Spaces:
Sleeping
Sleeping
test: fixed typo
Browse files
app.py
CHANGED
@@ -155,805 +155,6 @@ def get_top_images(slider_value: int, toggle_btn: bool) -> List[Image.Image]:
|
|
155 |
|
156 |
return _CACHE['top_images'][cache_key]
|
157 |
|
158 |
-
# Initialize data
|
159 |
-
data_dict, sae_data_dict = load_all_data(image_root="./data/image", pkl_root=pkl_root)
|
160 |
-
|
161 |
-
|
162 |
-
# def preload_activation(image_name):
|
163 |
-
# for model in ["CLIP"] + [f"MaPLE-{ds}" for ds in DATASET_LIST]:
|
164 |
-
# image_file = f"{pkl_root}/{model}/{image_name}.pkl.gz"
|
165 |
-
# with gzip.open(image_file, "rb") as f:
|
166 |
-
# preloaded_data[model] = pickle.load(f)
|
167 |
-
|
168 |
-
|
169 |
-
# def get_activation_distribution(image_name: str, model_type: str):
|
170 |
-
# activation = get_data(image_name, model_type)[0]
|
171 |
-
|
172 |
-
# noisy_features_indices = (
|
173 |
-
# (sae_data_dict["mean_acts"]["imagenet"] > 0.1).nonzero()[0].tolist()
|
174 |
-
# )
|
175 |
-
# activation[:, noisy_features_indices] = 0
|
176 |
-
|
177 |
-
# return activation
|
178 |
-
|
179 |
-
|
180 |
-
def get_grid_loc(evt, image):
|
181 |
-
# Get click coordinates
|
182 |
-
x, y = evt._data["index"][0], evt._data["index"][1]
|
183 |
-
|
184 |
-
cell_width = image.width // GRID_NUM
|
185 |
-
cell_height = image.height // GRID_NUM
|
186 |
-
|
187 |
-
grid_x = x // cell_width
|
188 |
-
grid_y = y // cell_height
|
189 |
-
return grid_x, grid_y, cell_width, cell_height
|
190 |
-
|
191 |
-
|
192 |
-
def highlight_grid(evt: gr.EventData, image_name):
|
193 |
-
image = data_dict[image_name]["image"]
|
194 |
-
grid_x, grid_y, cell_width, cell_height = get_grid_loc(evt, image)
|
195 |
-
|
196 |
-
highlighted_image = image.copy()
|
197 |
-
draw = ImageDraw.Draw(highlighted_image)
|
198 |
-
box = [
|
199 |
-
grid_x * cell_width,
|
200 |
-
grid_y * cell_height,
|
201 |
-
(grid_x + 1) * cell_width,
|
202 |
-
(grid_y + 1) * cell_height,
|
203 |
-
]
|
204 |
-
draw.rectangle(box, outline="red", width=3)
|
205 |
-
|
206 |
-
return highlighted_image
|
207 |
-
|
208 |
-
|
209 |
-
def load_image(img_name):
|
210 |
-
return Image.open(data_dict[img_name]["image_path"]).resize(
|
211 |
-
(IMAGE_SIZE, IMAGE_SIZE)
|
212 |
-
)
|
213 |
-
|
214 |
-
|
215 |
-
def plot_activations(
|
216 |
-
all_activation,
|
217 |
-
tile_activations=None,
|
218 |
-
grid_x=None,
|
219 |
-
grid_y=None,
|
220 |
-
top_k=5,
|
221 |
-
colors=("blue", "cyan"),
|
222 |
-
model_name="CLIP",
|
223 |
-
):
|
224 |
-
fig = go.Figure()
|
225 |
-
|
226 |
-
def _add_scatter_with_annotation(fig, activations, model_name, color, label):
|
227 |
-
fig.add_trace(
|
228 |
-
go.Scatter(
|
229 |
-
x=np.arange(len(activations)),
|
230 |
-
y=activations,
|
231 |
-
mode="lines",
|
232 |
-
name=label,
|
233 |
-
line=dict(color=color, dash="solid"),
|
234 |
-
showlegend=True,
|
235 |
-
)
|
236 |
-
)
|
237 |
-
top_neurons = np.argsort(activations)[::-1][:top_k]
|
238 |
-
for idx in top_neurons:
|
239 |
-
fig.add_annotation(
|
240 |
-
x=idx,
|
241 |
-
y=activations[idx],
|
242 |
-
text=str(idx),
|
243 |
-
showarrow=True,
|
244 |
-
arrowhead=2,
|
245 |
-
ax=0,
|
246 |
-
ay=-15,
|
247 |
-
arrowcolor=color,
|
248 |
-
opacity=0.7,
|
249 |
-
)
|
250 |
-
return fig
|
251 |
-
|
252 |
-
label = f"{model_name.split('-')[-0]} Image-level"
|
253 |
-
fig = _add_scatter_with_annotation(
|
254 |
-
fig, all_activation, model_name, colors[0], label
|
255 |
-
)
|
256 |
-
if tile_activations is not None:
|
257 |
-
label = f"{model_name.split('-')[-0]} Tile ({grid_x}, {grid_y})"
|
258 |
-
fig = _add_scatter_with_annotation(
|
259 |
-
fig, tile_activations, model_name, colors[1], label
|
260 |
-
)
|
261 |
-
|
262 |
-
fig.update_layout(
|
263 |
-
title="Activation Distribution",
|
264 |
-
xaxis_title="SAE latent index",
|
265 |
-
yaxis_title="Activation Value",
|
266 |
-
template="plotly_white",
|
267 |
-
)
|
268 |
-
fig.update_layout(
|
269 |
-
legend=dict(orientation="h", yanchor="middle", y=0.5, xanchor="center", x=0.5)
|
270 |
-
)
|
271 |
-
|
272 |
-
return fig
|
273 |
-
|
274 |
-
|
275 |
-
def get_activations(evt: gr.EventData, selected_image: str, model_name: str, colors):
