# Training configuration for RTX 4090 (24GB VRAM) with IMAGE-LEVEL splits
#
# Combines RTX 4090 hardware optimizations with image-level splitting and
# gentler contrastive learning for better generalization.
#
# Usage:
#   python train_clip_logo.py --config configs/cloud_rtx4090_image_split.yaml
#
# Estimated training time: 5-7 hours (more epochs than logo-level)
# Estimated cost on RunPod: ~$4

# Base model
base_model: "openai/clip-vit-large-patch14"

# Dataset paths
dataset_dir: "LogoDet-3K"
reference_dir: "reference_logos"
db_path: "test_data_mapping.db"

# Data split configuration - IMAGE LEVEL
# Each logo brand will have images in all splits, allowing the model
# to see some examples of each brand during training.
split_level: "image"
train_split: 0.7
val_split: 0.15
test_split: 0.15

# Larger batches for faster training on 24GB VRAM
batch_size: 32
logos_per_batch: 32
samples_per_logo: 4
gradient_accumulation_steps: 4  # Effective batch = 128
num_workers: 8

# Model architecture
lora_r: 16
lora_alpha: 32
lora_dropout: 0.1
freeze_layers: 12
use_gradient_checkpointing: true

# Training - GENTLER settings for better generalization
learning_rate: 5.0e-6           # Reduced from 1e-5
weight_decay: 0.01
warmup_steps: 500
max_epochs: 30                  # More epochs with slower learning
mixed_precision: true

# Loss - HIGHER temperature for softer contrastive learning
temperature: 0.15               # Increased from 0.07
loss_type: "infonce"
triplet_margin: 0.2             # Reduced from 0.3

# Early stopping - more patience with gentler learning
patience: 7
min_delta: 0.001

# Output - separate directory for image-split model
checkpoint_dir: "checkpoints_image_split"
output_dir: "models/logo_detection/clip_finetuned_image_split"
save_every_n_epochs: 2          # Save frequently for cloud

# Logging
log_every_n_steps: 10
eval_every_n_epochs: 1

seed: 42
use_hard_negatives: false
use_augmentation: true
augmentation_strength: "medium"