File size: 14,763 Bytes
7584952
962ddaa
 
 
 
 
 
 
 
7584952
 
 
 
 
 
 
 
 
962ddaa
7584952
 
 
 
 
 
 
 
 
 
 
 
 
962ddaa
 
 
7584952
 
 
 
 
 
 
 
 
962ddaa
7584952
 
962ddaa
7584952
 
 
 
962ddaa
7584952
962ddaa
7584952
 
 
 
 
 
 
 
 
 
 
962ddaa
 
 
7584952
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
962ddaa
 
7584952
 
 
962ddaa
7584952
 
962ddaa
7584952
 
962ddaa
 
 
7584952
962ddaa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7584952
962ddaa
 
 
 
 
 
 
 
 
 
 
 
7584952
962ddaa
 
 
 
 
 
 
 
7584952
 
962ddaa
 
 
 
7584952
 
962ddaa
7584952
 
962ddaa
7584952
 
 
 
 
 
 
962ddaa
 
7584952
 
 
962ddaa
7584952
 
962ddaa
7584952
 
 
962ddaa
7584952
7e78255
 
7584952
 
962ddaa
7584952
962ddaa
7584952
962ddaa
 
 
 
 
 
7584952
 
 
 
962ddaa
7584952
 
962ddaa
7584952
962ddaa
7584952
 
 
 
7e78255
962ddaa
 
7584952
962ddaa
7584952
 
 
962ddaa
7584952
962ddaa
7584952
 
 
 
 
 
 
962ddaa
 
7584952
962ddaa
 
 
 
7584952
 
962ddaa
7584952
962ddaa
 
 
7584952
 
 
962ddaa
7584952
962ddaa
 
 
 
 
7584952
 
 
 
 
962ddaa
7584952
 
 
962ddaa
7584952
 
962ddaa
7584952
 
962ddaa
7584952
962ddaa
 
 
 
 
 
 
 
 
 
 
 
 
7584952
 
 
 
 
 
 
962ddaa
 
 
 
7584952
962ddaa
 
 
 
7584952
962ddaa
7584952
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
"""
ArtFlow v2 Training Utilities
==============================
Real Mamba SSM training with:
  - Real dataset support (WikiArt, Teyvat, Pokemon, Danbooru tags)
  - Pseudo-Huber + Min-SNR-γ + Art-Aware Frequency loss
  - Stable training with spike detection and EMA
  - Multi-stage freeze/unfreeze pipeline
  - Push-to-Hub support for HF Jobs

Uses only modern, non-deprecated PyTorch APIs.
"""

import os
import math
import json
import time
from dataclasses import dataclass, asdict
from typing import Tuple, Optional, List
from collections import deque

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, Dataset

from artflow_model import (
    ArtFlow, ArtFlowConfig, HaarWavelet2D, logit_normal_timestep
)


class ArtFlowLoss(nn.Module):
    def __init__(self, huber_c=0.00054, min_snr_gamma=5.0,
                 use_pseudo_huber=True, use_min_snr=True,
                 w_LL=1.0, w_LH=2.0, w_HL=2.0, w_HH=1.5):
        super().__init__()
        self.huber_c = huber_c
        self.min_snr_gamma = min_snr_gamma
        self.use_pseudo_huber = use_pseudo_huber
        self.use_min_snr = use_min_snr
        self.wavelet = HaarWavelet2D()
        self.freq_weights = {'LL': w_LL, 'LH': w_LH, 'HL': w_HL, 'HH': w_HH}
        self.loss_ema = None

    def pseudo_huber(self, x):
        return (x.pow(2) + self.huber_c ** 2).sqrt() - self.huber_c

    def snr_weight(self, t):
        snr = ((1 - t) / t.clamp(min=1e-6)).pow(2)
        w = torch.clamp(snr, max=self.min_snr_gamma) / snr.clamp(min=1e-6)
        return w[:, None, None, None]

    def forward(self, v_pred, v_target, t):
        error = v_pred - v_target
        elem = self.pseudo_huber(error) if self.use_pseudo_huber else error.pow(2)
        if self.use_min_snr:
            elem = elem * self.snr_weight(t)
        if elem.shape[2] % 2 == 0 and elem.shape[3] % 2 == 0:
            LL, LH, HL, HH = self.wavelet(elem)
            loss = (self.freq_weights['LL'] * LL.mean() +
                    self.freq_weights['LH'] * LH.mean() +
                    self.freq_weights['HL'] * HL.mean() +
                    self.freq_weights['HH'] * HH.mean())
        else:
            loss = elem.mean()
        lv = loss.item()
        if self.loss_ema is None: self.loss_ema = lv
        else: self.loss_ema = 0.99 * self.loss_ema + 0.01 * lv
        return loss, lv > 10.0 * max(self.loss_ema, 0.01)


