File size: 16,647 Bytes
f0395ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2025 Hansheng Chen

import sys
import torch
import mmcv

from copy import deepcopy
from functools import partial
from mmgen.models.architectures.common import get_module_device
from mmgen.models.builder import MODULES

from . import GaussianFlow
from .piflow_policies import POLICY_CLASSES, GMFlowPolicy
from lakonlab.utils import module_eval


class PiFlowImitationBase(GaussianFlow):

    def __init__(self, *args, policy_type='GMFlow', policy_kwargs=None, **kwargs):
        super().__init__(*args, **kwargs)
        assert policy_type in POLICY_CLASSES, \
            f'Invalid policy: {policy_type}. Supported policies are {list(POLICY_CLASSES.keys())}.'
        self.policy_type = policy_type
        self.policy_class = partial(
            POLICY_CLASSES[policy_type], **policy_kwargs
        ) if policy_kwargs else POLICY_CLASSES[policy_type]

    def policy_rollout(
            self,
            x_t_start: torch.Tensor,  # (B, C, *, H, W)
            sigma_t_start: torch.Tensor,  # (B, 1, *, 1, 1)
            raw_t_start: torch.Tensor,  # (B, )
            raw_t_end: torch.Tensor,  # (B, )
            total_substeps: int,
            policy,
            seq_len=None):
        num_batches = x_t_start.size(0)
        ndim = x_t_start.dim()
        raw_t_start = raw_t_start.reshape(num_batches, *((ndim - 1) * [1]))
        raw_t_end = raw_t_end.reshape(num_batches, *((ndim - 1) * [1]))

        delta_raw_t = raw_t_start - raw_t_end
        num_substeps = (delta_raw_t * total_substeps).round().to(torch.long).clamp(min=1)
        substep_size = delta_raw_t / num_substeps
        max_num_substeps = num_substeps.max()

        raw_t = raw_t_start
        sigma_t = sigma_t_start
        x_t = x_t_start

        for substep_id in range(max_num_substeps.item()):
            u = policy.pi(x_t, sigma_t)
            
            raw_t_minus = (raw_t - substep_size).clamp(min=0)
            sigma_t_minus = self.timestep_sampler.warp_t(raw_t_minus, seq_len=seq_len)
            x_t_minus = x_t + u * (sigma_t_minus - sigma_t)
            
            active_mask = num_substeps > substep_id
            x_t = torch.where(active_mask, x_t_minus, x_t)
            sigma_t = torch.where(active_mask, sigma_t_minus, sigma_t)
            raw_t = torch.where(active_mask, raw_t_minus, raw_t)

        x_t_end = x_t
        sigma_t_end = sigma_t
        t_end = sigma_t_end.flatten() * self.num_timesteps
        return x_t_end, sigma_t_end, t_end

    def policy_average_u(
            self,
            x_t_start: torch.Tensor,  # (B, C, *, H, W)
            sigma_t_start: torch.Tensor,  # (B, 1, *, 1, 1)
            raw_t_start: torch.Tensor,  # (B, )
            raw_t_end: torch.Tensor,  # (B, )
            total_substeps: int,
            policy,
            seq_len=None,
            eps=1e-4):
        num_batches = x_t_start.size(0)
        ndim = x_t_start.dim()
        is_small_length = torch.round((raw_t_start - raw_t_end) * total_substeps) < 2
        pred_mean_u = pred_local_u = None
        if not is_small_length.all():  # mean velocity over the rollout length
            x_t_end, sigma_t_end, _ = self.policy_rollout(
                x_t_start, sigma_t_start, raw_t_start, raw_t_end, total_substeps,
                policy, seq_len=seq_len)
            pred_mean_u = (x_t_start - x_t_end) / (sigma_t_start - sigma_t_end).clamp(min=eps)
        if is_small_length.any():  # numerically stable local velocity
            pred_local_u = policy.pi(x_t_start, sigma_t_start)
        if pred_mean_u is None:
            pred_u = pred_local_u
        elif pred_local_u is None:
            pred_u = pred_mean_u
        else:
            pred_u = torch.where(
                is_small_length.reshape(num_batches, *((ndim - 1) * [1])), pred_local_u, pred_mean_u)
        return pred_u

