Samuel Stevens
commited on
Commit
·
af47b42
1
Parent(s):
6c9f92c
Add legend; add image uploader
Browse files
app.py
CHANGED
@@ -52,6 +52,44 @@ N_SAE_LATENTS = 2
|
|
52 |
N_LATENT_EXAMPLES = 4
|
53 |
"""Number of examples per SAE latent to show."""
|
54 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
##########
|
56 |
# Models #
|
57 |
##########
|
@@ -112,9 +150,9 @@ def load_tensors() -> tuple[
|
|
112 |
return top_img_i, top_values, mask
|
113 |
|
114 |
|
115 |
-
|
116 |
-
#
|
117 |
-
|
118 |
|
119 |
|
120 |
@jaxtyped(typechecker=beartype.beartype)
|
@@ -154,65 +192,43 @@ def add_highlights(
|
|
154 |
return Image.alpha_composite(img.convert("RGBA"), overlay)
|
155 |
|
156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
#######################
|
158 |
# Inference Functions #
|
159 |
#######################
|
160 |
|
161 |
|
162 |
@beartype.beartype
|
163 |
-
|
164 |
-
"""Represents an example image and its associated label.
|
165 |
-
|
166 |
-
Used to store examples of SAE latent activations for visualization.
|
167 |
-
"""
|
168 |
-
|
169 |
-
index: int
|
170 |
-
"""Dataset index."""
|
171 |
-
orig_url: str
|
172 |
-
"""The URL or path to access the original example image."""
|
173 |
-
highlighted_url: str
|
174 |
-
"""The URL or path to access the SAE-highlighted image."""
|
175 |
-
seg_url: str
|
176 |
-
"""Base64-encoded version of the colored segmentation map."""
|
177 |
-
|
178 |
-
|
179 |
-
@beartype.beartype
|
180 |
-
class SaeActivation(typing.TypedDict):
|
181 |
-
"""Represents the activation pattern of a single SAE latent across patches.
|
182 |
-
|
183 |
-
This captures how strongly a particular SAE latent fires on different patches of an input image.
|
184 |
-
"""
|
185 |
-
|
186 |
-
latent: int
|
187 |
-
"""The index of the SAE latent being measured."""
|
188 |
-
|
189 |
-
highlighted_url: str
|
190 |
-
"""The image with the colormaps applied."""
|
191 |
-
|
192 |
-
activations: list[float]
|
193 |
-
"""The activation values of this latent across different patches. Each value represents how strongly this latent fired on a particular patch."""
|
194 |
-
|
195 |
-
examples: list[Example]
|
196 |
-
"""Top examples for this latent."""
|
197 |
-
|
198 |
-
|
199 |
-
@beartype.beartype
|
200 |
-
def get_img(i: int) -> dict[str, object]:
|
201 |
img_sized = data.to_sized(data.get_img(i))
|
202 |
seg_sized = data.to_sized(data.get_seg(i))
|
203 |
seg_u8_sized = data.to_u8(seg_sized)
|
204 |
seg_img_sized = data.u8_to_img(seg_u8_sized)
|
205 |
|
206 |
return {
|
207 |
-
"index": i,
|
208 |
"orig_url": data.img_to_base64(img_sized),
|
209 |
"seg_url": data.img_to_base64(seg_img_sized),
|
|
|
210 |
}
|
211 |
|
212 |
|
213 |
@beartype.beartype
|
214 |
@torch.inference_mode
|
215 |
-
def get_sae_latents(
|
216 |
"""
|
217 |
Given a particular cell, returns some highlighted images showing what feature fires most on this cell.
