Spaces:
Running
Running
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,523 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
import numpy as np
|
| 3 |
+
import json
|
| 4 |
+
import pandas as pd
|
| 5 |
+
from openai import OpenAI
|
| 6 |
+
import yaml
|
| 7 |
+
from typing import Optional, List, Dict, Tuple, Any
|
| 8 |
+
from topk_sae import FastAutoencoder
|
| 9 |
+
import torch
|
| 10 |
+
import plotly.express as px
|
| 11 |
+
from collections import Counter
|
| 12 |
+
from huggingface_hub import hf_hub_download
|
| 13 |
+
import os
|
| 14 |
+
|
| 15 |
+
import os
|
| 16 |
+
print(os.getenv('MODEL_REPO_ID'))
|
| 17 |
+
|
| 18 |
+
# Constants
|
| 19 |
+
EMBEDDING_MODEL = "text-embedding-3-small"
|
| 20 |
+
d_model = 1536
|
| 21 |
+
n_dirs = d_model * 6
|
| 22 |
+
k = 64
|
| 23 |
+
auxk = 128
|
| 24 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
| 25 |
+
torch.set_grad_enabled(False)
|
| 26 |
+
|
| 27 |
+
# Function to download all necessary files
|
| 28 |
+
def download_all_files():
|
| 29 |
+
files_to_download = [
|
| 30 |
+
"astroPH_paper_metadata.csv",
|
| 31 |
+
"csLG_feature_analysis_results_64.json",
|
| 32 |
+
"astroPH_topk_indices_64_9216_int32.npy",
|
| 33 |
+
"astroPH_64_9216.pth",
|
| 34 |
+
"astroPH_topk_values_64_9216_float16.npy",
|
| 35 |
+
"csLG_abstract_texts.json",
|
| 36 |
+
"csLG_topk_values_64_9216_float16.npy",
|
| 37 |
+
"csLG_abstract_embeddings_float16.npy",
|
| 38 |
+
"csLG_paper_metadata.csv",
|
| 39 |
+
"csLG_64_9216.pth",
|
| 40 |
+
"astroPH_abstract_texts.json",
|
| 41 |
+
"astroPH_feature_analysis_results_64.json",
|
| 42 |
+
"csLG_topk_indices_64_9216_int32.npy",
|
| 43 |
+
"astroPH_abstract_embeddings_float16.npy"
|
| 44 |
+
]
|
| 45 |
+
|
| 46 |
+
for file in files_to_download:
|
| 47 |
+
local_path = os.path.join("data", file)
|
| 48 |
+
os.makedirs(os.path.dirname(local_path), exist_ok=True)
|
| 49 |
+
hf_hub_download(repo_id="charlieoneill/saerch-ai-data", filename=file, local_dir="data")
|
| 50 |
+
print(f"Downloaded {file}")
|
| 51 |
+
|
| 52 |
+
# Load configuration and initialize OpenAI client
|
| 53 |
+
download_all_files()
|
| 54 |
+
# config = yaml.safe_load(open('../config.yaml', 'r'))
|
| 55 |
+
# client = OpenAI(api_key=config['jwu_openai_key'])
|
| 56 |
+
|
| 57 |
+
# Load the API key from the environment variable
|
| 58 |
+
api_key = os.getenv('openai_key')
|
| 59 |
+
|
| 60 |
+
# Ensure the API key is set
|
| 61 |
+
if not api_key:
|
| 62 |
+
raise ValueError("The environment variable 'openai_key' is not set.")
