xwenKWS inference on chip
This project focuses on implementing Keyword Spotting (KWS) inference on a chip. Keyword Spotting Task is a classification Task: KWS aims to classify input signals to detect the presence or absence of specific keywords, such as "Hey Snips," in real-time.
Our Scope
General KWS Pipeline: The pipeline includes stages from the initial speech signal to the final decision.
Our Scope: We focus on the stages of the keyword spotting acoustic model right after feature extraction, ending just before posterior handling.
KWS Model with Dilated CNN
Data
We use the "Hey Snips" dataset, which includes various voice commands from different speakers, providing a robust dataset for training. This dataset contains thousands of audio samples of different keywords, recorded by multiple speakers, including background noise and variations in speech patterns.
Feature Extraction
Filter Bank: We use filter bank feature extraction to transform raw speech signals into a format suitable for the model. Filter banks decompose the signal into different frequency components, highlighting relevant features for speech recognition .After the filter bank, we padding the data so that it comes out the same size in each dimension.
Structure
The model consists of multiple layers: CMVN (Cepstral Mean and Variance Normalization), linear layers, ReLU activations, and dilated CNN blocks (repeated four times). Each dilated CNN block is followed by batch normalization and ReLU activations. The final layers include a linear layer and a sigmoid activation for binary classification.
Why Dilated CNN
Dilated Convolutional Neural Networks (CNNs) are chosen for their ability to handle temporal dependencies in speech signals. They capture features over larger time scales without significantly increasing computational load, making them ideal for real-time keyword spotting on limited-resource hardware.
Working process:
Model Training:
We train the deep learning model using the prepared dataset with Pytorch. This involves data preprocessing, model selection, and optimization to ensure high accuracy in keyword detection.
Implementation in Python with pure tensor: Rewrite and implement the trained model in a Python environment, ensuring it is free from external library dependencies. This makes integration with hardware more straightforward.
Weight Extraction:
Extract the model weights after the Python implementation. These weights are crucial for deploying the model on hardware.
Chats with 🤖:
We started with general questions about KWS. The prompt diverged to software and hardware design. AI in this case, not only act as a co-designer, but also a teacher who explains the fundamentals of IC design for me. First, a rough plan of modules was determined, then progress to designing the ISA and Finite State Machine. For each module, there were multiple attempts to generate the file, the first prompt and answer determines a big part of the quality of the code. The best way to generate a high quality code is to give it consice directions in the first place, with ISA, architecute and the python code all together. Another finding is, to give back the LLM it's generated code, or the code we implemented after chats, can be used to produce a better prompt, which can be used for the new chat. Each iteration(try) of fresh new version was labeled (both chat and verilog file), can be viewed in comparison. We use the LLM Claude to successfully create CMVN, ReLU, Linear module, with Verilog code and a python testbench in Cocotb. The dilated CNN and Sigmoid modules wasn't succesfully tested with it's calculation logic. They were not hardened. The most challenging part of this excersice was to generate a Linear module that performs linear transformation, in our case, a matrix multiplication of [50x20] x [20x20].This is the part where I found out LLMs don't really 'understand'(details at the bottom of the file).
How to read the Chats with Claude:
It's numbered! There isn't an export chat button for Claude so I pasted them in text files. Please find all the attachments I used to talk with Claude, including our lib free model, parameters, and a example prompt to generate new modules I developed over time in a seperate folder. I have pasted our ISA and FSM into a test file too so it can have a look during generating the new modules.
Prompt file directory
You can find everything I used to generate code from Claude. There's a subdirectory with the trained models in python. Inference_0508 is our first version with quantization, Inference_0510 is our second version with a input matrix size of 300x20 (300frames x 20 features), Inference_0519 is our final version with a input matrix size of 50x20. They were in html version because it's easier to check with browser, but we do have a jupyter notebook version attached.
Why Lib-free KWS?
The original plan was simple: train the model, feed into Claude, ask it to write the code, boom. Unfortunatly it didn't work. PyTorch's modules are encapsulated at too high level. PyTorch is a popular deep learning framework that provides high-level abstractions and modules for building neural networks. These abstractions and modules are designed to simplify the development process and make it easier to create complex models. However, the high-level nature of PyTorch's modules made it challenging for Language Models (LLMs) to understand the underlying computational logic. Writing it in a lib-free manner is to facilitate LLM's understanding of the computational logic and convert it into Verilog code for the KWS model, this means avoiding the use of high-level PyTorch modules and instead implementing the model using lower-level operations and primitives. By writing the model in a lib-free way, the computational logic becomes more explicit and transparent. This can help LLMs to better understand the step-by-step operations and data flow within the model.
Quantization?
We took a lot of radical paths to try things out. Quantization is one of them. The total weights of our first draft was around 400kB- pretty much impossible to fit in Caravel user area as register files. The first thing that came up to our mind was to quantize the model from float 32 to int 8. Reasons for not using quantization:
1/.PyTorch does not convert all data types to int8.
2/.During the quantization process, PyTorch modifies the model structure, making it impossible for us to extract (and implement as lib-free).
