pydec 0.2.0

Linear decomposition toolkit for neural network based on pytorch.

PyDec

PyDec is a linear decomposition toolkit for neural network based on PyTorch, which can decompose the tensor in the forward process into given components with a small amount of code. The result of decomposition can be applied to tasks such as attribution analysis.

Features:

• Fast. Compute decomposition in foward process and benefit from GPU acceleration.
• Real-time. Outputs attribution along with the model output results.
• Applicable to networks such as Transformer, CNN and RNN.

Examples

Attribution

Contribution Heat maps of the Roberta model (fine-tuned on SST-2). Warm colors indicate high contribution while cool colors indicate low contribution. The outputs of the model were positive, negative and positive, but the latter two samples did not match the labels.

Data flow visualization

Requirements and Installation

• PyTorch version >= 1.11.0
• Python version >= 3.7
• To install PyDec and develop locally:

git clone https://github.com/DoubleVII/pydec
cd pydec
pip install --editable ./

• To install the latest stable release:

pip install pydec

Getting Started

Example: deompose a tiny network

As a simple example, here’s a very simple model with two linear layers and an activation function. We’ll create an instance of it andand get the decomposition of the output by autotracing:

import torch

class TinyModel(torch.nn.Module):
    def __init__(self):
        super(TinyModel, self).__init__()

        self.linear1 = torch.nn.Linear(4, 10)
        self.activation = torch.nn.ReLU()
        self.linear2 = torch.nn.Linear(10, 2)

    def forward(self, x):
        x = self.linear1(x)
        x = self.activation(x)
        x = self.linear2(x)
        return x

tinymodel = TinyModel()

Given an input x, the output of the model is:

x = torch.rand(4)

print("Input tensor:")
print(x)

print("\n\nOutput tensor:")
print(tinymodel(x))

Out:

Input tensor:
tensor([0.7023, 0.3492, 0.7771, 0.0157])


Output tensor:
tensor([0.2751, 0.3626], grad_fn=<AddBackward0>)

To decompose the output, first wrap the model using pydec.autotracing.compile and then input the Composition initialized from x:

tinymodel = pydec.autotracing.compile(tinymodel)

c = pydec.zeros(x.size(), c_num=x.size(0))
c = pydec.diagonal_init(c, src=x, dim=0)

print("Input composition:")
print(c)

c_out = tinymodel(c)

print("\n\nOutput composition:")
print(c_out)

Out:

Input composition:
composition{
  components:
    tensor([0.7023, 0.0000, 0.0000, 0.0000]),
    tensor([0.0000, 0.3492, 0.0000, 0.0000]),
    tensor([0.0000, 0.0000, 0.7771, 0.0000]),
    tensor([0.0000, 0.0000, 0.0000, 0.0157]),
  residual:
    tensor([0., 0., 0., 0.])}


Output composition:
composition{
  components:
    tensor([-0.0418, -0.0296]),
    tensor([0.0566, 0.0332]),
    tensor([0.1093, 0.1147]),
    tensor([ 0.0015, -0.0018]),
  residual:
    tensor([0.1497, 0.2461]),
  grad_fn=<AddBackward0>}

Each component of the output composition represents the contribution of each feature in x to the output. Summing each component yields the tensor of original output:

print("Sum of each component:")
print(c_out.c_sum())

Out:

Sum of each component:
tensor([0.2751, 0.3626], grad_fn=<AddBackward0>)

To restore the ability of the model to forward tensor, use trace() to turn off autotracing:

tinymodel.trace(False)

print("Output tensor:")
print(tinymodel(x))

Out:

Output tensor:
tensor([0.2751, 0.3626], grad_fn=<AddBackward0>)

Documentation

The full documentation contains examples of implementations on realistic models, tutorials, notes and Python API.