fix network precision under specific situation #1391
Conversation
Codecov Report
@@ Coverage Diff @@
## devel #1391 +/- ##
==========================================
+ Coverage 75.53% 75.63% +0.09%
==========================================
Files 91 92 +1
Lines 7506 7531 +25
==========================================
+ Hits 5670 5696 +26
+ Misses 1836 1835 -1
Continue to review full report at Codecov.
|
njzjz
changed the title
| @@ -736,7 +736,7 @@ def _filter_lower( | |||
| if (not self.uniform_seed) and (self.seed is not None): self.seed += self.seed_shift | |||
| else: | |||
| # we can safely return the final xyz_scatter filled with zero directly | |||
| return tf.cast(tf.fill((natom, 4, outputs_size[-1]), 0.), GLOBAL_TF_FLOAT_PRECISION) | |||
| return tf.cast(tf.fill((natom, 4, outputs_size[-1]), 0.), self.filter_precision) | |||
There was a problem hiding this comment.
We cast the result back to global precision before return
There was a problem hiding this comment.
I don't see any cast in line 722 or line 747
There was a problem hiding this comment.
I don't see any
castin line 722 or line 747
747 should be cast back...
Why do we need to cast at L722?
There was a problem hiding this comment.
Instead of _filter_lower, I think we should cast back at line 839 before _filter returns.
There was a problem hiding this comment.
Why not writing a decorator for doing that? It casts the inputs of _filter to filter_precision and casts back to global precision when _filter returns
| """A decorator that casts and casts back the input | ||
| and output tensor of a method. |
There was a problem hiding this comment.
Could you please write more on the logic behind cast_precision?
- It casts i tensor from
GLOBAL_TF_FLOAT_PRECISIONto precision defined by propertyprecision. - It casts o tensor from
precisiontoGLOBAL_TF_FLOAT_PRECISION. - It checks the i/o list and only cast when an i or o is tensor and the tensor matches
GLOBAL_TF_FLOAT_PRECISIONorprecision, respectively.
| @@ -392,15 +392,15 @@ def _pass_filter(self, | |||
| [ 0, start_index* self.ndescrpt], | |||
| [-1, natoms[2+type_i]* self.ndescrpt] ) | |||
| inputs_i = tf.reshape(inputs_i, [-1, self.ndescrpt]) | |||
| layer = self._filter_r(tf.cast(inputs_i, self.filter_precision), type_i, name='filter_type_'+str(type_i)+suffix, natoms=natoms, reuse=reuse, trainable = trainable, activation_fn = self.filter_activation_fn) | |||
| layer = self._filter_r(self.filter_precision, type_i, name='filter_type_'+str(type_i)+suffix, natoms=natoms, reuse=reuse, trainable = trainable, activation_fn = self.filter_activation_fn) | |||
There was a problem hiding this comment.
The first variable should be "inputs_i" instead of “self.filter_precision”. It couldn't pass UT but it was merged into devel.
There was a problem hiding this comment.
In fact, this line is not covered by UT.
There was a problem hiding this comment.
Yes, it's covered by source/tests/test_model_compression_se_r.py, a UT just added in PR #1361.
* Finish model compression for se_r descriptor! * Improve error type. * Update compress.md * Improve error type. * Improve error type. * Improve exception handling and unittest. * Add gtest for tabulate_fusion se_r. * Fix variable mistake from #1391 Co-authored-by: huangliang <huangla@dp.tech>
No description provided.