Add passthrough IOHandlers to save and load models to and from memory

[davidsoergel]

Add passthrough IOHandlers that allow users to work with the tf.io.ModelArtifacts in-memory serialization format.

Partly addresses tensorflow/tfjs#495

FEATURE

---

This change is 

[caisq]

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

---

src/io/passthrough.ts, line 28 at r1 (raw file):

{}

Do you want to make this {}|ArrayBuffer, so that it is general enough to include FrozenModels?

---

src/io/passthrough.ts, line 28 at r1 (raw file):

 modelTopology

The ModelArtifacts interface actually allows modelTopology to be missing, in which case the artifacts are weights-only. I don't think it'll be too difficult to support that here.

---

src/io/passthrough.ts, line 33 at r1 (raw file):

return new Promise<ModelArtifacts>((resolve, reject) => {

Is new Promise ... necessary here? Can you simply do return {modelTopology: this.modelTopology} etc.?

---

src/io/passthrough.ts, line 34 at r1 (raw file):

this.weightData === undefined

Should this handle the case of this.weightData === null as well?

---

src/io/passthrough.ts, line 83 at r1 (raw file):

Creates an IOHandler that passes saved model artifacts to a callback.

Can you add a code example?

---

src/io/passthrough.ts, line 88 at r1 (raw file):

export function withHandler(
    saveHandler: (artifacts: ModelArtifacts) => SaveResult): IOHandler {

I think it would be better to call it "withSaveHandler" so that it is clearer whether this is for saving or loading.

---

src/io/passthrough_test.ts, line 90 at r1 (raw file):

555

Is it possible to not hard-code these numbers here to prevent possible future maintenance cost?

[davidsoergel]

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

---

src/io/passthrough.ts, line 28 at r1 (raw file):

Previously, caisq (Shanqing Cai) wrote…

{}

Do you want to make this {}|ArrayBuffer, so that it is general enough to include FrozenModels?

You bet, good catch

---

src/io/passthrough.ts, line 28 at r1 (raw file):

Previously, caisq (Shanqing Cai) wrote…

 modelTopology

The ModelArtifacts interface actually allows modelTopology to be missing, in which case the artifacts are weights-only. I don't think it'll be too difficult to support that here.

Done.

---

src/io/passthrough.ts, line 33 at r1 (raw file):

Previously, caisq (Shanqing Cai) wrote…

return new Promise<ModelArtifacts>((resolve, reject) => {

Is new Promise ... necessary here? Can you simply do return {modelTopology: this.modelTopology} etc.?

Oops, I overlooked the paste artifact, thanks

---

src/io/passthrough.ts, line 34 at r1 (raw file):

Previously, caisq (Shanqing Cai) wrote…

this.weightData === undefined

Should this handle the case of this.weightData === null as well?

Yep. What a ridiculous language :) I still get confused about {null, undefined} x {==, ===}. Anyway I think '!= null' covers both cases.

---

src/io/passthrough.ts, line 83 at r1 (raw file):

Previously, caisq (Shanqing Cai) wrote…

Creates an IOHandler that passes saved model artifacts to a callback.

Can you add a code example?

It seemed to me that the boilerplate for a minimal example overwhelmed the point here. Is this OK (eliding the details via '...')?

---

src/io/passthrough.ts, line 88 at r1 (raw file):

Previously, caisq (Shanqing Cai) wrote…

export function withHandler(
    saveHandler: (artifacts: ModelArtifacts) => SaveResult): IOHandler {

I think it would be better to call it "withSaveHandler" so that it is clearer whether this is for saving or loading.

Done.

---

src/io/passthrough_test.ts, line 90 at r1 (raw file):

Previously, caisq (Shanqing Cai) wrote…

555

Is it possible to not hard-code these numbers here to prevent possible future maintenance cost?

Done.

[caisq]

Great. Thanks!

[caisq]

Great

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @caisq, @nsthorat, and @dsmilkov)

[nsthorat]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @caisq, @nsthorat, and @dsmilkov)