Spaces:
Sleeping
Sleeping
File size: 10,042 Bytes
4c1c394 1361d88 62166b9 4c1c394 cc9ff34 314850b 9ea2428 4c1c394 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 | """
Dataset loading for the SAE Feature Explorer.
Both regular explorer_data.pt files and brain_meis.pt files are loaded
through a single function that produces the same dict schema. Derived
numpy arrays (freq, log_freq, live_mask, umap_backup, etc.) are
pre-computed at load time so callbacks never recompute them.
"""
import json
import os
from collections import OrderedDict
import numpy as np
import torch
from .args import args
from .state import _all_datasets
# ---------- Helpers ----------
def _build_basename_index(paths: list) -> dict:
"""Map both full basename and stem → image index for fast filename lookup."""
idx = {}
for i, p in enumerate(paths):
base = os.path.basename(p)
stem = os.path.splitext(base)[0]
idx[base] = i
idx[stem] = i
return idx
# ---------- Core loader ----------
def _load_dataset(path: str, label: str, *,
sae_url: str | None = None,
is_brain: bool = False,
thumb_dir: str = "") -> dict | None:
"""
Load one dataset file (.pt) and return a fully-populated dict.
Both explorer_data.pt and brain_meis.pt files are handled here.
The returned dict includes:
- all raw tensors from the file
- per-dataset derived arrays (freq, log_freq, live_mask, umap_backup, …)
- feature names and auto-interp labels read from JSON sidecars
- pre-computed heatmaps read from the _heatmaps.pt sidecar if present
"""
print(f"Loading [{label}] from {path} ...")
try:
d = torch.load(path, map_location='cpu', weights_only=False)
except Exception as err:
print(f" WARNING: failed to load {path}: {err}")
return None
# Resolve image paths for brain datasets where paths may be stored as basenames.
raw_paths = d.get('image_paths', [])
if is_brain and raw_paths and thumb_dir and (
not os.path.isabs(raw_paths[0]) or not os.path.exists(raw_paths[0])
):
image_paths = [os.path.join(thumb_dir, os.path.basename(p)) for p in raw_paths]
else:
image_paths = raw_paths
d_model = d['d_model']
nan2 = np.full((d_model, 2), np.nan, dtype=np.float32)
stem = os.path.splitext(path)[0]
# Feature names (manual labels)
names_file = (args.names_file if (path == args.data and args.names_file)
else stem + '_feature_names.json')
feature_names = {}
if os.path.exists(names_file):
with open(names_file) as f:
feature_names = {int(k): v for k, v in json.load(f).items()}
# Auto-interp labels
auto_interp_file = stem + '_auto_interp.json'
auto_interp_names = {}
if os.path.exists(auto_interp_file):
with open(auto_interp_file) as f:
auto_interp_names = {int(k): v for k, v in json.load(f).items()}
print(f" Loaded {len(auto_interp_names)} auto-interp labels")
# Core tensors
feature_frequency = d['feature_frequency']
feature_mean_act = d['feature_mean_act']
umap_coords = d['umap_coords'].numpy()
dict_umap_coords = (d['dict_umap_coords'].numpy()
if 'dict_umap_coords' in d else nan2)
# Derived arrays — computed once, stored in the dict
freq = feature_frequency.numpy()
mean_act = feature_mean_act.numpy()
log_freq = np.log10(freq + 1)
live_mask = ~np.isnan(umap_coords[:, 0])
live_indices = np.where(live_mask)[0]
dict_live_mask = ~np.isnan(dict_umap_coords[:, 0])
dict_live_indices = np.where(dict_live_mask)[0]
active_feats = [int(i) for i in range(d_model) if freq[i] > 0]
umap_backup = dict(
act_x = umap_coords[live_mask, 0].tolist(),
act_y = umap_coords[live_mask, 1].tolist(),
act_feat = live_indices.tolist(),
dict_x = dict_umap_coords[dict_live_mask, 0].tolist(),
dict_y = dict_umap_coords[dict_live_mask, 1].tolist(),
dict_feat= dict_live_indices.tolist(),
)
