JSON model import and export #469

alubbock · 2019-10-19T03:33:03Z

Add JSON model import and export capabilities. JSON
is a text-serializable format, which is good for storing
models within other objects, or for purposes like HTTP
APIs.

SimulationResult.save() now saves models in JSON
rather than pickle.

See discussion on #467 which prompted the need for
a durable and serializable model storage format.

Add JSON model import and export capabilities. JSON is a text-serializable format, which is good for storing models within other objects, or for purposes like HTTP APIs.

Replaces pickling the model

codecov-io · 2019-10-19T04:36:35Z

Codecov Report

❗ No coverage uploaded for pull request base (master@cf08a76). Click here to learn what that means.
The diff coverage is 89.4%.

@@            Coverage Diff            @@
##             master     #469   +/-   ##
=========================================
  Coverage          ?   80.08%           
=========================================
  Files             ?      101           
  Lines             ?    10343           
  Branches          ?        0           
=========================================
  Hits              ?     8283           
  Misses            ?     2060           
  Partials          ?        0

Impacted Files	Coverage Δ
pysb/export/__init__.py	`96% <ø> (ø)`
pysb/annotation.py	`93.33% <100%> (ø)`
pysb/tests/test_exporters.py	`90.47% <100%> (ø)`
pysb/simulator/base.py	`88.8% <50%> (ø)`
pysb/importers/json.py	`93.57% <93.57%> (ø)`
pysb/export/json.py	`98.48% <98.48%> (ø)`

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update cf08a76...e4d8e8f. Read the comment docs.

bgyori · 2019-10-21T14:48:53Z

This looks promising! One of the issues though, is that often, one of the main rationales behind pickling a PySB model is to save the result of having run bng.generate_equations on the model (which for certain models can be quite slow). The model can then be quickly reloaded and simulated without having to regenerate the reaction network. This is not immediately possible with the JSON approach, though perhaps could be optionally handled by encoding the equations in JSON as well.

Applies to both JSON export and import. Useful for caching rxn net to avoid network re-generation.

alubbock · 2019-10-22T02:17:29Z

@bgyori No problem, I've added an option for this: JsonExporter(model).export(include_netgen=True)

jmuhlich

This is impressive, and quickly turned around!

jmuhlich · 2019-11-08T21:04:17Z

pysb/simulator/base.py

+    # restore the 'model' weakrefs on all components
+    self.__dict__.update(state)
+    # Set "tags" attribute for older, pickled models
+    self.__dict__.setdefault('tags', ComponentSet())


Is this the only thing we would have to patch to support all potential old pickles that might be out there in saved SimulationResult hdf5 files, or are there other attributes we've added since? We should either fix everything or fix nothing, but I fear fixing everything might get out of hand.

This won't address the use case that brought up #467 anyway because they are doing explicit pickling/unpickling not saving SimulationResults.

This is the only change required to unpickle models or load HDF5 files, at least as far back as PySB v1.6.0 where the HDF5 format was introduced.

For the standalone pickle cases, the following snippet works. Let me know if you think I should remove the underscore prefix and/or document this somehwere explicitly.

from pysb.simulator.base import _patch_model_setstate with open('earm_1_0_pysb1.6.pkl', 'rb') as f: with _patch_model_setstate(): pickle.load(f)

jmuhlich · 2019-11-08T21:04:17Z

pysb/tests/test_exporters.py

+            # Check network generation and single-step integrator
+            if model.name not in ('pysb.examples.tutorial_b',
+                                  'pysb.examples.tutorial_c'):
+                ScipyOdeSimulator(m)


This just initializes the integrator. Based on the comment on line 112, did you mean to actually run it a bit?

The integrator RHS is run for t=0 during initialization, at least when using Cython (line 219 of scipyode.py). But I'll change this to enforce Cython, using compiler='cython' kwarg.

jmuhlich · 2019-11-08T21:04:17Z

pysb/export/json.py

+
+
+class PySBJSONEncoder(json.JSONEncoder):
+    FORMAT = 1


Can you write some documentation, perthaps on this class and/or the decoder, about what we need to think about as Model and the components gain new attributes in order to maintain proper compatibility here or bump the version number?

Also, to be a bit pedantic, is "format" the right name? I would think "version" or even "protocol" (to mimic pickle) would be more apt.

I've switched to protocol, rather than format.

I've added some documentation to cover the basics of the format. The protocol number should get bumped when new features are added which affect model simulation, or prevent it being loaded by the decoder. I don't think it's a bad thing if this happens, so long as the decoder maintains backwards compatibility.

jmuhlich · 2019-11-08T21:04:17Z

pysb/tests/test_exporters.py

+            # Check network generation and single-step integrator
+            if model.name not in ('pysb.examples.tutorial_b',
+                                  'pysb.examples.tutorial_c'):
+                ScipyOdeSimulator(m)


This is a reasonable smoke test to make sure the model isn't broken, but it doesn't demonstrate we got the same model back. We ought to actually check that the model is the same, but that's probably infeasible on every example unless we leverage some other exporter (and hope we don't end up with parallel bugs in both exporters). The first thing that comes to mind is to run each model and its re-imported version through the pysb_flat exporter and compare the strings. Of course the pysb_flat exporter doesn't have a round-trip fidelity test either, does it?

