Update app.py
Browse files
app.py
CHANGED
@@ -5,7 +5,7 @@ from PIL import Image
|
|
5 |
from torchvision import transforms
|
6 |
import gradio as gr
|
7 |
|
8 |
-
# Load model
|
9 |
birefnet = AutoModelForImageSegmentation.from_pretrained('zhengpeng7/BiRefNet_lite', trust_remote_code=True)
|
10 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
11 |
birefnet.to(device)
|
@@ -19,72 +19,85 @@ transform_image = transforms.Compose([
|
|
19 |
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
|
20 |
])
|
21 |
|
22 |
-
def
|
|
|
23 |
inputs = []
|
24 |
original_sizes = []
|
25 |
-
for img in
|
26 |
original_sizes.append(img.size)
|
27 |
inputs.append(transform_image(img))
|
28 |
input_tensor = torch.stack(inputs).to(device)
|
29 |
|
|
|
30 |
try:
|
31 |
with torch.no_grad():
|
32 |
preds = birefnet(input_tensor)[-1].sigmoid().cpu()
|
33 |
except torch.OutOfMemoryError:
|
|
|
34 |
torch.cuda.empty_cache()
|
35 |
return None
|
36 |
|
|
|
37 |
results = []
|
38 |
-
for i, img in enumerate(
|
39 |
pred = preds[i].squeeze()
|
40 |
pred_pil = transforms.ToPILImage()(pred)
|
41 |
mask = pred_pil.resize(original_sizes[i])
|
42 |
result = Image.new("RGBA", original_sizes[i], (0, 0, 0, 0))
|
43 |
result.paste(img, mask=mask)
|
44 |
results.append(result)
|
45 |
-
|
46 |
return results
|
47 |
|
48 |
def extract_objects(filepaths):
|
49 |
-
# Open all images
|
50 |
-
images = [Image.open(
|
51 |
-
|
52 |
-
# You can define a batch size here (e.g., batch_size = 5)
|
53 |
-
# This prevents trying to process all images at once if too large
|
54 |
-
batch_size = 5
|
55 |
-
batches = [images[i:i+batch_size] for i in range(0, len(images), batch_size)]
|
56 |
|
57 |
total_start = time.time()
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
67 |
else:
|
68 |
-
|
69 |
-
|
70 |
-
batch_times.append(f"Batch {b_idx+1}: {(b_end - b_start):.2f} s")
|
71 |
-
total_end = time.time()
|
72 |
|
73 |
-
|
74 |
-
f"Total request time: {total_end - total_start:.2f} s\n"
|
75 |
-
"Batch times:\n" + "\n".join(batch_times)
|
76 |
-
)
|
77 |
|
78 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
|
80 |
iface = gr.Interface(
|
81 |
fn=extract_objects,
|
82 |
inputs=gr.Files(label="Upload Multiple Images", type="filepath", file_count="multiple"),
|
83 |
outputs=[gr.Gallery(label="Processed Images"), gr.Textbox(label="Timing Info")],
|
84 |
-
title="BiRefNet
|
85 |
-
description="Upload
|
86 |
)
|
87 |
|
88 |
-
|
89 |
-
# You can adjust concurrency_count and max_size as needed.
|
90 |
-
iface.queue(concurrency_count=1, max_size=10).launch()
|
|
|
5 |
from torchvision import transforms
|
6 |
import gradio as gr
|
7 |
|
8 |
+
# Load the model
|
9 |
birefnet = AutoModelForImageSegmentation.from_pretrained('zhengpeng7/BiRefNet_lite', trust_remote_code=True)
|
10 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
11 |
birefnet.to(device)
|
|
|
19 |
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
|
20 |
])
|
21 |
|
22 |
+
def try_inference(images):
|
23 |
+
# Convert images to tensors
|
24 |
inputs = []
|
25 |
original_sizes = []
|
26 |
+
for img in images:
|
27 |
original_sizes.append(img.size)
|
28 |
inputs.append(transform_image(img))
|
29 |
input_tensor = torch.stack(inputs).to(device)
|
30 |
|
31 |
+
# Attempt inference
|
32 |
try:
|
33 |
with torch.no_grad():
|
34 |
preds = birefnet(input_tensor)[-1].sigmoid().cpu()
|
35 |
except torch.OutOfMemoryError:
|
36 |
+
# Clear CUDA cache and return None to indicate OOM
|
37 |
torch.cuda.empty_cache()
|
38 |
return None
|
39 |
|
40 |
+
# Post-process
|
41 |
results = []
|
42 |
+
for i, img in enumerate(images):
|
43 |
pred = preds[i].squeeze()
|
44 |
pred_pil = transforms.ToPILImage()(pred)
|
45 |
mask = pred_pil.resize(original_sizes[i])
|
46 |
result = Image.new("RGBA", original_sizes[i], (0, 0, 0, 0))
|
47 |
result.paste(img, mask=mask)
|
48 |
results.append(result)
|
|
|
49 |
return results
|
50 |
|
51 |
def extract_objects(filepaths):
|
52 |
+
# Open all images
|
53 |
+
images = [Image.open(p).convert("RGB") for p in filepaths]
|
|
|
|
|
|
|
|
|
|
|
54 |
|
55 |
total_start = time.time()
|
56 |
+
# If you have N images, start by trying them all
|
57 |
+
low = 1
|
58 |
+
high = len(images)
|
59 |
+
best = None
|
60 |
+
best_batch_size = 0
|
61 |
+
|
62 |
+
# Binary search to find max batch size that doesn't OOM
|
63 |
+
while low <= high:
|
64 |
+
mid = (low + high) // 2
|
65 |
+
batch_test = images[:mid]
|
66 |
+
|
67 |
+
start = time.time()
|
68 |
+
results = try_inference(batch_test)
|
69 |
+
end = time.time()
|
70 |
+
|
71 |
+
if results is not None:
|
72 |
+
# Succeeded with 'mid' batch size
|
73 |
+
best = results
|
74 |
+
best_batch_size = mid
|
75 |
+
low = mid + 1 # try a bigger batch
|
76 |
else:
|
77 |
+
# OOM, try smaller batch
|
78 |
+
high = mid - 1
|
|
|
|
|
79 |
|
80 |
+
total_end = time.time()
|
|
|
|
|
|
|
81 |
|
82 |
+
if best is None:
|
83 |
+
# Even a single image caused OOM
|
84 |
+
summary = "Could not process even a single image without OOM."
|
85 |
+
return [], summary
|
86 |
+
else:
|
87 |
+
# Process the final chosen batch size fully (we already have results)
|
88 |
+
summary = (
|
89 |
+
f"Total request time: {total_end - total_start:.2f} s\n"
|
90 |
+
f"Successfully processed {best_batch_size} images in a single batch.\n"
|
91 |
+
f"Could not handle more than {best_batch_size} images without OOM."
|
92 |
+
)
|
93 |
+
return best, summary
|
94 |
|
95 |
iface = gr.Interface(
|
96 |
fn=extract_objects,
|
97 |
inputs=gr.Files(label="Upload Multiple Images", type="filepath", file_count="multiple"),
|
98 |
outputs=[gr.Gallery(label="Processed Images"), gr.Textbox(label="Timing Info")],
|
99 |
+
title="BiRefNet Dynamic Batch OOM Test",
|
100 |
+
description="Upload images. The system will try to process all at once, and if OOM occurs, it will try smaller batches automatically, quickly finding the largest feasible batch size."
|
101 |
)
|
102 |
|
103 |
+
iface.launch()
|
|
|
|