image_classification¶
Image classification (ResNetv1-10 with CIFAR10)
Source code: image_classification.py
Pre-trained model: image_classification.mltk.zip
This was adapted from:
https://github.com/mlcommons/tiny/tree/master/benchmark/training/image_classification
https://github.com/SiliconLabs/platform_ml_models/blob/master/eembc/CIFAR10_ResNetv1/cifar10_main.py
Dataset¶
Model Topology¶
Performance (floating point model)¶
Accuracy - 86.2%
AUC - .989
Performance (quantized tflite model)¶
Accuracy - 86.1%
AUC - .988
Commands¶
# Do a "dry run" test training of the model
mltk train image_classification-test
# Train the model
mltk train image_classification
# Evaluate the trained model .tflite model
mltk evaluate image_classification --tflite
# Profile the model in the MVP hardware accelerator simulator
mltk profile image_classification --accelerator MVP
# Profile the model on a physical development board
mltk profile image_classification --accelerator MVP --device
Model Summary¶
mltk summarize image_classification --tflite
+-------+-----------------+-----------------+-----------------+-----------------------------------------------------+
| Index | OpCode | Input(s) | Output(s) | Config |
+-------+-----------------+-----------------+-----------------+-----------------------------------------------------+
| 0 | conv_2d | 32x32x3 (int8) | 32x32x16 (int8) | Padding:same stride:1x1 activation:relu |
| | | 3x3x3 (int8) | | |
| | | 16 (int32) | | |
| 1 | conv_2d | 32x32x16 (int8) | 32x32x16 (int8) | Padding:same stride:1x1 activation:relu |
| | | 3x3x16 (int8) | | |
| | | 16 (int32) | | |
| 2 | conv_2d | 32x32x16 (int8) | 32x32x16 (int8) | Padding:same stride:1x1 activation:none |
| | | 3x3x16 (int8) | | |
| | | 16 (int32) | | |
| 3 | add | 32x32x16 (int8) | 32x32x16 (int8) | Activation:relu |
| | | 32x32x16 (int8) | | |
| 4 | conv_2d | 32x32x16 (int8) | 16x16x32 (int8) | Padding:same stride:2x2 activation:relu |
| | | 3x3x16 (int8) | | |
| | | 32 (int32) | | |
| 5 | conv_2d | 16x16x32 (int8) | 16x16x32 (int8) | Padding:same stride:1x1 activation:none |
| | | 3x3x32 (int8) | | |
| | | 32 (int32) | | |
| 6 | conv_2d | 32x32x16 (int8) | 16x16x32 (int8) | Padding:same stride:2x2 activation:none |
| | | 1x1x16 (int8) | | |
| | | 32 (int32) | | |
| 7 | add | 16x16x32 (int8) | 16x16x32 (int8) | Activation:relu |
| | | 16x16x32 (int8) | | |
| 8 | conv_2d | 16x16x32 (int8) | 8x8x64 (int8) | Padding:same stride:2x2 activation:relu |
| | | 3x3x32 (int8) | | |
| | | 64 (int32) | | |
| 9 | conv_2d | 8x8x64 (int8) | 8x8x64 (int8) | Padding:same stride:1x1 activation:none |
| | | 3x3x64 (int8) | | |
| | | 64 (int32) | | |
| 10 | conv_2d | 16x16x32 (int8) | 8x8x64 (int8) | Padding:same stride:2x2 activation:none |
| | | 1x1x32 (int8) | | |
| | | 64 (int32) | | |
| 11 | add | 8x8x64 (int8) | 8x8x64 (int8) | Activation:relu |
| | | 8x8x64 (int8) | | |
| 12 | average_pool_2d | 8x8x64 (int8) | 1x1x64 (int8) | Padding:valid stride:8x8 filter:8x8 activation:none |
| 13 | reshape | 1x1x64 (int8) | 64 (int8) | BuiltinOptionsType=0 |
| | | 2 (int32) | | |
| 14 | fully_connected | 64 (int8) | 10 (int8) | Activation:none |
| | | 64 (int8) | | |
| | | 10 (int32) | | |
| 15 | softmax | 10 (int8) | 10 (int8) | BuiltinOptionsType=9 |
+-------+-----------------+-----------------+-----------------+-----------------------------------------------------+
Total MACs: 12.502 M
Total OPs: 25.122 M
Name: image_classification
Version: 1
Description: TinyML: Image classification - ResNetv1-10 with CIFAR10
Classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck
hash: d170adc21388920caa3f64ba22cd5b5d
date: 2022-02-04T19:20:20.062Z
runtime_memory_size: 53972
samplewise_norm.rescale: 0.0
samplewise_norm.mean_and_std: False
.tflite file size: 99.3kB
Model Profiling Report¶
# Profile on physical EFR32xG24 using MVP accelerator
mltk profile image_classification --device --accelerator MVP
Profiling Summary
Name: image_classification
Accelerator: MVP
Input Shape: 1x32x32x3
Input Data Type: int8
Output Shape: 1x10
Output Data Type: int8
Flash, Model File Size (bytes): 99.2k
RAM, Runtime Memory Size (bytes): 86.9k
Operation Count: 25.2M
Multiply-Accumulate Count: 12.4M
Layer Count: 16
Unsupported Layer Count: 0
Accelerator Cycle Count: 9.3M
CPU Cycle Count: 225.4k
CPU Utilization (%): 2.4
Clock Rate (hz): 78.0M
Time (s): 118.9m
Ops/s: 211.8M
MACs/s: 104.7M
Inference/s: 8.4
Model Layers
+-------+-----------------+--------+--------+------------+------------+----------+-------------------------+--------------+-----------------------------------------------------+
| Index | OpCode | # Ops | # MACs | Acc Cycles | CPU Cycles | Time (s) | Input Shape | Output Shape | Options |
+-------+-----------------+--------+--------+------------+------------+----------+-------------------------+--------------+-----------------------------------------------------+
| 0 | conv_2d | 933.9k | 442.4k | 583.4k | 30.6k | 7.5m | 1x32x32x3,16x3x3x3,16 | 1x32x32x16 | Padding:same stride:1x1 activation:relu |
| 1 | conv_2d | 4.8M | 2.4M | 1.8M | 30.1k | 22.3m | 1x32x32x16,16x3x3x16,16 | 1x32x32x16 | Padding:same stride:1x1 activation:relu |
| 2 | conv_2d | 4.7M | 2.4M | 1.8M | 30.0k | 22.3m | 1x32x32x16,16x3x3x16,16 | 1x32x32x16 | Padding:same stride:1x1 activation:none |
| 3 | add | 16.4k | 0 | 57.4k | 9.2k | 840.0u | 1x32x32x16,1x32x32x16 | 1x32x32x16 | Activation:relu |
| 4 | conv_2d | 2.4M | 1.2M | 883.8k | 15.3k | 11.2m | 1x32x32x16,32x3x3x16,32 | 1x16x16x32 | Padding:same stride:2x2 activation:relu |
| 5 | conv_2d | 4.7M | 2.4M | 1.6M | 30.2k | 21.0m | 1x16x16x32,32x3x3x32,32 | 1x16x16x32 | Padding:same stride:1x1 activation:none |
| 6 | conv_2d | 270.3k | 131.1k | 131.2k | 5.3k | 1.7m | 1x32x32x16,32x1x1x16,32 | 1x16x16x32 | Padding:same stride:2x2 activation:none |
| 7 | add | 8.2k | 0 | 28.7k | 5.3k | 450.0u | 1x16x16x32,1x16x16x32 | 1x16x16x32 | Activation:relu |
| 8 | conv_2d | 2.4M | 1.2M | 830.3k | 15.3k | 10.5m | 1x16x16x32,64x3x3x32,64 | 1x8x8x64 | Padding:same stride:2x2 activation:relu |
| 9 | conv_2d | 4.7M | 2.4M | 1.5M | 30.2k | 19.1m | 1x8x8x64,64x3x3x64,64 | 1x8x8x64 | Padding:same stride:1x1 activation:none |
| 10 | conv_2d | 266.2k | 131.1k | 115.4k | 5.3k | 1.5m | 1x16x16x32,64x1x1x32,64 | 1x8x8x64 | Padding:same stride:2x2 activation:none |
| 11 | add | 4.1k | 0 | 14.4k | 3.4k | 240.0u | 1x8x8x64,1x8x8x64 | 1x8x8x64 | Activation:relu |
| 12 | average_pool_2d | 4.2k | 0 | 2.2k | 3.9k | 60.0u | 1x8x8x64 | 1x1x1x64 | Padding:valid stride:8x8 filter:8x8 activation:none |
| 13 | reshape | 0 | 0 | 0 | 769.0 | 30.0u | 1x1x1x64,2 | 1x64 | Type=none |
| 14 | fully_connected | 1.3k | 640.0 | 1.0k | 2.1k | 30.0u | 1x64,10x64,10 | 1x10 | Activation:none |
| 15 | softmax | 50.0 | 0 | 0 | 8.2k | 120.0u | 1x10 | 1x10 | Type=softmaxoptions |
+-------+-----------------+--------+--------+------------+------------+----------+-------------------------+--------------+-----------------------------------------------------+
Model Diagram¶
mltk view image_classification --tflite