Or maybe create a minimal model that exercises every feature once, and explicitly check all the components before and after re-import? That might be a big job, though.

Indeed - I used PySBFlatExporter to manually check things were working as expected, but symbols within expressions don't appear in a deterministic order. I agree the best solution is probably to create a toy model that uses every feature once and check equality manually.

I ended up using check_model_against_component_list on all of the example models after round-tripping to JSON, which seems to do the job.

This is to avoid e.g. component names getting converted from strings to unicode. This is merely a test artifact, and doesn't seem worth fixing given we're dropping Python 2 support in under 2 months' time.

jmuhlich · 2019-11-13T15:10:41Z

pysb/core.py

        Component.__init__(self, name, _export)
+
+    @property
+    def value(self):


What's this new getter/setter for? I'm guessing round-trip testing exposed a code path where the normalization of .value to a float was getting bypassed?

Yes, in the earm_1_3.py model where .value is used with integers, but these values were getting converted to floats after round-tripped to JSON.

jmuhlich · 2019-11-13T15:10:41Z

pysb/examples/bngwiki_egfr_simple.py

@@ -39,7 +39,7 @@


 Monomer('EGF', ['R'])
-Monomer('EGFR', ['L','CR1','Y1068'], {'Y1068':('U','P')})
+Monomer('EGFR', ['L','CR1','Y1068'], {'Y1068': ['U', 'P']})


Maybe Monomer should just normalize sites and site_states.values() to list, but that can be done outside of this PR.

Agree, we should normalize in core.py.

jmuhlich · 2019-11-13T15:10:41Z

pysb/tests/test_exporters.py

@@ -112,7 +114,12 @@ def check_convert(model, format):
            # Check network generation and single-step integrator


As you pointed out in your previous comment, we actually just need to evaluate the RHS once which is different from stepping the integrator. This comment would be more accurate as something like "Check network generation and force RHS evaluation".

Comment fixed as requested

alubbock added 2 commits Oct 19, 2019

JSON model import and export

7928989

Add JSON model import and export capabilities. JSON is a text-serializable format, which is good for storing models within other objects, or for purposes like HTTP APIs.

SimulationResult.save() - use JSON for model

511e8ec

Replaces pickling the model

alubbock added 3 commits Oct 21, 2019

Fixes for JSON import

52add39

Add option to include rxn net, species in JSON

975ea09

Applies to both JSON export and import. Useful for caching rxn net to avoid network re-generation.

Fix Py2 unicode error

5f7c68e

alubbock mentioned this pull request Oct 22, 2019

Fix unpickling of models saved in PySB <1.10 #468

Closed

Merge branch 'master' into json

d60876a

jmuhlich requested changes Nov 8, 2019

View changes

alubbock added 5 commits Nov 8, 2019

Enforce Cython for testing JSON export/import

ac66c56

Replace "format" with "protocol" in JSON format

7eb4265

Documentation for JSON encoder, decoder

e4d8e8f

Check model components in JSON unit tests

2b3cf14

Don't check JSON round-tripped model on Py2

5282b9d

This is to avoid e.g. component names getting converted from strings to unicode. This is merely a test artifact, and doesn't seem worth fixing given we're dropping Python 2 support in under 2 months' time.

jmuhlich requested changes Nov 13, 2019

View changes

Clarify comment about RHS eval in JSON conv. tests

6c330b4

jmuhlich approved these changes Nov 13, 2019

View changes

alubbock merged commit 54ff9e4 into pysb:master Nov 13, 2019
2 checks passed

alubbock deleted the json branch Nov 13, 2019

alubbock mentioned this pull request Nov 14, 2019

New tags attribute causes error when unpickling older models #467

Closed

pysb / pysb Public

JSON model import and export #469

JSON model import and export #469

alubbock commented Oct 19, 2019

codecov-io commented Oct 19, 2019 •

edited

bgyori commented Oct 21, 2019

alubbock commented Oct 22, 2019

jmuhlich left a comment

jmuhlich Nov 8, 2019

alubbock Nov 8, 2019

jmuhlich Nov 8, 2019

alubbock Nov 8, 2019

jmuhlich Nov 8, 2019

alubbock Nov 8, 2019

jmuhlich Nov 8, 2019

alubbock Nov 8, 2019

alubbock Nov 12, 2019

jmuhlich Nov 13, 2019

alubbock Nov 13, 2019

jmuhlich Nov 13, 2019

alubbock Nov 13, 2019

jmuhlich Nov 13, 2019

alubbock Nov 13, 2019

		@@ -112,7 +114,12 @@ def check_convert(model, format):
		# Check network generation and single-step integrator

pysb / pysb Public

JSON model import and export #469

JSON model import and export #469

Conversation

alubbock commented Oct 19, 2019

codecov-io commented Oct 19, 2019 • edited

Codecov Report

bgyori commented Oct 21, 2019

alubbock commented Oct 22, 2019

jmuhlich left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

codecov-io commented Oct 19, 2019 •

edited