| { |
| "model": { |
| "visual_backbone": "convnext_base", |
| "text_encoder": "emilyalsentzer/Bio_ClinicalBERT", |
| "visual_feature_dim": 512, |
| "scene_graph_dim": 134, |
| "visual_embedding_dim": 646, |
| "hidden_size": 768, |
| "intermediate_size": 3072, |
| "num_hidden_layers": 6, |
| "num_attention_heads": 12, |
| "hidden_dropout_prob": 0.1, |
| "attention_probs_dropout_prob": 0.1, |
| "sim_layers": 2, |
| "num_regions": 30, |
| "num_entities": 22, |
| "region_embedding_dim": 64, |
| "entity_embedding_dim": 64, |
| "num_binary_classes": 2, |
| "num_category_classes": 14, |
| "num_region_classes": 26, |
| "num_severity_classes": 4, |
| "max_question_length": 128, |
| "vocab_size": 30522, |
| "qwen_model_id": null, |
| "use_quantization": null, |
| "lora_rank": 16, |
| "lora_alpha": 32, |
| "lora_target_modules": null, |
| "num_sg_tokens": 8, |
| "sg_generator": { |
| "name": "txrv_detr", |
| "num_entities": 22, |
| "num_regions": 30, |
| "num_relations": 10, |
| "num_queries": 20, |
| "d_model": 384, |
| "num_decoder_layers": 2, |
| "num_heads": 8, |
| "dropout": 0.1, |
| "input_size": 224, |
| "checkpoint": "./checkpoints/mimic-cxr-vqa-stage1-sg-only/best_model", |
| "frozen": true |
| } |
| }, |
| "data": { |
| "mimic_cxr_jpg_path": "/root/code/MASTERS-WORK/data/mimic-cxr-jpg", |
| "mimic_ext_cxr_qba_path": "/root/code/MASTERS-WORK/data/mimic-ext-cxr-qba", |
| "chexpert_labels_path": "", |
| "test_labels_csv_path": "", |
| "quality_grade": "B", |
| "view_filter": "frontal_only", |
| "question_types": null, |
| "use_exports": true, |
| "export_grade": "B_frontal", |
| "cache_dir": ".cache/dataset_samples", |
| "image_size": 224, |
| "normalize_mean": [ |
| 0.485, |
| 0.456, |
| 0.406 |
| ], |
| "normalize_std": [ |
| 0.229, |
| 0.224, |
| 0.225 |
| ], |
| "skip_question_types": null, |
| "min_localization_quality": 0, |
| "val_samples": 1000, |
| "val_quality_grade": null |
| }, |
| "training": { |
| "output_dir": "./checkpoints/mimic-cxr-vqa", |
| "phase": "alignment", |
| "batch_size_per_gpu": 2, |
| "gradient_accumulation_steps": 4, |
| "learning_rate": 5e-05, |
| "weight_decay": 0.01, |
| "warmup_ratio": 0.05, |
| "max_grad_norm": 1.0, |
| "num_epochs": 2, |
| "fp16": true, |
| "gradient_checkpointing": false, |
| "vqa_loss_weight": 0.0, |
| "generation_loss_weight": 1.0, |
| "chexpert_loss_weight": 0.05, |
| "scene_graph_loss_weight": 0.0, |
| "grounding_loss_weight": 0.0, |
| "binary_head_weight": 0.0, |
| "category_head_weight": 0.0, |
| "region_head_weight": 0.0, |
| "severity_head_weight": 0.0, |
| "logging_steps": 100, |
| "save_steps": 500, |
| "save_total_limit": 3, |
| "eval_steps": 2000, |
| "metric_for_best_model": "classification_accuracy", |
| "greater_is_better": true, |
| "early_stopping_patience": 5, |
| "dataloader_num_workers": 2, |
| "dataloader_pin_memory": true, |
| "dataloader_prefetch_factor": 2, |
| "dataloader_persistent_workers": true, |
| "hub_model_id": "KYAGABA/mimic-cxr-vqa-stage2-alignment", |
| "hub_private_repo": true, |
| "push_to_hub_strategy": "checkpoint", |
| "push_every_save": false, |
| "seed": 42, |
| "ema_decay": 0.0 |
| }, |
| "wandb": { |
| "enabled": true, |
| "project": "mimic-cxr-vqa", |
| "entity": "kyagabajonah", |
| "name": "stage2-alignment", |
| "group": "curriculum", |
| "tags": [ |
| "ssg-vqa", |
| "mimic-cxr", |
| "stage2", |
| "alignment", |
| "b-grade" |
| ], |
| "notes": "Stage 2: SG-LLM alignment (gen=1.0 primary, SG frozen). Resumes from Stage 1 best_model.", |
| "watch_model": false, |
| "watch_log_freq": 1000, |
| "log_model": true |
| }, |
| "deepspeed": { |
| "enabled": false, |
| "config_path": "configs/deepspeed_config.json", |
| "stage": 2 |
| } |
| } |
