Spaces:
Runtime error
Runtime error
Commit
·
0dc3eb6
1
Parent(s):
b887586
adding needed files
Browse files- pom/v_diffusion.py +168 -0
- sample.py +2 -2
pom/v_diffusion.py
ADDED
|
@@ -0,0 +1,168 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# v-diffusion codes for DDPM inpainting. May not be compatible with k-diffusion.
|
| 2 |
+
|
| 3 |
+
# @SuspectT's inpainting codes, Feb 25 2024
|
| 4 |
+
# shared w/ me over Discord:
|
| 5 |
+
# "that's the v-diffusion inpainting with ddpm
|
| 6 |
+
# optimal settings were around 100 steps for the scheduler
|
| 7 |
+
# (ts refering to timesteps here) and resamples was 4"
|
| 8 |
+
|
| 9 |
+
import torch
|
| 10 |
+
from torch import nn
|
| 11 |
+
from typing import Callable
|
| 12 |
+
from tqdm import trange
|
| 13 |
+
import math
|
| 14 |
+
import sys
|
| 15 |
+
|
| 16 |
+
# from kcrowson/v-diffusion-pytorch
|
| 17 |
+
def t_to_alpha_sigma(t):
|
| 18 |
+
"""Returns the scaling factors for the clean image and for the noise, given
|
| 19 |
+
a timestep."""
|
| 20 |
+
return torch.cos(t * math.pi / 2), torch.sin(t * math.pi / 2)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
#class DDPM(SamplerBase):
|
| 25 |
+
class DDPM():
|
| 26 |
+
|
| 27 |
+
def __init__(self, model_fn: Callable = None):
|
| 28 |
+
super().__init__()
|
| 29 |
+
|
| 30 |
+
def _step(
|
| 31 |
+
self, model_fn: Callable, x_t: torch.Tensor, step: int,
|
| 32 |
+
t_now: torch.Tensor, t_next: torch.Tensor,
|
| 33 |
+
callback: Callable, model_args, **sampler_args ) -> torch.Tensor:
|
| 34 |
+
|
| 35 |
+
alpha_now, sigma_now = t_to_alpha_sigma(t_now) # Get alpha / sigma for current timestep.
|
| 36 |
+
alpha_next, sigma_next = t_to_alpha_sigma(t_next) # Get alpha / sigma for next timestep.
|
| 37 |
+
|
| 38 |
+
v_t = model_fn(x_t, t_now.expand(x_t.shape[0]), **model_args) # Expand t to match batch_size which corresponds to x_t.shape[0]
|
| 39 |
+
|
| 40 |
+
eps_t = x_t * sigma_now + v_t * alpha_now
|
| 41 |
+
pred_t = x_t * alpha_now - v_t * sigma_now
|
| 42 |
+
|
| 43 |
+
if callback is not None:
|
| 44 |
+
callback({'step': step, 'x': x_t, 't': t_now, 'pred': pred_t, 'eps': eps_t})
|
| 45 |
+
|
| 46 |
+
return (pred_t * alpha_next + eps_t * sigma_next)
|
| 47 |
+
|
| 48 |
+
def _sample( self, model_fn: Callable, x_t: torch.Tensor, ts: torch.Tensor,
|
| 49 |
+
callback: Callable, model_args, **sampler_args ) -> torch.Tensor:
|
| 50 |
+
|
| 51 |
+
print("Using DDPM Sampler.")
