[Feat] Add the pre-computed weight vector to generate intrapersonal weight change using StyleGAN2

.gitignore

@@ -1,2 +1,5 @@
 __pycache__/
 .cache/
+out
+.DS_Store
+*.pth

.python-version

@@ -0,0 +1 @@
+3.12

.vscode/launch.json

@@ -0,0 +1,14 @@
+{
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "name": "Python: Debug Current File",
+            "type": "debugpy",
+            "request": "launch",
+            "program": "${file}",
+            "console": "integratedTerminal",
+            "cwd": "${workspaceFolder}",
+            "justMyCode": true
+        }
+    ]
+}

SAGEMAKER_SETUP.md

@@ -0,0 +1,170 @@
+# SageMaker Image Generation Setup Guide
+
+## What You Have Now
+
+✅ **requirements.txt** - All Python dependencies exported from your uv setup
+✅ **launch_sagemaker_generation.py** - SageMaker launcher script
+✅ **AWS credentials** - Configured via SSO (profile: Timothe)
+✅ **S3 bucket** - stylegan2-generated-images_v2
+✅ **IAM Role** - arn:aws:iam::183295429625:role/service-role/AmazonSageMaker-ExecutionRole-20241001T141781
+
+## How It Works
+
+1. **ScriptProcessor** uploads your entire repo to SageMaker
+2. Installs dependencies from requirements.txt
+3. Runs generate.py on GPU instances (ml.g4dn.xlarge)
+4. Each instance generates a portion of the 1M images
+5. Images are continuously uploaded to S3 as they're generated
+
+## Step-by-Step Instructions
+
+### Step 1: Install SageMaker Python SDK
+
+```bash
+uv pip install sagemaker boto3
+```
+
+### Step 2: (Optional) Upload Your Trained Model to S3
+
+If you have a custom trained model (not using NVIDIA's pretrained):
+
+```bash
+# Upload pre-trained VGG checkpoint to S3
+aws s3 cp /home/sagemaker-user/stylegan2-ada-pytorch/dex_age_classifier.pth \
+s3://emobot-prod-workspace-bucket/usecases/emobot-research/datasets/stylegan2-generated-images_v2/models/vgg.pth
+
+# Upload weight vector
+aws s3 cp /home/sagemaker-user/stylegan2-ada-pytorch/weight.npy \
+s3://emobot-prod-workspace-bucket/usecases/emobot-research/datasets/stylegan2-generated-images_v2/models/weight.npy
+```
+
+### Step 3: Run a Test Job (RECOMMENDED)
+
+Start with a small test to verify everything works:
+
+```bash
+python launch_sagemaker_generation.py --test
+```
+
+This will:
+- Generate only 100 images
+- Use 1 instance
+- Run for max 1 hour
+- Help verify your setup works
+
+### Step 4: Launch Full Production Job
+
+Once the test succeeds, launch the full job:
+
+```bash
+# Using NVIDIA's pretrained FFHQ model (no upload needed)
+python launch_sagemaker_generation.py \
+    --total-images 1000000 \
+    --instances 10 \
+    --role arn:aws:iam::183295429625:role/service-role/AmazonSageMaker-ExecutionRole-20241001T141781
+
+# OR with your custom model
+python launch_sagemaker_generation.py \
+    --total-images 1000000 \
+    --instances 10 \
+    --role arn:aws:iam::183295429625:role/service-role/AmazonSageMaker-ExecutionRole-20241001T141781 \
+    --model-s3 s3://stylegan2-generated-images_v2/models/model.pkl
+
+# OR with weight vector modulation
+python launch_sagemaker_generation.py \
+    --total-images 1000000 \
+    --instances 10 \
+    --role arn:aws:iam::183295429625:role/service-role/AmazonSageMaker-ExecutionRole-20241001T141781 \
+    --weight-vector-s3 s3://stylegan2-generated-images_v2/models/weight.npy
+```
+
+### Step 5: Monitor Progress
+
+Monitor your job at:
+https://console.aws.amazon.com/sagemaker/home?region=eu-west-3#/processing-jobs
+
+Or via CLI:
+```bash
+aws sagemaker list-processing-jobs --profile Timothe
+```
+
+### Step 6: Access Generated Images
+
+Images will be uploaded to:
+```
+s3://stylegan2-generated-images_v2/generated_images_YYYYMMDD_HHMMSS/
+```
+
+Download them:
+```bash
+aws s3 sync s3://stylegan2-generated-images_v2/generated_images_YYYYMMDD_HHMMSS/ ./downloaded_images/ --profile Timothe
+```
+
+## Configuration Options
+
+### Instance Types
+
+- **ml.g4dn.xlarge** (1 GPU, 16 GB RAM) - ~$0.736/hour - Default
+- **ml.g4dn.2xlarge** (1 GPU, 32 GB RAM) - ~$0.94/hour - More memory
+- **ml.g4dn.4xlarge** (1 GPU, 64 GB RAM) - ~$1.50/hour - If you need more RAM
+
+### Cost Estimation
+
+For 1M images with default settings:
+- 10 instances × ml.g4dn.xlarge × 24 hours = ~$176
+- Adjust instance_count and max_runtime based on your needs
+
+### Parallelization Strategy
+
+The script splits images evenly across instances:
+- **10 instances** → 100K images each
+- **20 instances** → 50K images each (faster but more expensive)
+- **5 instances** → 200K images each (cheaper but slower)
+
+## Customizing Generation
+
+Edit `launch_sagemaker_generation.py` to customize:
+
+```python
+launch_image_generation(
+    total_images=1_000_000,
+    instance_count=10,
+    truncation_psi=0.7,        # Adjust truncation
+    noise_mode='random',       # or 'const', 'none'
+    alphas="-5,0,5",          # For weight modulation
+    style_range=(0, 17),      # Style block range
+    max_runtime_hours=24,
+)
+```
+
+## Troubleshooting
+
+### Issue: "Could not auto-detect SageMaker role"
+**Solution:** Pass the role explicitly with `--role` flag
+
+### Issue: "Access Denied" when uploading to S3
+**Solution:** Check your IAM role has S3 write permissions:
+```bash
+aws iam get-role-policy --role-name AmazonSageMaker-ExecutionRole-20241001T141781 --policy-name S3Access --profile Timothe
+```
+
+### Issue: Job fails with "CUDA out of memory"
+**Solution:** Reduce batch size or use larger instance (ml.g4dn.2xlarge)
+
+### Issue: Dependencies fail to install
+**Solution:** The requirements.txt is comprehensive. If issues persist, you may need to create a custom Docker image.
+
+## Next Steps After Setup
+
+1. ✅ Run test job with `--test` flag
+2. ✅ Verify images appear in S3
+3. ✅ Launch full production job
+4. ✅ Monitor progress in SageMaker console
+5. ✅ Download and use generated images
+
+## Important Notes
+
+- **Continuous Upload**: Images upload to S3 as they're generated (no need to wait for job completion)
+- **Resumability**: If a job fails, you can restart with different seed ranges
+- **Cost Control**: Set `max_runtime_hours` to prevent runaway costs
+- **Region**: All resources are in eu-west-3 (Paris)

