Fine-tune in R: end-to-end example and some troubles

[statist-bhfz]

Hi,

I asked about fine-tune in R (#4817), and due to jeremiedb`s answers I finally was able to write complete R example: Dogs vs. Cats classification with mxnet and R. Please add this link to the list of tutorials (http://mxnet.io/tutorials/index.html and/or https://github.com/dmlc/mxnet/tree/master/example).

During this work I faced with new trouble. Unfortunately, all resnet models from http://data.dmlc.ml/models/imagenet/ completely failed after beginning of the training with some kind of memory allocation error. nvidia-smi utility shows that all memory (4Gb) is used even for batch_size = 1.

Then I tried inception-bn. It works fine, but my R session crashed after each iteration and checkpoint save.

I also noticed that iterators uses almost all available RAM and swap immediately after creating. Is it normal behavior?

[matt32106]

I worked on the same project. I have not tested all models but i am able to run the inception-bn-28-small model without memory issues with batch size 512 (i used symbol_inception-bn-28-small.R in the image-classification directory). I'll give a try to the resnet examples.

Maybe you should make a pull request for your example code (to go into example/image-classification), it is useful for R users.

[statist-bhfz]

Yes, inception-bn-28-small works good, but only with small images (my post (in Russian) about it). For images like 224x224 it is necessary to use another networks.

I'm not going to make pull request according to http://mxnet.io/community/contribute.html:

If you have blog posts about MXNet or tutorials that use MXNet, please tell us and we’ll add a link to them in the examples on GitHub.

Maybe I can add small vignette about fine-tune itself (without preprocessing steps) to https://github.com/dmlc/mxnet/tree/master/R-package/vignettes

[thirdwing]

@statist-bhfz A PR will be truly welcome.