-
Notifications
You must be signed in to change notification settings - Fork 0
/
clustering.py
388 lines (324 loc) · 13 KB
/
clustering.py
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
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
import base64
from io import BytesIO
import argparse
import umap
from PIL import Image
from bokeh.io import show, output_file
from bokeh.layouts import row
from bokeh.models import ColumnDataSource, CategoricalColorMapper, HoverTool
from bokeh.palettes import interp_palette
from bokeh.plotting import figure
from models import get_pipeline
from constants import *
from main import get_dataset
import torch
from sklearn.linear_model import LinearRegression
import pacmap
from torch.utils.data import Subset
from tqdm import tqdm
parser = argparse.ArgumentParser(description="Clusterer")
parser.add_argument(
"--model",
required=True,
type=str,
help="The classifier to run.",
choices=VALID_MODELS
)
parser.add_argument(
"--image-noun",
required=False,
type=str,
help="The image noun to use for clip models.",
default="photo"
)
parser.add_argument(
"--dataset",
required=False,
default="cifar10-test",
help="The name of the dataset that should be used.",
choices=["imagenet-val", "cifar10-test"]
)
parser.add_argument(
"--prefix-mod",
required=False,
type=str,
help="The prefix modifier for all ground_labels.",
default=""
)
parser.add_argument(
"--suffix-mod",
required=False,
type=str,
help="The suffix modifier for all ground_labels.",
default=""
)
parser.add_argument(
"--data-root",
required=False,
default="C:/ml_datasets",
type=str,
help="path containing all datasets (training and validation)"
)
parser.add_argument(
"--results-path",
required=False,
default="results",
type=str,
help="The path to store results."
)
parser.add_argument(
"--figures-path",
required=False,
default=FIGURES_PATH_DEFAULT,
type=str,
help="The path to store figures."
)
parser.add_argument(
"--use-random-confs",
default=False,
action=argparse.BooleanOptionalAction,
help="Whether to use random prediction confidences - mainly to test the validity of the scatter plot on the right."
)
class Flatten:
def __init__(self):
pass
def __call__(self, x):
return x.permute(1,2,0).flatten()
def embeddable_image1(data):
#img_data = data.reshape(3,32,32).permute(1,2,0)
#image = Image.fromarray(img_data.numpy(), mode='RGB').resize((64, 64), Image.Resampling.BICUBIC)
image = data.resize((64, 64), Image.Resampling.BICUBIC)
buffer = BytesIO()
image.save(buffer, format='png')
for_encoding = buffer.getvalue()
return 'data:image/png;base64,' + base64.b64encode(for_encoding).decode()
def get_pred_positions(top_preds, ground_labels, unique_labels):
output = torch.zeros(top_preds.shape[0]).int() # 0 means not in top k.
for label in unique_labels:
indices = torch.where(ground_labels == label)[0]
label_preds = top_preds[indices]
detections = torch.where(label_preds == label)
output[indices[detections[0].int()]] = detections[1].int()
return output
def generate_topk_diff(incorrect_confs, k=10, n_weight=2):
"""
Generates the topk diff measure
:param top_confs: N * M tensor where N is the number of probabilitity vectors to consider, '
and M is the length of each.
:param k: The top-k of confidence scores to consider. Must be less than or equal to M. Default is 10.
If M <= 10 then take k = M.
:return: Tensor of size N, with each entry corresponding to the topk weighted differences.
"""
r = 1/n_weight
k = max(incorrect_confs.shape[1], k)
top10_diff = r ** (torch.Tensor(list(range(k))))
top10_diff = top10_diff.repeat(len(incorrect_confs), 1)
top10_diff *= incorrect_confs
top10_diff = top10_diff[:, 0] - top10_diff[:, 1:].sum(dim=1)
return top10_diff
def avg_knn_dist(points, weighted_confs, k=10):
# points is a N by M tensor
k = min(k + 1, len(points))
dist_mat = torch.norm(points[:, None, :] - points[None, :, :], dim=2)#cosine_similarity(points[:, None, :], points[None, :, :], dim=-1, eps=1e-11)
dist_mat = dist_mat / dist_mat.max() # Shorter distance gives bigger weight.
