Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -5,6 +5,12 @@ from decord import VideoReader, cpu
|
|
5 |
from PIL import Image
|
6 |
from torchvision.transforms.functional import InterpolationMode
|
7 |
from transformers import AutoModel, AutoTokenizer
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
# Device Configuration
|
10 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
@@ -22,21 +28,6 @@ def build_transform(input_size):
|
|
22 |
])
|
23 |
return transform
|
24 |
|
25 |
-
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
|
26 |
-
best_ratio_diff = float('inf')
|
27 |
-
best_ratio = (1, 1)
|
28 |
-
area = width * height
|
29 |
-
for ratio in target_ratios:
|
30 |
-
target_aspect_ratio = ratio[0] / ratio[1]
|
31 |
-
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
|
32 |
-
if ratio_diff < best_ratio_diff:
|
33 |
-
best_ratio_diff = ratio_diff
|
34 |
-
best_ratio = ratio
|
35 |
-
elif ratio_diff == best_ratio_diff:
|
36 |
-
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
|
37 |
-
best_ratio = ratio
|
38 |
-
return best_ratio
|
39 |
-
|
40 |
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
|
41 |
orig_width, orig_height = image.size
|
42 |
aspect_ratio = orig_width / orig_height
|
@@ -46,16 +37,11 @@ def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbna
|
|
46 |
i * j <= max_num and i * j >= min_num)
|
47 |
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
|
48 |
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
target_width = image_size * target_aspect_ratio[0]
|
53 |
-
target_height = image_size * target_aspect_ratio[1]
|
54 |
-
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
|
55 |
-
|
56 |
resized_img = image.resize((target_width, target_height))
|
57 |
processed_images = []
|
58 |
-
for i in range(
|
59 |
box = (
|
60 |
(i % (target_width // image_size)) * image_size,
|
61 |
(i // (target_width // image_size)) * image_size,
|
@@ -64,13 +50,12 @@ def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbna
|
|
64 |
)
|
65 |
split_img = resized_img.crop(box)
|
66 |
processed_images.append(split_img)
|
67 |
-
assert len(processed_images) == blocks
|
68 |
if use_thumbnail and len(processed_images) != 1:
|
69 |
thumbnail_img = image.resize((image_size, image_size))
|
70 |
processed_images.append(thumbnail_img)
|
71 |
return processed_images
|
72 |
|
73 |
-
def load_image(image_file, input_size=448, max_num=12):
|
74 |
image = Image.open(image_file).convert('RGB')
|
75 |
transform = build_transform(input_size=input_size)
|
76 |
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
|
@@ -78,38 +63,6 @@ def load_image(image_file, input_size=448, max_num=12):
|
|
78 |
pixel_values = torch.stack(pixel_values).to(device)
|
79 |
return pixel_values
|
80 |
|
81 |
-
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
|
82 |
-
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
|
83 |
-
max_frame = len(vr) - 1
|
84 |
-
fps = float(vr.get_avg_fps())
|
85 |
-
|
86 |
-
pixel_values_list, num_patches_list = [], []
|
87 |
-
transform = build_transform(input_size=input_size)
|
88 |
-
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
|
89 |
-
for frame_index in frame_indices:
|
90 |
-
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
|
91 |
-
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
|
92 |
-
pixel_values = [transform(tile) for tile in img]
|
93 |
-
pixel_values = torch.stack(pixel_values)
|
94 |
-
num_patches_list.append(pixel_values.shape[0])
|
95 |
-
pixel_values_list.append(pixel_values)
|
96 |
-
pixel_values = torch.cat(pixel_values_list)
|
97 |
-
return pixel_values, num_patches_list
|
98 |
-
|
99 |
-
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
|
100 |
-
if bound:
|
101 |
-
start, end = bound[0], bound[1]
|
102 |
-
else:
|
103 |
-
start, end = -100000, 100000
|
104 |
-
start_idx = max(first_idx, round(start * fps))
|
105 |
-
end_idx = min(round(end * fps), max_frame)
|
106 |
-
seg_size = float(end_idx - start_idx) / num_segments
|
107 |
-
frame_indices = np.array([
|
108 |
-
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
|
109 |
-
for idx in range(num_segments)
|
110 |
-
])
|
111 |
-
return frame_indices
|
112 |
-
|
113 |
# Load Model
|
114 |
path = 'OpenGVLab/InternVL2_5-1B'
|
115 |
model = AutoModel.from_pretrained(
|
@@ -119,3 +72,14 @@ model = AutoModel.from_pretrained(
|
|
119 |
trust_remote_code=True
|
120 |
).eval().to(device)
|
121 |
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
from PIL import Image
|
6 |
from torchvision.transforms.functional import InterpolationMode
|
7 |
from transformers import AutoModel, AutoTokenizer
|
8 |
+
from fastapi import FastAPI, UploadFile, File
|
9 |
+
from typing import List
|
10 |
+
from io import BytesIO
|
11 |
+
|
12 |
+
# FastAPI app initialization
|
13 |
+
app = FastAPI()
|
14 |
|
15 |
# Device Configuration
|
16 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
28 |
])
|
29 |
return transform
|
30 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
|
32 |
orig_width, orig_height = image.size
|
33 |
aspect_ratio = orig_width / orig_height
|
|
|
37 |
i * j <= max_num and i * j >= min_num)
|
38 |
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
|
39 |
|
40 |
+
target_width = image_size * target_ratios[0][0]
|
41 |
+
target_height = image_size * target_ratios[0][1]
|
|
|
|
|
|
|
|
|
|
|
42 |
resized_img = image.resize((target_width, target_height))
|
43 |
processed_images = []
|
44 |
+
for i in range(target_ratios[0][0] * target_ratios[0][1]):
|
45 |
box = (
|
46 |
(i % (target_width // image_size)) * image_size,
|
47 |
(i // (target_width // image_size)) * image_size,
|
|
|
50 |
)
|
51 |
split_img = resized_img.crop(box)
|
52 |
processed_images.append(split_img)
|
|
|
53 |
if use_thumbnail and len(processed_images) != 1:
|
54 |
thumbnail_img = image.resize((image_size, image_size))
|
55 |
processed_images.append(thumbnail_img)
|
56 |
return processed_images
|
57 |
|
58 |
+
def load_image(image_file: BytesIO, input_size=448, max_num=12):
|
59 |
image = Image.open(image_file).convert('RGB')
|
60 |
transform = build_transform(input_size=input_size)
|
61 |
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
|
|
|
63 |
pixel_values = torch.stack(pixel_values).to(device)
|
64 |
return pixel_values
|
65 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
# Load Model
|
67 |
path = 'OpenGVLab/InternVL2_5-1B'
|
68 |
model = AutoModel.from_pretrained(
|
|
|
72 |
trust_remote_code=True
|
73 |
).eval().to(device)
|
74 |
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
|
75 |
+
|
76 |
+
@app.post("/predict")
|
77 |
+
async def predict(file: UploadFile = File(...), question: str = "Describe the image"):
|
78 |
+
# Load and preprocess the image
|
79 |
+
file_bytes = BytesIO(await file.read())
|
80 |
+
pixel_values = load_image(file_bytes)
|
81 |
+
|
82 |
+
# Generate a response
|
83 |
+
generation_config = dict(max_new_tokens=1024, do_sample=True)
|
84 |
+
response, _ = model.chat(tokenizer, pixel_values, question, generation_config)
|
85 |
+
return {"question": question, "response": response}
|