|
| 63 |
+
|
| 64 |
+
# Initialize the OpenAI client with the API key
|
| 65 |
+
client = OpenAI(api_key=api_key)
|
| 66 |
+
|
| 67 |
+
# Function to load data for a specific subject
|
| 68 |
+
def load_subject_data(subject):
|
| 69 |
+
# embeddings_path = f"data/{subject}_abstract_embeddings.npy"
|
| 70 |
+
# texts_path = f"data/{subject}_abstract_texts.json"
|
| 71 |
+
# feature_analysis_path = f"data/{subject}_feature_analysis_results_{k}.json"
|
| 72 |
+
# metadata_path = f'data/{subject}_paper_metadata.csv'
|
| 73 |
+
# topk_indices_path = f"data/{subject}_topk_indices_{k}_{n_dirs}.npy"
|
| 74 |
+
# topk_values_path = f"data/{subject}_topk_values_{k}_{n_dirs}.npy"
|
| 75 |
+
|
| 76 |
+
embeddings_path = f"data/{subject}_abstract_embeddings_float16.npy"
|
| 77 |
+
texts_path = f"data/{subject}_abstract_texts.json"
|
| 78 |
+
feature_analysis_path = f"data/{subject}_feature_analysis_results_{k}.json"
|
| 79 |
+
metadata_path = f'data/{subject}_paper_metadata.csv'
|
| 80 |
+
topk_indices_path = f"data/{subject}_topk_indices_{k}_{n_dirs}_int32.npy"
|
| 81 |
+
topk_values_path = f"data/{subject}_topk_values_{k}_{n_dirs}_float16.npy"
|
| 82 |
+
|
| 83 |
+
# abstract_embeddings = np.load(embeddings_path)
|
| 84 |
+
# with open(texts_path, 'r') as f:
|
| 85 |
+
# abstract_texts = json.load(f)
|
| 86 |
+
# with open(feature_analysis_path, 'r') as f:
|
| 87 |
+
# feature_analysis = json.load(f)
|
| 88 |
+
# df_metadata = pd.read_csv(metadata_path)
|
| 89 |
+
# topk_indices = np.load(topk_indices_path)
|
| 90 |
+
# topk_values = np.load(topk_values_path)
|
| 91 |
+
|
| 92 |
+
abstract_embeddings = np.load(embeddings_path).astype(np.float32) # Load float16 and convert to float32
|
| 93 |
+
with open(texts_path, 'r') as f:
|
| 94 |
+
abstract_texts = json.load(f)
|
| 95 |
+
with open(feature_analysis_path, 'r') as f:
|
| 96 |
+
feature_analysis = json.load(f)
|
| 97 |
+
df_metadata = pd.read_csv(metadata_path)
|
| 98 |
+
topk_indices = np.load(topk_indices_path) # Already in int32, no conversion needed
|
| 99 |
+
topk_values = np.load(topk_values_path).astype(np.float32)
|
| 100 |
+
|
| 101 |
+
model_filename = f"{subject}_64_9216.pth"
|
| 102 |
+
model_path = os.path.join("data", model_filename)
|
| 103 |
+
|
| 104 |
+
ae = FastAutoencoder(n_dirs, d_model, k, auxk, multik=0).to(device)
|
| 105 |
+
ae.load_state_dict(torch.load(model_path))
|
| 106 |
+
ae.eval()
|
| 107 |
+
|
| 108 |
+
weights = torch.load(model_path)
|
| 109 |
+
decoder = weights['decoder.weight'].cpu().numpy()
|
| 110 |
+
del weights
|
| 111 |
+
|
| 112 |
+
return {
|
| 113 |
+
'abstract_embeddings': abstract_embeddings,
|
| 114 |
+
'abstract_texts': abstract_texts,
|
| 115 |
+
'feature_analysis': feature_analysis,
|
| 116 |
+
'df_metadata': df_metadata,
|
| 117 |
+
'topk_indices': topk_indices,
|
| 118 |
+
'topk_values': topk_values,
|
| 119 |
+
'ae': ae,
|
| 120 |
+
'decoder': decoder
|
| 121 |
+
}
|
| 122 |
+
|
| 123 |
+
# Load data for both subjects
|
| 124 |
+
subject_data = {
|
| 125 |
+
'astroPH': load_subject_data('astroPH'),
|
| 126 |
+
'csLG': load_subject_data('csLG')
|
| 127 |
+
}
|
| 128 |
+
|
| 129 |
+
# Update existing functions to use the selected subject's data
|
| 130 |
+
def get_embedding(text: Optional[str], model: str = EMBEDDING_MODEL) -> Optional[np.ndarray]:
|
| 131 |
+
try:
|
| 132 |
+
embedding = client.embeddings.create(input=[text], model=model).data[0].embedding
|
| 133 |
+
return np.array(embedding, dtype=np.float32)
|
| 134 |
+
except Exception as e:
|
| 135 |
+
print(f"Error getting embedding: {e}")
|
| 136 |
+
return None
|
| 137 |
+
|
| 138 |
+
def intervened_hidden_to_intervened_embedding(topk_indices, topk_values, ae):
|
| 139 |
+
with torch.no_grad():
|
| 140 |
+
return ae.decode_sparse(topk_indices, topk_values)
