feat: add sequential model improvement roadmap (Phase 1 complete — yml config fixes)

.agents/AGENTS.md

@@ -30,3 +30,25 @@ All AI agents working on this codebase must adhere strictly to these rules:
      ```
    - This will copy the latest `AGENTS.md` to the vault's Rules directory and update the vault timestamp.
    - Vault location: `D:\AgentBrain\`
+
+7. **Mandatory Sequential Model Improvement Roadmap**:
+   - All model-related work MUST follow the sequential phase order defined in the task list artifact:
+     `C:\Users\admin\.gemini\antigravity-ide\brain\0bf6bec8-6164-477e-a32d-6f0b9ef577c6\task.md`
+   - **Before starting any model improvement task**, read `task.md` and identify the first incomplete task (`[ ]`).
+   - Do **NOT** skip phases or implement out-of-order. Each phase builds on the previous:
+     - **Phase 1** — Config fixes (yml) — COMPLETED ✅
+     - **Phase 2** — Resume & complete training run (iter 2k → 20k)
+     - **Phase 3** — ArcFace identity loss integration
+     - **Phase 4** — Dataset expansion verification & game character mixing
+     - **Phase 5** — Static INT8 ONNX quantization with calibration
+     - **Phase 6** — Stage II transformer fine-tune (GPU only, HF Space/Colab)
+     - **Phase 7** — A/B test UI & model selection dropdown
+   - After completing a task, mark it `[x]` in `task.md` before proceeding to the next.
+   - Reference the full improvement rationale in:
+     `C:\Users\admin\.gemini\antigravity-ide\brain\0bf6bec8-6164-477e-a32d-6f0b9ef577c6\model_improvement_proposals.md`
+
+8. **Always Update task.md During Model Work**:
+   - Mark tasks `[/]` (in progress) when starting them.
+   - Mark tasks `[x]` (done) when verified complete.
+   - Add notes under each task if you discover important findings (e.g., actual iteration count, loss values, timing).
+   - This ensures seamless handover between sessions and agents.

models/CodeFormer/options/CodeFormer_stage3_custom.yml

@@ -24,7 +24,7 @@ datasets:
     use_corrupt: true
     blur_kernel_size: 41
     use_motion_kernel: true
-    motion_kernel_prob: 0.05
+    motion_kernel_prob: 0.15
     kernel_list:
     - iso
     - aniso
@@ -36,13 +36,13 @@ datasets:
     - 10.0
     downsample_range:
     - 1.0
-    - 12.0
+    - 20.0
     noise_range:
     - 0.0
-    - 20.0
+    - 30.0
     jpeg_range:
-    - 50
-    - 100
+    - 10
+    - 70
     blur_sigma_large:
     - 1.0
     - 15.0
@@ -53,13 +53,13 @@ datasets:
     - 0.0
     - 30.0
     jpeg_range_large:
-    - 30
-    - 80
+    - 5
+    - 50
     latent_gt_path: null
     num_worker_per_gpu: 0
     batch_size_per_gpu: 1
-    dataset_enlarge_ratio: 1
-    prefetch_mode: cpu
+    dataset_enlarge_ratio: 5
+    prefetch_mode: null
 network_g:
   type: CodeFormer
   dim_embd: 512
@@ -97,7 +97,7 @@ path:
   param_key_g: params_ema
   strict_load_g: false
   pretrain_network_d: null
-  resume_state: D:\.gemini-scratch\custom-ai-enhancer\models\CodeFormer\experiments\20260708_201102_CodeFormer_stage3_custom\training_states\1600.state
+  resume_state: D:\.gemini-scratch\custom-ai-enhancer\models\CodeFormer\experiments\20260708_201102_CodeFormer_stage3_custom\training_states\2000.state
 train:
   use_hq_feat_loss: true
   feat_loss_weight: 1.0
@@ -121,11 +121,11 @@ train:
   scheduler:
     type: CosineAnnealingRestartLR
     periods:
-    - 150000
+    - 20000
     restart_weights:
     - 1
-    eta_min: 2.0e-05
-  total_iter: 20000
+    eta_min: 5.0e-06
+  total_iter: 2002
   warmup_iter: -1
   ema_decay: 0.997
   pixel_opt:

train_custom.py

@@ -2,7 +2,7 @@ import os
 import sys
 import cv2
 import torch
-torch.set_num_threads(4)
+torch.set_num_threads(8)  # Ryzen 7735HS has 8C/16T
 import yaml
 import subprocess
 import glob
@@ -60,7 +60,9 @@ def main():
             dataset = config["datasets"][phase]
             if device == "cpu":
                 dataset["num_worker_per_gpu"] = 0
-                dataset["prefetch_mode"] = "cpu"
+                # MUST be null on CPU — 'cpu' prefetch spawns multiprocessing
+                # workers which cause MemoryError/segfaults on Windows (Task 8).
+                dataset["prefetch_mode"] = None
             else:
                 dataset["num_worker_per_gpu"] = 4
                 
@@ -132,7 +134,7 @@ def main():
     # Force torch / BLAS backends to use limited CPU threads for stability
     for _k in ["OMP_NUM_THREADS", "MKL_NUM_THREADS", "OPENBLAS_NUM_THREADS",
                "NUMEXPR_NUM_THREADS", "VECLIB_MAXIMUM_THREADS"]:
-        env[_k] = "4"
+        env[_k] = "8"  # Match torch.set_num_threads — Ryzen 7735HS 8C/16T
     # Disable OpenCV threading and OpenCL runtime (prevents segfaults on Windows)
     env["OPENCV_OPENCL_RUNTIME"] = "disabled"
     env["OPENCV_THREAD_LIMIT"] = "1"