Spaces:
Runtime error
Runtime error
Damian Stewart
commited on
Commit
•
bf1e262
1
Parent(s):
52c8f3c
allow multiple train prompts
Browse files- app.py +49 -37
- memory_efficiency.py +4 -1
- train.py +75 -61
app.py
CHANGED
@@ -12,15 +12,20 @@ from train import train, training_should_cancel
|
|
12 |
import os
|
13 |
|
14 |
model_map = {}
|
15 |
-
|
|
|
|
|
16 |
global model_map
|
|
|
17 |
for model_file in os.listdir('models'):
|
18 |
path = 'models/' + model_file
|
19 |
if any([existing_path == path for existing_path in model_map.values()]):
|
20 |
continue
|
21 |
model_map[model_file] = path
|
22 |
-
|
23 |
-
model_map
|
|
|
|
|
24 |
|
25 |
ORIGINAL_SPACE_ID = 'baulab/Erasing-Concepts-In-Diffusion'
|
26 |
SPACE_ID = os.getenv('SPACE_ID')
|
@@ -64,6 +69,12 @@ class Demo:
|
|
64 |
|
65 |
with gr.Column(scale=1):
|
66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
67 |
self.prompt_input_infr = gr.Text(
|
68 |
placeholder="Enter prompt...",
|
69 |
label="Prompt",
|
@@ -104,12 +115,6 @@ class Demo:
|
|
104 |
interactive=True
|
105 |
)
|
106 |
|
107 |
-
self.base_repo_id_or_path_input_infr = gr.Text(
|
108 |
-
label="Base model",
|
109 |
-
value="CompVis/stable-diffusion-v1-4",
|
110 |
-
info="Path or huggingface repo id of the base model that this edit was done against"
|
111 |
-
)
|
112 |
-
|
113 |
with gr.Column(scale=2):
|
114 |
|
115 |
self.infr_button = gr.Button(
|
@@ -152,19 +157,10 @@ class Demo:
|
|
152 |
info="Image size for training, should match the model's native image size"
|
153 |
)
|
154 |
|
155 |
-
self.
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
maximum=32,
|
160 |
-
label="Sample generation batch size",
|
161 |
-
info="Batch size for sample generation, larger needs more VRAM"
|
162 |
-
)
|
163 |
-
|
164 |
-
self.prompt_input = gr.Text(
|
165 |
-
placeholder="Enter prompt...",
|
166 |
-
label="Prompt to Erase",
|
167 |
-
info="Prompt corresponding to concept to erase"
|
168 |
)
|
169 |
|
170 |
choices = ['ESD-x', 'ESD-self', 'ESD-u']
|
@@ -175,7 +171,7 @@ class Demo:
|
|
175 |
choices=choices,
|
176 |
value='ESD-x',
|
177 |
label='Train Method',
|
178 |
-
info='Method of training'
|
179 |
)
|
180 |
|
181 |
self.neg_guidance_input = gr.Number(
|
@@ -233,11 +229,21 @@ class Demo:
|
|
233 |
value='',
|
234 |
info="Negative prompts for use when generating sample images. One for each positive prompt, or leave empty for none."
|
235 |
)
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
241 |
|
242 |
with gr.Column(scale=1):
|
243 |
|
@@ -311,7 +317,7 @@ class Demo:
|
|
311 |
train_event = self.train_button.click(self.train, inputs = [
|
312 |
self.train_model_input,
|
313 |
self.train_img_size_input,
|
314 |
-
self.
|
315 |
self.train_method_input,
|
316 |
self.neg_guidance_input,
|
317 |
self.iterations_input,
|
@@ -346,9 +352,9 @@ class Demo:
|
|
346 |
|
347 |
def reload_models(self, model_dropdown):
|
348 |
current_model_name = model_dropdown
|
349 |
-
|
350 |
-
|
351 |
-
return [self.model_dropdown.update(choices=
|
352 |
|
353 |
def cancel_training(self):
|
354 |
if self.training:
|
@@ -356,7 +362,7 @@ class Demo:
|
|
356 |
print("cancellation requested...")
