Spaces:
Sleeping
Sleeping
+ added cmd line to msma
Browse files
app.py
CHANGED
@@ -6,12 +6,14 @@ import matplotlib.pyplot as plt
|
|
6 |
import numpy as np
|
7 |
import torch
|
8 |
|
9 |
-
from msma import
|
10 |
|
11 |
|
12 |
@cache
|
13 |
-
def load_model(preset="edm2-img64-s-fid", device='cpu'):
|
14 |
-
|
|
|
|
|
15 |
|
16 |
@cache
|
17 |
def load_reference_scores(model_dir):
|
@@ -38,24 +40,42 @@ def plot_against_reference(nll, ref_nll):
|
|
38 |
return fig
|
39 |
|
40 |
|
41 |
-
def
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
nll, pct, ref_nll = compute_gmm_likelihood(x.cpu(), model_dir=f"models/{preset}")
|
48 |
|
49 |
-
|
50 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
51 |
outstr = f"Anomaly score: {nll:.3f} / {pct:.2f} percentile"
|
52 |
-
|
|
|
|
|
|
|
53 |
|
54 |
|
55 |
demo = gr.Interface(
|
56 |
fn=run_inference,
|
57 |
inputs=["image"],
|
58 |
-
outputs=["text",
|
|
|
|
|
|
|
59 |
)
|
60 |
|
61 |
if __name__ == "__main__":
|
|
|
6 |
import numpy as np
|
7 |
import torch
|
8 |
|
9 |
+
from msma import ScoreFlow, config_presets
|
10 |
|
11 |
|
12 |
@cache
|
13 |
+
def load_model(modeldir, preset="edm2-img64-s-fid", device='cpu', outdir=None):
|
14 |
+
model = ScoreFlow(preset, device=device)
|
15 |
+
model.flow.load_state_dict(torch.load(f"{modeldir}/{preset}/flow.pt"))
|
16 |
+
return model
|
17 |
|
18 |
@cache
|
19 |
def load_reference_scores(model_dir):
|
|
|
40 |
return fig
|
41 |
|
42 |
|
43 |
+
def plot_heatmap(heatmap):
|
44 |
+
fig, ax = plt.subplots()
|
45 |
+
im = heatmap[0,0]
|
46 |
+
ax.imshow(im, cmap='gist_heat')
|
47 |
+
fig.tight_layout()
|
48 |
+
return fig
|
|
|
49 |
|
50 |
+
# def compute_scores
|
51 |
|
52 |
+
|
53 |
+
def run_inference(img, preset="edm2-img64-s-fid", device="cuda"):
|
54 |
+
|
55 |
+
with torch.inference_mode():
|
56 |
+
img = torch.from_numpy(img).permute(2,0,1).unsqueeze(0)
|
57 |
+
img = torch.nn.functional.interpolate(img, size=64, mode='bilinear')
|
58 |
+
img = img.to(device)
|
59 |
+
model = load_model(modeldir='models', preset=preset, device=device)
|
60 |
+
x = model.scorenet(img)
|
61 |
+
x = x.square().sum(dim=(2, 3, 4)) ** 0.5
|
62 |
+
img_likelihood = model(img).cpu().numpy()
|
63 |
+
nll, pct, ref_nll = compute_gmm_likelihood(x.cpu(), model_dir=f"models/{preset}")
|
64 |
+
|
65 |
outstr = f"Anomaly score: {nll:.3f} / {pct:.2f} percentile"
|
66 |
+
histplot = plot_against_reference(nll, ref_nll)
|
67 |
+
heatmapplot = plot_heatmap(img_likelihood)
|
68 |
+
|
69 |
+
return outstr, heatmapplot, histplot
|
70 |
|
71 |
|
72 |
demo = gr.Interface(
|
73 |
fn=run_inference,
|
74 |
inputs=["image"],
|
75 |
+
outputs=["text",
|
76 |
+
gr.Plot(label="Anomaly Heatmap"),
|
77 |
+
gr.Plot(label="Comparing to Imagenette"),
|
78 |
+
],
|
79 |
)
|
80 |
|
81 |
if __name__ == "__main__":
|
msma.py
CHANGED
@@ -3,6 +3,7 @@ import pickle
|
|
3 |
from functools import partial
|
4 |
from pickle import dump, load
|
5 |
|
|
|
6 |
import numpy as np
|
7 |
import PIL.Image
|
8 |
import torch
|
@@ -95,12 +96,12 @@ class EDMScorer(torch.nn.Module):
|
|
95 |
class ScoreFlow(torch.nn.Module):
|
96 |
def __init__(
|
97 |
self,
|
98 |
-
|
99 |
-
vectorize=False,
|
100 |
device="cpu",
|
101 |
):
|
102 |
super().__init__()
|
103 |
|
|
|
104 |
h = w = scorenet.net.img_resolution
|
105 |
c = scorenet.net.img_channels
|
106 |
num_sigmas = len(scorenet.sigma_steps)
|
@@ -134,9 +135,9 @@ def train_gmm(score_path, outdir, grid_search=False):
|
|
134 |
gm = GaussianMixture(
|
135 |
