Update app.py
Browse files
app.py
CHANGED
@@ -1,21 +1,17 @@
|
|
1 |
import streamlit as st
|
2 |
import torch
|
3 |
from PIL import Image
|
4 |
-
import os
|
5 |
import gc
|
6 |
-
from transformers import AutoProcessor
|
7 |
from peft import PeftModel
|
|
|
8 |
|
9 |
-
#
|
10 |
-
st.set_page_config(
|
11 |
-
page_title="Deepfake Image Analyzer",
|
12 |
-
page_icon="π",
|
13 |
-
layout="wide"
|
14 |
-
)
|
15 |
|
16 |
-
#
|
17 |
st.title("Deepfake Image Analyzer")
|
18 |
-
st.markdown("
|
19 |
|
20 |
# Function to free up memory
|
21 |
def free_memory():
|
@@ -23,51 +19,41 @@ def free_memory():
|
|
23 |
if torch.cuda.is_available():
|
24 |
torch.cuda.empty_cache()
|
25 |
|
26 |
-
#
|
27 |
-
def
|
28 |
-
if
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
|
|
|
|
37 |
|
|
|
38 |
@st.cache_resource
|
39 |
def load_model():
|
40 |
-
"""Load
|
41 |
try:
|
42 |
-
# Using your original base model
|
43 |
base_model_id = "unsloth/llama-3.2-11b-vision-instruct-unsloth-bnb-4bit"
|
44 |
|
45 |
# Load processor
|
46 |
processor = AutoProcessor.from_pretrained(base_model_id)
|
47 |
|
48 |
-
#
|
49 |
-
|
50 |
-
load_in_4bit=True,
|
51 |
-
bnb_4bit_compute_dtype=torch.float16,
|
52 |
-
bnb_4bit_use_double_quant=True,
|
53 |
-
bnb_4bit_quant_type="nf4",
|
54 |
-
bnb_4bit_quant_storage=torch.float16,
|
55 |
-
llm_int8_skip_modules=["lm_head"],
|
56 |
-
llm_int8_enable_fp32_cpu_offload=True
|
57 |
-
)
|
58 |
-
|
59 |
-
# Load the pre-quantized model with unsloth settings
|
60 |
-
model = AutoModelForCausalLM.from_pretrained(
|
61 |
base_model_id,
|
62 |
-
|
63 |
-
quantization_config=quantization_config,
|
64 |
torch_dtype=torch.float16,
|
65 |
-
|
66 |
-
low_cpu_mem_usage=True,
|
67 |
-
use_cache=True,
|
68 |
-
offload_folder="offload" # Enable disk offloading
|
69 |
)
|
70 |
|
|
|
|
|
|
|
71 |
# Load adapter
|
72 |
adapter_id = "saakshigupta/deepfake-explainer-1"
|
73 |
model = PeftModel.from_pretrained(model, adapter_id)
|
@@ -79,144 +65,110 @@ def load_model():
|
|
79 |
st.exception(e)
|
80 |
return None, None
|
81 |
|
82 |
-
#
|
83 |
-
def fix_processor_outputs(inputs):
|
84 |
-
"""Fix cross-attention mask dimensions if needed"""
|
85 |
-
if 'cross_attention_mask' in inputs and 0 in inputs['cross_attention_mask'].shape:
|
86 |
-
batch_size, seq_len, _, num_tiles = inputs['cross_attention_mask'].shape
|
87 |
-
visual_features = 6404 # The exact dimension used in training
|
88 |
-
new_mask = torch.ones(
|
89 |
-
(batch_size, seq_len, visual_features, num_tiles),
|
90 |
-
device=inputs['cross_attention_mask'].device
|
91 |
-
)
|
92 |
-
inputs['cross_attention_mask'] = new_mask
|
93 |
-
return True, inputs
|
94 |
-
return False, inputs
|
95 |
-
|
96 |
-
# Create sidebar with options
|
97 |
with st.sidebar:
|
98 |
-
st.header("
|
99 |
-
temperature = st.slider("Temperature",
|
100 |
-
|
101 |
-
max_length = st.slider("Maximum response length", min_value=100, max_value=1000, value=500, step=50)
|
102 |
|
103 |
-
|
104 |
-
|
105 |
-
|
|
|
106 |
height=100
|
107 |
)
|
108 |
-
|
109 |
-
st.markdown("### About")
|
110 |
-
st.markdown("""
|
111 |
-
This app uses a fine-tuned Llama 3.2 Vision model to detect and explain deepfakes.
|
112 |
-
|
113 |
-
The analyzer looks for:
|
114 |
-
- Inconsistencies in facial features
|
115 |
-
- Unusual lighting or shadows
|
116 |
-
- Unnatural blur patterns
|
117 |
-
- Artifacts around edges
|
118 |
-
- Texture inconsistencies
|
119 |
-
|
120 |
-
Model by [saakshigupta](https://huggingface.co/saakshigupta/deepfake-explainer-1)
|
121 |
-
""")
|
122 |
|
123 |
-
#
|
124 |
-
|
125 |
-
|
126 |
-
|
|
|
|
|
|
|
127 |
model, processor = load_model()
|
128 |
if model is not None and processor is not None:
|
129 |
st.session_state['model'] = model
|
130 |
st.session_state['processor'] = processor
|
131 |
-
st.success("Model loaded successfully!")
