File size: 12,301 Bytes
29d6673
 
9d7668d
29d6673
 
914e545
29d6673
c5ae9b8
29d6673
 
 
1a33e82
29d6673
7d99418
b86b199
29d6673
 
7d99418
74cf852
7d99418
 
 
 
 
 
 
 
 
 
 
 
74cf852
 
 
 
 
 
7d99418
 
 
 
 
 
74cf852
7d99418
 
 
 
 
 
74cf852
7d99418
 
 
 
 
 
 
9d58c9b
7d99418
 
9d58c9b
7d99418
 
 
 
9fef77b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8515975
29d6673
f0061b4
 
8515975
29d6673
 
 
 
 
cfb899a
29d6673
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
---
library_name: pytorch
license: bsd-3-clause
tags:
- backbone
- bu_auto
- android
pipeline_tag: image-classification

---

![](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/resnet50/web-assets/model_demo.png)

# ResNet50: Optimized for Qualcomm Devices

ResNet50 is a machine learning model that can classify images from the Imagenet dataset. It can also be used as a backbone in building more complex models for specific use cases.

This is based on the implementation of ResNet50 found [here](https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py).
This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.66.0/src/qai_hub_models/models/resnet50) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary).

Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) to run these models on a hosted Qualcomm® device.

## Getting Started
There are two ways to deploy this model on your device:

### Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

| Runtime | Precision | Chipset | SDK Versions | Download |
|---|---|---|---|---|
| ONNX | float | Universal | QAIRT 2.50, ONNX Runtime 1.30.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/resnet50/releases/v0.66.0/resnet50-onnx-float.zip)
| ONNX | w8a8 | Universal | QAIRT 2.50, ONNX Runtime 1.30.0 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/resnet50/releases/v0.66.0/resnet50-onnx-w8a8.zip)
| QNN_DLC | float | Universal | QAIRT 2.50 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/resnet50/releases/v0.66.0/resnet50-qnn_dlc-float.zip)
| QNN_DLC | w8a8 | Universal | QAIRT 2.50 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/resnet50/releases/v0.66.0/resnet50-qnn_dlc-w8a8.zip)
| TFLITE | float | Universal | QAIRT 2.50 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/resnet50/releases/v0.66.0/resnet50-tflite-float.zip)
| TFLITE | w8a8 | Universal | QAIRT 2.50 | [Download](https://qaihub-public-assets.s3.us-west-2.amazonaws.com/qai-hub-models/models/resnet50/releases/v0.66.0/resnet50-tflite-w8a8.zip)

For more device-specific assets and performance metrics, visit **[ResNet50 on Qualcomm® AI Hub](https://aihub.qualcomm.com/models/resnet50)**.


### Option 2: Export with Custom Configurations

Use the [Qualcomm® AI Hub Models](https://github.com/qualcomm/ai-hub-models/blob/v0.66.0/src/qai_hub_models/models/resnet50) Python library to compile and export the model with your own:
- Custom weights (e.g., fine-tuned checkpoints)
- Custom input shapes
- Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for [ResNet50 on GitHub](https://github.com/qualcomm/ai-hub-models/blob/v0.66.0/src/qai_hub_models/models/resnet50) for usage instructions.

## Model Details

**Model Type:** Model_use_case.image_classification

**Model Stats:**
- Input resolution: 224x224
- Model checkpoint: Imagenet
- Model size (float): 97.4 MB
- Model size (w8a8): 25.1 MB
- Number of parameters: 25.5M