|
276 |
-
activation = get_activation_distribution(selected_image, model_name)
|
277 |
-
all_activation = activation.mean(0)
|
278 |
-
|
279 |
-
tile_activations = None
|
280 |
-
grid_x = None
|
281 |
-
grid_y = None
|
282 |
-
|
283 |
-
if evt is not None:
|
284 |
-
if evt._data is not None:
|
285 |
-
image = data_dict[selected_image]["image"]
|
286 |
-
grid_x, grid_y, cell_width, cell_height = get_grid_loc(evt, image)
|
287 |
-
token_idx = grid_y * GRID_NUM + grid_x + 1
|
288 |
-
tile_activations = activation[token_idx]
|
289 |
-
|
290 |
-
fig = plot_activations(
|
291 |
-
all_activation,
|
292 |
-
tile_activations,
|
293 |
-
grid_x,
|
294 |
-
grid_y,
|
295 |
-
top_k=5,
|
296 |
-
model_name=model_name,
|
297 |
-
colors=colors,
|
298 |
-
)
|
299 |
-
return fig
|
300 |
-
|
301 |
-
|
302 |
-
def plot_activation_distribution(
|
303 |
-
evt: gr.EventData, selected_image: str, model_name: str
|
304 |
-
):
|
305 |
-
fig = make_subplots(
|
306 |
-
rows=2,
|
307 |
-
cols=1,
|
308 |
-
shared_xaxes=True,
|
309 |
-
subplot_titles=["CLIP Activation", f"{model_name} Activation"],
|
310 |
-
)
|
311 |
-
|
312 |
-
fig_clip = get_activations(
|
313 |
-
evt, selected_image, "CLIP", colors=("#00b4d8", "#90e0ef")
|
314 |
-
)
|
315 |
-
fig_maple = get_activations(
|
316 |
-
evt, selected_image, model_name, colors=("#ff5a5f", "#ffcad4")
|
317 |
-
)
|
318 |
-
|
319 |
-
def _attach_fig(fig, sub_fig, row, col, yref):
|
320 |
-
for trace in sub_fig.data:
|
321 |
-
fig.add_trace(trace, row=row, col=col)
|
322 |
-
|
323 |
-
for annotation in sub_fig.layout.annotations:
|
324 |
-
annotation.update(yref=yref)
|
325 |
-
fig.add_annotation(annotation)
|
326 |
-
return fig
|
327 |
-
|
328 |
-
fig = _attach_fig(fig, fig_clip, row=1, col=1, yref="y1")
|
329 |
-
fig = _attach_fig(fig, fig_maple, row=2, col=1, yref="y2")
|
330 |
-
|
331 |
-
fig.update_xaxes(title_text="SAE Latent Index", row=2, col=1)
|
332 |
-
fig.update_xaxes(title_text="SAE Latent Index", row=1, col=1)
|
333 |
-
fig.update_yaxes(title_text="Activation Value", row=1, col=1)
|
334 |
-
fig.update_yaxes(title_text="Activation Value", row=2, col=1)
|
335 |
-
fig.update_layout(
|
336 |
-
# height=500,
|
337 |
-
# title="Activation Distributions",
|
338 |
-
template="plotly_white",
|
339 |
-
showlegend=True,
|
340 |
-
legend=dict(orientation="h", yanchor="bottom", y=-0.2, xanchor="center", x=0.5),
|
341 |
-
margin=dict(l=20, r=20, t=40, b=20),
|
342 |
-
)
|
343 |
-
|
344 |
-
return fig
|
345 |
-
|
346 |
-
|
347 |
-
# def get_segmask(selected_image, slider_value, model_type):
|
348 |
-
# image = data_dict[selected_image]["image"]
|
349 |
-
# sae_act = get_data(selected_image, model_type)[0]
|
350 |
-
# temp = sae_act[:, slider_value]
|
351 |
-
# try:
|
352 |
-
# mask = torch.Tensor(temp[1:,].reshape(14, 14)).view(1, 1, 14, 14)
|
353 |
-
# except Exception as e:
|
354 |
-
# print(sae_act.shape, slider_value)
|
355 |
-
# mask = torch.nn.functional.interpolate(mask, (image.height, image.width))[0][
|
356 |
-
# 0
|
357 |
-
# ].numpy()
|
358 |
-
# mask = (mask - mask.min()) / (mask.max() - mask.min() + 1e-10)
|
359 |
-
|
360 |
-
# base_opacity = 30
|
361 |
-
# image_array = np.array(image)[..., :3]
|
362 |
-
# rgba_overlay = np.zeros((mask.shape[0], mask.shape[1], 4), dtype=np.uint8)
|
363 |
-
# rgba_overlay[..., :3] = image_array[..., :3]
|
364 |
-
|
365 |
-
# darkened_image = (image_array[..., :3] * (base_opacity / 255)).astype(np.uint8)
|
366 |
-
# rgba_overlay[mask == 0, :3] = darkened_image[mask == 0]
|
367 |
-
# rgba_overlay[..., 3] = 255 # Fully opaque
|
368 |
-
|
369 |
-
# return rgba_overlay
|
370 |
-
|
371 |
-
|
372 |
-
# def get_top_images(slider_value, toggle_btn):
|
373 |
-
# def _get_images(dataset_path):
|
374 |
-
# top_image_paths = [
|
375 |
-
# os.path.join(dataset_path, "imagenet", f"{slider_value}.jpg"),
|
376 |
-
# os.path.join(dataset_path, "imagenet-sketch", f"{slider_value}.jpg"),
|
377 |
-
# os.path.join(dataset_path, "caltech101", f"{slider_value}.jpg"),
|
378 |
-
# ]
|
379 |
-
# top_images = [
|
380 |
-
# (
|
381 |
-
# Image.open(path)
|
382 |
-
# if os.path.exists(path)
|
383 |
-
