VLM-Lens / demo /launch_gradio.py
marstin's picture
[martin-dev] add links
9373294
"""Gradio demo for visualizing VLM first token probability distributions with two images."""
from typing import Any, Dict, List, Optional, Tuple
import gradio as gr
from spaces import GPU
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn.functional as F
from matplotlib.figure import Figure
from matplotlib.text import Text
from PIL import Image
from demo.lookup import ModelVariants, get_model_info # noqa: E402
from src.main import get_model # noqa: E402
from src.models.base import ModelBase # noqa: E402
from src.models.config import Config, ModelSelection # noqa: E402
models_cache: Dict[str, Any] = {}
current_model_selection: Optional[ModelSelection] = None
def read_layer_spec(spec_file_path: str) -> List[str]:
"""Read available layers from the model spec file.
Args:
spec_file_path: Path to the model specification file.
Returns:
List of available layer names, skipping blank lines.
"""
try:
with open(spec_file_path, 'r', encoding='utf-8') as f:
lines = f.readlines()
# Filter out blank lines and strip whitespace
layers = [line.strip() for line in lines if line.strip()]
return layers
except FileNotFoundError:
print(f'Spec file not found: {spec_file_path}')
return ['Default layer (spec file not found)']
except Exception as e:
print(f'Error reading spec file: {str(e)}')
return ['Default layer (error reading spec)']
def update_layer_choices(model_choice: str) -> Tuple[gr.Dropdown, gr.Button]:
"""Update layer dropdown choices based on selected model.
Args:
model_choice: Selected model name.
Returns:
Updated dropdown component and button visibility.
"""
if not model_choice:
return gr.Dropdown(choices=[], visible=False), gr.Button(visible=False)
try:
# Convert string choice to ModelVariants enum
model_var = ModelVariants(model_choice.lower())
# Get model info and read layer spec
_, _, model_spec_path = get_model_info(model_var)
layers = read_layer_spec(model_spec_path)
# Return updated dropdown with layer choices and make button visible
return (
gr.Dropdown(
choices=layers,
label=f'Select Module for {model_choice}',
value=layers[0] if layers else None,
visible=True,
interactive=True
),
gr.Button('Analyze', variant='primary', visible=True)
)
except ValueError:
return (
gr.Dropdown(
choices=['Model not implemented'],
label='Select Module',
visible=True,
interactive=False
),
gr.Button('Analyze', variant='primary', visible=False)
)
except Exception as e:
return (
gr.Dropdown(
choices=[f'Error: {str(e)}'],
label='Select Module',
visible=True,
interactive=False
),
gr.Button('Analyze', variant='primary', visible=False)
)
def load_model(model_var: ModelVariants, config: Config) -> ModelBase:
"""Load the specified VLM and processor.
Args:
model_var: The model to load from ModelVariants enum.
config: The configuration object with model parameters.
Returns:
The loaded model instance.
Raises:
Exception: If model loading fails.
"""
global models_cache, current_model_selection
model_key = model_var.value
# Check if model is already loaded
if model_key in models_cache:
current_model_selection = model_var
return models_cache[model_key]
print(f'Loading {model_var.value} model...')
try:
model_selection = config.architecture
model = get_model(config.architecture, config)
# Cache the loaded model and processor
models_cache[model_key] = model
current_model_selection = model_selection
print(f'{model_selection.value} model loaded successfully!')
return model
except Exception as e:
print(f'Error loading model {model_selection.value}: {str(e)}')
raise
def get_single_image_probabilities(
instruction: str,
image: Image.Image,
vlm: ModelBase,
model_selection: ModelSelection,
top_k: int = 8
) -> Tuple[List[str], np.ndarray]:
"""Process a single image and return first token probabilities.
Args:
instruction: Text instruction for the model.
image: PIL Image to process.
vlm: Loaded model.
model_selection: The VLM being used.
top_k: Number of top tokens to return.
Returns:
Tuple containing list of top tokens and their probabilities.
