Upload folder using huggingface_hub
Browse files- DiT-MoE-B-8E2A/README.md +73 -0
- DiT-MoE-B-8E2A/model_index.json +1021 -0
- DiT-MoE-B-8E2A/pipeline.py +440 -0
- DiT-MoE-B-8E2A/scheduler/scheduler_config.json +13 -0
- DiT-MoE-B-8E2A/transformer/config.json +16 -0
- DiT-MoE-B-8E2A/transformer/diffusion_pytorch_model.safetensors +3 -0
- DiT-MoE-B-8E2A/transformer/transformer_dit_moe.py +511 -0
- DiT-MoE-B-8E2A/vae/config.json +29 -0
- DiT-MoE-B-8E2A/vae/diffusion_pytorch_model.safetensors +3 -0
- DiT-MoE-S-8E2A/README.md +2 -1
- DiT-MoE-S-8E2A/pipeline.py +1 -1
- README.md +4 -2
DiT-MoE-B-8E2A/README.md
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: diffusers
|
| 4 |
+
pipeline_tag: unconditional-image-generation
|
| 5 |
+
tags:
|
| 6 |
+
- diffusers
|
| 7 |
+
- dit-moe
|
| 8 |
+
- image-generation
|
| 9 |
+
- class-conditional
|
| 10 |
+
- imagenet
|
| 11 |
+
inference: true
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# DiT-MoE-B-8E2A
|
| 15 |
+
|
| 16 |
+
Self-contained Diffusers checkpoint for **DiT-B/2** with MoE routing, converted from [`feizhengcong/DiT-MoE`](https://huggingface.co/feizhengcong/DiT-MoE).
|
| 17 |
+
|
| 18 |
+
Each subfolder is a self-contained Diffusers model repo with:
|
| 19 |
+
|
| 20 |
+
- `model_index.json` (includes ImageNet `id2label`)
|
| 21 |
+
- `pipeline.py` (custom `DiTMoEPipeline`)
|
| 22 |
+
- `transformer/transformer_dit_moe.py` and weights
|
| 23 |
+
- `vae/diffusion_pytorch_model.safetensors`
|
| 24 |
+
- `scheduler/scheduler_config.json`
|
| 25 |
+
|
| 26 |
+
## ImageNet class labels
|
| 27 |
+
|
| 28 |
+
Each variant keeps an English `id2label` map in `model_index.json` (DiT-style).
|
| 29 |
+
|
| 30 |
+
- `pipe.id2label[207]` — `"golden retriever"`
|
| 31 |
+
- `pipe.get_label_ids("golden retriever")` — `[207]`
|
| 32 |
+
- `pipe(class_labels="golden retriever", ...)` — string labels resolved automatically
|
| 33 |
+
|
| 34 |
+
## Recommended inference (256×256)
|
| 35 |
+
|
| 36 |
+
| Setting | Value |
|
| 37 |
+
| --- | --- |
|
| 38 |
+
| Resolution | 256×256 |
|
| 39 |
+
| Sampler | DDIM |
|
| 40 |
+
| Steps | 50 |
|
| 41 |
+
| CFG scale | 4.0 |
|
| 42 |
+
| Dtype | `bfloat16` (recommended on Ampere+) |
|
| 43 |
+
| VAE | `stabilityai/sd-vae-ft-mse` (bundled under `vae/`) |
|
| 44 |
+
|
| 45 |
+
```python
|
| 46 |
+
from pathlib import Path
|
| 47 |
+
import torch
|
| 48 |
+
from diffusers import DiffusionPipeline
|
| 49 |
+
|
| 50 |
+
model_dir = Path("./DiT-MoE-B-8E2A").resolve()
|
| 51 |
+
pipe = DiffusionPipeline.from_pretrained(
|
| 52 |
+
str(model_dir),
|
| 53 |
+
local_files_only=True,
|
| 54 |
+
custom_pipeline=str(model_dir / "pipeline.py"),
|
| 55 |
+
trust_remote_code=True,
|
| 56 |
+
torch_dtype=torch.bfloat16,
|
| 57 |
+
)
|
| 58 |
+
pipe.to("cuda")
|
| 59 |
+
|
| 60 |
+
print(pipe.id2label[207])
|
| 61 |
+
print(pipe.get_label_ids("golden retriever"))
|
| 62 |
+
|
| 63 |
+
generator = torch.Generator(device="cuda").manual_seed(42)
|
| 64 |
+
image = pipe(
|
| 65 |
+
class_labels="golden retriever",
|
| 66 |
+
height=256,
|
| 67 |
+
width=256,
|
| 68 |
+
num_inference_steps=50,
|
| 69 |
+
guidance_scale=4.0,
|
| 70 |
+
generator=generator,
|
| 71 |
+
).images[0]
|
| 72 |
+
image.save("demo.png")
|
| 73 |
+
```
|
DiT-MoE-B-8E2A/model_index.json
ADDED
|
@@ -0,0 +1,1021 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_class_name": [
|
| 3 |
+
"pipeline",
|
| 4 |
+
"DiTMoEPipeline"
|
| 5 |
+
],
|
| 6 |
+
"_diffusers_version": "0.36.0",
|
| 7 |
+
"scheduler": [
|
| 8 |
+
"diffusers",
|
| 9 |
+
"DDIMScheduler"
|
| 10 |
+
],
|
| 11 |
+
"transformer": [
|
| 12 |
+
"transformer_dit_moe",
|
| 13 |
+
"DiTMoETransformer2DModel"
|
| 14 |
+
],
|
| 15 |
+
"vae": [
|
| 16 |
+
"diffusers",
|
| 17 |
+
"AutoencoderKL"
|
| 18 |
+
],
|
| 19 |
+
"id2label": {
|
| 20 |
+
"0": "tench, Tinca tinca",
|
| 21 |
+
"1": "goldfish, Carassius auratus",
|
| 22 |
+
"2": "great white shark, white shark, man-eater, man-eating shark, Carcharodon carcharias",
|
| 23 |
+
"3": "tiger shark, Galeocerdo cuvieri",
|
| 24 |
+
"4": "hammerhead, hammerhead shark",
|
| 25 |
+
"5": "electric ray, crampfish, numbfish, torpedo",
|
| 26 |
+
"6": "stingray",
|
| 27 |
+
"7": "cock",
|
| 28 |
+
"8": "hen",
|
| 29 |
+
"9": "ostrich, Struthio camelus",
|
| 30 |
+
"10": "brambling, Fringilla montifringilla",
|
| 31 |
+
"11": "goldfinch, Carduelis carduelis",
|
| 32 |
+
"12": "house finch, linnet, Carpodacus mexicanus",
|
| 33 |
+
"13": "junco, snowbird",
|
| 34 |
+
"14": "indigo bunting, indigo finch, indigo bird, Passerina cyanea",
|
| 35 |
+
"15": "robin, American robin, Turdus migratorius",
|
| 36 |
+
"16": "bulbul",
|
| 37 |
+
"17": "jay",
|
| 38 |
+
"18": "magpie",
|
| 39 |
+
"19": "chickadee",
|
| 40 |
+
"20": "water ouzel, dipper",
|
| 41 |
+
"21": "kite",
|
| 42 |
+
"22": "bald eagle, American eagle, Haliaeetus leucocephalus",
|
| 43 |
+
"23": "vulture",
|
| 44 |
+
"24": "great grey owl, great gray owl, Strix nebulosa",
|
| 45 |
+
"25": "European fire salamander, Salamandra salamandra",
|
| 46 |
+
"26": "common newt, Triturus vulgaris",
|
| 47 |
+
"27": "eft",
|
| 48 |
+
"28": "spotted salamander, Ambystoma maculatum",
|
| 49 |
+
"29": "axolotl, mud puppy, Ambystoma mexicanum",
|
| 50 |
+
"30": "bullfrog, Rana catesbeiana",
|
| 51 |
+
"31": "tree frog, tree-frog",
|
| 52 |
+
"32": "tailed frog, bell toad, ribbed toad, tailed toad, Ascaphus trui",
|
| 53 |
+
"33": "loggerhead, loggerhead turtle, Caretta caretta",
|
| 54 |
+
"34": "leatherback turtle, leatherback, leathery turtle, Dermochelys coriacea",
|
| 55 |
+
"35": "mud turtle",
|
| 56 |
+
"36": "terrapin",
|
| 57 |
+
"37": "box turtle, box tortoise",
|
| 58 |
+
"38": "banded gecko",
|
| 59 |
+
"39": "common iguana, iguana, Iguana iguana",
|
| 60 |
+
"40": "American chameleon, anole, Anolis carolinensis",
|
| 61 |
+
"41": "whiptail, whiptail lizard",
|
| 62 |
+
"42": "agama",
|
| 63 |
+
"43": "frilled lizard, Chlamydosaurus kingi",
|
| 64 |
+
"44": "alligator lizard",
|
| 65 |
+
"45": "Gila monster, Heloderma suspectum",
|
| 66 |
+
"46": "green lizard, Lacerta viridis",
|
| 67 |
+
"47": "African chameleon, Chamaeleo chamaeleon",
|
| 68 |
+
"48": "Komodo dragon, Komodo lizard, dragon lizard, giant lizard, Varanus komodoensis",
|
| 69 |
+
"49": "African crocodile, Nile crocodile, Crocodylus niloticus",
|
| 70 |
+
"50": "American alligator, Alligator mississipiensis",
|
| 71 |
+
"51": "triceratops",
|
| 72 |
+
"52": "thunder snake, worm snake, Carphophis amoenus",
|
| 73 |
+
"53": "ringneck snake, ring-necked snake, ring snake",
|
| 74 |
+
"54": "hognose snake, puff adder, sand viper",
|
| 75 |
+
"55": "green snake, grass snake",
|
| 76 |
+
"56": "king snake, kingsnake",
|
| 77 |
+
"57": "garter snake, grass snake",
|
| 78 |
+
"58": "water snake",
|
| 79 |
+
"59": "vine snake",
|
| 80 |
+
"60": "night snake, Hypsiglena torquata",
|
| 81 |
+
"61": "boa constrictor, Constrictor constrictor",
|
| 82 |
+
"62": "rock python, rock snake, Python sebae",
|
| 83 |
+
"63": "Indian cobra, Naja naja",
|
| 84 |
+
"64": "green mamba",
|
| 85 |
+
"65": "sea snake",
|
| 86 |
+
"66": "horned viper, cerastes, sand viper, horned asp, Cerastes cornutus",
|
| 87 |
+
"67": "diamondback, diamondback rattlesnake, Crotalus adamanteus",
|
| 88 |
+
"68": "sidewinder, horned rattlesnake, Crotalus cerastes",
|
| 89 |
+
"69": "trilobite",
|
| 90 |
+
"70": "harvestman, daddy longlegs, Phalangium opilio",
|
| 91 |
+
"71": "scorpion",
|
| 92 |
+
"72": "black and gold garden spider, Argiope aurantia",
|
| 93 |
+
"73": "barn spider, Araneus cavaticus",
|
| 94 |
+
"74": "garden spider, Aranea diademata",
|
| 95 |
+
"75": "black widow, Latrodectus mactans",
|
| 96 |
+
"76": "tarantula",
|
| 97 |
+
"77": "wolf spider, hunting spider",
|
| 98 |
+
"78": "tick",
|
| 99 |
+
"79": "centipede",
|
| 100 |
+
"80": "black grouse",
|
| 101 |
+
"81": "ptarmigan",
|
| 102 |
+
"82": "ruffed grouse, partridge, Bonasa umbellus",
|
| 103 |
+
"83": "prairie chicken, prairie grouse, prairie fowl",
|
| 104 |
+
"84": "peacock",
|
| 105 |
+
"85": "quail",
|
| 106 |
+
"86": "partridge",
|
| 107 |
+
"87": "African grey, African gray, Psittacus erithacus",
|
| 108 |
+
"88": "macaw",
|
| 109 |
+
"89": "sulphur-crested cockatoo, Kakatoe galerita, Cacatua galerita",
|
| 110 |
+
"90": "lorikeet",
|
| 111 |
+
"91": "coucal",
|
| 112 |
+
"92": "bee eater",
|
| 113 |
+
"93": "hornbill",
|
| 114 |
+
"94": "hummingbird",
|
| 115 |
+
"95": "jacamar",
|
| 116 |
+
"96": "toucan",
|
| 117 |
+
"97": "drake",
|
| 118 |
+
"98": "red-breasted merganser, Mergus serrator",
|
| 119 |
+
"99": "goose",
|
| 120 |
+
"100": "black swan, Cygnus atratus",
|
| 121 |
+
"101": "tusker",
|
| 122 |
+
"102": "echidna, spiny anteater, anteater",
|
| 123 |
+
"103": "platypus, duckbill, duckbilled platypus, duck-billed platypus, Ornithorhynchus anatinus",
|
| 124 |
+
"104": "wallaby, brush kangaroo",
|
| 125 |
+
"105": "koala, koala bear, kangaroo bear, native bear, Phascolarctos cinereus",
|
| 126 |
+
"106": "wombat",
|
| 127 |
+
"107": "jellyfish",
|
| 128 |
+
"108": "sea anemone, anemone",
|
| 129 |
+
"109": "brain coral",
|
| 130 |
+
"110": "flatworm, platyhelminth",
|
| 131 |
+
"111": "nematode, nematode worm, roundworm",
|
| 132 |
+
"112": "conch",
|
| 133 |
+
"113": "snail",
|
| 134 |
+
"114": "slug",
|
| 135 |
+
"115": "sea slug, nudibranch",
|
| 136 |
+
"116": "chiton, coat-of-mail shell, sea cradle, polyplacophore",
|
| 137 |
+
"117": "chambered nautilus, pearly nautilus, nautilus",
|
| 138 |
+
"118": "Dungeness crab, Cancer magister",
|
| 139 |
+
"119": "rock crab, Cancer irroratus",
|
| 140 |
+
"120": "fiddler crab",
|
| 141 |
+
"121": "king crab, Alaska crab, Alaskan king crab, Alaska king crab, Paralithodes camtschatica",
|
| 142 |
+
"122": "American lobster, Northern lobster, Maine lobster, Homarus americanus",
|
| 143 |
+
"123": "spiny lobster, langouste, rock lobster, crawfish, crayfish, sea crawfish",
|
| 144 |
+
"124": "crayfish, crawfish, crawdad, crawdaddy",
|
| 145 |
+
"125": "hermit crab",
|
| 146 |
+
"126": "isopod",
|
| 147 |
+
"127": "white stork, Ciconia ciconia",
|
| 148 |
+
"128": "black stork, Ciconia nigra",
|
| 149 |
+
"129": "spoonbill",
|
| 150 |
+
"130": "flamingo",
|
| 151 |
+
"131": "little blue heron, Egretta caerulea",
|
| 152 |
+
"132": "American egret, great white heron, Egretta albus",
|
| 153 |
+
"133": "bittern",
|
| 154 |
+
"134": "crane",
|
| 155 |
+
"135": "limpkin, Aramus pictus",
|
| 156 |
+
"136": "European gallinule, Porphyrio porphyrio",
|
| 157 |
+
"137": "American coot, marsh hen, mud hen, water hen, Fulica americana",
|
| 158 |
+
"138": "bustard",
|
| 159 |
+
"139": "ruddy turnstone, Arenaria interpres",
|
| 160 |
+
"140": "red-backed sandpiper, dunlin, Erolia alpina",
|
| 161 |
+
"141": "redshank, Tringa totanus",
|
| 162 |
+
"142": "dowitcher",
|
| 163 |
+
"143": "oystercatcher, oyster catcher",
|
| 164 |
+
"144": "pelican",
|
| 165 |
+
"145": "king penguin, Aptenodytes patagonica",
|
| 166 |
+
"146": "albatross, mollymawk",
|
| 167 |
+
"147": "grey whale, gray whale, devilfish, Eschrichtius gibbosus, Eschrichtius robustus",
|
| 168 |
+
"148": "killer whale, killer, orca, grampus, sea wolf, Orcinus orca",
|
| 169 |
+
"149": "dugong, Dugong dugon",
|
| 170 |
+
"150": "sea lion",
|
| 171 |
+
"151": "Chihuahua",
|
| 172 |
+
"152": "Japanese spaniel",
|
| 173 |
+
"153": "Maltese dog, Maltese terrier, Maltese",
|
| 174 |
+
"154": "Pekinese, Pekingese, Peke",
|
| 175 |
+
"155": "Shih-Tzu",
|
| 176 |
+
"156": "Blenheim spaniel",
|
| 177 |
+
"157": "papillon",
|
| 178 |
+
"158": "toy terrier",
|
| 179 |
+
"159": "Rhodesian ridgeback",
|
| 180 |