# else Image.new("RGB", (256, 256), (255, 255, 255))
|
384 |
-
# )
|
385 |
-
# for path in top_image_paths
|
386 |
-
# ]
|
387 |
-
# return top_images
|
388 |
-
|
389 |
-
# if toggle_btn:
|
390 |
-
# top_images = _get_images("./data/top_images_masked")
|
391 |
-
# else:
|
392 |
-
# top_images = _get_images("./data/top_images")
|
393 |
-
# return top_images
|
394 |
-
|
395 |
-
|
396 |
-
def show_activation_heatmap(selected_image, slider_value, model_type, toggle_btn=False):
|
397 |
-
slider_value = int(slider_value.split("-")[-1])
|
398 |
-
rgba_overlay = get_segmask(selected_image, slider_value, model_type)
|
399 |
-
top_images = get_top_images(slider_value, toggle_btn)
|
400 |
-
|
401 |
-
act_values = []
|
402 |
-
for dataset in ["imagenet", "imagenet-sketch", "caltech101"]:
|
403 |
-
act_value = sae_data_dict["mean_act_values"][dataset][slider_value, :5]
|
404 |
-
act_value = [str(round(value, 3)) for value in act_value]
|
405 |
-
act_value = " | ".join(act_value)
|
406 |
-
out = f"#### Activation values: {act_value}"
|
407 |
-
act_values.append(out)
|
408 |
-
|
409 |
-
return rgba_overlay, top_images, act_values
|
410 |
-
|
411 |
-
|
412 |
-
def show_activation_heatmap_clip(selected_image, slider_value, toggle_btn):
|
413 |
-
rgba_overlay, top_images, act_values = show_activation_heatmap(
|
414 |
-
selected_image, slider_value, "CLIP", toggle_btn
|
415 |
-
)
|
416 |
-
sleep(0.1)
|
417 |
-
return (
|
418 |
-
rgba_overlay,
|
419 |
-
top_images[0],
|
420 |
-
top_images[1],
|
421 |
-
top_images[2],
|
422 |
-
act_values[0],
|
423 |
-
act_values[1],
|
424 |
-
act_values[2],
|
425 |
-
)
|
426 |
-
|
427 |
-
|
428 |
-
def show_activation_heatmap_maple(selected_image, slider_value, model_name):
|
429 |
-
slider_value = int(slider_value.split("-")[-1])
|
430 |
-
rgba_overlay = get_segmask(selected_image, slider_value, model_name)
|
431 |
-
sleep(0.1)
|
432 |
-
return rgba_overlay
|
433 |
-
|
434 |
-
|
435 |
-
def get_init_radio_options(selected_image, model_name):
|
436 |
-
clip_neuron_dict = {}
|
437 |
-
maple_neuron_dict = {}
|
438 |
-
|
439 |
-
def _get_top_actvation(selected_image, model_name, neuron_dict, top_k=5):
|
440 |
-
activations = get_activation_distribution(selected_image, model_name).mean(0)
|
441 |
-
top_neurons = list(np.argsort(activations)[::-1][:top_k])
|
442 |
-
for top_neuron in top_neurons:
|
443 |
-
neuron_dict[top_neuron] = activations[top_neuron]
|
444 |
-
sorted_dict = dict(
|
445 |
-
sorted(neuron_dict.items(), key=lambda item: item[1], reverse=True)
|
446 |
-
)
|
447 |
-
return sorted_dict
|
448 |
-
|
449 |
-
clip_neuron_dict = _get_top_actvation(selected_image, "CLIP", clip_neuron_dict)
|
450 |
-
maple_neuron_dict = _get_top_actvation(
|
451 |
-
selected_image, model_name, maple_neuron_dict
|
452 |
-
)
|
453 |
-
|
454 |
-
radio_choices = get_radio_names(clip_neuron_dict, maple_neuron_dict)
|
455 |
-
|
456 |
-
return radio_choices
|
457 |
-
|
458 |
-
|
459 |
-
def get_radio_names(clip_neuron_dict, maple_neuron_dict):
|
460 |
-
clip_keys = list(clip_neuron_dict.keys())
|
461 |
-
maple_keys = list(maple_neuron_dict.keys())
|
462 |
-
|
463 |
-
common_keys = list(set(clip_keys).intersection(set(maple_keys)))
|
464 |
-
clip_only_keys = list(set(clip_keys) - (set(maple_keys)))
|
465 |
-
maple_only_keys = list(set(maple_keys) - (set(clip_keys)))
|
466 |
-
|
467 |
-
common_keys.sort(
|
468 |
-
key=lambda x: max(clip_neuron_dict[x], maple_neuron_dict[x]), reverse=True
|
469 |
-
)
|
470 |
-
clip_only_keys.sort(reverse=True)
|
471 |
-
maple_only_keys.sort(reverse=True)
|
472 |
-
|
473 |
-
out = []
|
474 |
-
out.extend([f"common-{i}" for i in common_keys[:5]])
|
475 |
-
out.extend([f"CLIP-{i}" for i in clip_only_keys[:5]])
|
476 |
-
out.extend([f"MaPLE-{i}" for i in maple_only_keys[:5]])
|
477 |
-
|
478 |
-
return out
|
479 |
-
|
480 |
-
|
481 |
-
def update_radio_options(evt: gr.EventData, selected_image, model_name):
|
482 |
-
def _sort_and_save_top_k(activations, neuron_dict, top_k=5):
|
483 |
-
top_neurons = list(np.argsort(activations)[::-1][:top_k])
|
484 |
-
for top_neuron in top_neurons:
|
485 |
-
neuron_dict[top_neuron] = activations[top_neuron]
