TSAlib: Support for Tensor Shape Annotations
Project description
Tensor Shape Annotations Library (tsalib)
Writing deep learning programs which manipulate multi-dimensional tensors (numpy
, pytorch
, keras
, tensorflow
, ...) requires you to carefully keep track of shapes of matrices/tensors. The Tensor Shape Annotation (TSA) library enables you to write first-class, library-independent, shape expressions over dimension variables to model matrix/tensor variable shapes.
TSAs enable us to label and verify tensor variables shapes as well as write more fluent shape transformations and tensor operations. Using TSAs enhances code clarity, accelerates debugging and improves overall developer productivity when writing tensor programs.
Detailed article here.
See Changelog here.
Introduction
Carrying around the tensor shapes in your head gets increasingly hard as programs become more complex, e.g., reshaping before a matmult
, examining/modifying deep pre-trained architectures (resnet
, densenet
, elmo
), designing new kinds of attention
mechanisms (multi-head attention
) or when creating a new RNN
cell. There is no principled way of shape specification and tracking inside code -- most developers resort to writing adhoc comments embedded in code to keep track of tensor shapes (see code from google-research/bert).
tsalib
comes to our rescue here. It allows you to write shape expressions over dimension variables describing the shape of tensor variables. These expressions can be used in multiple ways:
- as first-class annotations of tensor variables,
- to write
symbolic
shapeassert
ions and tensor constructors - to specify shape transformations (
reshape
,permute
,expand
) or tensor product operations (matmult
) succinctly.
TSAs expose the typically invisible tensor shape types, leading to improved productivity across the board.
Shape annotations/assertions turn out to be useful in many ways.
- They help us to quickly verify the variable shapes when writing new transformations or modifying existing modules.
- Assertions and annotations remain the same even if the concrete dimension lengths change.
- Faster debugging: if you annotate-as-you-go, the tensor variable shapes are explicit in code, readily available for a quick inspection. No more adhoc shape
print
ing when investigating obscure shape errors. - Do shape transformations using shorthand notation and avoid unwanted shape surgeries.
- Use TSAs to improve code clarity everywhere, even in your machine learning data pipelines.
- They serve as useful documentation to help others understand or extend your module.
Dimension Variables
Tensor shape annotations (TSAs) are constructed using dimension
variables --B
(Batch), C
(Channels), D
(EmbedDim) -- and arithmetic expressions (B*2
, C+D
) over them. Using tsalib
, you can define dimension variables customized to your architecture/program.
TSAs may be be represented as
- a tuple
(B,H,D)
[long form] - a string
'b,h,d'
(compact notation) (or simply'bhd'
) - a string with anonymous dimensions (
',h,'
is a 3-d tensor)
Here is an example snippet which uses TSAs in a pytorch
program to define, transform and verify tensor shapes. TSAs work seamlessly with arbitrary tensor libraries: numpy
, pytorch
, keras
, tensorflow
, mxnet
, etc.
from tsalib import dim_vars as dvs
from tsalib import permute_transform as pt
#declare dimension variables
B, C, H, W = dvs('Batch:32 Channels:3 Height:256 Width:256')
...
# create tensors using dimension variables (interpret dim vars as integers)
x: (B, C, H, W) = torch.randn(B, C, H, W)
# perform tensor transformations
x: (B, C, H // 2, W // 2) = maxpool(x)
# check symbolic assertions over TSAs, without knowing concrete shapes
assert x.size() == (B, C, H // 2, W // 2)
# super convenient reshapes!
x1 = x.view ((B,C, (H//2)*(W//2)))
assert x1.size() == (B, C, (H//2)*(W//2))
# permute using shorthand (einsum-like) notation,
# with anonymous dimensions
x: (B, C, H, W)
x1 = x.permute(pt(',c,,', ',,,c'))
assert x1.size() == (B, H, W, C)
# sequence of multiple transformations inline
# here: a sequence of a permute ('p') and view ('v') transformations
y = warp(x1, 'bhwc -> bchw -> b*c,h,w', 'pv')
assert y.size() == (B*C,H,W)
Installation
pip install [--upgrade] tsalib
Getting Started
See tests/test.py and tests/test_ext.py for complete examples of basic and extended usage.
from tsalib import dim_var as dv, dim_vars as dvs, dim_vars_shape as dvs2
import numpy as np
Declare Dimension Variables
#or declare dim vars with default integer values (optional)
B, C, D, H, W = dvs('Batch:48 Channels:3 EmbedDim:300 Height Width')
#or provide optional *shorthand* names for dim vars, default values
B, C, D, H, W = dvs('Batch(b):48 Channels(c):3 EmbedDim(d):300 Height(h) Width(w)')
# switch from using config constants to using dimension vars
B, C, D = dvs('Batch(b):{0} Channels(c):{1} EmbedDim(d):{2}'.format(config.batch_size, config.num_channels, config.embed_dim))
# TSAs are tuples over dimension variables
S1 = (B, C, D)
# we can always verify TSAs against concrete shapes
assert S1 == (48, 3, 300)
Use Dimension Variables to declare Tensors
Instead of scalar variables batch_size
, embed_dim
, use dimension variables B
, D
uniformly throughout your code.
B, D = dvs('Batch:48 EmbedDim:300')
#declare a 2-D tensor of shape(48, 300)
x = torch.randn(B, D)
#assertions over dimension variables (not exact values)
assert x.size() == (B, D)
Use TSAs to annotate variables on-the-go (Python 3)
a: (B, D) = np.array([[1., 2., 3.], [10., 9., 8.]]) #(Batch, EmbedDim): (2, 3)
b: (2, B, D) = np.stack([a, a]) #(2, Batch, EmbedDim): (2, 2, 3)
Arithmetic over dimension variables is supported. This enables easy tracking of shape changes across neural network layers.
v: (B, C, H, W) = torch.randn(B, C, h, w)
x : (B, C * 2, H//2, W//2) = torch.nn.conv2D(C, C*2, ...)(v)
Use TSAs to make matrix operations compact and explicit
Avoid explicit shape computations for reshaping
.
#use dimension variables directly
x = torch.ones(B, T, D)
x = x.view(B, T, 4, D//4)
In general, use tsalib.view_transform
to specify view changes declaratively.
x = np.ones((B, T, D))
from tsalib import view_transform as vt
#or, compact form:
x = x.reshape(vt('btd', 'b,t,4,d//4', x.shape)) #(20, 10, 300) -> (20, 10, 4, 75)
#or, super-compact, using anonymous dimensions:
x = x.reshape(vt(',,d', ',,4,d//4', x.shape))
Similarly, use tsalib.permute_transform
to compute permutation index order (no manual guess-n-check) from a declarative spec.
# long form:
perm_indices = permute_transform(src=(B,T,D,K), to=(D,T,B,K)) #(2, 1, 0, 3)
x = x.transpose(perm_indices) #(10, 50, 300, 30) -> (300, 50, 10, 30)
from tsalib import permute_transform as pt
#or, compactly:
x = x.transpose(pt('btdk', 'dtbk'))
#or, super-compact:
x = x.transpose(pt('b,,d,', 'd,,b,'))
Use dimension names instead of cryptic indices in reduction (mean
, max
, ...) operations.
from tsalib import drop_dims as dd
b: (2, B, D)
c: (2, D) = np.mean(b, axis=dd('2bd->2d')) #axis = 1
Sequence of shape transformations: warp
operator
The warp
operator allows squeezing in multiple shape transformations in a single line using the shorthand notation. The operator takes in 3 inputs, an input tensor, a sequence of shape transformations, and the corresponding transform types (view transform -> 'v', permute transform -> 'p').
x: 'btd' = torch.randn(B, T, D)
y = warp(x, 'btd -> b,t,4,d//4 -> b,4,t,d//4 ', 'vvp') #2 (v)iew transforms, then (p)ermute transform
assert(y.shape == (B,4,T,D//4))
Because it returns transformed tensors, the warp
operator is backend library-dependent. Currently supported backends are numpy
, tensorflow
and pytorch
. New backends can be added easily.
See tests/test.py and tests/test_ext.py for complete examples of basic and extended usage.
Examples
The examples directory contains TS annotations of a few well-known, complex neural architectures: Resnet, OpenAI Transformer. With TSAs, we can gain deeper and immediate insight into how the module works by scanning through the forward
function.
Dependencies
sympy
. A library for building symbolic expressions in Python.
Tested with Python 3.6. For writing type annotations inline, Python >= 3.5 is required.
Python >= 3.5 allows optional type annotations for variables. These annotations do not affect the program performance in any way.
Best Practices
- Convert all relevant config parameters into dimension variables. Use only the latter in your code.
- Avoid using
reshape
: useview
andtranspose
together. An inadvertentreshape
may not preserve your dimensions (axes). Usingview
to change shape protects against this: it throws an error if the dimensions being manipulated are not contiguous.
References
- Blog article introducing TSA.
- A proposal for designing a tensor library with named dimensions from ground-up. The TSA library takes care of some use cases, without requiring any change in the tensor libraries.
- Pytorch Issue on Names Axes here.
- Using einsum for tensor operations improves productivity and code readability. blog
- The Tile DSL uses indices ranging over dimension variables to write compact, library-independent tensor operations.
- The datashape library introduces a generic type system and grammar for structure data.
tsalib
focuses on shapes of homogeneous tensor data types only, with arithmetic support.
Contributors
Nishant Sinha, OffNote Labs. @medium, @twitter
Change Log
- [18 Nov 2018] Support for
warp
,drop_dims
. Backend modules fornumpy
,tensorflow
andtorch
added. - [9 Nov 2018] Support for shorthand notation in view/permute/expand transforms.
- [9 Nov 2018] Support for using TSA in assertions and tensor constructors (cast to integers).
- [25 Oct 2018] Initial Release
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.