This repo contains three methods for training and deploying a image classification task.
- In
Image_Classfication_Keras/, use keras to train a network, including training, testing codes. - In
Image_Classfication_Mobile, use TFmobile/TFLite to train, trained network can deploy on mobile devices. - In
Image_Classification_TensorflowSlim, use TFSlim as a base to train, focus on realize high accurate precision, run by Personal PC / server.
Usually, there are several steps for a Image_Classfication_work.
- open source data: There are many open source dataset from the world. like ImageNet, COCO, PASVAL, google Open Images, and some kaggle public competition dataset.
- collected by yourself data
- data argumentation
- tensoflow provided classication Pre-trained Models
- according to your application, make tradeoff between accuracy and speed
- consider intialization and generalization 👍👍👍
- Building powerful image classification models using very little data
- simple fine-tune actively training on Keras
- TensorFlow for Poets: How to train a custom image recognition model
- How to fine tune a image classifier on Tensorflow Flowers Data
- How to fine tune a image classifer on our own dataset
- TensorFlow-Slim image classification 👍👍👍
- optimizer selection
- calculate numeric metrics(Accuracy, Precision, Recall, F1)
- determine classification threshold (ROC Curve, PR curve)
- No Prediction bias
- TF Mobile and TF lite
- PC
- server
I fine trained a VGG16 gender recognition model based on my own dataset[22000].