|
486 |
-
|
487 |
-
def _get_top_actvation(evt, selected_image, model_name, neuron_dict):
|
488 |
-
all_activation = get_activation_distribution(selected_image, model_name)
|
489 |
-
image_activation = all_activation.mean(0)
|
490 |
-
_sort_and_save_top_k(image_activation, neuron_dict)
|
491 |
-
|
492 |
-
if evt is not None:
|
493 |
-
if evt._data is not None and isinstance(evt._data["index"], list):
|
494 |
-
image = data_dict[selected_image]["image"]
|
495 |
-
grid_x, grid_y, cell_width, cell_height = get_grid_loc(evt, image)
|
496 |
-
token_idx = grid_y * GRID_NUM + grid_x + 1
|
497 |
-
tile_activations = all_activation[token_idx]
|
498 |
-
_sort_and_save_top_k(tile_activations, neuron_dict)
|
499 |
-
|
500 |
-
sorted_dict = dict(
|
501 |
-
sorted(neuron_dict.items(), key=lambda item: item[1], reverse=True)
|
502 |
-
)
|
503 |
-
return sorted_dict
|
504 |
-
|
505 |
-
clip_neuron_dict = {}
|
506 |
-
maple_neuron_dict = {}
|
507 |
-
clip_neuron_dict = _get_top_actvation(evt, selected_image, "CLIP", clip_neuron_dict)
|
508 |
-
maple_neuron_dict = _get_top_actvation(
|
509 |
-
evt, selected_image, model_name, maple_neuron_dict
|
510 |
-
)
|
511 |
-
|
512 |
-
clip_keys = list(clip_neuron_dict.keys())
|
513 |
-
maple_keys = list(maple_neuron_dict.keys())
|
514 |
-
|
515 |
-
common_keys = list(set(clip_keys).intersection(set(maple_keys)))
|
516 |
-
clip_only_keys = list(set(clip_keys) - (set(maple_keys)))
|
517 |
-
maple_only_keys = list(set(maple_keys) - (set(clip_keys)))
|
518 |
-
|
519 |
-
common_keys.sort(
|
520 |
-
key=lambda x: max(clip_neuron_dict[x], maple_neuron_dict[x]), reverse=True
|
521 |
-
)
|
522 |
-
clip_only_keys.sort(reverse=True)
|
523 |
-
maple_only_keys.sort(reverse=True)
|
524 |
-
|
525 |
-
out = []
|
526 |
-
out.extend([f"common-{i}" for i in common_keys[:5]])
|
527 |
-
out.extend([f"CLIP-{i}" for i in clip_only_keys[:5]])
|
528 |
-
out.extend([f"MaPLE-{i}" for i in maple_only_keys[:5]])
|
529 |
-
|
530 |
-
radio_choices = gr.Radio(
|
531 |
-
choices=out, label="Top activating SAE latent", value=out[0]
|
532 |
-
)
|
533 |
-
sleep(0.1)
|
534 |
-
return radio_choices
|
535 |
-
|
536 |
-
|
537 |
-
def update_markdown(option_value):
|
538 |
-
latent_idx = int(option_value.split("-")[-1])
|
539 |
-
out_1 = f"## Segmentation mask for the selected SAE latent - {latent_idx}"
|
540 |
-
out_2 = f"## Top reference images for the selected SAE latent - {latent_idx}"
|
541 |
-
return out_1, out_2
|
542 |
-
|
543 |
-
|
544 |
-
def get_data(image_name, model_name):
|
545 |
-
pkl_root = "./data/out"
|
546 |
-
data_dir = f"{pkl_root}/{model_name}/{image_name}.pkl.gz"
|
547 |
-
with gzip.open(data_dir, "rb") as f:
|
548 |
-
data = pickle.load(f)
|
549 |
-
out = data
|
550 |
-
|
551 |
-
return out
|
552 |
-
|
553 |
-
|
554 |
-
def update_all(selected_image, slider_value, toggle_btn, model_name):
|
555 |
-
(
|
556 |
-
seg_mask_display,
|
557 |
-
top_image_1,
|
558 |
-
top_image_2,
|
559 |
-
top_image_3,
|
560 |
-
act_value_1,
|
561 |
-
act_value_2,
|
562 |
-
act_value_3,
|
563 |
-
) = show_activation_heatmap_clip(selected_image, slider_value, toggle_btn)
|
564 |
-
seg_mask_display_maple = show_activation_heatmap_maple(
|
565 |
-
selected_image, slider_value, model_name
|
566 |
-
)
|
567 |
-
markdown_display, markdown_display_2 = update_markdown(slider_value)
|
568 |
-
|
569 |
-
return (
|
570 |
-
seg_mask_display,
|
571 |
-
seg_mask_display_maple,
|
572 |
-
top_image_1,
|
573 |
-
top_image_2,
|
574 |
-
top_image_3,
|
575 |
-
act_value_1,
|
576 |
-
act_value_2,
|
577 |
-
act_value_3,
|
578 |
-
markdown_display,
|
579 |
-
markdown_display_2,
|
580 |
-
)
|
581 |
-
|
582 |
-
|
583 |
-
def load_all_data(image_root, pkl_root):
|
584 |
-
image_files = glob(f"{image_root}/*")
|
585 |
-
data_dict = {}
|
586 |
-
for image_file in image_files:
|
587 |
-
image_name = os.path.basename(image_file).split(".")[0]
|
588 |
-
if image_file not in data_dict:
|
589 |
-
data_dict[image_name] = {
|
590 |
-
"image": Image.open(image_file).resize((IMAGE_SIZE, IMAGE_SIZE)),
|
591 |
-
"image_path": image_file,
|
592 |
-
}
|
593 |
-
|
594 |
-
sae_data_dict = {}
|
