File size: 7,239 Bytes
6aef776 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 |
import os
import json
import pickle
import numpy as np
from tqdm import tqdm
import pandas as pd
from datetime import datetime
import yaml
# Import your actual modules exactly as app.py does
from utils.visualizations import get_instances, load_interp_space, compute_tsne_with_cache, compute_precomputed_regions
from utils.ui import update_task_display, instance_to_df
from utils.interp_space_utils import cached_generate_style_embedding, compute_g2v_features, compute_predicted_author
def load_config(path="config/config.yaml"):
with open(path, "r") as f:
return yaml.safe_load(f)
def precompute_all_caches(
models_to_test=None,
instances_to_process=None,
config_path="config/config.yaml",
force_regenerate=False
):
"""
Precompute all cache files using the EXACT same methods as app.py.
This follows the exact flow: load_task → update_task_display → run_visualization
"""
if models_to_test is None:
models_to_test = [
'gabrielloiseau/LUAR-MUD-sentence-transformers',
'gabrielloiseau/LUAR-CRUD-sentence-transformers',
'miladalsh/light-luar',
'AnnaWegmann/Style-Embedding'
]
print("=" * 60)
print("CACHE PRECOMPUTATION STARTED")
print(f"Timestamp: {datetime.now()}")
print(f"Models to test: {len(models_to_test)}")
print("=" * 60)
# Load configuration and instances EXACTLY like app.py
cfg = load_config(config_path)
print(f"Configuration loaded from {config_path}")
print(f"config : \n{cfg}")
instances, instance_ids = get_instances(cfg['instances_to_explain_path'])
interp = load_interp_space(cfg)
clustered_authors_df = interp['clustered_authors_df']
if instances_to_process is None:
instances_to_process = instance_ids
print(f"Processing {len(instances_to_process)} instances with {len(models_to_test)} models")
total_combinations = len(models_to_test) * len(instances_to_process)
current_combination = 0
cache_stats = {
'embeddings_generated': 0,
'tsne_computed': 0,
'regions_computed': 0,
'errors': []
}
for model_name in models_to_test:
print(f"\n{'=' * 40}")
print(f"PROCESSING MODEL: {model_name}")
print(f"{'=' * 40}")
for instance_id in tqdm(instances_to_process, desc=f"Processing instances for {model_name.split('/')[-1]}"):
current_combination += 1
try:
print(f"\n[{current_combination}/{total_combinations}] Processing Instance {instance_id}")
# STEP 1: Replicate the exact flow from load_button.click()
print(" → Replicating load_button.click() flow...")
# Create ground truth (using placeholder since we're caching)
ground_truth_author = None # Will be determined by the instance data
# Call update_task_display EXACTLY like app.py does
task_results = update_task_display(
mode="Predefined HRS Task", # Always use predefined for caching
iid=f"Task {instance_id}",
instances=instances,
background_df=clustered_authors_df,
mystery_file=None, # Not used for predefined
cand1_file=None, # Not used for predefined
cand2_file=None, # Not used for predefined
cand3_file=None, # Not used for predefined
true_author=ground_truth_author,
model_radio=model_name,
custom_model_input=""
)
# Extract the results exactly like app.py expects
(header_html, mystery_html, c0_html, c1_html, c2_html,
mystery_state, c0_state, c1_state, c2_state,
task_authors_embeddings_df, background_authors_embeddings_df,
predicted_author, ground_truth_author) = task_results
print(f" ✓ Embeddings generated for {len(task_authors_embeddings_df)} task authors")
print(f" ✓ Background embeddings: {len(background_authors_embeddings_df)} authors")
cache_stats['embeddings_generated'] += 1
# STEP 2: Replicate the exact flow from run_btn.click()
print(" → Replicating run_btn.click() flow...")
# Call visualize_clusters_plotly EXACTLY like app.py does
viz_results = visualize_clusters_plotly(
iid=int(instance_id),
cfg=cfg,
instances=instances,
model_radio=model_name,
custom_model_input="",
task_authors_df=task_authors_embeddings_df,
background_authors_embeddings_df=background_authors_embeddings_df,
pred_idx=predicted_author,
gt_idx=ground_truth_author
)
# Extract results exactly like app.py expects
(fig, style_names, bg_proj, bg_ids, bg_authors_df,
precomputed_regions_state, precomputed_regions_radio) = viz_results
print(f" ✓ t-SNE projection computed")
print(f" ✓ Precomputed regions generated")
cache_stats['tsne_computed'] += 1
cache_stats['regions_computed'] += 1
print(f" ✓ Instance {instance_id} with model {model_name} completed successfully")
except Exception as e:
error_msg = f"Error processing instance {instance_id} with model {model_name}: {str(e)}"
print(f" ✗ {error_msg}")
cache_stats['errors'].append(error_msg)
import traceback
traceback.print_exc()
continue
# Print final statistics
print("\n" + "=" * 60)
print("CACHE PRECOMPUTATION COMPLETED")
print("=" * 60)
print(f"Embeddings generated: {cache_stats['embeddings_generated']}")
print(f"t-SNE projections computed: {cache_stats['tsne_computed']}")
print(f"Region sets computed: {cache_stats['regions_computed']}")
print(f"Errors encountered: {len(cache_stats['errors'])}")
if cache_stats['errors']:
print("\nERROR DETAILS:")
for error in cache_stats['errors']:
print(f" - {error}")
return cache_stats
# Import the exact functions your app uses
from utils.visualizations import visualize_clusters_plotly
if __name__ == "__main__":
# Test with a small subset first
instances=[i for i in range(2)] # First 2 instances for testing
cache_stats = precompute_all_caches(
models_to_test=[
'gabrielloiseau/LUAR-MUD-sentence-transformers'
],
instances_to_process=instances,
force_regenerate=False
)
print(f"\nCache precomputation completed with {len(cache_stats['errors'])} errors.") |