Spaces:
Runtime error
Runtime error
Damian Stewart
commited on
Commit
·
2c1839c
1
Parent(s):
50b9662
cleanup and try to get cancellation working
Browse files- README.md +14 -0
- StableDiffuser.py +4 -2
- train.py +19 -3
README.md
CHANGED
@@ -10,7 +10,21 @@ pinned: false
|
|
10 |
license: mit
|
11 |
---
|
12 |
|
|
|
13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
# Erasing Concepts from Diffusion Models
|
16 |
|
|
|
10 |
license: mit
|
11 |
---
|
12 |
|
13 |
+
# A GUI with custom model support, validation, and sample generation for "Erasing Concepts from Diffusion Models"
|
14 |
|
15 |
+
Enables xformers, 8 bit AdamW via bitsandbytes, and AMP - editing SD1.5 models works with 16GB VRAM, and 2.5 models including the ESD-u training works with 24GB VRAM.
|
16 |
+
|
17 |
+
## Quick start
|
18 |
+
|
19 |
+
To run on vast.ai, use eg `pytorch/pytorch:2.0.1-cuda11.7-cudnn8-devel` - you need `-devel` for 8bit AdamW to work.
|
20 |
+
|
21 |
+
On the dev machine:
|
22 |
+
```
|
23 |
+
pip install -r requirements.txt
|
24 |
+
python app.py
|
25 |
+
```
|
26 |
+
|
27 |
+
then use the Gradio interface at port 7860.
|
28 |
|
29 |
# Erasing Concepts from Diffusion Models
|
30 |
|
StableDiffuser.py
CHANGED
@@ -107,6 +107,7 @@ class StableDiffuser(torch.nn.Module):
|
|
107 |
return latents
|
108 |
|
109 |
def get_cond_and_uncond_embeddings(self, prompts, negative_prompts=None, n_imgs=1):
|
|
|
110 |
text_tokens = self.text_tokenize(prompts)
|
111 |
text_embeddings = self.text_encode(text_tokens)
|
112 |
if negative_prompts is None:
|
@@ -115,8 +116,9 @@ class StableDiffuser(torch.nn.Module):
|
|
115 |
negative_prompts.append("")
|
116 |
unconditional_tokens = self.text_tokenize(negative_prompts)
|
117 |
unconditional_embeddings = self.text_encode(unconditional_tokens)
|
118 |
-
|
119 |
-
|
|
|
120 |
|
121 |
def predict_noise(self,
|
122 |
iteration,
|
|
|
107 |
return latents
|
108 |
|
109 |
def get_cond_and_uncond_embeddings(self, prompts, negative_prompts=None, n_imgs=1):
|
110 |
+
assert n_imgs == 1
|
111 |
text_tokens = self.text_tokenize(prompts)
|
112 |
text_embeddings = self.text_encode(text_tokens)
|
113 |
if negative_prompts is None:
|
|
|
116 |
negative_prompts.append("")
|
117 |
unconditional_tokens = self.text_tokenize(negative_prompts)
|
118 |
unconditional_embeddings = self.text_encode(unconditional_tokens)
|
119 |
+
combined_embeddings = [torch.cat([unconditional_embeddings[i:i+1], text_embeddings[i:i+1]]) for i in range(len(prompts))]
|
120 |
+
combined_embeddings = torch.cat(combined_embeddings)
|
121 |
+
return combined_embeddings
|
122 |
|
123 |
def predict_noise(self,
|
124 |
iteration,
|
train.py
CHANGED
@@ -34,7 +34,11 @@ def validate(diffuser: StableDiffuser, finetuner: FineTunedModel,
|
|
34 |
|
35 |
nsteps=50
|
36 |
num_validation_prompts = validation_embeddings.shape[0] // 2
|
37 |
-
|
|
|
|
|
|
|
|
|
38 |
accumulated_loss = None
|
39 |
this_validation_embeddings = validation_embeddings[i*2:i*2+2]
|
40 |
for j in range(val_count):
|
@@ -51,10 +55,14 @@ def validate(diffuser: StableDiffuser, finetuner: FineTunedModel,
|
|
51 |
loss = criteria(negative_latents, neutral_latents - (negative_guidance*(positive_latents - neutral_latents)))
|
52 |
accumulated_loss = (accumulated_loss or 0) + loss.item()
|
53 |
logger.add_scalar(f"loss/val_{i}", accumulated_loss/val_count, global_step=global_step)
|
|
|
54 |
|
55 |
num_samples = sample_embeddings.shape[0] // 2
|
56 |
-
for i in range(0, num_samples)
|
57 |
print(f'making sample {i}...')