|
| 141 |
+
|
| 142 |
+
# Function definitions for feature activation, co-occurrence, styling, etc.
|
| 143 |
+
def get_feature_activations(subject, feature_index, m=5, min_length=100):
|
| 144 |
+
abstract_texts = subject_data[subject]['abstract_texts']
|
| 145 |
+
abstract_embeddings = subject_data[subject]['abstract_embeddings']
|
| 146 |
+
topk_indices = subject_data[subject]['topk_indices']
|
| 147 |
+
topk_values = subject_data[subject]['topk_values']
|
| 148 |
+
|
| 149 |
+
doc_ids = abstract_texts['doc_ids']
|
| 150 |
+
abstracts = abstract_texts['abstracts']
|
| 151 |
+
|
| 152 |
+
feature_mask = topk_indices == feature_index
|
| 153 |
+
activated_indices = np.where(feature_mask.any(axis=1))[0]
|
| 154 |
+
activation_values = np.where(feature_mask, topk_values, 0).max(axis=1)
|
| 155 |
+
|
| 156 |
+
sorted_activated_indices = activated_indices[np.argsort(-activation_values[activated_indices])]
|
| 157 |
+
|
| 158 |
+
top_m_abstracts = []
|
| 159 |
+
top_m_indices = []
|
| 160 |
+
for i in sorted_activated_indices:
|
| 161 |
+
if len(abstracts[i]) > min_length:
|
| 162 |
+
top_m_abstracts.append((doc_ids[i], abstracts[i], activation_values[i]))
|
| 163 |
+
top_m_indices.append(i)
|
| 164 |
+
if len(top_m_abstracts) == m:
|
| 165 |
+
break
|
| 166 |
+
|
| 167 |
+
return top_m_abstracts
|
| 168 |
+
|
| 169 |
+
def calculate_co_occurrences(subject, target_index, n_features=9216):
|
| 170 |
+
topk_indices = subject_data[subject]['topk_indices']
|
| 171 |
+
|
| 172 |
+
mask = np.any(topk_indices == target_index, axis=1)
|
| 173 |
+
co_occurring_indices = topk_indices[mask].flatten()
|
| 174 |
+
co_occurrences = Counter(co_occurring_indices)
|
| 175 |
+
del co_occurrences[target_index]
|
| 176 |
+
result = np.zeros(n_features, dtype=int)
|
| 177 |
+
result[list(co_occurrences.keys())] = list(co_occurrences.values())
|
| 178 |
+
return result
|
| 179 |
+
|
| 180 |
+
def style_dataframe(df: pd.DataFrame, is_top: bool) -> pd.DataFrame:
|
| 181 |
+
cosine_values = df['Cosine similarity'].astype(float)
|
| 182 |
+
min_val = cosine_values.min()
|
| 183 |
+
max_val = cosine_values.max()
|
| 184 |
+
|
| 185 |
+
def color_similarity(val):
|
| 186 |
+
val = float(val)
|
| 187 |
+
# Normalize the value between 0 and 1
|
| 188 |
+
if is_top:
|
| 189 |
+
normalized_val = (val - min_val) / (max_val - min_val)
|
| 190 |
+
else:
|
| 191 |
+
# For bottom correlated, reverse the normalization
|
| 192 |
+
normalized_val = (max_val - val) / (max_val - min_val)
|
| 193 |
+
|
| 194 |
+
# Adjust the color intensity to avoid zero intensity
|
| 195 |
+
color_intensity = 0.2 + (normalized_val * 0.8) # This ensures the range is from 0.2 to 1.0
|
| 196 |
+
|
| 197 |
+
if is_top:
|
| 198 |
+
color = f'background-color: rgba(0, 255, 0, {color_intensity:.2f})'
|
| 199 |
+
else:
|
| 200 |
+
color = f'background-color: rgba(255, 0, 0, {color_intensity:.2f})'
|
| 201 |
+
return color
|
| 202 |
+
|
| 203 |
+
return df.style.applymap(color_similarity, subset=['Cosine similarity'])
|
| 204 |
+
|
| 205 |
+
def get_feature_from_index(subject, index):
|
| 206 |
+
feature = next((f for f in subject_data[subject]['feature_analysis'] if f['index'] == index), None)
