Expose iter_size solver option #744
Conversation
| $('#iter-size-option').show(); | ||
| } | ||
| {% endif %} | ||
| {% endfor %} |
There was a problem hiding this comment.
or maybe you can extend the structure on lines 414-416 with a can_accumulate_gradients field for a more direct access to framework properties?
There was a problem hiding this comment.
Oh yeah, good call. I also forgot to add this to the Generic path ...
| network = network, | ||
| random_seed = form.random_seed.data, | ||
| solver_type = form.solver_type.data, | ||
| shuffle = form.shuffle.data, |
There was a problem hiding this comment.
Sorry for the whitespace changes. It had to be done at some point.
|
Refactored to use the name |
|
That looks good to me - I haven't tested let me know if you want me to. |
|
Actually it would be great if you could do a quick smoketest. No need to validate convergence, just check that:
Assuming all is good, please merge. Thanks! |
| from digits.frameworks import CaffeFramework | ||
| if CaffeFramework().can_accumulate_gradients(): | ||
| solver.iter_size = self.batch_accumulation | ||
|
|
There was a problem hiding this comment.
I think you need to make this change in save_files_generic() too.
|
The classification path is working for me - there is a little omission in the generic path. |
Rename iter_size -> batch_accumulation (less Caffe-specific) Disable for Torch and older versions of Caffe
Expose iter_size solver option
Close #660
Replace #675
Expose
iter_sizesolver option (BVLC/caffe#1977) for NVcaffe >= 0.14./cc @drozdvadym