File size: 13,532 Bytes
dc9f917
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Train the amortized iterative controller against the frozen LeWM.



The controller never sees dataset actions as targets. It proposes a plan,

the frozen predictor says what that plan would cause, and the controller is

scored on how close the imagined outcome lands to the goal latent.

"""

import argparse
import json
import sys
import time
from pathlib import Path

import numpy as np
import torch
from torch.utils.data import DataLoader

sys.path.insert(0, str(Path(__file__).resolve().parents[1]))

from lejepa_control.controller import IterativeController  # noqa: E402
from lejepa_control.data import LatentGoalDataset, split_episodes  # noqa: E402
from lejepa_control.losses import BehaviorDensity, refinement_loss, support_loss  # noqa: E402
from lejepa_control.world_model import load_lewm  # noqa: E402


def parse_args():
    p = argparse.ArgumentParser()
    p.add_argument('--latents', default='data/latents')
    p.add_argument('--density', default='data/runs/density/density.pt')
    p.add_argument('--out', default='data/runs/controller')
    p.add_argument('--steps', type=int, default=20000)
    p.add_argument('--batch-size', type=int, default=32)
    p.add_argument('--lr', type=float, default=1e-4)
    p.add_argument('--weight-decay', type=float, default=1e-4)
    p.add_argument('--horizon', type=int, default=5)
    p.add_argument('--refinements', type=int, default=3)
    p.add_argument('--width', type=int, default=256)
    p.add_argument('--depth', type=int, default=4)
    p.add_argument('--heads', type=int, default=8)
    p.add_argument('--dropout', type=float, default=0.1)
    # exp7: tokens live directly in the world model's own latent coordinates
    # (width must equal latent_dim) instead of an arbitrary learned width
    p.add_argument('--no-latent-proj', action='store_true')
    # exp7: one fused operator Phi (single projection + one transformer,
    # depth doubled to match total F+G layers) instead of the split
    # consequence/refine two-network pipeline
    p.add_argument('--fused', action='store_true')
    p.add_argument(
        '--train-seed', type=int, default=None,
        help='seed torch/numpy at startup; default None keeps unseeded '
             'behavior (torch.manual_seed(0) below is unconditional and '
             'unrelated — this only seeds data/init variation across reps)',
    )
    p.add_argument('--alpha', type=float, default=0.05)
    p.add_argument('--lambda-support', type=float, default=0.01)
    # Horizon-matched arrival: index the goal term by the offset the goal was
    # relabeled from, instead of always scoring at block H. Off by default so
    # the original objective's ablations stay reproducible.
    p.add_argument('--arrival-hold', action='store_true')
    p.add_argument('--hold-weight', type=float, default=0.5)
    # in-process: the latent cache is resident, so workers would each copy
    # ~1 GB on spawn to save no real work
    p.add_argument('--workers', type=int, default=0)
    p.add_argument('--log-every', type=int, default=100)
    p.add_argument('--val-every', type=int, default=1000)
    p.add_argument(
        '--curriculum',
        default='0:2,0.25:3,0.5:5',
        help='fraction_of_training:max_goal_offset, comma separated',
    )
    p.add_argument('--wm-name', default='quentinll/lewm-pusht')
    return p.parse_args()


def parse_curriculum(spec, total_steps):
    stages = []
    for part in spec.split(','):
        frac, offset = part.split(':')
        stages.append((int(float(frac) * total_steps), int(offset)))
    return sorted(stages)


def current_offset(stages, step):
    offset = stages[0][1]
    for start, value in stages:
        if step >= start:
            offset = value
    return offset


@torch.no_grad()
def evaluate(controller, model, loader, device, max_batches=20):
    """Held-out terminal distance, per-refinement gain, and arrival profile.



    ``arrival`` is the distance at each sample's own goal offset q, which is

    what receding-horizon execution actually depends on; ``terminal`` is the

    distance at block H regardless of q. A controller that defers arrival

    scores well on terminal and badly on arrival.

