mirror of
https://github.com/explosion/spaCy.git
synced 2024-12-26 01:46:28 +03:00
Docs for new spacy-trf architectures (#8954)
* use TransformerModel.v2 in quickstart * update docs for new transformer architectures * bump spacy_transformers to 1.1.0 * Add new arguments spacy-transformers.TransformerModel.v3 * Mention that mixed-precision support is experimental * Describe delta transformers.Tok2VecTransformer versions * add dot * add dot, again * Update some more TransformerModel references v2 -> v3 * Add mixed-precision options to the training quickstart Disable mixed-precision training/prediction by default. * Update setup.cfg Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com> * Apply suggestions from code review Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com> * Update website/docs/usage/embeddings-transformers.md Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com> Co-authored-by: Daniël de Kok <me@danieldk.eu> Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
This commit is contained in:
Sofie Van Landeghem
Adriane Boyd
parent
9dfb12e29f
commit
eaa6798c66
|
|
@ -70,7 +70,7 @@ console_scripts =
|
|||
lookups =
|
||||
spacy_lookups_data>=1.0.2,<1.1.0
|
||||
transformers =
|
||||
spacy_transformers>=1.0.1,<1.1.0
|
||||
spacy_transformers>=1.0.1,<1.2.0
|
||||
ray =
|
||||
spacy_ray>=0.1.0,<1.0.0
|
||||
cuda =
|
||||
|
|
|
|||
|
|
@ -32,7 +32,7 @@ batch_size = {{ 128 if hardware == "gpu" else 1000 }}
|
|||
factory = "transformer"
|
||||
|
||||
[components.transformer.model]
|
||||
@architectures = "spacy-transformers.TransformerModel.v1"
|
||||
@architectures = "spacy-transformers.TransformerModel.v3"
|
||||
name = "{{ transformer["name"] }}"
|
||||
tokenizer_config = {"use_fast": true}
|
||||
|
||||
|
|
|
|||
|
|
@ -332,15 +332,18 @@ for details and system requirements.
|
|||
|
||||
</Infobox>
|
||||
|
||||
### spacy-transformers.TransformerModel.v1 {#TransformerModel}
|
||||
### spacy-transformers.TransformerModel.v3 {#TransformerModel}
|
||||
|
||||
> #### Example Config
|
||||
>
|
||||
> ```ini
|
||||
> [model]
|
||||
> @architectures = "spacy-transformers.TransformerModel.v1"
|
||||
> @architectures = "spacy-transformers.TransformerModel.v3"
|
||||
> name = "roberta-base"
|
||||
> tokenizer_config = {"use_fast": true}
|
||||
> transformer_config = {}
|
||||
> mixed_precision = true
|
||||
> grad_scaler_config = {"init_scale": 32768}
|
||||
>
|
||||
> [model.get_spans]
|
||||
> @span_getters = "spacy-transformers.strided_spans.v1"
|
||||
|
|
@ -367,11 +370,30 @@ in other components, see
|
|||
[Tok2VecTransformer](/api/architectures#Tok2VecTransformer).
|
||||
|
||||
| Name | Description |
|
||||
| ------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
|
||||
|----------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
|
||||
| `name` | Any model name that can be loaded by [`transformers.AutoModel`](https://huggingface.co/transformers/model_doc/auto.html#transformers.AutoModel). ~~str~~ |
|
||||
| `get_spans` | Function that takes a batch of [`Doc`](/api/doc) object and returns lists of [`Span`](/api) objects to process by the transformer. [See here](/api/transformer#span_getters) for built-in options and examples. ~~Callable[[List[Doc]], List[Span]]~~ |
|
||||
| `tokenizer_config` | Tokenizer settings passed to [`transformers.AutoTokenizer`](https://huggingface.co/transformers/model_doc/auto.html#transformers.AutoTokenizer). ~~Dict[str, Any]~~ |
|
||||
| `transformer_config` | Settings to pass to the transformers forward pass. ~~Dict[str, Any]~~ |
|
||||
| `mixed_precision` | Replace whitelisted ops by half-precision counterparts. Speeds up training and prediction on GPUs with [Tensor Cores](https://developer.nvidia.com/tensor-cores) and reduces GPU memory use. ~~bool~~ |
|
||||
| `grad_scaler_config` | Configuration to pass to `thinc.api.PyTorchGradScaler` during training when `mixed_precision` is enabled. ~~Dict[str, Any]~~ |
|
||||
| **CREATES** | The model using the architecture. ~~Model[List[Doc], FullTransformerBatch]~~ |
|
||||
| | |
|
||||
|
||||
<Infobox title="Mixed precision support" variant="warning">
|
||||
Mixed-precision support is currently an experimental feature.
