maintain single id tensor for each tensor list and tensor array

[pyu10055]

This will reduce tensor creation and possible tensor data read.

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

---

This change is 

[lina128]

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

---

tfjs-converter/src/executor/tensor_list.ts, line 104 at r1 (raw file):

    assertShapesMatchAllowUndefinedSize(
        elementShape, this.elementShape, 'TensorList shape mismatch: ');
    // return tidy(() => {

Remove.

---

tfjs-converter/src/executor/tensor_list.ts, line 129 at r1 (raw file):

    assertShapesMatchAllowUndefinedSize(
        tensor.shape, elementShape, 'TensorList shape mismatch: ');
    return tensor.reshape(elementShape);

Can we simply return tensor?

---

tfjs-converter/src/executor/tensor_list.ts, line 245 at r1 (raw file):

    }

    return tidy(() => {

Can we also remove the reshape and tidy here?

---

tfjs-converter/src/executor/tensor_list.ts, line 269 at r1 (raw file):

    }

    return tidy(() => {

Can we also remove the tidy and reshape?

[pyu10055]

Reviewable status: 0 of 1 approvals obtained (waiting on @lina128)

---

tfjs-converter/src/executor/tensor_list.ts, line 104 at r1 (raw file):

Previously, lina128 (Na Li) wrote…

Remove.

I have reverted the change, I think we still need to do the reshape, since the elementShape is an input, we need to honor that.

---

tfjs-converter/src/executor/tensor_list.ts, line 129 at r1 (raw file):

Previously, lina128 (Na Li) wrote…

Can we simply return tensor?

same reason as above.

---

tfjs-converter/src/executor/tensor_list.ts, line 245 at r1 (raw file):

Previously, lina128 (Na Li) wrote…

Can we also remove the reshape and tidy here?

same reason as above.

---

tfjs-converter/src/executor/tensor_list.ts, line 269 at r1 (raw file):

Previously, lina128 (Na Li) wrote…

Can we also remove the tidy and reshape?

same.

[lina128]

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

---

tfjs-converter/src/executor/tensor_list.ts, line 104 at r1 (raw file):

Previously, pyu10055 (Ping Yu) wrote…

I have reverted the change, I think we still need to do the reshape, since the elementShape is an input, we need to honor that.

But you have asserted elementShape === this.elementShape. This is basically saying I only allow the elementShape to be the tensor's original elementShape. So below logic is correct, but is unnecessary?

---

tfjs-converter/src/executor/tensor_list.ts, line 53 at r2 (raw file):

   */
  constructor(
      readonly tensors: Tensor[], readonly elementShape: number[],

Should we check whether the tensors truly have the elementShape and elementDtype?

[lina128]

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

---

tfjs-converter/src/executor/tensor_list.ts, line 245 at r1 (raw file):

Previously, pyu10055 (Ping Yu) wrote…

same reason as above.

Either way, should we keep consistent with tensor_array.ts? https://github.com/tensorflow/tfjs/blob/master/tfjs-converter/src/executor/tensor_array.ts#L205

[pyu10055]

Reviewable status: 0 of 1 approvals obtained (waiting on @lina128)

---

tfjs-converter/src/executor/tensor_list.ts, line 104 at r1 (raw file):

Previously, lina128 (Na Li) wrote…

But you have asserted elementShape === this.elementShape. This is basically saying I only allow the elementShape to be the tensor's original elementShape. So below logic is correct, but is unnecessary?

the check is compatible, they don't have to be exactly the same.

---

tfjs-converter/src/executor/tensor_list.ts, line 245 at r1 (raw file):

Previously, lina128 (Na Li) wrote…

Either way, should we keep consistent with tensor_array.ts? https://github.com/tensorflow/tfjs/blob/master/tfjs-converter/src/executor/tensor_array.ts#L205

TensorArray is slightly different, since there is no target tensor shape for any of these ops.

---

tfjs-converter/src/executor/tensor_list.ts, line 53 at r2 (raw file):

Previously, lina128 (Na Li) wrote…

Should we check whether the tensors truly have the elementShape and elementDtype?

good point.

[lina128]

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @lina128 and @pyu10055)

---

tfjs-converter/src/executor/tensor_list.ts, line 104 at r1 (raw file):

Previously, pyu10055 (Ping Yu) wrote…

the check is compatible, they don't have to be exactly the same.

Ah, I see. So the original size may have -1, the passed in elementShape gives the actual shape. Thanks for the explanation.