    """
    controller.eval()
    terminal, first, arrival, batches = 0.0, 0.0, 0.0, 0
    # mean distance profile over the plan, split by goal offset
    profile = torch.zeros(controller.horizon + 1, controller.horizon, device=device)
    counts = torch.zeros(controller.horizon + 1, device=device)
    for batch in loader:
        q = batch['goal_offset'].to(device).clamp(1, controller.horizon)
        out = controller(
            model,
            batch['context'].to(device),
            batch['past_actions'].to(device),
            batch['goal'].to(device),
        )
        d = out['distances'][-1]
        terminal += d[:, -1].mean().item()
        first += out['distances'][0][:, -1].mean().item()
        arrival += d.gather(1, (q - 1).unsqueeze(1)).squeeze(1).mean().item()
        profile.index_add_(0, q, d)
        counts.index_add_(0, q, torch.ones_like(q, dtype=d.dtype))
        batches += 1
        if batches >= max_batches:
            break
    controller.train()
    profile = (profile / counts.clamp(min=1).unsqueeze(1)).cpu()
    return {
        'terminal': terminal / batches,
        'first': first / batches,
        'arrival': arrival / batches,
        'profile': {
            q: [round(v, 4) for v in profile[q].tolist()]
            for q in range(1, controller.horizon + 1)
            if counts[q] > 0
        },
    }


def main():
    args = parse_args()
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    if args.train_seed is None:
        torch.manual_seed(0)  # current behavior: every run anchored at 0
    else:
        torch.manual_seed(args.train_seed)
        np.random.seed(args.train_seed)

    out_dir = Path(args.out)
    out_dir.mkdir(parents=True, exist_ok=True)

    stats = json.loads((Path(args.latents) / 'stats.json').read_text())
    latent_dim = stats['latent_dim']

    model = load_lewm(name=args.wm_name, device=device)  # frozen, eval, requires_grad_(False)

    # tanh bounds: raw PushT actions live in [-1, 1], and the world model was
    # trained on z-scored actions, so the bound moves into normalized space.
    a_std = torch.tensor(stats['action_std'])
    a_mean = torch.tensor(stats['action_mean'])
    controller = IterativeController(
        latent_dim=latent_dim,
        horizon=args.horizon,
        refinements=args.refinements,
        width=args.width,
        depth=args.depth,
        heads=args.heads,
        dropout=args.dropout,
        action_center=(-a_mean / a_std),
        action_scale=(1.0 / a_std),
        no_latent_proj=args.no_latent_proj,
        fused=args.fused,
    ).to(device)
    n_params = sum(p.numel() for p in controller.parameters())
    print(f'controller params: {n_params / 1e6:.2f}M')

    density, c95 = None, None
    if args.lambda_support > 0 and Path(args.density).exists():
        ckpt = torch.load(args.density, map_location=device)
        density = BehaviorDensity(
            latent_dim=latent_dim, components=ckpt['components']
        ).to(device)
        density.load_state_dict(ckpt['state_dict'])
        density.eval().requires_grad_(False)
        c95 = ckpt['c95']
        print(f'support model loaded, c95={c95:.4f}')
    else:
        print('no support model — training with goal loss only')

    train_eps, val_eps = split_episodes(stats['n_episodes'])
    train_set = LatentGoalDataset(
        args.latents, episodes=train_eps, horizon=args.horizon
    )
    val_set = LatentGoalDataset(
        args.latents, max_offset=5, episodes=val_eps, horizon=args.horizon
    )
    loader = DataLoader(
        train_set,
        batch_size=args.batch_size,
        shuffle=True,
        num_workers=args.workers,
        drop_last=True,
        persistent_workers=args.workers > 0,
        pin_memory=True,
    )
    val_loader = DataLoader(
        val_set, batch_size=args.batch_size, shuffle=True
    )

    opt = torch.optim.AdamW(
        controller.parameters(), lr=args.lr, weight_decay=args.weight_decay
    )
    sched = torch.optim.lr_scheduler.OneCycleLR(
        opt, max_lr=args.lr, total_steps=args.steps, pct_start=0.05
    )
    stages = parse_curriculum(args.curriculum, args.steps)
    print(f'curriculum: {stages}')

