patrickvonplaten
commited on
Commit
•
fe587de
1
Parent(s):
b5a4acb
Upload README.md
Browse files
README.md
CHANGED
@@ -1,8 +1,115 @@
|
|
1 |
---
|
|
|
|
|
|
|
|
|
2 |
tags:
|
3 |
-
-
|
|
|
|
|
|
|
4 |
---
|
5 |
|
6 |
-
#
|
7 |
-
|
8 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
language:
|
3 |
+
- en
|
4 |
+
datasets:
|
5 |
+
- c4
|
6 |
tags:
|
7 |
+
- deep-narrow
|
8 |
+
inference: false
|
9 |
+
|
10 |
+
license: apache-2.0
|
11 |
---
|
12 |
|
13 |
+
# T5-Efficient-SMALL-NL36 (Deep-Narrow version)
|
14 |
+
|
15 |
+
T5-Efficient-SMALL-NL36 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).
|
16 |
+
It is a *pretrained-only* checkpoint and was released with the
|
17 |
+
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
|
18 |
+
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
|
19 |
+
|
20 |
+
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
|
21 |
+
of similar parameter count.
|
22 |
+
|
23 |
+
To quote the paper:
|
24 |
+
|
25 |
+
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
|
26 |
+
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
|
27 |
+
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
|
28 |
+
> 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 find
|
30 |
+
> 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 36
|
32 |
+
> 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 to
|
35 |
+
> consider.
|
36 |
+
|
37 |
+
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
|
38 |
+
A sequence of word embeddings is therefore processed sequentially by each transformer block.
|
39 |
+
|
40 |
+
## Details model architecture
|
41 |
+
|
42 |
+
This model checkpoint - **t5-efficient-small-nl36** - is of model type **Small** with the following variations:
|
43 |
+
- **nl** is **36**
|
44 |
+
|
45 |
+
It has **280.87** million parameters and thus requires *ca.* **1123.47 MB** of memory in full precision (*fp32*)
|
46 |
+
or **561.74 MB** of memory in half precision (*fp16* or *bf16*).
|
47 |
+
|
48 |
+
A summary of the *original* T5 model architectures can be seen here:
|
49 |
+
|
50 |
+
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
|
51 |
+
| ----| ---- | ---- | ---- | ---- | ---- | ----|
|
52 |
+
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
|
53 |
+
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
|
54 |
+
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
|
55 |
+
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
|
56 |
+
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
|
57 |
+
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
|
58 |
+
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
|
59 |
+
|
60 |
+
whereas the following abbreviations are used:
|
61 |
+
|
62 |
+
| Abbreviation | Definition |
|
63 |
+
| ----| ---- |
|
64 |
+
| nl | Number of transformer blocks (depth) |
|
65 |
+
| dm | Dimension of embedding vector (output vector of transformers block) |
|
66 |
+
| kv | Dimension of key/value projection matrix |
|
67 |
+
| nh | Number of attention heads |
|
68 |
+
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
|
69 |
+
| el | Number of transformer blocks in the encoder (encoder depth) |
|
70 |
+
| dl | Number of transformer blocks in the decoder (decoder depth) |
|
71 |
+
| sh | Signifies that attention heads are shared |
|
72 |
+
| skv | Signifies that key-values projection matrices are tied |
|
73 |
+
|
74 |
+
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
|
75 |
+
|
76 |
+
## Pre-Training
|
77 |
+
|
78 |
+
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
|
79 |
+
the span-based masked language modeling (MLM) objective.
|
80 |
+
|
81 |
+
## Fine-Tuning
|
82 |
+
|
83 |
+
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
|
84 |
+
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
|
85 |
+
You can follow on of the following examples on how to fine-tune the model:
|
86 |
+
|
87 |
+
*PyTorch*:
|
88 |
+
|
89 |
+
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
|
90 |
+
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
|
91 |
+
- [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.
|
92 |
+
|
93 |
+
*Tensorflow*:
|
94 |
+
|
95 |
+
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
|
96 |
+
- [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.
|
97 |
+
|
98 |
+
*JAX/Flax*:
|
99 |
+
|
100 |
+
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
|
101 |
+
- [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.
|
102 |
+
|
103 |
+
## Downstream Performance
|
104 |
+
|
105 |
+
TODO: Add table if available
|
106 |
+
|
107 |
+
## Computational Complexity
|
108 |
+
|
109 |
+
TODO: Add table if available
|
110 |
+
|
111 |
+
## More information
|
112 |
+
|
113 |
+
We 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.
|
114 |
+
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
|
115 |
+
model 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.
|