Spaces:
Running
on
Zero
Running
on
Zero
File size: 13,626 Bytes
cc0c59d |
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 |
from typing import *
import copy
import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader
import numpy as np
from easydict import EasyDict as edict
from ..basic import BasicTrainer
from ...pipelines import samplers
from ...utils.general_utils import dict_reduce
from .mixins.classifier_free_guidance import ClassifierFreeGuidanceMixin
from .mixins.text_conditioned import TextConditionedMixin
from .mixins.image_conditioned import ImageConditionedMixin
class FlowMatchingTrainer(BasicTrainer):
"""
Trainer for diffusion model with flow matching objective.
Args:
models (dict[str, nn.Module]): Models to train.
dataset (torch.utils.data.Dataset): Dataset.
output_dir (str): Output directory.
load_dir (str): Load directory.
step (int): Step to load.
batch_size (int): Batch size.
batch_size_per_gpu (int): Batch size per GPU. If specified, batch_size will be ignored.
batch_split (int): Split batch with gradient accumulation.
max_steps (int): Max steps.
optimizer (dict): Optimizer config.
lr_scheduler (dict): Learning rate scheduler config.
elastic (dict): Elastic memory management config.
grad_clip (float or dict): Gradient clip config.
ema_rate (float or list): Exponential moving average rates.
fp16_mode (str): FP16 mode.
- None: No FP16.
- 'inflat_all': Hold a inflated fp32 master param for all params.
- 'amp': Automatic mixed precision.
fp16_scale_growth (float): Scale growth for FP16 gradient backpropagation.
finetune_ckpt (dict): Finetune checkpoint.
log_param_stats (bool): Log parameter stats.
i_print (int): Print interval.
i_log (int): Log interval.
i_sample (int): Sample interval.
i_save (int): Save interval.
i_ddpcheck (int): DDP check interval.
t_schedule (dict): Time schedule for flow matching.
sigma_min (float): Minimum noise level.
"""
def __init__(
self,
*args,
t_schedule: dict = {
'name': 'logitNormal',
'args': {
'mean': 0.0,
'std': 1.0,
}
},
sigma_min: float = 1e-5,
**kwargs
):
super().__init__(*args, **kwargs)
self.t_schedule = t_schedule
self.sigma_min = sigma_min
def diffuse(self, x_0: torch.Tensor, t: torch.Tensor, noise: Optional[torch.Tensor] = None) -> torch.Tensor:
"""
Diffuse the data for a given number of diffusion steps.
In other words, sample from q(x_t | x_0).
Args:
x_0: The [N x C x ...] tensor of noiseless inputs.
t: The [N] tensor of diffusion steps [0-1].
noise: If specified, use this noise instead of generating new noise.
Returns:
x_t, the noisy version of x_0 under timestep t.
"""
if noise is None:
noise = torch.randn_like(x_0)
assert noise.shape == x_0.shape, "noise must have same shape as x_0"
t = t.view(-1, *[1 for _ in range(len(x_0.shape) - 1)])
x_t = (1 - t) * x_0 + (self.sigma_min + (1 - self.sigma_min) * t) * noise
return x_t
def reverse_diffuse(self, x_t: torch.Tensor, t: torch.Tensor, noise: torch.Tensor) -> torch.Tensor:
"""
Get original image from noisy version under timestep t.
"""
assert noise.shape == x_t.shape, "noise must have same shape as x_t"
t = t.view(-1, *[1 for _ in range(len(x_t.shape) - 1)])
x_0 = (x_t - (self.sigma_min + (1 - self.sigma_min) * t) * noise) / (1 - t)
return x_0
def get_v(self, x_0: torch.Tensor, noise: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
"""
Compute the velocity of the diffusion process at time t.
"""
return (1 - self.sigma_min) * noise - x_0
def get_cond(self, cond, **kwargs):
"""
Get the conditioning data.
"""
return cond
def get_inference_cond(self, cond, **kwargs):
"""
Get the conditioning data for inference.
"""
return {'cond': cond, **kwargs}
def get_sampler(self, **kwargs) -> samplers.FlowEulerSampler:
"""
Get the sampler for the diffusion process.