|
| 52 |
+
steps = ts.size(0)
|
| 53 |
+
|
| 54 |
+
use_tqdm = sampler_args.get('use_tqdm')
|
| 55 |
+
use_range = trange if (use_tqdm if (use_tqdm != None) else False) else range
|
| 56 |
+
|
| 57 |
+
for step in use_range(steps - 1):
|
| 58 |
+
x_t = self._step( model_fn, x_t, step, ts[step], ts[step + 1],
|
| 59 |
+
lambda kwargs: callback(**dict(kwargs, steps=steps)) if(callback != None) else None,
|
| 60 |
+
model_args )
|
| 61 |
+
|
| 62 |
+
return x_t
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
def _inpaint(self,
|
| 66 |
+
model_fn: Callable, audio_source: torch.Tensor, mask: torch.Tensor,
|
| 67 |
+
ts: torch.Tensor, resamples: int, callback: Callable, model_args, **sampler_args
|
| 68 |
+
) -> torch.Tensor:
|
| 69 |
+
steps = ts.size(0)
|
| 70 |
+
batch_size = audio_source.size(0)
|
| 71 |
+
alphas, sigmas = t_to_alpha_sigma(ts)
|
| 72 |
+
|
| 73 |
+
# SHH: rescale audio_source to zero mean and unit variance
|
| 74 |
+
audio_source = (audio_source - audio_source.mean()) / audio_source.std()
|
| 75 |
+
|
| 76 |
+
x_t = audio_source
|
| 77 |
+
|
| 78 |
+
use_tqdm = sampler_args.get('use_tqdm')
|
| 79 |
+
use_range = trange if (use_tqdm if (use_tqdm != None) else False) else range
|
| 80 |
+
|
| 81 |
+
for step in use_range(steps - 1):
|
| 82 |
+
print("step, audio_source.min, audio_source.max, alphas[step], sigmas[step] = ", step, audio_source.min(), audio_source.max(), alphas[step], sigmas[step])
|
| 83 |
+
audio_source_noised = audio_source * alphas[step] + torch.randn_like(audio_source) * sigmas[step]
|
| 84 |
+
print("step, audio_source_noised.min, audio_source_noised.max = ", step, audio_source_noised.min(), audio_source_noised.max())
|
| 85 |
+
sigma_dt = torch.sqrt(sigmas[step] ** 2 - sigmas[step + 1] ** 2)
|
| 86 |
+
|
| 87 |
+
for re in range(resamples):
|
| 88 |
+
|
| 89 |
+
#x_t = audio_source_noised * mask + x_t * ~mask
|
| 90 |
+
x_t = audio_source_noised * mask + x_t * (1.0-mask)
|
| 91 |
+
|
| 92 |
+
# from ImageTransformerDenoiserModelV2:
|
| 93 |
+
# def forward(self, x, sigma, aug_cond=None, class_cond=None, mapping_cond=None):
|
| 94 |
+
#v_t = model_fn(x_t, ts[step].expand(batch_size), **model_args)
|
| 95 |
+
print("step, re, x_t.min, x_t.max , sigmas[step]= ", step, re, x_t.min(), x_t.max(), sigmas[step])
|
| 96 |
+
v_t = model_fn(x_t, sigmas[step].expand(batch_size), aug_cond=None, class_cond=None, mapping_cond=None)
|
| 97 |
+
print("step, re, v_t.min, v_t.max = ", step, re, v_t.min(), v_t.max())
|
| 98 |
+
if v_t.isnan().any():
|
| 99 |
+
print("v_t has NaNs.")
|
| 100 |
+
sys.exit(0)
|
| 101 |
+
|
| 102 |
+
eps_t = x_t * sigmas[step] + v_t * alphas[step]
|
| 103 |
+
pred_t = x_t * alphas[step] - v_t * sigmas[step]
|
| 104 |
+
|
| 105 |
+
if callback is not None:
|
| 106 |
+
callback({'steps': steps, 'step': step, 'x': x_t, 't': ts[step], 'pred': pred_t, 'eps': eps_t, 'res': re})
|
| 107 |
+
|
| 108 |
+
if(re < resamples - 1):
|
| 109 |
+
x_t = pred_t * alphas[step] + eps_t * sigmas[step + 1] + sigma_dt * torch.randn_like(x_t)
|
| 110 |
+
else:
|
| 111 |
+
x_t = pred_t * alphas[step + 1] + eps_t * sigmas[step + 1]
|
| 112 |
+
|
| 113 |
+
print("step, re, v_t.min, v_t.max, x_t.min, x_t.max = ", step, re, v_t.min(), v_t.max(), x_t.min(), x_t.max())
|
| 114 |
+
|
| 115 |
+
#sys.exit(0)
|
| 116 |
+
|
| 117 |
+
return (audio_source * mask + x_t * (1.0-mask))
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
def alpha_sigma_to_t(alpha, sigma):
|
| 121 |
+
"""Returns a timestep, given the scaling factors for the clean image and for
|
| 122 |
+
the noise."""
