…e.to

Fixes 1397

BREAKING(model): get rid of (flaky) `Model.introspection`

…o feat/joint-diarization-and-embedding

- fixes the dimension error between files id and probabilties arrays
- changes the way of how chunks for the embedding task are sampled
- creates two functions to draw chunks, one for each subtask

Tests are required to ensure that there are no bugs

For now this is a copy past from methods in segmentation task.

as computing this loss probably does not make sense in powerset
mode because first class (empty set of labels) does exactly this

as this instance attribute was not used

…` pipeline

Co-authored-by: Hervé BREDIN <[email protected]>

as these loop could break gradient flow and to optimize
the code

for now do the trick only for the diarization subtask

There was an issue when the number of speakers in a chunk was
greater than the maximum number per chunk set for the task.

these two methods were identical to the methods inherited from the
`SegmentationTaskMixin` class

and fix issue with the loss type during training

this version replace `StatsPool` by a concatenation of the last
outputs of TDNN (for the embedding part) and LSTM (for the diarization
part) and a LSTM layer

Now, this LSTM is bidirectionnal and has a hidden size of 1500, so
the outputs shape of this encoder is (b, s, 1500*2). This will allow
comparing with `StatsPool` version of the SPEED model

The goal of this method is to generate the data needed by the task
and save it on disk for future uses, for example by the `setup` method.
The objective is to avoid systematically recreating data on each process
at the beginning of a training

Now all the segmentations tasks in `pyannote` inherit the `SegmentationTask`
(previously `SegmentationTaskMixin`), which inherits the `Task` class. This
commit also adds a `prepared_data` attribute  to the `Task` class. That
attribute is a dict which contains all the prepared data by the `prepare_data`
method.

…note-audio into feat/data_preparation

One for the test of the `MultiLabelSegmentation` task, and
the other for the test of the `SupervisedRepresentationLearningWithArcFace`
task.

This eliminates the need to reload pickle data in setup when in the main process

…cenarios

…note-audio into feat/data_preparation

This issue occured when a list of classes was  specified during `MultiLabelSegmentation`
instanciation.

…iarization-and-embedding-with-prepared-data

…ding-with-prepared-data

* use npz archive instead pickle to save task data

* improve code readability

* improve(task): update numpy array dtypes

In order to use types whose size better machtes the contents of the arrays

* remove `end` entry from `annotated_regions` numpy array

This entry was redundant with the start and duration entries,
since `end` = `start` + `duration`.

* fix: allow data preparation to be finished when task has no validation

* improve: clear data lists after assignation to `self.prepared_data`

This is to avoid data redundancy in the `prepare_data` method

---------

Co-authored-by: clement-pages <[email protected]>

Now the joint task uses `prepare_data` and `setup` from core `Task` and
`SpeakerDiarization` task.

…' of github.com:clement-pages/pyannote-audio into feat/joint-diarization-and-embedding-with-prepared-data

…embedding-with-prepared-data

…mbedding`

…ddins

This new model is based on a `WeSpeakerResnet34` for the speaker embeddings
extraction part, and on `PyanNet` for (local) segmentation.

…ata`

…embedding-with-prepared-data

…ocol

…embedding-with-prepared-data

…-prepared-data

Now, the first `num_dia_samples` samples in a batch  are dedicated to the diarization
substak, and the remaining sample are for the embedding subtask

... and fix some bugs

…-prepared-data

…computation

…-prepared-data

…ding-with-prepared-data