TRANSFER LEARNING Using a pre trained network on images not in training set is known as transfer learning
DIFFERENT ARCHITECTURES WE CAN USE
THESE ARE THE TOP 1 AND TOP 2 ERROR
- The numbers such as VCG-11 where 11 are the no of layers.
- When we are using this we need to do the trade-off between accuracy and speed.
- They are massively deep.They have 100's of hidden layer
MODELS
USING DENSENET
- It has 121 different layers
LOADING AND ARCHITECTURE
- After loading we can see the architecture.We have features and classifiers
- The classifier has 1024 input features and 1000 output features
- Imagenet dataset has 1000 different classes.The no of output should be 1000 for these classes.
- The features would not be used and need to be freezed.
- We need feature part static, to this we would freeze our feature parameter as shown below
- This would increase processing speed as it would not keep track of features
BUILD CLASSIFIER
We would build our classifier as shown below
ATTACH TO OUR MODEL
MOVE TO GPU
- We can move all the computing to a GPU by specifying model.cuda . This would move all the parameters and model to the GPU
- To move the tensors to the GPU, we need to use image.cuda
- To specify cpu we need to change it to image.cpu, model.cpu
COMPARISON BETWEEN CPU AND GPU
- We would get a speed of over a 100 times
The article provides a practical introduction to transfer learning by showing how a pretrained image-classification network can be adapted to a new problem instead of training an entire deep learning architecture from scratch. The discussion of freezing pretrained features and replacing the classifier makes the workflow especially clear, while the trade-off between model depth, accuracy, and processing speed is an important consideration when selecting an architecture.
ReplyDeleteThe use of DenseNet and the explanation of its pretrained feature and classifier components also demonstrate how established architectures can provide a strong starting point for image-based applications. Image Processing Projects For Final Year is relevant for exploring practical project applications where these image-processing and deep-learning techniques can be applied.
The final step of attaching a newly built classifier to frozen pretrained features illustrates how transfer learning can reduce the amount of training required while adapting a model to a different set of classes. Deep Learning Projects for Final Year provides a broader context for applying architectures such as DenseNet and transfer-learning approaches to practical deep-learning projects.
ReplyDelete