something about small dataset for finetuning?

[jsjs0827]

How did you do the project? Can you give me a way to contact you for more details. I just doubt about finetune the network with only 6 pictures.What leads to the good result? The Imagenet parameters?Last but not least,can you give me some advice about transfer learning about cnn besides finetuning?

[radekosmulski]

You can find all the code I use in the repository. Please feel free to download it and run it yourself.

Imagenet has categories for multiple cat and dog breeds. What could be happening is that the original CNN stack already encodes the distinction 'this is a cat', 'this is a dog' and we are only learning to decode it. The extent to which this is happening is unknown and is dependent on whether the last layer of the CNN contains information on lower level features (shapes, textures) or higher level features (cat, dog).

This is open for speculation and further rigorous testing could be carried out. Also, there are nice papers on visualizing what the network sees at each of the layers, with Visualizing and Understanding Convolutional Networks being as far as I know the seminal one.

Having said that, I still find the results interesting in many ways.

As for the resources on transfer learning, I do not know of a better one that I know of than the Deep Learning For Coders, Part 1 which is available for free. I would encourage you to join the forum there as well as it is a great place for discussions.

If you have any further questions, please ask them here - maybe someone else will also find the conversation useful.

[jsjs0827]

I'm appreciated for hearing from you so fast.Meanwile,I need to say sorry for my terrible English and carelessness.I want to ask some advice addressed to the transfer learning in deep learning,but not about finetuning(also a way of transfer learning). Maybe I would write to you soon,after reading your codes. Appreciated again for your patient reply.