Spaces:
Runtime error
Runtime error
using groups for command line options
Browse files
msma.py
CHANGED
@@ -3,6 +3,7 @@ import os
|
|
3 |
import pickle
|
4 |
from functools import partial
|
5 |
from pickle import dump, load
|
|
|
6 |
|
7 |
import click
|
8 |
import numpy as np
|
@@ -20,6 +21,7 @@ import dnnlib
|
|
20 |
from dataset import ImageFolderDataset
|
21 |
from flowutils import PatchFlow
|
22 |
|
|
|
23 |
model_root = "https://nvlabs-fi-cdn.nvidia.com/edm2/posthoc-reconstructions"
|
24 |
|
25 |
config_presets = {
|
@@ -100,6 +102,7 @@ class ScoreFlow(torch.nn.Module):
|
|
100 |
self,
|
101 |
preset,
|
102 |
device="cpu",
|
|
|
103 |
):
|
104 |
super().__init__()
|
105 |
|
@@ -107,7 +110,7 @@ class ScoreFlow(torch.nn.Module):
|
|
107 |
h = w = scorenet.net.img_resolution
|
108 |
c = scorenet.net.img_channels
|
109 |
num_sigmas = len(scorenet.sigma_steps)
|
110 |
-
self.flow = PatchFlow((num_sigmas, c, h, w))
|
111 |
|
112 |
self.flow = self.flow.to(device)
|
113 |
self.scorenet = scorenet.to(device).requires_grad_(False)
|
@@ -187,7 +190,78 @@ def compute_gmm_likelihood(x_score, gmmdir):
|
|
187 |
return nll, percentile
|
188 |
|
189 |
|
190 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
191 |
dsobj = ImageFolderDataset(path=dataset_path, resolution=64)
|
192 |
refimg, reflabel = dsobj[0]
|
193 |
print(f"Loading dataset from {dataset_path}")
|
@@ -215,7 +289,40 @@ def cache_score_norms(preset, dataset_path, outdir, device="cpu"):
|
|
215 |
print(f"Computed score norms for {score_norms.shape[0]} samples")
|
216 |
|
217 |
|
218 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
219 |
dsobj = ImageFolderDataset(path=dataset_path, resolution=64)
|
220 |
refimg, reflabel = dsobj[0]
|
221 |
print(f"Loaded {len(dsobj)} samples from {dataset_path}")
|
@@ -238,7 +345,7 @@ def train_flow(dataset_path, preset, outdir, epochs=10, device="cuda"):
|
|
238 |
val_ds, batch_size=128, num_workers=4, prefetch_factor=2
|
239 |
)
|
240 |
|
241 |
-
model = ScoreFlow(preset, device=device)
|
242 |
opt = torch.optim.AdamW(model.flow.parameters(), lr=3e-4, weight_decay=1e-5)
|
243 |
train_step = partial(
|
244 |
PatchFlow.stochastic_step,
|
@@ -274,6 +381,10 @@ def train_flow(dataset_path, preset, outdir, epochs=10, device="cuda"):
|
|
274 |
with torch.inference_mode():
|
275 |
val_loss = eval_step(scores, x)
|
276 |
|
|
|
|
|
|
|
|
|
277 |
train_loss = train_step(scores, x)
|
278 |
|
279 |
if (step + 1) % 10 == 0:
|
@@ -297,117 +408,36 @@ def train_flow(dataset_path, preset, outdir, epochs=10, device="cuda"):
|
|
297 |
)
|
298 |
step += 1
|
299 |
|
300 |
-
#
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
|
306 |
-
|
307 |
-
|
308 |
-
|
309 |
-
|
310 |
-
|
311 |
-
|
312 |
-
model = build_model(device=device)
|
313 |
-
scores = model(x)
|
314 |
-
|
315 |
-
return scores
|
316 |
-
|
317 |
-
|
318 |
-
def test_flow_runner(preset, device="cpu", load_weights=None):
|
319 |
-
# f = "doge.jpg"
|
320 |
-
f = "goldfish.JPEG"
|
321 |
-
image = (PIL.Image.open(f)).resize((64, 64), PIL.Image.Resampling.LANCZOS)
|
322 |
-
image = np.array(image)
|
323 |
-
image = image.reshape(*image.shape[:2], -1).transpose(2, 0, 1)
|
324 |
-
x = torch.from_numpy(image).unsqueeze(0).to(device)
