google/fnet-base

1---2language: en3tags:4- fnet5license: apache-2.06datasets:7- c48---9 10# FNet base model11 12Pretrained model on English language using a masked language modeling (MLM) and next sentence prediction (NSP) objective. It was 13introduced in [this paper](https://arxiv.org/abs/2105.03824) and first released in [this repository](https://github.com/google-research/google-research/tree/master/f_net).14This model is cased: it makes a difference between english and English. The model achieves 0.58 accuracy on MLM objective and 0.80 on NSP objective.15 16Disclaimer: This model card has been written by [gchhablani](https://huggingface.co/gchhablani).17 18## Model description19 20FNet is a transformers model with attention replaced with fourier transforms. Hence, the inputs do not contain an `attention_mask`. It is pretrained on a large corpus of 21English data in a self-supervised fashion. This means it was pretrained on the raw texts only, with no humans labelling22them in any way (which is why it can use lots of publicly available data) with an automatic process to generate inputs and 23labels from those texts. More precisely, it was pretrained with two objectives:24 25- Masked language modeling (MLM): taking a sentence, the model randomly masks 15% of the words in the input then run26  the entire masked sentence through the model and has to predict the masked words. This is different from traditional27  recurrent neural networks (RNNs) that usually see the words one after the other, or from autoregressive models like28  GPT which internally mask the future tokens. It allows the model to learn a bidirectional representation of the29  sentence.30- Next sentence prediction (NSP): the models concatenates two masked sentences as inputs during pretraining. Sometimes31  they correspond to sentences that were next to each other in the original text, sometimes not. The model then has to32  predict if the two sentences were following each other or not.33 34This way, the model learns an inner representation of the English language that can then be used to extract features35useful for downstream tasks: if you have a dataset of labeled sentences for instance, you can train a standard36classifier using the features produced by the FNet model as inputs.37 38## Intended uses & limitations39 40You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to41be fine-tuned on a downstream task. See the [model hub](https://huggingface.co/models?filter=fnet) to look for42fine-tuned versions on a task that interests you.43 44Note that this model is primarily aimed at being fine-tuned on tasks that use the whole sentence (potentially masked)45to make decisions, such as sequence classification, token classification or question answering. For tasks such as text46generation you should look at model like GPT2.47 48## Training data49 50The FNet model was pretrained on [C4](https://huggingface.co/datasets/c4), a cleaned version of the Common Crawl dataset.51 52## Training procedure53 54### Preprocessing55 56The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 32,000. The inputs of the model are57then of the form:58 59```60[CLS] Sentence A [SEP] Sentence B [SEP]61```62 63With probability 0.5, sentence A and sentence B correspond to two consecutive sentences in the original corpus and in64the other cases, it's another random sentence in the corpus. Note that what is considered a sentence here is a65consecutive span of text usually longer than a single sentence. The only constrain is that the result with the two66"sentences" has a combined length of less than 512 tokens.67 68The details of the masking procedure for each sentence are the following:69- 15% of the tokens are masked.70- In 80% of the cases, the masked tokens are replaced by `[MASK]`.71- In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.72- In the 10% remaining cases, the masked tokens are left as is.73 74### Pretraining75 76FNet-base was trained on 4 cloud TPUs in Pod configuration (16 TPU chips total) for one million steps with a batch size77of 256. The sequence length was limited to 512 tokens. The optimizer78used is Adam with a learning rate of 1e-4, \\(\beta_{1} = 0.9\\) and \\(\beta_{2} = 0.999\\), a weight decay of 