Spaces:
Build error
Build error
Support training text encoder
Browse files- app.py +16 -1
- inference.py +20 -0
- lora +1 -1
- trainer.py +16 -4
app.py
CHANGED
@@ -83,6 +83,10 @@ def create_training_demo(trainer: Trainer,
|
|
83 |
num_training_steps = gr.Number(
|
84 |
label='Number of Training Steps', value=1000, precision=0)
|
85 |
learning_rate = gr.Number(label='Learning Rate', value=0.0001)
|
|
|
|
|
|
|
|
|
86 |
gradient_accumulation = gr.Number(
|
87 |
label='Number of Gradient Accumulation',
|
88 |
value=1,
|
@@ -113,6 +117,8 @@ def create_training_demo(trainer: Trainer,
|
|
113 |
concept_prompt,
|
114 |
num_training_steps,
|
115 |
learning_rate,
|
|
|
|
|
116 |
gradient_accumulation,
|
117 |
fp16,
|
118 |
use_8bit_adam,
|
@@ -136,6 +142,7 @@ def create_training_demo(trainer: Trainer,
|
|
136 |
def find_weight_files() -> list[str]:
|
137 |
curr_dir = pathlib.Path(__file__).parent
|
138 |
paths = sorted(curr_dir.rglob('*.pt'))
|
|
|
139 |
return [path.relative_to(curr_dir).as_posix() for path in paths]
|
140 |
|
141 |
|
@@ -165,6 +172,11 @@ def create_inference_demo(pipe: InferencePipeline) -> gr.Blocks:
|
|
165 |
maximum=2,
|
166 |
step=0.05,
|
167 |
value=1)
|
|
|
|
|
|
|
|
|
|
|
168 |
seed = gr.Slider(label='Seed',
|
169 |
minimum=0,
|
170 |
maximum=100000,
|
@@ -185,7 +197,8 @@ def create_inference_demo(pipe: InferencePipeline) -> gr.Blocks:
|
|
185 |
run_button = gr.Button('Generate')
|
186 |
|
187 |
gr.Markdown('''
|
188 |
-
- The pretrained models are trained with the concept prompt "style of sks".
|
|
|
189 |
''')
|
190 |
with gr.Column():
|
191 |
result = gr.Image(label='Result')
|
@@ -199,6 +212,7 @@ def create_inference_demo(pipe: InferencePipeline) -> gr.Blocks:
|
|
199 |
lora_weight_name,
|
200 |
prompt,
|
201 |
alpha,
|
|
|
202 |
seed,
|
203 |
num_steps,
|
204 |
guidance_scale,
|
@@ -211,6 +225,7 @@ def create_inference_demo(pipe: InferencePipeline) -> gr.Blocks:
|
|
211 |
lora_weight_name,
|
212 |
prompt,
|
213 |
alpha,
|
|
|
214 |
seed,
|
215 |
num_steps,
|
216 |
guidance_scale,
|
|
|
83 |
num_training_steps = gr.Number(
|
84 |
label='Number of Training Steps', value=1000, precision=0)
|
85 |
learning_rate = gr.Number(label='Learning Rate', value=0.0001)
|
86 |
+
train_text_encoder = gr.Checkbox(label='Train Text Encoder',
|
87 |
+
value=False)
|
88 |
+
learning_rate_text = gr.Number(
|
89 |
+
label='Learning Rate for Text Encoder', value=0.00005)
|
90 |
gradient_accumulation = gr.Number(
|
91 |
label='Number of Gradient Accumulation',
|
92 |
value=1,
|
|
|
117 |
concept_prompt,
|
118 |
num_training_steps,
|
119 |
learning_rate,
|
120 |
+
train_text_encoder,
|
121 |
+
learning_rate_text,
|
122 |
gradient_accumulation,
|
123 |
fp16,
|
124 |
use_8bit_adam,
|
|
|
142 |
def find_weight_files() -> list[str]:
|
143 |
curr_dir = pathlib.Path(__file__).parent
|
144 |
paths = sorted(curr_dir.rglob('*.pt'))
|
145 |
+
paths = [path for path in paths if not path.stem.endswith('.text_encoder')]
|
146 |
return [path.relative_to(curr_dir).as_posix() for path in paths]
|
147 |
|
148 |
|
|
|
172 |
maximum=2,
|
173 |
step=0.05,
|
174 |
value=1)
|
175 |
+
alpha_for_text = gr.Slider(label='Alpha for Text Encoder',
|
176 |
+
minimum=0,
|
177 |
+
maximum=2,
|
178 |
+
step=0.05,
|
179 |
+
value=1)
|
180 |
seed = gr.Slider(label='Seed',
|
181 |
minimum=0,
|
182 |
maximum=100000,
|
|
|
197 |
run_button = gr.Button('Generate')
|
198 |
|
199 |
gr.Markdown('''
|
200 |
+
- The pretrained models for "disney", "illust" and "pop" are trained with the concept prompt "style of sks".
