Don't include saved state in keras_model.weights #140
Conversation
drasmuss
force-pushed
the
model-checkpoint
branch
2 times, most recently
from
6fc1f6e to
3380717
Compare
| @@ -22,7 +22,7 @@ def configure_settings(**kwargs): | |||
| Parameters | |||
| ---------- | |||
| trainable : bool or None | |||
| Adds a parameter to Nengo Ensembles/Connections/Networks that controls | |||
| Adds a parameter to Nengo Ensembles/Connections that controls | |||
There was a problem hiding this comment.
Is this a change in function that this can't be set on networks, or just a fix to the documentation?
There was a problem hiding this comment.
Just a fix for the documentation (setting it on networks was removed a while ago, I just missed this docstring).
| non-trainable parameters of the network (this includes the internal | ||
| simulation state). | ||
| include_state : bool | ||
| If True (default False) also save the internal simulation state. |
There was a problem hiding this comment.
I would opt for still having a line for include_non_trainable, that basically says it's deprecated and equivalent to include_state.
There was a problem hiding this comment.
Looking at nengo.Simulator.trange and the switch in parameters there, we don't have documentation for both the old and new names, but we do have a "version changed" tag, so I've opted for that instead.
nengo_dl/simulator.py
Outdated
|
|
||
| vars = self.keras_model.weights | ||
| if include_state: | ||
| vars.extend(self.tensor_graph.saved_state.values()) |
There was a problem hiding this comment.
This will modify self.keras_model.weights, right? Is that a problem?
There was a problem hiding this comment.
Yeah good call probably safer to do this on a copy
nengo_dl/tensor_graph.py
Outdated
| "trainable": OrderedDict(), | ||
| "non_trainable": OrderedDict(), | ||
| "state": OrderedDict(), | ||
| } |
There was a problem hiding this comment.
I would make base_arrays an ordered dict just so it's deterministic when we call .items on it (right now, that's just in a logger message, but it's nice to have things consistent IMO).
CHANGES.rst
Outdated
| - Model parameters (e.g., connection weights) that are not trainable (because they've | ||
| been marked non-trainable by user or targeted by an online learning rule) will now | ||
| be treated separately from simulator state. For example, resetting the simulator | ||
| state will not reset those parameters, and the results of any online learning |
There was a problem hiding this comment.
Doesn't calling sim.reset reset online learning, though (as per test_online_learning_reset)? It's just sim.soft_reset with include_trainable=False that doesn't reset.
There was a problem hiding this comment.
Yeah I was intending "resetting the simulator state" to refer specifically to include_trainable=False, but that may not be clear.
|
|
||
| sim.reset() | ||
|
|
||
| assert np.allclose(w0, sim.data[conn].weights) |
There was a problem hiding this comment.
I'm going to group the lines a bit and add some short comments just making it clear what we expect for each case.
| assert np.allclose(sim_load.data[p1], sim_save.data[p0][10:]) | ||
| else: | ||
| assert not np.allclose(sim_load.data[p1], sim_save.data[p0][10:]) | ||
| assert np.allclose(sim_load.data[p1], sim_save.data[p0][:10]) |
There was a problem hiding this comment.
NTS: add a comment about this (took a minute of thinking to figure it out)
|
Done with my review, and pushed a fixup commit making the changes I've suggested above. If it looks good to you @drasmuss, and passes CI, I'll merge. EDIT: Oh, I did have one question above about the changelog, basically wondering if it's slightly unclear about when online learning weights are reset. |
|
Fixups all look good to me, I added a clarification to the changelog, if that looks good to you go ahead and merge! |
drasmuss
force-pushed
the
model-checkpoint
branch
from
abc9c16 to
4d64291
Compare
This makes it easier to re-use saved weights between models, as models with different saved state variables can still use the same saved Keras parameters.
Fixes #133