|
| 207 |
+
return feature
|
| 208 |
+
|
| 209 |
+
def visualize_feature(subject, index):
|
| 210 |
+
feature = next((f for f in subject_data[subject]['feature_analysis'] if f['index'] == index), None)
|
| 211 |
+
if feature is None:
|
| 212 |
+
return "Invalid feature index", None, None, None, None, None, None
|
| 213 |
+
|
| 214 |
+
output = f"# {feature['label']}\n\n"
|
| 215 |
+
output += f"* Pearson correlation: {feature['pearson_correlation']:.4f}\n\n"
|
| 216 |
+
output += f"* Density: {feature['density']:.4f}\n\n"
|
| 217 |
+
|
| 218 |
+
# Top m abstracts
|
| 219 |
+
top_m_abstracts = get_feature_activations(subject, index)
|
| 220 |
+
|
| 221 |
+
# Create dataframe for top abstracts
|
| 222 |
+
df_data = [
|
| 223 |
+
{"Title": m[1].split('\n\n')[0], "Activation value": f"{m[2]:.4f}"}
|
| 224 |
+
for m in top_m_abstracts
|
| 225 |
+
]
|
| 226 |
+
df_top_abstracts = pd.DataFrame(df_data)
|
| 227 |
+
|
| 228 |
+
# Activation value distribution
|
| 229 |
+
topk_indices = subject_data[subject]['topk_indices']
|
| 230 |
+
topk_values = subject_data[subject]['topk_values']
|
| 231 |
+
|
| 232 |
+
activation_values = np.where(topk_indices == index, topk_values, 0).max(axis=1)
|
| 233 |
+
fig2 = px.histogram(x=activation_values, nbins=50)
|
| 234 |
+
fig2.update_layout(
|
| 235 |
+
#title=f'{feature["label"]}',
|
| 236 |
+
xaxis_title='Activation value',
|
| 237 |
+
yaxis_title=None,
|
| 238 |
+
yaxis_type='log',
|
| 239 |
+
height=220,
|
| 240 |
+
)
|
| 241 |
+
|
| 242 |
+
# Correlated features
|
| 243 |
+
decoder = subject_data[subject]['decoder']
|
| 244 |
+
feature_vector = decoder[:, index]
|
| 245 |
+
decoder_without_feature = np.delete(decoder, index, axis=1)
|
| 246 |
+
cosine_similarities = np.dot(feature_vector, decoder_without_feature) / (np.linalg.norm(decoder_without_feature, axis=0) * np.linalg.norm(feature_vector))
|
| 247 |
+
|
| 248 |
+
topk = 5
|
| 249 |
+
topk_indices_cosine = np.argsort(-cosine_similarities)[:topk]
|
| 250 |
+
topk_values_cosine = cosine_similarities[topk_indices_cosine]
|
| 251 |
+
|
| 252 |
+
# Create dataframe for top 5 correlated features
|
| 253 |
+
df_top_correlated = pd.DataFrame({
|
| 254 |
+
"Feature": [get_feature_from_index(subject, i)['label'] for i in topk_indices_cosine],
|
| 255 |
+
"Cosine similarity": [f"{v:.4f}" for v in topk_values_cosine]
|
| 256 |
+
})
|
| 257 |
+
df_top_correlated_styled = style_dataframe(df_top_correlated, is_top=True)
|
| 258 |
+
|
| 259 |
+
bottomk = 5
|
| 260 |
+
bottomk_indices_cosine = np.argsort(cosine_similarities)[:bottomk]
|
| 261 |
+
bottomk_values_cosine = cosine_similarities[bottomk_indices_cosine]
|
| 262 |
+
|
| 263 |
+
# Create dataframe for bottom 5 correlated features
|
| 264 |
+
df_bottom_correlated = pd.DataFrame({
|
| 265 |
+
"Feature": [get_feature_from_index(subject, i)['label'] for i in bottomk_indices_cosine],
|
| 266 |
+
"Cosine similarity": [f"{v:.4f}" for v in bottomk_values_cosine]
|
| 267 |
+
})
|
| 268 |
+
df_bottom_correlated_styled = style_dataframe(df_bottom_correlated, is_top=False)
|
| 269 |
+
|
| 270 |
+
# Co-occurrences
|
| 271 |
+
co_occurrences = calculate_co_occurrences(subject, index)
|
