File size: 27,604 Bytes
0d0412d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
"""
Model Evaluation Pipeline with KSI v2.0 Metrics
================================================

Comprehensive evaluation script for trained badminton shot classification models.
Implements enhanced Kinematic Similarity Index (KSI) v2.0 for biomechanical
analysis and optional natural language coaching feedback generation.

Key Features:
    - Classification accuracy and confusion matrix analysis
    - KSI v2.0 biomechanical comparison against expert templates
    - Phase-aware scoring (preparation, loading, contact, follow-through)
    - Per-joint error analysis with confidence intervals
    - Velocity and acceleration derivative metrics
    - Optional NLP coaching report generation
    - Full MLflow experiment tracking integration

Evaluation Metrics:
    1. Classification Metrics
       - Test accuracy, precision, recall, F1
       - Confusion matrix visualization
       
    2. KSI v2.0 Metrics
       - Total KSI score (0-100, higher = better match)
       - Component breakdown (pose, velocity, acceleration)
       - Phase-specific scores
       - Per-joint error analysis
       - Ranking hinge score for class separation

Pipeline Position:
    train_pose.py / train_hybrid.py → [evaluate.py] → reports/
    
    Loads trained model and test data, performs comprehensive evaluation,
    and logs all metrics to MLflow for experiment tracking.

Dependencies:
    External: tensorflow, sklearn, matplotlib, seaborn, mlflow, numpy, yaml
    Internal: ksi_v2.EnhancedKSI, mlflow_utils, natural_language_coach

Configuration (params.yaml):
    pose_pipeline / hybrid_pipeline:
        model_path: Path to trained model
        data_path: Path to evaluation data
    ksi:
        weights: Component weighting for KSI calculation

Usage:
    python evaluate.py pose          # Evaluate pose model
    python evaluate.py hybrid --nlp  # Evaluate hybrid with NLP feedback

Author: IPD Research Team
Version: 2.0.0
"""

"""
# --- DETERMINISM FIXES (MUST BE BEFORE TF IMPORT) ---
import os
import sys

# Check for GPU flag early (before TF imports)
_use_gpu = '--gpu' in sys.argv

if not _use_gpu:
    # Force CPU mode for deterministic predictions
    os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
    os.environ['MEDIAPIPE_DISABLE_GPU'] = '1'
    print("🔒 Running in CPU mode for deterministic predictions (use --gpu to enable GPU)")

os.environ['TF_DETERMINISTIC_OPS'] = '1'
os.environ['TF_CUDNN_DETERMINISTIC'] = '1'
os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'

"""
# --- DETERMINISM FIXES (MUST BE BEFORE TF IMPORT) ---
import os
import sys

# Check for GPU flag early (before TF imports)
_use_gpu = '--gpu' in sys.argv

if not _use_gpu:
    # Force CPU mode for deterministic predictions
    os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
    os.environ['MEDIAPIPE_DISABLE_GPU'] = '1'
    print("🔒 Running in CPU mode for deterministic predictions (use --gpu to enable GPU)")

os.environ['TF_DETERMINISTIC_OPS'] = '1'
os.environ['TF_CUDNN_DETERMINISTIC'] = '1'
os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'

import argparse
import os
import yaml
import numpy as np
import json
import mlflow
import matplotlib.pyplot as plt
import seaborn as sns
from tensorflow.keras.models import load_model
from tensorflow.keras.utils import to_categorical
from sklearn.model_selection import train_test_split
from sklearn.metrics import confusion_matrix
from typing import Dict
from ksi_v2 import EnhancedKSI, ShotPhase
from mlflow_utils import MLflowRunManager

try:
    from natural_language_coach import generate_coaching_report
    NLP_AVAILABLE = True
except ImportError:
    NLP_AVAILABLE = False
    print("⚠️  natural_language_coach not available. NLP feedback will be skipped.")


def evaluate(pipeline_type: str, model_path: str = None, generate_nlp_feedback: bool = False, 
             nlp_skill_level: str = 'intermediate', max_nlp_samples: int = 5, auto_run_name: bool = False,
             data_path: str = None):
    """
    Evaluate model with enhanced KSI v2 metrics.
    Logs phase scores, confidence intervals, ranking hinge, and component breakdowns.
    