n_components=7, init_params="kmeans", covariance_type="full", max_iter=100000
|
136 |
)
|
|
|
137 |
|
138 |
if grid_search:
|
139 |
-
clf = Pipeline([("scaler", StandardScaler()), ("GMM", gm)])
|
140 |
param_grid = dict(
|
141 |
GMM__n_components=range(2, 11, 1),
|
142 |
)
|
@@ -184,10 +185,11 @@ def compute_gmm_likelihood(x_score, gmmdir):
|
|
184 |
return nll, percentile
|
185 |
|
186 |
|
187 |
-
def cache_score_norms(preset, dataset_path, device="cpu"):
|
188 |
dsobj = ImageFolderDataset(path=dataset_path, resolution=64)
|
189 |
refimg, reflabel = dsobj[0]
|
190 |
-
print(
|
|
|
191 |
dsloader = torch.utils.data.DataLoader(
|
192 |
dsobj, batch_size=48, num_workers=4, prefetch_factor=2
|
193 |
)
|
@@ -202,8 +204,8 @@ def cache_score_norms(preset, dataset_path, device="cpu"):
|
|
202 |
|
203 |
score_norms = torch.cat(score_norms, dim=0)
|
204 |
|
205 |
-
os.makedirs("
|
206 |
-
with open(f"
|
207 |
torch.save(score_norms, f)
|
208 |
|
209 |
print(f"Computed score norms for {score_norms.shape[0]} samples")
|
@@ -232,7 +234,7 @@ def train_flow(dataset_path, preset, device="cuda"):
|
|
232 |
val_ds, batch_size=48, num_workers=4, prefetch_factor=2
|
233 |
)
|
234 |
|
235 |
-
model = ScoreFlow(
|
236 |
opt = torch.optim.AdamW(model.flow.parameters(), lr=3e-4, weight_decay=1e-5)
|
237 |
train_step = partial(
|
238 |
PatchFlow.stochastic_step,
|
@@ -296,16 +298,15 @@ def test_runner(device="cpu"):
|
|
296 |
return scores
|
297 |
|
298 |
|
299 |
-
def test_flow_runner(device="cpu", load_weights=None):
|
300 |
-
f = "doge.jpg"
|
301 |
-
|
302 |
image = (PIL.Image.open(f)).resize((64, 64), PIL.Image.Resampling.LANCZOS)
|
303 |
image = np.array(image)
|
304 |
image = image.reshape(*image.shape[:2], -1).transpose(2, 0, 1)
|
305 |
x = torch.from_numpy(image).unsqueeze(0).to(device)
|
306 |
-
model = build_model(device=device)
|
307 |
|
308 |
-
score_flow = ScoreFlow(
|
309 |
|
310 |
if load_weights is not None:
|
311 |
score_flow.flow.load_state_dict(torch.load(load_weights))
|
@@ -323,13 +324,35 @@ def test_flow_runner(device="cpu", load_weights=None):
|
|
323 |
return
|
324 |
|
325 |
|
326 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
327 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
328 |
-
preset =
|
329 |
-
|
|
|
|
|
|
|
330 |
|
331 |
-
|
332 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
333 |
|
334 |
# cache_score_norms(
|
335 |
# preset=preset,
|
@@ -344,3 +367,6 @@ if __name__ == "__main__":
|
|
344 |
# s = s.to("cpu").numpy()
|
345 |
# nll, pct = compute_gmm_likelihood(s, gmmdir=f"out/msma/{preset}/")
|
346 |
# print(f"Anomaly score for image: {nll[0]:.3f} @ {pct*100:.2f} percentile")
|
|
|
|
|
|
|
|
3 |
from functools import partial
|
4 |
from pickle import dump, load
|
5 |
|
6 |
+
import click
|
7 |
import numpy as np
|
8 |
import PIL.Image
|
9 |
import torch
|
|
|
96 |
class ScoreFlow(torch.nn.Module):
|
97 |
def __init__(
|
98 |
self,
|
99 |
+
preset,
|
|
|
100 |
device="cpu",
|
101 |
):
|
102 |
super().__init__()
|
103 |
|
104 |
+
scorenet = build_model(preset)
|
105 |
h = w = scorenet.net.img_resolution
|
106 |
c = scorenet.net.img_channels
|
107 |
num_sigmas = len(scorenet.sigma_steps)
|
|
|
135 |
gm = GaussianMixture(
|
136 |
n_components=7, init_params="kmeans", covariance_type="full", max_iter=100000
|
137 |
)
|
138 |
+
clf = Pipeline([("scaler", StandardScaler()), ("GMM", gm)])
|
139 |
|
140 |
if grid_search:
|
|
|
141 |
param_grid = dict(
|
142 |
GMM__n_components=range(2, 11, 1),
|
143 |
)
|