    @staticmethod
    def get_shape_info(x):
        x_t_dst_shape = x.size()
        bs = x_t_dst_shape[0]
        ndim = len(x_t_dst_shape)
        seq_len = x.shape[2:].numel()
        return ndim, bs, seq_len

    def piid_segment(
            self, teacher, policy, x_t_src, raw_t_src, sigma_t_src, teacher_ratio, segment_size,
            teacher_kwargs, get_x_t_dst=False):
        eps = self.train_cfg.get('eps', 1e-4)
        total_substeps = self.train_cfg.get('total_substeps', 128)
        num_intermediate_states = self.train_cfg.get('num_intermediate_states', 2)
        window_substeps = self.train_cfg.get('window_substeps', 0)

        device = x_t_src.device
        ndim, bs, seq_len = self.get_shape_info(x_t_src)
        if not isinstance(segment_size, torch.Tensor):
            segment_size = torch.tensor(
                [segment_size], dtype=torch.float32, device=device)

        # window size ∆τ ≈ window_substeps / total_substeps
        num_substeps = (segment_size * total_substeps).round().to(torch.long).clamp(min=1)
        substep_size = segment_size / num_substeps
        window_size = torch.minimum(window_substeps * substep_size, segment_size)

        raw_t_dst = raw_t_src - segment_size

        policy_detached = policy.detach()
        if isinstance(policy_detached, GMFlowPolicy):
            gm_dropout = self.train_cfg.get('gm_dropout', 0.0)
            policy_detached.dropout_(gm_dropout)

        # time sampling for scheduled trajectory mixing
        assert not self.timestep_sampler.logit_normal_enable
        student_intervals = torch.rand(
            (bs, num_intermediate_states), device=device
        ) * ((1 - teacher_ratio) * (segment_size - window_size).unsqueeze(-1))
        student_intervals = torch.sort(student_intervals, dim=-1)[0]
        student_intervals = torch.diff(student_intervals, dim=-1, prepend=torch.zeros((bs, 1), device=device))

        teacher_intervals = torch.rand((bs, num_intermediate_states - 1), device=device)
        teacher_intervals = torch.sort(teacher_intervals, dim=-1)[0]
        teacher_intervals = torch.diff(
            teacher_intervals, dim=-1,
            prepend=torch.zeros((bs, 1), device=device),
            append=torch.ones(
                (bs, 1), device=device)
        ) * (teacher_ratio * (segment_size - window_size).unsqueeze(-1))

        x_t = x_t_src
        raw_t = raw_t_src
        sigma_t = sigma_t_src

        all_pred_u = []
        all_tgt_u = []
        all_timesteps = []

        for teacher_step_id in range(num_intermediate_states):
            raw_t_a = (raw_t - student_intervals[:, teacher_step_id]).clamp(min=0)
            raw_t_b = (raw_t_a - teacher_intervals[:, teacher_step_id]).clamp(min=0)

            with torch.no_grad(), module_eval(teacher):
                x_t_a, sigma_t_a, t_a = self.policy_rollout(
                    x_t, sigma_t, raw_t, raw_t_a, total_substeps,
                    policy_detached, seq_len=seq_len)
                tgt_u = teacher(return_u=True, x_t=x_t_a, t=t_a, **teacher_kwargs)
                all_tgt_u.append(tgt_u)
                all_timesteps.append(t_a)

            pred_u = self.policy_average_u(
                x_t_a, sigma_t_a, raw_t_a, raw_t_b - window_size, total_substeps,
                policy, seq_len=seq_len, eps=eps)
            all_pred_u.append(pred_u)

            sigma_t_b = self.timestep_sampler.warp_t(raw_t_b, seq_len=seq_len).reshape(bs, *((ndim - 1) * [1]))
            x_t = x_t_a + tgt_u * (sigma_t_b - sigma_t_a)
            raw_t = raw_t_b
            sigma_t = sigma_t_b

        loss_kwargs = dict(
            u_t_pred=torch.cat(all_pred_u, dim=0),
            u_t=torch.cat(all_tgt_u, dim=0),
            timesteps=torch.cat(all_timesteps, dim=0)
        )
        loss = self.flow_loss(loss_kwargs)

        if get_x_t_dst:
            with torch.no_grad():
                x_t_dst, _, _ = self.policy_rollout(
                    x_t, sigma_t, raw_t, raw_t_dst, total_substeps,
                    policy_detached, seq_len=seq_len)
        else:
            x_t_dst = None

        return loss, x_t_dst, raw_t_dst

    def forward_test(
            self, x_0=None, noise=None, guidance_scale=None,
            test_cfg_override=dict(), show_pbar=False, **kwargs):
        x_t_src = torch.randn_like(x_0) if noise is None else noise
        num_batches = x_t_src.size(0)
        seq_len = x_t_src.shape[2:].numel()  # h * w or t * h * w
        ori_dtype = x_t_src.dtype
        device = x_t_src.device
        x_t_src = x_t_src.float()
        ndim = x_t_src.dim()
        assert ndim in [4, 5], f'Invalid x_t_src shape: {x_t_src.shape}. Expected 4D or 5D tensor.'