|
201 |
+
- The pretrained model for "kiriko" is trained with the concept prompt "game character bnha". For this model, the text encoder is also trained.
|
202 |
''')
|
203 |
with gr.Column():
|
204 |
result = gr.Image(label='Result')
|
|
|
212 |
lora_weight_name,
|
213 |
prompt,
|
214 |
alpha,
|
215 |
+
alpha_for_text,
|
216 |
seed,
|
217 |
num_steps,
|
218 |
guidance_scale,
|
|
|
225 |
lora_weight_name,
|
226 |
prompt,
|
227 |
alpha,
|
228 |
+
alpha_for_text,
|
229 |
seed,
|
230 |
num_steps,
|
231 |
guidance_scale,
|
inference.py
CHANGED
@@ -32,6 +32,14 @@ class InferencePipeline:
|
|
32 |
curr_dir = pathlib.Path(__file__).parent
|
33 |
return curr_dir / name
|
34 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
def load_pipe(self, model_id: str, lora_filename: str) -> None:
|
36 |
weight_path = self.get_lora_weight_path(lora_filename)
|
37 |
if weight_path == self.weight_path:
|
@@ -47,6 +55,16 @@ class InferencePipeline:
|
|
47 |
pipe = pipe.to(self.device)
|
48 |
|
49 |
monkeypatch_lora(pipe.unet, lora_weight)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
50 |
self.pipe = pipe
|
51 |
|
52 |
def run(
|
@@ -55,6 +73,7 @@ class InferencePipeline:
|
|
55 |
lora_weight_name: str,
|
56 |
prompt: str,
|
57 |
alpha: float,
|
|
|
58 |
seed: int,
|
59 |
n_steps: int,
|
60 |
guidance_scale: float,
|
@@ -66,6 +85,7 @@ class InferencePipeline:
|
|
66 |
|
67 |
generator = torch.Generator(device=self.device).manual_seed(seed)
|
68 |
tune_lora_scale(self.pipe.unet, alpha) # type: ignore
|
|
|
69 |
out = self.pipe(prompt,
|
70 |
num_inference_steps=n_steps,
|
71 |
guidance_scale=guidance_scale,
|
|
|
32 |
curr_dir = pathlib.Path(__file__).parent
|
33 |
return curr_dir / name
|
34 |
|
35 |
+
@staticmethod
|
36 |
+
def get_lora_text_encoder_weight_path(path: pathlib.Path) -> str:
|
37 |
+
parent_dir = path.parent
|
38 |
+
stem = path.stem
|
39 |
+
text_encoder_filename = f'{stem}.text_encoder.pt'
|
40 |
+
path = parent_dir / text_encoder_filename
|
41 |
+
return path.as_posix() if path.exists() else ''
|
42 |
+
|
43 |
def load_pipe(self, model_id: str, lora_filename: str) -> None:
|
44 |
weight_path = self.get_lora_weight_path(lora_filename)
|
45 |
if weight_path == self.weight_path:
|
|
|
55 |
pipe = pipe.to(self.device)
|
56 |
|
57 |
monkeypatch_lora(pipe.unet, lora_weight)
|
58 |
+
|
59 |
+
lora_text_encoder_weight_path = self.get_lora_text_encoder_weight_path(
|
60 |
+
weight_path)
|
61 |
+
if lora_text_encoder_weight_path:
|
62 |
+
lora_text_encoder_weight = torch.load(