|
357 |
return [gr.update(value="Cancelling...", interactive=True)]
|
358 |
|
359 |
-
def train(self, repo_id_or_path, img_size,
|
360 |
use_adamw8bit=True, use_xformers=False, use_amp=False, use_gradient_checkpointing=False,
|
361 |
seed=-1, save_every=-1, sample_batch_size=1,
|
362 |
validation_prompts: str=None, sample_positive_prompts: str=None, sample_negative_prompts: str=None, validate_every_n_steps=-1,
|
@@ -365,7 +371,7 @@ class Demo:
|
|
365 |
|
366 |
:param repo_id_or_path:
|
367 |
:param img_size:
|
368 |
-
:param
|
369 |
:param train_method:
|
370 |
:param neg_guidance:
|
371 |
:param iterations:
|
@@ -386,7 +392,7 @@ class Demo:
|
|
386 |
if self.training:
|
387 |
return [gr.update(interactive=True, value='Train'), gr.update(value='Someone else is training... Try again soon'), None, gr.update()]
|
388 |
|
389 |
-
print(f"Training {repo_id_or_path} at {img_size} to remove '{
|
390 |
print(f" {train_method}, negative guidance {neg_guidance}, lr {lr}, {iterations} iterations.")
|
391 |
print(f" {'✅' if use_gradient_checkpointing else '❌'} gradient checkpointing")
|
392 |
print(f" {'✅' if use_amp else '❌'} AMP")
|
@@ -409,11 +415,12 @@ class Demo:
|
|
409 |
while True:
|
410 |
randn = torch.randint(1, 10000000, (1,)).item()
|
411 |
options = f'{"a8" if use_adamw8bit else ""}{"AM" if use_amp else ""}{"xf" if use_xformers else ""}{"gc" if use_gradient_checkpointing else ""}'
|
412 |
-
save_path = f"models/{
|
413 |
if not os.path.exists(save_path):
|
414 |
break
|
415 |
# repeat until a not-in-use path is found
|
416 |
|
|
|
417 |
validation_prompts = [] if validation_prompts is None else [p for p in validation_prompts.split('\n') if len(p)>0]
|
418 |
sample_positive_prompts = [] if sample_positive_prompts is None else [p for p in sample_positive_prompts.split('\n') if len(p)>0]
|
419 |
sample_negative_prompts = [] if sample_negative_prompts is None else sample_negative_prompts.split('\n')
|
@@ -425,7 +432,7 @@ class Demo:
|
|
425 |
self.training = True
|
426 |
self.train_cancel_button.update(interactive=True)
|
427 |
batch_size = 1 # other batch sizes are non-functional
|
428 |
-
save_path = train(repo_id_or_path, img_size,
|
429 |
use_adamw8bit, use_xformers, use_amp, use_gradient_checkpointing,
|
430 |
seed=int(seed), save_every_n_steps=int(save_every),
|
431 |
batch_size=int(batch_size), sample_batch_size=int(sample_batch_size),
|
@@ -476,6 +483,11 @@ class Demo:
|
|
476 |
model_path = model_map[model_name]
|
477 |
checkpoint = torch.load(model_path)
|
478 |
|
|
|
|
|
|
|
|
|
|
|
479 |
self.diffuser = StableDiffuser(scheduler='DDIM', repo_id_or_path=base_repo_id_or_path).to('cuda').eval().half()
|
480 |
finetuner = FineTunedModel.from_checkpoint(self.diffuser, checkpoint).eval().half()
|
481 |
|
|
|
12 |
import os
|
13 |
|
14 |
model_map = {}
|
15 |
+
model_names_list = []
|
16 |
+
|
17 |
+
def populate_global_model_map():
|
18 |
global model_map
|
19 |
+
global model_names_list
|
20 |
for model_file in os.listdir('models'):
|
21 |
path = 'models/' + model_file
