google/t5-efficient-tiny

1---2language:3- en4datasets:5- c46tags:7- deep-narrow8inference: false9 10license: apache-2.011---12 13# T5-Efficient-TINY (Deep-Narrow version)14 15T5-Efficient-TINY is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).16It is a *pretrained-only* checkpoint and was released with the17paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**18by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.19 20In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures21of similar parameter count.22 23To quote the paper:24 25> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased26> before considering any other forms of uniform scaling across other dimensions. This is largely due to27> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a28> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,29> a tall base model might also generally more efficient compared to a large model. We generally find30> that, regardless of size, even if absolute performance might increase as we continue to stack layers,31> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 3632> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,33> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,34> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to35> consider.36 37To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.38A sequence of word embeddings is therefore processed sequentially by each transformer block.39 40## Details model architecture41 42This model checkpoint - **t5-efficient-tiny** - is of model type **Tiny** with no variations.43It has **15.58** million parameters and thus requires *ca.* **62.32 MB** of memory in full precision (*fp32*)44 or  **31.16 MB** of memory in half precision (*fp16* or *bf16*).45 46A summary of the *original* T5 model architectures can be seen here:47 48| Model | nl (el/dl) | ff | dm | kv | nh | #Params|49| ----| ---- | ---- | ---- | ---- | ---- | ----|50| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|51| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|52| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|53| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|54| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|55| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|56| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|57 58whereas the following abbreviations are used:59 60| Abbreviation | Definition |61| ----| ---- |62| nl | Number of transformer blocks (depth) |63| dm | Dimension of embedding vector (output vector of transformers block) |64| kv | Dimension of key/value projection matrix |65| nh | Number of attention heads |66| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) | 67| el | Number of transformer blocks in the encoder (encoder depth) | 68| dl | Number of transformer blocks in the decoder (decoder depth) | 69| sh | Signifies that attention heads are shared | 70| skv | Signifies that key-values projection matrices are tied | 71 72If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.73 74## Pre-Training75 76The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using 77the span-based masked language modeling (MLM) objective.78 79## Fine-Tuning80 81**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.82The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.83You can follow on of the following examples on how to fine-tune the model:84 85*PyTorch*:86 87- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)88- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)89- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.90 91*Tensorflow*:92 93- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)94- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.95 96*JAX/Flax*:97 98- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)99- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.100 101## Downstream Performance102 103TODO: Add table if available104 105## Computational Complexity106 107TODO: Add table if available108 109## More information110 111We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.112As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv* 113model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.