|
325 |
-
|
326 |
-
score_flow = ScoreFlow(preset, device=device)
|
327 |
-
|
328 |
-
if load_weights is not None:
|
329 |
-
score_flow.flow.load_state_dict(torch.load(load_weights))
|
330 |
-
|
331 |
-
heatmap = score_flow(x)
|
332 |
-
print(heatmap.shape)
|
333 |
-
|
334 |
-
heatmap = score_flow(x).detach().cpu().numpy()
|
335 |
-
heatmap = (heatmap - heatmap.min()) / (heatmap.max() - heatmap.min()) * 255
|
336 |
-
im = PIL.Image.fromarray(heatmap[0, 0])
|
337 |
-
im.convert("RGB").save(
|
338 |
-
"heatmap.png",
|
339 |
-
)
|
340 |
-
|
341 |
-
return
|
342 |
|
|
|
|
|
343 |
|
344 |
-
@click.command()
|
345 |
|
346 |
-
#
|
347 |
-
|
348 |
-
|
349 |
-
|
350 |
-
|
351 |
-
|
352 |
-
|
353 |
-
)
|
354 |
-
|
355 |
-
|
356 |
-
|
357 |
-
|
358 |
-
|
359 |
-
required=True,
|
360 |
-
)
|
361 |
-
@click.option(
|
362 |
-
"--preset",
|
363 |
-
help="Configuration preset",
|
364 |
-
metavar="STR",
|
365 |
-
type=str,
|
366 |
-
default="edm2-img64-s-fid",
|
367 |
-
show_default=True,
|
368 |
-
)
|
369 |
-
@click.option(
|
370 |
-
"--data", help="Path to the dataset", metavar="ZIP|DIR", type=str, default=None
|
371 |
-
)
|
372 |
-
def cmdline(run, outdir, **opts):
|
373 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
374 |
-
preset = opts["preset"]
|
375 |
-
dataset_path = opts["data"]
|
376 |
-
|
377 |
-
if run in ["cache-scores", "train-flow"]:
|
378 |
-
assert opts["data"] is not None, "Provide path to dataset"
|
379 |
-
|
380 |
-
if run == "cache-scores":
|
381 |
-
cache_score_norms(
|
382 |
-
preset=preset, dataset_path=dataset_path, outdir=outdir, device=device
|
383 |
-
)
|
384 |
-
|
385 |
-
if run == "train-gmm":
|
386 |
-
train_gmm(
|
387 |
-
score_path=f"{outdir}/{preset}/imagenette_score_norms.pt",
|
388 |
-
outdir=f"{outdir}/{preset}",
|
389 |
-
grid_search=True,
|
390 |
-
)
|
391 |
-
|
392 |
-
if run == "train-flow":
|
393 |
-
train_flow(dataset_path, outdir=outdir, preset=preset, device=device)
|
394 |
-
test_flow_runner(preset, device=device, load_weights=f"{outdir}/{preset}/flow.pt")
|
395 |
-
|
396 |
-
# train_flow(imagenette_path, preset, device)
|
397 |
-
|
398 |
-
# cache_score_norms(
|
399 |
-
# preset=preset,
|
400 |
-
# dataset_path="/GROND_STOR/amahmood/datasets/img64/",
|
401 |
-
# device="cuda",
|
402 |
-
# )
|
403 |
-
# train_gmm(
|
404 |
-
# f"out/msma/{preset}_imagenette_score_norms.pt", outdir=f"out/msma/{preset}"
|
405 |
-
# )
|
406 |
-
# s = test_runner(device=device)
|
407 |
-
# s = s.square().sum(dim=(2, 3, 4)) ** 0.5
|
408 |
-
# s = s.to("cpu").numpy()
|
409 |
-
# nll, pct = compute_gmm_likelihood(s, gmmdir=f"out/msma/{preset}/")
|
410 |
-
# print(f"Anomaly score for image: {nll[0]:.3f} @ {pct*100:.2f} percentile")
|
411 |
|
412 |
|
413 |
if __name__ == "__main__":
|
|
|
3 |
import pickle
|
4 |
from functools import partial
|
5 |
from pickle import dump, load
|
6 |
+
from typing import Literal
|
7 |
|
8 |
import click
|
9 |
import numpy as np
|
|
|
21 |
from dataset import ImageFolderDataset
|
22 |
from flowutils import PatchFlow
|
23 |
|
24 |
+
DEVICE: Literal["cuda", "cpu"] = 'cpu'
|
25 |
model_root = "https://nvlabs-fi-cdn.nvidia.com/edm2/posthoc-reconstructions"
|
26 |
|
27 |