"""
# Generate prompt and process inputs
vlm.model.eval()
text = vlm._generate_prompt(instruction, has_images=True)
inputs = vlm._generate_processor_output(text, image)
for key in inputs:
if isinstance(inputs[key], torch.Tensor):
inputs[key] = inputs[key].to(vlm.config.device)
with torch.no_grad():
outputs = vlm.model.generate(
**inputs,
max_new_tokens=1, # Only generate first token
output_scores=True,
return_dict_in_generate=True,
do_sample=False
)
# Get the logits for the first generated token
first_token_logits = outputs.scores[0][0] # Shape: [vocab_size]
# Convert logits to probabilities
probabilities = torch.softmax(first_token_logits, dim=-1)
# Get top-k probabilities for visualization
top_probs, top_indices = torch.topk(probabilities, top_k)
# Convert tokens back to text
top_tokens = [vlm.processor.tokenizer.decode([idx.item()]) for idx in top_indices]
return top_tokens, top_probs.cpu().numpy()
def scale_figure_fonts(fig: Figure, factor: float = 1.5) -> None:
"""Multiply all text sizes in a Matplotlib Figure by `factor`.
Args:
fig: The Matplotlib Figure to scale.
factor: The scaling factor (e.g., 1.5 to increase by 50%).
"""
for ax in fig.get_axes():
# titles & axis labels
ax.title.set_fontsize(ax.title.get_fontsize() * factor)
ax.xaxis.label.set_size(ax.xaxis.label.get_size() * factor)
ax.yaxis.label.set_size(ax.yaxis.label.get_size() * factor)
# tick labels
for lbl in ax.get_xticklabels() + ax.get_yticklabels():
lbl.set_fontsize(lbl.get_fontsize() * factor)
# texts placed via ax.text(...) (e.g., numbers above bars / "No data" notes)
for t in ax.texts:
t.set_fontsize(t.get_fontsize() * factor)
# any stray Text artists attached to the figure (rare, but safe)
for t in fig.findobj(match=Text):
if t.figure is fig:
t.set_fontsize(t.get_fontsize() * factor)
def create_dual_probability_plot(
tokens1: List[str], probabilities1: np.ndarray,
tokens2: List[str], probabilities2: np.ndarray
) -> Figure:
"""Create a matplotlib plot comparing token probabilities from two images.
Args:
tokens1: List of token strings from first image.
probabilities1: Array of probability values from first image.
tokens2: List of token strings from second image.
probabilities2: Array of probability values from second image.
Returns:
Matplotlib Figure object.
"""
if len(tokens1) == 0 and len(tokens2) == 0:
fig, ax = plt.subplots(figsize=(15, 8))
ax.text(0.5, 0.5, 'No data to display',
horizontalalignment='center', verticalalignment='center')
ax.set_xlim(0, 1)
ax.set_ylim(0, 1)
return fig
# Unify y-range with padding (cap at 1.0)
max1 = float(np.max(probabilities1)) if len(tokens1) else 0.0
max2 = float(np.max(probabilities2)) if len(tokens2) else 0.0
y_upper = min(1.0, max(max1, max2) * 1.15 + 1e-6) # ~15% headroom
# Create subplots side by side with shared y
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(20, 12), sharey=True)
ax1.set_ylim(0, y_upper)
ax2.set_ylim(0, y_upper)
# Plot first image results
if len(tokens1) > 0:
bars1 = ax1.bar(range(len(tokens1)), probabilities1, color='lightcoral',
edgecolor='darkred', alpha=0.7)
ax1.set_xlabel('Tokens', fontsize=12)
ax1.set_ylabel('Probability', fontsize=12)
ax1.set_title('Image 1 - First Token Probabilities',
fontsize=14, fontweight='bold')
ax1.set_xticks(range(len(tokens1)))
ax1.set_xticklabels(tokens1, rotation=45, ha='right')
# Clamp label position so it stays inside the axes
for bar, prob in zip(bars1, probabilities1):
