Team NNoM submission for TinyML Contest (2022 ACM/IEEE TinyML Design Contest at ICCAD)
Contest result is out. NNoM ranks 10 out of 41 teams, overall a great result considering the limited time we put into the contest and the risks we took.
- We took a huge risk by using recurrent layers, which is not easy to train with fixed-point.
- NNoM framework took extra memories compared to CUBEMX-MX(not too much but have affect the scoring) because its extra evaluation tools and more layer supports.
Seeing the top ranking, there are indeed tricks can applied. For example, keeping the model small (inference time short + small memory) while losing some F_beta can still keep the final score. I am very satisfied with the result as this was the first contest we paticipated. Looking forward for the next one :-)
This model consist of a few 1D-Convolution layer (Inception structure) and 1 GRU recurrent layer.
The total trainable parameters is just under 2k a tiny model that can achieve 95+ score. The model is quantised into 8-bit ops and accelerated with CMSIS-NN backend. Can inference one test data in 34ms @ 80MHz Cortex M4, with the complexity only 220k MAC.
Code size
Program Size: Code=29944 RO-data=6376 RW-data=24 ZI-data=52784
Please skip if you are not willing to train a new model.
-
Copy
tinyml_contest_data_trainingto this folder. -
Run
dataset.py -
Run
main.pyto train, quantise and covert the mode to NNoM model file.
Tensorflow model will be generated under model folder.
NNoM model weights_contest.h will be generated under this folder.
Once we have the weights_contest.h generated, go to tinyml-contest-mcu, open the Keil project file, compile and download to the NUCLEO board.
Then run validation.py from the contest repository to validate the result.
Install Visual Studio, open the contest.sln to run it from VS.
OR
Install scons and GCC or msvc, then run scons in this folder.
During model conversion, the dataset is also exported, the pc program will read them and do validation on nnom model.
Due to the validation of the contest uses serial port to transmit data, the log of NNoM cannot be printed on the console. But we can run the same model from PC and check the memory and complexity of the model.
For example:
validation size: 7036882
Model version: 0.4.3
NNoM version 0.4.3
To disable logs, please void the marco 'NNOM_LOG(...)' in 'nnom_port.h'.
Data format: Channel last (HWC)
Start compiling model...
Layer(#) Activation output shape ops(MAC) mem(in, out, buf) mem blk lifetime
-------------------------------------------------------------------------------------------------
#1 Input - - ( 1,1250, 1,) ( 1250, 1250, 0) 4 - - - - - - -
#2 Conv2D - ReLU - ( 1, 625, 8,) 25k ( 1250, 5000, 0) 3 1 - - - - - -
#3 MaxPool - - ( 1, 156, 8,) ( 5000, 1248, 0) 2 1 1 1 - - - -
#4 Conv2D - ReLU - ( 1, 625, 4,) 22k ( 1250, 2500, 0) 2 1 - 1 - - - -
#5 MaxPool - - ( 1, 156, 4,) ( 2500, 624, 0) 1 1 1 1 1 - - -
#6 Conv2D - ReLU - ( 1, 625, 4,) 32k ( 1250, 2500, 0) 1 1 - 1 1 - - -
#7 MaxPool - - ( 1, 156, 4,) ( 2500, 624, 0) 1 1 1 1 1 - - -
#8 Concat - - ( 1, 156, 16,) ( 2496, 2496, 0) 1 - 1 1 1 - - -
#9 Conv2D - ReLU - ( 1, 156, 4,) 29k ( 2496, 624, 0) 1 1 - - - - - -
#10 MaxPool - - ( 1, 78, 4,) ( 624, 312, 0) 1 1 1 - - - - -
#11 RNN/GRU - - ( 78, 12, ) 111k ( 312, 936, 224) 1 1 1 - - - - -
#12 Flatten - - ( 936, ) ( 936, 936, 0) - 1 - - - - - -
#13 Dense - Sigmoid - ( 1, ) 936 ( 936, 1, 1872) 1 1 1 - - - - -
#14 Output - - ( 1, ) ( 1, 1, 0) - - 1 - - - - -
-------------------------------------------------------------------------------------------------
Memory cost by each block:
blk_0:2496 blk_1:5000 blk_2:624 blk_3:1248 blk_4:624 blk_5:0 blk_6:0 blk_7:0
Memory cost by network buffers: 9992 bytes
Total memory occupied: 15024 bytes
Processing 2%
Processing 4%
Processing 6%
Processing 9%
Processing 11%
Processing 13%
Processing 15%
Processing 18%
Processing 20%
Processing 22%
Processing 25%
Processing 27%
Processing 29%
Processing 31%
Processing 34%
Processing 36%
Processing 38%
Processing 40%
Processing 43%
Processing 45%
Processing 47%
Processing 50%
Processing 52%
Processing 54%
Processing 56%
Processing 59%
Processing 61%
Processing 63%
Processing 65%
Processing 68%
Processing 70%
Processing 72%
Processing 75%
Processing 77%
Processing 79%
Processing 81%
Processing 84%
Processing 86%
Processing 88%
Processing 91%
Processing 93%
Processing 95%
Processing 97%
Processing 100%
Prediction summary:
Test frames: 0
Test running time: 0 sec
Model running time: 0 ms
Print running stat..
Layer(#) - Time(us) ops(MACs) ops/us
--------------------------------------------------------
#1 Input - 0
#2 Conv2D - 0 25k
#3 MaxPool - 0
#4 Conv2D - 0 22k
#5 MaxPool - 0
#6 Conv2D - 0 32k
#7 MaxPool - 0
#8 Concat - 0
#9 Conv2D - 0 29k
#10 MaxPool - 0
#11 RNN/GRU - 0 111k
#12 Flatten - 0
#13 Dense - 0 936
#14 Output - 0
Summary:
Total ops (MAC): 222272(0.22M)
Prediction time :0us
Total memory:15024
Top 1 Accuracy on Tensorflow 96.57%
Top 1 Accuracy on NNoM 96.41%
Process finished with exit code 0
More information will be updated on https://github.com/majianjia/tinyml-contest and https://github.com/majianjia/nnom