    step, t0 = 0, time.perf_counter()
    running = {}
    controller.train()
    while step < args.steps:
        for batch in loader:
            offset = current_offset(stages, step)
            if train_set.max_offset != offset:
                train_set.set_max_offset(offset)

            ctx = batch['context'].to(device, non_blocking=True)
            past = batch['past_actions'].to(device, non_blocking=True)
            goal = batch['goal'].to(device, non_blocking=True)
            q = batch['goal_offset'].to(device, non_blocking=True)

            out = controller(model, ctx, past, goal)
            if args.arrival_hold:
                loss_refine = refinement_loss(
                    out['distances'],
                    goal_offset=q,
                    hold_weight=args.hold_weight,
                )
            else:
                loss_refine = refinement_loss(out['distances'], alpha=args.alpha)

            loss = loss_refine
            violation = torch.zeros((), device=device)
            if density is not None:
                # every refinement after the first, every horizon step
                contexts = torch.cat(out['contexts'][1:], dim=0).flatten(0, 1)
                blocks = torch.cat(out['plans'][1:], dim=0).flatten(0, 1)
                loss_support, violation = support_loss(
                    density, contexts, blocks, c95
                )
                loss = loss + args.lambda_support * loss_support
                running['support'] = (
                    running.get('support', 0.0) + loss_support.item()
                )

            opt.zero_grad(set_to_none=True)
            loss.backward()
            grad = torch.nn.utils.clip_grad_norm_(controller.parameters(), 1.0)
            opt.step()
            sched.step()

            d_first = out['distances'][0][:, -1].mean().item()
            d_last = out['distances'][-1][:, -1].mean().item()
            d_arrival = (
                out['distances'][-1]
                .gather(1, (q.clamp(1, args.horizon) - 1).unsqueeze(1))
                .squeeze(1)
                .mean()
                .item()
            )
            running['loss'] = running.get('loss', 0.0) + loss.item()
            running['terminal'] = running.get('terminal', 0.0) + d_last
            running['arrival'] = running.get('arrival', 0.0) + d_arrival
            running['gain'] = running.get('gain', 0.0) + (d_first - d_last)
            running['violation'] = (
                running.get('violation', 0.0) + violation.item()
            )
            running['grad'] = running.get('grad', 0.0) + grad.item()

            step += 1
            if step % args.log_every == 0:
                n = args.log_every
                rate = step / (time.perf_counter() - t0)
                msg = (
                    f'step {step:6d}  H_goal<={offset}  '
                    f'loss {running["loss"] / n:.4f}  '
                    f'terminal {running["terminal"] / n:.4f}  '
                    f'arrival {running["arrival"] / n:.4f}  '
                    f'gain {running["gain"] / n:+.4f}  '
                    f'viol {running["violation"] / n:.3f}  '
                    f'grad {running["grad"] / n:.2f}  '
                    f'{rate:.1f} it/s'
                )
                if 'support' in running:
                    msg += f'  support {running["support"] / n:.4f}'
                print(msg, flush=True)
                running = {}

            if step % args.val_every == 0 or step == args.steps:
                val = evaluate(controller, model, val_loader, device)
                print(
                    f'  [val] terminal K={args.refinements}: '
                    f'{val["terminal"]:.4f}  K=0: {val["first"]:.4f}  '
                    f'gain {val["first"] - val["terminal"]:+.4f}  '
                    f'arrival {val["arrival"]:.4f}  '
                    f'steps {controller.step_sizes_repr()}',
                    flush=True,
                )
                for q_val, prof in val['profile'].items():
                    print(f'    q={q_val}: {prof}', flush=True)
                torch.save(
                    {
                        'state_dict': controller.state_dict(),
                        'args': vars(args),
                        'step': step,
                        'val_terminal': val['terminal'],
                        'val_arrival': val['arrival'],
                        'val_profile': val['profile'],
                        'action_mean': stats['action_mean'],
                        'action_std': stats['action_std'],
                    },
                    out_dir / 'controller.pt',
                )

            if step >= args.steps:
                break

    print(f'done in {(time.perf_counter() - t0) / 60:.1f} min -> {out_dir}')


if __name__ == '__main__':
    main()