    Args:
        pipeline_type: 'pose' or 'hybrid'
        model_path: Optional path to model file (overrides params.yaml)
        generate_nlp_feedback: Whether to generate natural language coaching reports
        nlp_skill_level: Skill level for NLP coach ('beginner', 'intermediate', 'advanced', 'expert')
        max_nlp_samples: Maximum number of samples to generate detailed feedback for
        auto_run_name: If True, auto-generate MLflow run name without prompting
        data_path: Optional path to evaluation data (overrides params.yaml)
    """
    with open("params.yaml") as f:
        params = yaml.safe_load(f)
    cfg = params[f'{pipeline_type}_pipeline']
    ksi_cfg = params.get('ksi', {'weights': {'pose': 0.5, 'velocity': 0.3, 'acceleration': 0.2}})
    
    # Use provided model path or default from config
    if model_path:
        cfg['model_path'] = model_path
        run_suffix = f" (custom: {os.path.basename(model_path)})"
    else:
        run_suffix = ""
    
    # Use provided data path or default from config
    if data_path:
        cfg['data_path'] = data_path
    
    # Setup MLflow with interactive run manager
    exp_name = "Pose_LSTM_Experiment" if pipeline_type == 'pose' else "Hybrid_TCN_Experiment"
    run_manager = MLflowRunManager(exp_name)

    with run_manager.start_interactive_run(
        default_description=f"Evaluation of {pipeline_type} pipeline with KSI v2.0 metrics{run_suffix}",
        auto_name=auto_run_name
    ):
        # Log evaluation configuration
        mlflow.log_param("evaluation.model_path", cfg['model_path'])
        mlflow.log_param("evaluation.data_path", cfg['data_path'])
        mlflow.log_param("evaluation.pipeline_type", pipeline_type)
        
        # Display what we're evaluating on
        print(f"\n{'='*70}")
        print(f"📊 EVALUATION CONFIGURATION")
        print(f"{'='*70}")
        print(f"Pipeline: {pipeline_type}")
        print(f"Model: {cfg['model_path']}")
        print(f"Data: {cfg['data_path']}")
        print(f"{'='*70}\n")
        
        # 1. Load Data
        X, y = [], []
        raw_landmarks = []  # NEW: store raw landmarks for KSI
        if not os.path.exists(cfg['data_path']):
            print(f"Data path {cfg['data_path']} not found")
            return
        classes = sorted(os.listdir(cfg['data_path']))
        for i, cls in enumerate(classes):
            path = os.path.join(cfg['data_path'], cls)
            if not os.path.isdir(path):
                continue
            for f in os.listdir(path):
                if f.endswith('.npz'):
                    data = np.load(os.path.join(path, f))
                    X.append(data['features'])
                    y.append(i)
                    # NEW: load raw landmarks if available (for hybrid pipeline KSI)
                    if 'raw_landmarks' in data:
                        raw_landmarks.append(data['raw_landmarks'])
                    else:
                        raw_landmarks.append(None)
        
        if not X:
            print("No data loaded")
            return
        X, y_cat = np.array(X), to_categorical(y, len(classes))
        
        # Split data and raw landmarks together
        if raw_landmarks and any(lm is not None for lm in raw_landmarks):
            _, X_test, _, y_test, _, raw_test = train_test_split(
                X, y_cat, raw_landmarks, test_size=0.2, stratify=y, random_state=42
            )
            has_raw_landmarks = True
        else:
            _, X_test, _, y_test = train_test_split(X, y_cat, test_size=0.2, stratify=y, random_state=42)
            raw_test = None
            has_raw_landmarks = False
        
        # 2. Load Model & Predict
        model = load_model(cfg['model_path'])
        
        if pipeline_type == 'hybrid':
            cnn_dim = cfg['cnn_feature_dim']
            X_pose, X_cnn = X_test[..., :-cnn_dim], X_test[..., -cnn_dim:]
            loss, acc = model.evaluate([X_cnn, X_pose], y_test, verbose=0)
            y_pred = np.argmax(model.predict([X_cnn, X_pose]), axis=1)
            sample_pose = X_pose
        else:
            loss, acc = model.evaluate(X_test, y_test, verbose=0)
            y_pred = np.argmax(model.predict(X_test), axis=1)
            sample_pose = X_test