|
132 |
else:
|
133 |
-
st.error("Failed to load model
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
# Main content area - file uploader
|
139 |
-
uploaded_file = st.file_uploader("Choose an image...", type=["jpg", "jpeg", "png"])
|
140 |
-
|
141 |
-
# Check if model is loaded
|
142 |
-
model_loaded = 'model' in st.session_state and st.session_state['model'] is not None
|
143 |
-
|
144 |
-
if uploaded_file is not None:
|
145 |
-
# Display the image
|
146 |
-
image = Image.open(uploaded_file).convert('RGB')
|
147 |
-
st.image(image, caption="Uploaded Image", use_column_width=True)
|
148 |
|
149 |
-
#
|
150 |
-
if
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
st.
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
|
|
|
|
178 |
)
|
179 |
-
|
180 |
-
# Decode the output
|
181 |
-
response = processor.decode(output_ids[0], skip_special_tokens=True)
|
182 |
-
|
183 |
-
# Extract the actual response (removing the prompt)
|
184 |
-
if custom_prompt in response:
|
185 |
-
result = response.split(custom_prompt)[-1].strip()
|
186 |
-
else:
|
187 |
-
result = response
|
188 |
-
|
189 |
-
# Display result in a nice format
|
190 |
-
st.success("Analysis complete!")
|
191 |
-
|
192 |
-
# Show technical and non-technical explanations separately if they exist
|
193 |
-
if "Technical Explanation:" in result and "Non-Technical Explanation:" in result:
|
194 |
-
technical, non_technical = result.split("Non-Technical Explanation:")
|
195 |
-
technical = technical.replace("Technical Explanation:", "").strip()
|
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 |
-
st.
|
222 |
-
st.
|
|
|
1 |
import streamlit as st
|
2 |
import torch
|
3 |
from PIL import Image
|
|
|
4 |
import gc
|
5 |
+
from transformers import AutoProcessor
|
6 |
from peft import PeftModel
|
7 |
+
from unsloth import FastVisionModel
|
8 |
|
9 |
+
# Simple page config
|
10 |
+
st.set_page_config(page_title="Deepfake Analyzer", layout="wide")
|
|
|
|
|
|
|
|
|
11 |
|
12 |
+
# Minimal UI
|
13 |
st.title("Deepfake Image Analyzer")
|
14 |
+
st.markdown("This app analyzes images for signs of deepfake manipulation")
|
15 |
|
16 |
# Function to free up memory
|
17 |
def free_memory():
|
|
|
19 |
if torch.cuda.is_available():
|
20 |
torch.cuda.empty_cache()
|
21 |
|
22 |
+
# Function to fix cross-attention masks
|
23 |
+
def fix_processor_outputs(inputs):
|
24 |
+
"""Fix cross-attention mask dimensions if needed"""
|
25 |
+
if 'cross_attention_mask' in inputs and 0 in inputs['cross_attention_mask'].shape:
|
26 |
+
batch_size, seq_len, _, num_tiles = inputs['cross_attention_mask'].shape
|
27 |
+
visual_features = 6404 # The exact dimension used in training
|
28 |
+
new_mask = torch.ones(
|
29 |
+
(batch_size, seq_len, visual_features, num_tiles),
|
30 |
+
device=inputs['cross_attention_mask'].device
|
31 |
+
)
|
32 |
+
inputs['cross_attention_mask'] = new_mask
|
33 |
+
return True, inputs
|
34 |
+
return False, inputs
|
35 |
|
36 |
+
# Load model function
|
37 |
@st.cache_resource
|
38 |
def load_model():
|
39 |
+
"""Load model using Unsloth approach (similar to Colab)"""
|
40 |
try:
|
|
|
41 |
base_model_id = "unsloth/llama-3.2-11b-vision-instruct-unsloth-bnb-4bit"
|
42 |
|
43 |
# Load processor
|
44 |
processor = AutoProcessor.from_pretrained(base_model_id)
|
45 |
|
46 |
+
# Load model using Unsloth's FastVisionModel
|
47 |
+
model, _ = FastVisionModel.from_pretrained(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
base_model_id,
|
49 |
+
load_in_4bit=True,
|
|
|
50 |
torch_dtype=torch.float16,
|
51 |
+
device_map="auto"
|
|
|
|
|
|
|
52 |
)
|
53 |
|
54 |
+
# Set to inference mode
|
55 |
+
FastVisionModel.for_inference(model)
|
56 |
+
|
57 |
# Load adapter
|
58 |
adapter_id = "saakshigupta/deepfake-explainer-1"
|
59 |
model = PeftModel.from_pretrained(model, adapter_id)
|
|
|
65 |
st.exception(e)
|
66 |
return None, None
|
67 |
|
68 |
+
# Minimal sidebar
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
with st.sidebar:
|
70 |
+
st.header("Settings")
|
71 |
+
temperature = st.slider("Temperature", 0.1, 1.0, 0.7, 0.1)
|
72 |
+
max_length = st.slider("Max length", 100, 500, 300, 50)
|
|
|
73 |
|
74 |
+
# Instruction field
|
75 |
+
prompt = st.text_area(
|
76 |
+
"Analysis instruction",
|
77 |
+
value="Analyze this image and determine if it's a deepfake. Provide your reasoning.",
|
78 |
height=100
|
79 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
80 |
|
81 |
+
# Main content - two columns for clarity
|
82 |
+
col1, col2 = st.columns([1, 2])
|
83 |
+
|
84 |
+
with col1:
|
85 |
+
# Load model button
|
86 |
+
if st.button("1. Load Model"):
|
87 |
+
with st.spinner("Loading model... (this may take a minute)"):
|
88 |
model, processor = load_model()
|
89 |
if model is not None and processor is not None:
|
90 |
st.session_state['model'] = model
|
91 |
st.session_state['processor'] = processor
|
92 |
+
st.success("β Model loaded successfully!")
