Spaces:
Runtime error
Runtime error
Upload 6 files
Browse files
models/__init__.py
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .model_loader import ModelManager
|
2 |
+
from .vqa_inference import VQAInference
|
3 |
+
|
4 |
+
__all__ = ["ModelManager", "VQAInference"]
|
models/__pycache__/__init__.cpython-311.pyc
ADDED
Binary file (325 Bytes). View file
|
|
models/__pycache__/model_loader.cpython-311.pyc
ADDED
Binary file (3.92 kB). View file
|
|
models/__pycache__/vqa_inference.cpython-311.pyc
ADDED
Binary file (7.98 kB). View file
|
|
models/model_loader.py
ADDED
@@ -0,0 +1,98 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from transformers import (
|
3 |
+
BlipForQuestionAnswering,
|
4 |
+
BlipProcessor,
|
5 |
+
ViltForQuestionAnswering,
|
6 |
+
ViltProcessor,
|
7 |
+
)
|
8 |
+
|
9 |
+
|
10 |
+
class ModelManager:
|
11 |
+
"""
|
12 |
+
Class to manage loading and caching of various VQA models from Hugging Face
|
13 |
+
"""
|
14 |
+
|
15 |
+
def __init__(self, cache_dir=None):
|
16 |
+
"""
|
17 |
+
Initialize the model manager
|
18 |
+
|
19 |
+
Args:
|
20 |
+
cache_dir (str, optional): Directory to cache models. Defaults to None.
|
21 |
+
"""
|
22 |
+
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
23 |
+
self.cache_dir = cache_dir
|
24 |
+
self.models = {}
|
25 |
+
self.processors = {}
|
26 |
+
|
27 |
+
# Print device being used
|
28 |
+
print(f"Using device: {self.device}")
|
29 |
+
|
30 |
+
def load_blip(self):
|
31 |
+
"""
|
32 |
+
Load BLIP model for VQA
|
33 |
+
|
34 |
+
Returns:
|
35 |
+
tuple: (processor, model)
|
36 |
+
"""
|
37 |
+
if "blip" not in self.models:
|
38 |
+
print("Loading BLIP model for visual question answering...")
|
39 |
+
|
40 |
+
# Load processor and model
|
41 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base")
|
42 |
+
model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base")
|
43 |
+
|
44 |
+
# Move model to appropriate device
|
45 |
+
model.to(self.device)
|
46 |
+
|
47 |
+
# Store model and processor
|
48 |
+
self.models["blip"] = model
|
49 |
+
self.processors["blip"] = processor
|
50 |
+
|
51 |
+
print("BLIP model loaded successfully!")
|
52 |
+
|
53 |
+
return self.processors["blip"], self.models["blip"]
|
54 |
+
|
55 |
+
def load_vilt(self):
|
56 |
+
"""
|
57 |
+
Load ViLT model for VQA
|
58 |
+
|
59 |
+
Returns:
|
60 |
+
tuple: (processor, model)
|
61 |
+
"""
|
62 |
+
if "vilt" not in self.models:
|
63 |
+
print("Loading ViLT model for visual question answering...")
|
64 |
+
|
65 |
+
# Load processor and model
|
66 |
+
processor = ViltProcessor.from_pretrained("dandelin/vilt-b32-vqa")
|
67 |
+
model = ViltForQuestionAnswering.from_pretrained("dandelin/vilt-b32-vqa")
|
68 |
+
|
69 |
+
# Move model to appropriate device
|
70 |
+
model.to(self.device)
|
71 |
+
|
72 |
+
# Store model and processor
|
73 |
+
self.models["vilt"] = model
|
74 |
+
self.processors["vilt"] = processor
|
75 |
+
|
76 |
+
print("ViLT model loaded successfully!")
|
77 |
+
|
78 |
+
return self.processors["vilt"], self.models["vilt"]
|
79 |
+
|
80 |
+
def get_model(self, model_name="blip"):
|
81 |
+
"""
|
82 |
+
Get a model by name
|
83 |
+
|
84 |
+
Args:
|
85 |
+
model_name (str, optional): Name of model to load. Defaults to "blip".
