Dense dimension sparse input

[tabergma]

Proposed changes:
Do not set the output dimension of the sparse-to-dense layers to the same dimension as the dense features.

closes #6555

Status (please check what you already did):

• added some tests for the functionality
• updated the documentation
• updated the changelog (please check changelog for instructions)
• reformat files using black (please check Readme for instructions)

[github-actions]

Commit: 96a529d, The full report is available as an artifact.

Dataset: Carbon Bot

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) test: 42s, train: 3m45s, total: 4m27s	0.8388	0.6260	0.5894
Sparse + DIET(bow) + ResponseSelector(bow) test: 19s, train: 2m43s, total: 3m1s	0.7417	0.6260	0.4651

Dataset: Hermit

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) test: 1m23s, train: 23m19s, total: 24m41s	0.8866	0.7487	no data
Sparse + DIET(bow) + ResponseSelector(bow) test: 35s, train: 19m16s, total: 19m50s	0.8281	0.7487	no data

Dataset: Sara

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) test: 1m3s, train: 7m44s, total: 8m47s	0.8981	0.8683	0.9283
Sparse + DIET(bow) + ResponseSelector(bow) test: 27s, train: 5m18s, total: 5m44s	0.8139	0.8683	0.8522

[github-actions]

Commit: 83c9305, The full report is available as an artifact.

Dataset: Carbon Bot

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) test: 41s, train: 3m40s, total: 4m21s	0.8272	0.6260	0.6026
Sparse + DIET(bow) + ResponseSelector(bow) test: 21s, train: 2m43s, total: 3m4s	0.7437	0.6260	0.5033

Dataset: Hermit

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) test: 1m20s, train: 23m25s, total: 24m45s	0.8931	0.7487	no data
Sparse + DIET(bow) + ResponseSelector(bow) test: 35s, train: 19m25s, total: 20m0s	0.8290	0.7487	no data

Dataset: Sara

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) test: 1m3s, train: 7m51s, total: 8m54s	0.8981	0.8683	0.9196
Sparse + DIET(bow) + ResponseSelector(bow) test: 28s, train: 5m17s, total: 5m44s	0.8384	0.8683	0.8413

[github-actions]

Commit: 73f5d12, The full report is available as an artifact.

Dataset: Carbon Bot

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) test: 45s, train: 4m5s, total: 4m49s	0.8485	0.6260	0.5430
Sparse + DIET(bow) + ResponseSelector(bow) test: 20s, train: 2m53s, total: 3m13s	0.7223	0.6260	0.4967

Dataset: Hermit

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + DIET(bow) + ResponseSelector(bow) test: 37s, train: 20m0s, total: 20m36s	0.8253	0.7487	no data

Dataset: Sara

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) test: 1m7s, train: 8m13s, total: 9m19s	0.8962	0.8683	0.9217
Sparse + DIET(bow) + ResponseSelector(bow) test: 28s, train: 5m44s, total: 6m11s	0.8335	0.8683	0.8261

[github-actions]

/modeltest

include:
 - dataset: 
      - "Carbon Bot"
      - "Sara"
   config:
      - "Sparse + DIET(bow) + ResponseSelector(bow)"
      - "Sparse + DIET(seq) + ResponseSelector(bow)"
      - "Sparse + ConveRT + DIET(bow) + ResponseSelector(bow)"
      - "Sparse + ConveRT + DIET(seq) + ResponseSelector(bow)"
      - "Sparse + Spacy + DIET(bow) + ResponseSelector(bow)"
      - "Sparse + Spacy + DIET(seq) + ResponseSelector(bow)"

[github-actions]

Commit: 9e65197, The full report is available as an artifact.

Dataset: Carbon Bot

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) test: 43s, train: 3m50s, total: 4m32s	0.8369	0.6260	0.5960
Sparse + DIET(bow) + ResponseSelector(bow) test: 20s, train: 2m45s, total: 3m4s	0.7340	0.6260	0.5298

Dataset: Sara

Configuration	Intent Classification Micro F1	Entity Recognition Micro F1	Response Selection Micro F1
Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) test: 1m2s, train: 7m56s, total: 8m58s	0.8874	0.8683	0.9283
Sparse + DIET(bow) + ResponseSelector(bow) test: 28s, train: 5m34s, total: 6m1s	0.8325	0.8683	0.8522

[tabergma]

Results match those from the master branch (see https://metabase.rasa.com/dashboard/166)

[Ghostvv]

in response selector, 128 is not enough?

[Ghostvv]

since we changed it to predict labels, it is basically another DIET

[tabergma]

When I set it to something lower, the performance dropped (see the tables above).

[Ghostvv]

did it reduce train run time for DIET?

[tabergma]

The train time for Sparse + ConveRT + DIET(bow) + ResponseSelector(bow) increased for Sara and Carbon Bot by a couple of seconds. Without ConveRT it stayed the same.