|
22 |
if any([existing_path == path for existing_path in model_map.values()]):
|
23 |
continue
|
24 |
model_map[model_file] = path
|
25 |
+
model_names_list.clear()
|
26 |
+
model_names_list.extend(model_map.keys())
|
27 |
+
|
28 |
+
populate_global_model_map()
|
29 |
|
30 |
ORIGINAL_SPACE_ID = 'baulab/Erasing-Concepts-In-Diffusion'
|
31 |
SPACE_ID = os.getenv('SPACE_ID')
|
|
|
69 |
|
70 |
with gr.Column(scale=1):
|
71 |
|
72 |
+
self.base_repo_id_or_path_input_infr = gr.Text(
|
73 |
+
label="Base model",
|
74 |
+
value="CompVis/stable-diffusion-v1-4",
|
75 |
+
info="Path or huggingface repo id of the base model that this edit was done against"
|
76 |
+
)
|
77 |
+
|
78 |
self.prompt_input_infr = gr.Text(
|
79 |
placeholder="Enter prompt...",
|
80 |
label="Prompt",
|
|
|
115 |
interactive=True
|
116 |
)
|
117 |
|
|
|
|
|
|
|
|
|
|
|
|
|
118 |
with gr.Column(scale=2):
|
119 |
|
120 |
self.infr_button = gr.Button(
|
|
|
157 |
info="Image size for training, should match the model's native image size"
|
158 |
)
|
159 |
|
160 |
+
self.train_prompts_input = gr.Text(
|
161 |
+
placeholder="Enter prompts, one per line",
|
162 |
+
label="Prompts to Erase",
|
163 |
+
info="Prompts corresponding to concepts to erase, one per line"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
164 |
)
|
165 |
|
166 |
choices = ['ESD-x', 'ESD-self', 'ESD-u']
|
|
|
171 |
choices=choices,
|
172 |
value='ESD-x',
|
173 |
label='Train Method',
|
174 |
+
info='Method of training. ESD-x uses the least VRAM, and you may get OOM errors with the other methods.'
|
175 |
)
|
176 |
|
177 |
self.neg_guidance_input = gr.Number(
|
|
|
229 |
value='',
|
230 |
info="Negative prompts for use when generating sample images. One for each positive prompt, or leave empty for none."
|
231 |
)
|
232 |
+
|
233 |
+
with gr.Row():
|
234 |
+
self.train_sample_batch_size_input = gr.Slider(
|
235 |
+
value=1,
|
236 |
+
step=1,
|
237 |
+
minimum=1,
|
238 |
+
maximum=32,
|
239 |
+
label="Sample generation batch size",
|
240 |
+
info="Batch size for sample generation, larger needs more VRAM"
|
241 |
+
)
|
242 |
+
self.train_validate_every_n_steps = gr.Number(
|
243 |
+
label="Validate Every N Steps",
|
244 |
+
value=20,
|
245 |
+
info="Validation and sample generation will be run at intervals of this many steps"
|
246 |
+
)
|
247 |
|
248 |
with gr.Column(scale=1):
|
249 |
|
|
|
317 |
train_event = self.train_button.click(self.train, inputs = [
|
318 |
self.train_model_input,
|
319 |
self.train_img_size_input,
|
320 |
+
self.train_prompts_input,
|
321 |
self.train_method_input,
|
322 |
self.neg_guidance_input,
|
323 |
self.iterations_input,
|
|
|
352 |
|
353 |
def reload_models(self, model_dropdown):
|
354 |
current_model_name = model_dropdown
|
355 |
+
populate_global_model_map()
|
356 |
+
global model_names_list
|
357 |
+
return [self.model_dropdown.update(choices=model_names_list, value=current_model_name)]
|
358 |
|
359 |
def cancel_training(self):
|
360 |
if self.training:
|
|
|
362 |
print("cancellation requested...")