tops = torch.topk(dist_mat,k=k,dim=1, largest=False)
top_dists = tops.values[:, 1:]
top_idxs = tops.indices[:, 1:]
#top_confs = weighted_confs[top_idxs]
return torch.mean(top_dists, dim=1)
def weighted_avg_knn_dist(points, weighted_confs, k=None):
# points is a N by M tensor
if not k: k = len(points)
else: k = min(k + 1, len(points))
weighted_conf_diffs = 1 - torch.abs(weighted_confs[:,None] - weighted_confs)
dist_mat = torch.norm(points[:, None, :] - points[None, :, :], dim=2)#cosine_similarity(points[:, None, :], points[None, :, :], dim=-1, eps=1e-11)
dist_mat = dist_mat / dist_mat.max() # Shorter distance gives bigger weight.
tops = torch.topk(dist_mat, k=k, dim=1, largest=False)
top_dists = tops.values[:, 1:]
top_idxs = tops.indices[:, 1:]
selected_diffs = []
for row_idx in range(len(weighted_conf_diffs)):
current_weighted_conf_diff = weighted_conf_diffs[row_idx]
current_top_idxs = top_idxs[row_idx]
q = current_weighted_conf_diff[current_top_idxs]
selected_diffs.append(q)
selected_diffs = torch.stack(selected_diffs)
#top_confs = weighted_confs[top_idxs]
return torch.mean(top_dists * selected_diffs, dim=1)
def construct_random_data(N = 2000):
confs = []
for i in range(N):
a = []
accum = 1
rands = torch.rand(10)
for j in range(9):
inp = rands[j].item() * accum
a.append(inp)
accum -= inp
a.append(accum)
confs.append(a)
confs = torch.sort(torch.Tensor(confs), dim=1,descending=True).values
labels = torch.zeros(N).to(torch.int8)
top10preds = torch.ones(N, 10).int().to(torch.int16)
return labels, top10preds, confs
def generate_plots(classes_df, figures_file_path, colours, lr, k=None):
datasource = ColumnDataSource(classes_df)
color_mapping = CategoricalColorMapper(factors=list(map(str, range(len(colours)))),
palette=colours)
scatter_figure = figure(
title='PaCMAP projection of Incorrectly Predicted CIFAR10 Images',
# plot_width=600,
# plot_height=600,
tools=('pan, wheel_zoom, reset')
)
scatter_figure.add_tools(HoverTool(tooltips="""
<div>
<div>
<img src='@image' style='float: left; margin: 5px 5px 5px 5px'/>
</div>
<div>
<span style='font-size: 16px; color: #224499'>Pred_pos:</span>
<span style='font-size: 18px'>@pred_pos</span>
<br>
<span style='font-size: 16px; color: #224499'>Label:</span>
<span style='font-size: 18px'>@label_text</span>
<br>
<span style='font-size: 16px; color: #224499'>Top10_diff:</span>
<span style='font-size: 18px'>@top10_diff_str</span>
<br>
<table>
<tr>
<td><strong>@pred1</strong></td>
<td><strong>@conf1</strong></td>
</tr>
<tr>
<td>@pred2</td>
<td>@conf2</td>
</tr>
<tr>
<td>@pred3</td>
<td>@conf3</td>
</tr>
</table>
</div>
</div>
"""))
scatter_figure.circle(
'x',
'y',
source=datasource,
color=dict(field='conf1_idx', transform=color_mapping),
line_alpha=0.6,
fill_alpha=0.6,
size=4
)
scatter_figure1 = figure(
title=f'Top1 Confidence Difference vs. Average KNN dist for k={k}',
# plot_width=600,
# plot_height=600,
tools=('pan, wheel_zoom, reset')
)
scatter_figure1.circle(
'conf1_raw',
'k_nearest_avgs',
source=datasource,
# color=dict(field='top10_diff', transform=color_mapping),
line_alpha=0.6,
fill_alpha=0.6,
size=4
)
scatter_figure1.line(
classes_df["k_nearest_avgs"], lr.predict(classes_df["k_nearest_avgs"].to_numpy().reshape(-1,1))
)
output_file(str(figures_file_path), mode='inline')
show(row(scatter_figure, scatter_figure1))
def perform_umap(pipeline, images):
reducer = umap.UMAP(metric="cosine")
features = pipeline.get_image_features(images)
embedding = reducer.fit_transform(features)
return embedding
def perform_pacmap(image_features):
reducer = pacmap.PaCMAP()
embedding = reducer.fit_transform(image_features)
return embedding
def main(args):
# Load results file
colours = interp_palette(("#0000FF", "#FFFF00"), 256)