        # 3. Enhanced KSI v2 Calculation with contact-centered windowing
        # NEW: KSI now works with hybrid pipeline if raw_landmarks are available
        template_path = params['expert_pipeline']['output_path']
        ksi_results = {
            'avg_ksi_total': 0.0,
            'avg_ksi_weighted': 0.0,
            'avg_confidence_ci_width': 0.0,
            'avg_uncertainty_scalar': 0.0,
            'phase_scores': {},
            'component_scores': {},
            'reliable_count': 0,
            'total_samples': 0
        }
        
        # Check if we can do KSI evaluation
        can_do_ksi = os.path.exists(template_path) and (
            pipeline_type == 'pose' or (pipeline_type == 'hybrid' and has_raw_landmarks)
        )
        
        if can_do_ksi:
            print(f"Expert templates found at {template_path}. Computing KSI metrics...")
            templates = np.load(template_path, allow_pickle=True)
            
            # Check template format - detect if raw landmarks or enhanced features
            sample_template_key = next((k for k in templates.files if not k.startswith('_')), None)
            if sample_template_key is None:
                print(f"⚠ No valid templates found in {template_path}")
                can_do_ksi = False
            else:
                sample_template = templates[sample_template_key]
                # Raw landmarks: (T, 33, 3) or (T, 99) flattened
                # Enhanced features: (T, 32)
                template_is_raw_landmarks = (
                    sample_template.ndim == 3 and sample_template.shape[1:] == (33, 3)
                ) or (
                    sample_template.ndim == 2 and sample_template.shape[1] == 99
                )
                template_is_enhanced = sample_template.ndim == 2 and sample_template.shape[1] == 32
                
                if template_is_enhanced:
                    print(f"⚠ Templates are in 32-feature format (KSI v2 features), not raw landmarks.")
                    print(f"   KSI calculation requires raw landmarks (T, 33, 3). Regenerate templates with raw landmarks.")
                    can_do_ksi = False
                elif not template_is_raw_landmarks:
                    print(f"⚠ Unknown template format: shape={sample_template.shape}. Expected (T, 33, 3) or (T, 99).")
                    can_do_ksi = False
                else:
                    print(f"   Template format: {'(T, 33, 3)' if sample_template.ndim == 3 else '(T, 99)'} - OK")
        
        if can_do_ksi:
            # Initialize enhanced KSI calculator
            ksi_calc = EnhancedKSI(
                fps=params.get('fps', 30.0),
                contact_window_pre_frames=18,
                contact_window_post_frames=18,
                bootstrap_min=50,
                bootstrap_max=200,
                ranking_margin=0.05
            )
            
            ksi_totals = []
            ksi_weighted_list = []
            ci_widths = []
            uncertainty_scalars = []
            phase_score_accumulator = {p.value: [] for p in ShotPhase}
            component_accumulator = {'pose': [], 'velocity': [], 'acceleration': [], 'jerk': []}
            reliable_count = 0
            
            # Store individual results for NLP feedback
            individual_results = [] if generate_nlp_feedback else None
            
            # Sample for speed (evaluate up to 50 items)
            n_samples = min(50, len(sample_pose))
            evaluated_count = 0  # Track actually evaluated samples
            skipped_templates = set()  # Track missing templates to report once
            
            for i in range(n_samples):
                cls = classes[np.argmax(y_test[i])]
                
                # NEW: Try main template, then variants, prioritizing main
                template_key = None
                if cls in templates:
                    template_key = cls
                elif f'{cls}_variant1' in templates:
                    # If only variants exist, use variant1 (best quality)
                    template_key = f'{cls}_variant1'
                else:
                    # Try any key containing the class name
                    for key in templates.files:
                        if cls in key and not key.startswith('_'):
                            template_key = key
                            break
                
                if template_key is None:
                    if cls not in skipped_templates:
                        skipped_templates.add(cls)
                    continue
                
                # Get user landmarks - use raw_landmarks if available (hybrid), else reshape pose features
                if pipeline_type == 'hybrid' and raw_test is not None and raw_test[i] is not None:
                    user_lm = raw_test[i]  # Already (T, 33, 3)
                else:
                    # Pose pipeline: reshape from flattened features (T, 99) -> (T, 33, 3)
                    try:
                        user_lm = sample_pose[i].reshape(-1, 33, 3)
                    except ValueError as e:
                        print(f"⚠ Cannot reshape pose features to landmarks: {sample_pose[i].shape} -> (T, 33, 3)")
                        continue
                
