QA Ungraded Lab

Hi. I´m running the Transformer Network App: Question Answering ungraded lab in Vs Code. In this line

train_features = {x: train_ds.to_tensor(default_value=0, shape=[None,
tokenizer.model_max_length]) for x in [‘input_ids’, ‘attention_mask’]}

This error message comes up: AttributeError: ‘tensorflow.python.framework.ops.EagerTensor’ object has no attribute ‘to_tensor’

Probably a deprecated code?

In Jupyter runs flawless.


It’s very tricky to set up your own local environment, because you have to match all of the versions of tools that Coursera provides.

This code is slightly odd.

This used Hugging Face data set. And, if you look at a line above, you see

train_ds.set_format(type='tf', columns=columns_to_return)

With specifying “type='tf'”, both 'input_ids’ and 'attention_mask’ are already "Tensor". So, “to_tensor” does not make sense at all. This code may work on a particular version or combination of package(s), but, may not work in majority of environments.

Here is an action.

This code assumes ‘Ragged’ tensor, which includes a variety of records in different length. But, if you look at ‘input_ids’ and ‘attention_mask’, both are fixed length as below. (You can also see this is already a tf.Tensor.)


<tf.Tensor: shape=(1000, 26), dtype=int64, numpy=
array([[ 101, 1996, 2436, ..., 3829, 1029,  102],
       [ 101, 1996, 3829, ..., 1997, 1029,  102],
       [ 101, 1996, 3871, ..., 3871, 1029,  102],
       [ 101, 1996, 6797, ..., 5010, 1029,  102],
       [ 101, 1996, 6797, ..., 5723, 1029,  102],
       [ 101, 1996, 3829, ..., 1997, 1029,  102]])>

So, what we need to do is a zero-padding so that the length for each row is equal to “tokenizer.model_max_length”.

#train_features = {x: train_ds[x].to_tensor(default_value=0, shape=[None, tokenizer.model_max_length]) for x in ['input_ids', 'attention_mask']}
padding = tf.constant([[0,0],[0,tokenizer.model_max_length-train_ds['input_ids'].shape[1]]])
train_features = {x: tf.pad(train_ds[x], padding) for x in ['input_ids', 'attention_mask']}

Then, we can create the same “train_features” as we create on Coursera platform.

Finally, I could get the same result like this.

Then, the next question is, do we really need a padding ?
Usually, Transformer does inside it. So, let me try to simply remove “to_tensor” and zero padding like,

train_features = {x: train_ds[x] for x in ['input_ids', 'attention_mask']}

Then, I actually got the very similar result.

Of course, a quick testing to get “Garden” worked.

Anyway, I think the first implementation to add zero padding may be the first try. 2nd is an optional.

Hope this helps.

1 Like

You spot on ! it worked! thanks !