| 272 |
+
topk = 5
|
| 273 |
+
topk_indices_co_occurrence = np.argsort(-co_occurrences)[:topk]
|
| 274 |
+
topk_values_co_occurrence = co_occurrences[topk_indices_co_occurrence]
|
| 275 |
+
|
| 276 |
+
# Create dataframe for top 5 co-occurring features
|
| 277 |
+
df_co_occurrences = pd.DataFrame({
|
| 278 |
+
"Feature": [get_feature_from_index(subject, i)['label'] for i in topk_indices_co_occurrence],
|
| 279 |
+
"Co-occurrences": topk_values_co_occurrence
|
| 280 |
+
})
|
| 281 |
+
|
| 282 |
+
return output, df_top_abstracts, df_top_correlated_styled, df_bottom_correlated_styled, df_co_occurrences, fig2
|
| 283 |
+
|
| 284 |
+
# Modify the main interface function
|
| 285 |
+
def create_interface():
|
| 286 |
+
custom_css = """
|
| 287 |
+
#custom-slider-* {
|
| 288 |
+
background-color: #ffe6e6;
|
| 289 |
+
}
|
| 290 |
+
"""
|
| 291 |
+
|
| 292 |
+
with gr.Blocks(css=custom_css) as demo:
|
| 293 |
+
subject = gr.Dropdown(choices=['astroPH', 'csLG'], label="Select Subject", value='astroPH')
|
| 294 |
+
|
| 295 |
+
with gr.Tabs():
|
| 296 |
+
with gr.Tab("SAErch"):
|
| 297 |
+
input_text = gr.Textbox(label="input")
|
| 298 |
+
search_results_state = gr.State([])
|
| 299 |
+
feature_values_state = gr.State([])
|
| 300 |
+
feature_indices_state = gr.State([])
|
| 301 |
+
manually_added_features_state = gr.State([])
|
| 302 |
+
|
| 303 |
+
def update_search_results(feature_values, feature_indices, manually_added_features, current_subject):
|
| 304 |
+
ae = subject_data[current_subject]['ae']
|
| 305 |
+
abstract_embeddings = subject_data[current_subject]['abstract_embeddings']
|
| 306 |
+
abstract_texts = subject_data[current_subject]['abstract_texts']
|
| 307 |
+
df_metadata = subject_data[current_subject]['df_metadata']
|
| 308 |
+
|
| 309 |
+
# Combine manually added features with query-generated features
|
| 310 |
+
all_indices = []
|
| 311 |
+
all_values = []
|
| 312 |
+
|
| 313 |
+
# Add manually added features first
|
| 314 |
+
for index in manually_added_features:
|
| 315 |
+
if index not in all_indices:
|
| 316 |
+
all_indices.append(index)
|
| 317 |
+
all_values.append(feature_values[feature_indices.index(index)] if index in feature_indices else 0.0)
|
| 318 |
+
|
| 319 |
+
# Add remaining query-generated features
|
| 320 |
+
for index, value in zip(feature_indices, feature_values):
|
| 321 |
+
if index not in all_indices:
|
| 322 |
+
all_indices.append(index)
|
| 323 |
+
all_values.append(value)
|
| 324 |
+
|
| 325 |
+
# Reconstruct query embedding
|
| 326 |
+
topk_indices = torch.tensor(all_indices).to(device)
|
| 327 |
+
topk_values = torch.tensor(all_values).to(device)
|
| 328 |
+
|
| 329 |
+
intervened_embedding = intervened_hidden_to_intervened_embedding(topk_indices, topk_values, ae)
|
| 330 |
+
intervened_embedding = intervened_embedding.cpu().numpy().flatten()
|
| 331 |
+
|
| 332 |
+
# Perform similarity search
|
| 333 |
+
sims = np.dot(abstract_embeddings, intervened_embedding)
|
| 334 |
+
topk_indices_search = np.argsort(sims)[::-1][:10]
|
| 335 |
+
doc_ids = abstract_texts['doc_ids']
|
| 336 |
+
topk_doc_ids = [doc_ids[i] for i in topk_indices_search]
|
| 337 |
+
|
| 338 |
+
# Prepare search results
|