+
"160": "Afghan hound, Afghan",
|
| 181 |
+
"161": "basset, basset hound",
|
| 182 |
+
"162": "beagle",
|
| 183 |
+
"163": "bloodhound, sleuthhound",
|
| 184 |
+
"164": "bluetick",
|
| 185 |
+
"165": "black-and-tan coonhound",
|
| 186 |
+
"166": "Walker hound, Walker foxhound",
|
| 187 |
+
"167": "English foxhound",
|
| 188 |
+
"168": "redbone",
|
| 189 |
+
"169": "borzoi, Russian wolfhound",
|
| 190 |
+
"170": "Irish wolfhound",
|
| 191 |
+
"171": "Italian greyhound",
|
| 192 |
+
"172": "whippet",
|
| 193 |
+
"173": "Ibizan hound, Ibizan Podenco",
|
| 194 |
+
"174": "Norwegian elkhound, elkhound",
|
| 195 |
+
"175": "otterhound, otter hound",
|
| 196 |
+
"176": "Saluki, gazelle hound",
|
| 197 |
+
"177": "Scottish deerhound, deerhound",
|
| 198 |
+
"178": "Weimaraner",
|
| 199 |
+
"179": "Staffordshire bullterrier, Staffordshire bull terrier",
|
| 200 |
+
"180": "American Staffordshire terrier, Staffordshire terrier, American pit bull terrier, pit bull terrier",
|
| 201 |
+
"181": "Bedlington terrier",
|
| 202 |
+
"182": "Border terrier",
|
| 203 |
+
"183": "Kerry blue terrier",
|
| 204 |
+
"184": "Irish terrier",
|
| 205 |
+
"185": "Norfolk terrier",
|
| 206 |
+
"186": "Norwich terrier",
|
| 207 |
+
"187": "Yorkshire terrier",
|
| 208 |
+
"188": "wire-haired fox terrier",
|
| 209 |
+
"189": "Lakeland terrier",
|
| 210 |
+
"190": "Sealyham terrier, Sealyham",
|
| 211 |
+
"191": "Airedale, Airedale terrier",
|
| 212 |
+
"192": "cairn, cairn terrier",
|
| 213 |
+
"193": "Australian terrier",
|
| 214 |
+
"194": "Dandie Dinmont, Dandie Dinmont terrier",
|
| 215 |
+
"195": "Boston bull, Boston terrier",
|
| 216 |
+
"196": "miniature schnauzer",
|
| 217 |
+
"197": "giant schnauzer",
|
| 218 |
+
"198": "standard schnauzer",
|
| 219 |
+
"199": "Scotch terrier, Scottish terrier, Scottie",
|
| 220 |
+
"200": "Tibetan terrier, chrysanthemum dog",
|
| 221 |
+
"201": "silky terrier, Sydney silky",
|
| 222 |
+
"202": "soft-coated wheaten terrier",
|
| 223 |
+
"203": "West Highland white terrier",
|
| 224 |
+
"204": "Lhasa, Lhasa apso",
|
| 225 |
+
"205": "flat-coated retriever",
|
| 226 |
+
"206": "curly-coated retriever",
|
| 227 |
+
"207": "golden retriever",
|
| 228 |
+
"208": "Labrador retriever",
|
| 229 |
+
"209": "Chesapeake Bay retriever",
|
| 230 |
+
"210": "German short-haired pointer",
|
| 231 |
+
"211": "vizsla, Hungarian pointer",
|
| 232 |
+
"212": "English setter",
|
| 233 |
+
"213": "Irish setter, red setter",
|
| 234 |
+
"214": "Gordon setter",
|
| 235 |
+
"215": "Brittany spaniel",
|
| 236 |
+
"216": "clumber, clumber spaniel",
|
| 237 |
+
"217": "English springer, English springer spaniel",
|
| 238 |
+
"218": "Welsh springer spaniel",
|
| 239 |
+
"219": "cocker spaniel, English cocker spaniel, cocker",
|
| 240 |
+
"220": "Sussex spaniel",
|
| 241 |
+
"221": "Irish water spaniel",
|
| 242 |
+
"222": "kuvasz",
|
| 243 |
+
"223": "schipperke",
|
| 244 |
+
"224": "groenendael",
|
| 245 |
+
"225": "malinois",
|
| 246 |
+
"226": "briard",
|
| 247 |
+
"227": "kelpie",
|
| 248 |
+
"228": "komondor",
|
| 249 |
+
"229": "Old English sheepdog, bobtail",
|
| 250 |
+
"230": "Shetland sheepdog, Shetland sheep dog, Shetland",
|
| 251 |
+
"231": "collie",
|
| 252 |
+
"232": "Border collie",
|
| 253 |
+
"233": "Bouvier des Flandres, Bouviers des Flandres",
|
| 254 |
+
"234": "Rottweiler",
|
| 255 |
+
"235": "German shepherd, German shepherd dog, German police dog, alsatian",
|
| 256 |
+
"236": "Doberman, Doberman pinscher",
|
| 257 |
+
"237": "miniature pinscher",
|
| 258 |
+
"238": "Greater Swiss Mountain dog",
|
| 259 |
+
"239": "Bernese mountain dog",
|
| 260 |
+
"240": "Appenzeller",
|
| 261 |
+
"241": "EntleBucher",
|
| 262 |
+
"242": "boxer",
|
| 263 |
+
"243": "bull mastiff",
|
| 264 |
+
"244": "Tibetan mastiff",
|
| 265 |
+
"245": "French bulldog",
|
| 266 |
+
"246": "Great Dane",
|
| 267 |
+
"247": "Saint Bernard, St Bernard",
|
| 268 |
+
"248": "Eskimo dog, husky",
|
| 269 |
+
"249": "malamute, malemute, Alaskan malamute",
|
| 270 |
+
"250": "Siberian husky",
|
| 271 |
+
"251": "dalmatian, coach dog, carriage dog",
|
| 272 |
+
"252": "affenpinscher, monkey pinscher, monkey dog",
|
| 273 |
+
"253": "basenji",
|
| 274 |
+
"254": "pug, pug-dog",
|
| 275 |
+
"255": "Leonberg",
|
| 276 |
+
"256": "Newfoundland, Newfoundland dog",
|
| 277 |
+
"257": "Great Pyrenees",
|
| 278 |
+
"258": "Samoyed, Samoyede",
|
| 279 |
+
"259": "Pomeranian",
|
| 280 |
+
"260": "chow, chow chow",
|
| 281 |
+
"261": "keeshond",
|
| 282 |
+
"262": "Brabancon griffon",
|
| 283 |
+
"263": "Pembroke, Pembroke Welsh corgi",
|
| 284 |
+
"264": "Cardigan, Cardigan Welsh corgi",
|
| 285 |
+
"265": "toy poodle",
|
| 286 |
+
"266": "miniature poodle",
|
| 287 |
+
"267": "standard poodle",
|
| 288 |
+
"268": "Mexican hairless",
|
| 289 |
+
"269": "timber wolf, grey wolf, gray wolf, Canis lupus",
|
| 290 |
+
"270": "white wolf, Arctic wolf, Canis lupus tundrarum",
|
| 291 |
+
"271": "red wolf, maned wolf, Canis rufus, Canis niger",
|
| 292 |
+
"272": "coyote, prairie wolf, brush wolf, Canis latrans",
|
| 293 |
+
"273": "dingo, warrigal, warragal, Canis dingo",
|
| 294 |
+
"274": "dhole, Cuon alpinus",
|
| 295 |
+
"275": "African hunting dog, hyena dog, Cape hunting dog, Lycaon pictus",
|
| 296 |
+
"276": "hyena, hyaena",
|
| 297 |
+
"277": "red fox, Vulpes vulpes",
|
| 298 |
+
"278": "kit fox, Vulpes macrotis",
|
| 299 |
+
"279": "Arctic fox, white fox, Alopex lagopus",
|
| 300 |
+
"280": "grey fox, gray fox, Urocyon cinereoargenteus",
|
| 301 |
+
"281": "tabby, tabby cat",
|
| 302 |
+
"282": "tiger cat",
|
| 303 |
+
"283": "Persian cat",
|
| 304 |
+
"284": "Siamese cat, Siamese",
|
| 305 |
+
"285": "Egyptian cat",
|
| 306 |
+
"286": "cougar, puma, catamount, mountain lion, painter, panther, Felis concolor",
|
| 307 |
+
"287": "lynx, catamount",
|
| 308 |
+
"288": "leopard, Panthera pardus",
|
| 309 |
+
"289": "snow leopard, ounce, Panthera uncia",
|
| 310 |
+
"290": "jaguar, panther, Panthera onca, Felis onca",
|
| 311 |
+
"291": "lion, king of beasts, Panthera leo",
|
| 312 |
+
"292": "tiger, Panthera tigris",
|
| 313 |
+
"293": "cheetah, chetah, Acinonyx jubatus",
|
| 314 |
+
"294": "brown bear, bruin, Ursus arctos",
|
| 315 |
+
"295": "American black bear, black bear, Ursus americanus, Euarctos americanus",
|
| 316 |
+
"296": "ice bear, polar bear, Ursus Maritimus, Thalarctos maritimus",
|
| 317 |
+
"297": "sloth bear, Melursus ursinus, Ursus ursinus",
|
| 318 |
+
"298": "mongoose",
|
| 319 |
+
"299": "meerkat, mierkat",
|
| 320 |
+
"300": "tiger beetle",
|
| 321 |
+
"301": "ladybug, ladybeetle, lady beetle, ladybird, ladybird beetle",
|
| 322 |
+
"302": "ground beetle, carabid beetle",
|
| 323 |
+
"303": "long-horned beetle, longicorn, longicorn beetle",
|
| 324 |
+
"304": "leaf beetle, chrysomelid",
|
| 325 |
+
"305": "dung beetle",
|
| 326 |
+
"306": "rhinoceros beetle",
|
| 327 |
+
"307": "weevil",
|
| 328 |
+
"308": "fly",
|
| 329 |
+
"309": "bee",
|
| 330 |
+
"310": "ant, emmet, pismire",
|
| 331 |
+
"311": "grasshopper, hopper",
|
| 332 |
+
"312": "cricket",
|
| 333 |
+
"313": "walking stick, walkingstick, stick insect",
|
| 334 |
+
"314": "cockroach, roach",
|
| 335 |
+
"315": "mantis, mantid",
|
| 336 |
+
"316": "cicada, cicala",
|
| 337 |
+
"317": "leafhopper",
|
| 338 |
+
"318": "lacewing, lacewing fly",
|
| 339 |
+
"319": "dragonfly, darning needle, devil's darning needle, sewing needle, snake feeder, snake doctor, mosquito hawk, skeeter hawk",
|
| 340 |
+
"320": "damselfly",
|
| 341 |
+
"321": "admiral",
|
| 342 |
+
"322": "ringlet, ringlet butterfly",
|
| 343 |
+
"323": "monarch, monarch butterfly, milkweed butterfly, Danaus plexippus",
|
| 344 |
+
"324": "cabbage butterfly",
|
| 345 |
+
"325": "sulphur butterfly, sulfur butterfly",
|
| 346 |
+
"326": "lycaenid, lycaenid butterfly",
|
| 347 |
+
"327": "starfish, sea star",
|
| 348 |
+
"328": "sea urchin",
|
| 349 |
+
"329": "sea cucumber, holothurian",
|
| 350 |
+
"330": "wood rabbit, cottontail, cottontail rabbit",
|
| 351 |
+
"331": "hare",
|
| 352 |
+
"332": "Angora, Angora rabbit",
|
| 353 |
+
"333": "hamster",
|
| 354 |
+
"334": "porcupine, hedgehog",
|
| 355 |
+
"335": "fox squirrel, eastern fox squirrel, Sciurus niger",
|
| 356 |
+
"336": "marmot",
|
| 357 |
+
"337": "beaver",
|
| 358 |
+
"338": "guinea pig, Cavia cobaya",
|
| 359 |
+
"339": "sorrel",
|
| 360 |
+
"340": "zebra",
|
| 361 |
+
"341": "hog, pig, grunter, squealer, Sus scrofa",
|
| 362 |
+
"342": "wild boar, boar, Sus scrofa",
|
| 363 |
+
"343": "warthog",
|
| 364 |
+
"344": "hippopotamus, hippo, river horse, Hippopotamus amphibius",
|
| 365 |
+
"345": "ox",
|
| 366 |
+
"346": "water buffalo, water ox, Asiatic buffalo, Bubalus bubalis",
|
| 367 |
+
"347": "bison",
|
| 368 |
+
"348": "ram, tup",
|
| 369 |
+
"349": "bighorn, bighorn sheep, cimarron, Rocky Mountain bighorn, Rocky Mountain sheep, Ovis canadensis",
|
| 370 |
+
"350": "ibex, Capra ibex",
|
| 371 |
+
"351": "hartebeest",
|
| 372 |
+
"352": "impala, Aepyceros melampus",
|
| 373 |
+
"353": "gazelle",
|
| 374 |
+
"354": "Arabian camel, dromedary, Camelus dromedarius",
|
| 375 |
+
"355": "llama",
|
| 376 |
+
"356": "weasel",
|
| 377 |
+
"357": "mink",
|
| 378 |
+
"358": "polecat, fitch, foulmart, foumart, Mustela putorius",
|
| 379 |
+
"359": "black-footed ferret, ferret, Mustela nigripes",
|
| 380 |
+
"360": "otter",
|
| 381 |
+
"361": "skunk, polecat, wood pussy",
|
| 382 |
+
"362": "badger",
|
| 383 |
+
"363": "armadillo",
|
| 384 |
+
"364": "three-toed sloth, ai, Bradypus tridactylus",
|
| 385 |
+
"365": "orangutan, orang, orangutang, Pongo pygmaeus",
|
| 386 |
+
"366": "gorilla, Gorilla gorilla",
|
| 387 |
+
"367": "chimpanzee, chimp, Pan troglodytes",
|
| 388 |
+
"368": "gibbon, Hylobates lar",
|
| 389 |
+
"369": "siamang, Hylobates syndactylus, Symphalangus syndactylus",
|
| 390 |
+
"370": "guenon, guenon monkey",
|
| 391 |
+
"371": "patas, hussar monkey, Erythrocebus patas",
|
| 392 |
+
"372": "baboon",
|
| 393 |
+
"373": "macaque",
|
| 394 |
+
"374": "langur",
|
| 395 |
+
"375": "colobus, colobus monkey",
|
| 396 |
+
"376": "proboscis monkey, Nasalis larvatus",
|
| 397 |
+
"377": "marmoset",
|
| 398 |
+
"378": "capuchin, ringtail, Cebus capucinus",
|
| 399 |
+
"379": "howler monkey, howler",
|
| 400 |
+
"380": "titi, titi monkey",
|
| 401 |
+
"381": "spider monkey, Ateles geoffroyi",
|
| 402 |
+
"382": "squirrel monkey, Saimiri sciureus",
|
| 403 |
+
"383": "Madagascar cat, ring-tailed lemur, Lemur catta",
|
| 404 |
+
"384": "indri, indris, Indri indri, Indri brevicaudatus",
|
| 405 |
+
"385": "Indian elephant, Elephas maximus",
|
| 406 |
+
"386": "African elephant, Loxodonta africana",
|
| 407 |
+
"387": "lesser panda, red panda, panda, bear cat, cat bear, Ailurus fulgens",
|
| 408 |
+
"388": "giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca",
|
| 409 |
+
"389": "barracouta, snoek",
|
| 410 |
+
"390": "eel",
|
| 411 |
+
"391": "coho, cohoe, coho salmon, blue jack, silver salmon, Oncorhynchus kisutch",
|
| 412 |
+
"392": "rock beauty, Holocanthus tricolor",
|
| 413 |
+
"393": "anemone fish",
|
| 414 |
+
"394": "sturgeon",
|
| 415 |
+
"395": "gar, garfish, garpike, billfish, Lepisosteus osseus",
|
| 416 |
+
"396": "lionfish",
|
| 417 |
+
"397": "puffer, pufferfish, blowfish, globefish",
|
| 418 |
+
"398": "abacus",
|
| 419 |
+
"399": "abaya",
|
| 420 |
+
"400": "academic gown, academic robe, judge's robe",
|
| 421 |
+
"401": "accordion, piano accordion, squeeze box",
|
| 422 |
+
"402": "acoustic guitar",
|
| 423 |
+
"403": "aircraft carrier, carrier, flattop, attack aircraft carrier",
|
| 424 |
+
"404": "airliner",
