Checkpointing: Save/Load Parameters to Disk, Resume Training
The problem, from first principles
Training a real model can take hours or days — a crash, a preemption, or simply wanting to pause and resume later all mean the same thing: the current state of training (every parameter's current value, plus which epoch training had reached) needs to survive being written to disk and read back later, exactly.
From theory to code
Implement save_checkpoint(params, epoch, path) and load_checkpoint(path), a matched pair that write a model's parameters and its epoch number to a single file, and read them back exactly.
Constraints
params: a dict mapping parameter names (strings) to NumPy arrays of any shape/dtype.path: expected to already end in.npz.save_checkpointwrites every named array plus theepochinteger to one file atpath.load_checkpoint(path)returns(params, epoch):paramshas exactly the same name→array mapping that was saved,epochis a plain Pythonint.- A round trip (
load_checkpoint(path)immediately aftersave_checkpoint(params, epoch, path)) must reproduceparamsandepochexactly.
Hints
Open one at a time. Each gives away a little more than the last.
Hint 1
np.savez(path, **params) saves every dict entry as its own named array in one file — dict-unpacking with ** turns {"w": arr} into the keyword argument w=arr automatically.
Hint 2
Save the epoch under a key that could never collide with a real parameter name (like "__epoch__"), so load_checkpoint can tell "this one's the epoch" apart from "this one's a parameter" just by its key.
Keywords
No results yet
Run your code to check it against the examples, or submit to run every test.