File size: 8,339 Bytes
f86793b |
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 |
'''
-----------------------------------------------------------------------------
Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
NVIDIA CORPORATION and its licensors retain all intellectual property
and proprietary rights in and to this software, related documentation
and any modifications thereto. Any use, reproduction, disclosure or
distribution of this software and related documentation without an express
license agreement from NVIDIA CORPORATION is strictly prohibited.
-----------------------------------------------------------------------------
'''
import sys
sys.path.append('.')
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint
import numpy as np
from DIT.attention import SelfAttention, SelfAttentionqknorm, CrossAttention, CrossAttentionqnorm
class GEGLU(nn.Module):
def forward(self, x):
x, gates = x.chunk(2, dim = -1)
return x * F.gelu(gates)
class FeedForward(nn.Module):
def __init__(self, dim, mult=4):
super().__init__()
self.net = nn.Sequential(
nn.Linear(dim, dim * mult * 2),
GEGLU(),
nn.Linear(dim * mult, dim)
)
def forward(self, x):
return self.net(x)
class Timesteps(nn.Module):
def __init__(self, num_channels=256, flip_sin_to_cos=False, downscale_freq_shift=0, scale=1, max_period=10000):
super().__init__()
self.num_channels = num_channels
self.flip_sin_to_cos = flip_sin_to_cos
self.downscale_freq_shift = downscale_freq_shift
self.scale = scale
self.max_period = max_period
def forward(self, timesteps):
half_dim = self.num_channels // 2
exponent = -math.log(self.max_period) * torch.arange(start=0, end=half_dim, dtype=torch.float32, device=timesteps.device)
exponent = exponent / (half_dim - self.downscale_freq_shift)
emb = torch.exp(exponent)
emb = timesteps[:, None].float() * emb[None, :]
# scale embeddings
emb = self.scale * emb
# concat sine and cosine embeddings
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
# flip sine and cosine embeddings
if self.flip_sin_to_cos:
emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
# zero pad
if self.num_channels % 2 == 1:
emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
return emb
class TimestepEmbedding(nn.Module):
def __init__(
self,
in_channels: int,
time_embed_dim: int,
sample_proj_bias=True,
):
super().__init__()
self.linear_1 = nn.Linear(in_channels, time_embed_dim, sample_proj_bias)
self.act = F.silu
self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim, sample_proj_bias)
def forward(self, sample):
sample = self.linear_1(sample)
sample = self.act(sample)
sample = self.linear_2(sample)
return sample
# PixArtAlpha-style with removed cross-attention
class DiTLayer(nn.Module):
def __init__(self, dim, num_heads, gradient_checkpointing=True):
super().__init__()
self.dim = dim
self.num_heads = num_heads
self.gradient_checkpointing = gradient_checkpointing
self.norm1 = nn.LayerNorm(dim, eps=1e-6, elementwise_affine=False)
self.attn1 = SelfAttentionqknorm(dim, num_heads)
self.attn2 = CrossAttentionqnorm(dim, num_heads)
self.norm2 = nn.LayerNorm(dim, eps=1e-6, elementwise_affine=False)
self.norm3 = nn.LayerNorm(dim, eps=1e-6, elementwise_affine=False)
self.norm4 = nn.LayerNorm(dim, eps=1e-6, elementwise_affine=False)
self.ff = FeedForward(dim)
self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim ** 0.5)
def forward(self, x, t_adaln,mask=None,condition=None):
if self.training and self.gradient_checkpointing:
return checkpoint(self._forward, x, t_adaln,mask,condition, use_reentrant=False)
