move frozen model loading to use io handler

[pyu10055]

this would allow node.js user to load frozen model from file or http url.

fix tensorflow/tfjs#519

---

This change is 

[bileschi]

cool!

[caisq]

Reviewable status: 0 of 1 approvals obtained (waiting on @pyu10055 and @caisq)

---

src/executor/frozen_model.ts, line 35 at r1 (raw file):

  private handler: tfc.io.IOHandler;

Just a comment, no immediate action required: I found it perhaps a little strange that FrozenModel needs to be tied to an IOHandler. This means that a FrozenModel is and will always be loaded from elsewhere, i.e., not built in tfjs per se. Is that the case? If not, maybe it'll be good to refactor the IOHandler out of the class.

---

src/executor/frozen_model.ts, line 105 at r1 (raw file):

new Uint8Array(artifacts.modelTopology as ArrayBuffer)

Does this make a copy of the buffer or reuse it? If the former, it'll be nice to find a way to avoid it for performance reasons.

[pyu10055]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @pyu10055)

---

src/executor/frozen_model.ts, line 35 at r1 (raw file):

Previously, caisq (Shanqing Cai) wrote…

  private handler: tfc.io.IOHandler;

Just a comment, no immediate action required: I found it perhaps a little strange that FrozenModel needs to be tied to an IOHandler. This means that a FrozenModel is and will always be loaded from elsewhere, i.e., not built in tfjs per se. Is that the case? If not, maybe it'll be good to refactor the IOHandler out of the class.

I don't see frozen model will be generated directly by tfjs in near future. We can refactor it out, when there is such a need.

---

src/executor/frozen_model.ts, line 105 at r1 (raw file):

Previously, caisq (Shanqing Cai) wrote…

new Uint8Array(artifacts.modelTopology as ArrayBuffer)

Does this make a copy of the buffer or reuse it? If the former, it'll be nice to find a way to avoid it for performance reasons.

It should not be any copy, they should share the underlying buffer.