Update README.md
Browse files
README.md
CHANGED
@@ -21,4 +21,118 @@ image = pipe(
|
|
21 |
guidance_scale=7.0,
|
22 |
).images[0]
|
23 |
image
|
24 |
-
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
guidance_scale=7.0,
|
22 |
).images[0]
|
23 |
image
|
24 |
+
```
|
25 |
+
|
26 |
+
## Codes
|
27 |
+
```python
|
28 |
+
import importlib
|
29 |
+
|
30 |
+
import torch
|
31 |
+
import transformers
|
32 |
+
|
33 |
+
import diffusers
|
34 |
+
import rich
|
35 |
+
|
36 |
+
|
37 |
+
def get_original_model_configs(pipeline_cls: type[diffusers.DiffusionPipeline], pipeline_id: str):
|
38 |
+
pipeline_config: dict[str, list[str]] = pipeline_cls.load_config(pipeline_id)
|
39 |
+
model_configs = {}
|
40 |
+
|
41 |
+
for subfolder, import_strings in pipeline_config.items():
|
42 |
+
if subfolder.startswith("_"):
|
43 |
+
continue
|
44 |
+
module = importlib.import_module(".".join(import_strings[:-1]))
|
45 |
+
cls = getattr(module, import_strings[-1])
|
46 |
+
if issubclass(cls, transformers.PreTrainedModel):
|
47 |
+
config_class: transformers.PretrainedConfig = cls.config_class
|
48 |
+
config = config_class.from_pretrained(pipeline_id, subfolder=subfolder)
|
49 |
+
model_configs[subfolder] = config
|
50 |
+
elif issubclass(cls, diffusers.ModelMixin) and issubclass(cls, diffusers.ConfigMixin):
|
51 |
+
config = cls.load_config(pipeline_id, subfolder=subfolder)
|
52 |
+
model_configs[subfolder] = config
|
53 |
+
|
54 |
+
return model_configs
|
55 |
+
|
56 |
+
|
57 |
+
def load_pipeline(pipeline_cls: type[diffusers.DiffusionPipeline], pipeline_id: str, model_configs: dict[str, dict]):
|
58 |
+
pipeline_config: dict[str, list[str]] = pipeline_cls.load_config(pipeline_id)
|
59 |
+
components = {}
|
60 |
+
for subfolder, import_strings in pipeline_config.items():
|
61 |
+
if subfolder.startswith("_"):
|
62 |
+
continue
|
63 |
+
module = importlib.import_module(".".join(import_strings[:-1]))
|
64 |
+
cls = getattr(module, import_strings[-1])
|
65 |
+
print(f"Loading:", ".".join(import_strings))
|
66 |
+
if issubclass(cls, transformers.PreTrainedModel):
|
67 |
+
config = model_configs[subfolder]
|
68 |
+
component = cls(config)
|
69 |
+
elif issubclass(cls, transformers.PreTrainedTokenizerBase):
|
70 |
+
component = cls.from_pretrained(pipeline_id, subfolder=subfolder)
|
71 |
+
elif issubclass(cls, diffusers.ModelMixin) and issubclass(cls, diffusers.ConfigMixin):
|
72 |
+
config = model_configs[subfolder]
|
73 |
+
component = cls.from_config(config)
|
74 |
+
elif issubclass(cls, diffusers.SchedulerMixin) and issubclass(cls, diffusers.ConfigMixin):
|
75 |
+
component = cls.from_pretrained(pipeline_id, subfolder=subfolder)
|
76 |
+
else:
|
77 |
+
raise (f"unknown {subfolder}: {import_strings}")
|
78 |
+
components[subfolder] = component
|
79 |
+
pipeline = pipeline_cls(**components)
|
80 |
+
return pipeline
|
81 |
+
|
82 |
+
|
83 |
+
def get_pipeline():
|
84 |
+
torch.manual_seed(42)
|
85 |
+
pipeline_id = "stabilityai/stable-diffusion-3-medium-diffusers"
|
86 |
+
pipeline_cls = diffusers.StableDiffusion3Pipeline
|
87 |
+
model_configs = get_original_model_configs(pipeline_cls, pipeline_id)
|
88 |
+
rich.print(model_configs)
|
89 |
+
|
90 |
+
HIDDEN_SIZE = 8
|
91 |
+
|
92 |
+
model_configs["text_encoder"].hidden_size = HIDDEN_SIZE
|
93 |
+
model_configs["text_encoder"].intermediate_size = HIDDEN_SIZE * 2
|
94 |
+
model_configs["text_encoder"].num_attention_heads = 2
|
95 |
+
model_configs["text_encoder"].num_hidden_layers = 2
|
96 |
+
model_configs["text_encoder"].projection_dim = HIDDEN_SIZE
|
97 |
+
|
98 |
+
model_configs["text_encoder_2"].hidden_size = HIDDEN_SIZE
|
99 |
+
model_configs["text_encoder_2"].intermediate_size = HIDDEN_SIZE * 2
|
100 |
+
model_configs["text_encoder_2"].num_attention_heads = 2
|
101 |
+
model_configs["text_encoder_2"].num_hidden_layers = 2
|
102 |
+
model_configs["text_encoder_2"].projection_dim = HIDDEN_SIZE
|
103 |
+
|
104 |
+
model_configs["text_encoder_3"].d_model = HIDDEN_SIZE
|
105 |
+
model_configs["text_encoder_3"].d_ff = HIDDEN_SIZE * 2
|
106 |
+
model_configs["text_encoder_3"].d_kv = HIDDEN_SIZE // 2
|
107 |
+
model_configs["text_encoder_3"].num_heads = 2
|
108 |
+
model_configs["text_encoder_3"].num_layers = 2
|
109 |
+
|
110 |
+
model_configs["transformer"]["num_layers"] = 2
|
111 |
+
model_configs["transformer"]["num_attention_heads"] = 2
|
112 |
+
model_configs["transformer"]["attention_head_dim"] = HIDDEN_SIZE // 2
|
113 |
+
model_configs["transformer"]["pooled_projection_dim"] = HIDDEN_SIZE * 2
|
114 |
+
model_configs["transformer"]["joint_attention_dim"] = HIDDEN_SIZE
|
115 |
+
model_configs["transformer"]["caption_projection_dim"] = HIDDEN_SIZE
|
116 |
+
|
117 |
+
model_configs["vae"]["layers_per_block"] = 1
|
118 |
+
model_configs["vae"]["block_out_channels"] = [HIDDEN_SIZE] * 4
|
119 |
+
model_configs["vae"]["norm_num_groups"] = 2
|
120 |
+
model_configs["vae"]["latent_channels"] = 16
|
121 |
+
|
122 |
+
pipeline = load_pipeline(pipeline_cls, pipeline_id, model_configs)
|
123 |
+
return pipeline
|
124 |
+
|
125 |
+
|
126 |
+
pipeline = get_pipeline()
|
127 |
+
image = pipeline(
|
128 |
+
"hello world",
|
129 |
+
negative_prompt="runtime error",
|
130 |
+
num_inference_steps=2,
|
131 |
+
guidance_scale=7.0,
|
132 |
+
).images[0]
|
133 |
+
|
134 |
+
|
135 |
+
pipeline = pipeline.to(torch.float16)
|
136 |
+
pipeline.save_pretrained("/tmp/stable-diffusion-3-tiny-random")
|
137 |
+
pipeline.push_to_hub("yujiepan/stable-diffusion-3-tiny-random")
|
138 |
+
```
|