Hello,
Is there any way to apply transfer learning to a custom pre-trained model? Either a custom or an existing Pytorch pre-trained model for example. I followed the documentation on how to bring a pre-trained model using custom-learning-blocks and that seems to work fine. However, I did not found any specifications on transfer learning, is this functionality available? I’m using Image Classification.
Thanks,
Rodrigo Santa Comba
Can you share some more detail please? I’ll try get the tech team to help answering.
Best
Eoin
Thank you for the fast reply. My project involves training many CNNs and compare their accuracy, efficiency, memory etc. As EI doesn’t have that many models available for transfer learning for the memory specifications I need, under 1MB CPU Flash, I would like to use some pre-trained models I found online, already trained using ImageNet and other datasets. I would like to use transfer learning in the EI platform to make these models detect fire and smoke (Instead of the hundreds of different classes they detect), by using my own dataset and retraining only the last layers of these models.
I have been able to use a custom learning block to train a model, following this example: Custom learning blocks | Edge Impulse Documentation. However, I haven’t quite understood how to make the model keep the weights it has as they are and train only the last layers. I don’t know if I made myself very clear, English is not my first language and I am not at all an expert in ML, any help would be appreciated.