In this 2-hour long guided-project course, you will load a pretrained state of the art model CNN and you will train in PyTorch to classify radio signals with input as spectogram images. The data that you will use, consists of spectogram images (spectogram is a representation of audio signals) and there are targets such as ( Squiggle, Noises, Narrowband, etc). Furthermore, you will apply spectogram augmentation for classification task to augment spectogram images. Moreover, you are going to create train and evaluator function which will be helpful to write training loop. Lastly, you will use best trained model to classify radio signals given any 2D Spectogram of radio signal input images.
GD
Nice guided lab, however there are some content issues: 1. The last video is missing; 2. Some problem with certificates on loading the model.
HA
It was a wonderful project which not only covers a few concepts of signal processing but also sheds light on transfer learning with Pytorch.
Showing: 6 of 6
It was a wonderful project which not only covers a few concepts of signal processing but also sheds light on transfer learning with Pytorch.
This project was quit interesting
Nice guided lab, however there are some content issues: 1. The last video is missing; 2. Some problem with certificates on loading the model.
I feel the instructor put in very little effort. Usually in other courses the instructor provides an Completed Copy of the entire code .
There was an entire section of code and video which the instructor seems to have missed altogether.
The instructor seems to just copy the spec_augment library from some GitHub repository without showing which package to install. Overall I would say this was very poorly executed.
Very little discussion was spent into why spec-augment was used at all? Even a signle statement saying that the masking improves the Robustness of the Acoustic model (without the LM) on datasets such as Librispeech without any noise aware training shows good results would have sufficed.
Overall, this course looks reasonable useful. It needs some crucial improvements: (1) The training process did not improved model performance. The final accuracy values (for both training and validation datasets) were similar to their initial values. In other words, the model did not learned! (2) The course misses a final lecture to show how to make inferences. (3) make available the caption files. (4) Provide a complete and fully working version of the Jupyter Notebook.
If you are looking to learn how to implement pytorch dataloaders and neural networks the course covers those topics. If you are looking to learn anything about the RF signal processing/classification domain you will be disappointed.