Update training rock paper scisors (#93)

* feat: ✨ update with workshop code

* feat: ⬆️ upgrade Rock/Paper/Scisors to YOLOv11

Author: philippart-s

ai/rock-paper-scissors/application/Dockerfile

@@ -19,6 +19,3 @@ RUN chown -R 42420:42420 /workspace
 
 # run (CMD) the app.py
 CMD [ "streamlit" , "run" , "/workspace/app.py", "--server.address=0.0.0.0" ]
-
-
-#ENV HOME=/workspace

ai/rock-paper-scissors/training/README.md

@@ -5,12 +5,14 @@
 
 ### Set up
   - install `ovhai` CLI, see [documentation](https://help.ovhcloud.com/csm/en-gb-public-cloud-ai-cli-install-client?id=kb_article_view&sysparm_article=KB0047844)
-  - we assume that you have an S3* compatible bucket named `rock-paper-scissors` with all needed data from the demo Create a Notebook to [play to rock/paper/scissors](../../notebooks/YOLOV8/)
+  - we assume that you have an S3* compatible bucket named `rock-paper-scissors` with all needed data from the demo Create a Notebook to [play to rock/paper/scissors](../../rock-paper-scissors/notebooks/rock-paper-scissors.ipynb)
 
 ### Image build for AI Training
 
-	- build the image with the Python script: `docker build . -t ovhcom/rock-paper-scissors-training-job:1.0.0`
-	- push the image to the registry: `docker push ovhcom/rock-paper-scissors-training-job:1.0.0`
+	- build the image with the Python script: `docker build . -t <Shared Docker Registries>/rock-paper-scissors-training-job:1.0.0`
+	- push the image to the registry: `docker push <Shared Docker Registries>/rock-paper-scissors-training-job:1.0.0`
+
+> ℹ️ see the [documentation](https://help.ovhcloud.com/csm/fr-public-cloud-ai-manage-registries?id=kb_article_view&sysparm_article=KB0057958) for more information about the _Shared Docker Registries_ ℹ️
 
 ### AI Training Job creation 
 
@@ -22,7 +24,7 @@ ovhai job run \
 	--env NB_OF_EPOCHS=50 \
   --volume rock-paper-scissors-data@S3GRA/:/workspace/data:RW:cache \
 	--unsecure-http \
-	ovhcom/rock-paper-scissors-training-job:1.0.0
+	<Shared Docker Registries>/rock-paper-scissors-training-job:1.0.0
 ```
 
 You can follow the training with the logs: `ovhai job logs -f <job id>`

ai/rock-paper-scissors/training/train.py

@@ -3,39 +3,42 @@
 import shutil
 import os
 
-#######################################################################################################################
-## 🎯 The aim of this script is to do transfert learning on YOLOv8 model.                                            ##
-## ℹ️ Note on the environments variables:                                                                            ##
-##      - YOLO_MODEL (default value yolo11n.pt) is the YOLO model you want fine tune                                 ##
-##      - NB_OF_EPOCHS (default value: 50) is an environment variable passed to the Docker run command to specify    ##
-## the number of epochs                                                                                              ##
-##      - DEVICE_TO_USE (default value 0) is to specify to use GPU (0) or CPU (cpu)                                  ##
-##		  - PATH_TO_DATASET (default value is '/workspace/data/data.yaml') is to specify the path to the               ##
-## training dataset                                                                                                  ##
-##		  - PATH_TO_EXPORTED_MODEL (default value is '/workspace/data/') is to specify the path where export the       ##
-## trained model                                                                                                     ##
-#######################################################################################################################
+#########################################################################################################################
+## 🎯 The aim of this script is to do transfert learning on YOLOv11 model.                                             ##
+## ℹ️ Note on the environments variables:                                                                              ##
+##      - NB_OF_EPOCHS (default value: 50) is an environment variable passed to the Docker run command to specify      ##
+## the number of epochs                                                                                                ##
+##      - DEVICE_TO_USE (default value 0) is to specify to use GPU (0) or CPU (cpu)                                    ##
+##		  - PATH_TO_DATASET (default value is '/workspace/attendee/data.yaml') is to specify the path to the             ##
+## training dataset                                                                                                    ##
+##		  - PATH_TO_EXPORTED_MODEL (default value is '/workspace/attendee/') is to specify the path where export the     ##
+## trained model                                                                                                       ##
+##        - BATCH specifies the number of images used for one training iteration before updating the model's weights.  ##
+## A larger batch size can lead to faster training but requires more memory.                                           ## 
+##        - FREEZE allows to freeze certain layers of a pre-trained model. This way, these layers are kept unchanged   ##
+## during training, which allows to preserve knowledge from the pre-trained model.                                     ##
+#########################################################################################################################
 
 # ✅ Check configuration
 ultralytics.checks()
 
+# 🧠 Load a pretrained YOLO model
+model = YOLO('yolo11n.pt')
+
 # 🛠 Get configuration from environment variables
-yoloModel = os.getenv('YOLO_MODEL', 'yolo11n.pt')
 nbOfEpochs = os.getenv('NB_OF_EPOCHS', 50)
 deviceToUse = os.getenv('DEVICE_TO_USE', 0)
 pathToDataset = os.getenv('PATH_TO_DATASET', '/workspace/data/data.yaml')
 pathToExportedModel = os.getenv('PATH_TO_EXPORTED_MODEL', '/workspace/data/')
-print('YOLO model to use:', yoloModel)
+batch = os.getenv('BATCH', 64)
+freeze = os.getenv('FREEZE', 10)
 print('Number of epochs to set:', nbOfEpochs)
 print('Device to set:', deviceToUse)
 print('Path to the dataset to set:', pathToDataset)
 print('Path to the exported model to set:', pathToExportedModel)
 
-# 🧠 Load a pretrained YOLO model
-model = YOLO(yoloModel)
-
 # 💪 Train the model with new data ➡️ one GPU / NB_OF_EPOCHS iterations (epochs)
-model.train(data=pathToDataset, device=deviceToUse, epochs=int(nbOfEpochs), verbose=True)
+model.train(data=pathToDataset, device=deviceToUse, epochs=int(nbOfEpochs), verbose=True, batch=batch, freeze=freeze)
 
 # 💾 Save the model
 exportedMetaData = model.export()