Release AI-ModelZoo-4.0.0

README.md

@@ -1,10 +1,16 @@
-# IGN HAR model
+---
+license: other
+license_name: sla0044
+license_link: >-
+  https://github.com/STMicroelectronics/stm32ai-modelzoo/blob/main/human_activity_recognition/st_ign/ST_pretrainedmodel_custom_dataset/LICENSE.md
+---
+# ST_IGN HAR model
 
 ## **Use case** : `Human activity recognition`
 
 # Model description
 
-IGN is acronym of Ignatov, and is a convolutional neural network (CNN) based model for performing the human activity recognition (HAR) task based on the 3D accelerometer data. In this work we use a modified version of the IGN model presented in the [paper[2]](#2). It uses the 3D raw data with gravity rotation and supression filter as preprocessing. This is a light model with very small foot prints in terms of FLASH and RAM as well as computational requirements.
+IGN is acronym of Ignatov, and is a convolutional neural network (CNN) based model for performing the human activity recognition (HAR) task based on the 3D accelerometer data.  In this work we use a modified version of the IGN model presented in the [paper[2]](#2). The prefix `st_` denotes it is a variation of the model built by STMicroelectronics. It uses the 3D raw data with gravity rotation and supression filter as preprocessing. This is a light model with very small foot prints in terms of FLASH and RAM as well as computational requirements.
 
 This network supports any input size greater than (20 x 3 x 1) but we recommend to use at least (24 x 3 x 1), i.e. a window length of 24 samples. In this folder we provide IGN models trained with two different window lenghts [24 and 48].
 
@@ -48,16 +54,16 @@ For an input resolution of wl x 3 x 1 and P classes
 
 ## Metrics
 
-Measures are done with [STM32Cube.AI Dev Cloud version](https://stm32ai-cs.st.com/home) 10.0.0 with enabled input/output allocated options and balanced optimization. The inference time is reported is calculated using **STM32Cube.AI version 10.0.0**, on STM32 board **B-U585I-IOT02A** running at Frequency of **160 MHz**.
+Measures are done with [STEdge AI Dev Cloud version](https://stm32ai-cs.st.com/home) 3.0.0 with enabled input/output allocated options and balanced optimization. The inference time is reported is calculated using **STEdge AI version 3.0.0**, on STM32 board **B-U585I-IOT02A** running at Frequency of **160 MHz**.
 
 
 Reference memory footprint and inference times for IGN models are given in the table below. The accuracies are provided in the sections after for two datasets.
 
 
-| Model                                                                        |   Format  | Input Shape | Series  | Activation RAM (KiB) | Runtime RAM (KiB) | Weights Flash (KiB) | Code Flash (KiB) | Total RAM (KiB)| Total Flash (KiB) | Inference Time (msec) | STM32Cube.AI version  |
-|:-----------------------------------------------------------------------------:|:---------:|:-----------:|:-------:|:--------------------:|:-----------------:|:-------------------:|:----------------:|:--------------:|:-----------------:|:---------------------:|:---------------------:|
-| [IGN wl 24](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/ign/ST_pretrainedmodel_public_dataset/WISDM/ign_wl_24/ign_wl_24.h5) | FLOAT32   | 24 x 3 x 1  | STM32U5 | 2.03                 | 1.91              | 11.97               | 13.61            |  3.94          | 25.58          |   2.25    | 10.0.0            |
-| [IGN wl 48](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/ign/ST_pretrainedmodel_public_dataset/WISDM/ign_wl_48/ign_wl_48.h5) | FLOAT32   | 48 x 3 x 1  | STM32U5 | 4.56                 | 1.91              | 38.97               | 13.61            |  6.47          | 52.58          |   8.17    | 10.0.0            |
+| Model                                                                                     |   Format  | Input Shape | Series  | Activation RAM (KiB) | Runtime RAM (KiB) | Weights Flash (KiB) | Code Flash (KiB) | Total RAM (KiB)| Total Flash (KiB) | Inference Time (msec) | STEdge AI Core version  |
+|:-----------------------------------------------------------------------------------------:|:---------:|:-----------:|:-------:|:--------------------:|:-----------------:|:-------------------:|:----------------:|:--------------:|:-----------------:|:---------------------:|:---------------------:|
+| [st_ign_wl_24](./https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/st_ign/ST_pretrainedmodel_public_dataset/WISDM/st_ign_wl_24/st_ign_wl_24.keras) | FLOAT32   | 24 x 3 x 1  | STM32U5 | 2.88                 | 0.28              | 11.97               | 6.15            |  3.16          | 18.12            |   1.99                | 3.0.0                |
+| [st_ign_wl_48](./https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/st_ign/ST_pretrainedmodel_public_dataset/WISDM/st_ign_wl_48/st_ign_wl_48.keras) | FLOAT32   | 48 x 3 x 1  | STM32U5 | 9.91                 | 0.28              | 38.97               | 6.16            |  10.19         | 45.13            |   7.23                | 3.0.0                |
 