|
93 |
else:
|
94 |
+
st.error("Failed to load model")
|
95 |
+
|
96 |
+
# File uploader
|
97 |
+
uploaded_file = st.file_uploader("2. Upload an image", type=["jpg", "jpeg", "png"])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
|
99 |
+
# Display uploaded image
|
100 |
+
if uploaded_file is not None:
|
101 |
+
image = Image.open(uploaded_file).convert('RGB')
|
102 |
+
st.image(image, caption="Uploaded Image", use_column_width=True)
|
103 |
+
|
104 |
+
# Only enable analysis if model is loaded
|
105 |
+
model_loaded = 'model' in st.session_state and st.session_state['model'] is not None
|
106 |
+
|
107 |
+
if st.button("3. Analyze Image", disabled=not model_loaded):
|
108 |
+
if not model_loaded:
|
109 |
+
st.warning("Please load the model first")
|
110 |
+
else:
|
111 |
+
col2.subheader("Analysis Results")
|
112 |
+
with col2.spinner("Analyzing image..."):
|
113 |
+
try:
|
114 |
+
# Get model components
|
115 |
+
model = st.session_state['model']
|
116 |
+
processor = st.session_state['processor']
|
117 |
+
|
118 |
+
# Format message for analysis
|
119 |
+
messages = [
|
120 |
+
{"role": "user", "content": [
|
121 |
+
{"type": "image"},
|
122 |
+
{"type": "text", "text": prompt}
|
123 |
+
]}
|
124 |
+
]
|
125 |
+
|
126 |
+
# Apply chat template
|
127 |
+
input_text = processor.tokenizer.apply_chat_template(
|
128 |
+
messages,
|
129 |
+
add_generation_prompt=True
|
130 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
131 |
|
132 |
+
# Process with image
|
133 |
+
inputs = processor(
|
134 |
+
images=image,
|
135 |
+
text=input_text,
|
136 |
+
add_special_tokens=False,
|
137 |
+
return_tensors="pt"
|
138 |
+
).to(model.device)
|
139 |
|
140 |
+
# Apply the fix
|
141 |
+
fixed, inputs = fix_processor_outputs(inputs)
|
142 |
+
if fixed:
|
143 |
+
col2.info("Fixed cross-attention mask dimensions")
|
144 |
+
|
145 |
+
# Generate analysis
|
146 |
+
with torch.no_grad():
|
147 |
+
output_ids = model.generate(
|
148 |
+
**inputs,
|
149 |
+
max_new_tokens=max_length,
|
150 |
+
temperature=temperature,
|
151 |
+
top_p=0.9
|
152 |
+
)
|
153 |
+
|
154 |
+
# Decode the output
|
155 |
+
response = processor.tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
156 |
+
|
157 |
+
# Display results
|
158 |
+
col2.success("Analysis complete!")
|
159 |
+
col2.markdown(response)
|
160 |
+
|
161 |
+
# Free memory
|
162 |
+
free_memory()
|
163 |
+
|
164 |
+
except Exception as e:
|
165 |
+
col2.error(f"Error analyzing image: {str(e)}")
|
166 |
+
col2.exception(e)
|
167 |
+
elif not model_loaded:
|
168 |
+
st.info("Please load the model first (Step 1)")
|
169 |
+
else:
|
170 |
+
st.info("Please upload an image (Step 2)")
|
171 |
|
172 |
+
with col2:
|
173 |
+
if 'model' not in st.session_state:
|
174 |
+
st.info("π Follow the steps on the left to analyze an image")
|