dit.py

'''
-----------------------------------------------------------------------------
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')