|
| 425 |
+
"405": "airship, dirigible",
|
| 426 |
+
"406": "altar",
|
| 427 |
+
"407": "ambulance",
|
| 428 |
+
"408": "amphibian, amphibious vehicle",
|
| 429 |
+
"409": "analog clock",
|
| 430 |
+
"410": "apiary, bee house",
|
| 431 |
+
"411": "apron",
|
| 432 |
+
"412": "ashcan, trash can, garbage can, wastebin, ash bin, ash-bin, ashbin, dustbin, trash barrel, trash bin",
|
| 433 |
+
"413": "assault rifle, assault gun",
|
| 434 |
+
"414": "backpack, back pack, knapsack, packsack, rucksack, haversack",
|
| 435 |
+
"415": "bakery, bakeshop, bakehouse",
|
| 436 |
+
"416": "balance beam, beam",
|
| 437 |
+
"417": "balloon",
|
| 438 |
+
"418": "ballpoint, ballpoint pen, ballpen, Biro",
|
| 439 |
+
"419": "Band Aid",
|
| 440 |
+
"420": "banjo",
|
| 441 |
+
"421": "bannister, banister, balustrade, balusters, handrail",
|
| 442 |
+
"422": "barbell",
|
| 443 |
+
"423": "barber chair",
|
| 444 |
+
"424": "barbershop",
|
| 445 |
+
"425": "barn",
|
| 446 |
+
"426": "barometer",
|
| 447 |
+
"427": "barrel, cask",
|
| 448 |
+
"428": "barrow, garden cart, lawn cart, wheelbarrow",
|
| 449 |
+
"429": "baseball",
|
| 450 |
+
"430": "basketball",
|
| 451 |
+
"431": "bassinet",
|
| 452 |
+
"432": "bassoon",
|
| 453 |
+
"433": "bathing cap, swimming cap",
|
| 454 |
+
"434": "bath towel",
|
| 455 |
+
"435": "bathtub, bathing tub, bath, tub",
|
| 456 |
+
"436": "beach wagon, station wagon, wagon, estate car, beach waggon, station waggon, waggon",
|
| 457 |
+
"437": "beacon, lighthouse, beacon light, pharos",
|
| 458 |
+
"438": "beaker",
|
| 459 |
+
"439": "bearskin, busby, shako",
|
| 460 |
+
"440": "beer bottle",
|
| 461 |
+
"441": "beer glass",
|
| 462 |
+
"442": "bell cote, bell cot",
|
| 463 |
+
"443": "bib",
|
| 464 |
+
"444": "bicycle-built-for-two, tandem bicycle, tandem",
|
| 465 |
+
"445": "bikini, two-piece",
|
| 466 |
+
"446": "binder, ring-binder",
|
| 467 |
+
"447": "binoculars, field glasses, opera glasses",
|
| 468 |
+
"448": "birdhouse",
|
| 469 |
+
"449": "boathouse",
|
| 470 |
+
"450": "bobsled, bobsleigh, bob",
|
| 471 |
+
"451": "bolo tie, bolo, bola tie, bola",
|
| 472 |
+
"452": "bonnet, poke bonnet",
|
| 473 |
+
"453": "bookcase",
|
| 474 |
+
"454": "bookshop, bookstore, bookstall",
|
| 475 |
+
"455": "bottlecap",
|
| 476 |
+
"456": "bow",
|
| 477 |
+
"457": "bow tie, bow-tie, bowtie",
|
| 478 |
+
"458": "brass, memorial tablet, plaque",
|
| 479 |
+
"459": "brassiere, bra, bandeau",
|
| 480 |
+
"460": "breakwater, groin, groyne, mole, bulwark, seawall, jetty",
|
| 481 |
+
"461": "breastplate, aegis, egis",
|
| 482 |
+
"462": "broom",
|
| 483 |
+
"463": "bucket, pail",
|
| 484 |
+
"464": "buckle",
|
| 485 |
+
"465": "bulletproof vest",
|
| 486 |
+
"466": "bullet train, bullet",
|
| 487 |
+
"467": "butcher shop, meat market",
|
| 488 |
+
"468": "cab, hack, taxi, taxicab",
|
| 489 |
+
"469": "caldron, cauldron",
|
| 490 |
+
"470": "candle, taper, wax light",
|
| 491 |
+
"471": "cannon",
|
| 492 |
+
"472": "canoe",
|
| 493 |
+
"473": "can opener, tin opener",
|
| 494 |
+
"474": "cardigan",
|
| 495 |
+
"475": "car mirror",
|
| 496 |
+
"476": "carousel, carrousel, merry-go-round, roundabout, whirligig",
|
| 497 |
+
"477": "carpenter's kit, tool kit",
|
| 498 |
+
"478": "carton",
|
| 499 |
+
"479": "car wheel",
|
| 500 |
+
"480": "cash machine, cash dispenser, automated teller machine, automatic teller machine, automated teller, automatic teller, ATM",
|
| 501 |
+
"481": "cassette",
|
| 502 |
+
"482": "cassette player",
|
| 503 |
+
"483": "castle",
|
| 504 |
+
"484": "catamaran",
|
| 505 |
+
"485": "CD player",
|
| 506 |
+
"486": "cello, violoncello",
|
| 507 |
+
"487": "cellular telephone, cellular phone, cellphone, cell, mobile phone",
|
| 508 |
+
"488": "chain",
|
| 509 |
+
"489": "chainlink fence",
|
| 510 |
+
"490": "chain mail, ring mail, mail, chain armor, chain armour, ring armor, ring armour",
|
| 511 |
+
"491": "chain saw, chainsaw",
|
| 512 |
+
"492": "chest",
|
| 513 |
+
"493": "chiffonier, commode",
|
| 514 |
+
"494": "chime, bell, gong",
|
| 515 |
+
"495": "china cabinet, china closet",
|
| 516 |
+
"496": "Christmas stocking",
|
| 517 |
+
"497": "church, church building",
|
| 518 |
+
"498": "cinema, movie theater, movie theatre, movie house, picture palace",
|
| 519 |
+
"499": "cleaver, meat cleaver, chopper",
|
| 520 |
+
"500": "cliff dwelling",
|
| 521 |
+
"501": "cloak",
|
| 522 |
+
"502": "clog, geta, patten, sabot",
|
| 523 |
+
"503": "cocktail shaker",
|
| 524 |
+
"504": "coffee mug",
|
| 525 |
+
"505": "coffeepot",
|
| 526 |
+
"506": "coil, spiral, volute, whorl, helix",
|
| 527 |
+
"507": "combination lock",
|
| 528 |
+
"508": "computer keyboard, keypad",
|
| 529 |
+
"509": "confectionery, confectionary, candy store",
|
| 530 |
+
"510": "container ship, containership, container vessel",
|
| 531 |
+
"511": "convertible",
|
| 532 |
+
"512": "corkscrew, bottle screw",
|
| 533 |
+
"513": "cornet, horn, trumpet, trump",
|
| 534 |
+
"514": "cowboy boot",
|
| 535 |
+
"515": "cowboy hat, ten-gallon hat",
|
| 536 |
+
"516": "cradle",
|
| 537 |
+
"517": "crane",
|
| 538 |
+
"518": "crash helmet",
|
| 539 |
+
"519": "crate",
|
| 540 |
+
"520": "crib, cot",
|
| 541 |
+
"521": "Crock Pot",
|
| 542 |
+
"522": "croquet ball",
|
| 543 |
+
"523": "crutch",
|
| 544 |
+
"524": "cuirass",
|
| 545 |
+
"525": "dam, dike, dyke",
|
| 546 |
+
"526": "desk",
|
| 547 |
+
"527": "desktop computer",
|
| 548 |
+
"528": "dial telephone, dial phone",
|
| 549 |
+
"529": "diaper, nappy, napkin",
|
| 550 |
+
"530": "digital clock",
|
| 551 |
+
"531": "digital watch",
|
| 552 |
+
"532": "dining table, board",
|
| 553 |
+
"533": "dishrag, dishcloth",
|
| 554 |
+
"534": "dishwasher, dish washer, dishwashing machine",
|
| 555 |
+
"535": "disk brake, disc brake",
|
| 556 |
+
"536": "dock, dockage, docking facility",
|
| 557 |
+
"537": "dogsled, dog sled, dog sleigh",
|
| 558 |
+
"538": "dome",
|
| 559 |
+
"539": "doormat, welcome mat",
|
| 560 |
+
"540": "drilling platform, offshore rig",
|
| 561 |
+
"541": "drum, membranophone, tympan",
|
| 562 |
+
"542": "drumstick",
|
| 563 |
+
"543": "dumbbell",
|
| 564 |
+
"544": "Dutch oven",
|
| 565 |
+
"545": "electric fan, blower",
|
| 566 |
+
"546": "electric guitar",
|
| 567 |
+
"547": "electric locomotive",
|
| 568 |
+
"548": "entertainment center",
|
| 569 |
+
"549": "envelope",
|
| 570 |
+
"550": "espresso maker",
|
| 571 |
+
"551": "face powder",
|
| 572 |
+
"552": "feather boa, boa",
|
| 573 |
+
"553": "file, file cabinet, filing cabinet",
|
| 574 |
+
"554": "fireboat",
|
| 575 |
+
"555": "fire engine, fire truck",
|
| 576 |
+
"556": "fire screen, fireguard",
|
| 577 |
+
"557": "flagpole, flagstaff",
|
| 578 |
+
"558": "flute, transverse flute",
|
| 579 |
+
"559": "folding chair",
|
| 580 |
+
"560": "football helmet",
|
| 581 |
+
"561": "forklift",
|
| 582 |
+
"562": "fountain",
|
| 583 |
+
"563": "fountain pen",
|
| 584 |
+
"564": "four-poster",
|
| 585 |
+
"565": "freight car",
|
| 586 |
+
"566": "French horn, horn",
|
| 587 |
+
"567": "frying pan, frypan, skillet",
|
| 588 |
+
"568": "fur coat",
|
| 589 |
+
"569": "garbage truck, dustcart",
|
| 590 |
+
"570": "gasmask, respirator, gas helmet",
|
| 591 |
+
"571": "gas pump, gasoline pump, petrol pump, island dispenser",
|
| 592 |
+
"572": "goblet",
|
| 593 |
+
"573": "go-kart",
|
| 594 |
+
"574": "golf ball",
|
| 595 |
+
"575": "golfcart, golf cart",
|
| 596 |
+
"576": "gondola",
|
| 597 |
+
"577": "gong, tam-tam",
|
| 598 |
+
"578": "gown",
|
| 599 |
+
"579": "grand piano, grand",
|
| 600 |
+
"580": "greenhouse, nursery, glasshouse",
|
| 601 |
+
"581": "grille, radiator grille",
|
| 602 |
+
"582": "grocery store, grocery, food market, market",
|
| 603 |
+
"583": "guillotine",
|
| 604 |
+
"584": "hair slide",
|
| 605 |
+
"585": "hair spray",
|
| 606 |
+
"586": "half track",
|
| 607 |
+
"587": "hammer",
|
| 608 |
+
"588": "hamper",
|
| 609 |
+
"589": "hand blower, blow dryer, blow drier, hair dryer, hair drier",
|
| 610 |
+
"590": "hand-held computer, hand-held microcomputer",
|
| 611 |
+
"591": "handkerchief, hankie, hanky, hankey",
|
| 612 |
+
"592": "hard disc, hard disk, fixed disk",
|
| 613 |
+
"593": "harmonica, mouth organ, harp, mouth harp",
|
| 614 |
+
"594": "harp",
|
| 615 |
+
"595": "harvester, reaper",
|
| 616 |
+
"596": "hatchet",
|
| 617 |
+
"597": "holster",
|
| 618 |
+
"598": "home theater, home theatre",
|
| 619 |
+
"599": "honeycomb",
|
| 620 |
+
"600": "hook, claw",
|
| 621 |
+
"601": "hoopskirt, crinoline",
|
| 622 |
+
"602": "horizontal bar, high bar",
|
| 623 |
+
"603": "horse cart, horse-cart",
|
| 624 |
+
"604": "hourglass",
|
| 625 |
+
"605": "iPod",
|
| 626 |
+
"606": "iron, smoothing iron",
|
| 627 |
+
"607": "jack-o'-lantern",
|
| 628 |
+
"608": "jean, blue jean, denim",
|
| 629 |
+
"609": "jeep, landrover",
|
| 630 |
+
"610": "jersey, T-shirt, tee shirt",
|
| 631 |
+
"611": "jigsaw puzzle",
|
| 632 |
+
"612": "jinrikisha, ricksha, rickshaw",
|
| 633 |
+
"613": "joystick",
|
| 634 |
+
"614": "kimono",
|
| 635 |
+
"615": "knee pad",
|
| 636 |
+
"616": "knot",
|
| 637 |
+
"617": "lab coat, laboratory coat",
|
| 638 |
+
"618": "ladle",
|
| 639 |
+
"619": "lampshade, lamp shade",
|
| 640 |
+
"620": "laptop, laptop computer",
|
| 641 |
+
"621": "lawn mower, mower",
|
| 642 |
+
"622": "lens cap, lens cover",
|
| 643 |
+
"623": "letter opener, paper knife, paperknife",
|
| 644 |
+
"624": "library",
|
| 645 |
+
"625": "lifeboat",
|
| 646 |
+
"626": "lighter, light, igniter, ignitor",
|
| 647 |
+
"627": "limousine, limo",
|
| 648 |
+
"628": "liner, ocean liner",
|
| 649 |
+
"629": "lipstick, lip rouge",
|
| 650 |
+
"630": "Loafer",
|
| 651 |
+
"631": "lotion",
|
| 652 |
+
"632": "loudspeaker, speaker, speaker unit, loudspeaker system, speaker system",
|
| 653 |
+
"633": "loupe, jeweler's loupe",
|
| 654 |
+
"634": "lumbermill, sawmill",
|
| 655 |
+
"635": "magnetic compass",
|
| 656 |
+
"636": "mailbag, postbag",
|
| 657 |
+
"637": "mailbox, letter box",
|
| 658 |
+
"638": "maillot",
|
| 659 |
+
"639": "maillot, tank suit",
|
| 660 |
+
"640": "manhole cover",
|
| 661 |
+
"641": "maraca",
|
| 662 |
+
"642": "marimba, xylophone",
|
| 663 |
+
"643": "mask",
|
| 664 |
+
"644": "matchstick",
|
| 665 |
+
"645": "maypole",
|
| 666 |
+
"646": "maze, labyrinth",
|
| 667 |
+
"647": "measuring cup",
|
| 668 |
+
"648": "medicine chest, medicine cabinet",
|
| 669 |
+
"649": "megalith, megalithic structure",
|
| 670 |
+
"650": "microphone, mike",
|
| 671 |
+
"651": "microwave, microwave oven",
|
| 672 |
+
"652": "military uniform",
|
| 673 |
+
"653": "milk can",
|
| 674 |
+
"654": "minibus",
|
| 675 |
+
"655": "miniskirt, mini",
|
| 676 |
+
"656": "minivan",
|
| 677 |
+
"657": "missile",
|
| 678 |
+
"658": "mitten",
|
| 679 |
+
"659": "mixing bowl",
|
| 680 |
+
"660": "mobile home, manufactured home",
|
| 681 |
+
"661": "Model T",
|
| 682 |
+
"662": "modem",
|
| 683 |
+
"663": "monastery",
|
| 684 |
+
"664": "monitor",
|
| 685 |
+
"665": "moped",
|
| 686 |
+
"666": "mortar",
|
| 687 |
+
"667": "mortarboard",
|
| 688 |
+
"668": "mosque",
|
| 689 |
+
"669": "mosquito net",
|
| 690 |
+
"670": "motor scooter, scooter",
|
| 691 |
+
"671": "mountain bike, all-terrain bike, off-roader",
|
| 692 |
+
"672": "mountain tent",
|
| 693 |
+
"673": "mouse, computer mouse",
|
| 694 |
+
"674": "mousetrap",
|
| 695 |
+
"675": "moving van",
|
| 696 |
+
"676": "muzzle",
|
| 697 |
+
"677": "nail",
|
| 698 |
+
"678": "neck brace",
|
| 699 |
+
"679": "necklace",
|
| 700 |
+
"680": "nipple",
|
| 701 |
+
"681": "notebook, notebook computer",
|
| 702 |
+
"682": "obelisk",
|
| 703 |
+
"683": "oboe, hautboy, hautbois",
