* Add spacy.TextCatParametricAttention.v1
This layer provides is a simplification of the ensemble classifier that
only uses paramteric attention. We have found empirically that with a
sufficient amount of training data, using the ensemble classifier with
BoW does not provide significant improvement in classifier accuracy.
However, plugging in a BoW classifier does reduce GPU training and
inference performance substantially, since it uses a GPU-only kernel.
* Fix merge fallout
* Add TextCatReduce.v1
This is a textcat classifier that pools the vectors generated by a
tok2vec implementation and then applies a classifier to the pooled
representation. Three reductions are supported for pooling: first, max,
and mean. When multiple reductions are enabled, the reductions are
concatenated before providing them to the classification layer.
This model is a generalization of the TextCatCNN model, which only
supports mean reductions and is a bit of a misnomer, because it can also
be used with transformers. This change also reimplements TextCatCNN.v2
using the new TextCatReduce.v1 layer.
* Doc fixes
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Fully specify `TextCatCNN` <-> `TextCatReduce` equivalence
* Move TextCatCNN docs to legacy, in prep for moving to spacy-legacy
* Add back a test for TextCatCNN.v2
* Replace TextCatCNN in pipe configurations and templates
* Add an infobox to the `TextCatReduce` section with an `TextCatCNN` anchor
* Add last reduction (`use_reduce_last`)
* Remove non-working TextCatCNN Netlify redirect
* Revert layer changes for the quickstart
* Revert one more quickstart change
* Remove unused import
* Fix docstring
* Fix setting name in error message
---------
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* Update `TextCatBOW` to use the fixed `SparseLinear` layer
A while ago, we fixed the `SparseLinear` layer to use all available
parameters: https://github.com/explosion/thinc/pull/754
This change updates `TextCatBOW` to `v3` which uses the new
`SparseLinear_v2` layer. This results in a sizeable improvement on a
text categorization task that was tested.
While at it, this `spacy.TextCatBOW.v3` also adds the `length_exponent`
option to make it possible to change the hidden size. Ideally, we'd just
have an option called `length`. But the way that `TextCatBOW` uses
hashes results in a non-uniform distribution of parameters when the
length is not a power of two.
* Replace TexCatBOW `length_exponent` parameter by `length`
We now round up the length to the next power of two if it isn't
a power of two.
* Remove some tests for TextCatBOW.v2
* Fix missing import
* Add note on score_weight if using a non-default span_key for SpanCat.
* Fix formatting.
* Fix formatting.
* Fix typo.
* Use warning infobox.
* Fix infobox formatting.
* add span key option for CLI evaluation
* Rephrase CLI help to refer to Doc.spans instead of spancat
* Rephrase docs to refer to Doc.spans instead of spancat
---------
Co-authored-by: Adriane Boyd <adrianeboyd@gmail.com>
* fix construction example
* shorten task-specific factory list
* small edits to HF models
* small edit to API models
* typo
* fix space
Co-authored-by: Raphael Mitsch <r.mitsch@outlook.com>
---------
Co-authored-by: Raphael Mitsch <r.mitsch@outlook.com>
* initial
* initial documentation run
* fix typo
* Remove mentions of Torchscript and quantization
Both are disabled in the initial release of `spacy-curated-transformers`.
* Fix `piece_encoder` entries
* Remove `spacy-transformers`-specific warning
* Fix duplicate entries in tables
* Doc fixes
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Remove type aliases
* Fix copy-paste typo
* Change `debug pieces` version tag to `3.7`
* Set curated transformers API version to `3.7`
* Fix transformer listener naming
* Add docs for `init fill-config-transformer`
* Update CLI command invocation syntax
* Update intro section of the pipeline component docs
* Fix source URL
* Add a note to the architectures section about the `init fill-config-transformer` CLI command
* Apply suggestions from code review
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Update CLI command name, args
* Remove hyphen from the `curated-transformers.mdx` filename
* Fix links
* Remove placeholder text
* Add text to the model/tokenizer loader sections
* Fill in the `DocTransformerOutput` section
* Formatting fixes
* Add curated transformer page to API docs sidebar
* More formatting fixes
* Remove TODO comment
* Remove outdated info about default config
* Apply suggestions from code review
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
* Add link to HF model hub
* `prettier`
---------
Co-authored-by: Madeesh Kannan <shadeMe@users.noreply.github.com>
Co-authored-by: Sofie Van Landeghem <svlandeg@users.noreply.github.com>
SpaCy's HashEmbedCNN layer performs convolutions over tokens to produce
contextualized embeddings using a `MaxoutWindowEncoder` layer. These
convolutions are implemented using Thinc's `expand_window` layer, which
concatenates `window_size` neighboring sequence items on either side of
the sequence item being processed. This is repeated across `depth`
convolutional layers.
For example, consider the sequence "ABCDE" and a `MaxoutWindowEncoder`
layer with a context window of 1 and a depth of 2. We'll focus on the
token "C". We can visually represent the contextual embedding produced
for "C" as:
```mermaid
flowchart LR
A0(A<sub>0</sub>)
B0(B<sub>0</sub>)
C0(C<sub>0</sub>)
D0(D<sub>0</sub>)
E0(E<sub>0</sub>)
B1(B<sub>1</sub>)
C1(C<sub>1</sub>)
D1(D<sub>1</sub>)
C2(C<sub>2</sub>)
A0 --> B1
B0 --> B1
C0 --> B1
B0 --> C1
C0 --> C1
D0 --> C1
C0 --> D1
D0 --> D1
E0 --> D1
B1 --> C2
C1 --> C2
D1 --> C2
```
Described in words, this graph shows that before the first layer of the
convolution, the "receptive field" centered at each token consists only
of that same token. That is to say, that we have a receptive field of 1.
The first layer of the convolution adds one neighboring token on either
side to the receptive field. Since this is done on both sides, the
receptive field increases by 2, giving the first layer a receptive field
of 3. The second layer of the convolutions adds an _additional_
neighboring token on either side to the receptive field, giving a final
receptive field of 5.
However, this doesn't match the formula currently given in the docs,
which read:
> The receptive field of the CNN will be
> `depth * (window_size * 2 + 1)`, so a 4-layer network with a window
> size of `2` will be sensitive to 20 words at a time.
Substituting in our depth of 2 and window size of 1, this formula gives
us a receptive field of:
```
depth * (window_size * 2 + 1)
= 2 * (1 * 2 + 1)
= 2 * (2 + 1)
= 2 * 3
= 6
```
This not only doesn't match our computations from above, it's also an
even number! This is suspicious, since the receptive field is supposed
to be centered on a token, and not between tokens. Generally, this
formula results in an even number for any even value of `depth`.
The error in this formula is that the adjustment for the center token
is multiplied by the depth, when it should occur only once. The
corrected formula, `depth * window_size * 2 + 1`, gives the correct
value for our small example from above:
```
depth * window_size * 2 + 1
= 2 * 1 * 2 + 1
= 4 + 1
= 5
```
These changes update the docs to correct the receptive field formula and
the example receptive field size.