|
86 |
+
Options: "blip", "vilt"
|
87 |
+
|
88 |
+
Returns:
|
89 |
+
tuple: (processor, model)
|
90 |
+
"""
|
91 |
+
if model_name.lower() == "blip":
|
92 |
+
return self.load_blip()
|
93 |
+
elif model_name.lower() == "vilt":
|
94 |
+
return self.load_vilt()
|
95 |
+
else:
|
96 |
+
raise ValueError(
|
97 |
+
f"Unknown model: {model_name}. Available models: blip, vilt"
|
98 |
+
)
|
models/vqa_inference.py
ADDED
@@ -0,0 +1,152 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import io
|
2 |
+
import os
|
3 |
+
import traceback
|
4 |
+
import torch
|
5 |
+
from PIL import Image, UnidentifiedImageError
|
6 |
+
from .model_loader import ModelManager
|
7 |
+
|
8 |
+
|
9 |
+
class VQAInference:
|
10 |
+
"""
|
11 |
+
Class to perform inference with Visual Question Answering models
|
12 |
+
"""
|
13 |
+
|
14 |
+
def __init__(self, model_name="blip", cache_dir=None):
|
15 |
+
"""
|
16 |
+
Initialize the VQA inference
|
17 |
+
|
18 |
+
Args:
|
19 |
+
model_name (str, optional): Name of model to use. Defaults to "blip".
|
20 |
+
cache_dir (str, optional): Directory to cache models. Defaults to None.
|
21 |
+
"""
|
22 |
+
self.model_name = model_name
|
23 |
+
self.model_manager = ModelManager(cache_dir=cache_dir)
|
24 |
+
self.processor, self.model = self.model_manager.get_model(model_name)
|
25 |
+
self.device = self.model_manager.device
|
26 |
+
|
27 |
+
def predict(self, image, question):
|
28 |
+
"""
|
29 |
+
Perform VQA prediction on an image with a question
|
30 |
+
|
31 |
+
Args:
|
32 |
+
image (PIL.Image.Image or str): Image to analyze or path to image
|
33 |
+
question (str): Question to ask about the image
|
34 |
+
|
35 |
+
Returns:
|
36 |
+
str: Answer to the question
|
37 |
+
"""
|
38 |
+
# Handle image input - could be a file path or PIL Image
|
39 |
+
if isinstance(image, str):
|
40 |
+
try:
|
41 |
+
# Check if file exists
|
42 |
+
if not os.path.exists(image):
|
43 |
+
raise FileNotFoundError(f"Image file not found: {image}")
|
44 |
+
|
45 |
+
# Try multiple approaches to load the image
|
46 |
+
try:
|
47 |
+
# Try the standard approach first
|
48 |
+
image = Image.open(image).convert("RGB")
|
49 |
+
print(
|
50 |
+
f"Successfully opened image: {image.size}, mode: {image.mode}"
|
51 |
+
)
|
52 |
+
except Exception as img_err:
|
53 |
+
print(
|
54 |
+
f"Standard image loading failed: {img_err}, trying alternative method..."