|
| 704 |
+
"684": "ocarina, sweet potato",
|
| 705 |
+
"685": "odometer, hodometer, mileometer, milometer",
|
| 706 |
+
"686": "oil filter",
|
| 707 |
+
"687": "organ, pipe organ",
|
| 708 |
+
"688": "oscilloscope, scope, cathode-ray oscilloscope, CRO",
|
| 709 |
+
"689": "overskirt",
|
| 710 |
+
"690": "oxcart",
|
| 711 |
+
"691": "oxygen mask",
|
| 712 |
+
"692": "packet",
|
| 713 |
+
"693": "paddle, boat paddle",
|
| 714 |
+
"694": "paddlewheel, paddle wheel",
|
| 715 |
+
"695": "padlock",
|
| 716 |
+
"696": "paintbrush",
|
| 717 |
+
"697": "pajama, pyjama, pj's, jammies",
|
| 718 |
+
"698": "palace",
|
| 719 |
+
"699": "panpipe, pandean pipe, syrinx",
|
| 720 |
+
"700": "paper towel",
|
| 721 |
+
"701": "parachute, chute",
|
| 722 |
+
"702": "parallel bars, bars",
|
| 723 |
+
"703": "park bench",
|
| 724 |
+
"704": "parking meter",
|
| 725 |
+
"705": "passenger car, coach, carriage",
|
| 726 |
+
"706": "patio, terrace",
|
| 727 |
+
"707": "pay-phone, pay-station",
|
| 728 |
+
"708": "pedestal, plinth, footstall",
|
| 729 |
+
"709": "pencil box, pencil case",
|
| 730 |
+
"710": "pencil sharpener",
|
| 731 |
+
"711": "perfume, essence",
|
| 732 |
+
"712": "Petri dish",
|
| 733 |
+
"713": "photocopier",
|
| 734 |
+
"714": "pick, plectrum, plectron",
|
| 735 |
+
"715": "pickelhaube",
|
| 736 |
+
"716": "picket fence, paling",
|
| 737 |
+
"717": "pickup, pickup truck",
|
| 738 |
+
"718": "pier",
|
| 739 |
+
"719": "piggy bank, penny bank",
|
| 740 |
+
"720": "pill bottle",
|
| 741 |
+
"721": "pillow",
|
| 742 |
+
"722": "ping-pong ball",
|
| 743 |
+
"723": "pinwheel",
|
| 744 |
+
"724": "pirate, pirate ship",
|
| 745 |
+
"725": "pitcher, ewer",
|
| 746 |
+
"726": "plane, carpenter's plane, woodworking plane",
|
| 747 |
+
"727": "planetarium",
|
| 748 |
+
"728": "plastic bag",
|
| 749 |
+
"729": "plate rack",
|
| 750 |
+
"730": "plow, plough",
|
| 751 |
+
"731": "plunger, plumber's helper",
|
| 752 |
+
"732": "Polaroid camera, Polaroid Land camera",
|
| 753 |
+
"733": "pole",
|
| 754 |
+
"734": "police van, police wagon, paddy wagon, patrol wagon, wagon, black Maria",
|
| 755 |
+
"735": "poncho",
|
| 756 |
+
"736": "pool table, billiard table, snooker table",
|
| 757 |
+
"737": "pop bottle, soda bottle",
|
| 758 |
+
"738": "pot, flowerpot",
|
| 759 |
+
"739": "potter's wheel",
|
| 760 |
+
"740": "power drill",
|
| 761 |
+
"741": "prayer rug, prayer mat",
|
| 762 |
+
"742": "printer",
|
| 763 |
+
"743": "prison, prison house",
|
| 764 |
+
"744": "projectile, missile",
|
| 765 |
+
"745": "projector",
|
| 766 |
+
"746": "puck, hockey puck",
|
| 767 |
+
"747": "punching bag, punch bag, punching ball, punchball",
|
| 768 |
+
"748": "purse",
|
| 769 |
+
"749": "quill, quill pen",
|
| 770 |
+
"750": "quilt, comforter, comfort, puff",
|
| 771 |
+
"751": "racer, race car, racing car",
|
| 772 |
+
"752": "racket, racquet",
|
| 773 |
+
"753": "radiator",
|
| 774 |
+
"754": "radio, wireless",
|
| 775 |
+
"755": "radio telescope, radio reflector",
|
| 776 |
+
"756": "rain barrel",
|
| 777 |
+
"757": "recreational vehicle, RV, R.V.",
|
| 778 |
+
"758": "reel",
|
| 779 |
+
"759": "reflex camera",
|
| 780 |
+
"760": "refrigerator, icebox",
|
| 781 |
+
"761": "remote control, remote",
|
| 782 |
+
"762": "restaurant, eating house, eating place, eatery",
|
| 783 |
+
"763": "revolver, six-gun, six-shooter",
|
| 784 |
+
"764": "rifle",
|
| 785 |
+
"765": "rocking chair, rocker",
|
| 786 |
+
"766": "rotisserie",
|
| 787 |
+
"767": "rubber eraser, rubber, pencil eraser",
|
| 788 |
+
"768": "rugby ball",
|
| 789 |
+
"769": "rule, ruler",
|
| 790 |
+
"770": "running shoe",
|
| 791 |
+
"771": "safe",
|
| 792 |
+
"772": "safety pin",
|
| 793 |
+
"773": "saltshaker, salt shaker",
|
| 794 |
+
"774": "sandal",
|
| 795 |
+
"775": "sarong",
|
| 796 |
+
"776": "sax, saxophone",
|
| 797 |
+
"777": "scabbard",
|
| 798 |
+
"778": "scale, weighing machine",
|
| 799 |
+
"779": "school bus",
|
| 800 |
+
"780": "schooner",
|
| 801 |
+
"781": "scoreboard",
|
| 802 |
+
"782": "screen, CRT screen",
|
| 803 |
+
"783": "screw",
|
| 804 |
+
"784": "screwdriver",
|
| 805 |
+
"785": "seat belt, seatbelt",
|
| 806 |
+
"786": "sewing machine",
|
| 807 |
+
"787": "shield, buckler",
|
| 808 |
+
"788": "shoe shop, shoe-shop, shoe store",
|
| 809 |
+
"789": "shoji",
|
| 810 |
+
"790": "shopping basket",
|
| 811 |
+
"791": "shopping cart",
|
| 812 |
+
"792": "shovel",
|
| 813 |
+
"793": "shower cap",
|
| 814 |
+
"794": "shower curtain",
|
| 815 |
+
"795": "ski",
|
| 816 |
+
"796": "ski mask",
|
| 817 |
+
"797": "sleeping bag",
|
| 818 |
+
"798": "slide rule, slipstick",
|
| 819 |
+
"799": "sliding door",
|
| 820 |
+
"800": "slot, one-armed bandit",
|
| 821 |
+
"801": "snorkel",
|
| 822 |
+
"802": "snowmobile",
|
| 823 |
+
"803": "snowplow, snowplough",
|
| 824 |
+
"804": "soap dispenser",
|
| 825 |
+
"805": "soccer ball",
|
| 826 |
+
"806": "sock",
|
| 827 |
+
"807": "solar dish, solar collector, solar furnace",
|
| 828 |
+
"808": "sombrero",
|
| 829 |
+
"809": "soup bowl",
|
| 830 |
+
"810": "space bar",
|
| 831 |
+
"811": "space heater",
|
| 832 |
+
"812": "space shuttle",
|
| 833 |
+
"813": "spatula",
|
| 834 |
+
"814": "speedboat",
|
| 835 |
+
"815": "spider web, spider's web",
|
| 836 |
+
"816": "spindle",
|
| 837 |
+
"817": "sports car, sport car",
|
| 838 |
+
"818": "spotlight, spot",
|
| 839 |
+
"819": "stage",
|
| 840 |
+
"820": "steam locomotive",
|
| 841 |
+
"821": "steel arch bridge",
|
| 842 |
+
"822": "steel drum",
|
| 843 |
+
"823": "stethoscope",
|
| 844 |
+
"824": "stole",
|
| 845 |
+
"825": "stone wall",
|
| 846 |
+
"826": "stopwatch, stop watch",
|
| 847 |
+
"827": "stove",
|
| 848 |
+
"828": "strainer",
|
| 849 |
+
"829": "streetcar, tram, tramcar, trolley, trolley car",
|
| 850 |
+
"830": "stretcher",
|
| 851 |
+
"831": "studio couch, day bed",
|
| 852 |
+
"832": "stupa, tope",
|
| 853 |
+
"833": "submarine, pigboat, sub, U-boat",
|
| 854 |
+
"834": "suit, suit of clothes",
|
| 855 |
+
"835": "sundial",
|
| 856 |
+
"836": "sunglass",
|
| 857 |
+
"837": "sunglasses, dark glasses, shades",
|
| 858 |
+
"838": "sunscreen, sunblock, sun blocker",
|
| 859 |
+
"839": "suspension bridge",
|
| 860 |
+
"840": "swab, swob, mop",
|
| 861 |
+
"841": "sweatshirt",
|
| 862 |
+
"842": "swimming trunks, bathing trunks",
|
| 863 |
+
"843": "swing",
|
| 864 |
+
"844": "switch, electric switch, electrical switch",
|
| 865 |
+
"845": "syringe",
|
| 866 |
+
"846": "table lamp",
|
| 867 |
+
"847": "tank, army tank, armored combat vehicle, armoured combat vehicle",
|
| 868 |
+
"848": "tape player",
|
| 869 |
+
"849": "teapot",
|
| 870 |
+
"850": "teddy, teddy bear",
|
| 871 |
+
"851": "television, television system",
|
| 872 |
+
"852": "tennis ball",
|
| 873 |
+
"853": "thatch, thatched roof",
|
| 874 |
+
"854": "theater curtain, theatre curtain",
|
| 875 |
+
"855": "thimble",
|
| 876 |
+
"856": "thresher, thrasher, threshing machine",
|
| 877 |
+
"857": "throne",
|
| 878 |
+
"858": "tile roof",
|
| 879 |
+
"859": "toaster",
|
| 880 |
+
"860": "tobacco shop, tobacconist shop, tobacconist",
|
| 881 |
+
"861": "toilet seat",
|
| 882 |
+
"862": "torch",
|
| 883 |
+
"863": "totem pole",
|
| 884 |
+
"864": "tow truck, tow car, wrecker",
|
| 885 |
+
"865": "toyshop",
|
| 886 |
+
"866": "tractor",
|
| 887 |
+
"867": "trailer truck, tractor trailer, trucking rig, rig, articulated lorry, semi",
|
| 888 |
+
"868": "tray",
|
| 889 |
+
"869": "trench coat",
|
| 890 |
+
"870": "tricycle, trike, velocipede",
|
| 891 |
+
"871": "trimaran",
|
| 892 |
+
"872": "tripod",
|
| 893 |
+
"873": "triumphal arch",
|
| 894 |
+
"874": "trolleybus, trolley coach, trackless trolley",
|
| 895 |
+
"875": "trombone",
|
| 896 |
+
"876": "tub, vat",
|
| 897 |
+
"877": "turnstile",
|
| 898 |
+
"878": "typewriter keyboard",
|
| 899 |
+
"879": "umbrella",
|
| 900 |
+
"880": "unicycle, monocycle",
|
| 901 |
+
"881": "upright, upright piano",
|
| 902 |
+
"882": "vacuum, vacuum cleaner",
|
| 903 |
+
"883": "vase",
|
| 904 |
+
"884": "vault",
|
| 905 |
+
"885": "velvet",
|
| 906 |
+
"886": "vending machine",
|
| 907 |
+
"887": "vestment",
|
| 908 |
+
"888": "viaduct",
|
| 909 |
+
"889": "violin, fiddle",
|
| 910 |
+
"890": "volleyball",
|
| 911 |
+
"891": "waffle iron",
|
| 912 |
+
"892": "wall clock",
|
| 913 |
+
"893": "wallet, billfold, notecase, pocketbook",
|
| 914 |
+
"894": "wardrobe, closet, press",
|
| 915 |
+
"895": "warplane, military plane",
|
| 916 |
+
"896": "washbasin, handbasin, washbowl, lavabo, wash-hand basin",
|
| 917 |
+
"897": "washer, automatic washer, washing machine",
|
| 918 |
+
"898": "water bottle",
|
| 919 |
+
"899": "water jug",
|
| 920 |
+
"900": "water tower",
|
| 921 |
+
"901": "whiskey jug",
|
| 922 |
+
"902": "whistle",
|
| 923 |
+
"903": "wig",
|
| 924 |
+
"904": "window screen",
|
| 925 |
+
"905": "window shade",
|
| 926 |
+
"906": "Windsor tie",
|
| 927 |
+
"907": "wine bottle",
|
| 928 |
+
"908": "wing",
|
| 929 |
+
"909": "wok",
|
| 930 |
+
"910": "wooden spoon",
|
| 931 |
+
"911": "wool, woolen, woollen",
|
| 932 |
+
"912": "worm fence, snake fence, snake-rail fence, Virginia fence",
|
| 933 |
+
"913": "wreck",
|
| 934 |
+
"914": "yawl",
|
| 935 |
+
"915": "yurt",
|
| 936 |
+
"916": "web site, website, internet site, site",
|
| 937 |
+
"917": "comic book",
|
| 938 |
+
"918": "crossword puzzle, crossword",
|
| 939 |
+
"919": "street sign",
|
| 940 |
+
"920": "traffic light, traffic signal, stoplight",
|
| 941 |
+
"921": "book jacket, dust cover, dust jacket, dust wrapper",
|
| 942 |
+
"922": "menu",
|
| 943 |
+
"923": "plate",
|
| 944 |
+
"924": "guacamole",
|
| 945 |
+
"925": "consomme",
|
| 946 |
+
"926": "hot pot, hotpot",
|
| 947 |
+
"927": "trifle",
|
| 948 |
+
"928": "ice cream, icecream",
|
| 949 |
+
"929": "ice lolly, lolly, lollipop, popsicle",
|
| 950 |
+
"930": "French loaf",
|
| 951 |
+
"931": "bagel, beigel",
|
| 952 |
+
"932": "pretzel",
|
| 953 |
+
"933": "cheeseburger",
|
| 954 |
+
"934": "hotdog, hot dog, red hot",
|
| 955 |
+
"935": "mashed potato",
|
| 956 |
+
"936": "head cabbage",
|
| 957 |
+
"937": "broccoli",
|
| 958 |
+
"938": "cauliflower",
|
| 959 |
+
"939": "zucchini, courgette",
|
| 960 |
+
"940": "spaghetti squash",
|
| 961 |
+
"941": "acorn squash",
|
| 962 |
+
"942": "butternut squash",
|
| 963 |
+
"943": "cucumber, cuke",
|
| 964 |
+
"944": "artichoke, globe artichoke",
|
| 965 |
+
"945": "bell pepper",
|
| 966 |
+
"946": "cardoon",
|
| 967 |
+
"947": "mushroom",
|
| 968 |
+
"948": "Granny Smith",
|
| 969 |
+
"949": "strawberry",
|
| 970 |
+
"950": "orange",
|
| 971 |
+
"951": "lemon",
|
| 972 |
+
"952": "fig",
|
| 973 |
+
"953": "pineapple, ananas",
|
| 974 |
+
"954": "banana",
|
| 975 |
+
"955": "jackfruit, jak, jack",
|
| 976 |
+
"956": "custard apple",
|
| 977 |
+
"957": "pomegranate",
|
| 978 |
+
"958": "hay",
|
| 979 |
+
"959": "carbonara",
|
| 980 |
+
"960": "chocolate sauce, chocolate syrup",
|
| 981 |
+
"961": "dough",
|
| 982 |
+
"962": "meat loaf, meatloaf",
|
| 983 |
+
"963": "pizza, pizza pie",
|
| 984 |
+
"964": "potpie",
|
| 985 |
+
"965": "burrito",
|
| 986 |
+
"966": "red wine",
|
| 987 |
+
"967": "espresso",
|
| 988 |
+
"968": "cup",
|
| 989 |
+
"969": "eggnog",
|
| 990 |
+
"970": "alp",
|
| 991 |
+
"971": "bubble",
|
| 992 |
+
"972": "cliff, drop, drop-off",
|
| 993 |
+
"973": "coral reef",
|
| 994 |
+
"974": "geyser",
|
| 995 |
+
"975": "lakeside, lakeshore",
|
| 996 |
+
"976": "promontory, headland, head, foreland",
|
| 997 |
+
"977": "sandbar, sand bar",
|
| 998 |
+
"978": "seashore, coast, seacoast, sea-coast",
|
| 999 |
+
"979": "valley, vale",
|
| 1000 |
+
"980": "volcano",
|
