File size: 10,932 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
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
#!/usr/bin/env python3
"""
Experiment A: Missing-modality robustness for scene recognition (T1).

Train a late-fusion Transformer on all 5 modalities with random per-sample
modality dropout. At test time, systematically evaluate every modality subset
(single modalities, leave-one-out, and full set) by zeroing out the
slices of the concatenated input tensor that correspond to the dropped
modalities.

Reuses: experiments.dataset.get_dataloaders, experiments.models.build_model,
and the pretrained-backbone-transfer helper from train_exp1.py.
"""

import os
import sys
import json
import time
import random
import argparse
import itertools
import numpy as np
import torch
import torch.nn as nn
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 get_dataloaders, NUM_CLASSES
from nets.models import build_model
from tasks.train_exp1 import (
    set_seed, apply_augmentation, _load_and_freeze_backbone,
)


def modality_slices(modality_dims):
    """Return {mod_name: (start, end)} byte-offsets into the concatenated feature dim."""
    slices = {}
    off = 0
    for name, dim in modality_dims.items():
        slices[name] = (off, off + dim)
        off += dim
    return slices


def mask_modalities(x, slices, active_mods):
    """Zero out the slices of x corresponding to modalities NOT in active_mods.

    x: (B, T, F_total)
    Returns a new tensor; does not mutate x in place.
    """
    if set(active_mods) == set(slices.keys()):
        return x
    x2 = x.clone()
    for name, (s, e) in slices.items():
        if name not in active_mods:
            x2[..., s:e] = 0.0
    return x2


def train_one_epoch_with_dropout(model, loader, criterion, optimizer, device,
                                 slices, mod_dropout_p=0.0,
                                 augment=False, noise_std=0.1, time_mask_ratio=0.1):
    """Train one epoch. With probability mod_dropout_p, for each training sample
    independently drop a random non-empty subset of modalities.

    Strategy: for each sample, flip an independent Bernoulli(p) per modality;
    if ALL modalities would be dropped, keep one at random.
    """
    model.train()
    mods = list(slices.keys())
    total_loss = 0.0
    all_preds, all_labels = [], []

    for x, y, mask, _ in loader:
        x, y, mask = x.to(device), y.to(device), mask.to(device)
        if augment:
            x = apply_augmentation(x, mask, noise_std, time_mask_ratio)

        if mod_dropout_p > 0:
            B = x.size(0)
            for i in range(B):
                dropped = [m for m in mods if random.random() < mod_dropout_p]
                # ensure at least one modality survives
                if len(dropped) == len(mods):
                    dropped = random.sample(dropped, len(dropped) - 1)
                for m in dropped:
                    s, e = slices[m]
                    x[i, :, s:e] = 0.0

        optimizer.zero_grad()
        logits = model(x, mask)
        loss = criterion(logits, y)
        loss.backward()
        torch.nn.utils.clip_grad_norm_(
            [p for p in model.parameters() if p.requires_grad], 1.0
        )
        optimizer.step()

        total_loss += loss.item() * y.size(0)
        all_preds.extend(logits.argmax(dim=1).cpu().numpy())
        all_labels.extend(y.cpu().numpy())

    n = len(all_labels)
    return total_loss / n, accuracy_score(all_labels, all_preds)


@torch.no_grad()
def evaluate_with_mask(model, loader, criterion, device, slices, active_mods):
    model.eval()
    total_loss = 0.0
    all_preds, all_labels = [], []
    for x, y, mask, _ in loader:
        x, y, mask = x.to(device), y.to(device), mask.to(device)
        x = mask_modalities(x, slices, set(active_mods))
        logits = model(x, mask)
        loss = criterion(logits, y)
        total_loss += loss.item() * y.size(0)
        all_preds.extend(logits.argmax(dim=1).cpu().numpy())
        all_labels.extend(y.cpu().numpy())
    n = len(all_labels)
    if n == 0:
        return 0.0, 0.0, 0.0, np.zeros((NUM_CLASSES, NUM_CLASSES), dtype=int)
    acc = accuracy_score(all_labels, all_preds)
    f1 = f1_score(all_labels, all_preds, average='macro', zero_division=0)
    cm = confusion_matrix(all_labels, all_preds, labels=list(range(NUM_CLASSES)))
    return total_loss / n, acc, f1, cm


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(',')
    print(f"Model: {args.model} | Fusion: {args.fusion} | Modalities: {modalities}")
    print(f"Training dropout p={args.mod_dropout_p}")

    train_loader, val_loader, test_loader, info = get_dataloaders(
        modalities, batch_size=args.batch_size, downsample=args.downsample
    )
    if info['val_size'] == 0:
        val_loader = test_loader
    print(f"Train: {info['train_size']}, Test: {info['test_size']}")
    print(f"Feature dim: {info['feat_dim']}, Modality dims: {info['modality_dims']}")

    slices = modality_slices(info['modality_dims'])
    print(f"Modality slices: {slices}")

    model = build_model(
        args.model, args.fusion, info['feat_dim'],
        info['modality_dims'], info['num_classes'],
        hidden_dim=args.hidden_dim, proj_dim=args.proj_dim,
        late_agg=args.late_agg,
    ).to(device)