        cfg = deepcopy(self.test_cfg)
        cfg.update(test_cfg_override)

        total_substeps = cfg.get('total_substeps', self.num_timesteps)
        eps = cfg.get('eps', 1e-4)
        nfe = cfg['nfe']
        final_step_size_scale = max(cfg.get('final_step_size_scale', 1.0), eps)
        base_segment_size = 1 / (nfe - 1 + final_step_size_scale)

        raw_t_src = torch.ones((num_batches,), dtype=torch.float32, device=device)
        sigma_t_src = self.timestep_sampler.warp_t(raw_t_src, seq_len=seq_len).reshape(
            num_batches, *((ndim - 1) * [1]))
        t_src = sigma_t_src.flatten() * self.num_timesteps

        if show_pbar:
            pbar = mmcv.ProgressBar(self.distill_steps)

        # ========== Main sampling loop ==========
        for step_id in range(nfe):
            is_final_step = step_id == nfe - 1
            if is_final_step:
                segment_size = base_segment_size * final_step_size_scale
            else:
                segment_size = base_segment_size

            raw_t_dst = raw_t_src - segment_size

            denoising_output = self.pred(x_t_src, t_src, **kwargs)
            policy = self.policy_class(
                denoising_output, x_t_src, sigma_t_src, eps=eps)
            if isinstance(policy, GMFlowPolicy) and not is_final_step:
                temperature = cfg.get('temperature', 1.0)
                policy.temperature_(temperature)

            x_t_dst, sigma_t_dst, t_dst = self.policy_rollout(
                x_t_src, sigma_t_src, raw_t_src, raw_t_dst, total_substeps,
                policy, seq_len=seq_len)

            x_t_src = x_t_dst
            raw_t_src = raw_t_dst
            sigma_t_src = sigma_t_dst
            t_src = t_dst

            if show_pbar:
                pbar.update()

        if show_pbar:
            sys.stdout.write('\n')

        return x_t_src.to(ori_dtype)


@MODULES.register_module()
class PiFlowImitation(PiFlowImitationBase):

    def sample_t(self, num_batches, ndim, seq_len=None, device=None):
        eps = self.train_cfg.get('eps', 1e-4)
        nfe = self.train_cfg['nfe']

        final_step_size_scale = max(self.train_cfg.get('final_step_size_scale', 1.0), eps)
        one_minus_final_scale = 1 - final_step_size_scale
        base_segment_size = 1 / (nfe - one_minus_final_scale)
        final_step_size = final_step_size_scale * base_segment_size

        raw_t = self.timestep_sampler(
            num_batches, warp_t=False, scale_t=False, device=device).clamp(min=eps)
        raw_t_src_idx = torch.ceil(
            raw_t / base_segment_size + one_minus_final_scale).clamp(min=1)
        if isinstance(nfe, torch.Tensor):
            raw_t_src_idx = torch.minimum(raw_t_src_idx, nfe)
        else:
            raw_t_src_idx = raw_t_src_idx.clamp(max=nfe)
        raw_t_src = ((raw_t_src_idx - one_minus_final_scale) * base_segment_size).clamp(min=eps, max=1)
        is_final_step = raw_t_src_idx == 1
        segment_size = torch.where(
            is_final_step, final_step_size, base_segment_size)

        sigma_t_src = self.timestep_sampler.warp_t(raw_t_src, seq_len=seq_len).reshape(
            num_batches, *((ndim - 1) * [1]))
        t_src = sigma_t_src.flatten() * self.num_timesteps
        return raw_t_src, sigma_t_src, t_src, segment_size

    def forward_train(self, x_0, teacher=None, teacher_kwargs=dict(), running_status=None, **kwargs):
        device = get_module_device(self)
        num_batches = x_0.size(0)
        seq_len = x_0.shape[2:].numel()  # h * w or t * h * w
        ndim = x_0.dim()
        assert ndim in [4, 5], f'Invalid x_0 shape: {x_0.shape}. Expected 4D or 5D tensor.'