| 339 |
+
search_results = []
|
| 340 |
+
for doc_id in topk_doc_ids:
|
| 341 |
+
metadata = df_metadata[df_metadata['arxiv_id'] == doc_id].iloc[0]
|
| 342 |
+
title = metadata['title'].replace('[', '').replace(']', '')
|
| 343 |
+
search_results.append([
|
| 344 |
+
title,
|
| 345 |
+
int(metadata['citation_count']),
|
| 346 |
+
int(metadata['year'])
|
| 347 |
+
])
|
| 348 |
+
|
| 349 |
+
return search_results, all_values, all_indices
|
| 350 |
+
|
| 351 |
+
@gr.render(inputs=[input_text, search_results_state, feature_values_state, feature_indices_state, manually_added_features_state, subject])
|
| 352 |
+
def show_components(text, search_results, feature_values, feature_indices, manually_added_features, current_subject):
|
| 353 |
+
if len(text) == 0:
|
| 354 |
+
return gr.Markdown("## No Input Provided")
|
| 355 |
+
|
| 356 |
+
if not search_results or text != getattr(show_components, 'last_query', None):
|
| 357 |
+
show_components.last_query = text
|
| 358 |
+
query_embedding = get_embedding(text)
|
| 359 |
+
|
| 360 |
+
ae = subject_data[current_subject]['ae']
|
| 361 |
+
with torch.no_grad():
|
| 362 |
+
recons, z_dict = ae(torch.tensor(query_embedding).unsqueeze(0).to(device))
|
| 363 |
+
topk_indices = z_dict['topk_indices'][0].cpu().numpy()
|
| 364 |
+
topk_values = z_dict['topk_values'][0].cpu().numpy()
|
| 365 |
+
|
| 366 |
+
feature_values = topk_values.tolist()
|
| 367 |
+
feature_indices = topk_indices.tolist()
|
| 368 |
+
search_results, feature_values, feature_indices = update_search_results(feature_values, feature_indices, manually_added_features, current_subject)
|
| 369 |
+
|
| 370 |
+
with gr.Row():
|
| 371 |
+
with gr.Column(scale=2):
|
| 372 |
+
df = gr.Dataframe(
|
| 373 |
+
headers=["Title", "Citation Count", "Year"],
|
| 374 |
+
value=search_results,
|
| 375 |
+
label="Top 10 Search Results"
|
| 376 |
+
)
|
| 377 |
+
|
| 378 |
+
feature_search = gr.Textbox(label="Search Feature Labels")
|
| 379 |
+
feature_matches = gr.CheckboxGroup(label="Matching Features", choices=[])
|
| 380 |
+
add_button = gr.Button("Add Selected Features")
|
| 381 |
+
|
| 382 |
+
def search_feature_labels(search_text):
|
| 383 |
+
if not search_text:
|
| 384 |
+
return gr.CheckboxGroup(choices=[])
|
| 385 |
+
matches = [f"{f['label']} ({f['index']})" for f in subject_data[current_subject]['feature_analysis'] if search_text.lower() in f['label'].lower()]
|
| 386 |
+
return gr.CheckboxGroup(choices=matches[:10])
|
| 387 |
+
|
| 388 |
+
feature_search.change(search_feature_labels, inputs=[feature_search], outputs=[feature_matches])
|
| 389 |
+
|
| 390 |
+
def on_add_features(selected_features, current_values, current_indices, manually_added_features):
|
| 391 |
+
if selected_features:
|
| 392 |
+
new_indices = [int(f.split('(')[-1].strip(')')) for f in selected_features]
|
| 393 |
+
|
| 394 |
+
# Add new indices to manually_added_features if they're not already there
|
| 395 |
+
manually_added_features = list(dict.fromkeys(manually_added_features + new_indices))
|
| 396 |
+
|
| 397 |
+
return gr.CheckboxGroup(value=[]), current_values, current_indices, manually_added_features
|
| 398 |
+