0.01,79learning rate warmup for 10,000 steps and linear decay of the learning rate after.80 81## Evaluation results82 83FNet-base was fine-tuned and evaluated on the validation data of the [GLUE benchamrk](https://huggingface.co/datasets/glue). The results of the official model (written in Flax) can be seen in Table 1 on page 7 of [the official paper](https://arxiv.org/abs/2105.03824).84 85For comparison, this model (ported to PyTorch) was fine-tuned and evaluated using the [official Hugging Face GLUE evaluation scripts](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification#glue-tasks) alongside [bert-base-cased](https://hf.co/models/bert-base-cased) for comparison.86The training was done on a single 16GB NVIDIA Tesla V100 GPU. For MRPC/WNLI, the models were trained for 5 epochs, while for other tasks, the models were trained for 3 epochs. A sequence length of 512 was used with batch size 16 and learning rate 2e-5.87 88The following table summarizes the results for [fnet-base](https://huggingface.co/google/fnet-base) (called *FNet (PyTorch) - Reproduced*) and [bert-base-cased](https://hf.co/models/bert-base-cased) (called *Bert (PyTorch) - Reproduced*) in terms of **fine-tuning** speed. The format is *hour:min:seconds*. **Note** that the authors compared **pre-traning** speed in [the official paper](https://arxiv.org/abs/2105.03824) instead.89 90| Task/Model | FNet-base (PyTorch) |Bert-base (PyTorch)|91|:----:|:-----------:|:----:|92| MNLI-(m/mm) | [06:40:55](https://huggingface.co/gchhablani/fnet-base-finetuned-mnli) | [09:52:33](https://huggingface.co/gchhablani/bert-base-cased-finetuned-mnli)|93| QQP  | [06:21:16](https://huggingface.co/gchhablani/fnet-base-finetuned-qqp) | [09:25:01](https://huggingface.co/gchhablani/bert-base-cased-finetuned-qqp) |94| QNLI  | [01:48:22](https://huggingface.co/gchhablani/fnet-base-finetuned-qnli)  | [02:40:22](https://huggingface.co/gchhablani/bert-base-cased-finetuned-qnli)|95| SST-2 | [01:09:27](https://huggingface.co/gchhablani/fnet-base-finetuned-sst2) | [01:42:17](https://huggingface.co/gchhablani/bert-base-cased-finetuned-sst2)|96| CoLA  | [00:09:47](https://huggingface.co/gchhablani/fnet-base-finetuned-cola) | [00:14:20](https://huggingface.co/gchhablani/bert-base-cased-finetuned-cola)|97| STS-B | [00:07:09](https://huggingface.co/gchhablani/fnet-base-finetuned-stsb) | [00:10:24](https://huggingface.co/gchhablani/bert-base-cased-finetuned-stsb)|98| MRPC  | [00:07:48](https://huggingface.co/gchhablani/fnet-base-finetuned-mrpc) | [00:11:12](https://huggingface.co/gchhablani/bert-base-cased-finetuned-mrpc)|99| RTE  | [00:03:24](https://huggingface.co/gchhablani/fnet-base-finetuned-rte) | [00:04:51](https://huggingface.co/gchhablani/bert-base-cased-finetuned-rte)|100| WNLI | [00:02:37](https://huggingface.co/gchhablani/fnet-base-finetuned-wnli) | [00:03:23](https://huggingface.co/gchhablani/bert-base-cased-finetuned-wnli)|101| SUM | 16:30:45 | 24:23:56 |102 103On average the PyTorch version of FNet-base requires *ca.* 32% less time for GLUE fine-tuning on GPU.104 105The following table summarizes the results for [fnet-base](https://huggingface.co/google/fnet-base) (called *FNet (PyTorch) - Reproduced*) and [bert-base-cased](https://hf.co/models/bert-base-cased) (called *Bert (PyTorch) - Reproduced*) in terms of performance and compares it to the reported performance of the official FNet-base model (called *FNet (Flax) - Official*). Note that the training hyperparameters of the reproduced models were not the same as the official model, so the performance may differ significantly for some tasks (for example: CoLA). 