-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathswiss_dino_evaluation.py
279 lines (233 loc) · 12.2 KB
/
swiss_dino_evaluation.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
################################################################################
# Copyright (c) 2024 Samsung Electronics Co., Ltd.
#
# Author(s):
# Kirill Paramonov ([email protected])
# Jia-Xing Zhong ([email protected])
# Umberto Michieli ([email protected])
# Jijoong Moon ([email protected])
# Mete Ozay ([email protected])
#
# Licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) License, (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at https://creativecommons.org/licenses/by-nc-sa/4.0
# Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and limitations under the License.
# For conditions of distribution and use, see the accompanying LICENSE.md file.
################################################################################
import os
import argparse
import logging
import torch
import tqdm
import time
import shutil
import numpy as np
import pandas as pd
from sklearn.metrics import average_precision_score
from torchvision.transforms import ConvertImageDtype, Normalize, Resize, Compose
from torchvision.transforms.functional import InterpolationMode
from torchvision.io import ImageReadMode, read_image
from swiss.dataset_builders.icubworld import ICWDatasetBuilder, generate_icb_episodes
from swiss.dataset_builders.perseg import PerSegDatasetBuilder, generate_perseg_episodes
from swiss.kmeans_utils import kmeans_map_from_feature_map_full
from swiss.feature_extractor_utils import generate_features_dino
from swiss.engine import SwissEngine
# Schema for the few-shot detection dataset
IMAGE_PATH_COLUMN_NAME = "image_path"
LABEL_COLUMN_NAME = "labels"
IMAGE_RESIZE = 448
RANDOM_SEED = 42
PERCENTILE_FOR_ADAPTIVE_THRESHOLD = 5
COORDINATE_SCALING_FACTOR = 200
KMEANS_CLUSTER_NUM_QUERY = 150
def process_dataset(
dataset, fs_episode, dataset_classes, draw_masks=False,
feature_extractor_type="vit_b", output_masks_dir="output_dir",
annotation_type='bbox', refine_patch_maps=False, verbose=False):
device = torch.device("cuda")
# Initializing DINOv2 model
image_transforms = Compose([
Resize([IMAGE_RESIZE, IMAGE_RESIZE], interpolation=InterpolationMode.BICUBIC, antialias=True),
ConvertImageDtype(torch.float),
Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
)
if feature_extractor_type == "vit_s":
feature_extractor_model = torch.hub.load('facebookresearch/dinov2', 'dinov2_vits14')
elif feature_extractor_type == "vit_b":
feature_extractor_model = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitb14')
elif feature_extractor_type == "vit_l":
feature_extractor_model = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14')
feature_extractor_model.to(device)
feature_extractor_model.eval()
feature_map_generator = (lambda images: generate_features_dino(
images, feature_extractor_model, image_transforms, device=device))
swiss_engine = SwissEngine(
dataset_classes=dataset_classes, annotation_type=annotation_type,
refine_patch_maps = refine_patch_maps,
percentile_for_adaptive_threshold=PERCENTILE_FOR_ADAPTIVE_THRESHOLD)
if draw_masks:
output_dir_base = os.path.join(os.getcwd(), output_masks_dir)
output_dir_supp = os.path.join(output_dir_base, "AA_support_dir")
if os.path.isdir(output_dir_base):
try:
shutil.rmtree(output_dir_base)
except OSError as e:
print("Error: %s - %s." % (e.filename, e.strerror))
os.mkdir(output_dir_base)
os.mkdir(output_dir_supp)