|
218 |
"""
|
@@ -222,9 +238,7 @@ def get_sae_latents(img_i: int, patches: list[int]) -> list[SaeActivation]:
|
|
222 |
split_vit, vit_transform = modeling.load_vit(DEVICE)
|
223 |
sae = load_sae(DEVICE)
|
224 |
|
225 |
-
|
226 |
-
|
227 |
-
x_BCWH = vit_transform(img)[None, ...].to(DEVICE)
|
228 |
|
229 |
x_BPD = split_vit.forward_start(x_BCWH)
|
230 |
x_BPD = (
|
@@ -274,10 +288,10 @@ def get_sae_latents(img_i: int, patches: list[int]) -> list[SaeActivation]:
|
|
274 |
)
|
275 |
|
276 |
examples.append({
|
277 |
-
"index": i_im,
|
278 |
"orig_url": data.img_to_base64(img_sized),
|
279 |
"highlighted_url": data.img_to_base64(highlighted_sized),
|
280 |
"seg_url": data.img_to_base64(seg_img_sized),
|
|
|
281 |
})
|
282 |
|
283 |
sae_activations.append({
|
@@ -288,12 +302,12 @@ def get_sae_latents(img_i: int, patches: list[int]) -> list[SaeActivation]:
|
|
288 |
return sae_activations
|
289 |
|
290 |
|
|
|
291 |
@torch.inference_mode
|
292 |
-
def get_orig_preds(
|
293 |
-
img = data.get_img(i)
|
294 |
split_vit, vit_transform = modeling.load_vit(DEVICE)
|
295 |
|
296 |
-
x_BCWH = vit_transform(img)[None, ...].to(DEVICE)
|
297 |
|
298 |
x_BPD = split_vit.forward_start(x_BCWH)
|
299 |
x_BPD = split_vit.forward_end(x_BPD)
|
@@ -304,11 +318,10 @@ def get_orig_preds(i: int) -> dict[str, object]:
|
|
304 |
logits_WHC = clf(x_WHD)
|
305 |
|
306 |
pred_WH = logits_WHC.argmax(axis=-1)
|
307 |
-
# preds = einops.rearrange(pred_WH, "w h -> (w h)").tolist()
|
308 |
return {
|
309 |
-
"index": i,
|
310 |
"orig_url": data.img_to_base64(data.to_sized(img)),
|
311 |
"seg_url": data.img_to_base64(data.u8_to_img(upsample(pred_WH))),
|
|
|
312 |
}
|
313 |
|
314 |
|
@@ -333,16 +346,15 @@ def map_range(
|
|
333 |
|
334 |
@beartype.beartype
|
335 |
@torch.inference_mode
|
336 |
-
def get_mod_preds(
|
337 |
latents = {int(k): float(v) for k, v in latents.items()}
|
338 |
-
img = data.get_img(i)
|
339 |
|
340 |
split_vit, vit_transform = modeling.load_vit(DEVICE)
|
341 |
sae = load_sae(DEVICE)
|
342 |
_, top_values, _ = load_tensors()
|
343 |
clf = load_clf()
|
344 |
|
345 |
-
x_BCWH = vit_transform(img)[None, ...].to(DEVICE)
|
346 |
x_BPD = split_vit.forward_start(x_BCWH)
|
347 |
x_hat_BPD, f_x_BPS, _ = sae(x_BPD)
|
348 |
|
@@ -375,27 +387,12 @@ def get_mod_preds(i: int, latents: dict[str, int | float]) -> dict[str, object]:
|
|
375 |
pred_WH = logits_WHC.argmax(axis=-1)
|
376 |
# pred_WH = einops.rearrange(pred_P, "(w h) -> w h", w=16, h=16)
|
377 |
return {
|
378 |
-
"index": i,
|
379 |
"orig_url": data.img_to_base64(data.to_sized(img)),
|
380 |
"seg_url": data.img_to_base64(data.u8_to_img(upsample(pred_WH))),
|
|
|
381 |
}
|
382 |
|
383 |
|
384 |
-
@jaxtyped(typechecker=beartype.beartype)
|
385 |
-
@torch.inference_mode
|
386 |
-
def upsample(
|
387 |
-
x_WH: Int[Tensor, "width_ps height_ps"],
|
388 |
-
) -> UInt8[Tensor, "width_px height_px"]:
|
389 |
-
return (
|
390 |
-
torch.nn.functional.interpolate(
|
391 |
-
x_WH.view((1, 1, 16, 16)).float(),
|
392 |
-
scale_factor=28,
|
393 |
-
)
|
394 |
-
.view((448, 448))
|
395 |
-
.type(torch.uint8)
|
396 |
-
)
|
397 |
-
|
398 |
-
|
399 |
with gr.Blocks() as demo:
|
400 |
###########
|
401 |
# get-img #
|
@@ -418,13 +415,19 @@ with gr.Blocks() as demo:
|
|
418 |
|
419 |
# Inputs
|
420 |
patches_json = gr.JSON(label="Patches", value=[])
|
|
|
|
|
|
|
|
|
|
|
|
|
421 |
# Outputs
|
422 |
get_sae_latents_out = gr.JSON(label="get_sae_latents_out", value=[])
|
423 |
|
424 |
get_sae_latents_btn = gr.Button(value="Get SAE Latents")
|
425 |
get_sae_latents_btn.click(
|
426 |
get_sae_latents,
|
427 |
-
inputs=[
|
428 |
outputs=[get_sae_latents_out],
|
429 |
api_name="get-sae-latents",
|
430 |
)
|
@@ -439,7 +442,7 @@ with gr.Blocks() as demo:
|
|
439 |
get_pred_labels_btn = gr.Button(value="Get Predictions")
|
440 |
get_pred_labels_btn.click(
|
441 |
get_orig_preds,
|
442 |
-
inputs=[
|
443 |
outputs=[get_orig_preds_out],
|
444 |
api_name="get-orig-preds",
|
445 |
)
|
@@ -457,7 +460,7 @@ with gr.Blocks() as demo:
|
|
457 |
get_pred_labels_btn = gr.Button(value="Get Predictions")
|
458 |
get_pred_labels_btn.click(
|
459 |
get_mod_preds,
|
460 |
-
inputs=[
|
461 |
outputs=[get_mod_preds_out],
|
462 |
api_name="get-mod-preds",
|
463 |
)
|
|
|
52 |
N_LATENT_EXAMPLES = 4
|
53 |
"""Number of examples per SAE latent to show."""