|
||||
</Infobox>
|
||||
|
||||
<Accordion title="Previous versions of spacy-transformers.TransformerModel" spaced>
|
||||
|
||||
* The `transformer_config` argument was added in
|
||||
`spacy-transformers.TransformerModel.v2`.
|
||||
* The `mixed_precision` and `grad_scaler_config` arguments were added in
|
||||
`spacy-transformers.TransformerModel.v3`.
|
||||
|
||||
The other arguments are shared between all versions.
|
||||
|
||||
</Accordion>
|
||||
|
||||
### spacy-transformers.TransformerListener.v1 {#TransformerListener}
|
||||
|
||||
|
|
@ -403,16 +425,19 @@ a single token vector given zero or more wordpiece vectors.
|
|||
| `upstream` | A string to identify the "upstream" `Transformer` component to communicate with. By default, the upstream name is the wildcard string `"*"`, but you could also specify the name of the `Transformer` component. You'll almost never have multiple upstream `Transformer` components, so the wildcard string will almost always be fine. ~~str~~ |
|
||||
| **CREATES** | The model using the architecture. ~~Model[List[Doc], List[Floats2d]]~~ |
|
||||
|
||||
### spacy-transformers.Tok2VecTransformer.v1 {#Tok2VecTransformer}
|
||||
### spacy-transformers.Tok2VecTransformer.v3 {#Tok2VecTransformer}
|
||||
|
||||
> #### Example Config
|
||||
>
|
||||
> ```ini
|
||||
> [model]
|
||||
> @architectures = "spacy-transformers.Tok2VecTransformer.v1"
|
||||
> @architectures = "spacy-transformers.Tok2VecTransformer.v3"
|
||||
> name = "albert-base-v2"
|
||||
> tokenizer_config = {"use_fast": false}
|
||||
> transformer_config = {}
|
||||
> grad_factor = 1.0
|
||||
> mixed_precision = true
|
||||
> grad_scaler_config = {"init_scale": 32768}
|
||||
> ```
|
||||
|
||||
Use a transformer as a [`Tok2Vec`](/api/tok2vec) layer directly. This does
|
||||
|
|
@ -422,13 +447,32 @@ object, but it's a **simpler solution** if you only need the transformer within
|
|||
one component.
