Missing modality handling

configs/config.yaml

@@ -0,0 +1,16 @@
+# DeepDiet Training Configuration
+# Use Hydra to manage experiments: python src/train_hydra.py
+# Override from command line: python src/train_hydra.py model.side_aggregation=attention training.lr=5e-5
+
+defaults:
+  - _self_
+  - model: default
+  - training: default
+  - data: default
+  - logging: default
+
+# Experiment metadata
+experiment:
+  name: null  # Auto-generated if null
+  tags: []
+  notes: ""

configs/logging/default.yaml

@@ -0,0 +1,13 @@
+# Logging configuration
+# Override: python src/train_hydra.py logging.wandb=true
+
+# Weights & Biases
+wandb: true
+wandb_project: deepdiet
+wandb_entity: null  # Your W&B username/team (null = default)
+
+# TensorBoard (always enabled)
+tensorboard: true
+
+# Advanced metrics frequency
+advanced_metrics_freq: 2

configs/logging/local.yaml

@@ -0,0 +1,9 @@
+# Local-only logging (no W&B)
+# Use: python src/train_hydra.py logging=local
+
+wandb: false
+wandb_project: deepdiet
+wandb_entity: null
+
+tensorboard: true
+advanced_metrics_freq: 2

configs/model/attention.yaml

@@ -0,0 +1,21 @@
+# Attention-based model configuration
+# Use: python src/train_hydra.py model=attention
+
+# Modalities
+use_side_frames: true
+use_overhead: true
+use_depth: true
+allow_missing_modalities: false
+
+# Use temporal attention for side frames
+side_aggregation: attention
+
+# LSTM hidden (not used with attention, but kept for compatibility)
+lstm_hidden: 640
+
+# Encoder settings
+freeze_encoders: false
+unfreeze_epoch: 10
+encoder_lr_multiplier: 0.1
+
+chunk_size: 4

configs/model/default.yaml

@@ -0,0 +1,22 @@
+# Model configuration
+# Override: python src/train_hydra.py model=attention
+
+# Modalities to use
+use_side_frames: true
+use_overhead: true
+use_depth: true
+allow_missing_modalities: false
+
+# Side frame aggregation: lstm, attention, or mean
+side_aggregation: lstm
+
+# LSTM-specific (ignored if using attention/mean)
+lstm_hidden: 640
+
+# Encoder settings
+freeze_encoders: false
+unfreeze_epoch: 10
+encoder_lr_multiplier: 0.1
+
+# Processing
+chunk_size: 4  # Frames processed at once (lower = less memory)

configs/model/side_only.yaml

@@ -0,0 +1,16 @@
+# Side frames only configuration (ablation study)
+# Use: python src/train_hydra.py model=side_only
+
+use_side_frames: true
+use_overhead: false
+use_depth: false
+allow_missing_modalities: false
+
+side_aggregation: attention
+lstm_hidden: 640
+
+freeze_encoders: false
+unfreeze_epoch: 10
+encoder_lr_multiplier: 0.1
+
+chunk_size: 4

configs/training/default.yaml

@@ -0,0 +1,17 @@
+# Training configuration
+# Override: python src/train_hydra.py training.lr=5e-5
+
+epochs: 20
+batch_size: 4
+lr: 1e-4
+weight_decay: 1e-4
+
+# Learning rate schedule
+lr_decay_epochs: 5
+lr_decay_factor: 0.5
+
+# Gradient clipping
+max_grad_norm: 5.0
+
+# Data loading
+num_workers: 2

configs/training/fast.yaml

@@ -0,0 +1,13 @@
+# Fast training configuration for debugging/testing
+# Use: python src/train_hydra.py training=fast
+
+epochs: 3
+batch_size: 2
+lr: 1e-4
+weight_decay: 1e-4
+
+lr_decay_epochs: 5
+lr_decay_factor: 0.5
+max_grad_norm: 5.0
+
+num_workers: 0  # Easier debugging

configs/training/long.yaml

@@ -0,0 +1,13 @@
+# Long training configuration
+# Use: python src/train_hydra.py training=long
+
+epochs: 50
+batch_size: 8
+lr: 1e-4
+weight_decay: 1e-4
+
+lr_decay_epochs: 10
+lr_decay_factor: 0.5
+max_grad_norm: 5.0
+
+num_workers: 4