106 107| Task/Model | Metric | FNet-base (PyTorch) | Bert-base (PyTorch) | FNet-Base (Flax - official) |108|:----:|:-----------:|:----:|:-----------:|:----:|109| MNLI-(m/mm) | Accuracy or Match/Mismatch  | [76.75](https://huggingface.co/gchhablani/fnet-base-finetuned-mnli) | [84.10](https://huggingface.co/gchhablani/bert-base-cased-finetuned-mnli) | 72/73 |110| QQP | mean(Accuracy,F1)  | [86.5](https://huggingface.co/gchhablani/fnet-base-finetuned-qqp) | [89.26](https://huggingface.co/gchhablani/bert-base-cased-finetuned-qqp) | 83 |111| QNLI | Accuracy  | [84.39](https://huggingface.co/gchhablani/fnet-base-finetuned-qnli) | [90.99](https://huggingface.co/gchhablani/bert-base-cased-finetuned-qnli) | 80 | 112| SST-2 | Accuracy  | [89.45](https://huggingface.co/gchhablani/fnet-base-finetuned-sst2) | [92.32](https://huggingface.co/gchhablani/bert-base-cased-finetuned-sst2) | 95 |113| CoLA | Matthews corr or Accuracy | [35.94](https://huggingface.co/gchhablani/fnet-base-finetuned-cola) | [59.57](https://huggingface.co/gchhablani/bert-base-cased-finetuned-cola) | 69 |114| STS-B | Spearman corr. | [82.19](https://huggingface.co/gchhablani/fnet-base-finetuned-stsb) | [88.98](https://huggingface.co/gchhablani/bert-base-cased-finetuned-stsb) | 79 | 115| MRPC | mean(F1/Accuracy) | [81.15](https://huggingface.co/gchhablani/fnet-base-finetuned-mrpc) | [88.15](https://huggingface.co/gchhablani/bert-base-cased-finetuned-mrpc) | 76 | 116| RTE | Accuracy | [62.82](https://huggingface.co/gchhablani/fnet-base-finetuned-rte) | [67.15](https://huggingface.co/gchhablani/bert-base-cased-finetuned-rte) | 63 |117| WNLI | Accuracy | [54.93](https://huggingface.co/gchhablani/fnet-base-finetuned-wnli) | [46.48](https://huggingface.co/gchhablani/bert-base-cased-finetuned-wnli) | - |118| Avg  | - | 72.7 | 78.6 | 76.7 |119 120We can see that FNet-base achieves around 93% of BERT-base's performance on average.121 122For more details, please refer to the checkpoints linked with the scores. On overview of all fine-tuned checkpoints of the following table can be accessed [here](https://huggingface.co/models?other=fnet-bert-base-comparison).123 124### How to use125 126You can use this model directly with a pipeline for masked language modeling:127 128**Note: The mask filling pipeline doesn't work exactly as the original model performs masking after converting to tokens. In masking pipeline an additional space is added after the [MASK].**129 130```python131>>> from transformers import FNetForMaskedLM, FNetTokenizer, pipeline132>>> tokenizer = FNetTokenizer.from_pretrained("google/fnet-base")133>>> model = FNetForMaskedLM.from_pretrained("google/fnet-base")134>>> unmasker = pipeline('fill-mask', model=model, tokenizer=tokenizer)135>>> unmasker("Hello I'm a [MASK] model.")136 137[138    {"sequence": "hello i'm a new model.", "score": 0.12073223292827606, "token": 351, "token_str": "new"},139    {"sequence": "hello i'm a first model.", "score": 0.08501081168651581, "token": 478, "token_str": "first"},140    {"sequence": "hello i'm a next model.", "score": 0.060546260327100754, "token": 1037, "token_str": "next"},141    {"sequence": "hello i'm a last model.", "score": 0.038265593349933624, "token": 813, "token_str": "last"},142    {"sequence": "hello i'm a sister model.", "score": 0.033868927508592606, "token": 6232, "token_str": "sister"},143]144 145```146 147Here is how to use this model to get the features of a given text in PyTorch:148 149**Note: You must specify the maximum sequence length to be 512 and truncate/pad to the same length because the original model has no attention mask and considers all the hidden states during forward pass.**150 151```python152from transformers import FNetTokenizer, FNetModel153tokenizer = FNetTokenizer.from_pretrained("google/fnet-base")154model = FNetModel.from_pretrained("google/fnet-base")155text = "Replace me by any text you'd like."156encoded_input = tokenizer(text, return_tensors='pt', padding='max_length', truncation=True, max_length=512)157output = model(**encoded_input)158```159 160### BibTeX entry and citation info161 162```bibtex163@article{DBLP:journals/corr/abs-2105-03824,164  author    = {James Lee{-}Thorp and165               Joshua Ainslie and166               Ilya Eckstein and167               Santiago Onta{\~{n}}{\'{o}}n},168  title     = {FNet: Mixing Tokens with Fourier Transforms},169  journal   = {CoRR},170  volume    = {abs/2105.03824},171  year      = {2021},172  url       = {https://arxiv.org/abs/2105.03824},173  archivePrefix = {arXiv},174  eprint    = {2105.03824},175  timestamp = {Fri, 14 May 2021 12:13:30 +0200},176  biburl    = {https://dblp.org/rec/journals/corr/abs-2105-03824.bib},177  bibsource = {dblp computer science bibliography, https://dblp.org}178}179```180 181## Contributions182Thanks to [@gchhablani](https://huggingface.co/gchhablani) for adding this model.