|
| 123 |
+
return torch.atan2(sigma, alpha) / math.pi * 2
|
| 124 |
+
|
| 125 |
+
def log_snr_to_alpha_sigma(log_snr):
|
| 126 |
+
"""Returns the scaling factors for the clean image and for the noise, given
|
| 127 |
+
the log SNR for a timestep."""
|
| 128 |
+
return log_snr.sigmoid().sqrt(), log_snr.neg().sigmoid().sqrt()
|
| 129 |
+
|
| 130 |
+
def get_ddpm_schedule(ddpm_t):
|
| 131 |
+
"""Returns timesteps for the noise schedule from the DDPM paper."""
|
| 132 |
+
log_snr = -torch.special.expm1(1e-4 + 10 * ddpm_t**2).log()
|
| 133 |
+
alpha, sigma = log_snr_to_alpha_sigma(log_snr)
|
| 134 |
+
return alpha_sigma_to_t(alpha, sigma)
|
| 135 |
+
|
| 136 |
+
|
| 137 |
+
#class LogSchedule(SchedulerBase):
|
| 138 |
+
class LogSchedule():
|
| 139 |
+
def __init__(self, device:torch.device = None):
|
| 140 |
+
super().__init__(device)
|
| 141 |
+
|
| 142 |
+
def create(self, steps: int, first: float = 1, last: float = 0, device: torch.device = None, scheduler_args = {'min_log_snr': -10, 'max_log_snr': 10}) -> torch.Tensor:
|
| 143 |
+
ramp = torch.linspace(first, last, steps, device = device if (device != None) else self.device)
|
| 144 |
+
min_log_snr = scheduler_args.get('min_log_snr')
|
| 145 |
+
max_log_snr = scheduler_args.get('max_log_snr')
|
| 146 |
+
return self.get_log_schedule(
|
| 147 |
+
ramp,
|
| 148 |
+
min_log_snr if min_log_snr!=None else -10,
|
| 149 |
+
max_log_snr if max_log_snr!=None else 10,
|
| 150 |
+
)
|
| 151 |
+
|
| 152 |
+
def get_log_schedule(self, t, min_log_snr=-10, max_log_snr=10):
|
| 153 |
+
log_snr = t * (min_log_snr - max_log_snr) + max_log_snr
|
| 154 |
+
alpha = log_snr.sigmoid().sqrt()
|
| 155 |
+
sigma = log_snr.neg().sigmoid().sqrt()
|
| 156 |
+
return torch.atan2(sigma, alpha) / math.pi * 2 # this returns a timestep?
|
| 157 |
+
|
| 158 |
+
|
| 159 |
+
#class CrashSchedule(SchedulerBase):
|
| 160 |
+
class CrashSchedule():
|
| 161 |
+
def __init__(self, device:torch.device = None):
|
| 162 |
+
super().__init__(device)
|
| 163 |
+
|
| 164 |
+
def create(self, steps: int, first: float = 1, last: float = 0, device: torch.device = None, scheduler_args = None) -> torch.Tensor:
|
| 165 |
+
ramp = torch.linspace(first, last, steps, device = device if (device != None) else self.device)
|
| 166 |
+
sigma = torch.sin(ramp * math.pi / 2) ** 2
|
| 167 |
+
alpha = (1 - sigma**2) ** 0.5
|
| 168 |
+
return torch.atan2(sigma, alpha) / math.pi * 2 # this returns a timestep?
|
sample.py
CHANGED
|
@@ -16,9 +16,9 @@ from torchvision import transforms
|
|
| 16 |
|
| 17 |
import k_diffusion as K
|
| 18 |
|
| 19 |
-
from
|
| 20 |
#CHORD_BORDER = 8 # chord border size in pixels
|
| 21 |
-
from
|
| 22 |
|
| 23 |
|
| 24 |
# ---- my mangled sampler that includes repaint
|
|
|
|
| 16 |
|
| 17 |
import k_diffusion as K
|
| 18 |
|
| 19 |
+
from pom.v_diffusion import DDPM, LogSchedule, CrashSchedule
|
| 20 |
#CHORD_BORDER = 8 # chord border size in pixels
|
| 21 |
+
from pom.chords import CHORD_BORDER, img_batch_to_seq_emb, ChordSeqEncoder
|
| 22 |
|
| 23 |
|
| 24 |
# ---- my mangled sampler that includes repaint
|