"""
return samplers.FlowEulerSampler(self.sigma_min)
def vis_cond(self, **kwargs):
"""
Visualize the conditioning data.
"""
return {}
def sample_t(self, batch_size: int) -> torch.Tensor:
"""
Sample timesteps.
"""
if self.t_schedule['name'] == 'uniform':
t = torch.rand(batch_size)
elif self.t_schedule['name'] == 'logitNormal':
mean = self.t_schedule['args']['mean']
std = self.t_schedule['args']['std']
t = torch.sigmoid(torch.randn(batch_size) * std + mean)
else:
raise ValueError(f"Unknown t_schedule: {self.t_schedule['name']}")
return t
def training_losses(
self,
x_0: torch.Tensor,
cond=None,
**kwargs
) -> Tuple[Dict, Dict]:
"""
Compute training losses for a single timestep.
Args:
x_0: The [N x C x ...] tensor of noiseless inputs.
cond: The [N x ...] tensor of additional conditions.
kwargs: Additional arguments to pass to the backbone.
Returns:
a dict with the key "loss" containing a tensor of shape [N].
may also contain other keys for different terms.
"""
noise = torch.randn_like(x_0)
t = self.sample_t(x_0.shape[0]).to(x_0.device).float()
x_t = self.diffuse(x_0, t, noise=noise)
cond = self.get_cond(cond, **kwargs)
pred = self.training_models['denoiser'](x_t, t * 1000, cond, **kwargs)
assert pred.shape == noise.shape == x_0.shape
target = self.get_v(x_0, noise, t)
terms = edict()
terms["mse"] = F.mse_loss(pred, target)
terms["loss"] = terms["mse"]
# log loss with time bins
mse_per_instance = np.array([
F.mse_loss(pred[i], target[i]).item()
for i in range(x_0.shape[0])
])
time_bin = np.digitize(t.cpu().numpy(), np.linspace(0, 1, 11)) - 1
for i in range(10):
if (time_bin == i).sum() != 0:
terms[f"bin_{i}"] = {"mse": mse_per_instance[time_bin == i].mean()}
return terms, {}
@torch.no_grad()
def run_snapshot(
self,
num_samples: int,
batch_size: int,
verbose: bool = False,
) -> Dict:
dataloader = DataLoader(
copy.deepcopy(self.dataset),
batch_size=batch_size,
shuffle=True,
num_workers=0,
collate_fn=self.dataset.collate_fn if hasattr(self.dataset, 'collate_fn') else None,
)
# inference
sampler = self.get_sampler()
sample_gt = []
sample = []
cond_vis = []
for i in range(0, num_samples, batch_size):
batch = min(batch_size, num_samples - i)
data = next(iter(dataloader))
data = {k: v[:batch].cuda() if isinstance(v, torch.Tensor) else v[:batch] for k, v in data.items()}
noise = torch.randn_like(data['x_0'])
sample_gt.append(data['x_0'])
cond_vis.append(self.vis_cond(**data))
del data['x_0']
args = self.get_inference_cond(**data)
res = sampler.sample(
self.models['denoiser'],
noise=noise,
**args,
steps=50, cfg_strength=3.0, verbose=verbose,
)
sample.append(res.samples)
sample_gt = torch.cat(sample_gt, dim=0)
sample = torch.cat(sample, dim=0)
sample_dict = {
'sample_gt': {'value': sample_gt, 'type': 'sample'},
'sample': {'value': sample, 'type': 'sample'},
}
sample_dict.update(dict_reduce(cond_vis, None, {
'value': lambda x: torch.cat(x, dim=0),
'type': lambda x: x[0],
}))
return sample_dict
class FlowMatchingCFGTrainer(ClassifierFreeGuidanceMixin, FlowMatchingTrainer):
"""
Trainer for diffusion model with flow matching objective and classifier-free guidance.
Args:
models (dict[str, nn.Module]): Models to train.
dataset (torch.utils.data.Dataset): Dataset.
output_dir (str): Output directory.
load_dir (str): Load directory.
step (int): Step to load.
batch_size (int): Batch size.
batch_size_per_gpu (int): Batch size per GPU. If specified, batch_size will be ignored.
batch_split (int): Split batch with gradient accumulation.
max_steps (int): Max steps.
optimizer (dict): Optimizer config.
lr_scheduler (dict): Learning rate scheduler config.
elastic (dict): Elastic memory management config.
grad_clip (float or dict): Gradient clip config.
ema_rate (float or list): Exponential moving average rates.
fp16_mode (str): FP16 mode.