    # Optional pretrained backbone loading (per-modality)
    if args.pretrained_dir:
        for i, mod in enumerate(modalities):
            pt_path = os.path.join(args.pretrained_dir,
                                   f"transformer_{mod}_early", "model_best.pt")
            if os.path.exists(pt_path):
                _load_and_freeze_backbone(model, pt_path, i, args.fusion)
            else:
                print(f"  WARN: no pretrained ckpt for {mod} at {pt_path}")

    total = sum(p.numel() for p in model.parameters())
    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    print(f"Params: {trainable:,}/{total:,}")

    class_weights = info['class_weights'].to(device)
    criterion = nn.CrossEntropyLoss(weight=class_weights,
                                    label_smoothing=args.label_smoothing)

    optimizer = torch.optim.Adam(
        filter(lambda p: p.requires_grad, model.parameters()),
        lr=args.lr, weight_decay=args.weight_decay,
    )
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
        optimizer, mode='min', factor=0.5, patience=7, min_lr=1e-6,
    )

    mod_str = '-'.join(modalities)
    exp_name = f"{args.model}_{mod_str}_{args.fusion}_drop{args.mod_dropout_p}_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_val_loss = float('inf')
    best_epoch = 0
    patience_counter = 0

    for epoch in range(1, args.epochs + 1):
        t0 = time.time()
        train_loss, train_acc = train_one_epoch_with_dropout(
            model, train_loader, criterion, optimizer, device,
            slices=slices, mod_dropout_p=args.mod_dropout_p,
            augment=args.augment,
        )
        # Validate on FULL modalities (baseline performance)
        val_loss, val_acc, val_f1, _ = evaluate_with_mask(
            model, val_loader, criterion, device, slices, modalities,
        )
        scheduler.step(val_loss)
        print(f"  E{epoch:3d} | tr_loss {train_loss:.4f} tr_acc {train_acc:.4f} | "
              f"va_loss {val_loss:.4f} va_acc {val_acc:.4f} va_f1 {val_f1:.4f} | "
              f"{time.time()-t0:.1f}s")
        if val_loss < best_val_loss:
            best_val_loss = val_loss
            best_epoch = epoch
            patience_counter = 0
            torch.save(model.state_dict(), os.path.join(out_dir, 'model_best.pt'))
        else:
            patience_counter += 1
        if patience_counter >= args.patience:
            print(f"  Early stop at epoch {epoch} (best {best_epoch})")
            break

    # Restore best model
    model.load_state_dict(torch.load(os.path.join(out_dir, 'model_best.pt'),
                                     weights_only=True))

    # Systematic evaluation: full, leave-one-out, and all singletons
    print("\n=== Robustness Evaluation ===")
    eval_configs = []
    eval_configs.append(('full', modalities))
    for m in modalities:
        remaining = [x for x in modalities if x != m]
        eval_configs.append((f'drop_{m}', remaining))
    for m in modalities:
        eval_configs.append((f'only_{m}', [m]))

    results_matrix = {}
    for name, active in eval_configs:
        _, acc, f1, _ = evaluate_with_mask(
            model, test_loader, criterion, device, slices, active,
        )
        results_matrix[name] = {'active': active, 'acc': float(acc), 'f1': float(f1)}
        print(f"  {name:<15s} mods={active} | acc {acc:.4f} f1 {f1:.4f}")

    results = {
        'experiment': exp_name,
        'training_dropout_p': args.mod_dropout_p,
        'seed': args.seed,
        'best_epoch': best_epoch,
        'eval_configs': results_matrix,
        'train_size': info['train_size'],
        'test_size': info['test_size'],
        'modality_dims': info['modality_dims'],
        'args': vars(args),
    }
    with open(os.path.join(out_dir, 'results.json'), 'w') as f:
        json.dump(results, f, indent=2, ensure_ascii=False)
    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('--modalities', type=str, default='mocap,emg,eyetrack,imu,pressure')
    p.add_argument('--fusion', type=str, default='late')
    p.add_argument('--late_agg', type=str, default='mean')
    p.add_argument('--mod_dropout_p', type=float, default=0.3,
                   help='Per-modality independent dropout prob at training time')
    p.add_argument('--pretrained_dir', type=str, default='',
                   help='Directory with pretrained single-modality ckpts')
    p.add_argument('--epochs', type=int, default=100)
    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('--proj_dim', type=int, default=0)
    p.add_argument('--downsample', type=int, default=5)
    p.add_argument('--patience', type=int, default=15)
    p.add_argument('--label_smoothing', type=float, default=0.1)
    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()