celeb和mnist上最简单的训练结果

CIntellifusion · May 9, 2024 · ce9a888 · ce9a888
1 parent f3a62a6
commit ce9a888
Show file tree

Hide file tree

Showing 8 changed files with 51 additions and 14 deletions.
diff --git a/README.md b/README.md
@@ -3,33 +3,62 @@ A simple diffusion framework for comprehensive understanding.
 
 Carry out a series training and record the training results. 
 
-Compare the result with papers. 
+Compare the result with papers.
+
+
+
+> we don’t understand what we can’t create. 
 
 # Task 1 Unconditional Diffusion
 
-- [ ] Implement a DDPM and DDIM that can generate Celeb64Image 
-- [ ] Unet Scalable Unet 
+- [x] Implement a DDPM and DDIM
 
-# Task 2 Conditional Diffusion
+- [x] Training On Mnist 
+
+![mnist_200epoch](README/mnist_200epoch.gif)
+
+- [x]  that can generate Celeb64Image  20240509 
+
+![celeb_200epoch_32pix](UnconditionalDiffusion/training_results/celeb_200epoch_32pix.gif)
+
+- [ ] Scalable Unet to 128 - 256 - 512 
+
+
+
+for further details please refer to `./UnoncditionalDiffusion/experiments.pptx`  and `./UnoncditionalDiffusion/experiments.md` 
 
-- [ ] pretrained text model for condition
-- [ ] different condition type: vanilla , token , cross attention etc. 
-- [ ] multiple condition : zero-conv(controlnet)
 
 
+## Engineer Issues
 
-# Task3 Latent Diffusion
+在VAE和UnconditionalDiffusion的代码中，`DataModule`和`Trainer`类重复度是非常高的。 在开发的过程中，用工程文件的形式可以保证代码的一致性。 但是会增加代码的复杂度。 
+
+
+
+第二点: 可以利用`importlib` +`omega_conf` 来简化类初始化。 参考VideoCrafter的实现。 
+
+
+
+# Task 2 Latent Diffusion
 
 - [ ] vae for compression 
+  - [x] train on mnist 
+
 - [ ] diffusion on other latent space: text , audio , mesh , etc.
 
-# Task4 Diffusion Transformer
+# Task 3Conditional Diffusion
+
+- [ ] pretrained text model for condition
+- [ ] different condition type: vanilla , token , cross attention etc. 
+- [ ] multiple condition : zero-conv(controlnet)
+
+# Task 4 Diffusion Transformer
 
 - [ ] replace Unet with a transformer 
 
 
 
-# Task5 Video Diffusion
+# Task 5 Video Diffusion
 
 - [ ] insert temporal layer into diffusion 
 

diff --git a/README/celeb_200epoch_32pix.gif b/README/celeb_200epoch_32pix.gif
diff --git a/README/mnist_200epoch.gif b/README/mnist_200epoch.gif
diff --git a/UnconditionalDiffusion/experiment.pptx b/UnconditionalDiffusion/experiment.pptx
diff --git a/UnconditionalDiffusion/experiments.md b/UnconditionalDiffusion/experiments.md
@@ -1,9 +1,5 @@
 # Unconditional Diffusion
 
-## 1 Dataset
-
-Celeb64 example: 
-
 # Debug
 
 
@@ -30,6 +26,12 @@ lambda: x: (x-0.5)/2
 
 
 
+现在只用ToTensor做预处理就足够 
+
+
+
+
+
 ## 学习率调整
 
 应该先用不调整学习率方法进行训练
@@ -40,5 +42,11 @@ lambda: x: (x-0.5)/2
 
 
 
+损失函数的绝对值和采样方法有很大的关系。 扩散模型中，模型是在拟合一个条件概率分布。 如果前向的概率分布错了那么拟合的损失也会受影响。 
+
+
+
+损失函数的收敛速度也和采样方法有很大的关系。 
+
 
 
diff --git a/UnconditionalDiffusion/training_results/celeb_1200peoch.gif b/UnconditionalDiffusion/training_results/celeb_1200peoch.gif
diff --git a/UnconditionalDiffusion/training_results/celeb_200epoch_32pix.gif b/UnconditionalDiffusion/training_results/celeb_200epoch_32pix.gif
diff --git a/UnconditionalDiffusion/training_results/mnist_200epoch.gif b/UnconditionalDiffusion/training_results/mnist_200epoch.gif