- None: No FP16.
- 'inflat_all': Hold a inflated fp32 master param for all params.
- 'amp': Automatic mixed precision.
fp16_scale_growth (float): Scale growth for FP16 gradient backpropagation.
finetune_ckpt (dict): Finetune checkpoint.
log_param_stats (bool): Log parameter stats.
i_print (int): Print interval.
i_log (int): Log interval.
i_sample (int): Sample interval.
i_save (int): Save interval.
i_ddpcheck (int): DDP check interval.
t_schedule (dict): Time schedule for flow matching.
sigma_min (float): Minimum noise level.
p_uncond (float): Probability of dropping conditions.
"""
pass
class TextConditionedFlowMatchingCFGTrainer(TextConditionedMixin, FlowMatchingCFGTrainer):
"""
Trainer for text-conditioned diffusion model with flow matching objective and classifier-free guidance.
Args:
models (dict[str, nn.Module]): Models to train.
dataset (torch.utils.data.Dataset): Dataset.
output_dir (str): Output directory.
load_dir (str): Load directory.
step (int): Step to load.
batch_size (int): Batch size.
batch_size_per_gpu (int): Batch size per GPU. If specified, batch_size will be ignored.
batch_split (int): Split batch with gradient accumulation.
max_steps (int): Max steps.
optimizer (dict): Optimizer config.
lr_scheduler (dict): Learning rate scheduler config.
elastic (dict): Elastic memory management config.
grad_clip (float or dict): Gradient clip config.
ema_rate (float or list): Exponential moving average rates.
fp16_mode (str): FP16 mode.
- None: No FP16.
- 'inflat_all': Hold a inflated fp32 master param for all params.
- 'amp': Automatic mixed precision.
fp16_scale_growth (float): Scale growth for FP16 gradient backpropagation.
finetune_ckpt (dict): Finetune checkpoint.
log_param_stats (bool): Log parameter stats.
i_print (int): Print interval.
i_log (int): Log interval.
i_sample (int): Sample interval.
i_save (int): Save interval.
i_ddpcheck (int): DDP check interval.
t_schedule (dict): Time schedule for flow matching.
sigma_min (float): Minimum noise level.
p_uncond (float): Probability of dropping conditions.
text_cond_model(str): Text conditioning model.
"""
pass
class ImageConditionedFlowMatchingCFGTrainer(ImageConditionedMixin, FlowMatchingCFGTrainer):
"""
Trainer for image-conditioned diffusion model with flow matching objective and classifier-free guidance.
Args:
models (dict[str, nn.Module]): Models to train.
dataset (torch.utils.data.Dataset): Dataset.
output_dir (str): Output directory.
load_dir (str): Load directory.
step (int): Step to load.
batch_size (int): Batch size.
batch_size_per_gpu (int): Batch size per GPU. If specified, batch_size will be ignored.
batch_split (int): Split batch with gradient accumulation.
max_steps (int): Max steps.
optimizer (dict): Optimizer config.
lr_scheduler (dict): Learning rate scheduler config.
elastic (dict): Elastic memory management config.
grad_clip (float or dict): Gradient clip config.
ema_rate (float or list): Exponential moving average rates.
fp16_mode (str): FP16 mode.
- None: No FP16.
- 'inflat_all': Hold a inflated fp32 master param for all params.
- 'amp': Automatic mixed precision.
fp16_scale_growth (float): Scale growth for FP16 gradient backpropagation.
finetune_ckpt (dict): Finetune checkpoint.
log_param_stats (bool): Log parameter stats.
i_print (int): Print interval.
i_log (int): Log interval.
i_sample (int): Sample interval.
i_save (int): Save interval.
i_ddpcheck (int): DDP check interval.
t_schedule (dict): Time schedule for flow matching.
sigma_min (float): Minimum noise level.
p_uncond (float): Probability of dropping conditions.
image_cond_model (str): Image conditioning model.
"""
pass
|