Wav2Vec2: ValueError: Unable to create tensor, you should probably activate padding with 'padding=True' to have batched tensors with the same length

opus111 · May 23, 2022, 11:32pm

I am trying to fine tune facebook/wav2vec2-large-xlsr-53 models but am running into the above error.

I have read the various fixes and related bugs e.g.

github.com/huggingface/transformers

ValueError When Padding Long Sequences To Small Max Length (Pytorch and Tensorflow)

opened 04:11PM - 17 Jan 22 UTC

closed 05:41PM - 17 Jan 22 UTC

darien-schettler

## Environment info  - `transformers` version: 4.15.0 - Platform: Linux-5.4.144+-x86_64-with-Ubuntu-18.04-bionic - Python version: 3.7.12 - PyTorch version (GPU?): 1.10.0+cu111 (False) - Tensorflow version (GPU?): 2.7.0 (False) - Flax version (CPU?/GPU?/TPU?): not installed (NA) - Jax version: not installed - JaxLib version: not installed - Using GPU in script?: NO - Using distributed or parallel set-up in script?: NO ### Who can help - @Rocketknight1 - @SaulLu - @sgugger ## Information I am following along with the HuggingFace Transformers course. At the [end of Chapter 2](https://huggingface.co/course/chapter2/6?fw=tf), there is information concerning how to use **`AutoTokenizer`** to perform certain tasks (tokenized single/multiple sequences, pad, truncate, convert to other libraries). The problem arises when using: * [the official example colab](https://colab.research.google.com/github/huggingface/notebooks/blob/master/course/chapter2/section6_tf.ipynb): * This code executes flawlessly because at no point do we ever have the following conditions: * **`return_tensors="pt" or return_tensors="tf"`** * **`padding="max_length"`** * **`max_length=<a number smaller than the number of tokens in the longest sequence>`** * [my self contained example](https://colab.research.google.com/drive/1_Zt5I-pBAIIaiJRL9-JTg4BuOWR8a8LY?usp=sharing): * I simply coded a simple self-contained example to show the error and when it appears * I am 99% sure this has to do with the handling of ragged tensors. * As when you pass a **`max_length`** value smaller than the longest sequence, the default behaviour is to leave this sequence alone (no truncation). * This can result in tensors of different lengths. * This causes problems as these tensors cannot be batched together. **Similar errors can also occur when performing truncation and passing other combinations of arguments.** <br> The tasks I am working on is: * I am replicating the entire HuggingFace course for Kaggle consumption (I'm doing both Pytorch and Tensorflow versions) ## To reproduce Steps to reproduce the behavior: 1. Open the [**self-contained example**](https://colab.research.google.com/drive/1_Zt5I-pBAIIaiJRL9-JTg4BuOWR8a8LY?usp=sharing) 2. Run the code cells ## Code Snippet **Code That Works** ```python tokenizer(["I've been waiting for a HuggingFace course my whole life.", "So have I!"], padding='max_length', max_length=8) ``` **Code That Breaks** ```python tokenizer(["I've been waiting for a HuggingFace course my whole life.", "So have I!"], return_tensors="tf", padding='max_length', max_length=8) ``` **Two main errors:** ```python ValueError: Can't convert non-rectangular Python sequence to Tensor. ``` ```python ValueError: Unable to create tensor, you should probably activate truncation and/or padding with 'padding=True' 'truncation=True' to have batched tensors with the same length. ``` **Example of how to handle this situation** ```python tokenizer(["I've been waiting for a HuggingFace course my whole life.", "So have I!"], return_tensors="tf", truncation=True, padding='max_length', max_length=8) ``` ## Expected behavior I'm not 100% sure what the desired functionality is in this case. However, I believe that throwing an error is probably not the correct action. Also, the error message does not make 100% sense. I would think that delivering a warning and coercing the argument **`truncation`** to be **`True`** would be an acceptable option. If this is of interest, I can create a pull request for this feature. **Thanks in advance!**

I suspect this is actually a bug in the Wav2Vec2 processor. None of the suggested fixes seem to do anything e.g padding=True in various places.

Has anyone encountered this with these models, and.do you have any insight or suggestions

Thank you
Peter

lnxdx · November 22, 2023, 6:23pm

I’m facing the same problem. Did you find the solution?