data_root = args.data_root
dataset_full = args.dataset
dataset_name, dataset_split = dataset_full.split("-")
model_name = args.model
model_type, weights_name = model_name_parser(model_name)
results_path = args.results_path
image_noun = args.image_noun
prefix_mod = args.prefix_mod
suffix_mod = args.suffix_mod
figures_path = args.figures_path
use_random_confs = args.use_random_confs
results_file_path = get_output_path(results_path, dataset_full, model_type, weights_name,
image_noun, prefix_mod, suffix_mod)
figures_file_path = get_output_path(figures_path, dataset_full, model_type, weights_name,
image_noun, prefix_mod, suffix_mod, filetype="html")
figures_file_path.parent.mkdir(parents=True, exist_ok=True)
if not use_random_confs:
results = torch.load(str(results_file_path))
labels = results["labels"]
top10preds = results["top10preds"].to(torch.int16)
top10confs = results["top10confs"]
else:
labels, top10preds, top10confs = construct_random_data(4000)
top1preds = top10preds[:, 0]
#top1confs = top10confs[:, 0]
incorrect = torch.where(top1preds != labels)[0]
dataset_conf, _ = get_dataset("cifar10",
"test",
DATA_PATH_DEFAULT,
indices=incorrect)
dataset = dataset_conf()
pipeline = get_pipeline(model_type, weights_name, dataset_name).to(DEVICE)
incorrect_preds = top10preds[incorrect]
incorrect_confs = top10confs[incorrect]
incorrect_images = []
actual_labels = []
for image, label in dataset:
incorrect_images.append(image)
actual_labels.append(label)
actual_labels = torch.tensor(actual_labels)
pred_poses = get_pred_positions(incorrect_preds, actual_labels, list(range(10)))
# torch.set_printoptions(profile="full")
image_features = pipeline.get_image_features(dataset_conf)
if not use_random_confs:
embedding = perform_pacmap(image_features)
else:
embedding = torch.rand(4000,2)
classes_df = pd.DataFrame(embedding, columns=["x", "y"])
# classes_df["dist_from_centre"]
classes_df["pred_pos"] = pred_poses.tolist()
classes_df["pred_pos"] = classes_df["pred_pos"].astype(str)
classes_df["label"] = actual_labels
classes_df["label"] = classes_df["label"].astype(str)
classes_df["label_text"] = [str(CIFAR10_LABELS_TEXT[x.item()]) for x in actual_labels]
classes_df["image"] = list(map(embeddable_image1, incorrect_images))
classes_df["pred1"] = [str(CIFAR10_LABELS_TEXT[x.item()]) for x in incorrect_preds[:, 0]]
classes_df["pred2"] = [str(CIFAR10_LABELS_TEXT[x.item()]) for x in incorrect_preds[:, 1]]
classes_df["pred3"] = [str(CIFAR10_LABELS_TEXT[x.item()]) for x in incorrect_preds[:, 2]]
classes_df["conf1_raw"] = incorrect_confs[:, 0].tolist()
classes_df["conf1"] = classes_df["conf1_raw"].astype(str)
classes_df["conf2"] = [str(x.item()) for x in incorrect_confs[:, 1]]
classes_df["conf3"] = [str(x.item()) for x in incorrect_confs[:, 2]]
top10_diff = generate_topk_diff(incorrect_confs)
classes_df["top10_diff_original"] = top10_diff
classes_df["top10_diff_str"] = classes_df["top10_diff_original"].astype(str)
classes_df["top10_diff"] = classes_df["top10_diff_original"] * len(colours)
classes_df["top10_diff"] = classes_df["top10_diff"].astype(int).astype(str)
classes_df["conf1_idx"] = (classes_df["conf1_raw"] * len(colours)).astype(int).astype(str)
k = 100
classes_df["k_nearest_avgs"] = avg_knn_dist(image_features, top10_diff, k=k)
lr = LinearRegression().fit(classes_df["k_nearest_avgs"].to_numpy().reshape(-1,1), classes_df["conf1_raw"].to_numpy())
#classes_df["lr_xs"] = [min(classes_df["k_nearest_avgs"]), max(classes_df["k_nearest_avgs"])]
#classes_df["lr_ys"] = lr.coef_ * classes_df["lr_xs"] + lr.intercept_
generate_plots(classes_df,figures_file_path,colours, lr, k=k)
if __name__ == "__main__":
#labels, preds, confs = construct_random_data()
args = parser.parse_args()
main(args)