| 1001 |
+
"981": "ballplayer, baseball player",
|
| 1002 |
+
"982": "groom, bridegroom",
|
| 1003 |
+
"983": "scuba diver",
|
| 1004 |
+
"984": "rapeseed",
|
| 1005 |
+
"985": "daisy",
|
| 1006 |
+
"986": "yellow lady's slipper, yellow lady-slipper, Cypripedium calceolus, Cypripedium parviflorum",
|
| 1007 |
+
"987": "corn",
|
| 1008 |
+
"988": "acorn",
|
| 1009 |
+
"989": "hip, rose hip, rosehip",
|
| 1010 |
+
"990": "buckeye, horse chestnut, conker",
|
| 1011 |
+
"991": "coral fungus",
|
| 1012 |
+
"992": "agaric",
|
| 1013 |
+
"993": "gyromitra",
|
| 1014 |
+
"994": "stinkhorn, carrion fungus",
|
| 1015 |
+
"995": "earthstar",
|
| 1016 |
+
"996": "hen-of-the-woods, hen of the woods, Polyporus frondosus, Grifola frondosa",
|
| 1017 |
+
"997": "bolete",
|
| 1018 |
+
"998": "ear, spike, capitulum",
|
| 1019 |
+
"999": "toilet tissue, toilet paper, bathroom tissue"
|
| 1020 |
+
}
|
| 1021 |
+
}
|
DiT-MoE-B-8E2A/pipeline.py
ADDED
|
@@ -0,0 +1,440 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2026 The HuggingFace Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
|
| 15 |
+
import importlib
|
| 16 |
+
import inspect
|
| 17 |
+
import json
|
| 18 |
+
import sys
|
| 19 |
+
from pathlib import Path
|
| 20 |
+
from typing import Dict, List, Optional, Tuple, Union
|
| 21 |
+
|
| 22 |
+
import torch
|
| 23 |
+
|
| 24 |
+
from diffusers.image_processor import VaeImageProcessor
|
| 25 |
+
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
|
| 26 |
+
from diffusers.schedulers import DDIMScheduler, KarrasDiffusionSchedulers
|
| 27 |
+
from diffusers.utils import replace_example_docstring
|
| 28 |
+
from diffusers.utils.torch_utils import randn_tensor
|
| 29 |
+
|
| 30 |
+
DEFAULT_NATIVE_RESOLUTION = 256
|
| 31 |
+
|
| 32 |
+
EXAMPLE_DOC_STRING = """
|
| 33 |
+
Examples:
|
| 34 |
+
```py
|
| 35 |
+
>>> from pathlib import Path
|
| 36 |
+
>>> import torch
|
| 37 |
+
>>> from diffusers import DiffusionPipeline
|
| 38 |
+
|
| 39 |
+
>>> model_dir = Path("./DiT-MoE-S-8E2A").resolve()
|
| 40 |
+
>>> pipe = DiffusionPipeline.from_pretrained(
|
| 41 |
+
... str(model_dir),
|
| 42 |
+
... local_files_only=True,
|
| 43 |
+
... custom_pipeline=str(model_dir / "pipeline.py"),
|
| 44 |
+
... trust_remote_code=True,
|
| 45 |
+
... torch_dtype=torch.bfloat16,
|
| 46 |
+
... )
|
| 47 |
+
>>> pipe.to("cuda")
|
| 48 |
+
|
| 49 |
+
>>> print(pipe.id2label[207])
|
| 50 |
+
>>> print(pipe.get_label_ids("golden retriever"))
|
| 51 |
+
|
| 52 |
+
>>> class_id = pipe.get_label_ids("golden retriever")[0]
|
| 53 |
+
>>> generator = torch.Generator(device="cuda").manual_seed(42)
|
| 54 |
+
>>> image = pipe(
|
| 55 |
+
... class_labels=class_id,
|
| 56 |
+
... height=256,
|
| 57 |
+
... width=256,
|
| 58 |
+
... num_inference_steps=50,
|
| 59 |
+
... guidance_scale=4.0,
|
| 60 |
+
... generator=generator,
|
| 61 |
+
... ).images[0]
|
| 62 |
+
>>> image.save("demo.png")
|
| 63 |
+
```
|
| 64 |
+
"""
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
class DiTMoEPipeline(DiffusionPipeline):
|
| 68 |
+
r"""
|
| 69 |
+
Pipeline for class-conditional image generation with DiT-MoE.
|
| 70 |
+
|
| 71 |
+
Supports DDIM diffusion sampling and rectified-flow (RF) sampling.
|
| 72 |
+
Each checkpoint keeps an English `id2label` map in `model_index.json` (DiT-style).
|
| 73 |
+
"""
|
| 74 |
+
|
| 75 |
+
model_cpu_offload_seq = "transformer->vae"
|
| 76 |
+
_optional_components = ["vae"]
|
| 77 |
+
|
| 78 |
+
def __init__(
|
| 79 |
+
self,
|
| 80 |
+
transformer,
|
| 81 |
+
scheduler: KarrasDiffusionSchedulers,
|
| 82 |
+
vae=None,
|
| 83 |
+
id2label: Optional[Dict[Union[int, str], str]] = None,
|
| 84 |
+
null_class_id: Optional[int] = None,
|
| 85 |
+
):
|
| 86 |
+
super().__init__()
|
| 87 |
+
self.register_modules(transformer=transformer, scheduler=scheduler, vae=vae)
|
| 88 |
+
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
|
| 89 |
+
self._use_rectified_flow = scheduler.__class__.__name__ == "DiTMoEFlowMatchScheduler"
|
| 90 |
+
|
| 91 |
+
if null_class_id is None:
|
| 92 |
+
null_class_id = int(getattr(self.transformer.config, "num_classes", 1000))
|
| 93 |
+
self.register_to_config(null_class_id=int(null_class_id))
|
| 94 |
+
|
| 95 |
+
self._id2label = self._normalize_id2label(id2label)
|
| 96 |
+
self.labels = self._build_label2id(self._id2label)
|
| 97 |
+
|
| 98 |
+
@property
|
| 99 |
+
def vae_scale_factor(self) -> int:
|
| 100 |
+
if self.vae is None:
|
| 101 |
+
return 8
|
| 102 |
+
block_out_channels = getattr(self.vae.config, "block_out_channels", None)
|
| 103 |
+
if block_out_channels:
|
| 104 |
+
return int(2 ** (len(block_out_channels) - 1))
|
| 105 |
+
return 8
|
| 106 |
+
|
| 107 |
+
@classmethod
|
| 108 |
+
def from_pretrained(cls, pretrained_model_name_or_path=None, subfolder=None, **kwargs):
|
| 109 |
+
r"""Load a self-contained variant folder locally or from the Hub."""
|
| 110 |
+
repo_root = Path(__file__).resolve().parent
|
| 111 |
+
|
| 112 |
+
if pretrained_model_name_or_path in (None, "", "."):
|
| 113 |
+
variant = repo_root
|
| 114 |
+
elif (
|
| 115 |
+
isinstance(pretrained_model_name_or_path, str)
|
| 116 |
+
and "/" in pretrained_model_name_or_path
|
| 117 |
+
and not Path(pretrained_model_name_or_path).exists()
|
| 118 |
+
):
|
| 119 |
+
from huggingface_hub import snapshot_download
|
| 120 |
+
|
| 121 |
+
hub_kwargs = dict(kwargs.pop("hub_kwargs", {}))
|
| 122 |
+
if subfolder:
|
| 123 |
+
hub_kwargs.setdefault("allow_patterns", [f"{subfolder}/**"])
|
| 124 |
+
cache_dir = snapshot_download(pretrained_model_name_or_path, **hub_kwargs)
|
| 125 |
+
variant = Path(cache_dir) / subfolder if subfolder else Path(cache_dir)
|
| 126 |
+
else:
|
| 127 |
+
variant = Path(pretrained_model_name_or_path)
|
| 128 |
+
if not variant.is_absolute():
|
| 129 |
+
candidate = (Path.cwd() / variant).resolve()
|
| 130 |
+
variant = candidate if candidate.exists() else (repo_root / variant).resolve()
|
| 131 |
+
if subfolder:
|
| 132 |
+
variant = variant / subfolder
|
| 133 |
+
|
| 134 |
+
id2label_override = kwargs.pop("id2label", None)
|
| 135 |
+
null_class_id_override = kwargs.pop("null_class_id", None)
|
| 136 |
+
model_kwargs = dict(kwargs)
|
| 137 |
+
inserted: List[str] = []
|
| 138 |
+
|
| 139 |
+
def _load_component(folder: str, module_name: str, class_name: str):
|
| 140 |
+
comp_dir = variant / folder
|
| 141 |
+
module_path = comp_dir / f"{module_name}.py"
|
| 142 |
+
has_weights = (comp_dir / "config.json").exists() or (comp_dir / "scheduler_config.json").exists()
|
| 143 |
+
if not module_path.exists() or not has_weights:
|
| 144 |
+
return None
|
| 145 |
+
|
| 146 |
+
comp_path = str(comp_dir)
|
| 147 |
+
if comp_path not in sys.path:
|
| 148 |
+
sys.path.insert(0, comp_path)
|
| 149 |
+
inserted.append(comp_path)
|
| 150 |
+
|
| 151 |
+
module = importlib.import_module(module_name)
|
| 152 |
+
component_cls = getattr(module, class_name)
|
| 153 |
+
return component_cls.from_pretrained(str(comp_dir), **model_kwargs)
|
| 154 |
+
|
| 155 |
+
try:
|
| 156 |
+
transformer = _load_component("transformer", "transformer_dit_moe", "DiTMoETransformer2DModel")
|
| 157 |
+
if transformer is None:
|
| 158 |
+
raise ValueError(f"No loadable transformer found under {variant}")
|
| 159 |
+
|
| 160 |
+
scheduler = None
|
| 161 |
+
scheduler_config_path = variant / "scheduler" / "scheduler_config.json"
|
| 162 |
+
scheduler_dir = variant / "scheduler"
|
| 163 |
+
if scheduler_config_path.exists():
|
| 164 |
+
scheduler_config = json.loads(scheduler_config_path.read_text(encoding="utf-8"))
|
| 165 |
+
scheduler_class = scheduler_config.get("_class_name", "DDIMScheduler")
|
| 166 |
+
if scheduler_class == "DiTMoEFlowMatchScheduler":
|
| 167 |
+
scheduler_path = str(scheduler_dir)
|
| 168 |
+
if scheduler_path not in sys.path:
|
| 169 |
+
sys.path.insert(0, scheduler_path)
|
| 170 |
+
inserted.append(scheduler_path)
|
| 171 |
+
scheduler_module = importlib.import_module("scheduling_flow_match_dit_moe")
|
| 172 |
+
scheduler_cls = getattr(scheduler_module, "DiTMoEFlowMatchScheduler")
|
| 173 |
+
scheduler = scheduler_cls.from_pretrained(str(variant), subfolder="scheduler")
|
| 174 |
+
else:
|
| 175 |
+
scheduler = DDIMScheduler.from_pretrained(str(variant), subfolder="scheduler")
|
| 176 |
+
if scheduler is None:
|
| 177 |
+
scheduler = DDIMScheduler(num_train_timesteps=1000)
|
| 178 |
+
|
| 179 |
+
vae = None
|
| 180 |
+
vae_dir = variant / "vae"
|
| 181 |
+
if vae_dir.exists() and (vae_dir / "config.json").exists():
|
| 182 |
+
from diffusers import AutoencoderKL
|
| 183 |
+
|
| 184 |
+
vae = AutoencoderKL.from_pretrained(str(vae_dir), **model_kwargs)
|
| 185 |
+
|
| 186 |
+
pipeline_config = cls._read_pipeline_config_from_model_index(str(variant))
|
| 187 |
+
id2label = id2label_override or pipeline_config.get("id2label")
|
| 188 |
+
null_class_id = null_class_id_override if null_class_id_override is not None else pipeline_config.get(
|
| 189 |
+
"null_class_id"
|
| 190 |
+
)
|
| 191 |
+
pipe = cls(
|
| 192 |
+
transformer=transformer,
|
| 193 |
+
scheduler=scheduler,
|
| 194 |
+
vae=vae,
|
| 195 |
+
id2label=id2label,
|
| 196 |
+
null_class_id=null_class_id,
|
| 197 |
+
)
|
| 198 |
+
if hasattr(pipe, "register_to_config"):
|
| 199 |
+
pipe.register_to_config(_name_or_path=str(variant))
|
| 200 |
+
return pipe
|
| 201 |
+
finally:
|
| 202 |
+
for comp_path in inserted:
|
| 203 |
+
if comp_path in sys.path:
|
| 204 |
+
sys.path.remove(comp_path)
|
| 205 |
+
|
| 206 |
+
@staticmethod
|
| 207 |
+
def _normalize_id2label(id2label: Optional[Dict[Union[int, str], str]]) -> Dict[int, str]:
|
| 208 |
+
if not id2label:
|
| 209 |
+
return {}
|
| 210 |
+
return {int(key): value for key, value in id2label.items()}
|
| 211 |
+
|
| 212 |
+
@staticmethod
|
| 213 |
+
def _read_pipeline_config_from_model_index(variant_path: Optional[str]) -> Dict[str, object]:
|
| 214 |
+
if not variant_path:
|
| 215 |
+
return {}
|
| 216 |
+
variant_dir = Path(variant_path).resolve()
|
| 217 |
+
model_index_path = variant_dir / "model_index.json"
|
| 218 |
+
if not model_index_path.exists():
|
| 219 |
+
return {}
|
| 220 |
+
raw = json.loads(model_index_path.read_text(encoding="utf-8"))
|
| 221 |
+
config: Dict[str, object] = {}
|
| 222 |
+
id2label = raw.get("id2label")
|
| 223 |
+
if isinstance(id2label, dict):
|
| 224 |
+
config["id2label"] = {int(key): value for key, value in id2label.items()}
|
| 225 |
+
if "null_class_id" in raw:
|
| 226 |
+
config["null_class_id"] = int(raw["null_class_id"])
|
| 227 |
+
return config
|
| 228 |
+
|
| 229 |
+
@staticmethod
|
| 230 |
+
def _build_label2id(id2label: Dict[int, str]) -> Dict[str, int]:
|
| 231 |
+
label2id: Dict[str, int] = {}
|
| 232 |
+
for class_id, value in id2label.items():
|
| 233 |
+
for synonym in value.split(","):
|
| 234 |
+
synonym = synonym.strip()
|
| 235 |
+
if synonym:
|
| 236 |
+
label2id[synonym] = int(class_id)
|
| 237 |
+
return dict(sorted(label2id.items()))
|
| 238 |
+
|
| 239 |
+
@property
|
| 240 |
+
def id2label(self) -> Dict[int, str]:
|
| 241 |
+
return self._id2label
|
| 242 |
+
|
| 243 |
+
def get_label_ids(self, label: Union[str, List[str]]) -> List[int]:
|
| 244 |
+
r"""Map English ImageNet labels to class ids."""
|
| 245 |
+
labels = [label] if isinstance(label, str) else label
|
| 246 |
+
if not self.labels:
|
| 247 |
+
raise ValueError("No id2label mapping is available in this checkpoint.")
|
| 248 |
+
missing = [item for item in labels if item not in self.labels]
|
| 249 |
+
if missing:
|
| 250 |
+
preview = ", ".join(list(self.labels.keys())[:8])
|
| 251 |
+
raise ValueError(f"Unknown labels: {missing}. Example valid labels: {preview}, ...")