|
63 |
+
lora_text_encoder_weight_path, map_location=self.device)
|
64 |
+
monkeypatch_lora(pipe.text_encoder,
|
65 |
+
lora_text_encoder_weight,
|
66 |
+
target_replace_module=['CLIPAttention'])
|
67 |
+
|
68 |
self.pipe = pipe
|
69 |
|
70 |
def run(
|
|
|
73 |
lora_weight_name: str,
|
74 |
prompt: str,
|
75 |
alpha: float,
|
76 |
+
alpha_for_text: float,
|
77 |
seed: int,
|
78 |
n_steps: int,
|
79 |
guidance_scale: float,
|
|
|
85 |
|
86 |
generator = torch.Generator(device=self.device).manual_seed(seed)
|
87 |
tune_lora_scale(self.pipe.unet, alpha) # type: ignore
|
88 |
+
tune_lora_scale(self.pipe.text_encoder, alpha_for_text) # type: ignore
|
89 |
out = self.pipe(prompt,
|
90 |
num_inference_steps=n_steps,
|
91 |
guidance_scale=guidance_scale,
|
lora
CHANGED
@@ -1 +1 @@
|
|
1 |
-
Subproject commit
|
|
|
1 |
+
Subproject commit 26787a09bff4ebcb08f0ad4e848b67bce4389a7a
|
trainer.py
CHANGED
@@ -54,10 +54,20 @@ class Trainer:
|
|
54 |
out_path = self.instance_data_dir / f'{i:03d}.jpg'
|
55 |
image.save(out_path, format='JPEG', quality=100)
|
56 |
|
57 |
-
def run(
|
58 |
-
|
59 |
-
|
60 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
if not torch.cuda.is_available():
|
62 |
raise gr.Error('CUDA is not available.')
|
63 |
|
@@ -93,6 +103,8 @@ class Trainer:
|
|
93 |
command += ' --mixed_precision fp16 '
|
94 |
if use_8bit_adam:
|
95 |
command += ' --use_8bit_adam'
|
|
|
|
|
96 |
|
97 |
with open(self.output_dir / 'train.sh', 'w') as f:
|
98 |
command_s = ' '.join(command.split())
|
|
|
54 |
out_path = self.instance_data_dir / f'{i:03d}.jpg'
|
55 |
image.save(out_path, format='JPEG', quality=100)
|
56 |
|
57 |
+
def run(
|
58 |
+
self,
|
59 |
+
base_model: str,
|
60 |
+
resolution_s: str,
|
61 |
+
concept_images: list | None,
|
62 |
+
concept_prompt: str,
|
63 |
+
n_steps: int,
|
64 |
+
learning_rate: float,
|
65 |
+
train_text_encoder: bool,
|
66 |
+
learning_rate_text: float,
|
67 |
+
gradient_accumulation: int,
|
68 |
+
fp16: bool,
|
69 |
+
use_8bit_adam: bool,
|
70 |
+
) -> tuple[dict, str]:
|
71 |
if not torch.cuda.is_available():
|
72 |
raise gr.Error('CUDA is not available.')
|
73 |
|
|
|
103 |
command += ' --mixed_precision fp16 '
|
104 |
if use_8bit_adam:
|
105 |
command += ' --use_8bit_adam'
|
106 |
+
if train_text_encoder:
|
107 |
+
command += f' --train_text_encoder --learning_rate_text={learning_rate_text} --color_jitter'
|
108 |
|
109 |
with open(self.output_dir / 'train.sh', 'w') as f:
|
110 |
command_s = ' '.join(command.split())
|