 
 
@@ -68,14 +74,14 @@ Reference memory footprint and inference times for IGN models are given in the t
 Dataset details: A custom dataset and not publically available, Number of classes: 5 [Stationary, Walking, Jogging, Biking, Vehicle]. **(We kept only 4, [Stationary, Walking, Jogging, Biking]) and removed Driving**, Number of input frames:  81,151 (for wl = 24), and 40,575 for (wl = 48).
 
 
-| Model                                                                                        | Format | Resolution | Accuracy (%)|
-|:--------------------------------------------------------------------------------------------:|:------:|:----------:|:-----------:|
-| [IGN wl 24](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/ign/ST_pretrainedmodel_public_dataset/mobility_v1/ign_wl_24/ign_wl_24.h5)          | FLOAT32| 24 x 3 x 1 | 94.64       |
-| [IGN wl 48](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/ign/ST_pretrainedmodel_public_dataset/mobility_v1/ign_wl_48/ign_wl_48.h5)          | FLOAT32| 48 x 3 x 1 | 95.01       |
+| Model                                                                                            | Format | Resolution | Accuracy (%)|
+|:------------------------------------------------------------------------------------------------:|:------:|:----------:|:-----------:|
+| [st_ign_wl_24](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/st_ign/ST_pretrainedmodel_custom_dataset/mobility_v1/st_ign_wl_24/st_ign_wl_24.keras)  | FLOAT32| 24 x 3 x 1 | 95.04       |
+| [st_ign_wl_48](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/st_ign/ST_pretrainedmodel_custom_dataset/mobility_v1/st_ign_wl_48/st_ign_wl_48.keras)  | FLOAT32| 48 x 3 x 1 | 94.29       |
 
 Confusion matrix for IGN wl 24 with Float32 weights for mobility_v1 dataset is given below.
 
-![plot](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/ign/doc/img/mobility_v1_ign_wl_24_confusion_matrix.png)
+![plot](./doc/img/mobility_v1_st_ign_wl_24_confusion_matrix.png)
 
 
 ### Accuracy with WISDM dataset
@@ -83,10 +89,10 @@ Confusion matrix for IGN wl 24 with Float32 weights for mobility_v1 dataset is g
 
 Dataset details: [link](([WISDM]("https://www.cis.fordham.edu/wisdm/dataset.php"))) , License [CC BY 2.0](https://creativecommons.org/licenses/by/2.0/) , Quotation[[1]](#1) , Number of classes: 4 (we are combining [Upstairs and Downstairs into Stairs] and [Standing and Sitting into Stationary]), Number of samples: 45,579 (at wl = 24), and 22,880 (at wl = 48).
 
-| Model                                                                                 | Format  | Resolution |  Accuracy (%) |
-|:-------------------------------------------------------------------------------------:|:-------:|:----------:|:-------------:|
-| [IGN wl 24]((https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/ign/ST_pretrainedmodel_public_datase/WISDM/ign_wl_24/ign_wl_24.h5)           | FLOAT32 | 24 x 3 x 1 | 91.7         |
-| [IGN wl 48]((https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/ign/ST_pretrainedmodel_public_datase/WISDM/ign_wl_48/ign_wl_48.h5)           | FLOAT32 | 48 x 3 x 1 | 93.67          |
+| Model                                                                                    | Format  | Resolution |  Accuracy (%) |
+|:----------------------------------------------------------------------------------------:|:-------:|:----------:|:-------------:|
+| [st_ign_wl_24](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/st_ign/ST_pretrainedmodel_public_dataset/WISDM/st_ign_wl_24/st_ign_wl_24.keras)  | FLOAT32 | 24 x 3 x 1 | 91.78          |
+| [st_ign_wl_48](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/human_activity_recognition/st_ign/ST_pretrainedmodel_public_dataset/WISDM/st_ign_wl_48/st_ign_wl_48.keras)  | FLOAT32 | 48 x 3 x 1 | 94.09         |
 
 
 ## Retraining and Integration in a simple example: