tsalib 0.1.3

TSAlib: Support for Tensor Shape Annotations

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 shape assertions 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 printing 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)

Examples

The examples directory contains tsalib 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.

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))

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:{batch_size} EmbedDim:{embed_dim}}')
#declare a 2-D tensor of shape(48, 300)
x = torch.randn(B, D)
#assertions over dimension variables (code unchanged even if dim sizes change)
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 ', 'vp') #(v)iew, 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 backend.py).

See tests/test.py and tests/test_ext.py for complete examples of basic and extended usage.

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 : use view and transpose together. An inadvertent reshape may not preserve your dimensions (axes). Using view 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 for numpy, tensorflow and torch 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