h = bar.get_height()
y = min(h + 0.02 * y_upper, y_upper * 0.98)
ax1.text(bar.get_x() + bar.get_width()/2., y, f'{prob:.3f}',
ha='center', va='bottom', fontsize=9)
ax1.grid(axis='y', alpha=0.3)
else:
ax1.text(0.5, 0.5, 'No data for Image 1',
horizontalalignment='center', verticalalignment='center')
ax1.set_xlim(0, 1)
ax1.set_ylim(0, 1)
# Plot second image results
if len(tokens2) > 0:
bars2 = ax2.bar(range(len(tokens2)), probabilities2, color='skyblue',
edgecolor='navy', alpha=0.7)
ax2.set_xlabel('Tokens', fontsize=12)
ax2.set_ylabel('Probability', fontsize=12)
ax2.set_title('Image 2 - First Token Probabilities',
fontsize=14, fontweight='bold')
ax2.set_xticks(range(len(tokens2)))
ax2.set_xticklabels(tokens2, rotation=45, ha='right')
for bar, prob in zip(bars2, probabilities2):
h = bar.get_height()
y = min(h + 0.02 * y_upper, y_upper * 0.98)
ax2.text(bar.get_x() + bar.get_width()/2., y, f'{prob:.3f}',
ha='center', va='bottom', fontsize=9)
ax2.grid(axis='y', alpha=0.3)
else:
ax2.text(0.5, 0.5, 'No data for Image 2',
horizontalalignment='center', verticalalignment='center')
ax2.set_xlim(0, 1)
ax2.set_ylim(0, 1)
# Give extra space for rotated tick labels
fig.tight_layout()
fig.subplots_adjust(bottom=0.18)
return fig
def get_module_similarity_pooled(
vlm: ModelBase,
module_name: str,
image1: Image.Image,
image2: Image.Image,
instruction: str,
pooling: str = 'mean'
) -> float:
"""Compute cosine similarity with optional pooling strategies.
Args:
vlm: The loaded VLM (ModelBase instance).
module_name: The layer/module name to extract features from.
image1: First PIL Image.
image2: Second PIL Image.
instruction: Text instruction for the model.
pooling: Pooling strategy - 'mean', 'max', 'cls', or 'none'.
Returns:
Cosine similarity value between the two embeddings.
Raises:
ValueError: If feature extraction fails or module not found.
"""
embeddings = {}
target_module = None
def hook_fn(
module: torch.nn.Module,
input: Any,
output: Any
) -> None:
"""Forward hook to capture module output.
Args:
module: The module being hooked.
input: The input to the module.
output: The output from the module.
"""
if isinstance(output, tuple):
embeddings['activation'] = output[0].detach()
else:
embeddings['activation'] = output.detach()
# Find and register hook
for name, module in vlm.model.named_modules():
if name == module_name:
target_module = module
hook_handle = module.register_forward_hook(hook_fn)
break
if target_module is None:
raise ValueError(f"Module '{module_name}' not found in model")
try:
vlm.model.eval()
# Extract embedding for image1
text = vlm._generate_prompt(instruction, has_images=True)
inputs1 = vlm._generate_processor_output(text, image1)
for key in inputs1:
if isinstance(inputs1[key], torch.Tensor):
inputs1[key] = inputs1[key].to(vlm.config.device)
embeddings.clear()
with torch.no_grad():
_ = vlm.model(**inputs1)
if 'activation' not in embeddings:
raise ValueError('Failed to extract features for image1')
embedding1 = embeddings['activation']
# Extract embedding for image2
inputs2 = vlm._generate_processor_output(text, image2)
for key in inputs2:
if isinstance(inputs2[key], torch.Tensor):
inputs2[key] = inputs2[key].to(vlm.config.device)
embeddings.clear()
with torch.no_grad():
_ = vlm.model(**inputs2)
if 'activation' not in embeddings:
raise ValueError('Failed to extract features for image2')
embedding2 = embeddings['activation']
