Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- pytorch
|
| 5 |
+
- diffusers
|
| 6 |
+
- class-conditional-image-generation
|
| 7 |
+
- diffusion-models-class
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
# This model is a diffusion model for generate number.
|
| 13 |
+
|
| 14 |
+
## Usage
|
| 15 |
+
|
| 16 |
+
```python
|
| 17 |
+
# 定义模型
|
| 18 |
+
class ClassConditionalUnet(nn.Module):
|
| 19 |
+
def __init__(self, num_classes=10, class_emb_size=4):
|
| 20 |
+
super().__init__()
|
| 21 |
+
# 将数字所属的类别映射到一个长度为class_emb_size的特征向量
|
| 22 |
+
self.class_emb = nn.Embedding(num_classes, class_emb_size)
|
| 23 |
+
# self.model就是一个不带条件的unet模型,在这里给他添加额外的输入通道,用于接收条件信息
|
| 24 |
+
self.model = UNet2DModel(
|
| 25 |
+
sample_size=28, #生成的图像是28*28
|
| 26 |
+
in_channels=1 + class_emb_size, #加入额外的输入通道
|
| 27 |
+
out_channels=1, # 输入单通道黑白数字图
|
| 28 |
+
layers_per_block=2, # 设置一个unet模块有多少个残差连接层
|
| 29 |
+
block_out_channels=(32, 64, 64),
|
| 30 |
+
down_block_types=(
|
| 31 |
+
"DownBlock2D", #普通的ResNet下采样模块
|
| 32 |
+
"AttnDownBlock2D", #含有spatial self-attention的下采样和模块
|
| 33 |
+
"AttnDownBlock2D",
|
| 34 |
+
|
| 35 |
+
),
|
| 36 |
+
up_block_types=(
|
| 37 |
+
"AttnUpBlock2D", #含有spatial self-attention的ResNet上采样模块
|
| 38 |
+
"AttnUpBlock2D",
|
| 39 |
+
"UpBlock2D",
|
| 40 |
+
),
|
| 41 |
+
)
|
| 42 |
+
|
| 43 |
+
def forward(self, x, t, class_labels):
|
| 44 |
+
bs, ch, w, h = x.shape
|
| 45 |
+
class_cond = self.class_emb(class_labels) # 将类别映射为向量形式
|
| 46 |
+
class_cond = class_cond.view(bs, class_cond.shape[1], 1, 1).expand(bs, class_cond.shape[1], w,
|
| 47 |
+
h) # 拓展张量形状
|
| 48 |
+
net_input = torch.cat((class_cond, x), dim=1)
|
| 49 |
+
return self.model(net_input, t).sample
|
| 50 |
+
ckpt = torch.load("class_cond_unet.pth", map_location="cpu")
|
| 51 |
+
|
| 52 |
+
model = ClassConditionalUnet(
|
| 53 |
+
num_classes=ckpt["num_classes"],
|
| 54 |
+
class_emb_size=ckpt["class_emb_size"]
|
| 55 |
+
)
|
| 56 |
+
|
| 57 |
+
model.load_state_dict(ckpt["model_state_dict"])
|
| 58 |
+
model.eval()
|
| 59 |
+
noise_scheduler = DDPMScheduler.from_pretrained("Dirry525/class_num_generator")
|
| 60 |
+
```
|