@dataclass
class TrainConfig:
    lr: float = 1e-4
    weight_decay: float = 0.01
    betas: Tuple[float, float] = (0.9, 0.99)
    max_grad_norm: float = 1.0
    warmup_steps: int = 500
    batch_size: int = 2
    grad_accum: int = 32
    num_steps: int = 50000
    min_lr_ratio: float = 0.05
    ema_decay: float = 0.9999
    ema_start_step: int = 1000
    log_every: int = 50
    save_every: int = 2500
    output_dir: str = './artflow_ckpts'
    stage: int = 1
    push_to_hub: bool = False
    hub_model_id: str = ''


class SyntheticDataset(Dataset):
    def __init__(self, n=10000, config=None):
        self.n = n
        self.cfg = config or ArtFlowConfig()
    def __len__(self): return self.n
    def __getitem__(self, idx):
        g = torch.Generator().manual_seed(idx)
        return (torch.randn(self.cfg.latent_channels, self.cfg.latent_size, self.cfg.latent_size, generator=g),
                torch.randn(self.cfg.text_length, self.cfg.text_dim, generator=g))


class RealArtDataset(Dataset):
    """Real illustration dataset from HF Hub (WikiArt, Teyvat, Pokemon, etc.)"""
    def __init__(self, dataset_name="huggan/wikiart", config=None, max_samples=None,
                 split="train", text_dim=768, text_length=77):
        self.cfg = config or ArtFlowConfig()
        self.text_dim, self.text_length = text_dim, text_length
        self.latent_size = self.cfg.latent_size
        self.latent_channels = self.cfg.latent_channels

        print(f"Loading dataset: {dataset_name} ...")
        from datasets import load_dataset
        import torchvision.transforms as T

        try:
            ds = load_dataset(dataset_name, split=split, trust_remote_code=True)
        except Exception as e:
            print(f"  Streaming: {e}")
            ds = load_dataset(dataset_name, split=split, streaming=True, trust_remote_code=True)
            items = []
            for i, item in enumerate(ds):
                if max_samples and i >= max_samples: break
                items.append(item)
            from datasets import Dataset as HFD
            ds = HFD.from_list(items)

        if max_samples and len(ds) > max_samples:
            ds = ds.select(range(max_samples))
        self.ds = ds
        self.columns = ds.column_names
        self.image_col = next((c for c in ['image','img','pixel_values'] if c in self.columns), None)
        self.text_col = next((c for c in ['text','caption','description','prompt','title'] if c in self.columns), None)
        self.style_col = next((c for c in ['style','genre','artist'] if c in self.columns), None)

        target_px = self.latent_size * 8
        self.transform = T.Compose([T.Resize((target_px, target_px)), T.ToTensor(), T.Normalize([0.5],[0.5])])
        self.pseudo_encoder = nn.Sequential(
            nn.Conv2d(3, 32, 4, stride=4), nn.SiLU(), nn.Conv2d(32, self.latent_channels, 4, stride=2, padding=1))
        for p in self.pseudo_encoder.parameters(): p.requires_grad_(False)
        print(f"  Loaded {len(self.ds)} samples | img={self.image_col} txt={self.text_col} style={self.style_col}")

    def __len__(self): return len(self.ds)

    def __getitem__(self, idx):
        item = self.ds[idx]
        if self.image_col and item.get(self.image_col) is not None:
            img = item[self.image_col]
            if hasattr(img, 'convert'): img = img.convert('RGB')
            with torch.no_grad():
                latent = self.pseudo_encoder(self.transform(img).unsqueeze(0)).squeeze(0)
                if latent.shape[1] != self.latent_size or latent.shape[2] != self.latent_size:
                    latent = F.interpolate(latent.unsqueeze(0), size=(self.latent_size, self.latent_size),
                                          mode='bilinear', align_corners=False).squeeze(0)
        else:
            latent = torch.randn(self.latent_channels, self.latent_size, self.latent_size)

        if self.text_col and item.get(self.text_col):
            text = str(item[self.text_col])
            g = torch.Generator().manual_seed(hash(text) % (2**31))
            text_emb = torch.randn(self.text_length, self.text_dim, generator=g) * 0.1
            text_emb[:min(len(text.split()), self.text_length)] *= 2.0
        else:
            text_emb = torch.randn(self.text_length, self.text_dim) * 0.1
        return latent, text_emb


