File size: 9,959 Bytes
b4b2877
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""
Experiment F: Zero-shot scene generalization.

Leave-one-scene-out evaluation on T1 (scene recognition). For each of the 8
scenes S_k, train on the remaining 7 scenes across all train+test
volunteers, then evaluate on scene S_k only (all volunteers). Since the
held-out scene was never seen during training, the held-out scene's samples
should be distributed over the remaining 7 classes -- so we report the
fraction of held-out samples that get classified into the single nearest
remaining class (dominant neighbor) and macro-F1 on the 7 seen scenes
during training+eval on mixed scenes.

Simpler protocol: train 8-class classifier but WITHOUT scene S_k in the
training set. Evaluate on full test set (all 8 scenes). Measure what the
holdout scene gets misclassified to -- reveals scene similarity and
generalization behavior.
"""

import os
import sys
import json
import time
import argparse
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from sklearn.metrics import accuracy_score, f1_score, confusion_matrix

sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from data.dataset import (
    MultimodalSceneDataset, collate_fn, TRAIN_VOLS, TEST_VOLS, SCENE_LABELS,
    NUM_CLASSES,
)
from nets.models import build_model
from tasks.train_exp1 import set_seed, apply_augmentation


def filter_dataset_by_scene(ds, excluded_scene):
    """Return indices of samples NOT from the excluded scene."""
    idxs = []
    for i, info in enumerate(ds.sample_info):
        if f"/{excluded_scene}" not in info:
            idxs.append(i)
    return idxs


def run_experiment(args):
    set_seed(args.seed)
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    print(f"Device: {device}")
    modalities = args.modalities.split(',')
    held_out = args.held_out_scene
    assert held_out in SCENE_LABELS, f"Unknown scene: {held_out}"
    print(f"Held-out scene: {held_out} (= class {SCENE_LABELS[held_out]})")

    # Full train/test datasets
    print("Loading train data...")
    full_train = MultimodalSceneDataset(TRAIN_VOLS, modalities, args.downsample)
    stats = full_train.get_stats()
    print("Loading test data...")
    full_test = MultimodalSceneDataset(TEST_VOLS, modalities, args.downsample,
                                       stats=stats)

    # Filter train to exclude the held-out scene
    train_idx = filter_dataset_by_scene(full_train, held_out)
    print(f"Train size (7 seen scenes): {len(train_idx)}/{len(full_train)}")

    # For test, split into "seen" (not held-out) and "unseen" (held-out)
    test_seen_idx = filter_dataset_by_scene(full_test, held_out)
    test_unseen_idx = [i for i in range(len(full_test))
                       if i not in test_seen_idx]
    print(f"Test seen: {len(test_seen_idx)} unseen: {len(test_unseen_idx)}")

    train_sub = torch.utils.data.Subset(full_train, train_idx)
    test_seen_sub = torch.utils.data.Subset(full_test, test_seen_idx)
    test_unseen_sub = torch.utils.data.Subset(full_test, test_unseen_idx)

    train_loader = DataLoader(train_sub, batch_size=args.batch_size, shuffle=True,
                              collate_fn=collate_fn)
    test_seen_loader = DataLoader(test_seen_sub, batch_size=args.batch_size,
                                  shuffle=False, collate_fn=collate_fn)
    test_unseen_loader = DataLoader(test_unseen_sub, batch_size=args.batch_size,
                                    shuffle=False, collate_fn=collate_fn)

    # Build model -- keep 8-class head (we train on only 7 seen classes but
    # leave the held-out logit available; it will predict ~0 since never seen)
    model = build_model(
        args.model, args.fusion, full_train.feat_dim,
        full_train.modality_dims, NUM_CLASSES,
        hidden_dim=args.hidden_dim, proj_dim=0, late_agg='mean',
    ).to(device)
    n_params = sum(p.numel() for p in model.parameters())
    print(f"Params: {n_params:,}")

    # Re-weight: give zero weight to held-out class
    class_weights = full_train.get_class_weights().clone().to(device)
    class_weights[SCENE_LABELS[held_out]] = 0.0
    criterion = nn.CrossEntropyLoss(weight=class_weights, label_smoothing=0.1,
                                    ignore_index=SCENE_LABELS[held_out])
    optimizer = torch.optim.Adam(model.parameters(), lr=args.lr,
                                 weight_decay=args.weight_decay)
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
        optimizer, mode='min', factor=0.5, patience=5, min_lr=1e-6,
    )

    exp_name = f"zs_{args.model}_{'-'.join(modalities)}_hold_{held_out}_seed{args.seed}"
    if args.tag:
        exp_name += f"_{args.tag}"
    out_dir = os.path.join(args.output_dir, exp_name)
    os.makedirs(out_dir, exist_ok=True)

    best_seen_f1 = 0.0
    best_state = None
    best_epoch = 0
    patience_counter = 0

    for epoch in range(1, args.epochs + 1):
        t0 = time.time()
        model.train()
        tr_loss, n = 0.0, 0
        for x, y, mask, _ in train_loader:
            x, y, mask = x.to(device), y.to(device), mask.to(device)
            if args.augment:
                x = apply_augmentation(x, mask, 0.1, 0.1)
            optimizer.zero_grad()
            logits = model(x, mask)
            loss = criterion(logits, y)
            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
            optimizer.step()
            tr_loss += loss.item() * y.size(0)
            n += y.size(0)
        tr_loss /= max(n, 1)

