Decouple the computational batch size and minibatch size by accumulating gradients

[shelhamer]

Accumulate gradients across batches through the iter_size solver field. With this setting batch_size: 16 with iter_size: 1 and batch_size: 4 with iter_size: 4 are equivalent.

master edition of #1663.

• deduplicate solver logic: done by Deduplicate solver regularization, logging, and local rates and decays #2518
• adjust local_rate and local_decay according to iter_size normalize gradients by iter_size
• test equality of updates for batch size equivalents

Historical context:
From @longjon

This PRs the gradient accumulation branch living at https://github.com/shelhamer/caffe/tree/accum-grad. I took a lighter approach here than the one there: parameter gradients are always accumulated, there is no other option. The gradient checker is made correct by zero-initing parameter diffs.

Issues:

• This changes the behavior of Backward. External code that used Backward is likely to break, if there is any.
• I think this breaks solvers other than SGDSolver, but haven't thought carefully about that yet.

From @jeffdonahue

Have we thought about how to handle the case when we're sharing parameters but using different learning rates? I would be okay with simply disallowing that case since it would probably be a pretty weird thing to do. Otherwise the only other way I can think to handle it is pretty messy -- we could have a a special case where, e.g. if blobs_lr is 2 in one layer but 1 in all others, the Net could prescale (by a factor of 2) the top_diff for the layer with blobs_lr 2 by 2... Actually, even that wouldn't work if the layer has other shared param blobs that don't also have the same relative LR...

From @shelhamer

Always accumulating is simple and good, but let's review the weight sharing and solvers issues before merging.

[tnarihi]

I think this does not work correctly. Diving by iter_size should be applied before accumulating parameter decays.

[tnarihi]

Multiplying local_decay by iter_size should be okay?

Dtype local_decay = weight_decay * net_params_weight_decay[param_id] * this->param_.iter_size();

[longjon]

Ah... good point.

[shelhamer]

To clarify: the local_decay needs to be multiplied by the iter_size because the update will include the product of local_rate and local_decay. That is, the update by https://github.com/BVLC/caffe/blob/master/src/caffe/solver.cpp#L497-L499 is computed after weight decay is included on https://github.com/BVLC/caffe/blob/master/src/caffe/solver.cpp#L479-L483. As is, weight decay is defined per iteration so should not be scaled by the effective batch size of batch_size * iter_size.

[tnarihi]

Commented on the diff.

By the way, I am not understanding very well about what @jeffdonahue mentions. Is there any relations between this PR and weight sharing. Gradient accumulations among shared parameters are computed independently (since diffs are not shared), and applying learning rates are also independent (then, accumulated to owners). If my understanding is correct, accumulating gradient does not matter for weight sharing. I may be missing something..
Actually, for my purpose, very recently I have implemented accumulating gradient using a simple way but it needs additional memory (It doesn't change the behavior of Backward). tnarihi/caffe@d016dbd
If my understanding is incorrect, my branch also doesn't work in the special case Jeff mentions.

Anyway, the idea of always accumulating is very good (less memory) if the issues are solved. Both issues do not matter to me since I usually use SGDSolver and I will notice the behavior change of Backward, but it might matter to others :P I will move to use this when this PR is done.

[jeffdonahue]

Oops, I haven't actually stepped through this myself, but I think you're totally right @tnarihi -- there shouldn't be an issue with weight sharing in this implementation. I was confusing it with my version -- I had rebased my RNN PR (#1873) on this, and then just threw the additional changes to net.cpp I had made in my recurrent branch into a single non-descript commit (jeffdonahue@8983e6f). That commit should really probably have been 3 separate commits with good clarifying messages, as they were pretty substantial now that I look back on it... The first is that I indeed added ShareDiff to the parameter sharing behavior, for the exact reason you mention -- it's essential for memory efficiency in the RNN/LSTMLayers added there (which are implemented as unrolled networks). The other is that I removed the Net's shared parameter gradient accumulation since they're sharing diffs, which gives a bit of a speedup but probably doesn't really matter much because it's just an axpy (but O(T) of them for unrolled recurrent nets...). Those changes do assume lr_mult (the new blobs_lr) is the same for all shared parameters (I should probably have added that that PR is not currently mergable...)

Besides the other issues Takuya mentioned, I now think this is strictly good (i.e. it doesn't break anything that works now) and should be merged. Maybe I'll write a new PR based on this, or a commit to append to this one, that does ShareDiff (and skips the Net's accumulation) for any shared parameters where all lr_mults match. Not the prettiest solution, but I think the speed/memory savings is worth the additional complexity in this case.

[tnarihi]

I see. Thanks Jeff! Sharing diff for weight sharing is nice for memory consumption. I think to restrict all lr_mults match should be okay.

The other thing is, I think, to notifying developers (especially for developers working on PR regarding layers that has parameter updates) that Backward behavior change is necessary. To add a test checking if all layers work in gradient accumulation is better, but it seems difficult...

[jeffdonahue]

The other thing is, I think, to notifying developers (especially for developers working on PR regarding layers that has parameter updates) that Backward behavior change is necessary. To add a test checking if all layers work in gradient accumulation is better, but it seems difficult...

Another good point. At some point I had modified the gradient checker to check accumulation (by adding some random noise to the param blob diffs, calling Backward, then subtracting the noise before checking the result) -- I can try to dig that up to add to this.

[tnarihi]

That sounds nice idea. Abstracting Backward testing could solve this issue. Currently gradient checker seems to play the role.

[longjon]

Actually I have thought of a simpler way to implement this that is independent of gradient accumulation. Maybe it is too tricky, maybe not. Will update.

(I am still mildly in favor of always accumulating gradients, disallowing different lr_mults, and simplifying weight sharing. Ideally one implements lr_mult by having a backward net that is slightly different from one's forward net.)

[sguada]

@shelhamer @jeffdonahue @longjon what is happening with this PR, I think we need to find a solution and merge it as soon as possible. Actually I thought it was already merged since the solution has been around for a while.

[shelhamer]

Accumulating gradients includes subtleties with regards to scaling gradients and hyperparameters w.r.t. to the effective batch size vs. the computational batch size batch_size and iteration size iter_size.

For merge, this needs a test that compares the updates computed by batch_size: 16 and iter_size: 1 with batch_size: 4 and iter_size: 4 for instance. We have merely reasoned our way to correctness.

[longjon]

Right, this should be fine after @shelhamer's list. My idea for a simpler implementation did not pan out; it would only have worked for SGD with momentum.

[shelhamer]

0e7a078 makes the normalization for accumulation more obvious and fixes the issue with AdaGrad by normalizing the gradient before the update and history are computed.

However, when gradients are accumulated there's overhead for this separate scaling step. The time to update CaffeNet parameters for batch_size: 128 and iter_size: 2 rises 1.3x to 3.1 ms from 2.3 ms. If we truly care about the 1 ms / iter AdaGradSolver::ComputeUpdateValue() could be hacked instead.

[shelhamer]

Merging for control of memory usage now that this is simple and tested.

@sguada sorry for the wait!

[hli2020]

I think the PReLU layer also needs to accumulate the gradients. @shelhamer

[tnarihi]

Here is my implementation of PReLU gradient accumulation: tnarihi@4d3fbd5
@shelhamer You can cherry-pick it (code may not be clean).

[shelhamer]

@tnarihi sorry, I missed the PReLU accumulation when merging accumulating gradients. Could you send a PR for that particular patch? Thanks.

Thanks for the comment @hli2020

[shelhamer]

Oh sorry, I missed the cherry pick + commit ID. That'll work fine.

[gcr]

Is this normalization correct?

Doing this will reduce the gradient by a factor of iter_size compared to computing the gradient over an entire batch. If I'm interpreting this correctly, learning rates should be multiplied by iter_size to overcome this existing code.

Or: Is learning rate automatically scaled by the batch size elsewhere, and this code is necessary to account for the effective increase in the batch size?

[shelhamer]

It is done this way due to the separation of Net and Solver but it is correct. Net normalizes by the (computation) batch size but only Solver knows about iter_size so it does the portion of the normalization needed to handle accumulation.