|
363 |
return [gr.update(value="Cancelling...", interactive=True)]
|
364 |
|
365 |
+
def train(self, repo_id_or_path, img_size, prompts, train_method, neg_guidance, iterations, lr,
|
366 |
use_adamw8bit=True, use_xformers=False, use_amp=False, use_gradient_checkpointing=False,
|
367 |
seed=-1, save_every=-1, sample_batch_size=1,
|
368 |
validation_prompts: str=None, sample_positive_prompts: str=None, sample_negative_prompts: str=None, validate_every_n_steps=-1,
|
|
|
371 |
|
372 |
:param repo_id_or_path:
|
373 |
:param img_size:
|
374 |
+
:param prompts:
|
375 |
:param train_method:
|
376 |
:param neg_guidance:
|
377 |
:param iterations:
|
|
|
392 |
if self.training:
|
393 |
return [gr.update(interactive=True, value='Train'), gr.update(value='Someone else is training... Try again soon'), None, gr.update()]
|
394 |
|
395 |
+
print(f"Training {repo_id_or_path} at {img_size} to remove '{prompts}'.")
|
396 |
print(f" {train_method}, negative guidance {neg_guidance}, lr {lr}, {iterations} iterations.")
|
397 |
print(f" {'✅' if use_gradient_checkpointing else '❌'} gradient checkpointing")
|
398 |
print(f" {'✅' if use_amp else '❌'} AMP")
|
|
|
415 |
while True:
|
416 |
randn = torch.randint(1, 10000000, (1,)).item()
|
417 |
options = f'{"a8" if use_adamw8bit else ""}{"AM" if use_amp else ""}{"xf" if use_xformers else ""}{"gc" if use_gradient_checkpointing else ""}'
|
418 |
+
save_path = f"models/{prompts[0].lower().replace(' ', '')}_{train_method}_ng{neg_guidance}_lr{lr}_iter{iterations}_seed{seed}_{options}__{randn}.pt"
|
419 |
if not os.path.exists(save_path):
|
420 |
break
|
421 |
# repeat until a not-in-use path is found
|
422 |
|
423 |
+
prompts = [p for p in prompts.split('\n') if len(p)>0]
|
424 |
validation_prompts = [] if validation_prompts is None else [p for p in validation_prompts.split('\n') if len(p)>0]
|
425 |
sample_positive_prompts = [] if sample_positive_prompts is None else [p for p in sample_positive_prompts.split('\n') if len(p)>0]
|
426 |
sample_negative_prompts = [] if sample_negative_prompts is None else sample_negative_prompts.split('\n')
|
|
|
432 |
self.training = True
|
433 |
self.train_cancel_button.update(interactive=True)
|
434 |
batch_size = 1 # other batch sizes are non-functional
|
435 |
+
save_path = train(repo_id_or_path, img_size, prompts, modules, frozen, iterations, neg_guidance, lr, save_path,
|
436 |
use_adamw8bit, use_xformers, use_amp, use_gradient_checkpointing,
|
437 |
seed=int(seed), save_every_n_steps=int(save_every),
|
438 |
batch_size=int(batch_size), sample_batch_size=int(sample_batch_size),
|
|
|
483 |
model_path = model_map[model_name]
|
484 |
checkpoint = torch.load(model_path)
|
485 |
|
486 |
+
if type(prompt) is str:
|
487 |
+
prompt = [prompt]
|
488 |
+
if type(negative_prompt) is str:
|
489 |
+
negative_prompt = [negative_prompt]
|
490 |
+
|
491 |
self.diffuser = StableDiffuser(scheduler='DDIM', repo_id_or_path=base_repo_id_or_path).to('cuda').eval().half()
|
492 |
finetuner = FineTunedModel.from_checkpoint(self.diffuser, checkpoint).eval().half()
|
493 |
|
memory_efficiency.py
CHANGED
@@ -66,10 +66,13 @@ class MemoryEfficiencyWrapper:
|
|
66 |
growth_interval=25,
|
67 |
)
|
68 |
|
69 |
-
def
|
70 |
self.grad_scaler.scale(loss).backward()
|
|
|
|
|
71 |