## Performance Summary
| Model | Runtime | Precision | Chipset | Inference Time (ms) | Peak Memory Range (MB) | Primary Compute Unit
|---|---|---|---|---|---|---
| ResNet50 | ONNX | float | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 1.07 ms | 0 - 53 MB | NPU
| ResNet50 | ONNX | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.224 ms | 0 - 50 MB | NPU
| ResNet50 | ONNX | float | Snapdragon® X2 Elite | 0.964 ms | 2 - 2 MB | NPU
| ResNet50 | ONNX | float | Snapdragon® X Elite | 1.922 ms | 50 - 50 MB | NPU
| ResNet50 | ONNX | float | Snapdragon® 8 Gen 3 Mobile | 1.434 ms | 0 - 82 MB | NPU
| ResNet50 | ONNX | float | Snapdragon® 8 Gen 1 Mobile | 2.889 ms | 0 - 67 MB | NPU
| ResNet50 | ONNX | float | Qualcomm® Dragonwing™ IQ-8275 | 3.3 ms | 0 - 5 MB | NPU
| ResNet50 | ONNX | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 1.981 ms | 0 - 71 MB | NPU
| ResNet50 | ONNX | float | Qualcomm® QCS8450 | 2.889 ms | 0 - 67 MB | NPU
| ResNet50 | ONNX | float | Qualcomm® Dragonwing™ IQ-9075 | 3.0 ms | 1 - 4 MB | NPU
| ResNet50 | ONNX | float | Qualcomm® Dragonwing™ IQ-X7181 | 1.922 ms | 50 - 50 MB | NPU
| ResNet50 | ONNX | float | Qualcomm® Dragonwing™ Q-8750 | 1.224 ms | 0 - 50 MB | NPU
| ResNet50 | ONNX | w8a8 | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 0.559 ms | 0 - 58 MB | NPU
| ResNet50 | ONNX | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.578 ms | 0 - 53 MB | NPU
| ResNet50 | ONNX | w8a8 | Snapdragon® X2 Elite | 0.409 ms | 1 - 1 MB | NPU
| ResNet50 | ONNX | w8a8 | Snapdragon® X Elite | 0.843 ms | 25 - 25 MB | NPU
| ResNet50 | ONNX | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.674 ms | 0 - 89 MB | NPU
| ResNet50 | ONNX | w8a8 | Snapdragon® 8 Gen 1 Mobile | 1.055 ms | 0 - 90 MB | NPU
| ResNet50 | ONNX | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 2.64 ms | 0 - 3 MB | NPU
| ResNet50 | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 0.977 ms | 0 - 4 MB | NPU
| ResNet50 | ONNX | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.91 ms | 0 - 30 MB | NPU
| ResNet50 | ONNX | w8a8 | Qualcomm® QCS8450 | 1.055 ms | 0 - 90 MB | NPU
| ResNet50 | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 1.018 ms | 0 - 3 MB | NPU
| ResNet50 | ONNX | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 0.843 ms | 25 - 25 MB | NPU
| ResNet50 | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 6.219 ms | 0 - 170 MB | NPU
| ResNet50 | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 1.134 ms | 0 - 59 MB | NPU
| ResNet50 | ONNX | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.578 ms | 0 - 53 MB | NPU
| ResNet50 | ONNX | w8a8 | Snapdragon® 7 Gen 4 Mobile | 1.134 ms | 0 - 59 MB | NPU
| ResNet50 | QNN_DLC | float | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 1.096 ms | 1 - 53 MB | NPU
| ResNet50 | QNN_DLC | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.281 ms | 0 - 47 MB | NPU
| ResNet50 | QNN_DLC | float | Snapdragon® X2 Elite | 1.225 ms | 1 - 1 MB | NPU
| ResNet50 | QNN_DLC | float | Snapdragon® X Elite | 2.348 ms | 1 - 1 MB | NPU
| ResNet50 | QNN_DLC | float | Snapdragon® 8 Gen 3 Mobile | 1.617 ms | 0 - 79 MB | NPU
| ResNet50 | QNN_DLC | float | Snapdragon® 8 Gen 1 Mobile | 3.475 ms | 0 - 67 MB | NPU
| ResNet50 | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-8275 | 3.36 ms | 1 - 4 MB | NPU
| ResNet50 | QNN_DLC | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 2.245 ms | 1 - 2 MB | NPU
| ResNet50 | QNN_DLC | float | Qualcomm® SA8775P | 3.316 ms | 1 - 50 MB | NPU
| ResNet50 | QNN_DLC | float | Qualcomm® SA8650P | 3.316 ms | 1 - 50 MB | NPU
| ResNet50 | QNN_DLC | float | Qualcomm® SA8255P | 3.316 ms | 1 - 50 MB | NPU
| ResNet50 | QNN_DLC | float | Qualcomm® QCS8450 | 3.475 ms | 0 - 67 MB | NPU
| ResNet50 | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-9075 | 3.168 ms | 3 - 5 MB | NPU
| ResNet50 | QNN_DLC | float | Qualcomm® Dragonwing™ IQ-X7181 | 2.348 ms | 1 - 1 MB | NPU
| ResNet50 | QNN_DLC | float | Qualcomm® Dragonwing™ Q-8750 | 1.281 ms | 0 - 47 MB | NPU