|
54 |
|
55 |
+
|
56 |
+
@beartype.beartype
|
57 |
+
class Example(typing.TypedDict):
|
58 |
+
"""Represents an example image and its associated label.
|
59 |
+
|
60 |
+
Used to store examples of SAE latent activations for visualization.
|
61 |
+
"""
|
62 |
+
|
63 |
+
orig_url: str
|
64 |
+
"""The URL or path to access the original example image."""
|
65 |
+
highlighted_url: typing.NotRequired[str]
|
66 |
+
"""The URL or path to access the SAE-highlighted image."""
|
67 |
+
seg_url: str
|
68 |
+
"""Base64-encoded version of the colored segmentation map."""
|
69 |
+
classes: list[int]
|
70 |
+
"""Unique list of all classes in the seg_url."""
|
71 |
+
|
72 |
+
|
73 |
+
@beartype.beartype
|
74 |
+
class SaeActivation(typing.TypedDict):
|
75 |
+
"""Represents the activation pattern of a single SAE latent across patches.
|
76 |
+
|
77 |
+
This captures how strongly a particular SAE latent fires on different patches of an input image.
|
78 |
+
"""
|
79 |
+
|
80 |
+
latent: int
|
81 |
+
"""The index of the SAE latent being measured."""
|
82 |
+
|
83 |
+
highlighted_url: str
|
84 |
+
"""The image with the colormaps applied."""
|
85 |
+
|
86 |
+
activations: list[float]
|
87 |
+
"""The activation values of this latent across different patches. Each value represents how strongly this latent fired on a particular patch."""
|
88 |
+
|
89 |
+
examples: list[Example]
|
90 |
+
"""Top examples for this latent."""
|
91 |
+
|
92 |
+
|
93 |
##########
|
94 |
# Models #
|
95 |
##########
|
|
|
150 |
return top_img_i, top_values, mask
|
151 |
|
152 |
|
153 |
+
###########
|
154 |
+
# Imaging #
|
155 |
+
###########
|
156 |
|
157 |
|
158 |
@jaxtyped(typechecker=beartype.beartype)
|
|
|
192 |
return Image.alpha_composite(img.convert("RGBA"), overlay)
|
193 |
|
194 |
|
195 |
+
@jaxtyped(typechecker=beartype.beartype)
|
196 |
+
@torch.inference_mode
|
197 |
+
def upsample(
|
198 |
+
x_WH: Int[Tensor, "width_ps height_ps"],
|
199 |
+
) -> UInt8[Tensor, "width_px height_px"]:
|
200 |
+
return (
|
201 |
+
torch.nn.functional.interpolate(
|
202 |
+
x_WH.view((1, 1, 16, 16)).float(),
|
203 |
+
scale_factor=28,
|
204 |
+
)
|
205 |
+
.view((448, 448))
|
206 |
+
.type(torch.uint8)
|
207 |
+
)
|
208 |
+
|
209 |
+
|
210 |
#######################
|
211 |
# Inference Functions #
|
212 |
#######################
|
213 |
|
214 |
|
215 |
@beartype.beartype
|
216 |
+
def get_img(i: int) -> Example:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
217 |
img_sized = data.to_sized(data.get_img(i))
|
218 |
seg_sized = data.to_sized(data.get_seg(i))
|
219 |
seg_u8_sized = data.to_u8(seg_sized)
|
220 |
seg_img_sized = data.u8_to_img(seg_u8_sized)
|
221 |
|
222 |
return {
|
|
|
223 |
"orig_url": data.img_to_base64(img_sized),
|
224 |
"seg_url": data.img_to_base64(seg_img_sized),
|
225 |
+
"classes": data.to_classes(seg_u8_sized),
|
226 |
}
|
227 |
|
228 |
|
229 |
@beartype.beartype
|
230 |
@torch.inference_mode
|
231 |
+
def get_sae_latents(img: Image.Image, patches: list[int]) -> list[SaeActivation]:
|
232 |
"""
|
233 |
Given a particular cell, returns some highlighted images showing what feature fires most on this cell.