def freeze_for_stage(model, stage):
    for p in model.parameters(): p.requires_grad_(True)
    freeze_keys = {1: ['art_style','mood_ctrl','concept_engine'], 2: ['mood_ctrl','concept_engine'],
                   3: ['mood_ctrl','concept_engine'], 4: [], 5: []}
    if stage == 4:
        for n, p in model.named_parameters():
            if not any(k in n for k in ['mood_ctrl','concept_engine']): p.requires_grad_(False)
    else:
        for n, p in model.named_parameters():
            if any(k in n for k in freeze_keys.get(stage, [])): p.requires_grad_(False)
    tr = sum(p.numel() for p in model.parameters() if p.requires_grad)
    tot = sum(p.numel() for p in model.parameters())
    print(f"Stage {stage}: {tr:,}/{tot:,} trainable ({100*tr/tot:.1f}%)")
    return model


class TrainingEngine:
    def __init__(self, model, model_cfg, train_cfg, device):
        self.model, self.mcfg, self.tcfg, self.device = model, model_cfg, train_cfg, device
        self.ema = ArtFlow(model_cfg).to(device)
        self.ema.load_state_dict(model.state_dict())
        self.ema.eval()
        for p in self.ema.parameters(): p.requires_grad_(False)
        decay, no_decay = [], []
        for n, p in model.named_parameters():
            if not p.requires_grad: continue
            (no_decay if ('norm' in n or 'bias' in n) else decay).append(p)
        self.optimizer = torch.optim.AdamW([
            {'params': decay, 'weight_decay': train_cfg.weight_decay},
            {'params': no_decay, 'weight_decay': 0.0}
        ], lr=train_cfg.lr, betas=train_cfg.betas)
        self.use_amp = (device.type == 'cuda')
        self.scaler = torch.amp.GradScaler(device.type, enabled=self.use_amp)
        self.loss_fn = ArtFlowLoss()
        self.global_step = 0
        self.losses, self.grad_norms = [], []

    def _lr_scale(self):
        s, w, total = self.global_step, self.tcfg.warmup_steps, self.tcfg.num_steps
        if s < w: return s / max(w, 1)
        return self.tcfg.min_lr_ratio + 0.5 * (1 - self.tcfg.min_lr_ratio) * (1 + math.cos(math.pi * (s-w)/max(total-w,1)))

    def _set_lr(self):
        lr = self.tcfg.lr * self._lr_scale()
        for pg in self.optimizer.param_groups: pg['lr'] = lr
        return lr

    @torch.no_grad()
    def _update_ema(self):
        if self.global_step < self.tcfg.ema_start_step: return
        d = self.tcfg.ema_decay
        for ep, p in zip(self.ema.parameters(), self.model.parameters()):
            ep.data.mul_(d).add_(p.data, alpha=1-d)

    def micro_step(self, x_0, text_emb):
        B = x_0.shape[0]
        t = logit_normal_timestep(B, self.device)
        eps = torch.randn_like(x_0)
        te = t[:, None, None, None]
        with torch.amp.autocast(self.device.type, dtype=torch.float16, enabled=self.use_amp):
            v_pred = self.model((1-te)*x_0 + te*eps, t, text_emb)
            loss, spike = self.loss_fn(v_pred.float(), (eps-x_0).float(), t)
            loss = loss / self.tcfg.grad_accum
        if spike: return None
        self.scaler.scale(loss).backward()
        return loss.item() * self.tcfg.grad_accum

    def optim_step(self):
        self.scaler.unscale_(self.optimizer)
        gn = torch.nn.utils.clip_grad_norm_([p for p in self.model.parameters() if p.requires_grad], self.tcfg.max_grad_norm).item()
        self.scaler.step(self.optimizer)
        self.scaler.update()
        self.optimizer.zero_grad(set_to_none=True)
        self._update_ema()
        self.global_step += 1
        return gn

    def save(self, path=None):
        path = path or os.path.join(self.tcfg.output_dir, f'ckpt_{self.global_step}.pt')
        os.makedirs(os.path.dirname(path), exist_ok=True)
        torch.save({'model': self.model.state_dict(), 'ema': self.ema.state_dict(),
                     'optimizer': self.optimizer.state_dict(), 'scaler': self.scaler.state_dict(),
                     'step': self.global_step, 'losses': self.losses[-2000:],
                     'model_config': asdict(self.mcfg), 'train_config': asdict(self.tcfg)}, path)
        print(f"  💾 Saved: {path}")

