Spaces:
Build error
Build error
Update app.py
Browse files
app.py
CHANGED
|
@@ -149,25 +149,13 @@ models_rbm = core.Models(
|
|
| 149 |
models_rbm.generator.eval().requires_grad_(False)
|
| 150 |
|
| 151 |
def infer(style_description, ref_style_file, caption):
|
| 152 |
-
# Move models to the correct device
|
| 153 |
-
models_rbm.effnet.to(device)
|
| 154 |
-
models_rbm.generator.to(device)
|
| 155 |
-
if low_vram:
|
| 156 |
-
models_rbm.previewer.to(device)
|
| 157 |
-
|
| 158 |
-
# Also, revalidate data types and devices for key tensors
|
| 159 |
-
def check_and_move(tensor):
|
| 160 |
-
if tensor is not None and tensor.device != device:
|
| 161 |
-
return tensor.to(device)
|
| 162 |
-
return tensor
|
| 163 |
-
|
| 164 |
clear_gpu_cache() # Clear cache before inference
|
| 165 |
|
| 166 |
-
height
|
| 167 |
-
width
|
| 168 |
-
batch_size
|
| 169 |
-
output_file
|
| 170 |
-
|
| 171 |
stage_c_latent_shape, stage_b_latent_shape = calculate_latent_sizes(height, width, batch_size=batch_size)
|
| 172 |
|
| 173 |
extras.sampling_configs['cfg'] = 4
|
|
@@ -180,26 +168,24 @@ def infer(style_description, ref_style_file, caption):
|
|
| 180 |
extras_b.sampling_configs['timesteps'] = 10
|
| 181 |
extras_b.sampling_configs['t_start'] = 1.0
|
| 182 |
|
| 183 |
-
|
| 184 |
-
ref_style = resize_image(PIL.Image.open(ref_style_file).convert("RGB"))
|
| 185 |
-
ref_style = ref_style.unsqueeze(0).expand(batch_size, -1, -1, -1).to(device)
|
| 186 |
|
| 187 |
batch = {'captions': [caption] * batch_size}
|
| 188 |
batch['style'] = ref_style
|
| 189 |
|
| 190 |
x0_style_forward = models_rbm.effnet(extras.effnet_preprocess(ref_style.to(device)))
|
| 191 |
|
| 192 |
-
conditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=False, eval_image_embeds=True, eval_style=True, eval_csd=False)
|
| 193 |
-
unconditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=True, eval_image_embeds=False)
|
| 194 |
conditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=False)
|
| 195 |
unconditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=True)
|
| 196 |
|
| 197 |
if low_vram:
|
| 198 |
-
#
|
| 199 |
models_to(models_rbm, device="cpu", excepts=["generator", "previewer"])
|
| 200 |
|
| 201 |
-
# Stage C reverse process
|
| 202 |
-
with torch.cuda.amp.autocast():
|
| 203 |
sampling_c = extras.gdf.sample(
|
| 204 |
models_rbm.generator, conditions, stage_c_latent_shape,
|
| 205 |
unconditions, device=device,
|
|
@@ -216,24 +202,19 @@ def infer(style_description, ref_style_file, caption):
|
|
| 216 |
|
| 217 |
clear_gpu_cache() # Clear cache between stages
|
| 218 |
|
| 219 |
-
#
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
with torch.no_grad(), torch.cuda.amp.autocast(dtype=torch.bfloat16):
|
| 225 |
-
conditions_b['effnet'] = sampled_c.to(device)
|
| 226 |
-
unconditions_b['effnet'] = torch.zeros_like(sampled_c).to(device)
|
| 227 |
-
|
| 228 |
sampling_b = extras_b.gdf.sample(
|
| 229 |
models_b.generator, conditions_b, stage_b_latent_shape,
|
| 230 |
unconditions_b, device=device, **extras_b.sampling_configs,
|
| 231 |
)
|
| 232 |
for (sampled_b, _, _) in tqdm(sampling_b, total=extras_b.sampling_configs['timesteps']):
|
| 233 |
sampled_b = sampled_b
|
| 234 |
-
sampled = models_b.stage_a.decode(sampled_b).float()
|
| 235 |
|
| 236 |
-
# Post-process and save the image
|
| 237 |
sampled = torch.cat([
|
| 238 |