|
| 252 |
+
return [self.labels[item] for item in labels]
|
| 253 |
+
|
| 254 |
+
def _normalize_class_labels(
|
| 255 |
+
self,
|
| 256 |
+
class_labels: Union[int, str, List[Union[int, str]], torch.Tensor],
|
| 257 |
+
) -> List[int]:
|
| 258 |
+
if isinstance(class_labels, torch.Tensor):
|
| 259 |
+
class_labels = class_labels.detach().cpu().tolist()
|
| 260 |
+
if isinstance(class_labels, int):
|
| 261 |
+
return [class_labels]
|
| 262 |
+
if isinstance(class_labels, str):
|
| 263 |
+
return self.get_label_ids(class_labels)
|
| 264 |
+
if not class_labels:
|
| 265 |
+
raise ValueError("`class_labels` cannot be empty.")
|
| 266 |
+
if isinstance(class_labels[0], str):
|
| 267 |
+
return self.get_label_ids(class_labels) # type: ignore[arg-type]
|
| 268 |
+
return [int(class_id) for class_id in class_labels] # type: ignore[union-attr]
|
| 269 |
+
|
| 270 |
+
def prepare_latents(
|
| 271 |
+
self,
|
| 272 |
+
batch_size: int,
|
| 273 |
+
num_channels: int,
|
| 274 |
+
height: int,
|
| 275 |
+
width: int,
|
| 276 |
+
dtype: torch.dtype,
|
| 277 |
+
device: torch.device,
|
| 278 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]],
|
| 279 |
+
) -> torch.Tensor:
|
| 280 |
+
latent_height = height // self.vae_scale_factor
|
| 281 |
+
latent_width = width // self.vae_scale_factor
|
| 282 |
+
return randn_tensor(
|
| 283 |
+
(batch_size, num_channels, latent_height, latent_width),
|
| 284 |
+
generator=generator,
|
| 285 |
+
device=device,
|
| 286 |
+
dtype=dtype,
|
| 287 |
+
)
|
| 288 |
+
|
| 289 |
+
@staticmethod
|
| 290 |
+
def prepare_extra_step_kwargs(
|
| 291 |
+
scheduler: KarrasDiffusionSchedulers,
|
| 292 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]],
|
| 293 |
+
eta: float,
|
| 294 |
+
) -> Dict[str, object]:
|
| 295 |
+
kwargs: Dict[str, object] = {}
|
| 296 |
+
step_params = set(inspect.signature(scheduler.step).parameters.keys())
|
| 297 |
+
if "eta" in step_params:
|
| 298 |
+
kwargs["eta"] = eta
|
| 299 |
+
if "generator" in step_params:
|
| 300 |
+
kwargs["generator"] = generator
|
| 301 |
+
return kwargs
|
| 302 |
+
|
| 303 |
+
def _apply_cfg(self, model_output: torch.Tensor, guidance_scale: float) -> torch.Tensor:
|
| 304 |
+
if guidance_scale <= 1.0:
|
| 305 |
+
return model_output
|
| 306 |
+
cond, uncond = model_output.chunk(2)
|
| 307 |
+
if self.transformer.learn_sigma:
|
| 308 |
+
eps_cond, rest_cond = cond[:, : self.transformer.in_channels], cond[:, self.transformer.in_channels :]
|
| 309 |
+
eps_uncond, _rest_uncond = uncond[:, : self.transformer.in_channels], uncond[:, self.transformer.in_channels :]
|
| 310 |
+
eps = eps_uncond + guidance_scale * (eps_cond - eps_uncond)
|
| 311 |
+
guided = torch.cat([eps, rest_cond], dim=1)
|
| 312 |
+
else:
|
| 313 |
+
guided = uncond + guidance_scale * (cond - uncond)
|
| 314 |
+
return guided
|
| 315 |
+
|
| 316 |
+
def decode_latents(self, latents: torch.Tensor, output_type: str = "pil"):
|
| 317 |
+
if self.vae is None:
|
| 318 |
+
if output_type == "latent":
|
| 319 |
+
return latents
|
| 320 |
+
raise ValueError("Cannot decode latents without a VAE.")
|
| 321 |
+
scaling_factor = getattr(self.vae.config, "scaling_factor", 0.18215)
|
| 322 |
+
latents = latents / scaling_factor
|
| 323 |
+
if output_type == "latent":
|
| 324 |
+
return latents
|
| 325 |
+
image = self.vae.decode(latents).sample
|
| 326 |
+
return self.image_processor.postprocess(image, output_type=output_type)
|
| 327 |
+
|
| 328 |
+
@torch.inference_mode()
|
| 329 |
+
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
| 330 |
+
def __call__(
|
| 331 |
+
self,
|
| 332 |
+
class_labels: Union[int, str, List[Union[int, str]], torch.Tensor] = 207,
|
| 333 |
+
height: Optional[int] = None,
|
| 334 |
+
width: Optional[int] = None,
|
| 335 |
+
num_inference_steps: int = 50,
|
| 336 |
+
guidance_scale: float = 4.0,
|
| 337 |
+
eta: float = 0.0,
|
| 338 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
| 339 |
+
output_type: str = "pil",
|
| 340 |
+
return_dict: bool = True,
|
| 341 |
+
) -> Union[ImagePipelineOutput, Tuple]:
|
| 342 |
+
r"""
|
| 343 |
+
Generate class-conditional samples from a DiT-MoE checkpoint.
|
| 344 |
+
|
| 345 |
+
Examples:
|
| 346 |
+
<!-- this section is replaced by replace_example_docstring -->
|
| 347 |
+
"""
|
| 348 |
+
class_labels_list = self._normalize_class_labels(class_labels)
|
| 349 |
+
batch_size = len(class_labels_list)
|
| 350 |
+
native_size = int(getattr(self.transformer.config, "input_size", 32)) * self.vae_scale_factor
|
| 351 |
+
height = native_size if height is None else int(height)
|
| 352 |
+
width = native_size if width is None else int(width)
|
| 353 |
+
|
| 354 |
+
if height % self.vae_scale_factor != 0 or width % self.vae_scale_factor != 0:
|
| 355 |
+
raise ValueError(
|
| 356 |
+
f"`height` and `width` must be divisible by {self.vae_scale_factor}, got ({height}, {width})."
|
| 357 |
+
)
|
| 358 |
+
if output_type not in {"pil", "np", "pt", "latent"}:
|
| 359 |
+
raise ValueError(f"Unsupported `output_type`: {output_type}")
|
| 360 |
+
|
| 361 |
+
device = self._execution_device
|
| 362 |
+
model_dtype = next(self.transformer.parameters()).dtype
|
| 363 |
+
latents = self.prepare_latents(
|
| 364 |
+
batch_size,
|
| 365 |
+
self.transformer.config.in_channels,
|
| 366 |
+
height,
|
| 367 |
+
width,
|
| 368 |
+
model_dtype,
|
| 369 |
+
device,
|
| 370 |
+
generator,
|
| 371 |
+
)
|
| 372 |
+
|
| 373 |
+
do_cfg = guidance_scale > 1.0
|
| 374 |
+
class_labels_tensor = torch.tensor(class_labels_list, device=device, dtype=torch.long)
|
| 375 |
+
if do_cfg:
|
| 376 |
+
latents = torch.cat([latents, latents], dim=0)
|
| 377 |
+
null_class = int(self.config.null_class_id)
|
| 378 |
+
uncond = torch.full((batch_size,), null_class, device=device, dtype=torch.long)
|
| 379 |
+
labels = torch.cat([class_labels_tensor, uncond], dim=0)
|
| 380 |
+
else:
|
| 381 |
+
labels = class_labels_tensor
|
| 382 |
+
|
| 383 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(self.scheduler, generator=generator, eta=eta)
|
| 384 |
+
|
| 385 |
+
if self._use_rectified_flow:
|
| 386 |
+
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
| 387 |
+
timesteps = self.scheduler.timesteps
|
| 388 |
+
for index, timestep in enumerate(self.progress_bar(timesteps)):
|
| 389 |
+
next_timestep = timesteps[index + 1] if index + 1 < len(timesteps) else torch.tensor(
|
| 390 |
+
0.0, device=device
|
| 391 |
+
)
|
| 392 |
+
timestep_batch = torch.full((labels.shape[0],), float(timestep), device=device, dtype=model_dtype)
|
| 393 |
+
model_output = self.transformer(latents, timestep_batch, labels, return_dict=True).sample
|
| 394 |
+
model_output = self.transformer.split_velocity(model_output)
|
| 395 |
+
if do_cfg:
|
| 396 |
+
model_output = self._apply_cfg(model_output, guidance_scale)
|
| 397 |
+
latents_cfg = latents[:batch_size]
|
| 398 |
+
else:
|
| 399 |
+
latents_cfg = latents
|
| 400 |
+
step_output = self.scheduler.step(
|
| 401 |
+
model_output[:batch_size] if do_cfg else model_output,
|
| 402 |
+
timestep_batch[:batch_size] if do_cfg else timestep_batch,
|
| 403 |
+
latents_cfg,
|
| 404 |
+
next_timestep=next_timestep,
|
| 405 |
+
).prev_sample
|
| 406 |
+
latents = step_output if not do_cfg else torch.cat([step_output, step_output], dim=0)
|
| 407 |
+
latents = latents[:batch_size]
|
| 408 |
+
else:
|
| 409 |
+
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
| 410 |
+
for timestep in self.progress_bar(self.scheduler.timesteps):
|
| 411 |
+
timestep_batch = torch.full((labels.shape[0],), float(timestep), device=device, dtype=model_dtype)
|
| 412 |
+
latent_model_input = self.scheduler.scale_model_input(latents, timestep)
|
| 413 |
+
model_output = self.transformer(latent_model_input, timestep_batch, labels, return_dict=True).sample
|
| 414 |
+
if do_cfg:
|
| 415 |
+
model_output = self._apply_cfg(model_output, guidance_scale)
|
| 416 |
+
latents_input = latents[:batch_size]
|
| 417 |
+
model_output = model_output[:batch_size]
|
| 418 |
+
else:
|
| 419 |
+
latents_input = latents
|
| 420 |
+
if self.transformer.learn_sigma:
|
| 421 |
+
model_output, _ = torch.split(model_output, self.transformer.in_channels, dim=1)
|
| 422 |
+
latents = self.scheduler.step(
|
| 423 |
+
model_output,
|
| 424 |
+
timestep,
|
| 425 |
+
latents_input,
|
| 426 |
+
**extra_step_kwargs,
|
| 427 |
+
).prev_sample
|
| 428 |
+
if do_cfg:
|
| 429 |
+
latents = torch.cat([latents, latents], dim=0)
|
| 430 |
+
|
| 431 |
+
image = self.decode_latents(latents, output_type=output_type)
|
| 432 |
+
self.maybe_free_model_hooks()
|
| 433 |
+
if not return_dict:
|
| 434 |
+
return (image,)
|
| 435 |
+
return ImagePipelineOutput(images=image)
|
| 436 |
+
|
| 437 |
+
|
| 438 |
+
DiTMoEPipelineOutput = ImagePipelineOutput
|
| 439 |
+
|
| 440 |
+
__all__ = ["DiTMoEPipeline", "DiTMoEPipelineOutput"]
|
DiT-MoE-B-8E2A/scheduler/scheduler_config.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_class_name": "DDIMScheduler",
|
| 3 |
+
"_diffusers_version": "0.36.0",
|
| 4 |
+
"beta_end": 0.02,
|
| 5 |
+
"beta_schedule": "linear",
|
| 6 |
+
"beta_start": 0.0001,
|
| 7 |
+
"clip_sample": false,
|
| 8 |
+
"num_train_timesteps": 1000,
|
| 9 |
+
"prediction_type": "epsilon",
|
| 10 |
+
"set_alpha_to_one": true,
|
| 11 |
+
"steps_offset": 0,
|
| 12 |
+
"trained_betas": null
|
| 13 |
+
}
|
DiT-MoE-B-8E2A/transformer/config.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_class_name": "DiTMoETransformer2DModel",
|
| 3 |
+
"class_dropout_prob": 0.1,
|
| 4 |
+
"depth": 12,
|
| 5 |
+
"hidden_size": 768,
|
| 6 |
+
"in_channels": 4,
|
| 7 |
+
"input_size": 32,
|
| 8 |
+
"learn_sigma": true,
|
| 9 |
+
"num_classes": 1000,
|
| 10 |
+
"num_experts": 8,
|
| 11 |
+
"num_experts_per_tok": 2,
|
| 12 |
+
"num_heads": 12,
|
| 13 |
+
"patch_size": 2,
|
| 14 |
+
"pretraining_tp": 2,
|
| 15 |
+
"use_flash_attn": false
|
| 16 |
+
}
|
DiT-MoE-B-8E2A/transformer/diffusion_pytorch_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b515c7400d23980bc73f178cde2d7b35bfbb1dc0e43ccf24284f981e7fac40af
|
| 3 |
+
size 3183493168
|
DiT-MoE-B-8E2A/transformer/transformer_dit_moe.py
ADDED
|
@@ -0,0 +1,511 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
| 2 |
+
# Copyright 2026 The HuggingFace Team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
|
| 7 |
+
from dataclasses import dataclass
|
| 8 |
+
import math
|
| 9 |
+
from typing import Optional, Tuple, Union
|
| 10 |
+
|
| 11 |
+
import numpy as np
|
| 12 |
+
import torch
|
| 13 |
+
import torch.nn as nn
|
| 14 |
+
import torch.nn.functional as F
|
| 15 |
+
|
| 16 |
+
try:
|
| 17 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
| 18 |
+
from diffusers.models.modeling_utils import ModelMixin
|
| 19 |
+
from diffusers.utils import BaseOutput
|
| 20 |
+
except Exception: # pragma: no cover
|
| 21 |
+
class BaseOutput(dict):
|
| 22 |
+
def __post_init__(self):
|
| 23 |
+
self.update(self.__dict__)
|
| 24 |
+
|
| 25 |
+
class _Config(dict):
|
| 26 |
+
def __getattr__(self, key):
|
| 27 |
+
try:
|
| 28 |
+
return self[key]
|
| 29 |
+
except KeyError as error:
|
| 30 |
+
raise AttributeError(key) from error
|
| 31 |
+
|
| 32 |
+
class ConfigMixin:
|
| 33 |
+
config_name = "config.json"
|
| 34 |
+
|
| 35 |
+
class ModelMixin(nn.Module):
|
| 36 |
+
pass
|
| 37 |
+
|
| 38 |
+
def register_to_config(init):
|
| 39 |
+
def wrapper(self, *args, **kwargs):
|
| 40 |
+
import inspect
|
| 41 |
+
|
| 42 |
+
signature = inspect.signature(init)
|
| 43 |
+
bound = signature.bind(self, *args, **kwargs)
|
| 44 |
+
bound.apply_defaults()
|
| 45 |
+
self.config = _Config({key: value for key, value in bound.arguments.items() if key != "self"})
|
| 46 |
+
init(self, *args, **kwargs)
|
| 47 |
+
|
| 48 |
+
return wrapper
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
try:
|
| 52 |
+
from timm.models.vision_transformer import Attention
|
| 53 |
+
except Exception: # pragma: no cover
|
| 54 |
+
Attention = None
|
| 55 |
+
|
| 56 |
+
try:
|
| 57 |
+
import flash_attn
|
| 58 |
+
|
| 59 |
+
if hasattr(flash_attn, "__version__") and int(flash_attn.__version__[0]) == 2:
|
| 60 |
+
from flash_attn.modules.mha import FlashSelfAttention
|
| 61 |
+
else:
|
| 62 |
+
from flash_attn.modules.mha import FlashSelfAttention
|
| 63 |
+
except Exception:
|
| 64 |
+
FlashSelfAttention = None
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
@dataclass
|
| 68 |
+
class DiTMoETransformer2DModelOutput(BaseOutput):
|
| 69 |
+
sample: torch.FloatTensor
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
|
| 73 |
+
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
class DiTMoEPatchEmbed(nn.Module):
|
| 77 |
+
def __init__(self, input_size: int, patch_size: int, in_channels: int, embed_dim: int, bias: bool = True):
|
| 78 |
+
super().__init__()
|
| 79 |
+
self.input_size = (input_size, input_size)
|
| 80 |
+
self.patch_size = (patch_size, patch_size)
|
| 81 |
+
self.num_patches = (input_size // patch_size) ** 2
|
| 82 |
+
self.proj = nn.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
|
| 83 |
+
|
| 84 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 85 |
+
hidden_states = self.proj(hidden_states)
|
| 86 |
+
return hidden_states.flatten(2).transpose(1, 2)
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
class DiTMoETimestepEmbedder(nn.Module):
|
| 90 |
+
def __init__(self, hidden_size: int, frequency_embedding_size: int = 256):
|
| 91 |
+
super().__init__()
|
| 92 |
+
self.mlp = nn.Sequential(
|
| 93 |
+
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
|
| 94 |
+
nn.SiLU(),
|
| 95 |
+
nn.Linear(hidden_size, hidden_size, bias=True),
|
| 96 |
+
)
|
| 97 |
+
self.frequency_embedding_size = frequency_embedding_size
|
| 98 |
+
|
| 99 |
+
@staticmethod
|
| 100 |
+
def timestep_embedding(timesteps: torch.Tensor, dim: int, max_period: int = 10000):
|
| 101 |
+
half = dim // 2
|
| 102 |
+
freqs = torch.exp(
|
| 103 |
+
-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=timesteps.device) / half
|
| 104 |
+
)
|
| 105 |
+
args = timesteps[:, None].float() * freqs[None]
|
| 106 |
+
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
| 107 |
+
if dim % 2:
|
| 108 |
+
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
| 109 |
+
return embedding
|
| 110 |
+
|
| 111 |
+
def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
|
| 112 |
+
t_freq = self.timestep_embedding(timesteps, self.frequency_embedding_size)
|
| 113 |
+
weight_dtype = self.mlp[0].weight.dtype
|
| 114 |
+
return self.mlp(t_freq.to(dtype=weight_dtype))
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
class DiTMoELabelEmbedder(nn.Module):
|
| 118 |
+
def __init__(self, num_classes: int, hidden_size: int, dropout_prob: float):
|
| 119 |
+
super().__init__()
|
| 120 |
+
use_cfg_embedding = dropout_prob > 0
|
| 121 |
+
self.embedding_table = nn.Embedding(num_classes + int(use_cfg_embedding), hidden_size)
|
| 122 |
+
self.num_classes = num_classes
|
| 123 |
+
self.dropout_prob = dropout_prob
|
| 124 |
+
|
| 125 |
+
def token_drop(self, labels: torch.Tensor, force_drop_ids: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 126 |
+
if force_drop_ids is None:
|
| 127 |
+
drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
|
| 128 |
+
else:
|
| 129 |
+
drop_ids = force_drop_ids == 1
|
| 130 |
+
return torch.where(drop_ids, self.num_classes, labels)
|
| 131 |
+
|
| 132 |
+
def forward(
|
| 133 |
+
self,
|
| 134 |
+
labels: torch.Tensor,
|
| 135 |
+
train: bool = False,
|
| 136 |
+
force_drop_ids: Optional[torch.Tensor] = None,
|
| 137 |
+
) -> torch.Tensor:
|
| 138 |
+
use_dropout = self.dropout_prob > 0
|
| 139 |
+
if (train and use_dropout) or (force_drop_ids is not None):
|
| 140 |
+
labels = self.token_drop(labels, force_drop_ids)
|
| 141 |
+
return self.embedding_table(labels)
|
| 142 |
+
|
| 143 |
+
|
| 144 |
+
class MoEGate(nn.Module):
|
| 145 |
+
def __init__(self, embed_dim: int, num_experts: int = 16, num_experts_per_tok: int = 2, aux_loss_alpha: float = 0.01):
|
| 146 |
+
super().__init__()
|
| 147 |
+
self.top_k = num_experts_per_tok
|
| 148 |
+
self.n_routed_experts = num_experts
|
| 149 |
+
self.scoring_func = "softmax"
|
| 150 |
+
self.alpha = aux_loss_alpha
|
| 151 |
+
self.seq_aux = False
|
| 152 |
+
self.norm_topk_prob = False
|
| 153 |
+
self.gating_dim = embed_dim
|
| 154 |
+
self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim)))
|
| 155 |
+
self.reset_parameters()
|
| 156 |
+
|
| 157 |
+
def reset_parameters(self) -> None:
|
| 158 |
+
nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
|
| 159 |
+
|
| 160 |
+
def forward(self, hidden_states: torch.Tensor):
|
| 161 |
+
bsz, seq_len, hidden_dim = hidden_states.shape
|
| 162 |
+
hidden_states = hidden_states.view(-1, hidden_dim)
|
| 163 |
+
logits = F.linear(hidden_states, self.weight, None)
|
| 164 |
+
scores = logits.softmax(dim=-1)
|
| 165 |
+
topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False)
|
| 166 |
+
if self.top_k > 1 and self.norm_topk_prob:
|
| 167 |
+
topk_weight = topk_weight / (topk_weight.sum(dim=-1, keepdim=True) + 1e-20)
|
| 168 |
+
|
| 169 |
+
if self.training and self.alpha > 0.0:
|
| 170 |
+
topk_idx_for_aux_loss = topk_idx.view(bsz, -1)
|
| 171 |
+
mask_ce = F.one_hot(topk_idx_for_aux_loss.view(-1), num_classes=self.n_routed_experts)
|
| 172 |
+
ce = mask_ce.float().mean(0)
|
| 173 |
+
pi = scores.mean(0)
|
| 174 |
+
fi = ce * self.n_routed_experts
|
| 175 |
+
aux_loss = (pi * fi).sum() * self.alpha
|
| 176 |
+
else:
|
| 177 |
+
aux_loss = None
|
| 178 |
+
return topk_idx, topk_weight, aux_loss
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
class AddAuxiliaryLoss(torch.autograd.Function):
|
| 182 |
+
@staticmethod
|
| 183 |
+
def forward(ctx, x, loss):
|
| 184 |
+
assert loss.numel() == 1
|
| 185 |
+
ctx.dtype = loss.dtype
|
| 186 |
+
ctx.required_aux_loss = loss.requires_grad
|
| 187 |
+
return x
|
| 188 |
+
|
| 189 |
+
@staticmethod
|
| 190 |
+
def backward(ctx, grad_output):
|
| 191 |
+
grad_loss = None
|
| 192 |
+
if ctx.required_aux_loss:
|
| 193 |
+
grad_loss = torch.ones(1, dtype=ctx.dtype, device=grad_output.device)
|
| 194 |
+
return grad_output, grad_loss
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
class MoeMLP(nn.Module):
|
| 198 |
+
def __init__(self, hidden_size: int, intermediate_size: int, pretraining_tp: int = 2):
|
| 199 |
+
super().__init__()
|
| 200 |
+
self.hidden_size = hidden_size
|
| 201 |
+
self.intermediate_size = intermediate_size
|
| 202 |
+
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
| 203 |
+
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
| 204 |
+
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
| 205 |
+
self.act_fn = nn.SiLU()
|
| 206 |
+
self.pretraining_tp = pretraining_tp
|
| 207 |
+
|
| 208 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 209 |
+
if self.pretraining_tp > 1:
|
| 210 |
+
slice_size = self.intermediate_size // self.pretraining_tp
|
| 211 |
+
gate_proj_slices = self.gate_proj.weight.split(slice_size, dim=0)
|
| 212 |
+
up_proj_slices = self.up_proj.weight.split(slice_size, dim=0)
|
| 213 |
+
down_proj_slices = self.down_proj.weight.split(slice_size, dim=1)
|
| 214 |
+
gate_proj = torch.cat([F.linear(x, gate_proj_slices[i]) for i in range(self.pretraining_tp)], dim=-1)
|
| 215 |
+
up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.pretraining_tp)], dim=-1)
|
| 216 |
+
intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice_size, dim=-1)
|
| 217 |
+
down_proj = [F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.pretraining_tp)]
|
| 218 |
+
return sum(down_proj)
|
| 219 |
+
return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
| 220 |
+
|
| 221 |
+
|
| 222 |
+
class SparseMoeBlock(nn.Module):
|
| 223 |
+
def __init__(
|
| 224 |
+
self,
|
| 225 |
+
embed_dim: int,
|
| 226 |
+
mlp_ratio: int = 4,
|
| 227 |
+
num_experts: int = 16,
|
| 228 |
+
num_experts_per_tok: int = 2,
|
| 229 |
+
pretraining_tp: int = 2,
|
| 230 |
+
):
|
| 231 |
+
super().__init__()
|
| 232 |
+
self.num_experts_per_tok = num_experts_per_tok
|
| 233 |
+
self.experts = nn.ModuleList(
|
| 234 |
+
[
|
| 235 |
+
MoeMLP(
|
| 236 |
+
hidden_size=embed_dim,
|
| 237 |
+
intermediate_size=mlp_ratio * embed_dim,
|
| 238 |
+
pretraining_tp=pretraining_tp,
|
| 239 |
+
)
|
| 240 |
+
for _ in range(num_experts)
|
| 241 |
+
]
|
| 242 |
+
)
|
| 243 |
+
self.gate = MoEGate(embed_dim=embed_dim, num_experts=num_experts, num_experts_per_tok=num_experts_per_tok)
|
| 244 |
+
self.n_shared_experts = 2
|
| 245 |
+
if self.n_shared_experts is not None:
|
| 246 |
+
intermediate_size = embed_dim * self.n_shared_experts
|
| 247 |
+
self.shared_experts = MoeMLP(
|
| 248 |
+
hidden_size=embed_dim,
|
| 249 |
+
intermediate_size=intermediate_size,
|
| 250 |
+
pretraining_tp=pretraining_tp,
|
| 251 |
+
)
|
| 252 |
+
|
| 253 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 254 |
+
identity = hidden_states
|
| 255 |
+
orig_shape = hidden_states.shape
|
| 256 |
+
topk_idx, topk_weight, aux_loss = self.gate(hidden_states)
|
| 257 |
+
hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
|
| 258 |
+
flat_topk_idx = topk_idx.view(-1)
|
| 259 |
+
if self.training:
|
| 260 |
+
hidden_states = hidden_states.repeat_interleave(self.num_experts_per_tok, dim=0)
|
| 261 |
+
y = torch.empty_like(hidden_states, dtype=hidden_states.dtype)
|
| 262 |
+
for index, expert in enumerate(self.experts):
|
| 263 |
+
y[flat_topk_idx == index] = expert(hidden_states[flat_topk_idx == index]).float()
|
| 264 |
+
y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
|
| 265 |
+
y = y.view(*orig_shape)
|
| 266 |
+
y = AddAuxiliaryLoss.apply(y, aux_loss)
|
| 267 |
+
else:
|
| 268 |
+
y = self.moe_infer(hidden_states, flat_topk_idx, topk_weight.view(-1, 1)).view(*orig_shape)
|
| 269 |
+
if self.n_shared_experts is not None:
|
| 270 |
+
y = y + self.shared_experts(identity)
|
| 271 |
+
return y
|
| 272 |
+
|
| 273 |
+
@torch.no_grad()
|
| 274 |
+
def moe_infer(self, x: torch.Tensor, flat_expert_indices: torch.Tensor, flat_expert_weights: torch.Tensor):
|
| 275 |
+
expert_cache = torch.zeros_like(x)
|
| 276 |
+
idxs = flat_expert_indices.argsort()
|
| 277 |
+
tokens_per_expert = flat_expert_indices.bincount().cpu().numpy().cumsum(0)
|
| 278 |
+
token_idxs = idxs // self.num_experts_per_tok
|
| 279 |
+
for index, end_idx in enumerate(tokens_per_expert):
|
| 280 |
+
start_idx = 0 if index == 0 else tokens_per_expert[index - 1]
|
| 281 |
+
if start_idx == end_idx:
|
| 282 |
+
continue
|
| 283 |
+
expert = self.experts[index]
|
| 284 |
+
exp_token_idx = token_idxs[start_idx:end_idx]
|
| 285 |
+
expert_tokens = x[exp_token_idx]
|
| 286 |
+
expert_out = expert(expert_tokens)
|
| 287 |
+
expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])
|
| 288 |
+
expert_cache = expert_cache.to(expert_out.dtype)
|
| 289 |
+
expert_cache.scatter_reduce_(
|
| 290 |
+
0,
|
| 291 |
+
exp_token_idx.view(-1, 1).repeat(1, x.shape[-1]),
|
| 292 |
+
expert_out,
|
| 293 |
+
reduce="sum",
|
| 294 |
+
)
|
| 295 |
+
return expert_cache
|
| 296 |
+
|
| 297 |
+
|
| 298 |
+
class FlashSelfMHAModified(nn.Module):
|
| 299 |
+
def __init__(self, dim: int, num_heads: int, qkv_bias: bool = True, qk_norm: bool = False):
|
| 300 |
+
super().__init__()
|
| 301 |
+
if FlashSelfAttention is None:
|
| 302 |
+
raise ImportError("flash_attn is required when use_flash_attn=True.")
|
| 303 |
+
self.dim = dim
|
| 304 |
+
self.num_heads = num_heads
|
| 305 |
+
self.head_dim = dim // num_heads
|
| 306 |
+
self.Wqkv = nn.Linear(dim, 3 * dim, bias=qkv_bias)
|
| 307 |
+
self.q_norm = nn.LayerNorm(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
|
| 308 |
+
self.k_norm = nn.LayerNorm(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
|
| 309 |
+
self.inner_attn = FlashSelfAttention(attention_dropout=0.0)
|
| 310 |
+
self.out_proj = nn.Linear(dim, dim, bias=qkv_bias)
|
| 311 |
+
self.proj_drop = nn.Dropout(0.0)
|
| 312 |
+
|
| 313 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 314 |
+
batch_size, seq_len, dim = x.shape
|
| 315 |
+
qkv = self.Wqkv(x).view(batch_size, seq_len, 3, self.num_heads, self.head_dim)
|
| 316 |
+
query, key, value = qkv.unbind(dim=2)
|
| 317 |
+
query = self.q_norm(query).to(dtype=torch.float16)
|
| 318 |
+
key = self.k_norm(key).to(dtype=torch.float16)
|
| 319 |
+
qkv = torch.stack([query, key, value], dim=2)
|
| 320 |
+
context = self.inner_attn(qkv)
|
| 321 |
+
out = self.out_proj(context.view(batch_size, seq_len, dim))
|
| 322 |
+
return self.proj_drop(out)
|
| 323 |
+
|
| 324 |
+
|
| 325 |
+
class DiTMoEBlock(nn.Module):
|
| 326 |
+
def __init__(
|
| 327 |
+
self,
|
| 328 |
+
hidden_size: int,
|
| 329 |
+
num_heads: int,
|
| 330 |
+
mlp_ratio: float = 4.0,
|
| 331 |
+
num_experts: int = 8,
|
| 332 |
+
num_experts_per_tok: int = 2,
|
| 333 |
+
pretraining_tp: int = 2,
|
| 334 |
+
use_flash_attn: bool = False,
|
| 335 |
+
):
|
| 336 |
+
super().__init__()
|
| 337 |
+
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
| 338 |
+
if use_flash_attn:
|
| 339 |
+
self.attn = FlashSelfMHAModified(hidden_size, num_heads=num_heads, qkv_bias=True, qk_norm=True)
|
| 340 |
+
else:
|
| 341 |
+
if Attention is None:
|
| 342 |
+
raise ImportError("timm is required when use_flash_attn=False.")