age_estimator.py

@@ -0,0 +1,34 @@
+import torch
+import torch.nn.functional as F 
+from vgg import VGG
+
+
+class AgePredictor:
+    def __init__(self, vgg_path: str):
+        self.age_net = VGG()
+        ckpt = torch.load(vgg_path, map_location="cpu")['state_dict']
+        ckpt = {k.replace('-', '_'): v for k, v in ckpt.items()}
+        self.age_net.load_state_dict(ckpt)
+        self.age_net.cuda()
+        self.age_net.eval()
+        self.min_age = 0
+        self.max_age = 100
+
+    def __get_predicted_age(self, age_pb):
+        predict_age_pb = F.softmax(age_pb)
+        predict_age = torch.zeros(age_pb.size(0)).type_as(predict_age_pb)
+        for i in range(age_pb.size(0)):
+            for j in range(age_pb.size(1)):
+                predict_age[i] += j * predict_age_pb[i][j]
+        return predict_age
+
+    def extract_age(self, x):
+        x = F.interpolate(x, size=(224, 224), mode='bilinear', align_corners=False)  # shape: [1, 3, 224, 224]
+        predict_age_pb = self.age_net(x)['fc8']
+        predicted_age = self.__get_predicted_age(predict_age_pb)
+        return predicted_age
+
+    def __call__(self, img: torch.tensor) -> float | None:
+        """Allow instance to be called directly."""
+        return self.extract_age(img)[0]
+