df_columns = ["supp_class_name", "swiss_iou", "swiss_ap", "swiss_acc"]
results_df = pd.DataFrame(columns=df_columns)
results_df = results_df.set_index('supp_class_name')
logging.info(f"Starting evaluation")
support_rows = dataset.select(fs_episode['support'])
supp_labels = support_rows[LABEL_COLUMN_NAME]
supp_img_paths = support_rows[IMAGE_PATH_COLUMN_NAME]
supp_images = [read_image(img_path, mode=ImageReadMode.RGB) for img_path in supp_img_paths]
supp_feature_maps, _ = feature_map_generator(supp_images)
query_rows = dataset.select(fs_episode['query'])
query_labels = query_rows[LABEL_COLUMN_NAME]
query_img_paths = query_rows[IMAGE_PATH_COLUMN_NAME]
query_images = [read_image(img_path, mode=ImageReadMode.RGB) for img_path in query_img_paths]
if annotation_type == 'bbox':
BOUNDING_BOX_COLUMN_NAME = "bounding_box"
supp_bboxes = support_rows[BOUNDING_BOX_COLUMN_NAME]
query_bboxes = query_rows[BOUNDING_BOX_COLUMN_NAME]
elif annotation_type == 'segmap':
SEGMENTATION_PATH_COLUMN_NAME = "segmentation_path"
supp_seg_paths = support_rows[SEGMENTATION_PATH_COLUMN_NAME]
query_seg_paths = query_rows[SEGMENTATION_PATH_COLUMN_NAME]
else:
raise ValueError('Unknown annotation type')
# Query feature generation
generate_start_time = time.time()
query_feature_maps, _ = feature_map_generator(query_images)
if verbose:
logging.info(f"Finished generating query features, avg time: {(time.time()-generate_start_time)/len(query_rows)}")
# Kmeans clustering
if refine_patch_maps:
kmeans_start_time = time.time()
query_kmeans_maps = []
for qfm in tqdm.tqdm(query_feature_maps, desc=f"Query clustering"):
query_kmeans_maps.append(
kmeans_map_from_feature_map_full(
qfm, cluster_num=KMEANS_CLUSTER_NUM_QUERY, coord_scaling_factor=COORDINATE_SCALING_FACTOR))
logging.info(f"Finished kmeans for queries, avg time: {(time.time()-kmeans_start_time)/len(query_rows)}")
logging.info("")
gt_logits = np.zeros((len(support_rows), len(query_rows)))
swiss_iou = np.zeros((len(support_rows), len(query_rows)))
swiss_logits = np.zeros((len(support_rows), len(query_rows)))
s_q_pairs = [(s_ind, q_ind) for s_ind in range(len(support_rows)) for q_ind in range(len(query_rows))]
supp_start_time = time.time()
for s_row_ind, q_row_ind in tqdm.tqdm(s_q_pairs, desc=f"Processing support/query pairs."):
class_name = dataset_classes[supp_labels[s_row_ind]]
populate_support_kwargs = {
'image': supp_images[s_row_ind], 'label': supp_labels[s_row_ind],
'feature_map': supp_feature_maps[s_row_ind]}
if annotation_type=='bbox':
populate_support_kwargs['bounding_box'] = supp_bboxes[s_row_ind]
elif annotation_type == 'segmap':
populate_support_kwargs['segmentation_map_path'] = supp_seg_paths[s_row_ind]
swiss_engine.populate_support_sample(**populate_support_kwargs)
if draw_masks:
swiss_engine.draw_support_mask(draw_path=output_dir_supp, support_image_path=supp_img_paths[s_row_ind])
output_dir_ep_supp = os.path.join(output_dir_base, f'supp_{class_name}')
os.mkdir(output_dir_ep_supp)
predict_query_kwargs = {'feature_map': query_feature_maps[q_row_ind]}
if refine_patch_maps:
predict_query_kwargs['kmeans_map'] = query_kmeans_maps[q_row_ind]
pred_logit, pred_seg_mask = swiss_engine.predict_mask_for_query_sample(**predict_query_kwargs)
is_true_positive_sample = (supp_labels[s_row_ind] == query_labels[q_row_ind])
if draw_masks and is_true_positive_sample:
swiss_engine.draw_predicted_query_mask(
draw_path=output_dir_ep_supp, query_image_path=query_img_paths[q_row_ind],
query_label=query_labels[q_row_ind])