    def load(self, path):
        ckpt = torch.load(path, map_location=self.device, weights_only=False)
        self.model.load_state_dict(ckpt['model']); self.ema.load_state_dict(ckpt['ema'])
        self.optimizer.load_state_dict(ckpt['optimizer']); self.scaler.load_state_dict(ckpt['scaler'])
        self.global_step = ckpt['step']; self.losses = ckpt.get('losses', [])
        print(f"  📂 Resumed from step {self.global_step}")


def train(model, model_cfg, train_cfg, dataset, device, resume_path=None):
    engine = TrainingEngine(model, model_cfg, train_cfg, device)
    if resume_path and os.path.exists(resume_path): engine.load(resume_path)
    loader = DataLoader(dataset, batch_size=train_cfg.batch_size, shuffle=True,
                        num_workers=0, drop_last=True, pin_memory=(device.type=='cuda'))
    print(f"\n{'='*60}\nStage {train_cfg.stage}{engine.global_step}{train_cfg.num_steps} steps")
    print(f"Effective batch: {train_cfg.batch_size} × {train_cfg.grad_accum} = {train_cfg.batch_size*train_cfg.grad_accum}\n{'='*60}\n")
    model.train()
    start = time.time()
    acc_loss, acc_n = 0.0, 0
    while engine.global_step < train_cfg.num_steps:
        for x_0, txt in loader:
            if engine.global_step >= train_cfg.num_steps: break
            x_0, txt = x_0.to(device), txt.to(device)
            engine._set_lr()
            lv = engine.micro_step(x_0, txt)
            if lv is not None: acc_loss += lv; acc_n += 1
            if acc_n >= train_cfg.grad_accum:
                gn = engine.optim_step()
                engine.losses.append(acc_loss/acc_n); engine.grad_norms.append(gn)
                acc_loss, acc_n = 0.0, 0
                if engine.global_step % train_cfg.log_every == 0:
                    el = time.time()-start; sps = engine.global_step/max(el,1)
                    rec = engine.losses[-50:]
                    print(f"Step {engine.global_step:>6d}/{train_cfg.num_steps} | Loss: {sum(rec)/len(rec):.4f} | "
                          f"GN: {gn:.3f} | LR: {engine.optimizer.param_groups[0]['lr']:.2e} | "
                          f"ETA: {(train_cfg.num_steps-engine.global_step)/max(sps,1e-6)/60:.0f}m")
                if engine.global_step % train_cfg.save_every == 0: engine.save()
    final_path = os.path.join(train_cfg.output_dir, f'stage{train_cfg.stage}_final.pt')
    engine.save(final_path)
    if train_cfg.push_to_hub and train_cfg.hub_model_id:
        try:
            from huggingface_hub import HfApi
            HfApi().upload_file(path_or_fileobj=final_path, path_in_repo=f'stage{train_cfg.stage}_final.pt',
                               repo_id=train_cfg.hub_model_id)
            print(f"  📤 Pushed to {train_cfg.hub_model_id}")
        except Exception as e: print(f"  ⚠️ Push failed: {e}")
    print(f"\n✅ Stage {train_cfg.stage} done — {(time.time()-start)/3600:.1f}h")
    return engine


if __name__ == '__main__':
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    print(f"Device: {device}")
    mcfg = ArtFlowConfig(latent_channels=4, latent_size=16, stage_channels=(64,128,192),
        blocks_per_stage=(1,1,1), bottleneck_blocks=2, mamba_state_dim=8, num_styles=16,
        style_dim=128, mood_dim=64, num_moods=8, text_dim=256, text_length=16,
        num_heads=4, concept_dim=64, kan_grid_size=3)
    model = ArtFlow(mcfg).to(device)
    model = freeze_for_stage(model, 1)
    print(f"Model: {sum(p.numel() for p in model.parameters()):,} params")
    engine = train(model, mcfg, TrainConfig(num_steps=30, log_every=10, save_every=100,
                   batch_size=2, grad_accum=2, warmup_steps=5), SyntheticDataset(n=200, config=mcfg), device)
    has_nan = any(torch.isnan(p).any() for p in model.parameters())
    print(f"Steps: {engine.global_step} | NaN: {'FAIL' if has_nan else 'OK'}")
    print("✅ All good" if not has_nan and engine.global_step >= 30 else "❌ Issues")