                # Load expert template and reshape if needed
                expert_template = templates[template_key]
                try:
                    if expert_template.ndim == 3 and expert_template.shape[1:] == (33, 3):
                        expert_lm = expert_template  # Already (T, 33, 3)
                    elif expert_template.ndim == 2 and expert_template.shape[1] == 99:
                        expert_lm = expert_template.reshape(-1, 33, 3)  # (T, 99) -> (T, 33, 3)
                    else:
                        print(f"⚠ Cannot convert template '{template_key}' shape {expert_template.shape} to landmarks")
                        continue
                except ValueError as e:
                    print(f"⚠ Template reshape failed for '{template_key}': {e}")
                    continue
                
                # Calculate enhanced KSI
                result = ksi_calc.calculate(
                    expert_landmarks=expert_lm,
                    user_landmarks=user_lm,
                    weights=ksi_cfg['weights'],
                    baseline_ksi=None  # Could pass previous user score for ranking hinge
                )
                
                ksi_totals.append(result.ksi_total)
                ksi_weighted_list.append(result.ksi_weighted)
                
                # Confidence metrics
                if result.confidence:
                    ci_width = result.confidence.get('ci_95_upper', 0) - result.confidence.get('ci_95_lower', 0)
                    ci_widths.append(ci_width)
                    uncertainty_scalars.append(result.confidence.get('uncertainty_scalar', 0))
                    if result.confidence.get('reliable', False):
                        reliable_count += 1
                
                # Phase scores
                for phase, score in result.phase_scores.items():
                    if phase in phase_score_accumulator:
                        phase_score_accumulator[phase].append(score)
                
                # Component scores
                for comp in ['pose', 'velocity', 'acceleration', 'jerk']:
                    if comp in result.components:
                        component_accumulator[comp].append(result.components[comp])
                
                # Store individual result for NLP feedback (only if prediction is correct)
                if generate_nlp_feedback:
                    predicted_cls = classes[y_pred[i]] if i < len(y_pred) else None
                    if predicted_cls == cls:  # Only store correctly predicted samples
                        individual_results.append({
                            'sample_idx': i,
                            'class': cls,
                            'predicted_class': predicted_cls,
                            'ksi_result': result,
                            'ksi_total': result.ksi_total
                        })
                
                evaluated_count += 1
            
            # Report skipped templates
            if skipped_templates:
                print(f"   ⚠ Missing templates for classes: {sorted(skipped_templates)}")
            
            # Aggregate results
            ksi_results['avg_ksi_total'] = float(np.mean(ksi_totals)) if ksi_totals else 0.0
            ksi_results['avg_ksi_weighted'] = float(np.mean(ksi_weighted_list)) if ksi_weighted_list else 0.0
            ksi_results['avg_confidence_ci_width'] = float(np.mean(ci_widths)) if ci_widths else 0.0
            ksi_results['avg_uncertainty_scalar'] = float(np.mean(uncertainty_scalars)) if uncertainty_scalars else 0.0
            ksi_results['reliable_count'] = reliable_count
            ksi_results['total_samples'] = evaluated_count  # Use actual evaluated count, not attempted
            
            for phase, scores in phase_score_accumulator.items():
                if scores:
                    ksi_results['phase_scores'][phase] = float(np.mean(scores))
            
            for comp, scores in component_accumulator.items():
                if scores:
                    ksi_results['component_scores'][comp] = float(np.mean(scores))
            
            # Generate natural language coaching reports
            if generate_nlp_feedback and individual_results and NLP_AVAILABLE:
                print(f"\n{'='*70}")
                print(f"GENERATING NATURAL LANGUAGE COACHING FEEDBACK")
                print(f"{'='*70}")
                print(f"   Found {len(individual_results)} correctly predicted samples")
                
                # DEBUG: Show KSI scores distribution
                ksi_scores = [r['ksi_total'] for r in individual_results]
                print(f"   KSI scores range: {min(ksi_scores):.3f} - {max(ksi_scores):.3f}")
                
                os.makedirs("coaching_reports", exist_ok=True)
                