595 |
-
with open("./data/sae_data/mean_acts.pkl", "rb") as f:
|
596 |
-
data = pickle.load(f)
|
597 |
-
sae_data_dict["mean_acts"] = data
|
598 |
-
|
599 |
-
sae_data_dict["mean_act_values"] = {}
|
600 |
-
for dataset in ["imagenet", "imagenet-sketch", "caltech101"]:
|
601 |
-
with gzip.open(f"./data/sae_data/mean_act_values_{dataset}.pkl.gz", "rb") as f:
|
602 |
-
data = pickle.load(f)
|
603 |
-
sae_data_dict["mean_act_values"][dataset] = data
|
604 |
-
|
605 |
-
return data_dict, sae_data_dict
|
606 |
-
|
607 |
-
|
608 |
-
# data_dict, sae_data_dict = load_all_data(image_root="./data/image", pkl_root=pkl_root)
|
609 |
-
default_image_name = "christmas-imagenet"
|
610 |
-
|
611 |
-
|
612 |
-
with gr.Blocks(
|
613 |
-
theme=gr.themes.Citrus(),
|
614 |
-
css="""
|
615 |
-
.image-row .gr-image { margin: 0 !important; padding: 0 !important; }
|
616 |
-
.image-row img { width: auto; height: 50px; } /* Set a uniform height for all images */
|
617 |
-
""",
|
618 |
-
) as demo:
|
619 |
-
with gr.Row():
|
620 |
-
with gr.Column():
|
621 |
-
# Left View: Image selection and click handling
|
622 |
-
gr.Markdown("## Select input image and patch on the image")
|
623 |
-
image_selector = gr.Dropdown(
|
624 |
-
choices=list(data_dict.keys()),
|
625 |
-
value=default_image_name,
|
626 |
-
label="Select Image",
|
627 |
-
)
|
628 |
-
image_display = gr.Image(
|
629 |
-
value=data_dict[default_image_name]["image"],
|
630 |
-
type="pil",
|
631 |
-
interactive=True,
|
632 |
-
)
|
633 |
-
|
634 |
-
# Update image display when a new image is selected
|
635 |
-
image_selector.change(
|
636 |
-
fn=lambda img_name: data_dict[img_name]["image"],
|
637 |
-
inputs=image_selector,
|
638 |
-
outputs=image_display,
|
639 |
-
)
|
640 |
-
image_display.select(
|
641 |
-
fn=highlight_grid, inputs=[image_selector], outputs=[image_display]
|
642 |
-
)
|
643 |
-
|
644 |
-
with gr.Column():
|
645 |
-
gr.Markdown("## SAE latent activations of CLIP and MaPLE")
|
646 |
-
model_options = [f"MaPLE-{dataset_name}" for dataset_name in DATASET_LIST]
|
647 |
-
model_selector = gr.Dropdown(
|
648 |
-
choices=model_options,
|
649 |
-
value=model_options[0],
|
650 |
-
label="Select adapted model (MaPLe)",
|
651 |
-
)
|
652 |
-
init_plot = plot_activation_distribution(
|
653 |
-
None, default_image_name, model_options[0]
|
654 |
-
)
|
655 |
-
neuron_plot = gr.Plot(
|
656 |
-
label="Neuron Activation", value=init_plot, show_label=False
|
657 |
-
)
|
658 |
-
|
659 |
-
image_selector.change(
|
660 |
-
fn=plot_activation_distribution,
|
661 |
-
inputs=[image_selector, model_selector],
|
662 |
-
outputs=neuron_plot,
|
663 |
-
)
|
664 |
-
image_display.select(
|
665 |
-
fn=plot_activation_distribution,
|
666 |
-
inputs=[image_selector, model_selector],
|
667 |
-
outputs=neuron_plot,
|
668 |
-
)
|
669 |
-
model_selector.change(
|
670 |
-
fn=load_image, inputs=[image_selector], outputs=image_display
|
671 |
-
)
|
672 |
-
model_selector.change(
|
673 |
-
fn=plot_activation_distribution,
|
674 |
-
inputs=[image_selector, model_selector],
|
675 |
-
outputs=neuron_plot,
|
676 |
-
)
|
677 |
-
|
678 |
-
with gr.Row():
|
679 |
-
with gr.Column():
|
680 |
-
radio_names = get_init_radio_options(default_image_name, model_options[0])
|
681 |
-
|
682 |
-
feautre_idx = radio_names[0].split("-")[-1]
|
683 |
-
markdown_display = gr.Markdown(
|
684 |
-
f"## Segmentation mask for the selected SAE latent - {feautre_idx}"
|
685 |
-
)
|
686 |
-
init_seg, init_tops, init_values = show_activation_heatmap(
|
687 |
-
default_image_name, radio_names[0], "CLIP"
|
688 |
-
)
|
689 |
-
|
690 |
-
gr.Markdown("### Localize SAE latent activation using CLIP")
|
691 |
-
seg_mask_display = gr.Image(value=init_seg, type="pil", show_label=False)
|
692 |
-
init_seg_maple, _, _ = show_activation_heatmap(
|
693 |
-
default_image_name, radio_names[0], model_options[0]
|
694 |
-
)
|
695 |
-
gr.Markdown("### Localize SAE latent activation using MaPLE")
|
696 |
-
seg_mask_display_maple = gr.Image(
|
697 |
-
value=init_seg_maple, type="pil", show_label=False
|
698 |
-
)
|
699 |
-
|
700 |
-
with gr.Column():
|
701 |
-
gr.Markdown("## Top activating SAE latent index")