config_presets = {
|
|
|
102 |
self,
|
103 |
preset,
|
104 |
device="cpu",
|
105 |
+
**flow_kwargs
|
106 |
):
|
107 |
super().__init__()
|
108 |
|
|
|
110 |
h = w = scorenet.net.img_resolution
|
111 |
c = scorenet.net.img_channels
|
112 |
num_sigmas = len(scorenet.sigma_steps)
|
113 |
+
self.flow = PatchFlow((num_sigmas, c, h, w), **flow_kwargs)
|
114 |
|
115 |
self.flow = self.flow.to(device)
|
116 |
self.scorenet = scorenet.to(device).requires_grad_(False)
|
|
|
190 |
return nll, percentile
|
191 |
|
192 |
|
193 |
+
@torch.inference_mode
|
194 |
+
def test_runner(device="cpu"):
|
195 |
+
# f = "doge.jpg"
|
196 |
+
f = "goldfish.JPEG"
|
197 |
+
image = (PIL.Image.open(f)).resize((64, 64), PIL.Image.Resampling.LANCZOS)
|
198 |
+
image = np.array(image)
|
199 |
+
image = image.reshape(*image.shape[:2], -1).transpose(2, 0, 1)
|
200 |
+
x = torch.from_numpy(image).unsqueeze(0).to(device)
|
201 |
+
model = build_model(device=device)
|
202 |
+
scores = model(x)
|
203 |
+
|
204 |
+
return scores
|
205 |
+
|
206 |
+
|
207 |
+
def test_flow_runner(preset, device="cpu", load_weights=None):
|
208 |
+
# f = "doge.jpg"
|
209 |
+
f = "goldfish.JPEG"
|
210 |
+
image = (PIL.Image.open(f)).resize((64, 64), PIL.Image.Resampling.LANCZOS)
|
211 |
+
image = np.array(image)
|
212 |
+
image = image.reshape(*image.shape[:2], -1).transpose(2, 0, 1)
|
213 |
+
x = torch.from_numpy(image).unsqueeze(0).to(device)
|
214 |
+
|
215 |
+
score_flow = ScoreFlow(preset, device=device)
|
216 |
+
|
217 |
+
if load_weights is not None:
|
218 |
+
score_flow.flow.load_state_dict(torch.load(load_weights))
|
219 |
+
|
220 |
+
heatmap = score_flow(x)
|
221 |
+
print(heatmap.shape)
|
222 |
+
|
223 |
+
heatmap = score_flow(x).detach().cpu().numpy()
|
224 |
+
heatmap = (heatmap - heatmap.min()) / (heatmap.max() - heatmap.min()) * 255
|
225 |
+
im = PIL.Image.fromarray(heatmap[0, 0])
|
226 |
+
im.convert("RGB").save(
|
227 |
+
"heatmap.png",
|
228 |
+
)
|
229 |
+
|
230 |
+
return
|
231 |
+
|
232 |
+
|
233 |
+
@click.group()
|
234 |
+
def cmdline():
|
235 |
+
global DEVICE
|
236 |
+
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
237 |
+
|
238 |
+
|
239 |
+
@cmdline.command(name="cache-scores")
|
240 |
+
@click.option(
|
241 |
+
"--preset",
|
242 |
+
help="Configuration preset",
|
243 |
+
metavar="STR",
|
244 |
+
type=str,
|
245 |
+
default="edm2-img64-s-fid",
|
246 |
+
show_default=True,
|
247 |
+
)
|
248 |
+
@click.option(
|
249 |
+
"--dataset_path",
|
250 |
+
help="Path to the dataset",
|
251 |
+
metavar="ZIP|DIR",
|
252 |
+
type=str,
|
253 |
+
default=None,
|
254 |
+
)
|
255 |
+
@click.option(
|
256 |
+
"--outdir",
|
257 |
+
help="Where to load/save the results",
|
258 |
+
metavar="DIR",
|
259 |
+
type=str,
|
260 |
+
required=True,
|
261 |
+
)
|
262 |
+
|
263 |
+
def cache_score_norms(preset, dataset_path, outdir):
|
264 |
+
device = DEVICE
|
265 |
dsobj = ImageFolderDataset(path=dataset_path, resolution=64)
|
266 |
refimg, reflabel = dsobj[0]
|
267 |
print(f"Loading dataset from {dataset_path}")
|
|
|
289 |
print(f"Computed score norms for {score_norms.shape[0]} samples")
|
290 |
|
291 |
|
292 |
+
@cmdline.command(name="train-flow")
|
293 |
+
@click.option(
|
294 |
+
"--dataset_path",
