[example] finetuning CaffeNet on Flickr Style data #970
Conversation
There was a problem hiding this comment.
Give the model a distinct name like StyleNet, CaffeFTFlickrStyle, or whatever.
|
@shelhamer ready for review |
There was a problem hiding this comment.
This should snapshot to the example dir, like the other examples.
There was a problem hiding this comment.
disagree, i think model snapshots should be in data/ and trained models in models/ (coming up in zoo)
There was a problem hiding this comment.
Oh, I only just saw this -- I would have replied here. I'm not comfortable with filing snapshots as "data" as they are models, merely intermediate ones on the way to the "final" model, which is only chosen by accepting a given validation error in most cases.
|
Thanks for the example Sergey! A few comments on framing the merits of fine-tuning, a path issue, and suggestion for comments in the prototxt and this is ready. |
|
disagree on paths, but rest is done |
|
Why would a model go to data? No model saves to data. Not CaffeNet from On Wednesday, August 27, 2014, Sergey Karayev notifications@github.com
|
|
Further this example uniquely runs from the Caffe root instead of the example's dir. Is there a reason for that choice? These choices should be uniform across examples. |
|
I think it makes a lot more sense to run everything from caffe root than to deal with |
Makes sense. We can standardize this in follow-up commits.
I'm open to this. Of course the choice isn't in the scope of this PR but let's talk offline. |
|
Take my last commit as compromise |
4278286 to
c01f07a
Compare
- further detail merits of fine-tuning: less starving for itme and data - set random seed for reproducing the tutorial - 1557 train / 382 test split is more indicative of training quality than splits of 200 images
Fine-tuning tutorial: adapt CaffeNet to Flickr Style data
|
Thanks for the fine-tuning tutorial on a neat task Sergey! I made a few edits to the text and expanded the fine-tuning to 2,000 images for a better indication of training progress. |
|
Fixed paths and test accuracy along with including a few details in a857001. |
Fine-tuning tutorial: adapt CaffeNet to Flickr Style data
Fine-tuning tutorial: adapt CaffeNet to Flickr Style data
This example shows how to fine-tune the BVLC-distributed CaffeNet model on a different dataset: Flickr Style, which has style category labels.
The dataset is distributed as a list of URLs with corresponding labels.
Using a script, we will download a small subset of the data, split it into train and val portinos, and use the ImageDataLayer to load images directly from the filesystem.