Spaces:
Paused
Paused
File size: 6,445 Bytes
e9d1f07 | 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 | import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from inspect import isfunction
from functools import partial
import numpy as np
from tqdm.auto import tqdm
from .forward import *
from .samplers import *
from .backbones.unet_convnext import *
class DenoisingDiffusionProcess(nn.Module):
def __init__(self,
generated_channels=3,
loss_fn=F.mse_loss,
schedule='linear',
num_timesteps=1000,
sampler=None
):
super().__init__()
# Basic Params
self.generated_channels=generated_channels
self.num_timesteps=num_timesteps
self.loss_fn=loss_fn
# Forward Process Used for Training
self.forward_process=GaussianForwardProcess(num_timesteps=self.num_timesteps,
schedule=schedule)
self.model=UnetConvNextBlock(dim=64,
dim_mults=(1,2,4,8),
channels=self.generated_channels,
out_dim=self.generated_channels,
with_time_emb=True)
# defaults to a DDPM sampler if None is provided
self.sampler=DDPM_Sampler(num_timesteps=self.num_timesteps) if sampler is None else sampler
@torch.no_grad()
def forward(self,
shape=(256,256),
batch_size=1,
sampler=None,
verbose=False
):
"""
forward() function triggers a complete inference cycle
A custom sampler can be provided as an argument!
"""
# read dimensions
b,c,h,w=batch_size,self.generated_channels,*shape
device=next(self.model.parameters()).device
# select sampler
if sampler is None:
sampler=self.sampler
else:
sampler.to(device)
# time steps list
num_timesteps=sampler.num_timesteps
it=reversed(range(0, num_timesteps))
x_t = torch.randn([b, self.generated_channels, h, w],device=device)
for i in tqdm(it, desc='diffusion sampling', total=num_timesteps) if verbose else it:
t = torch.full((b,), i, device=device, dtype=torch.long)
z_t=self.model(x_t,t) # prediction of noise
x_t = sampler(x_t,t,z_t) # prediction of next state
return x_t
def p_loss(self,output):
"""
Assumes output is in [-1,+1] range
"""
b,c,h,w=output.shape
device=output.device
# loss for training
# input is the optional condition
t = torch.randint(0, self.forward_process.num_timesteps, (b,), device=device).long()
output_noisy, noise=self.forward_process(output,t,return_noise=True)
# reverse pass
noise_hat = self.model(output_noisy, t)
# apply loss
return self.loss_fn(noise, noise_hat)
class DenoisingDiffusionConditionalProcess(nn.Module):
def __init__(self,
generated_channels=4,
condition_channels=12,
loss_fn=F.mse_loss,
schedule='linear',
num_timesteps=1000,
sampler=None
):
super().__init__()
# Basic Params
self.generated_channels=generated_channels
self.condition_channels=condition_channels
self.num_timesteps=num_timesteps
self.loss_fn=loss_fn
# Forward Process
self.forward_process=GaussianForwardProcess(num_timesteps=self.num_timesteps,
schedule=schedule)
# Neural Network Backbone
self.model=UnetConvNextBlock(dim=128,
dim_mults = (1,2,4,8),
channels=self.generated_channels + condition_channels,
out_dim=self.generated_channels,
with_time_emb=True)
# defaults to a DDPM sampler if None is provided
self.sampler=DDPM_Sampler(num_timesteps=self.num_timesteps) if sampler is None else sampler
@torch.no_grad()
def forward(self,
condition,
sampler=None,
verbose=False
):
"""
forward() function triggers a complete inference cycle
A custom sampler can be provided as an argument!
"""
# read dimensions
b,c,h,w=condition.shape
device=next(self.model.parameters()).device
condition=condition.to(device)
# select sampler
if sampler is None:
sampler=self.sampler
else:
sampler.to(device)
# time steps list
num_timesteps=sampler.num_timesteps
it=reversed(range(0, num_timesteps))
x_t = torch.randn([b, self.generated_channels, h, w],device=device)
for i in tqdm(it, desc='diffusion sampling', total=num_timesteps) if verbose else it:
t = torch.full((b,), i, device=device, dtype=torch.long)
model_input=torch.cat([x_t,condition],1).to(device)
z_t=self.model(model_input,t) # prediction of noise
x_t=self.sampler(x_t,t,z_t) # prediction of next state
return x_t
def p_loss(self,output,condition):
"""
Assumes output and input are in [-1,+1] range
"""
b,c,h,w=output.shape
device=output.device
# loss for training
# input is the optional condition
t = torch.randint(0, self.forward_process.num_timesteps, (b,), device=device).long()
output_noisy, noise=self.forward_process(output,t,return_noise=True)
# reverse pass
model_input=torch.cat([output_noisy,condition],1).to(device)
noise_hat = self.model(model_input, t)
# apply loss
return self.loss_fn(noise, noise_hat) |