|
55 |
+
)
|
56 |
+
|
57 |
+
# Try alternative approach with binary mode explicitly
|
58 |
+
with open(image, "rb") as img_file:
|
59 |
+
img_data = img_file.read()
|
60 |
+
image = Image.open(io.BytesIO(img_data)).convert("RGB")
|
61 |
+
print(
|
62 |
+
f"Alternative image loading succeeded: {image.size}, mode: {image.mode}"
|
63 |
+
)
|
64 |
+
|
65 |
+
except UnidentifiedImageError as e:
|
66 |
+
# Specific error when image format cannot be identified
|
67 |
+
raise ValueError(f"Cannot identify image format: {str(e)}")
|
68 |
+
except Exception as e:
|
69 |
+
# Provide detailed error information
|
70 |
+
error_details = traceback.format_exc()
|
71 |
+
print(f"Error details: {error_details}")
|
72 |
+
raise ValueError(f"Could not open image file: {str(e)}")
|
73 |
+
|
74 |
+
# Make sure image is a PIL Image
|
75 |
+
if not isinstance(image, Image.Image):
|
76 |
+
raise ValueError("Image must be a PIL Image or a file path")
|
77 |
+
|
78 |
+
# Process based on model type
|
79 |
+
if self.model_name.lower() == "blip":
|
80 |
+
return self._predict_with_blip(image, question)
|
81 |
+
elif self.model_name.lower() == "vilt":
|
82 |
+
return self._predict_with_vilt(image, question)
|
83 |
+
else:
|
84 |
+
raise ValueError(f"Prediction not implemented for model: {self.model_name}")
|
85 |
+
|
86 |
+
def _predict_with_blip(self, image, question):
|
87 |
+
"""
|
88 |
+
Perform prediction with BLIP model
|
89 |
+
|
90 |
+
Args:
|
91 |
+
image (PIL.Image.Image): Image to analyze
|
92 |
+
question (str): Question to ask about the image
|
93 |
+
|
94 |
+
Returns:
|
95 |
+
str: Answer to the question
|
96 |
+
"""
|
97 |
+
try:
|
98 |
+
# Process image and text inputs
|
99 |
+
inputs = self.processor(
|
100 |
+
images=image, text=question, return_tensors="pt"
|
101 |
+
).to(self.device)
|
102 |
+
|
103 |
+
# Generate answer
|
104 |
+
with torch.no_grad():
|
105 |
+
outputs = self.model.generate(**inputs)
|
106 |
+
|
107 |
+
# Decode the output to text
|
108 |
+
answer = self.processor.decode(outputs[0], skip_special_tokens=True)
|
109 |
+
|
110 |
+
return answer
|
111 |
+
except Exception as e:
|
112 |
+
error_details = traceback.format_exc()
|
113 |
+
print(f"Error in BLIP prediction: {str(e)}")
|
114 |
+
print(f"Error details: {error_details}")
|
115 |
+
raise RuntimeError(f"BLIP model prediction failed: {str(e)}")
|
116 |
+
|
117 |
+
def _predict_with_vilt(self, image, question):
|
118 |
+
"""
|
119 |
+
Perform prediction with ViLT model
|
120 |
+
|
121 |
+
Args:
|
122 |
+
image (PIL.Image.Image): Image to analyze
|
123 |
+
question (str): Question to ask about the image
|
124 |
+
|
125 |
+
Returns:
|
126 |
+
str: Answer to the question
|
127 |
+
"""
|
128 |
+
try:
|
129 |
+
# Process image and text inputs
|
130 |
+
encoding = self.processor(images=image, text=question, return_tensors="pt")
|
131 |
+
|
132 |
+
# Move inputs to device
|
133 |
+
for k, v in encoding.items():
|
134 |
+
encoding[k] = v.to(self.device)
|
135 |
+
|
136 |
+
# Forward pass
|
137 |
+
with torch.no_grad():
|
138 |
+
outputs = self.model(**encoding)
|
139 |
+
logits = outputs.logits
|
140 |
+
|
141 |
+
# Get the predicted answer idx
|
142 |
+
idx = logits.argmax(-1).item()
|
143 |
+
|
144 |
+
# Convert to answer text
|
145 |
+
answer = self.model.config.id2label[idx]
|
146 |
+
|
147 |
+
return answer
|
148 |
+
except Exception as e:
|
149 |
+
error_details = traceback.format_exc()
|
150 |
+
print(f"Error in ViLT prediction: {str(e)}")
|
151 |
+
print(f"Error details: {error_details}")
|
152 |
+
raise RuntimeError(f"ViLT model prediction failed: {str(e)}")
|