torch.nn.functional.interpolate(ref_style.cpu(), size=(height, width)),
|
| 239 |
sampled.cpu(),
|
|
@@ -253,8 +234,6 @@ def infer(style_description, ref_style_file, caption):
|
|
| 253 |
|
| 254 |
return output_file # Return the path to the saved image
|
| 255 |
|
| 256 |
-
|
| 257 |
-
|
| 258 |
import gradio as gr
|
| 259 |
|
| 260 |
gr.Interface(
|
|
|
|
| 149 |
models_rbm.generator.eval().requires_grad_(False)
|
| 150 |
|
| 151 |
def infer(style_description, ref_style_file, caption):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 152 |
clear_gpu_cache() # Clear cache before inference
|
| 153 |
|
| 154 |
+
height=1024
|
| 155 |
+
width=1024
|
| 156 |
+
batch_size=1
|
| 157 |
+
output_file='output.png'
|
| 158 |
+
|
| 159 |
stage_c_latent_shape, stage_b_latent_shape = calculate_latent_sizes(height, width, batch_size=batch_size)
|
| 160 |
|
| 161 |
extras.sampling_configs['cfg'] = 4
|
|
|
|
| 168 |
extras_b.sampling_configs['timesteps'] = 10
|
| 169 |
extras_b.sampling_configs['t_start'] = 1.0
|
| 170 |
|
| 171 |
+
ref_style = resize_image(PIL.Image.open(ref_style_file).convert("RGB")).unsqueeze(0).expand(batch_size, -1, -1, -1).to(device)
|
|
|
|
|
|
|
| 172 |
|
| 173 |
batch = {'captions': [caption] * batch_size}
|
| 174 |
batch['style'] = ref_style
|
| 175 |
|
| 176 |
x0_style_forward = models_rbm.effnet(extras.effnet_preprocess(ref_style.to(device)))
|
| 177 |
|
| 178 |
+
conditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=False, eval_image_embeds=True, eval_style=True, eval_csd=False)
|
| 179 |
+
unconditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=True, eval_image_embeds=False)
|
| 180 |
conditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=False)
|
| 181 |
unconditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=True)
|
| 182 |
|
| 183 |
if low_vram:
|
| 184 |
+
# The sampling process uses more vram, so we offload everything except two modules to the cpu.
|
| 185 |
models_to(models_rbm, device="cpu", excepts=["generator", "previewer"])
|
| 186 |
|
| 187 |
+
# Stage C reverse process.
|
| 188 |
+
with torch.cuda.amp.autocast(): # Use mixed precision
|
| 189 |
sampling_c = extras.gdf.sample(
|
| 190 |
models_rbm.generator, conditions, stage_c_latent_shape,
|
| 191 |
unconditions, device=device,
|
|
|
|
| 202 |
|
| 203 |
clear_gpu_cache() # Clear cache between stages
|
| 204 |
|
| 205 |
+
# Stage B reverse process.
|
| 206 |
+
with torch.no_grad(), torch.cuda.amp.autocast(dtype=torch.bfloat16):
|
| 207 |
+
conditions_b['effnet'] = sampled_c
|
| 208 |
+
unconditions_b['effnet'] = torch.zeros_like(sampled_c)
|
| 209 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
| 210 |
sampling_b = extras_b.gdf.sample(
|
| 211 |
models_b.generator, conditions_b, stage_b_latent_shape,
|
| 212 |
unconditions_b, device=device, **extras_b.sampling_configs,
|
| 213 |
)
|
| 214 |
for (sampled_b, _, _) in tqdm(sampling_b, total=extras_b.sampling_configs['timesteps']):
|
| 215 |
sampled_b = sampled_b
|
| 216 |
+
sampled = models_b.stage_a.decode(sampled_b).float()
|
| 217 |
|
|
|
|
| 218 |
sampled = torch.cat([
|
| 219 |
torch.nn.functional.interpolate(ref_style.cpu(), size=(height, width)),
|
| 220 |
sampled.cpu(),
|
|
|
|
| 234 |
|
| 235 |
return output_file # Return the path to the saved image
|
| 236 |
|
|
|
|
|
|
|
| 237 |
import gradio as gr
|
| 238 |
|
| 239 |
gr.Interface(
|