replace the weights node name with scaled weights node

[pyu10055]

Fixes #2055

When the batchnorm ops share the same weights, the scaled weights calculated by batchnorm folding will generate duplicated nodes.
This fix uses the conv2d op node name as the base to avoid the duplication.

To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.

---

This change is 

[nsthorat]

Reviewed 1 of 3 files at r1.
Reviewable status: 0 of 1 approvals obtained (waiting on @dsmilkov, @nsthorat, and @pyu10055)

---

tfjs-converter/python/tensorflowjs/converters/fold_batch_norms_test.py, line 144 at r1 (raw file):

        self.assertNotEqual("FusedBatchNorm", node.op)

  def testFoldFusedBatchNormsV3(self):

this seems okay for now but weren't we talking about having some testing infra that lets you define a graph, run the converter, and then execute it (maybe in node) to make sure these things are no-ops?

for now this is cool, can you just make sure this gives identical results for a real model before and after this change?

[pyu10055]

Reviewable status: 0 of 1 approvals obtained (waiting on @dsmilkov and @nsthorat)

---

tfjs-converter/python/tensorflowjs/converters/fold_batch_norms_test.py, line 144 at r1 (raw file):

Previously, nsthorat (Nikhil Thorat) wrote…

this seems okay for now but weren't we talking about having some testing infra that lets you define a graph, run the converter, and then execute it (maybe in node) to make sure these things are no-ops?

for now this is cool, can you just make sure this gives identical results for a real model before and after this change?

we have integration tests that verifies the model inference results in the tfjs/integration folder, it requires a new pip released first.