

   2. download and unzip the dataset to home/user/SLC_chen/dataset/128_128

      **notice:

      • You can choose what dataset size to use in script
      • Original images didn't provided with download link cause its too huge

2. Move into dir SLC_chen/experiment_space/, and do your experiments.

SkinLesionClassify()

Class SkinLesionClassify() is for process control. The class takes control of high level oprations in deep learning experiments.

Class SkinLesionClassify control the below process:

• Prepare data
• Set workpaths
• Data generate
• Feed into training or validation
• Get scores
• Save necessary intermediate data
• Save checkpoint
• Create log

The whole training process is as simple as

import SkinLesionClassify
slc = SkinLesionClassify.SkinLesionClassify()
slc.train()

But in the process, there are so many parameters that too trivial to log. Thus for control and reproduce, all configurable parameters are loaded from .conf file when init the SkinLesionClassify class like this:

**notice: .conf file must be provided.

config_file = "./SLC.conf"
slc = SkinLesionClassify.SkinLesionClassify(config_file)

You can take a look of .conf file in code/SLC_template.conf. It listed all parameters can be changed.

SLCModel

Class SLCModel() is for model control. The class has a Keras.model() object as main part.

SLCModel() can do:

• Load metrics and loss
• Compile Keras model object
• Change model

The model control is as:

import SLC_models
slc_model = SLC_models.SLCModel('config_file')
slc_model.loadModel('model_name', **kwargs)

For elastic change of model object, here is the only **kwargs didn't zipped in config file. In one set of similar experiments, the **kwargs may be the most frequently changed parameter, but don't worry, it will be saved into a log file part from config_file.

Example

import SkinLesionClassify
import SLC_models

config_file = "./SLC.conf"
slc = SkinLesionClassify.SkinLesionClassify(config_file)
slc.resizeImages()
slc_model = SLC_models.SLCModel(config_file)
slc_model.loadModel('resnet_v1', depth = 20)
slc.train(slc_model)
slc.evaluation(slc_model)

Notice ⛔️

• /code folder should be more stastic and .gitignore file will ignore tracking operations in /experiment_space
• data files can be orgnized like:
  • /dataset/HAM10000_metadata.csv,/origin,/resized_{}_{}