|
295 |
+
help="Path to the dataset",
|
296 |
+
metavar="ZIP|DIR",
|
297 |
+
type=str,
|
298 |
+
default=None,
|
299 |
+
)
|
300 |
+
@click.option(
|
301 |
+
"--outdir",
|
302 |
+
help="Where to load/save the results",
|
303 |
+
metavar="DIR",
|
304 |
+
type=str,
|
305 |
+
required=True,
|
306 |
+
)
|
307 |
+
@click.option(
|
308 |
+
"--preset",
|
309 |
+
help="Configuration preset",
|
310 |
+
metavar="STR",
|
311 |
+
type=str,
|
312 |
+
default="edm2-img64-s-fid",
|
313 |
+
show_default=True,
|
314 |
+
)
|
315 |
+
@click.option(
|
316 |
+
"--num_flows",
|
317 |
+
help="Number of normalizing flow functions in the PatchFlow model",
|
318 |
+
metavar="INT",
|
319 |
+
type=int,
|
320 |
+
default=4,
|
321 |
+
show_default=True,
|
322 |
+
)
|
323 |
+
def train_flow(dataset_path, preset, outdir, epochs=10, **flow_kwargs):
|
324 |
+
print("using device:", DEVICE)
|
325 |
+
device = DEVICE
|
326 |
dsobj = ImageFolderDataset(path=dataset_path, resolution=64)
|
327 |
refimg, reflabel = dsobj[0]
|
328 |
print(f"Loaded {len(dsobj)} samples from {dataset_path}")
|
|
|
345 |
val_ds, batch_size=128, num_workers=4, prefetch_factor=2
|
346 |
)
|
347 |
|
348 |
+
model = ScoreFlow(preset, device=device, **flow_kwargs)
|
349 |
opt = torch.optim.AdamW(model.flow.parameters(), lr=3e-4, weight_decay=1e-5)
|
350 |
train_step = partial(
|
351 |
PatchFlow.stochastic_step,
|
|
|
381 |
with torch.inference_mode():
|
382 |
val_loss = eval_step(scores, x)
|
383 |
|
384 |
+
# Log details about model
|
385 |
+
writer.add_graph(model.flow.flows, (torch.zeros(1, scores.shape[1], device=device),
|
386 |
+
torch.zeros(1, model.flow.position_encoding.cached_penc.shape[-1], device=device)))
|
387 |
+
|
388 |
train_loss = train_step(scores, x)
|
389 |
|
390 |
if (step + 1) % 10 == 0:
|
|
|
408 |
)
|
409 |
step += 1
|
410 |
|
411 |
+
# Squeeze the juice
|
412 |
+
best_ckpt = torch.load(f"{experiment_dir}/flow.pt")
|
413 |
+
model.flow.load_state_dict(best_ckpt)
|
414 |
+
for i, (x, _) in enumerate(testiter):
|
415 |
+
x = x.to(device)
|
416 |
+
scores = model.scorenet(x)
|
417 |
+
train_loss = train_step(scores, x)
|
418 |
+
writer.add_scalar("loss/train", train_loss, step)
|
419 |
+
pbar.set_description(
|
420 |
+
f"(Tuning) Step: {step:d} - Train: {train_loss:.3f} - Val: {val_loss:.3f}"
|
421 |
+
)
|
422 |
+
step += 1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
423 |
|
424 |
+
torch.save(model.flow.state_dict(), f"{experiment_dir}/flow.pt")
|
425 |
+
writer.close()
|
426 |
|
|
|
427 |
|
428 |
+
# cache_score_norms(
|
429 |
+
# preset=preset,
|
430 |
+
# dataset_path="/GROND_STOR/amahmood/datasets/img64/",
|
431 |
+
# device="cuda",
|
432 |
+
# )
|
433 |
+
# train_gmm(
|
434 |
+
# f"out/msma/{preset}_imagenette_score_norms.pt", outdir=f"out/msma/{preset}"
|
435 |
+
# )
|
436 |
+
# s = test_runner(device=device)
|
437 |
+
# s = s.square().sum(dim=(2, 3, 4)) ** 0.5
|
438 |
+
# s = s.to("cpu").numpy()
|
439 |
+
# nll, pct = compute_gmm_likelihood(s, gmmdir=f"out/msma/{preset}/")
|
440 |
+
# print(f"Anomaly score for image: {nll[0]:.3f} @ {pct*100:.2f} percentile")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
441 |
|
442 |
|
443 |
if __name__ == "__main__":
|