|
|
185 |
return nll, percentile
|
186 |
|
187 |
|
188 |
+
def cache_score_norms(preset, dataset_path, outdir, device="cpu"):
|
189 |
dsobj = ImageFolderDataset(path=dataset_path, resolution=64)
|
190 |
refimg, reflabel = dsobj[0]
|
191 |
+
print(f"Loading dataset from {dataset_path}")
|
192 |
+
print(f"Number of Samples: {len(dsobj)} - shape: {refimg.shape}, dtype: {refimg.dtype}, labels {reflabel}")
|
193 |
dsloader = torch.utils.data.DataLoader(
|
194 |
dsobj, batch_size=48, num_workers=4, prefetch_factor=2
|
195 |
)
|
|
|
204 |
|
205 |
score_norms = torch.cat(score_norms, dim=0)
|
206 |
|
207 |
+
os.makedirs(f"{outdir}/{preset}/", exist_ok=True)
|
208 |
+
with open(f"{outdir}/{preset}/imagenette_score_norms.pt", "wb") as f:
|
209 |
torch.save(score_norms, f)
|
210 |
|
211 |
print(f"Computed score norms for {score_norms.shape[0]} samples")
|
|
|
234 |
val_ds, batch_size=48, num_workers=4, prefetch_factor=2
|
235 |
)
|
236 |
|
237 |
+
model = ScoreFlow(preset, device=device)
|
238 |
opt = torch.optim.AdamW(model.flow.parameters(), lr=3e-4, weight_decay=1e-5)
|
239 |
train_step = partial(
|
240 |
PatchFlow.stochastic_step,
|
|
|
298 |
return scores
|
299 |
|
300 |
|
301 |
+
def test_flow_runner(preset, device="cpu", load_weights=None):
|
302 |
+
# f = "doge.jpg"
|
303 |
+
f = "goldfish.JPEG"
|
304 |
image = (PIL.Image.open(f)).resize((64, 64), PIL.Image.Resampling.LANCZOS)
|
305 |
image = np.array(image)
|
306 |
image = image.reshape(*image.shape[:2], -1).transpose(2, 0, 1)
|
307 |
x = torch.from_numpy(image).unsqueeze(0).to(device)
|
|
|
308 |
|
309 |
+
score_flow = ScoreFlow(preset, device=device)
|
310 |
|
311 |
if load_weights is not None:
|
312 |
score_flow.flow.load_state_dict(torch.load(load_weights))
|
|
|
324 |
return
|
325 |
|
326 |
|
327 |
+
@click.command()
|
328 |
+
|
329 |
+
# Main options.
|
330 |
+
@click.option('--run', help='Which function to run',
|
331 |
+
type=click.Choice(['cache-scores', 'train-flow', 'train-gmm'], case_sensitive=False)
|
332 |
+
)
|
333 |
+
@click.option('--outdir', help='Where to load/save the results', metavar='DIR', type=str, required=True)
|
334 |
+
@click.option('--preset', help='Configuration preset', metavar='STR', type=str, default='edm2-img64-s-fid', show_default=True)
|
335 |
+
@click.option('--data', help='Path to the dataset', metavar='ZIP|DIR', type=str, default=None)
|
336 |
+
def cmdline(run, outdir, **opts):
|
337 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
338 |
+
preset = opts['preset']
|
339 |
+
dataset_path = opts['data']
|
340 |
+
|
341 |
+
if run in ['cache-scores', 'train-flow']:
|
342 |
+
assert opts['data'] is not None, "Provide path to dataset"
|
343 |
|
344 |
+
if run == "cache-scores":
|
345 |
+
cache_score_norms(preset=preset, dataset_path=dataset_path, outdir=outdir, device=device)
|
346 |
+
|
347 |
+
if run == "train-gmm":
|
348 |
+
train_gmm(
|
349 |
+
score_path=f"{outdir}/{preset}/imagenette_score_norms.pt",
|
350 |
+
outdir=f"{outdir}/{preset}",
|
351 |
+
grid_search=True,
|
352 |
+
)
|
353 |
+
|
354 |
+
# test_flow_runner("cuda", f"out/msma/{preset}/flow.pt")
|
355 |
+
# train_flow(imagenette_path, preset, device)
|
356 |
|
357 |
# cache_score_norms(
|
358 |
# preset=preset,
|
|
|
367 |
# s = s.to("cpu").numpy()
|
368 |
# nll, pct = compute_gmm_likelihood(s, gmmdir=f"out/msma/{preset}/")
|
369 |
# print(f"Anomaly score for image: {nll[0]:.3f} @ {pct*100:.2f} percentile")
|
370 |
+
|
371 |
+
if __name__ == "__main__":
|
372 |
+
cmdline()
|