fix: Prune configs and renaming

Author: Linked-Liszt

configs/norm_train.yaml

@@ -1,17 +1,18 @@
-# AlphaDiffract trainer configuration (paper-aligned defaults provided here)
-# This file is required by src/trainer/train_paper.py. It contains all parameters with no script-side defaults.
+# AlphaDiffract trainer configuration — ConvNeXt (paper-matching lightweight variant)
+# Use with: PYTHONPATH=src python -m trainer.train_paper configs/trainer_convnext_paper.yaml
 
 # --- Data / Manifests ---
-manifest_dir: "../../data/manifests"
-dataset_root: "../../data/dataset"     # used when auto_generate_manifests is true
+manifest_dir: "../../../ad_data/manifests"
+dataset_root: "../../../ad_data/data/dataset"
+extra_val_file: "rruff.jsonl"
 auto_generate_manifests: true
 train_ratio: 0.8
 val_ratio: 0.1
 test_ratio: 0.1
 seed: 42
 
 # --- DataLoader ---
-batch_size: 256                  # paper used 64
+batch_size: 64                  # match OG run (64 per process)
 num_workers: 8
 pin_memory: true
 persistent_workers: true
@@ -31,22 +32,33 @@ labels_key_map:
   lp_alpha: "_cell_angle_alpha"
   lp_beta: "_cell_angle_beta"
   lp_gamma: "_cell_angle_gamma"
-dtype: "float32"                # one of: float32, float64, float16, bfloat16
-mmap_mode: null                 # NumPy memmap mode: 'r', 'r+', 'w+', or null to disable
-floor_at_zero: True             # Clamp negative counts to 0 before any normalization
-normalize_log1p: True          # If true, apply log1p(x) to compress dynamic range
+dtype: "float32"
+mmap_mode: null
+floor_at_zero: true
+normalize_log1p: False          # paper used log1p preprocessing
+model_type: "multiscale"
 
-# --- Model architecture ---
-depths: [3, 3, 9, 3]
-dims: [80, 160, 320, 640]
-kernel_sizes: [7, 7, 7, 7]
-strides: [4, 2, 2, 2]
+# --- ConvNeXt (OG-equivalent configuration) ---
+# 3 stages; one block per stage; large kernels; stride-5 downsampling
+# Matches OG multiscale_cnn_cls_regr_convnextBlock_LeakyReLU.json exactly
+depths: [1, 1, 1]
+dims: [80, 80, 80]
+kernel_sizes: [100, 50, 25]
+strides: [5, 5, 5]
 dropout_rate: 0.3
-layer_scale_init_value: 1.0e-6
-drop_path_rate: 0.1
+# OG uses layer_scale_init_value=0 (disabled)
+layer_scale_init_value: 0.0
+# OG uses constant drop_path_rate=0.3 (not ramped)
+drop_path_rate: 0.3
+ramped_dropout_rate: false
+block_type: "convnext"
+pooling_type: "average"
+final_pool: true
+use_batchnorm: false
+output_type: "flatten"
 
 # Heads
-head_dropout: 0.2
+head_dropout: 0.5
 cs_hidden: [2300, 1150]
 sg_hidden: [2300, 1150]
 lp_hidden: [512, 256]
@@ -58,7 +70,7 @@ num_lp_outputs: 6
 
 # LP output bounds
 lp_bounds_min: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
-lp_bounds_max: [500.0, 500.0, 500.0, 180.0, 180.0, 180.0]
+lp_bounds_max: [300.0, 300.0, 300.0, 180.0, 180.0, 180.0]
 bound_lp_with_sigmoid: true
 
 # Loss weights
@@ -68,30 +80,39 @@ lambda_lp: 1.0
 
 # Optional GEMD term on SG
 gemd_mu: 0.0
-gemd_distance_matrix_path:          # e.g., "path/to/space_group_distance_matrix.npy" to enable GEMD
+gemd_distance_matrix_path: null
 
-# Optimizer
-lr: 0.00015                     # paper used 2e-4
-weight_decay: 0.01             # paper used 0.01
+# Optimizer (paper): AdamW, lr=2e-4, wd=0.01
+lr: 0.0002
+weight_decay: 0.01
 use_adamw: true
 gradient_clip_val: 1.0
 gradient_clip_algorithm: "norm"
 
+# --- Noise augmentation (training split only; matches paper) ---
+# If provided, noise is applied dynamically per-sample in the DataModule using the same
+# sequencing as the paper: Poisson -> normalize -> add Gaussian -> renormalize -> rescale.
+# Set ranges to None to disable.
+noise_poisson_range: [1.0, 100.0]
+noise_gaussian_range: [0.001, 0.1]
+
+# Standardize after noise to match OG CLI (--standardize-to 0 100)
+standardize_to: [0.0, 100.0]
 # --- Logging ---
-logger: "mlflow"                  # 'csv' or 'mlflow'
-csv_logger_name: "model_logs"
-mlflow_experiment_name: "OpenAlphaDiffract_ConvFUll"
-mlflow_tracking_uri: null      # null uses MLflow default (file:./mlruns)
-mlflow_run_name: "OpenAlphaDiffract_Run"
+logger: "mlflow"
+csv_logger_name: "model_logs_convnext_paper"
+mlflow_experiment_name: "AlphaDiffract_Paper_ConvNeXt"
+mlflow_tracking_uri: null
+mlflow_run_name: "ConvNeXt_Paper_Run"
 
 # --- Trainer settings ---
-default_root_dir: "outputs/model"
-max_epochs: 50
+default_root_dir: "outputs/convnext_paper"
+max_epochs: 100
 accumulate_grad_batches: 1
-precision: "bf16-mixed"        # e.g., '32', '16-mixed', 'bf16-mixed'
+precision: "32"                # match OG (AMP disabled)
 accelerator: "gpu"
 devices: 1
-log_every_n_steps: 50
+log_every_n_steps: 200
 deterministic: false
 benchmark: true
 
@@ -102,5 +123,5 @@ save_top_k: 1
 every_n_epochs: 1
 
 # --- Evaluation ---
-resume_from:                    # e.g., "outputs/paper_model/checkpoints/epochXYZ.ckpt"
+resume_from:
 test_after_train: true