|
|
|
|
|
|
|
58 |
with finetuner:
|
59 |
pipeline = StableDiffusionPipeline(vae=diffuser.vae,
|
60 |
text_encoder=diffuser.text_encoder,
|
@@ -93,6 +101,8 @@ def train(repo_id_or_path, img_size, prompt, modules, freeze_modules, iterations
|
|
93 |
neutral_latents = None
|
94 |
positive_latents = None
|
95 |
|
|
|
|
|
96 |
nsteps = 50
|
97 |
print(f"using img_size of {img_size}")
|
98 |
diffuser = StableDiffuser(scheduler='DDIM', repo_id_or_path=repo_id_or_path, native_img_size=img_size).to('cuda')
|
@@ -137,13 +147,19 @@ def train(repo_id_or_path, img_size, prompt, modules, freeze_modules, iterations
|
|
137 |
seed = random.randint(0, 2 ** 30)
|
138 |
set_seed(int(seed))
|
139 |
|
|
|
|
|
|
|
|
|
|
|
|
|
140 |
prev_losses = []
|
141 |
start_loss = None
|
142 |
max_prev_loss_count = 10
|
143 |
try:
|
144 |
for i in pbar:
|
145 |
if training_should_cancel:
|
146 |
-
print("
|
147 |
return None
|
148 |
|
149 |
with torch.no_grad():
|
|
|
34 |
|
35 |
nsteps=50
|
36 |
num_validation_prompts = validation_embeddings.shape[0] // 2
|
37 |
+
|
38 |
+
for i in tqdm(range(num_validation_prompts))
|
39 |
+
if training_should_cancel:
|
40 |
+
print("cancel requested, bailing")
|
41 |
+
return
|
42 |
accumulated_loss = None
|
43 |
this_validation_embeddings = validation_embeddings[i*2:i*2+2]
|
44 |
for j in range(val_count):
|
|
|
55 |
loss = criteria(negative_latents, neutral_latents - (negative_guidance*(positive_latents - neutral_latents)))
|
56 |
accumulated_loss = (accumulated_loss or 0) + loss.item()
|
57 |
logger.add_scalar(f"loss/val_{i}", accumulated_loss/val_count, global_step=global_step)
|
58 |
+
pbar.step()
|
59 |
|
60 |
num_samples = sample_embeddings.shape[0] // 2
|
61 |
+
for i in tqdm(range(0, num_samples));
|
62 |
print(f'making sample {i}...')
|
63 |
+
if training_should_cancel:
|
64 |
+
print("cancel requested, bailing")
|
65 |
+
return
|
66 |
with finetuner:
|
67 |
pipeline = StableDiffusionPipeline(vae=diffuser.vae,
|
68 |
text_encoder=diffuser.text_encoder,
|
|
|
101 |
neutral_latents = None
|
102 |
positive_latents = None
|
103 |
|
104 |
+
global training_should_cancel
|
105 |
+
|
106 |
nsteps = 50
|
107 |
print(f"using img_size of {img_size}")
|
108 |
diffuser = StableDiffuser(scheduler='DDIM', repo_id_or_path=repo_id_or_path, native_img_size=img_size).to('cuda')
|
|
|
147 |
seed = random.randint(0, 2 ** 30)
|
148 |
set_seed(int(seed))
|
149 |
|
150 |
+
validate(diffuser, finetuner,
|
151 |
+
validation_embeddings=validation_embeddings,
|
152 |
+
sample_embeddings=sample_embeddings,
|
153 |
+
neutral_embeddings=neutral_text_embeddings,
|
154 |
+
logger=logger, use_amp=False, global_step=0)
|
155 |
+
|
156 |
prev_losses = []
|
157 |
start_loss = None
|
158 |
max_prev_loss_count = 10
|
159 |
try:
|
160 |
for i in pbar:
|
161 |
if training_should_cancel:
|
162 |
+
print("cancel requested, bailing")
|
163 |
return None
|
164 |
|
165 |
with torch.no_grad():
|