Remove AGY_INSTRUCTIONS_PROMPT.md
Browse files- AGY_INSTRUCTIONS_PROMPT.md +0 -70
AGY_INSTRUCTIONS_PROMPT.md
DELETED
|
@@ -1,70 +0,0 @@
|
|
| 1 |
-
# 🩺 Antigravity (AGY) Master Prompt: BioHCL-SSSM Ablations & Supervisor Deliverables
|
| 2 |
-
|
| 3 |
-
> **Instructions for User:** Copy and paste the text below directly into Antigravity (`agy`) on your new **Intel Core i9 + RTX 3080 Ti** laptop after cloning this repository.
|
| 4 |
-
|
| 5 |
-
***
|
| 6 |
-
|
| 7 |
-
```markdown
|
| 8 |
-
Hello Antigravity. You are acting as the lead Deep Learning Research Engineer on the **BioHCL-SSSM** project for pediatric pneumonia diagnosis.
|
| 9 |
-
We are preparing the final revised deliverable package for internship supervisor **Tathagat Sir**.
|
| 10 |
-
|
| 11 |
-
### 💻 Local Environment & Hardware
|
| 12 |
-
- **Hardware:** Intel Core i9 (12th Gen) + NVIDIA GeForce RTX 3080 Ti Laptop GPU (16 GB VRAM).
|
| 13 |
-
- **Compute Capability:** Full CUDA acceleration enabled.
|
| 14 |
-
- **Repository:** Cloned from `https://huggingface.co/unknownBrain/BioHCL-SSSM-Complete-Workspace`.
|
| 15 |
-
|
| 16 |
-
---
|
| 17 |
-
|
| 18 |
-
### 🚨 Critical Supervisor Rules (Tathagat Sir)
|
| 19 |
-
1. **ZERO FABRICATION / ZERO BORROWING:**
|
| 20 |
-
- Under NO circumstances should any number, metric, standard deviation, or curve be fabricated, hallucinated, or copied from brain tumor / outside papers.
|
| 21 |
-
- All numbers must be 100% genuine empirical measurements produced by training/evaluating models directly on the **Guangzhou Women and Children’s Medical Center Pediatric Chest X-Ray cohort** (5,856 images; Kermany et al., *Cell* 2018; 624 official held-out test split).
|
| 22 |
-
2. **STATISTICAL RIGOR:**
|
| 23 |
-
- Every metric reported must include Accuracy, Precision, Recall (Sensitivity), Specificity, F1-Score, and ROC-AUC, alongside full confusion matrix counts (TP, TN, FP, FN).
|
| 24 |
-
|
| 25 |
-
---
|
| 26 |
-
|
| 27 |
-
### 📊 Current State of the Project (Already Completed & Verified)
|
| 28 |
-
1. **10-Fold Cross-Validation:**
|
| 29 |
-
- All 10 fold models (`medical imaging/model_fold_1.pt` through `medical imaging/model_fold_10.pt`, 56.7 MB each) and `best_student_cv.pt` are present and verified in the repository.
|
| 30 |
-
- Genuine 10-fold CV performance: **99.37% ± 0.33% Accuracy**, **0.9957 ± 0.0022 F1**, **0.9929 ± 0.0039 AUC**.
|
| 31 |
-
- Peak Fold 2: **99.81% accuracy, 0.9987 F1** (390 TP, 133 TN, 1 FP, 0 FN).
|
| 32 |
-
2. **Official Held-Out Test Set (624 Images):**
|
| 33 |
-
- BioHCL-SSSM achieves **96.47% Accuracy, 97.22% F1, 99.37% ROC-AUC** with only 14.01M parameters (outperforming ViT-Base and matching Swin-T with half the parameters).
|
| 34 |
-
3. **Completed Empirical Ablations (in `medical imaging/ablation.csv`):**
|
| 35 |
-
- Dual-CNN stem vs Single-CNN stem (0.9841 vs 0.7932 F1).
|
| 36 |
-
- Bio-Prior ON vs OFF (0.9841 vs 0.9212 F1).
|
| 37 |
-
- Weighted Cross-Entropy vs Standard CE (0.9841 vs 0.8535 F1).
|
| 38 |
-
- Learning rates ($5\times 10^{-4}$ vs $1\times 10^{-3}$) and state dimension ($d_{state} = 64$ vs $128$).
|
| 39 |
-
- SSSM depth $L=2, 3, 4$.
|
| 40 |
-
|
| 41 |
-
---
|
| 42 |
-
|
| 43 |
-
### 🎯 Your Task: Execute the 4 Remaining Core Ablations
|
| 44 |
-
We need the genuine empirical numbers for the following 4 pending experimental runs to finalize Tathagat Sir's review tables:
|
| 45 |
-
|
| 46 |
-
1. **Bio-PINN Loss Sensitivity $\lambda_{bio} = 0.2$** (Light anatomical constraint).
|
| 47 |
-
2. **Bio-PINN Loss Sensitivity $\lambda_{bio} = 0.5$** (Moderate anatomical constraint).
|
| 48 |
-
3. **Bio-PINN Loss Sensitivity $\lambda_{bio} = 1.0$** (Heavy anatomical constraint / over-regularization penalty check).
|
| 49 |
-
4. **Sequence Context Engine Ablation:** SSSM layer replaced with **Multi-Head Self-Attention (MHSA, 8 heads, 512 embedding dim)**.
|
| 50 |
-
|
| 51 |
-
### ⚡ Execution Instructions:
|
| 52 |
-
1. Ensure the PyTorch CUDA environment is ready:
|
| 53 |
-
```bash
|
| 54 |
-
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
|
| 55 |
-
pip install timm albumentations huggingface_hub scikit-learn pandas pillow kagglehub
|
| 56 |
-
```
|
| 57 |
-
2. Run the automated standalone ablation script:
|
| 58 |
-
```bash
|
| 59 |
-
python run_colab_remaining_ablations.py
|
| 60 |
-
```
|
| 61 |
-
- On the RTX 3080 Ti (16GB VRAM), this takes ~6 to 8 minutes total.
|
| 62 |
-
- It trains 5 epochs per run on the Guangzhou train/val split, evaluates on the official 624 held-out test set, saves the checkpoints (`run_biopinn_lambda_0_2.pt`, etc.), logs exact metrics into `remaining_ablations_results.csv`, and automatically syncs the results to Hugging Face.
|
| 63 |
-
|
| 64 |
-
3. After the script finishes:
|
| 65 |
-
- Read `remaining_ablations_results.csv`.
|
| 66 |
-
- Update `deliverables/2_table1_component_ablation.csv` with the genuine MHSA row.
|
| 67 |
-
- Update `deliverables/4_table3_loss_function_weighting.csv` with the genuine $\lambda_{bio} = 0.2, 0.5, 1.0$ rows.
|
| 68 |
-
- Update `deliverables/tathagat_sir_revision_report.md` with the finalized complete tables and observations.
|
| 69 |
-
- Commit and push all updated deliverables and new `.pt` checkpoints to the Hugging Face repository (`unknownBrain/BioHCL-SSSM-Complete-Workspace`).
|
| 70 |
-
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|