|
702 |
-
|
703 |
-
radio_choices = gr.Radio(
|
704 |
-
choices=radio_names,
|
705 |
-
label="Top activating SAE latent",
|
706 |
-
interactive=True,
|
707 |
-
value=radio_names[0],
|
708 |
-
)
|
709 |
-
toggle_btn = gr.Checkbox(label="Show segmentation mask", value=False)
|
710 |
-
|
711 |
-
markdown_display_2 = gr.Markdown(
|
712 |
-
f"## Top reference images for the selected SAE latent - {feautre_idx}"
|
713 |
-
)
|
714 |
-
|
715 |
-
gr.Markdown("### ImageNet")
|
716 |
-
top_image_1 = gr.Image(
|
717 |
-
value=init_tops[0], type="pil", label="ImageNet", show_label=False
|
718 |
-
)
|
719 |
-
act_value_1 = gr.Markdown(init_values[0])
|
720 |
-
|
721 |
-
gr.Markdown("### ImageNet-Sketch")
|
722 |
-
top_image_2 = gr.Image(
|
723 |
-
value=init_tops[1],
|
724 |
-
type="pil",
|
725 |
-
label="ImageNet-Sketch",
|
726 |
-
show_label=False,
|
727 |
-
)
|
728 |
-
act_value_2 = gr.Markdown(init_values[1])
|
729 |
-
|
730 |
-
gr.Markdown("### Caltech101")
|
731 |
-
top_image_3 = gr.Image(
|
732 |
-
value=init_tops[2], type="pil", label="Caltech101", show_label=False
|
733 |
-
)
|
734 |
-
act_value_3 = gr.Markdown(init_values[2])
|
735 |
-
|
736 |
-
image_display.select(
|
737 |
-
fn=update_radio_options,
|
738 |
-
inputs=[image_selector, model_selector],
|
739 |
-
outputs=[radio_choices],
|
740 |
-
)
|
741 |
-
|
742 |
-
model_selector.change(
|
743 |
-
fn=update_radio_options,
|
744 |
-
inputs=[image_selector, model_selector],
|
745 |
-
outputs=[radio_choices],
|
746 |
-
)
|
747 |
-
|
748 |
-
image_selector.select(
|
749 |
-
fn=update_radio_options,
|
750 |
-
inputs=[image_selector, model_selector],
|
751 |
-
outputs=[radio_choices],
|
752 |
-
)
|
753 |
-
|
754 |
-
radio_choices.change(
|
755 |
-
fn=update_all,
|
756 |
-
inputs=[image_selector, radio_choices, toggle_btn, model_selector],
|
757 |
-
outputs=[
|
758 |
-
seg_mask_display,
|
759 |
-
seg_mask_display_maple,
|
760 |
-
top_image_1,
|
761 |
-
top_image_2,
|
762 |
-
top_image_3,
|
763 |
-
act_value_1,
|
764 |
-
act_value_2,
|
765 |
-
act_value_3,
|
766 |
-
markdown_display,
|
767 |
-
markdown_display_2,
|
768 |
-
],
|
769 |
-
)
|
770 |
-
|
771 |
-
toggle_btn.change(
|
772 |
-
fn=show_activation_heatmap_clip,
|
773 |
-
inputs=[image_selector, radio_choices, toggle_btn],
|
774 |
-
outputs=[
|
775 |
-
seg_mask_display,
|
776 |
-
top_image_1,
|
777 |
-
top_image_2,
|
778 |
-
top_image_3,
|
779 |
-
act_value_1,
|
780 |
-
act_value_2,
|
781 |
-
act_value_3,
|
782 |
-
],
|
783 |
-
)
|
784 |
-
|
785 |
-
# Launch the app
|
786 |
-
# demo.queue()
|
787 |
-
# demo.launch()
|
788 |
-
|
789 |
-
|
790 |
-
if __name__ == "__main__":
|
791 |
-
demo.queue() # Enable queuing for better handling of concurrent users
|
792 |
-
demo.launch(
|
793 |
-
server_name="0.0.0.0", # Allow external access
|
794 |
-
server_port=7860,
|
795 |
-
share=False, # Set to True if you want to create a public URL
|
796 |
-
show_error=True,
|
797 |
-
# Optimize concurrency
|
798 |
-
max_threads=8, # Adjust based on your CPU cores
|
799 |
-
)
|
800 |
-
import gzip
|
801 |
-
import os
|
802 |
-
import pickle
|
803 |
-
from glob import glob
|
804 |
-
from time import sleep
|
805 |
-
|
806 |
-
from functools import lru_cache
|
807 |
-
import concurrent.futures
|
808 |
-
from typing import Dict, Tuple, List
|
809 |
-
|
810 |
-
import gradio as gr
|
811 |
-
import numpy as np
|
812 |
-
import plotly.graph_objects as go
|
813 |
-
import torch
|
814 |
-
from PIL import Image, ImageDraw
|
815 |
-
from plotly.subplots import make_subplots
|
816 |
-
|
817 |
-
IMAGE_SIZE = 400
|
818 |
-
DATASET_LIST = ["imagenet", "oxford_flowers", "ucf101", "caltech101", "dtd", "eurosat"]
|
819 |
-
GRID_NUM = 14
|
820 |
-
pkl_root = "./data/out"
|
821 |
-
preloaded_data = {}
|
822 |
-
|
823 |
-
|
824 |
-
# Global cache for data
|
825 |
-
_CACHE = {
|
826 |
-
'data_dict': {},
|
827 |
-
'sae_data_dict': {},
|
828 |
-
'model_data': {},
|
829 |
-
'segmasks': {},
|
830 |
-
'top_images': {}
|
831 |
-
}
|
832 |
-
|
833 |
-
def load_all_data(image_root: str, pkl_root: str) -> Tuple[Dict, Dict]:
|
834 |
-
"""Load all data with optimized parallel processing."""
