Fix Cannot convert [array()] to EagerTensor of dtype int64

[pavi-ninjaac]

While running the model.prepare_tf_dataset() method,
it raises the error below:

TypeError: Cannot convert [array([322.,   1.])] to EagerTensor of dtype int64

This happens, in "DataCollatorForSeq2Seq" function when we are try
to convert the labels to tensors. While converting the labels to tensors,
the labels can be in the format of list of list or list of ndarrays.
There is no problem converting the list of list lables. There is a problem
when the list of ndarrays are float values(like below).

[array([322.,   1.])]

so the exception raises while trying to convert this label to tensors using
below code.

batch["labels"] = tf.constant(batch["labels"], dtype=tf.int64)

The labels are always integer values, so this got converted to float
values in the label padding operation below.

batch["labels"] = [
                    call(label)
                    if padding_side == "right"
                    else np.concatenate([[self.label_pad_token_id] * (max_label_length - len(label)), label])
                    for label in labels
                    ]

Here we have 2 cases:
1 - Concatenating an array having integer padding token value with labels.
2 - Concatenating an empty array with labels.

---

case 1: Concatenating an array having integer padding token value with labels.
WORKS EXPECTED:

label = np.array([233, 1])
max_label_length = 4
label_pad_token_id = -100
np.concatenate([[label_pad_token_id] * (max_label_length - len(label)), label])
o/p:
array([-100, -100,  233,    1])

---

Case 2: Concatenating an empty array with labels.
GIVES THE ISSUE:
This scenorio can happen when the label has the maximum label length -- No padding needed.

label = np.array([233, 1])
max_label_length = 2
label_pad_token_id = -100
np.concatenate([[label_pad_token_id] * (max_label_length - len(label)), label])
o/p:
array([233.,   1.])

---

Solution:

We need to concatenate a ndarray of dtype int with labels.

AFTER FIX:

case 1:

label = np.array([233, 1])
max_label_length = 4
label_pad_token_id = -100
np.concatenate([np.array([label_pad_token_id] * (max_label_length - len(label)), dtype=np.int64),label])

o/p:
array([-100, -100,  233,    1])

case 2:

label = np.array([233, 1])
max_label_length = 2
label_pad_token_id = -100
np.concatenate([np.array([label_pad_token_id] * (max_label_length - len(label)), dtype=np.int64),label])

o/p:
array([233,   1])

What does this PR do?

Fixes # (issue)
31112

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[LysandreJik]

cc @muellerzr

[pavi-ninjaac]

found a better solution, let me commit that changes.

[pavi-ninjaac]

@muellerzr i am done, please check at your own time.

[muellerzr]

Thanks for doing this! The solution makes perfect sense here imo. Can you go and rebase off main so tests can pass on the ci?

cc @amyeroberts :)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[amyeroberts]

LGTM - thanks for digging into this and fixing!