This repo provides common building blocks for RETURNN, such as models or networks, network creation code, datasets, etc.
RETURNN originally used dicts to define the network (model, computation graph). The network consists of layers, where each layer represents a block of operations and potentially also parameters. In here, we adopt many conventions by PyTorch or functional Keras and other frameworks, such that you use pure Python code to define the network (model, computation graph). Further, a module instance does not represent the actual computation but only once you call it with actual inputs, then it will perform the actual computation (create a RETURNN layer, or the corresponding RETURNN layer dict).
See the wiki for a starting point for documentation.
from returnn_common import nn
class MyModelBlock(nn.Module):
def __init__(self, dim: nn.Dim, *,
hidden: nn.Dim = nn.FeatureDim("hidden", 2048),
dropout: float = 0.1):
super().__init__()
self.layer_norm = nn.LayerNorm(dim)
self.linear_hidden = nn.Linear(dim, hidden)
self.linear_out = nn.Linear(hidden, dim)
self.dropout = dropout
def __call__(self, x: nn.Tensor) -> nn.Tensor:
y = self.layer_norm(x)
y = self.linear_hidden(y)
y = nn.sigmoid(y)
y = self.linear_out(y)
y = nn.dropout(y, dropout=self.dropout, axis=nn.any_feature_dim)
return x + yIn case you want to have this three times separately now:
class MyModel(nn.Module):
def __init__(self, dim: nn.Dim):
super().__init__()
self.block1 = MyModelBlock(dim)
self.block2 = MyModelBlock(dim)
self.block3 = MyModelBlock(dim)
def __call__(self, x: nn.Tensor) -> nn.Tensor:
x = self.block1(x)
x = self.block2(x)
x = self.block3(x)
return xOr if you want to share the parameters but run this three times:
class MyModel(nn.Module):
def __init__(self, dim: nn.Dim):
super().__init__()
self.block = MyModelBlock(dim)
def __call__(self, x: nn.Tensor) -> nn.Tensor:
x = self.block(x)
x = self.block(x)
x = self.block(x)
return xWhen this is integrated as part of a Sisyphus recipe,
the common way people use it is similar as for i6_experiments,
i.e. you would git clone this repo into your recipe directory.
See i6_experiments.
Earlier, this was intended to be used for the RETURNN returnn.import_ mechanism.
See returnn #436 for initial import_ discussions.
See #2 for discussions on import_ usage here.
Note that this might not be the preferred usage pattern anymore but this is up to you.
Usage example for config:
from returnn.import_ import import_
test = import_("github.com/rwth-i6/returnn_common", "test.py", "20210602-1bc6822")
print(test.hello())You can also make use of auto-completion features in your editor (e.g. PyCharm).
Add ~/returnn/_pkg_import to your Python paths,
and use this alternative code:
from returnn.import_ import import_
import_("github.com/rwth-i6/returnn_common", ".", "20210602-1bc6822")
from returnn_import.github_com.rwth_i6.returnn_common.v20210302133012_01094bef2761 import test
print(test.hello())During development of a new feature in returnn-experiments,
you would use a special None placeholder for the version,
such that you can directly work in the checked out repo.
The config code looks like this:
from returnn.import_ import import_
import_("github.com/rwth-i6/returnn_common", ".", None)
from returnn_import.github_com.rwth_i6.returnn_common.dev import test
print(test.hello())You would also edit the code in ~/returnn/pkg/...,
and once finished, you would commit and push to returnn_common,
and then change the config to that specific version (date & commit).
These are the ideas behind the recipes. If you want to contribute, please try to follow them. (If something is unclear, or even in general, better speak with someone before you do changes, or add something.)
This is supposed to be simple.
Functions or classes can have some options
with reasonable defaults.
This should not become too complicated.
E.g. a function to return a Librispeech corpus
should not be totally generic to cover every possible case.
When it doesn't fit your use case,
instead of making the function more complicated,
just provide your alternative LibrispeechCustomX class.
There should be reasonable defaults.
E.g. just Librispeech() will give you some reasonable dataset.
E.g. maybe ~5 arguments per function is ok
(and each argument should have some good default),
but it should not be much more.
Better just make separate functions instead
even when there is some amount of duplicate code
(make_transformer which creates a standard Transformer,
vs make_linformer which creates a Linformer, etc.).
It should be simple to use functions
as basic building blocks to build sth more complex.
E.g. when you implement the Transformer model
(put that to models/segmental/transformer.py)
make functions make_trafo_enc_block
and make_trafo_encoder(num_layers=..., ...)
in models/encoder/transformer.py,
and then make_transformer_decoder and make_transformer
in models/segmental/transformer.py.
That makes parts of it easily reusable.
Break it down as much as it is reasonable.
The building blocks will naturally depend on each other.
In most cases, you should use relative imports
to make use of other building blocks,
and not import_.
Small files (e.g. vocabularies up to a certain size <100kb or so) could be directly put to the repository next to the Python files. This should be kept minimal and only be used for the most common files. (E.g. our Librispeech BPE vocab is stored.) The repository should stay small, so try to avoid this if this is not really needed.
For any larger files or other files,
the idea is that this can easily be used across different systems.
So there would be a common directory structure
in some directory which could be some symlinks elsewhere.
(We could also provide some scripts to simplify handling this.)
To refer to such a file path, use the functions in data.py.
Python 3.7+. See #43.
Recent RETURNN (>=2022), needs behavior version >=12.