|
835 |
-
# Load images in parallel
|
836 |
-
with concurrent.futures.ThreadPoolExecutor() as executor:
|
837 |
-
image_files = glob(f"{image_root}/*")
|
838 |
-
future_to_file = {
|
839 |
-
executor.submit(_load_image_file, image_file): image_file
|
840 |
-
for image_file in image_files
|
841 |
-
}
|
842 |
-
|
843 |
-
for future in concurrent.futures.as_completed(future_to_file):
|
844 |
-
image_file = future_to_file[future]
|
845 |
-
image_name = os.path.basename(image_file).split(".")[0]
|
846 |
-
result = future.result()
|
847 |
-
if result is not None:
|
848 |
-
_CACHE['data_dict'][image_name] = result
|
849 |
-
|
850 |
-
# Load SAE data
|
851 |
-
with open("./data/sae_data/mean_acts.pkl", "rb") as f:
|
852 |
-
_CACHE['sae_data_dict']["mean_acts"] = pickle.load(f)
|
853 |
-
|
854 |
-
# Load mean act values in parallel
|
855 |
-
datasets = ["imagenet", "imagenet-sketch", "caltech101"]
|
856 |
-
_CACHE['sae_data_dict']["mean_act_values"] = {}
|
857 |
-
|
858 |
-
with concurrent.futures.ThreadPoolExecutor() as executor:
|
859 |
-
future_to_dataset = {
|
860 |
-
executor.submit(_load_mean_act_values, dataset): dataset
|
861 |
-
for dataset in datasets
|
862 |
-
}
|
863 |
-
|
864 |
-
for future in concurrent.futures.as_completed(future_to_dataset):
|
865 |
-
dataset = future_to_dataset[future]
|
866 |
-
result = future.result()
|
867 |
-
if result is not None:
|
868 |
-
_CACHE['sae_data_dict']["mean_act_values"][dataset] = result
|
869 |
-
|
870 |
-
return _CACHE['data_dict'], _CACHE['sae_data_dict']
|
871 |
-
|
872 |
-
def _load_image_file(image_file: str) -> Dict:
|
873 |
-
"""Helper function to load a single image file."""
|
874 |
-
try:
|
875 |
-
image = Image.open(image_file).resize((IMAGE_SIZE, IMAGE_SIZE))
|
876 |
-
return {
|
877 |
-
"image": image,
|
878 |
-
"image_path": image_file,
|
879 |
-
}
|
880 |
-
except Exception as e:
|
881 |
-
print(f"Error loading {image_file}: {e}")
|
882 |
-
return None
|
883 |
-
|
884 |
-
def _load_mean_act_values(dataset: str) -> np.ndarray:
|
885 |
-
"""Helper function to load mean act values for a dataset."""
|
886 |
-
try:
|
887 |
-
with gzip.open(f"./data/sae_data/mean_act_values_{dataset}.pkl.gz", "rb") as f:
|
888 |
-
return pickle.load(f)
|
889 |
-
except Exception as e:
|
890 |
-
print(f"Error loading mean act values for {dataset}: {e}")
|
891 |
-
return None
|
892 |
-
|
893 |
-
@lru_cache(maxsize=1024)
|
894 |
-
def get_data(image_name: str, model_name: str) -> np.ndarray:
|
895 |
-
"""Cached function to get model data."""
|
896 |
-
cache_key = f"{model_name}_{image_name}"
|
897 |
-
if cache_key not in _CACHE['model_data']:
|
898 |
-
data_dir = f"{pkl_root}/{model_name}/{image_name}.pkl.gz"
|
899 |
-
with gzip.open(data_dir, "rb") as f:
|
900 |
-
_CACHE['model_data'][cache_key] = pickle.load(f)
|
901 |
-
return _CACHE['model_data'][cache_key]
|
902 |
-
|
903 |
-
@lru_cache(maxsize=1024)
|
904 |
-
def get_activation_distribution(image_name: str, model_type: str) -> np.ndarray:
|
905 |
-
"""Cached function to get activation distribution."""
|
906 |
-
activation = get_data(image_name, model_type)[0]
|
907 |
-
noisy_features_indices = (
|
908 |
-
(_CACHE['sae_data_dict']["mean_acts"]["imagenet"] > 0.1).nonzero()[0].tolist()
|
909 |
-
)
|
910 |
-
activation[:, noisy_features_indices] = 0
|
911 |
-
return activation
|
912 |
-
|
913 |
-
@lru_cache(maxsize=1024)
|
914 |
-
def get_segmask(selected_image: str, slider_value: int, model_type: str) -> np.ndarray:
|
915 |
-
"""Cached function to get segmentation mask."""
|
916 |
-
cache_key = f"{selected_image}_{slider_value}_{model_type}"
|
917 |
-
if cache_key not in _CACHE['segmasks']:
|
918 |
-
image = _CACHE['data_dict'][selected_image]["image"]
|
919 |
-
sae_act = get_data(selected_image, model_type)[0]
|
920 |
-
temp = sae_act[:, slider_value]
|
921 |
-
|
922 |
-
mask = torch.Tensor(temp[1:].reshape(14, 14)).view(1, 1, 14, 14)
|
923 |
-
mask = torch.nn.functional.interpolate(mask, (image.height, image.width))[0][0].numpy()
|
924 |
-
mask = (mask - mask.min()) / (mask.max() - mask.min() + 1e-10)
|
925 |
-
|
926 |
-
base_opacity = 30
|
927 |
-
image_array = np.array(image)[..., :3]
|
928 |
-
rgba_overlay = np.zeros((mask.shape[0], mask.shape[1], 4), dtype=np.uint8)
|
929 |
-
rgba_overlay[..., :3] = image_array[..., :3]
|
930 |
-
|
931 |
-
darkened_image = (image_array[..., :3] * (base_opacity / 255)).astype(np.uint8)
|
932 |
-
rgba_overlay[mask == 0, :3] = darkened_image[mask == 0]
|
933 |
-
rgba_overlay[..., 3] = 255
|
934 |
-
|
935 |
-
_CACHE['segmasks'][cache_key] = rgba_overlay
|
936 |
-
|
937 |
-
return _CACHE['segmasks'][cache_key]
|
938 |
-
|
939 |
-
@lru_cache(maxsize=1024)
|
940 |
-
def get_top_images(slider_value: int, toggle_btn: bool) -> List[Image.Image]:
|
941 |
-
"""Cached function to get top images."""