|
||||
|
||||
| Name | Description |
|
||||
| ------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
|
||||
|----------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
|
||||
| `get_spans` | Function that takes a batch of [`Doc`](/api/doc) object and returns lists of [`Span`](/api) objects to process by the transformer. [See here](/api/transformer#span_getters) for built-in options and examples. ~~Callable[[List[Doc]], List[Span]]~~ |
|
||||
| `tokenizer_config` | Tokenizer settings passed to [`transformers.AutoTokenizer`](https://huggingface.co/transformers/model_doc/auto.html#transformers.AutoTokenizer). ~~Dict[str, Any]~~ |
|
||||
| `transformer_config` | Settings to pass to the transformers forward pass. ~~Dict[str, Any]~~ |
|
||||
| `pooling` | A reduction layer used to calculate the token vectors based on zero or more wordpiece vectors. If in doubt, mean pooling (see [`reduce_mean`](https://thinc.ai/docs/api-layers#reduce_mean)) is usually a good choice. ~~Model[Ragged, Floats2d]~~ |
|
||||
| `grad_factor` | Reweight gradients from the component before passing them upstream. You can set this to `0` to "freeze" the transformer weights with respect to the component, or use it to make some components more significant than others. Leaving it at `1.0` is usually fine. ~~float~~ |
|
||||
| `mixed_precision` | Replace whitelisted ops by half-precision counterparts. Speeds up training and prediction on GPUs with [Tensor Cores](https://developer.nvidia.com/tensor-cores) and reduces GPU memory use. ~~bool~~ |
|
||||
| `grad_scaler_config` | Configuration to pass to `thinc.api.PyTorchGradScaler` during training when `mixed_precision` is enabled. ~~Dict[str, Any]~~ |
|
||||
| **CREATES** | The model using the architecture. ~~Model[List[Doc], List[Floats2d]]~~ |
|
||||
|
||||
|
||||
<Infobox title="Mixed precision support" variant="warning">
|
||||
Mixed-precision support is currently an experimental feature.
|
||||
</Infobox>
|
||||
|
||||
<Accordion title="Previous versions of spacy-transformers.Tok2VecTransformer" spaced>
|
||||
|
||||
* The `transformer_config` argument was added in
|
||||
`spacy-transformers.Tok2VecTransformer.v2`.
|
||||
* The `mixed_precision` and `grad_scaler_config` arguments were added in
|
||||
`spacy-transformers.Tok2VecTransformer.v3`.
|
||||
|
||||
The other arguments are shared between all versions.
|
||||
|
||||
</Accordion>
|
||||
|
||||
## Pretraining architectures {#pretrain source="spacy/ml/models/multi_task.py"}
|
||||
|
||||
The spacy `pretrain` command lets you initialize a `Tok2Vec` layer in your
|
||||
|
|
|
|||
|
|
@ -92,9 +92,12 @@ https://github.com/explosion/spacy-transformers/blob/master/spacy_transformers/p
|
|||
> # Construction via add_pipe with custom config
|
||||
> config = {
|
||||
> "model": {
|
||||
> "@architectures": "spacy-transformers.TransformerModel.v1",
|
||||
> "@architectures": "spacy-transformers.TransformerModel.v3",
|
||||
> "name": "bert-base-uncased",
|
||||
> "tokenizer_config": {"use_fast": True}
|
||||
> "tokenizer_config": {"use_fast": True},
|
||||
> "transformer_config": {"output_attentions": True},
|
||||
> "mixed_precision": True,
|
||||
> "grad_scaler_config": {"init_scale": 32768}
|
||||
> }
|
||||
> }
|
||||
> trf = nlp.add_pipe("transformer", config=config)
|
||||
|
|
|
|||
|
|
@ -351,7 +351,7 @@ factory = "transformer"
|
|||
max_batch_items = 4096
|
||||
|
||||
[components.transformer.model]
|
||||
@architectures = "spacy-transformers.TransformerModel.v1"
|
||||
@architectures = "spacy-transformers.TransformerModel.v3"
|
||||
name = "bert-base-cased"
|
||||
tokenizer_config = {"use_fast": true}
|
||||
|
||||
|
|
@ -367,7 +367,7 @@ The `[components.transformer.model]` block describes the `model` argument passed
|
|||
to the transformer component. It's a Thinc
|
||||
[`Model`](https://thinc.ai/docs/api-model) object that will be passed into the
|
||||
component. Here, it references the function
|
||||
[spacy-transformers.TransformerModel.v1](/api/architectures#TransformerModel)
|
||||
[spacy-transformers.TransformerModel.v3](/api/architectures#TransformerModel)
|
||||
registered in the [`architectures` registry](/api/top-level#registry). If a key
|
||||
in a block starts with `@`, it's **resolved to a function** and all other
|
||||
settings are passed to the function as arguments. In this case, `name`,
|
||||
|
|
|
|||
Loading…
Reference in New Issue
Block a user