|
234 |
"""
|
|
|
238 |
split_vit, vit_transform = modeling.load_vit(DEVICE)
|
239 |
sae = load_sae(DEVICE)
|
240 |
|
241 |
+
x_BCWH = vit_transform(img.convert("RGB"))[None, ...].to(DEVICE)
|
|
|
|
|
242 |
|
243 |
x_BPD = split_vit.forward_start(x_BCWH)
|
244 |
x_BPD = (
|
|
|
288 |
)
|
289 |
|
290 |
examples.append({
|
|
|
291 |
"orig_url": data.img_to_base64(img_sized),
|
292 |
"highlighted_url": data.img_to_base64(highlighted_sized),
|
293 |
"seg_url": data.img_to_base64(seg_img_sized),
|
294 |
+
"classes": data.to_classes(seg_u8_sized),
|
295 |
})
|
296 |
|
297 |
sae_activations.append({
|
|
|
302 |
return sae_activations
|
303 |
|
304 |
|
305 |
+
@beartype.beartype
|
306 |
@torch.inference_mode
|
307 |
+
def get_orig_preds(img: Image.Image) -> Example:
|
|
|
308 |
split_vit, vit_transform = modeling.load_vit(DEVICE)
|
309 |
|
310 |
+
x_BCWH = vit_transform(img.convert("RGB"))[None, ...].to(DEVICE)
|
311 |
|
312 |
x_BPD = split_vit.forward_start(x_BCWH)
|
313 |
x_BPD = split_vit.forward_end(x_BPD)
|
|
|
318 |
logits_WHC = clf(x_WHD)
|
319 |
|
320 |
pred_WH = logits_WHC.argmax(axis=-1)
|
|
|
321 |
return {
|
|
|
322 |
"orig_url": data.img_to_base64(data.to_sized(img)),
|
323 |
"seg_url": data.img_to_base64(data.u8_to_img(upsample(pred_WH))),
|
324 |
+
"classes": data.to_classes(pred_WH),
|
325 |
}
|
326 |
|
327 |
|
|
|
346 |
|
347 |
@beartype.beartype
|
348 |
@torch.inference_mode
|
349 |
+
def get_mod_preds(img: Image.Image, latents: dict[str, int | float]) -> Example:
|
350 |
latents = {int(k): float(v) for k, v in latents.items()}
|
|
|
351 |
|
352 |
split_vit, vit_transform = modeling.load_vit(DEVICE)
|
353 |
sae = load_sae(DEVICE)
|
354 |
_, top_values, _ = load_tensors()
|
355 |
clf = load_clf()
|
356 |
|
357 |
+
x_BCWH = vit_transform(img.convert("RGB"))[None, ...].to(DEVICE)
|
358 |
x_BPD = split_vit.forward_start(x_BCWH)
|
359 |
x_hat_BPD, f_x_BPS, _ = sae(x_BPD)
|
360 |
|
|
|
387 |
pred_WH = logits_WHC.argmax(axis=-1)
|
388 |
# pred_WH = einops.rearrange(pred_P, "(w h) -> w h", w=16, h=16)
|
389 |
return {
|
|
|
390 |
"orig_url": data.img_to_base64(data.to_sized(img)),
|
391 |
"seg_url": data.img_to_base64(data.u8_to_img(upsample(pred_WH))),
|
392 |
+
"classes": data.to_classes(pred_WH),
|
393 |
}
|
394 |
|
395 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
396 |
with gr.Blocks() as demo:
|
397 |
###########
|
398 |
# get-img #
|
|
|
415 |
|
416 |
# Inputs
|
417 |
patches_json = gr.JSON(label="Patches", value=[])
|
418 |
+
input_img = gr.Image(
|
419 |
+
label="Input Image",
|
420 |
+
sources=["upload", "clipboard"],
|
421 |
+
type="pil",
|
422 |
+
interactive=True,
|
423 |
+
)
|
424 |
# Outputs
|
425 |
get_sae_latents_out = gr.JSON(label="get_sae_latents_out", value=[])
|
426 |
|
427 |
get_sae_latents_btn = gr.Button(value="Get SAE Latents")
|
428 |
get_sae_latents_btn.click(
|
429 |
get_sae_latents,
|
430 |
+
inputs=[input_img, patches_json],
|
431 |
outputs=[get_sae_latents_out],
|
432 |
api_name="get-sae-latents",
|
433 |
)
|
|
|
442 |
get_pred_labels_btn = gr.Button(value="Get Predictions")
|
443 |
