Fix release script.

[lina128]

This PR fixes issues found during release. Note, the issues were found right at the spot and were fixed right away. The release versions were correctly published, so the fixes in this PR were tested in the release of 1.7.1.

This PR has below changes:

1. Update dependency list, because Layers is recently added as a dependency for Node: https://github.com/tensorflow/tfjs/blob/master/tfjs-node/package.json#L50

2. node-gpu needs an additional step to build.

3. wasm needs an additional step to build.

4. The publish script has to use this command to run YARN_REGISTRY="https://registry.npmjs.org/" yarn publish-npm. The previous try that adds npm registry information in the last step works on my macbook but not on workstation. On workstation, it is still using yarn registry. Above command works on both macbook and workstation.

---

This change is 

[kangyizhang]

Thank you Na for putting this together which makes the release process much more smooth!

Reviewable status: 0 of 1 approvals obtained (waiting on @dsmilkov, @kangyizhang, @lina128, @nsthorat, and @pyu10055)

---

scripts/release-util.ts, line 51 at r1 (raw file):

export const NODE_PHASE: Phase = {
  packages: ['tfjs-node', 'tfjs-node-gpu'],
  deps: ['tfjs', 'tfjs-core', 'tfjs-layers'],

I saw that tfjs-layers is added as devDependency in https://github.com/tensorflow/tfjs/pull/2854/files#diff-48e8607fcaefdc233be729e07c18f766R48. Can you provide some context why we did this? tfjs-node depends on the tfjs union package, so it should already have tfjs-layers and tfjs-core installed as dependency.

[lina128]

Thank you Na for putting this together which makes the release process much more smooth!

Reviewable status: 0 of 1 approvals obtained (waiting on @dsmilkov, @kangyizhang, @lina128, @nsthorat, and @pyu10055)

scripts/release-util.ts, line 51 at r1 (raw file):

export const NODE_PHASE: Phase = {
  packages: ['tfjs-node', 'tfjs-node-gpu'],
  deps: ['tfjs', 'tfjs-core', 'tfjs-layers'],

I saw that tfjs-layers is added as devDependency in https://github.com/tensorflow/tfjs/pull/2854/files#diff-48e8607fcaefdc233be729e07c18f766R48. Can you provide some context why we did this? tfjs-node depends on the tfjs union package, so it should already have tfjs-layers and tfjs-core installed as dependency.

Hi Kangyi, thanks for the review. We added tfjs-core and tfjs-layers because they were directly used in the code instead of through tfjs, see https://github.com/tensorflow/tfjs/blob/master/tfjs-node/src/io/file_system_test.ts#L19-L20. If they are directly used, they need to be in the dependency list, see this discussion: #2762 and #2761

[dsmilkov]

Reviewable status: 0 of 1 approvals obtained (waiting on @dsmilkov, @kangyizhang, @nsthorat, and @pyu10055)

---

scripts/release-util.ts, line 51 at r1 (raw file):

Previously, lina128 (Na Li) wrote…

Hi Kangyi, thanks for the review. We added tfjs-core and tfjs-layers because they were directly used in the code instead of through tfjs, see https://github.com/tensorflow/tfjs/blob/master/tfjs-node/src/io/file_system_test.ts#L19-L20. If they are directly used, they need to be in the dependency list, see this discussion: #2762 and #2761

Na, let's add what you just said in a comment above the deps line. Say something like: "Node has unit tests that use tf.layers to test serialization of models in node."

Something to think about for later: Ideally that unit test should live in layers, because it depends on platform specific stuff (node writing files to disk), nothing fundamental to the node backend.

[dsmilkov]

LGTM, with 1 comment to follow up.

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

[lina128]

Node has unit tests that use tf.layers to test serialization of models in node.

Done. Thank you for review, Daniel!

[kangyizhang]

Reviewable status: complete! 2 of 1 approvals obtained (waiting on @dsmilkov, @kangyizhang, @lina128, @nsthorat, and @pyu10055)

---

scripts/release-util.ts, line 51 at r1 (raw file):

Previously, dsmilkov (Daniel Smilkov) wrote…

Na, let's add what you just said in a comment above the deps line. Say something like: "Node has unit tests that use tf.layers to test serialization of models in node."

Something to think about for later: Ideally that unit test should live in layers, because it depends on platform specific stuff (node writing files to disk), nothing fundamental to the node backend.

Thank you Na for the explanation!