        num_decay_iters = self.train_cfg.get('num_decay_iters', 0)
        if num_decay_iters > 0:
            teacher_ratio = 1 - min(running_status['iteration'], num_decay_iters) / num_decay_iters
            log_vars = dict(teacher_ratio=teacher_ratio)
        else:
            teacher_ratio = 0.0
            log_vars = dict()

        raw_t_src, sigma_t_src, t_src, segment_size = self.sample_t(
            num_batches, ndim, seq_len=seq_len, device=device)
        noise = torch.randn_like(x_0)
        x_t_src, _, _ = self.sample_forward_diffusion(x_0, t_src, noise)

        denoising_output = self.pred(x_t_src, t_src, **kwargs)
        policy = self.policy_class(denoising_output, x_t_src, sigma_t_src)

        loss_diffusion, _, _ = self.piid_segment(
            teacher, policy, x_t_src, raw_t_src, sigma_t_src, teacher_ratio, segment_size,
            teacher_kwargs)

        loss = loss_diffusion
        log_vars.update(self.flow_loss.log_vars)
        log_vars.update(loss_diffusion=float(loss_diffusion))

        return loss, log_vars


@MODULES.register_module()
class PiFlowImitationDataFree(PiFlowImitationBase):

    is_multistep = True

    def forward_initialize(
            self, x_0, teacher=None, teacher_kwargs=dict(), running_status=None, **kwargs):
        device = get_module_device(self)
        num_batches = x_0.size(0)  # x_0 is a dummy input

        num_decay_iters = self.train_cfg.get('num_decay_iters', 0)
        if num_decay_iters > 0:
            teacher_ratio = 1 - min(running_status['iteration'], num_decay_iters) / num_decay_iters
            log_vars = dict(teacher_ratio=teacher_ratio)
        else:
            teacher_ratio = 0.0
            log_vars = dict()

        x_t_src = torch.randn_like(x_0)
        raw_t_src = torch.ones((num_batches,), dtype=torch.float32, device=device)
        step_states = dict(
            step_id=0,
            terminate=False,
            detachable=True,
            teacher_ratio=teacher_ratio,
            x_t_src=x_t_src,
            raw_t_src=raw_t_src,
        )

        return step_states, log_vars

    def forward_train(
            self, x_0, step_states=None, teacher=None, teacher_kwargs=dict(), running_status=None, **kwargs):
        step_id = step_states['step_id']
        teacher_ratio = step_states['teacher_ratio']
        x_t_src = step_states['x_t_src']
        raw_t_src = step_states['raw_t_src']

        num_batches = x_t_src.size(0)
        seq_len = x_t_src.shape[2:].numel()
        ndim = x_t_src.dim()
        assert ndim in [4, 5], f'Invalid x_t_src shape: {x_t_src.shape}. Expected 4D or 5D tensor.'

        eps = self.train_cfg.get('eps', 1e-4)
        nfe = self.train_cfg['nfe']
        final_step_size_scale = max(self.train_cfg.get('final_step_size_scale', 1.0), eps)
        base_segment_size = 1 / (nfe - 1 + final_step_size_scale)
        is_final_step = step_id == nfe - 1
        if is_final_step:
            segment_size = base_segment_size * final_step_size_scale
        else:
            segment_size = base_segment_size

        sigma_t_src = self.timestep_sampler.warp_t(raw_t_src, seq_len=seq_len).reshape(
            num_batches, *((ndim - 1) * [1]))
        t_src = sigma_t_src.flatten() * self.num_timesteps

        denoising_output = self.pred(x_t_src, t_src, **kwargs)
        policy = self.policy_class(denoising_output, x_t_src, sigma_t_src)

        step_loss_diffusion, x_t_dst, raw_t_dst = self.piid_segment(
            teacher, policy, x_t_src, raw_t_src, sigma_t_src, teacher_ratio, segment_size,
            teacher_kwargs, get_x_t_dst=True)

        loss_diffusion = step_loss_diffusion * segment_size  # Weighing by segment size
        loss = loss_diffusion
        log_vars = {k: v * segment_size for k, v in self.flow_loss.log_vars.items()}
        log_vars.update({
            'loss_diffusion': float(loss_diffusion),
            f'loss_diffusion_step{step_id}': float(step_loss_diffusion)
        })

        if step_id < nfe - 1:
            step_states.update(
                step_id=step_id + 1,
                x_t_src=x_t_dst,
                raw_t_src=raw_t_dst)
        else:
            step_states.update(terminate=True)

        return loss, log_vars, step_states

    def forward(self, x_0=None, return_step_states=False, **kwargs):
        if return_step_states:
            return self.forward_initialize(x_0=x_0, **kwargs)
        else:
            return super().forward(x_0=x_0, **kwargs)