get_pred_labels_btn.click(
|
444 |
get_orig_preds,
|
445 |
+
inputs=[input_img],
|
446 |
outputs=[get_orig_preds_out],
|
447 |
api_name="get-orig-preds",
|
448 |
)
|
|
|
460 |
get_pred_labels_btn = gr.Button(value="Get Predictions")
|
461 |
get_pred_labels_btn.click(
|
462 |
get_mod_preds,
|
463 |
+
inputs=[input_img, latents_json],
|
464 |
outputs=[get_mod_preds_out],
|
465 |
api_name="get-mod-preds",
|
466 |
)
|
data.py
CHANGED
@@ -8,7 +8,7 @@ import beartype
|
|
8 |
import einops.layers.torch
|
9 |
import numpy as np
|
10 |
import requests
|
11 |
-
from jaxtyping import UInt8, jaxtyped
|
12 |
from PIL import Image
|
13 |
from torch import Tensor
|
14 |
from torchvision.transforms import v2
|
@@ -48,12 +48,13 @@ def make_colors() -> UInt8[np.ndarray, "n 3"]:
|
|
48 |
random.Random(42).shuffle(colors)
|
49 |
colors = np.array(colors, dtype=np.uint8)
|
50 |
|
51 |
-
# Fixed colors
|
52 |
colors[2] = np.array([201, 249, 255], dtype=np.uint8)
|
53 |
colors[4] = np.array([151, 204, 4], dtype=np.uint8)
|
54 |
colors[13] = np.array([104, 139, 88], dtype=np.uint8)
|
55 |
colors[16] = np.array([54, 48, 32], dtype=np.uint8)
|
56 |
colors[26] = np.array([45, 125, 210], dtype=np.uint8)
|
|
|
57 |
colors[46] = np.array([238, 185, 2], dtype=np.uint8)
|
58 |
colors[52] = np.array([88, 91, 86], dtype=np.uint8)
|
59 |
colors[72] = np.array([76, 46, 5], dtype=np.uint8)
|
@@ -97,6 +98,12 @@ def u8_to_img(map: UInt8[Tensor, "width height"]) -> Image.Image:
|
|
97 |
return Image.fromarray(colored)
|
98 |
|
99 |
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
@beartype.beartype
|
101 |
def img_to_base64(img: Image.Image) -> str:
|
102 |
buf = io.BytesIO()
|
|
|
8 |
import einops.layers.torch
|
9 |
import numpy as np
|
10 |
import requests
|
11 |
+
from jaxtyping import Integer, UInt8, jaxtyped
|
12 |
from PIL import Image
|
13 |
from torch import Tensor
|
14 |
from torchvision.transforms import v2
|
|
|
48 |
random.Random(42).shuffle(colors)
|
49 |
colors = np.array(colors, dtype=np.uint8)
|
50 |
|
51 |
+
# Fixed colors. Must be synced with Segmentation.elm.
|
52 |
colors[2] = np.array([201, 249, 255], dtype=np.uint8)
|
53 |
colors[4] = np.array([151, 204, 4], dtype=np.uint8)
|
54 |
colors[13] = np.array([104, 139, 88], dtype=np.uint8)
|
55 |
colors[16] = np.array([54, 48, 32], dtype=np.uint8)
|
56 |
colors[26] = np.array([45, 125, 210], dtype=np.uint8)
|
57 |
+
colors[29] = np.array([116, 142, 84], dtype=np.uint8)
|
58 |
colors[46] = np.array([238, 185, 2], dtype=np.uint8)
|
59 |
colors[52] = np.array([88, 91, 86], dtype=np.uint8)
|
60 |
colors[72] = np.array([76, 46, 5], dtype=np.uint8)
|
|
|
98 |
return Image.fromarray(colored)
|
99 |
|
100 |
|
101 |
+
@jaxtyped(typechecker=beartype.beartype)
|
102 |
+
def to_classes(map: Integer[Tensor, "width height"]) -> list[int]:
|
103 |
+
# Integer is any signed or unsigned int: https://docs.kidger.site/jaxtyping/api/array/#dtype
|
104 |
+
return list(set(map.view(-1).tolist()))
|
105 |
+
|
106 |
+
|
107 |
@beartype.beartype
|
108 |
def img_to_base64(img: Image.Image) -> str:
|
109 |
buf = io.BytesIO()
|