                # Pick best sample (highest KSI score) - this gives most interesting feedback
                individual_results.sort(key=lambda x: x['ksi_total'], reverse=True)
                best_sample = individual_results[0]
                samples_to_generate = [best_sample]
                
                sample = samples_to_generate[0]
                cls = sample['class']
                ksi_result = sample['ksi_result']
                sample_idx = sample['sample_idx']
                
                print(f"\n📝 Generating coaching report for: {cls} (KSI: {ksi_result.ksi_total:.3f})")
                
                try:
                    
                    # Generate simplified coaching report
                    report = generate_coaching_report(
                        ksi_result=ksi_result,
                        shot_type_str=cls,
                        skill_level_str=nlp_skill_level,
                        user_name=None,
                        output_format='text',
                        simplified=True  # Remove weekly plan, shorten output
                    )
                    
                    # Save report
                    report_filename = f"coaching_reports/{cls}_ksi{ksi_result.ksi_total:.3f}_report.txt"
                    with open(report_filename, 'w') as f:
                        f.write(report)
                    
                    print(f"   ✅ Saved: {report_filename}")
                    
                    # Also save JSON version
                    json_report = generate_coaching_report(
                        ksi_result=ksi_result,
                        shot_type_str=cls,
                        skill_level_str=nlp_skill_level,
                        output_format='json',
                        simplified=True
                    )
                    json_filename = f"coaching_reports/{cls}_ksi{ksi_result.ksi_total:.3f}_report.json"
                    with open(json_filename, 'w') as f:
                        f.write(json_report)
                    
                    # Log to MLflow
                    mlflow.log_artifact(report_filename)
                    print(f"   📊 Logged to MLflow")
                        
                except Exception as e:
                    print(f"   ⚠️ Failed to generate report: {e}")
                    import traceback
                    traceback.print_exc()
                
                print(f"\n✅ Generated coaching report in coaching_reports/")
                print(f"{'='*70}\n")
        elif pipeline_type == 'hybrid' and not has_raw_landmarks:
            print(f"⚠️  Raw landmarks not found in hybrid data. Run preprocessing again to enable KSI evaluation.")
            print(f"    (Old data format detected - missing 'raw_landmarks' in .npz files)")
        else:
            print(f"⚠ Expert templates not found at {template_path}. Skipping KSI metrics.")

        # 4. Logging to MLflow
        mlflow.log_metric("test_accuracy", acc)
        mlflow.log_metric("test_loss", loss)
        mlflow.log_metric("ksi_total", ksi_results['avg_ksi_total'])
        mlflow.log_metric("ksi_weighted", ksi_results['avg_ksi_weighted'])
        mlflow.log_metric("ksi_ci_width", ksi_results['avg_confidence_ci_width'])
        mlflow.log_metric("ksi_uncertainty", ksi_results['avg_uncertainty_scalar'])
        mlflow.log_metric("ksi_reliable_ratio", 
                          ksi_results['reliable_count'] / max(1, ksi_results['total_samples']))
        
        # Log phase scores
        for phase, score in ksi_results['phase_scores'].items():
            mlflow.log_metric(f"phase_{phase}", score)
        
        # Log component scores
        for comp, score in ksi_results['component_scores'].items():
            mlflow.log_metric(f"ksi_{comp}", score)
        
        # 5. Confusion Matrix
        cm = confusion_matrix(np.argmax(y_test, axis=1), y_pred)
        plt.figure(figsize=(8, 6))
        sns.heatmap(cm, annot=True, fmt='d', xticklabels=classes, yticklabels=classes, cmap='Blues')
        plt.title(f"Confusion Matrix - {pipeline_type.upper()}")
        
        os.makedirs("dvclive", exist_ok=True)
        # Use production filename for tuned models, pipeline-specific otherwise
        cm_filename = "production_confusion_matrix.png" if "tuned" in cfg['model_path'] else f"{pipeline_type}_confusion_matrix.png"
        cm_path = os.path.join("dvclive", cm_filename)
        plt.savefig(cm_path)
        plt.close()
        mlflow.log_artifact(cm_path)
        
        # 6. KSI Analysis Plot
        if ksi_results['phase_scores'] or ksi_results['component_scores']:
            fig, axes = plt.subplots(1, 2, figsize=(12, 5))
            