3/.The accuracy of the model decreases after quantization:our model has a lot weights data that as small as 10^-8, int 8 doesn't provide enough range and granularity for the weights, a lot of weights became 0 if we convert them to int8. In fact, in later stage, we tried to convert the weights to int 16 and the granularity was still not enough. (It's calculated with 2^-x, x is the number of bits you reserved for the fractional parts) Now let me explain a bit more on point 2: When applying quantization to a PyTorch model, the framework modify the model's structure to accommodate the quantization process by adding quantization-specific layers and operations, such as quantization and dequantization nodes. That means the even though during the calculation operation data can be in int 8 or int 16, after the operation and before the next operation they were convert back to float 32. This method ensures a smaller model and fast processing speed but doesnt solve our problem of having massive weights information. Also, quantization and dequantization nodes highly dependant on pytorch encapsulation- consider limited time we had, we gave up unwrapping Quantization nodes to Lib-free python and decided to focus on preserving the original model's structure and accuracy without Quantization.
Implementation in Verilog:
Translate the Python implementation and extracted weights into Verilog for hardware implementation. This step involves designing the hardware description language suitable for chip design. I ran into some unexpected diffulties too, such as error from Efabless SRAM or rookie mistakes like adapoting a Full-Wrapper Flattening strategy but ran out of RAM on my own computer.
By implementing KWS inference on a chip, we aim to create a low-latency, efficient system capable of recognizing voice commands in real-time. This project leverages the latest advancements in deep learning and practical hardware implementation techniques to develop responsive and intelligent voice-controlled devices.
Hardware modules
Overview
This is a simple outline for our RISC-V based KWS accelerator.
Fixed-point arithmetic is adopted to simplify the hardwar design. To preserve the the fractional part of the data, we converted our weights and inputs to 32 bits fixed-point 1.7.24 format.That is 1 Sign bit, 7 bits for the integer part, 24 bits for the fractional part.
The input features and weights are initialized on the SRAM inside the user project area.
The major components include
1. A finite state machine as a controller for fetching instructions and orchestrating other modules
(CNN component below)
2.CMVN pool Cepstral mean and variance normalization pool. The input data is normalized and flattened here.
3.Linear module is adopted to compute matrix multiplications between two matrices of 50x20 and 20x20 for linear transformation.
4.ReLU function module to perform ReLU function
5.Dilated CNN module to perform CNN with systolic array (to be tested)
6.Batch normalize module to be tested)
7.Sigmoid Activation module (to be tested)
Results for the hardware modules
The hardened modules, CMVN(normalising pool), ReLU, Linear, are all successfully tested before hardening. The Linear module was rewrite last minute with a register file because I couldn’t figure out how to correctly incorporate the Efabless SRAM 😅. Rest of the calculation modules will have their Verilog code provided, on the roadmap I would like to finish in later times. I have attached my attempt and code with each attempt.
Roadmap
Dilated CNN module to perform CNN with systolic array (to be tested)
Batch normalize module to be tested)
Sigmoid Activation module (to be tested)
Contributing
Xiaoyu Wen -- https://github.com/aaalce --www.linkedin.com/in/xiaoyu-wen-4b5974165
Hongzhi Hou --https://github.com/hoho1st --https://www.linkedin.com/in/hongzhi-hou-033b32186
Jian Sun --https://github.com/jsun1006
Feel free to contact us if you have any questions!
Some personal thoughts
In 2023 me and my friend(Hongzhi Hou, the other author of the project) was stunned by power of LLM that transformed human learning. We decided to each learn a random skill with LLM from scratch as a fun experiment. In March 2024 I came across this challenge on Efabless, hence picking IC design, as my goal. It is something I was alway curious about but felt intimidated due to its complexity. The challenge provided a perfect start with real project and goals, it also accelerated my learning by giving it a time constraint with pressure. The real difficulties from this exercise was to decide, which direction i should go, what shall I focus on when everything is unknown.
Throughout the journey, I have a few findings about LLM.
1 LLM might be perfect for debugging and spotting details we miss, it lacks of logic when a real concept needs to be applied. One of the things I spent the most time doing was to explain to Claude the calculation logic, such as what Matrix Multiplication is, and how shall it be done in a data stream like ours. In the beginning I wasn’t aware it doesn’t truly ‘understand’, because it can explain matrix multiplication in perfect sentences. After realising its shortcomings, I have spent substantial amount of time, giving examples to it how the calculation and data selection logic should work, Claude seemed to understand after some examples, and gave back the correct answers if I ask in hypothetical terms, but it would be unable to produce the correct Verilog logic, unable to match everything on the correct cycle, unless I explicitly stated. Nor spot the apparent mistakes such as the wrong dimensions of the output matrix.
2 LLM is pretty bad as maths. Basic arithmetic. I wouldn’t trust their answers if the numbers get big or equations gets long.
3 They get stuck in loops sometimes, so you need to either feed it some completely new information or open a new chat.
4 They will give you relatively similar answers, or designs in this case, for a KWS accelerator, across several LLM. Unless you start to read papers or feed it information elsewhere, they won’t give you anything ‘new’. Therefore I’m skeptical about its ability to create- or let’s say, innovate.
Acknowledgments
| :exclamation: Important Note |
|---|
Please fill in your project documentation in this README.md file
Refer to README for a quickstart of how to use caravel_user_project
Refer to README for this sample project documentation.
Refer to the following readthedocs for how to add cocotb tests to your project.
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [yyyy] [name of copyright owner]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
