Hugging Face's logo

STMicroelectronics
/
deep_layer_aggregation

Image Classification
Model card Files
xet
Community
FBAGSTM commited on
Commit
5ddbe01
·
verified ·
1 Parent(s): 78e9ec3

Release AI-ModelZoo-4.0.0

Browse files
Files changed (1) hide show
  1. README.md +124 -3
README.md CHANGED
@@ -1,3 +1,124 @@
1
- ---
2
- license: apache-2.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ pipeline_tag: image-classification
4
+ ---
5
+ # DLA (Deep Layer Aggregation)
6
+
7
+ ## **Use case** : `Image classification`
8
+
9
+ # Model description
10
+
11
+
12
+ Deep Layer Aggregation (DLA) networks introduce iterative and hierarchical structures for aggregating features across layers. DLA extends standard architectures by merging features from different depths and resolutions, enabling **richer semantic and spatial information flow**.
13
+
14
+ DLA employs **Hierarchical Deep Aggregation (HDA)** to merge feature hierarchies combining features from different depths, and **Iterative Deep Aggregation (IDA)** to progressively refine resolution and semantic information. The dense connections enable gradient flow and feature reuse across the network.
15
+
16
+ DLA is particularly well-suited for applications requiring multi-scale feature representation, such as semantic segmentation and object detection.
17
+
18
+ (source: https://arxiv.org/abs/1707.06484)
19
+
20
+ The model is quantized to **int8** using **ONNX Runtime** and exported for efficient deployment.
21
+
22
+ ## Network information
23
+
24
+
25
+ | Network Information | Value |
26
+ |--------------------|-------|
27
+ | Framework | Torch |
28
+ | MParams | ~1.04–1.25 M |
29
+ | Quantization | Int8 |
30
+ | Provenance | https://github.com/ucbdrive/dla |
31
+ | Paper | https://arxiv.org/abs/1707.06484 |
32
+
33
+ ## Network inputs / outputs
34
+
35
+
36
+ For an image resolution of NxM and P classes
37
+
38
+ | Input Shape | Description |
39
+ | ----- | ----------- |
40
+ | (1, N, M, 3) | Single NxM RGB image with UINT8 values between 0 and 255 |
41
+
42
+ | Output Shape | Description |
43
+ | ----- | ----------- |
44
+ | (1, P) | Per-class confidence for P classes in FLOAT32|
45
+
46
+
47
+ ## Recommended platforms
48
+
49
+
50
+ | Platform | Supported | Recommended |
51
+ |----------|-----------|-----------|
52
+ | STM32L0 |[]|[]|
53
+ | STM32L4 |[]|[]|
54
+ | STM32U5 |[]|[]|
55
+ | STM32H7 |[]|[]|
56
+ | STM32MP1 |[]|[]|
57
+ | STM32MP2 |[]|[]|
58
+ | STM32N6 |[x]|[x]|
59
+
60
+ # Performances
61
+
62
+ ## Metrics
63
+
64
+ - Measures are done with default STEdgeAI Core configuration with enabled input / output allocated option.
65
+ - All the models are trained from scratch on Imagenet dataset
66
+
67
+ ### Reference **NPU** memory footprint on Imagenet dataset (see Accuracy for details on dataset)
68
+ | Model | Dataset | Format | Resolution | Series | Internal RAM (KiB) | External RAM (KiB) | Weights Flash (KiB) | STEdgeAI Core version |
69
+ |-------|---------|--------|------------|--------|--------------|--------------|---------------|----------------------|
70
+ | [dla46xc_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46xc_pt_224/dla46xc_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6 | 2361 | 6272 | 1036.41 | 3.0.0 |
71
+ | [dla46c_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46c_pt_224/dla46c_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6 | 2361 | 6272 | 1266.66 | 3.0.0 |
72
+ | [dla60xc_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla60xc_pt_224/dla60xc_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6 | 2361 | 6272 | 1278.52 | 3.0.0 |
73
+
74
+
75
+
76
+ ### Reference **NPU** inference time on Imagenet dataset (see Accuracy for details on dataset)
77
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
78
+ |-------|---------|--------|--------|------------|-------|-----------------|-------------------|---------------------|
79
+ | [dla46c_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46c_pt_224/dla46c_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 184.23 | 5.43 | 3.0.0 |
80
+ | [dla46xc_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46xc_pt_224/dla46xc_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 186.36 | 5.37 | 3.0.0 |
81
+ | [dla60xc_pt_224](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla60xc_pt_224/dla60xc_pt_224_qdq_int8.onnx) | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 187.54 | 5.33 | 3.0.0 |
82
+
83
+
84
+
85
+
86
+ ### Accuracy with Imagenet dataset
87
+
88
+ | Model | Format | Resolution | Top 1 Accuracy |
89
+ | --- | --- | --- | --- |
90
+ | [dla46c_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46c_pt_224/dla46c_pt_224.onnx) | Float | 224x224x3 | 65.03 % |
91
+ | [dla46c_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46c_pt_224/dla46c_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 64.43 % |
92
+ | [dla46xc_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46xc_pt_224/dla46xc_pt_224.onnx) | Float | 224x224x3 | 66.50 % |
93
+ | [dla46xc_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46xc_pt_224/dla46xc_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 66.06 % |
94
+ | [dla60xc_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla60xc_pt_224/dla60xc_pt_224.onnx) | Float | 224x224x3 | 68.30 % |
95
+ | [dla60xc_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla60xc_pt_224/dla60xc_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 67.73 % |
96
+
97
+
98
+ Dataset details: [link](https://www.image-net.org)
99
+ Number of classes: 1000.
100
+ To perform the quantization, we calibrated the activations with a random subset of the training set.
101
+ For the sake of simplicity, the accuracy reported here was estimated on the 50000 labelled images of the validation set.
102
+
103
+ | Model | Format | Resolution | Top 1 Accuracy |
104
+ | --- | --- | --- | --- |
105
+ | [dla46c_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46c_pt_224/dla46c_pt_224.onnx) | Float | 224x224x3 | 65.03 % |
106
+ | [dla46c_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46c_pt_224/dla46c_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 64.43 % |
107
+ | [dla46xc_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46xc_pt_224/dla46xc_pt_224.onnx) | Float | 224x224x3 | 66.50 % |
108
+ | [dla46xc_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla46xc_pt_224/dla46xc_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 66.06 % |
109
+ | [dla60xc_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla60xc_pt_224/dla60xc_pt_224.onnx) | Float | 224x224x3 | 68.30 % |
110
+ | [dla60xc_pt](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/image_classification/dla_pt/Public_pretrainedmodel_public_dataset/Imagenet/dla60xc_pt_224/dla60xc_pt_224_qdq_int8.onnx) | Int8 | 224x224x3 | 67.73 % |
111
+
112
+
113
+
114
+ ## Retraining and Integration in a simple example:
115
+
116
+ Please refer to the stm32ai-modelzoo-services GitHub [here](https://github.com/STMicroelectronics/stm32ai-modelzoo-services)
117
+
118
+
119
+
120
+ # References
121
+
122
+ <a id="1">[1]</a> - **Dataset**: Imagenet (ILSVRC 2012) — https://www.image-net.org/
123
+
124
+ <a id="2">[2]</a> - **Model**: Deep Layer Aggregation — https://github.com/ucbdrive/dla