Computer vision (3): retraining one’s own data model with induction-v3 model


Retraining your data model with inception-v3


Modern image recognition models have millions of parameters. Training from scratch requires a large number of sample data and consumes huge computing resources (hundreds of GPUs). Therefore, the cost of retrain a model by transfer learning is low,Using concept-v3 as a trained model to achieve their own image classification and recognition

Introduction to the directory of inception model file:

Data directory: the data to be trained is placed in this directory: induction_ Model: put the download inception model in this directory:



test_ Images: after the training is completed, test the directory of pictures


After the data preparation is completed, perform the following steps:

Step 1:


Run batch file retrain.bat


python ^
--bottleneck_ Dir bottleneck ^ ා generate the data of each training picture
--how_ many_ training_ Steps 200
--model_ dir inception_ Model
--output_ graph output_ graph.pb  ^Output model after training
--output_ labels output_ labels.txt  ^Output label after training
--image_dir data



Run batch file


Step 2:

After training, the test is as follows




The test results are as follows:





Conclusion: if the test objects are not similar, the recognition rate is very high, but if the objects are similar in shape, the recognition rate is not high

Source code: Extraction code: JIT