self.grad_scaler.step(optimizer)
|
72 |
self.grad_scaler.update()
|
|
|
73 |
|
74 |
def __exit__(self, exc_type, exc_value, tb):
|
75 |
if exc_type is not None:
|
|
|
66 |
growth_interval=25,
|
67 |
)
|
68 |
|
69 |
+
def backward(self, loss):
|
70 |
self.grad_scaler.scale(loss).backward()
|
71 |
+
|
72 |
+
def step(self, optimizer):
|
73 |
self.grad_scaler.step(optimizer)
|
74 |
self.grad_scaler.update()
|
75 |
+
optimizer.zero_grad(set_to_none=True)
|
76 |
|
77 |
def __exit__(self, exc_type, exc_value, tb):
|
78 |
if exc_type is not None:
|
train.py
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
import os.path
|
2 |
import random
|
3 |
import multiprocessing
|
|
|
4 |
|
5 |
from accelerate.utils import set_seed
|
6 |
from diffusers import StableDiffusionPipeline
|
@@ -34,11 +35,13 @@ def validate(diffuser: StableDiffuser, finetuner: FineTunedModel,
|
|
34 |
set_seed(validation_seed)
|
35 |
criteria = torch.nn.MSELoss()
|
36 |
negative_guidance = 1
|
37 |
-
val_count = 5
|
38 |
|
39 |
nsteps=50
|
40 |
num_validation_batches = validation_embeddings.shape[0] // (batch_size*2)
|
41 |
|
|
|
|
|
|
|
42 |
for i in tqdm(range(num_validation_batches)):
|
43 |
if training_should_cancel.acquire(block=False):
|
44 |
print("cancel requested, bailing")
|
@@ -58,9 +61,11 @@ def validate(diffuser: StableDiffuser, finetuner: FineTunedModel,
|
|
58 |
|
59 |
loss = criteria(negative_latents, neutral_latents - (negative_guidance*(positive_latents - neutral_latents)))
|
60 |
accumulated_loss = (accumulated_loss or 0) + loss.item()
|
|
|
61 |
logger.add_scalar(f"loss/val_{i}", accumulated_loss/val_count, global_step=global_step)
|
|
|
62 |
|
63 |
-
num_sample_batches = sample_embeddings.shape[0]
|
64 |
for i in tqdm(range(0, num_sample_batches)):
|
65 |
print(f'making sample batch {i}...')
|
66 |
if training_should_cancel.acquire(block=False):
|
@@ -82,9 +87,9 @@ def validate(diffuser: StableDiffuser, finetuner: FineTunedModel,
|
|
82 |
images = pipeline(prompt_embeds=batch_prompt_embeds, #sample_embeddings[i*2+1:i*2+2],
|
83 |
negative_prompt_embeds=batch_negative_prompt_embeds, # sample_embeddings[i*2:i*2+1],
|
84 |
num_inference_steps=50)
|
85 |
-
for
|
86 |
-
image_tensor = transforms.ToTensor()(
|
87 |
-
logger.add_image(f"samples/{i*sample_batch_size+
|
88 |
|
89 |
"""
|
90 |
with finetuner, torch.cuda.amp.autocast(enabled=use_amp):
|
@@ -97,20 +102,12 @@ def validate(diffuser: StableDiffuser, finetuner: FineTunedModel,
|
|
97 |
|
98 |
torch.cuda.empty_cache()
|
99 |
|
100 |
-
def train(repo_id_or_path, img_size,
|
101 |
use_adamw8bit=True, use_xformers=True, use_amp=True, use_gradient_checkpointing=False, seed=-1,
|
102 |
batch_size=1, sample_batch_size=1,
|
103 |
save_every_n_steps=-1, validate_every_n_steps=-1,
|
104 |
validation_prompts=[], sample_positive_prompts=[], sample_negative_prompts=[]):
|
105 |
|
106 |
-
diffuser = None
|
107 |
-
loss = None
|
108 |
-
optimizer = None
|
109 |
-
finetuner = None
|
110 |
-
negative_latents = None
|
111 |
-
neutral_latents = None
|
112 |
-
positive_latents = None
|
113 |
-
|
114 |
nsteps = 50
|
115 |
print(f"using img_size of {img_size}")
|
116 |
diffuser = StableDiffuser(scheduler='DDIM', repo_id_or_path=repo_id_or_path, native_img_size=img_size).to('cuda')
|