entry = {
'label': label,
'path': path,
'image_paths': image_paths,
'basename_index': _build_basename_index(image_paths),
'd_model': d_model,
'n_images': d.get('n_images', len(image_paths)),
'patch_grid': d.get('patch_grid', 16),
'image_size': d.get('image_size', 224),
'backbone': d.get('backbone', 'dinov2'),
'top_img_idx': d['top_img_idx'],
'top_img_act': d['top_img_act'],
'mean_img_idx': d.get('mean_img_idx', d['top_img_idx']),
'mean_img_act': d.get('mean_img_act', d['top_img_act']),
'nsd_top_img_idx': d.get('nsd_top_img_idx'),
'nsd_top_img_act': d.get('nsd_top_img_act'),
'nsd_mean_img_idx': d.get('nsd_mean_img_idx'),
'nsd_mean_img_act': d.get('nsd_mean_img_act'),
'feature_frequency': feature_frequency,
'feature_mean_act': feature_mean_act,
'umap_coords': umap_coords,
'dict_umap_coords': dict_umap_coords,
'clip_embeds': d.get('clip_feature_embeds'),
'nsd_clip_embeds': d.get('nsd_clip_feature_embeds'),
'sae_url': sae_url,
'inference_cache': OrderedDict(),
'names_file': names_file,
'auto_interp_file': auto_interp_file,
'feature_names': feature_names,
'auto_interp_names': auto_interp_names,
# Pre-computed derived arrays
'freq': freq,
'mean_act': mean_act,
'log_freq': log_freq,
'live_mask': live_mask,
'live_indices': live_indices,
'dict_live_mask': dict_live_mask,
'dict_live_indices': dict_live_indices,
'active_feats': active_feats,
'umap_backup': umap_backup,
}
# Brain MEI sidecar (disabled — re-enable after running precompute_brain_response_meis.py)
# brain_sidecar = stem + '_brain_meis.pt'
# if os.path.exists(brain_sidecar):
# print(f" Loading brain MEI sidecar {os.path.basename(brain_sidecar)} ...")
# bm = torch.load(brain_sidecar, map_location='cpu', weights_only=False)
# entry['brain_top_img_idx'] = bm.get('brain_top_img_idx')
# entry['brain_top_img_act'] = bm.get('brain_top_img_act')
# else:
# entry['brain_top_img_idx'] = None
# entry['brain_top_img_act'] = None
# Heatmaps sidecar
sidecar = stem + '_heatmaps.pt'
if os.path.exists(sidecar):
print(f" Loading heatmaps sidecar {os.path.basename(sidecar)} ...")
hm = torch.load(sidecar, map_location='cpu', weights_only=not is_brain)
entry['top_heatmaps'] = hm.get('top_heatmaps')
entry['mean_heatmaps'] = hm.get('mean_heatmaps')
entry['nsd_top_heatmaps'] = hm.get('nsd_top_heatmaps')
entry['nsd_mean_heatmaps'] = hm.get('nsd_mean_heatmaps')
entry['heatmap_patch_grid'] = hm.get('patch_grid', d.get('patch_grid', 16))
has_hm = 'yes'
else:
entry['top_heatmaps'] = None
entry['mean_heatmaps'] = None
entry['nsd_top_heatmaps'] = None
entry['nsd_mean_heatmaps'] = None
entry['heatmap_patch_grid'] = d.get('patch_grid', 16)
has_hm = 'no'
# Brain render sidecar (precomputed compact phi map PNGs)
brain_render_sidecar = stem + '_brain_renders.pt'
if os.path.exists(brain_render_sidecar):
print(f" Loading brain render sidecar {os.path.basename(brain_render_sidecar)} ...")
br = torch.load(brain_render_sidecar, map_location='cpu', weights_only=False)
entry['phi_map_cache'] = {int(k): v for k, v in br.items()}
print(f" Cached {len(entry['phi_map_cache'])} phi map renders")
else:
entry['phi_map_cache'] = {}
# Cortical profile sidecar (precomputed 4-view cortical profile PNGs)
cortical_sidecar = stem + '_cortical_profiles.pt'
if os.path.exists(cortical_sidecar):
print(f" Loading cortical profile sidecar {os.path.basename(cortical_sidecar)} ...")
cp = torch.load(cortical_sidecar, map_location='cpu', weights_only=False)
entry['cortical_profile_cache'] = {int(k): v for k, v in cp.items()}
print(f" Cached {len(entry['cortical_profile_cache'])} cortical profiles")
else:
entry['cortical_profile_cache'] = {}
# GT brain render sidecar (precomputed fMRI response PNGs per NSD image)
gt_brain_sidecar = stem + '_gt_brain_renders.pt'
if os.path.exists(gt_brain_sidecar):
print(f" Loading GT brain render sidecar {os.path.basename(gt_brain_sidecar)} ...")
gb = torch.load(gt_brain_sidecar, map_location='cpu', weights_only=False)
entry['gt_brain_cache'] = {int(k): v for k, v in gb.items()}
print(f" Cached {len(entry['gt_brain_cache'])} GT brain renders")
else:
entry['gt_brain_cache'] = {}
has_clip = 'yes' if entry['clip_embeds'] is not None else 'no'
print(f" d={d_model}, n={entry['n_images']}, backbone={entry['backbone']}, "
f"clip={has_clip}, heatmaps={has_hm}")
return entry
# ---------- Public entry point ----------
def load_all_datasets():
"""Load all datasets specified by CLI args into _all_datasets."""
# Primary dataset (always required)
_all_datasets.append(
_load_dataset(args.data, args.primary_label, sae_url=args.sae_url)
)
# Optional NSD brain dataset
if args.brain_data:
if os.path.exists(args.brain_data):
entry = _load_dataset(
args.brain_data, args.brain_label,
is_brain=True,
thumb_dir=args.brain_thumbnails or '',
)
if entry is not None:
_all_datasets.append(entry)
else:
print(f"[Brain] WARNING: --brain-data not found: {args.brain_data}")
|