Customize Ready Model in DEEPCRAFT™ Studio
A Ready Model project is a starting point. After you download the project from DEEPCRAFT™ Studio, you can generate code for the shipped model and deploy the model onto an Infineon board, or customize the model in Studio to your device, environment, and product requirements. Refer to Getting started to know how to create the project, and Customize a Ready Model in Studio to follow the machine learning workflow.
- The Ready Model is imported into DEEPCRAFT™ Studio in evaluation mode. Triggering a new training job does not retrain the model. It evaluates the existing model against the available
train,validation, andtestdata splits. To retrain the current architecture, re-import the.h5file using the Import Architecture (H5) button, and select the modality that matches your project. - Performance figures and operating recommendations apply only to the original Ready Model. If you retrain the model, change the architecture, or change the parameters, the ready model performance reports no longer apply.
Follow the machine learning workflow in DEEPCRAFT™ Studio
Complete the following steps in order. Each step opens the matching DEEPCRAFT™ Studio documentation.
Step 1: Collect data
Collect additional data that matches your target device, microphone placement, and background environment.
- Refer to Real-time Data Collection to know how to stream data from Infineon boards.
- For audio Ready Models, you can collect audio with a built-in PC microphone. Refer to Real-Time Audio data collection and labeling using PC/laptop Microphone to know more.
- For audio collection on Infineon kits, refer to the microphone topics for the PSOC™ Edge E84 AI Kit and PSOC™ 6 AI Evaluation Kit.
- For gesture collection on Infineon kits, refer to the radar topics for the PSOC™ Edge E84 AI Kit and PSOC™ 6 AI Evaluation Kit.
Step 2: Prepare and label data
Add the new recordings to the project, label events, and keep balanced train, validation, and test splits.
- Refer to Data Preparation to know the steps after you collect data.
- Refer to Data Labeling to know how to label data in Studio.
- Refer to Bring your own data if you already have recordings outside Studio.
Step 3: Preprocess data
Keep the shipped preprocessor unless your new data requires a change. If you change preprocessing, evaluate the model again before you deploy. Refer to Preprocessing to know more.
Step 4: Train the model
Re-import the .h5 architecture as described above, then start a training job on your updated dataset.Refer to Model Training to know how to generate a model, start a training job, and download the trained files.
Step 5: Evaluate the model
Evaluate model performance in Graph UX using the project data and, if required, live streaming.
- Refer to Model Evaluation to know more.
- Refer to Evaluating a classification model using Graph UX for the Graph UX evaluation workflow.
The Graph UX project in CodeGenGraphUX/ can also be used for model evaluation.
Step 6: Generate code
Generate optimized C code from the Ready Model Graph UX project or from the Classic UI.
- Refer to Code generation for Ready Models for the Ready Model-specific Graph UX, Classic UI, and quantized workflows.
If you modify the training dataset and the project includes calibrationData.npz, regenerate that file so quantization reflects the updated data.
Step 7: Deploy the model
Deploy the generated code to the target Infineon AI hardware. Refer to Deploy model on PSOC™ 6 and PSOC™ Edge boards for ModusToolbox™ deployment.
The following deployment examples are available for Audio and Gesture Classification Ready Models:
Support
If you need support, submit a ticket on the Infineon community forum .