Weights sharding for Keras saving

[nkovela1]

This PR adds weights sharding initial functionality to the Keras saving/loading APIs, which are accessed by passing the sharded=True flag to the corresponding saving/loading calls.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 67.76860% with 39 lines in your changes are missing coverage. Please review.

Project coverage is 75.61%. Comparing base (c8700f4) to head (cfbb761).
Report is 88 commits behind head on master.

Files	Patch %	Lines
keras/saving/saving_lib.py	66.66%	26 Missing and 13 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #19286      +/-   ##
==========================================
- Coverage   80.14%   75.61%   -4.53%     
==========================================
  Files         341      365      +24     
  Lines       36163    39909    +3746     
  Branches     7116     7747     +631     
==========================================
+ Hits        28982    30177    +1195     
- Misses       5578     8054    +2476     
- Partials     1603     1678      +75     

Flag	Coverage Δ
keras	75.46% <67.76%> (-4.53%)	⬇️
keras-jax	59.71% <67.76%> (-3.35%)	⬇️
keras-numpy	54.30% <66.11%> (-2.79%)	⬇️
keras-tensorflow	61.21% <67.76%> (-3.44%)	⬇️
keras-torch	60.29% <53.71%> (-3.58%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[fchollet]

Shouldn't max_size be an int? e.g. in MB?

[fchollet]

Why pick a convnet for a large model?

[fchollet]

Why should sharded be configurable here -- wouldn't it just depend on the file and the model?

[mattdangerw]

Talked with Neel a bit about this, but one idea, building off the recent change Francois made with __getitem__/__setitem__...

• Create h5 groups on variable write, instead of eagerly when creating a H5Entry. (Side note, this would clean up our h5 entires, no empty groups for layers without weights.)
• H5Entry is just a dict like object that proxies calls get/set calls to parent H5Store.
• On write, H5Store could just keep a running list of how big the shard is currently, and "roll over" to a new shard as soon as the next variable would be bigger than shard limit.
• On read, H5Store could just check every shard file for the weight is try to load (as checking is cheap, reading is slow).

Pseudocode:

write(path, key, value):
    if self.current_shard_size + value.nbytes > self.shard_size:
        close current shard
        open new shard file
        self.current_shard_size = 0
    group = create parent groups if needed
    self.current_shard_size += value.nbytes
    group[key] = value

read(path, key):
    for file in shards:
        if path in file:
            group = file[path]
            if key in group:
                return group[key]

This could be fairly simple. Avoid the need for a separate class if we want (though we still could), allow splitting up individual layer weight across shards (important if you have one big layer).

This could even allow avoiding the json file entirely I think? Supporting something like this:

# If shard_size is set, pass a format string as path?
filenames = model.save_weights("./model_{}.weights.h5", shard_size="10GB")
# Load weights handles loading a list of files, and checking all files for the variables.
model.load_weights(filenames)

This last bit is optional, just though it was interesting. What do people think?