Skip to content

week16_복습과제_김효민 #90

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Kim-Hyomin wants to merge 1 commit into
base: Week_16
Choose a base branch
from
Open

week16_복습과제_김효민 #90

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
187 changes: 187 additions & 0 deletions Week16_복습과제_김효민.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,187 @@
{
"nbformat": 4,
"nbformat_minor": 0,
"metadata": {
"colab": {
"provenance": []
},
"kernelspec": {
"name": "python3",
"display_name": "Python 3"
},
"language_info": {
"name": "python"
}
},
"cells": [
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"id": "lOtuNF8jgL8o"
},
"outputs": [],
"source": [
"def forward(self, x, temb):\n",
" \"\"\"\n",
" x: latent feature map (B, C, H, W)\n",
" - VAE encoder가 만든 latent에서 denoising을 수행하는 feature\n",
" temb: timestep embedding (B, temb_dim)\n",
" - diffusion step t → sinusoidal embedding → dense → nonlinearity → dense\n",
" - U-Net 전체에서 noise level/time 정보를 전달하기 위한 vector\n",
"\n",
" 이 블록은 U-Net의 한 레이어에서 residual path 연산을 수행합니다.\n",
" LDM U-Net에서는 이 블록이 down/up sampling, mid layer 등에 반복적으로 쓰입니다. :contentReference[oaicite:2]{index=2}\n",
" \"\"\"\n",
"\n",
" # ================= Residual Branch 시작 =================\n",
" h = x # 입력 복사 (residual 합을 위한 기본 분기)\n",
"\n",
" # ----- 1. 정규화(norm) → 활성화(nonlinearity) → conv1 -----\n",
" h = self.norm1(h)\n",
" h = nonlinearity(h) # 보통 SiLU(x) = x * sigmoid(x)\n",
" h = self.conv1(h) # latent 공간에서의 convolution\n",
"\n",
" # ----- 2. Time embedding 주입 (noise scale conditioning) -----\n",
" if temb is not None:\n",
" # temb: (B, temb_dim) → nonlinearity → linear projection → (B, C)\n",
" # → [:, :, None, None]로 reshape하여 spatial 전체에 broadcast\n",
" # Latent space 전체를 noise level 조건으로 강화\n",
" h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]\n",
"\n",
" # ----- 3. 정규화(norm) → 활성화 → dropout → conv2 -----\n",
" h = self.norm2(h)\n",
" h = nonlinearity(h)\n",
" h = self.dropout(h) # regularization\n",
" h = self.conv2(h)\n",
"\n",
" # ================= Shortcut (residual) 처리 =================\n",
" # in_channels != out_channels 인 경우 채널 mismatch 보정\n",
" # U-Net down/up block에서 채널 수가 바뀔 때 필요\n",
" if self.in_channels != self.out_channels:\n",
" if self.use_conv_shortcut:\n",
" # conv_shortcut: 일반 3×3 conv\n",
" x = self.conv_shortcut(x)\n",
" else:\n",
" # nin_shortcut: 1×1 conv(Network-in-Network)\n",
" x = self.nin_shortcut(x)\n",
"\n",
" # ================= 최종 residual 합 =================\n",
" return x + h\n"
]
},
{
"cell_type": "code",
"source": [
"def forward(self, x, t=None, context=None):\n",
" \"\"\"\n",
" x : noisy latent feature map (B, C, H, W)\n",
" - VAE encoder 결과 z에 noise가 섞인 상태\n",
" t : diffusion timestep (B,)\n",
" - noise level을 나타내는 정수 timestep\n",
" context : conditioning 정보 (optional)\n",
" - LDM에서는 보통 text embedding (e.g. CLIP text encoder output)\n",
" - spatially aligned된 경우 channel 방향으로 concat\n",
" \"\"\"\n",
"\n",
" # =========================================================\n",
" # 1. Context conditioning (early fusion)\n",
" # =========================================================\n",
" if context is not None:\n",
" # context가 latent와 spatially aligned되어 있다고 가정\n",
" # channel 차원으로 concat하여 조건 정보를 입력에 직접 결합\n",
" x = torch.cat((x, context), dim=1)\n",
"\n",
" # =========================================================\n",
" # 2. Timestep embedding 생성\n",
" # =========================================================\n",
" if self.use_timestep:\n",
" # diffusion model에서는 timestep conditioning이 필수\n",
" assert t is not None\n",
"\n",
" # t → sinusoidal embedding (B, ch)\n",
" temb = get_timestep_embedding(t, self.ch)\n",
"\n",
" # MLP 1층\n",
" temb = self.temb.dense[0](temb)\n",
" temb = nonlinearity(temb)\n",
"\n",
" # MLP 2층\n",
" temb = self.temb.dense[1](temb)\n",
" else:\n",
" temb = None\n",
"\n",
" # =========================================================\n",
" # 3. Downsampling path (Encoder)\n",
" # =========================================================\n",
" # skip connection에 사용할 feature들을 저장\n",
" hs = [self.conv_in(x)] # 입력 latent를 feature space로 사상\n",
"\n",
" for i_level in range(self.num_resolutions):\n",
" for i_block in range(self.num_res_blocks):\n",
" # ResNet block (앞에서 설명한 그 block)\n",
" h = self.down[i_level].block[i_block](hs[-1], temb)\n",
"\n",
" # 해당 resolution에서 attention을 쓰는 경우\n",
" if len(self.down[i_level].attn) > 0:\n",
" h = self.down[i_level].attn[i_block](h)\n",
"\n",
" # skip connection 저장\n",
" hs.append(h)\n",
"\n",
" # 마지막 resolution이 아니면 spatial downsampling\n",
" if i_level != self.num_resolutions - 1:\n",
" hs.append(self.down[i_level].downsample(hs[-1]))\n",
"\n",
" # =========================================================\n",
" # 4. Middle block (Bottleneck)\n",
" # =========================================================\n",
" h = hs[-1]\n",
"\n",
" # ResNet block\n",
" h = self.mid.block_1(h, temb)\n",
"\n",
" # Self-Attention (global context 통합)\n",
" h = self.mid.attn_1(h)\n",
"\n",
" # ResNet block\n",
" h = self.mid.block_2(h, temb)\n",
"\n",
" # =========================================================\n",
" # 5. Upsampling path (Decoder)\n",
" # =========================================================\n",
" for i_level in reversed(range(self.num_resolutions)):\n",
" for i_block in range(self.num_res_blocks + 1):\n",
" # skip connection과 concat (U-Net 핵심)\n",
" h = self.up[i_level].block[i_block](\n",
" torch.cat([h, hs.pop()], dim=1),\n",
" temb\n",
" )\n",
"\n",
" # attention이 있는 경우 적용\n",
" if len(self.up[i_level].attn) > 0:\n",
" h = self.up[i_level].attn[i_block](h)\n",
"\n",
" # 마지막 level이 아니면 upsampling\n",
" if i_level != 0:\n",
" h = self.up[i_level].upsample(h)\n",
"\n",
" # =========================================================\n",
" # 6. Output projection\n",
" # =========================================================\n",
" h = self.norm_out(h)\n",
" h = nonlinearity(h)\n",
"\n",
" # 최종 출력: noise prediction (ε̂ or v̂)\n",
" h = self.conv_out(h)\n",
"\n",
" return h\n"
],
"metadata": {
"id": "1mk1TJ9xgbRV"
},
"execution_count": null,
"outputs": []
}
]
}