            # Phase scores
            if ksi_results['phase_scores']:
                phases = list(ksi_results['phase_scores'].keys())
                p_scores = list(ksi_results['phase_scores'].values())
                axes[0].barh(phases, p_scores, color=['gray', 'blue', 'orange', 'red', 'green'][:len(phases)])
                axes[0].set_xlim([0, 1])
                axes[0].set_xlabel('Average Score')
                axes[0].set_title('Phase Scores (Avg)')
            
            # Component scores
            if ksi_results['component_scores']:
                comps = list(ksi_results['component_scores'].keys())
                c_scores = list(ksi_results['component_scores'].values())
                axes[1].bar(comps, c_scores, color=['steelblue', 'coral', 'seagreen', 'orchid'][:len(comps)])
                axes[1].set_ylim([0, 1])
                axes[1].set_ylabel('Average Score')
                axes[1].set_title('Component Scores (Avg)')
            
            plt.suptitle(f"KSI v2 Analysis - {pipeline_type.upper()} | "
                         f"KSI: {ksi_results['avg_ksi_total']:.3f} ± {ksi_results['avg_confidence_ci_width']:.3f}")
            plt.tight_layout()
            ksi_plot_path = f"dvclive/{pipeline_type}_ksi_analysis.png"
            plt.savefig(ksi_plot_path)
            plt.close()
            mlflow.log_artifact(ksi_plot_path)

        # 7. Save Metrics for DVC
        metrics = {
            "accuracy": float(acc),
            "loss": float(loss),
            "ksi_total": ksi_results['avg_ksi_total'],
            "ksi_weighted": ksi_results['avg_ksi_weighted'],
            "ksi_ci_width": ksi_results['avg_confidence_ci_width'],
            "ksi_uncertainty": ksi_results['avg_uncertainty_scalar'],
            "ksi_reliable_ratio": ksi_results['reliable_count'] / max(1, ksi_results['total_samples']),
            "phase_scores": ksi_results['phase_scores'],
            "component_scores": ksi_results['component_scores']
        }
        
        # Determine metrics filename based on model path
        metrics_filename = "production_metrics.json" if "tuned" in cfg['model_path'] else f"{pipeline_type}_metrics.json"
        metrics_path = os.path.join("dvclive", metrics_filename)
        
        with open(metrics_path, "w") as f:
            json.dump(metrics, f, indent=2)

        print(f"\n{'='*60}")
        print(f"{pipeline_type.upper()} EVALUATION (KSI v2)")
        print(f"{'='*60}")
        print(f"Accuracy: {acc:.4f} | Loss: {loss:.4f}")
        print(f"KSI Total: {ksi_results['avg_ksi_total']:.4f}")
        print(f"KSI Weighted: {ksi_results['avg_ksi_weighted']:.4f}")
        print(f"CI Width: {ksi_results['avg_confidence_ci_width']:.4f}")
        print(f"Uncertainty: {ksi_results['avg_uncertainty_scalar']:.4f}")
        print(f"Reliable: {ksi_results['reliable_count']}/{ksi_results['total_samples']}")
        print(f"Phase scores: {ksi_results['phase_scores']}")
        print(f"Component scores: {ksi_results['component_scores']}")
        print(f"{'='*60}\n")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--type", choices=['pose', 'hybrid'], required=True)
    parser.add_argument("--model", type=str, help="Optional: custom model path (overrides params.yaml)")
    parser.add_argument("--data", type=str, help="Optional: custom data path for evaluation (overrides params.yaml)")
    parser.add_argument("--nlp", action='store_true', help="Generate natural language coaching reports")
    parser.add_argument("--nlp-skill", type=str, default='intermediate', 
                        choices=['beginner', 'intermediate', 'advanced', 'expert'],
                        help="Skill level for natural language feedback (default: intermediate)")
    parser.add_argument("--nlp-samples", type=int, default=5,
                        help="Max number of samples to generate detailed feedback for (default: 5)")
    parser.add_argument("--auto-name", action='store_true', 
                        help="Auto-generate MLflow run name without prompting")
    parser.add_argument("--gpu", action='store_true', help="Use GPU for inference (faster but less deterministic)")
    args = parser.parse_args()
    evaluate(args.type, args.model, args.nlp, args.nlp_skill, args.nlp_samples, args.auto_name, args.data)