# Apply pooling strategy
if pooling == 'mean':
# Mean pooling across sequence dimension
if embedding1.dim() >= 2:
embedding1_pooled = embedding1.mean(dim=1)
embedding2_pooled = embedding2.mean(dim=1)
else:
embedding1_pooled = embedding1
embedding2_pooled = embedding2
elif pooling == 'max':
# Max pooling across sequence dimension
if embedding1.dim() >= 2:
embedding1_pooled = embedding1.max(dim=1)[0]
embedding2_pooled = embedding2.max(dim=1)[0]
else:
embedding1_pooled = embedding1
embedding2_pooled = embedding2
elif pooling == 'cls':
# Use first token (CLS token)
if embedding1.dim() >= 2:
embedding1_pooled = embedding1[:, 0, :]
embedding2_pooled = embedding2[:, 0, :]
else:
embedding1_pooled = embedding1
embedding2_pooled = embedding2
elif pooling == 'none':
# Flatten without pooling
embedding1_pooled = embedding1.reshape(embedding1.shape[0], -1)
embedding2_pooled = embedding2.reshape(embedding2.shape[0], -1)
else:
raise ValueError(f'Unknown pooling strategy: {pooling}')
# Ensure 2D shape [batch, features]
if embedding1_pooled.dim() == 1:
embedding1_pooled = embedding1_pooled.unsqueeze(0)
embedding2_pooled = embedding2_pooled.unsqueeze(0)
# Compute cosine similarity
similarity = F.cosine_similarity(embedding1_pooled, embedding2_pooled, dim=1)
similarity_value = float(similarity.mean().cpu().item())
return similarity_value
finally:
hook_handle.remove()
@GPU(duration=120)
def process_dual_inputs(
model_choice: str,
selected_layer: str,
instruction: str,
image1: Optional[Image.Image],
image2: Optional[Image.Image],
top_k: int = 8
) -> Tuple[Optional[Figure], str]:
"""Main function to process dual inputs and return comparison plot.
Args:
model_choice: String name of the selected model.
selected_layer: String name of the selected layer.
instruction: Text instruction for the model.
image1: First PIL Image to process, can be None.
image2: Second PIL Image to process, can be None.
top_k: Number of top tokens to display.
Returns:
Tuple containing the plot figure and info text.
"""
if image1 is None and image2 is None:
return None, 'Please upload at least one image.'
if not instruction.strip():
return None, 'Please provide an instruction.'
if not model_choice:
return None, 'Please select a model.'
if not selected_layer:
return None, 'Please select a layer.'
try:
# Initialize a config
model_var = ModelVariants(model_choice.lower())
model_selection, model_path, _ = get_model_info(model_var)
config = Config(model_selection, model_path, selected_layer, instruction)
config.model = {
'torch_dtype': torch.float16,
'low_cpu_mem_usage': True,
'device_map': 'auto'
}
# Load the model
model = load_model(model_var, config)
# Handle cases where only one image is provided
if image1 is None:
image1 = image2
tokens1, probs1 = [], np.array([])
tokens2, probs2 = get_single_image_probabilities(
instruction, image2, model, model_selection, top_k
)
elif image2 is None:
image2 = image1
tokens1, probs1 = get_single_image_probabilities(
instruction, image1, model, model_selection, top_k
)
tokens2, probs2 = [], np.array([])
else:
tokens1, probs1 = get_single_image_probabilities(
instruction, image1, model, model_selection, top_k
)
tokens2, probs2 = get_single_image_probabilities(
instruction, image2, model, model_selection, top_k
)
if len(tokens1) == 0 and len(tokens2) == 0:
return None, 'Error: Could not process the inputs. Please check the model loading.'
