[tfjs-node] Fix bug in node.tensorBoard() callback re initialEpoch

[caisq]

Fixes #3705

Previously, the tensorBoard callback in tfjs-node does not honor
the initialEpoch arg passed to the fit() call that uses the
callback. It always incorrectly starts from 0.
This CL fixes this bug by using the epoch arg passed to onEpochEnd()
instead of an epochsSeen counter maintained by the callback object
itself.

---

This change is 

[pyu10055]

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

---

tfjs-node/src/callbacks.ts, line 215 at r1 (raw file):

        this.batchesSeen++;
        if (this.args.updateFreq !== 'epoch') {
          this.logMetrics(logs, 'batch_', this.batchesSeen);

should the batch number be updated the same way as epoch?

[caisq]

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

---

tfjs-node/src/callbacks.ts, line 215 at r1 (raw file):

Previously, pyu10055 (Ping Yu) wrote…

should the batch number be updated the same way as epoch?

Good question. The behavior of tf.keras in Python is that batch numbers logged by TensorBoard callback with update_freq='batch' doesn't reflect the initial_epoch arg, even though the epoch numbers do. For instance, I tested the following code:

import numpy as np
import tensorflow as tf

model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(units=1, input_shape=(4,)))
model.compile(loss="mse", optimizer="sgd")

xs = np.ones([8000, 4])
ys = np.zeros([8000, 1])

model.fit(xs, ys, epochs=3)

callback = tf.keras.callbacks.TensorBoard(
    "/tmp/initial_epochs_logdir", update_freq="batch")
model.fit(xs,
          ys,
          batch_size=40,
          epochs=6,
          initial_epoch=3,
          callbacks=[callback])

Here the tensorboard scalar log "batch_loss" starts at step 1, instead of a larger number that reflects the batches that have already happened in the previous (first) model.fit() call. Therefore the behavior in tfjs-node code here is correct: it keeps track of the batch number by itself.

[pyu10055]

Reviewed 2 of 2 files at r1.
Reviewable status: complete! 1 of 1 approvals obtained (waiting on @lina128 and @pyu10055)

[caisq]

@pyu10055 @lina128 Any idea why the CI of tfjs-node is failing? It seems to be unrelated to this PR.

[pyu10055]

The error seems to related to the lint error on the test file:

ERROR: (no-any) /workspace/tfjs-node/src/tensorboard_test.ts[309, 18]: Type declaration of 'any' loses type-safety. Consider replacing it with a more precise type.
ERROR: (no-any) /workspace/tfjs-node/src/tensorboard_test.ts[310, 18]: Type declaration of 'any' loses type-safety. Consider replacing it with a more precise type.
ERROR: (no-any) /workspace/tfjs-node/src/tensorboard_test.ts[311, 53]: Type declaration of 'any' loses type-safety. Consider replacing it with a more precise type.
ERROR: (no-any) /workspace/tfjs-node/src/tensorboard_test.ts[312, 51]: Type declaration of 'any' loses type-safety. Consider replacing it with a more precise type.

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

[lina128]

Hi @caisq , it complains the newly added test: maybe skip lint?
ERROR: (no-any) /workspace/tfjs-node/src/tensorboard_test.ts[309, 18]: Type declaration of 'any' loses type-safety. Consider replacing it with a more precise type.
ERROR: (no-any) /workspace/tfjs-node/src/tensorboard_test.ts[310, 18]: Type declaration of 'any' loses type-safety. Consider replacing it with a more precise type.
ERROR: (no-any) /workspace/tfjs-node/src/tensorboard_test.ts[311, 53]: Type declaration of 'any' loses type-safety. Consider replacing it with a more precise type.
ERROR: (no-any) /workspace/tfjs-node/src/tensorboard_test.ts[312, 51]: Type declaration of 'any' loses type-safety. Consider replacing it with a more precise type.

[caisq]

@lina128 @pyu10055 Ah - good find. Sorry I didn't notice that.