Transformers: ValueError: You have to specify either input_ids or inputs_embeds!

[innat]

Details

I'm quite new to NLP task. However, I was trying to train the T5-large model and set things as follows. But unfortunately, I've got an error.

def build_model(transformer, max_len=512):
    input_word_ids = Input(shape=(max_len,), dtype=tf.int32, name="input_word_ids")
    sequence_output = transformer(input_word_ids)[0]
    cls_token = sequence_output[:, 0, :]
    out = Dense(1, activation='sigmoid')(cls_token)
    model = Model(inputs=input_word_ids, outputs=out)
    return model

model = build_model(transformer_layer, max_len=MAX_LEN)

It thorws

ValueError: in converted code:
ValueError                                Traceback (most recent call last)
<ipython-input-19-8ad6e68cd3f5> in <module>
----> 5     model = build_model(transformer_layer, max_len=MAX_LEN)
      6 
      7 model.summary()

<ipython-input-17-e001ed832ed6> in build_model(transformer, max_len)
     31     """
     32     input_word_ids = Input(shape=(max_len,), dtype=tf.int32, name="input_word_ids")
---> 33     sequence_output = transformer(input_word_ids)[0]
     34     cls_token = sequence_output[:, 0, :]
     35     out = Dense(1, activation='sigmoid')(cls_token)
ValueError: You have to specify either input_ids or inputs_embeds

[patrickvonplaten]

Hi @innat,

T5 is an encoder-decoder model so you will have to provide both input_ids and decoder_input_ids to the model. Maybe taking a look at the T5 docs (especially the "Examples") can help you :-)

[patrickvonplaten]

Just noticed that the Examples docstring for TF T5 was wrong. Is fixed with #3636 .

[innat]

@patrickvonplaten
hello, sorry to bother you. Would you please justify the following piece of code:

Imports

from transformers import TFAutoModel, AutoTokenizer

# First load the real tokenizer
tokenizer = AutoTokenizer.from_pretrained('t5-small')
transformer_layer = TFAutoModel.from_pretrained('t5-small')

Define Encoder

def encode(texts, tokenizer, maxlen=512):
    enc_di = tokenizer.batch_encode_plus(
        texts, 
        return_attention_masks=False, 
        return_token_type_ids=False,
        pad_to_max_length=True,
        max_length=maxlen
    )
    return np.array(enc_di['input_ids'])

# tokenized
x_train = encode('text', tokenizer, maxlen=200)
y_train

Define Model and Call

def build_mod(transformer, max_len=512):
    input_word_ids = Input(shape=(max_len,), dtype=tf.int32, name="input_word_ids")
    sequence_output = transformer(input_word_ids)[0]
    cls_token = sequence_output[:, 0, :]
    out = Dense(1, activation='sigmoid')(cls_token)

    model = Model(inputs=input_word_ids, outputs=out)
    model.compile(Adam(lr=1e-5), loss='binary_crossentropy', metrics=['accuracy'])

    return model

# calling
model = build_model(transformer_layer, max_len=200)

Now, according to the docstring, should I do,

outputs = model(input_ids=x_train, decoder_input_ids=x_train)[0]

?

[patrickvonplaten]

I'm not 100% sure what you want to do here exactly. T5 is always trained in a text-to-text format. We have a section here on how to train T5: https://huggingface.co/transformers/model_doc/t5.html#training

Otherwise I'd recommend taking a look at the official paper.

[enzoampil]

@patrickvonplaten Thanks for this. I encountered the same issue and this resolved it!

I'm wondering if it makes sense to make the error message capture the requirement of having both input_ids and decoder_input_ids since this is an encoder-decoder model? This may make the fix clearer for users of encoder decoder models in the future.

I.e., for encoded-decoder models, switch the error message from:

ValueError: You have to specify either input_ids or inputs_embeds

to:

ValueError: You have to specify either (input_ids and decoder_input_ids) or inputs_embeds

I can sent this as a PR as well if you think it makes sense!

[patrickvonplaten]

Hi @enzoampil,

A PR for a cleaner Error message would be nice if you feel like it :-). It would be good if the error message could change between ValueError: You have to specify either input_ids or inputs_embeds if self.is_decoder == False and ValueError: You have to specify either decoder_input_ids or decoder_inputs_embeds if self.is_decoder == True. So adding a simple if statement to the error message is definitely a good idea!