return gr.CheckboxGroup(value=[]), current_values, current_indices, manually_added_features
|
| 399 |
+
|
| 400 |
+
add_button.click(
|
| 401 |
+
on_add_features,
|
| 402 |
+
inputs=[feature_matches, feature_values_state, feature_indices_state, manually_added_features_state],
|
| 403 |
+
outputs=[feature_matches, feature_values_state, feature_indices_state, manually_added_features_state]
|
| 404 |
+
)
|
| 405 |
+
|
| 406 |
+
with gr.Column(scale=1):
|
| 407 |
+
update_button = gr.Button("Update Results")
|
| 408 |
+
sliders = []
|
| 409 |
+
for i, (value, index) in enumerate(zip(feature_values, feature_indices)):
|
| 410 |
+
feature = next((f for f in subject_data[current_subject]['feature_analysis'] if f['index'] == index), None)
|
| 411 |
+
label = f"{feature['label']} ({index})" if feature else f"Feature {index}"
|
| 412 |
+
|
| 413 |
+
# Add prefix and change color for manually added features
|
| 414 |
+
if index in manually_added_features:
|
| 415 |
+
label = f"[Custom] {label}"
|
| 416 |
+
slider = gr.Slider(minimum=0, maximum=1, step=0.01, value=value, label=label, key=f"slider-{index}", elem_id=f"custom-slider-{index}")
|
| 417 |
+
else:
|
| 418 |
+
slider = gr.Slider(minimum=0, maximum=1, step=0.01, value=value, label=label, key=f"slider-{index}")
|
| 419 |
+
|
| 420 |
+
sliders.append(slider)
|
| 421 |
+
|
| 422 |
+
def on_slider_change(*values):
|
| 423 |
+
manually_added_features = values[-1]
|
| 424 |
+
slider_values = list(values[:-1])
|
| 425 |
+
|
| 426 |
+
# Reconstruct feature_indices based on the order of sliders
|
| 427 |
+
reconstructed_indices = [int(slider.label.split('(')[-1].split(')')[0]) for slider in sliders]
|
| 428 |
+
|
| 429 |
+
new_results, new_values, new_indices = update_search_results(slider_values, reconstructed_indices, manually_added_features, current_subject)
|
| 430 |
+
return new_results, new_values, new_indices, manually_added_features
|
| 431 |
+
|
| 432 |
+
update_button.click(
|
| 433 |
+
on_slider_change,
|
| 434 |
+
inputs=sliders + [manually_added_features_state],
|
| 435 |
+
outputs=[search_results_state, feature_values_state, feature_indices_state, manually_added_features_state]
|
| 436 |
+
)
|
| 437 |
+
|
| 438 |
+
return [df, feature_search, feature_matches, add_button, update_button] + sliders
|
| 439 |
+
|
| 440 |
+
with gr.Tab("Feature Visualisation"):
|
| 441 |
+
gr.Markdown("# Feature Visualiser")
|
| 442 |
+
with gr.Row():
|
| 443 |
+
feature_search = gr.Textbox(label="Search Feature Labels")
|
| 444 |
+
feature_matches = gr.CheckboxGroup(label="Matching Features", choices=[])
|
| 445 |
+
visualize_button = gr.Button("Visualize Feature")
|
| 446 |
+
|
| 447 |
+
feature_info = gr.Markdown()
|
| 448 |
+
abstracts_heading = gr.Markdown("## Top 5 Abstracts")
|
| 449 |
+
top_abstracts = gr.Dataframe(
|
| 450 |
+
headers=["Title", "Activation value"],
|
| 451 |
+
interactive=False
|
| 452 |
+
)
|
| 453 |
+
|
| 454 |
+
gr.Markdown("## Correlated Features")
|
| 455 |
+
with gr.Row():
|
| 456 |
+
with gr.Column(scale=1):
|
| 457 |
+
gr.Markdown("### Top 5 Correlated Features")
|
| 458 |
+
top_correlated = gr.Dataframe(
|
| 459 |
+
headers=["Feature", "Cosine similarity"],
|
| 460 |
+
