torch-dag 0.1.3

torch-dag is a repository in which we implement a graph-like representation for torch models so that we have a unified structure for every model.

1. What is torch-dag?

torch-dag is a repository in which we implement a graph-like representation for torch models so that we have a unified structure for every model. This structure is a directed acyclic graph (DAG), which we call a DagModule. We do this so that:

1. We can easily run graph modification algorithms on the whole model without the need to edit any code defining the model. For example:

• we can easily switch all activation functions in the model with just a couple lines of code
• we can easily insert new vertices (modules) into the model based on some predefined pattern, e.g., add batch norm layers after every convolution provided there is none present immediately after the convolution
• we can introduce hierarchy into the model by wrapping some parts of the graph -> this is very useful if one wants to run some Neural Architecture Search algorithms

2. We can build a model once in the plain pytorch way, then convert it to a DagModule, save it and load it without having to have the source code for model building. This is extremely useful - imagine all the trouble one has to go through when one needs to modify the model source code manually to some simple stuff like changing some of the model layers.

The conversion between a regular torch.nn.Module instance and DagModule uses torch.FX. We decide not to use torch.FX graph representation directly for the following reasons:

• there is no serialization/deserialization for torch.FX GraphModule
• the graph generated by torch.FX is often not a computational deep learning graph one would like to work with since it contains a huge number of purely python operations as nodes
• there is no notion of hierarchy/nesting in GraphModule
• there is often a number of ways to represent the same operation in torch.FX GraphModule (think of addition as x+y, torch.add(x, y), which would have a very different representation in GraphModule), which leads to plenty of issues when one tries to implement some method to modify graph based on the presence of additions -> one has to think of all the different ways in which addition can be implemented in torch!

Note: Not every torch.nn.Module model can be converted to a DagModule! We mostly share the limitations of torch.FX in this respect, although we try to overcome that by explicitly considering some torch.FX-unfriendly modules as atomic modules of our representation.

2. Why do we need a unified DAG-like representation?

The short answer is to run algorithms for stuff like model compression. In torch-dag we implement a channel pruning algorithm that aspires to be architecture agnostic and we try really hard to make it work for as many architectures as possible. The exhaustive list of the timm models we support is given here.

3. Installation

3.1 Installation from cloned source code

In the cloned project's root (where this README.md file is) run

pip install -e .

3.2 Installation from PyPI

pip install torch-dag

4. Basics

If you have a toch.nn.Module model and you want to convert it to a DagModule simply run:

import torch_dag as td
dag_model = td.build_from_unstructured_module(model)

For details and more extended documentation see How to convert torch.nn.Module instances to DagModule?

5. Model compression algorithms

5.1 Channel-pruning

A jupyter notebook with a toy example of channel pruning can be viewed here. If you want to read a more detailed intro to channel pruning with torch-dag havce a look at pruning readme.

This is the algorithm we spent plenty of time developing and refining. It helped us to WIN Mobile AI 2022 Single-Image Depth Estimation on Mobile Devices.

Supported Models and Limitations

---

Current model coverage timm==0.9.5

	num models	percentage
all models	945	100.00%
convertible models	690	73.02%
channel prunable_models	585	61.90%

---

At this point we support plenty of convolutional models and a subset of vision transformer architectures. A full list of supported timm models and a proportion of FLOPS that can be removed in each model can be seen channel pruning supported models.

We do NOT support models that cannot be traced using torch.FX (there are notable exceptions, that require additional processing like vit* and deit3* models, which are supported) see limitations-of-symbolic-tracing. If you have a custom model that can be traced:

my_module = Model()
from torch.fx import symbolic_trace
# Symbolic tracing frontend - captures the semantics of the module
symbolic_traced : torch.fx.GraphModule = symbolic_trace(my_module)

then chances are it will be convertible to a DagModule and some part of it can be pruned! How much precisely?

Results in ImageNet1k for some timm models.

NOTE: as a proxy for model size we often use FLOPs normalized by input resolution expressed in thousands (kmapp for short - thousands of FLOPs per pixel), i.e.: kmapp(model) = FLOPs(model) / (H * W * 1000), where (H, W) is input resolution.

---

fbnetv3_g.ra2_in1k

FBNETV3 is a highly optimized architecture so further slimming down of it is challenging. Nonetheless, the results are still pretty good. We prune fbnetv3g and then fine-tune the pruned model for 200 epochs.

Results on ImageNet1K validation set.

model	acc (224x224)	acc (240x240)	GFLOPs (224x224)	params (m)	FLOPs reduction
fbnetv3g	0.8061	0.8132	2.14	16.6
fbnetv3d	0.7856	0.7927	1.04	10.3
fbnetv3b	0.7812	0.7871	0.845	8.6
m16	0.7742	0.7793	0.799	7.8	↓ 62.5%
m22	0.7920	0.7955	1.2	10.5	↓ 43.9%
m28	0.79686	0.8016	1.396	11.6	↓ 34.8%
m32	0.803	0.8025	1.61	13.1	↓ 24.8%

convnextv2_nano.fcmae_ft_in22k_in1k

ConvNextV2 is a purely convolutional model that is inspired by the design of Vision Transformers. We prune convnextv2_nano.fcmae_ft_in22k_in1k and then fine-tune the pruned model for 200 epochs.

Results on ImageNet1K validation set.

model	acc (224x224)	GFLOPs	params (m)	FLOPs reduction
baseline	0.8197	4.91	15.6
m0.5	0.8155	2.64	9.4	↓ 46.2%
m0.3	0.7922	1.68	6.3	↓ 65.8%
m0.2	0.7531	0.93	3.8	↓ 81.0%

To see much more results have a look here

TODO Block-pruning

TODO Learnable Low Rank Compression (LLRC)