|
942 |
-
cache_key = f"{slider_value}_{toggle_btn}"
|
943 |
-
if cache_key not in _CACHE['top_images']:
|
944 |
-
dataset_path = "./data/top_images_masked" if toggle_btn else "./data/top_images"
|
945 |
-
paths = [
|
946 |
-
os.path.join(dataset_path, dataset, f"{slider_value}.jpg")
|
947 |
-
for dataset in ["imagenet", "imagenet-sketch", "caltech101"]
|
948 |
-
]
|
949 |
-
|
950 |
-
_CACHE['top_images'][cache_key] = [
|
951 |
-
Image.open(path) if os.path.exists(path) else Image.new("RGB", (256, 256), (255, 255, 255))
|
952 |
-
for path in paths
|
953 |
-
]
|
954 |
-
|
955 |
-
return _CACHE['top_images'][cache_key]
|
956 |
-
|
957 |
|
958 |
# def preload_activation(image_name):
|
959 |
# for model in ["CLIP"] + [f"MaPLE-{ds}" for ds in DATASET_LIST]:
|
@@ -1413,9 +614,6 @@ def preload_all_model_data():
|
|
1413 |
except Exception as e:
|
1414 |
print(f"Error preloading {cache_key}: {e}")
|
1415 |
|
1416 |
-
# Add to initialization
|
1417 |
-
preload_all_model_data()
|
1418 |
-
|
1419 |
def precompute_activations():
|
1420 |
"""Precompute and cache common activation patterns"""
|
1421 |
print("Precomputing activations...")
|
@@ -1425,11 +623,7 @@ def precompute_activations():
|
|
1425 |
cache_key = f"activation_{model_name}_{image_name}"
|
1426 |
_CACHE['precomputed_activations'][cache_key] = activation.mean(0)
|
1427 |
|
1428 |
-
# Add to _CACHE initialization
|
1429 |
-
_CACHE['precomputed_activations'] = {}
|
1430 |
|
1431 |
-
# Add to initialization
|
1432 |
-
precompute_activations()
|
1433 |
|
1434 |
def precompute_segmasks():
|
1435 |
"""Precompute common segmentation masks"""
|
@@ -1444,13 +638,6 @@ def precompute_segmasks():
|
|
1444 |
except Exception as e:
|
1445 |
print(f"Error precomputing mask {cache_key}: {e}")
|
1446 |
|
1447 |
-
# Add to initialization
|
1448 |
-
precompute_segmasks()
|
1449 |
-
|
1450 |
-
|
1451 |
-
data_dict, sae_data_dict = load_all_data(image_root="./data/image", pkl_root=pkl_root)
|
1452 |
-
default_image_name = "christmas-imagenet"
|
1453 |
-
|
1454 |
|
1455 |
with gr.Blocks(
|
1456 |
theme=gr.themes.Citrus(),
|
@@ -1672,6 +859,18 @@ if __name__ == "__main__":
|
|
1672 |
import threading
|
1673 |
start_memory_monitor()
|
1674 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1675 |
# Launch the app with memory-optimized settings
|
1676 |
demo.queue(max_size=min(20, int(total_ram_gb))) # Scale queue with RAM
|
1677 |
demo.launch(
|
|
|
155 |
|
156 |
return _CACHE['top_images'][cache_key]
|
157 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
158 |
|
159 |
# def preload_activation(image_name):
|
160 |
# for model in ["CLIP"] + [f"MaPLE-{ds}" for ds in DATASET_LIST]:
|
|
|
614 |
except Exception as e:
|
615 |
print(f"Error preloading {cache_key}: {e}")
|
616 |
|
|
|
|
|
|
|
617 |
def precompute_activations():
|
618 |
"""Precompute and cache common activation patterns"""
|
619 |
print("Precomputing activations...")
|
|
|
623 |
cache_key = f"activation_{model_name}_{image_name}"
|
624 |
_CACHE['precomputed_activations'][cache_key] = activation.mean(0)
|
625 |
|
|
|
|
|
626 |
|
|
|
|
|
627 |
|
628 |
def precompute_segmasks():
|
629 |
"""Precompute common segmentation masks"""
|
|
|
638 |
except Exception as e:
|
639 |
print(f"Error precomputing mask {cache_key}: {e}")
|
640 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
641 |
|
642 |
with gr.Blocks(
|
643 |
theme=gr.themes.Citrus(),
|
|
|
859 |
import threading
|
860 |
start_memory_monitor()
|
861 |
|
862 |
+
|
863 |
+
# Add to initialization
|
864 |
+
preload_all_model_data()
|
865 |
+
|
866 |
+
_CACHE['precomputed_activations'] = {}
|
867 |
+
precompute_activations()
|
868 |
+
precompute_segmasks()
|
869 |
+
|
870 |
+
data_dict, sae_data_dict = load_all_data(image_root="./data/image", pkl_root=pkl_root)
|
871 |
+
default_image_name = "christmas-imagenet"
|
872 |
+
|
873 |
+
|
874 |
# Launch the app with memory-optimized settings
|
875 |
demo.queue(max_size=min(20, int(total_ram_gb))) # Scale queue with RAM
|
876 |
demo.launch(
|