|
| 343 |
+
self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True)
|
| 344 |
+
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
| 345 |
+
self.moe = SparseMoeBlock(hidden_size, int(mlp_ratio), num_experts, num_experts_per_tok, pretraining_tp)
|
| 346 |
+
self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True))
|
| 347 |
+
|
| 348 |
+
def forward(self, hidden_states: torch.Tensor, conditioning: torch.Tensor) -> torch.Tensor:
|
| 349 |
+
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(conditioning).chunk(
|
| 350 |
+
6, dim=1
|
| 351 |
+
)
|
| 352 |
+
hidden_states = hidden_states + gate_msa.unsqueeze(1) * self.attn(
|
| 353 |
+
modulate(self.norm1(hidden_states), shift_msa, scale_msa)
|
| 354 |
+
)
|
| 355 |
+
hidden_states = hidden_states + gate_mlp.unsqueeze(1) * self.moe(
|
| 356 |
+
modulate(self.norm2(hidden_states), shift_mlp, scale_mlp)
|
| 357 |
+
)
|
| 358 |
+
return hidden_states
|
| 359 |
+
|
| 360 |
+
|
| 361 |
+
class DiTMoEFinalLayer(nn.Module):
|
| 362 |
+
def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
|
| 363 |
+
super().__init__()
|
| 364 |
+
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
|
| 365 |
+
self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
|
| 366 |
+
self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
|
| 367 |
+
|
| 368 |
+
def forward(self, hidden_states: torch.Tensor, conditioning: torch.Tensor) -> torch.Tensor:
|
| 369 |
+
shift, scale = self.adaLN_modulation(conditioning).chunk(2, dim=1)
|
| 370 |
+
hidden_states = modulate(self.norm_final(hidden_states), shift, scale)
|
| 371 |
+
return self.linear(hidden_states)
|
| 372 |
+
|
| 373 |
+
|
| 374 |
+
def get_2d_sincos_pos_embed(embed_dim: int, grid_size: int, cls_token: bool = False, extra_tokens: int = 0):
|
| 375 |
+
grid_h = np.arange(grid_size, dtype=np.float32)
|
| 376 |
+
grid_w = np.arange(grid_size, dtype=np.float32)
|
| 377 |
+
grid = np.meshgrid(grid_w, grid_h)
|
| 378 |
+
grid = np.stack(grid, axis=0).reshape([2, 1, grid_size, grid_size])
|
| 379 |
+
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
|
| 380 |
+
if cls_token and extra_tokens > 0:
|
| 381 |
+
pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
|
| 382 |
+
return pos_embed
|
| 383 |
+
|
| 384 |
+
|
| 385 |
+
def get_2d_sincos_pos_embed_from_grid(embed_dim: int, grid: np.ndarray):
|
| 386 |
+
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])
|
| 387 |
+
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])
|
| 388 |
+
return np.concatenate([emb_h, emb_w], axis=1)
|
| 389 |
+
|
| 390 |
+
|
| 391 |
+
def get_1d_sincos_pos_embed_from_grid(embed_dim: int, pos: np.ndarray):
|
| 392 |
+
omega = np.arange(embed_dim // 2, dtype=np.float64)
|
| 393 |
+
omega /= embed_dim / 2.0
|
| 394 |
+
omega = 1.0 / 10000**omega
|
| 395 |
+
pos = pos.reshape(-1)
|
| 396 |
+
out = np.einsum("m,d->md", pos, omega)
|
| 397 |
+
return np.concatenate([np.sin(out), np.cos(out)], axis=1)
|
| 398 |
+
|
| 399 |
+
|
| 400 |
+
class DiTMoETransformer2DModel(ModelMixin, ConfigMixin):
|
| 401 |
+
config_name = "config.json"
|
| 402 |
+
|
| 403 |
+
@register_to_config
|
| 404 |
+
def __init__(
|
| 405 |
+
self,
|
| 406 |
+
input_size: int = 32,
|
| 407 |
+
patch_size: int = 2,
|
| 408 |
+
in_channels: int = 4,
|
| 409 |
+
hidden_size: int = 1152,
|
| 410 |
+
depth: int = 28,
|
| 411 |
+
num_heads: int = 16,
|
| 412 |
+
mlp_ratio: float = 4.0,
|
| 413 |
+
class_dropout_prob: float = 0.1,
|
| 414 |
+
num_classes: int = 1000,
|
| 415 |
+
num_experts: int = 8,
|
| 416 |
+
num_experts_per_tok: int = 2,
|
| 417 |
+
pretraining_tp: int = 2,
|
| 418 |
+
learn_sigma: bool = True,
|
| 419 |
+
use_flash_attn: bool = False,
|
| 420 |
+
):
|
| 421 |
+
super().__init__()
|
| 422 |
+
self.learn_sigma = learn_sigma
|
| 423 |
+
self.in_channels = in_channels
|
| 424 |
+
self.out_channels = in_channels * 2 if learn_sigma else in_channels
|
| 425 |
+
self.patch_size = patch_size
|
| 426 |
+
self.num_heads = num_heads
|
| 427 |
+
self.num_classes = num_classes
|
| 428 |
+
|
| 429 |
+
self.x_embedder = DiTMoEPatchEmbed(input_size, patch_size, in_channels, hidden_size, bias=True)
|
| 430 |
+
self.t_embedder = DiTMoETimestepEmbedder(hidden_size)
|
| 431 |
+
self.y_embedder = DiTMoELabelEmbedder(num_classes, hidden_size, class_dropout_prob)
|
| 432 |
+
self.pos_embed = nn.Parameter(torch.zeros(1, self.x_embedder.num_patches, hidden_size), requires_grad=False)
|
| 433 |
+
self.blocks = nn.ModuleList(
|
| 434 |
+
[
|
| 435 |
+
DiTMoEBlock(
|
| 436 |
+
hidden_size,
|
| 437 |
+
num_heads,
|
| 438 |
+
mlp_ratio=mlp_ratio,
|
| 439 |
+
num_experts=num_experts,
|
| 440 |
+
num_experts_per_tok=num_experts_per_tok,
|
| 441 |
+
pretraining_tp=pretraining_tp,
|
| 442 |
+
use_flash_attn=use_flash_attn,
|
| 443 |
+
)
|
| 444 |
+
for _ in range(depth)
|
| 445 |
+
]
|
| 446 |
+
)
|
| 447 |
+
self.final_layer = DiTMoEFinalLayer(hidden_size, patch_size, self.out_channels)
|
| 448 |
+
self.initialize_weights()
|
| 449 |
+
|
| 450 |
+
def initialize_weights(self):
|
| 451 |
+
def _basic_init(module):
|
| 452 |
+
if isinstance(module, nn.Linear):
|
| 453 |
+
nn.init.xavier_uniform_(module.weight)
|
| 454 |
+
if module.bias is not None:
|
| 455 |
+
nn.init.constant_(module.bias, 0)
|
| 456 |
+
|
| 457 |
+
self.apply(_basic_init)
|
| 458 |
+
pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches**0.5))
|
| 459 |
+
self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
|
| 460 |
+
w = self.x_embedder.proj.weight.data
|
| 461 |
+
nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
|
| 462 |
+
nn.init.constant_(self.x_embedder.proj.bias, 0)
|
| 463 |
+
nn.init.normal_(self.y_embedder.embedding_table.weight, std=0.02)
|
| 464 |
+
nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
|
| 465 |
+
nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
|
| 466 |
+
for block in self.blocks:
|
| 467 |
+
nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
|
| 468 |
+
nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
|
| 469 |
+
nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
|
| 470 |
+
nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
|
| 471 |
+
nn.init.constant_(self.final_layer.linear.weight, 0)
|
| 472 |
+
nn.init.constant_(self.final_layer.linear.bias, 0)
|
| 473 |
+
|
| 474 |
+
def unpatchify(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 475 |
+
channels = self.out_channels
|
| 476 |
+
patch_size = self.patch_size
|
| 477 |
+
height = width = int(hidden_states.shape[1] ** 0.5)
|
| 478 |
+
hidden_states = hidden_states.reshape(hidden_states.shape[0], height, width, patch_size, patch_size, channels)
|
| 479 |
+
hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
|
| 480 |
+
return hidden_states.reshape(hidden_states.shape[0], channels, height * patch_size, width * patch_size)
|
| 481 |
+
|
| 482 |
+
def forward(
|
| 483 |
+
self,
|
| 484 |
+
hidden_states: torch.Tensor,
|
| 485 |
+
timestep: Union[torch.Tensor, float, int],
|
| 486 |
+
class_labels: torch.LongTensor,
|
| 487 |
+
return_dict: bool = True,
|
| 488 |
+
) -> Union[DiTMoETransformer2DModelOutput, torch.Tensor]:
|
| 489 |
+
if not torch.is_tensor(timestep):
|
| 490 |
+
timestep = torch.tensor([timestep], device=hidden_states.device, dtype=hidden_states.dtype)
|
| 491 |
+
timestep = timestep.to(device=hidden_states.device, dtype=torch.float32)
|
| 492 |
+
if timestep.ndim == 0:
|
| 493 |
+
timestep = timestep.unsqueeze(0)
|
| 494 |
+
if timestep.shape[0] == 1 and hidden_states.shape[0] > 1:
|
| 495 |
+
timestep = timestep.expand(hidden_states.shape[0])
|
| 496 |
+
|
| 497 |
+
class_labels = class_labels.to(device=hidden_states.device, dtype=torch.long).reshape(-1)
|
| 498 |
+
hidden_states = self.x_embedder(hidden_states) + self.pos_embed
|
| 499 |
+
conditioning = self.t_embedder(timestep) + self.y_embedder(class_labels, train=self.training)
|
| 500 |
+
for block in self.blocks:
|
| 501 |
+
hidden_states = block(hidden_states, conditioning)
|
| 502 |
+
hidden_states = self.final_layer(hidden_states, conditioning)
|
| 503 |
+
hidden_states = self.unpatchify(hidden_states)
|
| 504 |
+
if not return_dict:
|
| 505 |
+
return hidden_states
|
| 506 |
+
return DiTMoETransformer2DModelOutput(sample=hidden_states)
|
| 507 |
+
|
| 508 |
+
def split_velocity(self, sample: torch.Tensor) -> torch.Tensor:
|
| 509 |
+
if self.learn_sigma:
|
| 510 |
+
sample, _ = sample.chunk(2, dim=1)
|
| 511 |
+
return sample
|
DiT-MoE-B-8E2A/vae/config.json
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_class_name": "AutoencoderKL",
|
| 3 |
+
"_diffusers_version": "0.4.2",
|
| 4 |
+
"act_fn": "silu",
|
| 5 |
+
"block_out_channels": [
|
| 6 |
+
128,
|
| 7 |
+
256,
|
| 8 |
+
512,
|
| 9 |
+
512
|
| 10 |
+
],
|
| 11 |
+
"down_block_types": [
|
| 12 |
+
"DownEncoderBlock2D",
|
| 13 |
+
"DownEncoderBlock2D",
|
| 14 |
+
"DownEncoderBlock2D",
|
| 15 |
+
"DownEncoderBlock2D"
|
| 16 |
+
],
|
| 17 |
+
"in_channels": 3,
|
| 18 |
+
"latent_channels": 4,
|
| 19 |
+
"layers_per_block": 2,
|
| 20 |
+
"norm_num_groups": 32,
|
| 21 |
+
"out_channels": 3,
|
| 22 |
+
"sample_size": 256,
|
| 23 |
+
"up_block_types": [
|
| 24 |
+
"UpDecoderBlock2D",
|
| 25 |
+
"UpDecoderBlock2D",
|
| 26 |
+
"UpDecoderBlock2D",
|
| 27 |
+
"UpDecoderBlock2D"
|
| 28 |
+
]
|
| 29 |
+
}
|
DiT-MoE-B-8E2A/vae/diffusion_pytorch_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f3944a3bb4da8058dd3d48327d8da8d875d5ea780833ade1b095bc4049536bac
|
| 3 |
+
size 334643252
|
DiT-MoE-S-8E2A/README.md
CHANGED
|
@@ -39,6 +39,7 @@ Each variant keeps an English `id2label` map in `model_index.json` (DiT-style).
|
|
| 39 |
| Sampler | DDIM |
|
| 40 |
| Steps | 50 |
|
| 41 |
| CFG scale | 4.0 |
|
|
|
|
| 42 |
| VAE | `stabilityai/sd-vae-ft-mse` (bundled under `vae/`) |
|
| 43 |
|
| 44 |
```python
|
|
@@ -52,7 +53,7 @@ pipe = DiffusionPipeline.from_pretrained(
|
|
| 52 |
local_files_only=True,
|
| 53 |
custom_pipeline=str(model_dir / "pipeline.py"),
|
| 54 |
trust_remote_code=True,
|
| 55 |
-
torch_dtype=torch.
|
| 56 |
)
|
| 57 |
pipe.to("cuda")
|
| 58 |
|
|
|
|
| 39 |
| Sampler | DDIM |
|
| 40 |
| Steps | 50 |
|
| 41 |
| CFG scale | 4.0 |
|
| 42 |
+
| Dtype | `bfloat16` (recommended on Ampere+) |
|
| 43 |
| VAE | `stabilityai/sd-vae-ft-mse` (bundled under `vae/`) |
|
| 44 |
|
| 45 |
```python
|
|
|
|
| 53 |
local_files_only=True,
|
| 54 |
custom_pipeline=str(model_dir / "pipeline.py"),
|
| 55 |
trust_remote_code=True,
|
| 56 |
+
torch_dtype=torch.bfloat16,
|
| 57 |
)
|
| 58 |
pipe.to("cuda")
|
| 59 |
|
DiT-MoE-S-8E2A/pipeline.py
CHANGED
|
@@ -42,7 +42,7 @@ EXAMPLE_DOC_STRING = """
|
|
| 42 |
... local_files_only=True,
|
| 43 |
... custom_pipeline=str(model_dir / "pipeline.py"),
|
| 44 |
... trust_remote_code=True,
|
| 45 |
-
... torch_dtype=torch.
|
| 46 |
... )
|
| 47 |
>>> pipe.to("cuda")
|
| 48 |
|
|
|
|
| 42 |
... local_files_only=True,
|
| 43 |
... custom_pipeline=str(model_dir / "pipeline.py"),
|
| 44 |
... trust_remote_code=True,
|
| 45 |
+
... torch_dtype=torch.bfloat16,
|
| 46 |
... )
|
| 47 |
>>> pipe.to("cuda")
|
| 48 |
|
README.md
CHANGED
|
@@ -34,7 +34,7 @@ Each variant keeps an English `id2label` map directly in its own `model_index.js
|
|
| 34 |
| Checkpoint | Path | Resolution | Sampler |
|
| 35 |
| --- | --- | --- | --- |
|
| 36 |
| DiT-MoE-S/2-8E2A | `./DiT-MoE-S-8E2A` | 256×256 | DDIM |
|
| 37 |
-
| DiT-MoE-B/2-8E2A |
|
| 38 |
| DiT-MoE-XL/2-8E2A | *(convert with `--rectified-flow`)* | 256×256 | RF |
|
| 39 |
| DiT-MoE-G/2-16E2A | *(convert with `--rectified-flow --num-experts 16`)* | 256×256 | RF |
|
| 40 |
|
|
@@ -56,6 +56,8 @@ python scripts/convert_dit_moe_to_diffusers.py \
|
|
| 56 |
|
| 57 |
## Inference
|
| 58 |
|
|
|
|
|
|
|
| 59 |
```python
|
| 60 |
from pathlib import Path
|
| 61 |
import torch
|
|
@@ -67,7 +69,7 @@ pipe = DiffusionPipeline.from_pretrained(
|
|
| 67 |
local_files_only=True,
|
| 68 |
custom_pipeline=str(model_dir / "pipeline.py"),
|
| 69 |
trust_remote_code=True,
|
| 70 |
-
torch_dtype=torch.
|
| 71 |
)
|
| 72 |
pipe.to("cuda")
|
| 73 |
|
|
|
|
| 34 |
| Checkpoint | Path | Resolution | Sampler |
|
| 35 |
| --- | --- | --- | --- |
|
| 36 |
| DiT-MoE-S/2-8E2A | `./DiT-MoE-S-8E2A` | 256×256 | DDIM |
|
| 37 |
+
| DiT-MoE-B/2-8E2A | `./DiT-MoE-B-8E2A` | 256×256 | DDIM |
|
| 38 |
| DiT-MoE-XL/2-8E2A | *(convert with `--rectified-flow`)* | 256×256 | RF |
|
| 39 |
| DiT-MoE-G/2-16E2A | *(convert with `--rectified-flow --num-experts 16`)* | 256×256 | RF |
|
| 40 |
|
|
|
|
| 56 |
|
| 57 |
## Inference
|
| 58 |
|
| 59 |
+
Use `torch.bfloat16` on Ampere+ GPUs (default in examples and `sample_dit_moe.py`).
|
| 60 |
+
|
| 61 |
```python
|
| 62 |
from pathlib import Path
|
| 63 |
import torch
|
|
|
|
| 69 |
local_files_only=True,
|
| 70 |
custom_pipeline=str(model_dir / "pipeline.py"),
|
| 71 |
trust_remote_code=True,
|
| 72 |
+
torch_dtype=torch.bfloat16,
|
| 73 |
)
|
| 74 |
pipe.to("cuda")
|
| 75 |
|