interactive=False
|
| 461 |
+
)
|
| 462 |
+
with gr.Column(scale=1):
|
| 463 |
+
gr.Markdown("### Bottom 5 Correlated Features")
|
| 464 |
+
bottom_correlated = gr.Dataframe(
|
| 465 |
+
headers=["Feature", "Cosine similarity"],
|
| 466 |
+
interactive=False
|
| 467 |
+
)
|
| 468 |
+
|
| 469 |
+
with gr.Row():
|
| 470 |
+
with gr.Column(scale=1):
|
| 471 |
+
gr.Markdown("## Top 5 Co-occurring Features")
|
| 472 |
+
co_occurring_features = gr.Dataframe(
|
| 473 |
+
headers=["Feature", "Co-occurrences"],
|
| 474 |
+
interactive=False
|
| 475 |
+
)
|
| 476 |
+
with gr.Column(scale=1):
|
| 477 |
+
gr.Markdown(f"## Activation Value Distribution")
|
| 478 |
+
activation_dist = gr.Plot()
|
| 479 |
+
|
| 480 |
+
def search_feature_labels(search_text, current_subject):
|
| 481 |
+
if not search_text:
|
| 482 |
+
return gr.CheckboxGroup(choices=[])
|
| 483 |
+
matches = [f"{f['label']} ({f['index']})" for f in subject_data[current_subject]['feature_analysis'] if search_text.lower() in f['label'].lower()]
|
| 484 |
+
return gr.CheckboxGroup(choices=matches[:10])
|
| 485 |
+
|
| 486 |
+
feature_search.change(search_feature_labels, inputs=[feature_search, subject], outputs=[feature_matches])
|
| 487 |
+
|
| 488 |
+
def on_visualize(selected_features, current_subject):
|
| 489 |
+
if not selected_features:
|
| 490 |
+
return "Please select a feature to visualize.", None, None, None, None, None, "", []
|
| 491 |
+
|
| 492 |
+
# Extract the feature index from the selected feature string
|
| 493 |
+
feature_index = int(selected_features[0].split('(')[-1].strip(')'))
|
| 494 |
+
feature_info, top_abstracts, top_correlated, bottom_correlated, co_occurring_features, activation_dist = visualize_feature(current_subject, feature_index)
|
| 495 |
+
|
| 496 |
+
# Return the visualization results along with empty values for search box and checkbox
|
| 497 |
+
return feature_info, top_abstracts, top_correlated, bottom_correlated, co_occurring_features, activation_dist, "", []
|
| 498 |
+
|
| 499 |
+
visualize_button.click(
|
| 500 |
+
on_visualize,
|
| 501 |
+
inputs=[feature_matches, subject],
|
| 502 |
+
outputs=[feature_info, top_abstracts, top_correlated, bottom_correlated, co_occurring_features, activation_dist, feature_search, feature_matches]
|
| 503 |
+
)
|
| 504 |
+
|
| 505 |
+
# Add logic to update components when subject changes
|
| 506 |
+
def on_subject_change(new_subject):
|
| 507 |
+
# Clear all states and return empty values for all components
|
| 508 |
+
return [], [], [], [], "", [], "", [], None, None, None, None, None, None
|
| 509 |
+
|
| 510 |
+
subject.change(
|
| 511 |
+
on_subject_change,
|
| 512 |
+
inputs=[subject],
|
| 513 |
+
outputs=[search_results_state, feature_values_state, feature_indices_state, manually_added_features_state,
|
| 514 |
+
input_text, feature_matches, feature_search, feature_matches,
|
| 515 |
+
feature_info, top_abstracts, top_correlated, bottom_correlated, co_occurring_features, activation_dist]
|
| 516 |
+
)
|
| 517 |
+
|
| 518 |
+
return demo
|
| 519 |
+
|
| 520 |
+
# Launch the interface
|
| 521 |
+
if __name__ == "__main__":
|
| 522 |
+
demo = create_interface()
|
| 523 |
+
demo.launch()
|