        # Eval on seen (7 classes) and unseen (held-out)
        model.eval()
        def run_eval(loader):
            preds, ys, losses = [], [], 0.0
            nn_ = 0
            with torch.no_grad():
                for x, y, mask, _ in loader:
                    x, y, mask = x.to(device), y.to(device), mask.to(device)
                    logits = model(x, mask)
                    losses += criterion(logits, y).item() * y.size(0)
                    nn_ += y.size(0)
                    preds.extend(logits.argmax(dim=1).cpu().numpy())
                    ys.extend(y.cpu().numpy())
            return preds, ys, losses / max(nn_, 1)

        seen_preds, seen_ys, seen_loss = run_eval(test_seen_loader)
        uns_preds, uns_ys, _ = run_eval(test_unseen_loader)

        seen_acc = accuracy_score(seen_ys, seen_preds)
        seen_f1 = f1_score(seen_ys, seen_preds, average='macro',
                           labels=[c for c in range(NUM_CLASSES)
                                   if c != SCENE_LABELS[held_out]],
                           zero_division=0)
        uns_pred_counts = np.bincount(uns_preds, minlength=NUM_CLASSES)
        # What does the unseen scene get mapped to?
        dominant = int(np.argmax(uns_pred_counts))
        dominant_frac = float(uns_pred_counts[dominant] / max(len(uns_preds), 1))
        held_out_pred_frac = float(uns_pred_counts[SCENE_LABELS[held_out]] /
                                   max(len(uns_preds), 1))

        scheduler.step(seen_loss)

        print(f"  E{epoch:3d} | tr {tr_loss:.4f} te {seen_loss:.4f} | "
              f"seen_acc {seen_acc:.3f} f1 {seen_f1:.3f} | "
              f"unseen -> {dominant} ({dominant_frac:.2f}) "
              f"held_out_predicted_frac {held_out_pred_frac:.3f} | "
              f"{time.time()-t0:.1f}s")

        if seen_f1 > best_seen_f1:
            best_seen_f1 = seen_f1
            best_state = {k: v.cpu().clone() for k, v in model.state_dict().items()}
            best_epoch = epoch
            patience_counter = 0
            best_metrics = {
                'seen_acc': float(seen_acc),
                'seen_f1': float(seen_f1),
                'unseen_dominant_class': int(dominant),
                'unseen_dominant_frac': float(dominant_frac),
                'unseen_pred_hist': uns_pred_counts.tolist(),
                'n_unseen': len(uns_preds),
                'held_out_pred_frac': float(held_out_pred_frac),
            }
        else:
            patience_counter += 1
        if patience_counter >= args.patience:
            print(f"  Early stop (best epoch {best_epoch})")
            break

    if best_state is not None:
        torch.save(best_state, os.path.join(out_dir, 'model_best.pt'))

    results = {
        'experiment': exp_name,
        'model': args.model,
        'modalities': modalities,
        'held_out_scene': held_out,
        'held_out_label': SCENE_LABELS[held_out],
        'seed': args.seed,
        'best_epoch': best_epoch,
        'best_metrics': best_metrics,
        'train_size': len(train_sub),
        'test_seen_size': len(test_seen_sub),
        'test_unseen_size': len(test_unseen_sub),
        'args': vars(args),
    }
    with open(os.path.join(out_dir, 'results.json'), 'w') as f:
        json.dump(results, f, indent=2)
    print(f"Saved: {out_dir}/results.json")
    return results


def main():
    p = argparse.ArgumentParser()
    p.add_argument('--model', type=str, default='transformer')
    p.add_argument('--fusion', type=str, default='early')
    p.add_argument('--modalities', type=str, default='mocap,emg,imu')
    p.add_argument('--held_out_scene', type=str, required=True,
                   help='One of s1..s8')
    p.add_argument('--epochs', type=int, default=60)
    p.add_argument('--batch_size', type=int, default=16)
    p.add_argument('--lr', type=float, default=1e-3)
    p.add_argument('--weight_decay', type=float, default=1e-4)
    p.add_argument('--hidden_dim', type=int, default=128)
    p.add_argument('--downsample', type=int, default=5)
    p.add_argument('--patience', type=int, default=12)
    p.add_argument('--augment', action='store_true')
    p.add_argument('--seed', type=int, default=42)
    p.add_argument('--output_dir', type=str, required=True)
    p.add_argument('--tag', type=str, default='')
    args = p.parse_args()
    run_experiment(args)


if __name__ == '__main__':
    main()