Developpement of transfert learning methods for custom image recognition
###Run the code
run using the following command : python35 transfert_learning_vgg16.py
-To train the model we have used the '.fit' keras method and we have fixed the 'epochs=10' only for the test.
So you can try with other value higher than 10 to improve the accuracy of the model
-The saved trained model are in the forlder 'output_trained_model'
The dataset are in this path : Transfert_Learning_using_Resnet50/data.zip
(when you use tranfert learning based on vgg16, please move the folder data.zip in the same that of vgg16)
In the case oF TL using vgg16 architecture (with fine tuning):
-Training time is arround 73 minutes
-loss=0.3300, accuracy: 97.5309%
In the case oF TL using resnet50 architecture (with fine tuning):
-Training time is arround 37 minutes
-loss=0.1147, accuracy: 96.9136%
-The analysis of the training, loss, and validation as function of epoch numberare in:
Fig_train_loss_vs_val_loss.png and Fig_train_acc_vs_val_acc.png (see folder analysis_results)
- We should improve the computing time in the training step by using GPU