else:
return self._forward(x, t_adaln,mask,condition)
def _forward(self, x, t_adaln,mask=None,condition=None):
# x: [B, N, C], hidden states
# t_adaln: [B, 6, C], timestep embedding of adaln
# return: [B, N, C], updated hidden states
B, N, C = x.shape
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None] + t_adaln).chunk(6, dim=1)
# Self-attention block
# x1 = self.norm1(x)
x1 = self.norm1(x) * (1 + scale_msa) + shift_msa
x = x + gate_msa * self.norm2(self.attn1(x1,mask))
x = x + self.attn2(x, condition,mask)
# x = self.norm3(x)
x2 = self.norm3(x) * (1 + scale_mlp) + shift_mlp
x = x + gate_mlp * self.norm4(self.ff(x2))
return x
class DiT(nn.Module):
def __init__(self, hidden_dim=1024, num_heads=16, latent_size=2048, latent_dim=64, num_layers=24, gradient_checkpointing=True,length=50):
super().__init__()
self.length=length
self.latent_dim = latent_dim
# project in
self.proj_in = nn.Linear(latent_dim, hidden_dim)
self.hidden_dim=hidden_dim
# positional encoding (just use a learnable positional encoding)
# self.pos_embed = nn.Parameter(torch.randn(1, self.length, hidden_dim) / hidden_dim ** 0.5)
# timestep encoding
self.timestep_embed = Timesteps(num_channels=256)
self.timestep_proj = TimestepEmbedding(256, hidden_dim)
self.adaln_linear = nn.Linear(hidden_dim, hidden_dim * 6, bias=True)
self.num_faces = nn.Embedding(1000, hidden_dim)
self.num_linear = nn.Linear(hidden_dim, hidden_dim * 6, bias=True)
# transformer layers
self.layers = nn.ModuleList([DiTLayer(hidden_dim, num_heads, gradient_checkpointing) for _ in range(num_layers)])
# project out
self.norm_out = nn.LayerNorm(hidden_dim, eps=1e-6, elementwise_affine=False)
self.scale_shift_table = nn.Parameter(torch.randn(2, hidden_dim) / hidden_dim ** 0.5)
self.proj_out = nn.Linear(hidden_dim, latent_dim)
def forward(self, x, times,mask=None,condition=None,num_faces=None):
# x: [B, N, C], hidden states
# t: [B,], timestep
# return: [B, N, C], updated hidden states
# mask=None
if num_faces is None:
num_faces = mask.sum(-1)
B, N, C = x.shape
# project in
x = self.proj_in(x)
# positional encoding
# x = x + self.pos_embed
# timestep encoding
t_emb = self.timestep_embed(times)
t_emb = self.timestep_proj(t_emb) # [B, C]
t_adaln = self.adaln_linear(F.silu(t_emb)).view(B, 6, -1) # [B, 6, C]
t_adaln=t_adaln+self.num_linear(self.num_faces(num_faces)).view(B, 6, -1)
# transformer layers
for layer in self.layers:
x = layer(x, t_adaln,mask,condition)
# project out
shift, scale = (self.scale_shift_table[None] + t_emb[:, None]).chunk(2, dim=1)
x = self.norm_out(x)
x = x * (1 + scale) + shift
x = self.proj_out(x)
return x
if __name__ == '__main__':
# import kiui
# from kiui.nn.utils import count_parameters
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = DiT(gradient_checkpointing=True).to(device)
print(model)
# total, trainable = count_parameters(model)
# print(f'[INFO] param total: {total/1024**2:.2f}M, trainable: {trainable/1024**2:.2f}M')
# test forward
x = torch.randn(16, 2048, 64, device=device, dtype=torch.float16)
times = torch.randint(0, 1000, (16,), device=device)
with torch.autocast(device_type='cuda', dtype=torch.float16):
y = model(x, times)
# kiui.lo(y)
mem_free, mem_total = torch.cuda.mem_get_info()
print(f'[INFO] mem forward: {(mem_total-mem_free)/1024**3:.2f}/{mem_total/1024**3:.2f}G')
# test backward
loss = y.mean()
loss.backward()
mem_free, mem_total = torch.cuda.mem_get_info()
print(f'[INFO] mem backward: {(mem_total-mem_free)/1024**3:.2f}/{mem_total/1024**3:.2f}G') |