@@ -118,7 +115,7 @@ def train(repo_id_or_path, img_size, prompt, modules, freeze_modules, iterations
|
|
118 |
|
119 |
memory_efficiency_wrapper = MemoryEfficiencyWrapper(diffuser=diffuser, use_amp=use_amp, use_xformers=use_xformers,
|
120 |
use_gradient_checkpointing=use_gradient_checkpointing )
|
121 |
-
with memory_efficiency_wrapper:
|
122 |
diffuser.train()
|
123 |
finetuner = FineTunedModel(diffuser, modules, frozen_modules=freeze_modules)
|
124 |
if use_adamw8bit:
|
@@ -139,7 +136,7 @@ def train(repo_id_or_path, img_size, prompt, modules, freeze_modules, iterations
|
|
139 |
|
140 |
with torch.no_grad():
|
141 |
neutral_text_embeddings = diffuser.get_cond_and_uncond_embeddings([''], n_imgs=1)
|
142 |
-
|
143 |
validation_embeddings = diffuser.get_cond_and_uncond_embeddings(validation_prompts, n_imgs=1)
|
144 |
sample_embeddings = diffuser.get_cond_and_uncond_embeddings(sample_positive_prompts, sample_negative_prompts, n_imgs=1)
|
145 |
|
@@ -173,51 +170,68 @@ def train(repo_id_or_path, img_size, prompt, modules, freeze_modules, iterations
|
|
173 |
start_loss = None
|
174 |
max_prev_loss_count = 10
|
175 |
try:
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
with torch.no_grad():
|
182 |
-
optimizer.zero_grad()
|
183 |
-
|
184 |
-
iteration = torch.randint(1, nsteps - 1, (1,)).item()
|
185 |
-
|
186 |
-
with finetuner:
|
187 |
-
diffused_latents = get_diffused_latents(diffuser, nsteps, positive_text_embeddings, iteration, use_amp)
|
188 |
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
221 |
|
222 |
if save_every_n_steps > 0 and ((i+1) % save_every_n_steps) == 0:
|
223 |
torch.save(finetuner.state_dict(), save_path + f"__step_{i+1}.pt")
|
@@ -231,7 +245,7 @@ def train(repo_id_or_path, img_size, prompt, modules, freeze_modules, iterations
|
|
231 |
torch.save(finetuner.state_dict(), save_path)
|
232 |
return save_path
|
233 |
finally:
|
234 |
-
del diffuser,
|
235 |
torch.cuda.empty_cache()
|
236 |
|
237 |
|
|
|
1 |
import os.path
|
2 |
import random
|
3 |
import multiprocessing
|
4 |
+
import math
|
5 |
|
6 |
from accelerate.utils import set_seed
|
7 |
from diffusers import StableDiffusionPipeline
|
|
|
35 |
set_seed(validation_seed)
|
36 |
criteria = torch.nn.MSELoss()
|
37 |
negative_guidance = 1
|
|
|
38 |
|
39 |
nsteps=50
|
40 |
num_validation_batches = validation_embeddings.shape[0] // (batch_size*2)
|
41 |
|
42 |
+
val_count = max(1, 5 // num_validation_batches)
|
43 |
+
|
44 |
+
val_total_loss = 0
|
45 |
for i in tqdm(range(num_validation_batches)):
|
46 |
if training_should_cancel.acquire(block=False):
|
47 |
print("cancel requested, bailing")
|
|
|
61 |
|
62 |
loss = criteria(negative_latents, neutral_latents - (negative_guidance*(positive_latents - neutral_latents)))
|
63 |
accumulated_loss = (accumulated_loss or 0) + loss.item()
|
64 |
+
val_total_loss += loss.item()
|
65 |
logger.add_scalar(f"loss/val_{i}", accumulated_loss/val_count, global_step=global_step)
|
66 |
+
logger.add_scalar(f"loss/_val_all_combined", val_total_loss/(val_count*num_validation_batches), global_step=global_step)
|
67 |
|
68 |
+
num_sample_batches = int(math.ceil(sample_embeddings.shape[0] / (sample_batch_size*2)))
|
69 |
for i in tqdm(range(0, num_sample_batches)):
|
70 |
print(f'making sample batch {i}...')