# Create comparison plot
plot = create_dual_probability_plot(
tokens1, probs1, tokens2, probs2
)
scale_figure_fonts(plot, factor=1.25)
# Create info text
info_text = f'Model: {model_choice.upper()}\n'
info_text += f'Top-K: {top_k}\n'
info_text += f"Instruction: '{instruction}'\n\n"
if len(tokens1) > 0:
info_text += f"Image 1 - Top token: '{tokens1[0]}' (probability: {probs1[0]:.4f})\n"
else:
info_text += 'Image 1 - No data\n'
if len(tokens2) > 0:
info_text += f"Image 2 - Top token: '{tokens2[0]}' (probability: {probs2[0]:.4f})\n"
else:
info_text += 'Image 2 - No data\n'
if len(tokens1) > 0 and len(tokens2) > 0:
info_text += f'\nLayer: {selected_layer}\n'
similarity = get_module_similarity_pooled(model, selected_layer, image1, image2, instruction)
info_text += f'Cosine similarity between Image 1 and 2: {similarity:.3f}\n'
return plot, info_text
except ValueError as e:
return None, f'Invalid model selection: {str(e)}'
except Exception as e:
return None, f'Error: {str(e)}'
def create_demo() -> gr.Blocks:
"""Create and configure the Gradio demo interface for dual image comparison.
Returns:
Configured Gradio Blocks interface.
"""
with gr.Blocks(title='VLM-Lens Visualizer') as demo:
gr.Markdown("""
# VLM-Lens (EMNLP 2025 System Demonstration)
## [arXiv](https://arxiv.org/abs/2510.02292) | [GitHub](https://github.com/compling-wat/vlm-lens)
This beta version processes an instruction with up to two images through various VLMs,
computes cosine similarity between their embeddings at a specified layer,
and visualizes the probability distribution of the first token in the response for each image.
**Instructions:**
1. Select a VLM from the dropdown
2. Select a layer from the available embedding layers
3. Upload two images for comparison
4. Enter your instruction/question about the images
5. Adjust the number of top tokens to display (1-20)
6. Click "Analyze" to see the first token probability distributions side by side
**Note:** You can upload just one image if you prefer single image analysis.
""")
with gr.Row():
with gr.Column():
model_dropdown = gr.Dropdown(
choices=[v.value.capitalize() for v in ModelVariants],
label='Select VLM',
value=None,
interactive=True
)
layer_dropdown = gr.Dropdown(
choices=[],
label='Select Module',
visible=False,
interactive=True
)
instruction_input = gr.Textbox(
label='Instruction',
placeholder='Describe what you see in this image...',
lines=3
)
top_k_slider = gr.Slider(
minimum=1,
maximum=20,
value=8,
step=1,
label='Number of Top Tokens to Display',
info='Select how many top probability tokens to show in the visualization'
)
with gr.Row():
image1_input = gr.Image(
label='Upload Image 1',
type='pil'
)
image2_input = gr.Image(
label='Upload Image 2',
type='pil'
)
analyze_btn = gr.Button('Analyze', variant='primary', visible=False)
with gr.Column():
plot_output = gr.Plot(label='First Token Probability Distribution Comparison')
info_output = gr.Textbox(
label='Analysis Info',
lines=8,
interactive=False
)
# Set up event handlers
model_dropdown.change(
fn=update_layer_choices,
inputs=[model_dropdown],
outputs=[layer_dropdown, analyze_btn]
)
analyze_btn.click(
fn=process_dual_inputs,
inputs=[model_dropdown, layer_dropdown, instruction_input, image1_input, image2_input, top_k_slider],
outputs=[plot_output, info_output]
)
# Add examples
gr.Examples(
examples=[
['What is in this image? Describe in one word.', None, None],
['Describe the main object in the picture in one word.', None, None],
['What color is the dominant object? Describe in one word.', None, None],
],
inputs=[instruction_input, image1_input, image2_input]
)
return demo
if __name__ == '__main__':
# Create and launch the demo
demo = create_demo()
demo.launch(
share=True,
server_name='0.0.0.0',
server_port=7860
)