[enzoampil]

Got it will do. Thanks for the pointers! 😄

[ratthachat]

Hi, I also got the same error when training seq2seq on tf.keras and I could not follow the example you provide on https://huggingface.co/transformers/model_doc/t5.html#training (this example is for pytorch I think)

I create x_encoder asinput_ids and x_decoder_in for decoder_input_ids

model = TFT5Model.from_pretrained('t5-base')
model.compile('adam',loss='sparse_binary_crossentropy')

So when I want to train the model I simply do
model.fit({'input_ids': x_encoder, 'decoder_input_ids': x_decoder_in})

where I clearly provide input_ids , but still got this error message :
ValueError: You have to specify either input_ids or inputs_embeds

Note that changing input from dict to list got the same error. Changing model from TFT5Model to TFT5ForConditionalGeneration got the same error. Changing loss to BCE got the same error.

Moreover, changing input to only one array
model.fit({'input_ids': x_encoder})
is also error :

ValueError: No data provided for "decoder_input_ids". Need data for each key in: ['decoder_input_ids', 'input_ids']

[ratthachat]

In class TFT5Model(TFT5PreTrainedModel):

I found this line (899-900):
```

    # retrieve arguments
    input_ids = kwargs.get("inputs", None)

    ```

Shouldn't it be kwargs.get("input_ids", None) ??

[patrickvonplaten]

@ratthachat - thanks for you message!
We definitely need to provide more TF examples for the T5 Model. I want to tackle this problem in ~2 weeks.

In TF we use the naming convention inputs, so the you should change to model.fit({"inputs": x_encoder}) . I very much agree that the error message is quite misleading and correct it in this PR: #4401.

[ratthachat]

Thanks for your consideration, Patrick!

[ratthachat]

@patrickvonplaten Sorry to tag you in this old thread, but is there any official T5 TF example (as you mentioned in the last thread)?

[patrickvonplaten]

@ratthachat - no worries, we should definitely add more TF T5 examples and we still don't have a good TF T5 notebook.
I am moving the discussion to the forum and if no one answers I will spent some time coping a T5 PT notebook to TF.

[HarrisDePerceptron]

Hi @patrickvonplaten i wanted to fine tune using T5 using TF 2.0 but its soo confusing at each end as compared to pytorch which is really well documented all current examples (community + offcial) are for pytorch. is the work for TFT5 notebook underway?

[patrickvonplaten]

Okey, seems like no-one has a complete TF T5 notebook. I will start working on it this week: https://discuss.huggingface.co/t/how-to-train-t5-with-tensorflow/641/6

Should be done by next week sometime :-)

[prashant-kikani]

Hi @patrickvonplaten
Please help me with this error.

I'm doing inference with a T5-base model which I finetuned on GLUE tasks.

It's giving error like
ValueError: You have to specify either decoder_input_ids or decoder_inputs_embeds

While doing inference, we just need to provide input_ids for the encoder right?
Why do we need decoder_input_ids?

And as it's inference, my labels will also be None.
So, this part will not execute.
decoder_input_ids = self._shift_right(labels)

Waiting for your reply.
Thank you.

[HarrisDePerceptron]

@prashant-kikani it is indeed a strange behavior. have you tried passing input_ids to decoder_input_ids like:

input_ids = tokenizer(..., return_tensor='tf')   # replace pt for pytorch
outputs= model(input_ids=input_ids, decoder_input_ids=input_ids)

assert len(outputs)==3, 'must return 3 tensors when inferencing'

[prashant-kikani]

Hi @HarrisDePerceptron
We can do it & it's giving some output also. But it's not the right thing to do.

You see, T5 which Transformer itself, is a text to text model.
So, it can do inference in linear time by matrix multiplication when label is available.

But, when label is not available, we need to go sequentially by doing forward pass in decoder for each word till </s> doesn't come.
We need to concatenate last output of decoder with new input if decoder each time.

What do you think?

[ratthachat]

@prashant-kikani @HarrisDePerceptron

For decoder_input_ids , we just need to put a single BOS token so that the decoder will know that this is the beginning of the output sentence. (Even in GLUE task, T5 still looks at every output label as a complete sentence )

We can see a concrete example by looking at the function
prepare_inputs_for_generation which is called by model.generate
(generate function is here : https://github.com/huggingface/transformers/blob/master/src/transformers/generation_tf_utils.py )

See line 298 in the above link :

if self.config.is_encoder_decoder:
            if decoder_start_token_id is None:
                decoder_start_token_id = bos_token_id

and line 331:

# create empty decoder_input_ids
            input_ids = (
                tf.ones(
                    (effective_batch_size * num_beams, 1),
                    dtype=tf.int32,
                )
                * decoder_start_token_id
            )

and see T5's prepare_inputs_for_generation which change the above input_ids into decoder_input_ids implementation at :
https://github.com/huggingface/transformers/blob/08f534d2da47875a4b7eb1c125cfa7f0f3b79642/src/transformers/modeling_tf_t5.py#L1367