|
71 |
if training_should_cancel.acquire(block=False):
|
|
|
87 |
images = pipeline(prompt_embeds=batch_prompt_embeds, #sample_embeddings[i*2+1:i*2+2],
|
88 |
negative_prompt_embeds=batch_negative_prompt_embeds, # sample_embeddings[i*2:i*2+1],
|
89 |
num_inference_steps=50)
|
90 |
+
for image_index, image in enumerate(images.images):
|
91 |
+
image_tensor = transforms.ToTensor()(image)
|
92 |
+
logger.add_image(f"samples/{i*sample_batch_size+image_index}", img_tensor=image_tensor, global_step=global_step)
|
93 |
|
94 |
"""
|
95 |
with finetuner, torch.cuda.amp.autocast(enabled=use_amp):
|
|
|
102 |
|
103 |
torch.cuda.empty_cache()
|
104 |
|
105 |
+
def train(repo_id_or_path, img_size, prompts, modules, freeze_modules, iterations, negative_guidance, lr, save_path,
|
106 |
use_adamw8bit=True, use_xformers=True, use_amp=True, use_gradient_checkpointing=False, seed=-1,
|
107 |
batch_size=1, sample_batch_size=1,
|
108 |
save_every_n_steps=-1, validate_every_n_steps=-1,
|
109 |
validation_prompts=[], sample_positive_prompts=[], sample_negative_prompts=[]):
|
110 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
111 |
nsteps = 50
|
112 |
print(f"using img_size of {img_size}")
|
113 |
diffuser = StableDiffuser(scheduler='DDIM', repo_id_or_path=repo_id_or_path, native_img_size=img_size).to('cuda')
|
|
|
115 |
|
116 |
memory_efficiency_wrapper = MemoryEfficiencyWrapper(diffuser=diffuser, use_amp=use_amp, use_xformers=use_xformers,
|
117 |
use_gradient_checkpointing=use_gradient_checkpointing )
|
118 |
+
with (((((memory_efficiency_wrapper))))):
|
119 |
diffuser.train()
|
120 |
finetuner = FineTunedModel(diffuser, modules, frozen_modules=freeze_modules)
|
121 |
if use_adamw8bit:
|
|
|
136 |
|
137 |
with torch.no_grad():
|
138 |
neutral_text_embeddings = diffuser.get_cond_and_uncond_embeddings([''], n_imgs=1)
|
139 |
+
all_positive_text_embeddings = diffuser.get_cond_and_uncond_embeddings(prompts, n_imgs=1)
|
140 |
validation_embeddings = diffuser.get_cond_and_uncond_embeddings(validation_prompts, n_imgs=1)
|
141 |
sample_embeddings = diffuser.get_cond_and_uncond_embeddings(sample_positive_prompts, sample_negative_prompts, n_imgs=1)
|
142 |
|
|
|
170 |
start_loss = None
|
171 |
max_prev_loss_count = 10
|
172 |
try:
|
173 |
+
loss=None
|
174 |
+
negative_latents=None
|
175 |
+
neutral_latents=None
|
176 |
+
positive_latents=None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
177 |
|
178 |
+
num_prompts = all_positive_text_embeddings.shape[0] // 2
|
179 |
+
for i in pbar:
|
180 |
+
try:
|
181 |
+
loss = None
|
182 |
+
negative_latents = None
|
183 |
+
positive_latents = None
|
184 |
+
neutral_latents = None
|
185 |
+