| ResNet50 | QNN_DLC | float | Qualcomm® SA8295P | 3.548 ms | 0 - 32 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 0.486 ms | 0 - 48 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.534 ms | 0 - 51 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Snapdragon® X2 Elite | 0.444 ms | 0 - 0 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Snapdragon® X Elite | 0.939 ms | 0 - 0 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.657 ms | 0 - 77 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Snapdragon® 8 Gen 1 Mobile | 1.054 ms | 0 - 82 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 2.96 ms | 2 - 4 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 0.943 ms | 0 - 3 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.901 ms | 0 - 2 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® SA8775P | 1.086 ms | 0 - 52 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® SA8650P | 1.086 ms | 0 - 52 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® SA8255P | 1.086 ms | 0 - 52 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® QCS8450 | 1.054 ms | 0 - 82 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 0.965 ms | 0 - 2 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ IQ-X7181 | 0.939 ms | 0 - 0 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 6.443 ms | 0 - 161 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 1.222 ms | 0 - 55 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.534 ms | 0 - 51 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Qualcomm® SA8295P | 1.309 ms | 0 - 47 MB | NPU
| ResNet50 | QNN_DLC | w8a8 | Snapdragon® 7 Gen 4 Mobile | 1.222 ms | 0 - 55 MB | NPU
| ResNet50 | TFLITE | float | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 1.107 ms | 0 - 72 MB | NPU
| ResNet50 | TFLITE | float | Snapdragon® 8 Elite For Galaxy Mobile | 1.257 ms | 0 - 69 MB | NPU
| ResNet50 | TFLITE | float | Snapdragon® 8 Gen 3 Mobile | 1.592 ms | 0 - 118 MB | NPU
| ResNet50 | TFLITE | float | Snapdragon® 8 Gen 1 Mobile | 3.471 ms | 0 - 103 MB | NPU
| ResNet50 | TFLITE | float | Qualcomm® Dragonwing™ IQ-8275 | 3.334 ms | 0 - 53 MB | NPU
| ResNet50 | TFLITE | float | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 2.203 ms | 0 - 3 MB | NPU
| ResNet50 | TFLITE | float | Qualcomm® QCS8450 | 3.471 ms | 0 - 103 MB | NPU
| ResNet50 | TFLITE | float | Qualcomm® Dragonwing™ IQ-9075 | 3.132 ms | 0 - 52 MB | NPU
| ResNet50 | TFLITE | float | Qualcomm® Dragonwing™ Q-8750 | 1.257 ms | 0 - 69 MB | NPU
| ResNet50 | TFLITE | float | Qualcomm® SA7255P | 10.595 ms | 0 - 71 MB | NPU
| ResNet50 | TFLITE | float | Qualcomm® SA8295P | 3.501 ms | 0 - 54 MB | NPU
| ResNet50 | TFLITE | w8a8 | Snapdragon® 8 Elite Gen 5 For Galaxy Mobile | 0.453 ms | 0 - 46 MB | NPU
| ResNet50 | TFLITE | w8a8 | Snapdragon® 8 Elite For Galaxy Mobile | 0.477 ms | 0 - 43 MB | NPU
| ResNet50 | TFLITE | w8a8 | Snapdragon® 8 Gen 3 Mobile | 0.585 ms | 0 - 78 MB | NPU
| ResNet50 | TFLITE | w8a8 | Snapdragon® 8 Gen 1 Mobile | 0.921 ms | 0 - 86 MB | NPU
| ResNet50 | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS6490 | 2.662 ms | 0 - 27 MB | NPU
| ResNet50 | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-8275 | 0.835 ms | 0 - 28 MB | NPU
| ResNet50 | TFLITE | w8a8 | Qualcomm® Dragonwing™ QCS8550 (Proxy) | 0.752 ms | 0 - 2 MB | NPU
| ResNet50 | TFLITE | w8a8 | Qualcomm® QCS8450 | 0.921 ms | 0 - 86 MB | NPU
| ResNet50 | TFLITE | w8a8 | Qualcomm® Dragonwing™ IQ-9075 | 0.839 ms | 0 - 27 MB | NPU
| ResNet50 | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-6690 | 5.867 ms | 0 - 157 MB | NPU
| ResNet50 | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-7790 | 1.08 ms | 0 - 53 MB | NPU
| ResNet50 | TFLITE | w8a8 | Qualcomm® Dragonwing™ Q-8750 | 0.477 ms | 0 - 43 MB | NPU
| ResNet50 | TFLITE | w8a8 | Qualcomm® SA7255P | 1.759 ms | 0 - 47 MB | NPU
| ResNet50 | TFLITE | w8a8 | Qualcomm® SA8295P | 1.154 ms | 0 - 45 MB | NPU
| ResNet50 | TFLITE | w8a8 | Snapdragon® 7 Gen 4 Mobile | 1.08 ms | 0 - 53 MB | NPU

## License
* The license for the original implementation of ResNet50 can be found
  [here](https://github.com/pytorch/vision/blob/main/LICENSE).

## References
* [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385)
* [Source Model Implementation](https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py)

## Community
* Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI.
* For questions or feedback please [reach out to us](mailto:[email protected]).