diffused_latents = None
|
186 |
+
for j in tqdm(range(num_prompts)):
|
187 |
+
positive_text_embeddings = all_positive_text_embeddings[j*2:j*2+2]
|
188 |
+
if training_should_cancel.acquire(block=False):
|
189 |
+
print("cancel requested, bailing")
|
190 |
+
return None
|
191 |
+
|
192 |
+
with torch.no_grad():
|
193 |
+
optimizer.zero_grad()
|
194 |
+
|
195 |
+
iteration = torch.randint(1, nsteps - 1, (1,)).item()
|
196 |
+
|
197 |
+
with finetuner:
|
198 |
+
diffused_latents = get_diffused_latents(diffuser, nsteps, positive_text_embeddings, iteration, use_amp)
|
199 |
+
|
200 |
+
iteration = int(iteration / nsteps * 1000)
|
201 |
+
|
202 |
+
with autocast(enabled=use_amp):
|
203 |
+
positive_latents = diffuser.predict_noise(iteration, diffused_latents, positive_text_embeddings, guidance_scale=1)
|
204 |
+
neutral_latents = diffuser.predict_noise(iteration, diffused_latents, neutral_text_embeddings, guidance_scale=1)
|
205 |
+
|
206 |
+
with finetuner:
|
207 |
+
with autocast(enabled=use_amp):
|
208 |
+
negative_latents = diffuser.predict_noise(iteration, diffused_latents, positive_text_embeddings, guidance_scale=1)
|
209 |
+
|
210 |
+
positive_latents.requires_grad = False
|
211 |
+
neutral_latents.requires_grad = False
|
212 |
+
|
213 |
+
# loss = criteria(e_n, e_0) works the best try 5000 epochs
|
214 |
+
loss = criteria(negative_latents, neutral_latents - (negative_guidance*(positive_latents - neutral_latents)))
|
215 |
+
memory_efficiency_wrapper.backward(loss)
|
216 |
+
|
217 |
+
logger.add_scalar("loss", loss.item(), global_step=i)
|
218 |
+
|
219 |
+
# print moving average loss
|
220 |
+
prev_losses.append(loss.detach().clone())
|
221 |
+
if len(prev_losses) > max_prev_loss_count:
|
222 |
+
prev_losses.pop(0)
|
223 |
+
if start_loss is None:
|
224 |
+
start_loss = prev_losses[-1]
|
225 |
+
if len(prev_losses) >= max_prev_loss_count:
|
226 |
+
moving_average_loss = sum(prev_losses) / len(prev_losses)
|
227 |
+
print(
|
228 |
+
f"step {i}: loss={loss.item()} (avg={moving_average_loss.item()}, start ∆={(moving_average_loss - start_loss).item()}")
|
229 |
+
else:
|
230 |
+
print(f"step {i}: loss={loss.item()}")
|
231 |
+
|
232 |
+
memory_efficiency_wrapper.step(optimizer)
|
233 |
+
finally:
|
234 |
+
del loss, negative_latents, positive_latents, neutral_latents, diffused_latents
|
235 |
|
236 |
if save_every_n_steps > 0 and ((i+1) % save_every_n_steps) == 0:
|
237 |
torch.save(finetuner.state_dict(), save_path + f"__step_{i+1}.pt")
|
|
|
245 |
torch.save(finetuner.state_dict(), save_path)
|
246 |
return save_path
|
247 |
finally:
|
248 |
+
del diffuser, optimizer, finetuner
|
249 |
torch.cuda.empty_cache()
|
250 |
|
251 |
|