# Updating metrics
gt_logits[s_row_ind, q_row_ind] = is_true_positive_sample
swiss_logits[s_row_ind, q_row_ind] = pred_logit
if annotation_type == 'bbox':
swiss_iou[s_row_ind, q_row_ind] = swiss_engine.evaluate_iou_bbox(
pred_seg_mask, query_bboxes[q_row_ind], query_images[q_row_ind])
elif annotation_type == 'segmap':
swiss_iou[s_row_ind, q_row_ind] = swiss_engine.evaluate_iou_segmap(
pred_seg_mask, query_seg_paths[q_row_ind])
if verbose:
logging.info(f"Finished support/query comparison, avg time: {(time.time()-supp_start_time)/len(s_q_pairs)}")
logging.info("")
# Computing metrics
swiss_pred_labels = np.argmax(swiss_logits, axis=0)
for s_ind in range(len(support_rows)):
s_class_name = dataset_classes[supp_labels[s_ind]]
gt_mask = gt_logits[s_ind].astype(bool)
swiss_avg_iou = np.mean(swiss_iou[s_ind][gt_mask])
swiss_acc = np.mean(swiss_pred_labels[gt_mask]==s_ind)
swiss_ap = average_precision_score(gt_mask, swiss_logits[s_ind])
results_df.loc[s_class_name] = [swiss_avg_iou, swiss_ap, swiss_acc]
if verbose:
logging.info(f"Class: {s_class_name}, SwissIOU: {swiss_avg_iou}, ")
logging.info(f"SwissAP: {swiss_ap}, SwissAcc: {swiss_acc}")
logging.info("")
logging.info(f"Finished evaluation")
logging.info(f"Current average SwissIOU is {np.mean(results_df['swiss_iou'])}")
logging.info(f"Current average SwissAP is {np.mean(results_df['swiss_ap'])}")
logging.info(f"Current average SwissACC is {np.mean(results_df['swiss_acc'])}")
return results_df
def main(args):
with open(args.log_file, 'w') as f:
f.close()
logging.basicConfig(
filename=args.log_file, filemode='a', format='%(message)s',
datefmt='%H:%M:%S', level=logging.INFO)
logging.info(
f"Starting SWISS DINO comparison on {args.dataset_name} dataset. "
f"Feature extractor DINOv2_{args.fe_model_type}")
logging.info(
f"Hyperparameters: PERCENTILE_FOR_ADAPTIVE_THRESHOLD {PERCENTILE_FOR_ADAPTIVE_THRESHOLD}")
if args.refine_patch_maps:
logging.info(
f"Refinement hyperparameters: KMEANS_CLUSTER_NUM_QUERY {KMEANS_CLUSTER_NUM_QUERY}, "
f"COORDINATE_SCALING_FACTOR {COORDINATE_SCALING_FACTOR} ")
logging.info("")
if args.dataset_name == 'icubworld':
dataset_builder = ICWDatasetBuilder(data_dir=args.data_dir)
fs_dataset = dataset_builder.as_dataset(split="validation")
fs_episodes = generate_icb_episodes(
fs_dataset, episode_num=1, n_way=None,
seed=RANDOM_SEED, cluttered_scenes=args.cluttered_scenes)
annotation_type = 'bbox'
elif args.dataset_name == 'perseg':
dataset_builder = PerSegDatasetBuilder(data_dir=args.data_dir)
fs_dataset = dataset_builder.as_dataset(split="validation")
fs_episodes = generate_perseg_episodes(
fs_dataset, episode_num=1, n_way=None, seed=RANDOM_SEED)
annotation_type = 'segmap'
dataset_classes = dataset_builder.classes
# Assume one episode for evaluation
fs_episode = fs_episodes[0]
process_dataset(
fs_dataset, fs_episode, dataset_classes,
draw_masks=args.draw_masks,
feature_extractor_type=args.fe_model_type,
output_masks_dir=args.output_masks_dir,
annotation_type=annotation_type,
refine_patch_maps=args.refine_patch_maps,
verbose=args.verbose)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Evaluate SwissDINO on datasets.')
parser.add_argument('--dataset_name', type=str) # options: perseg, icubworld
parser.add_argument('--data_dir', type=str)
parser.add_argument('--fe_model_type', type=str) # options: vit_s, vit_b, vit_l
parser.add_argument('--draw_masks', action='store_true')
parser.set_defaults(draw_masks=False)
parser.add_argument('--output_masks_dir', type=str, default="output_dir")
parser.add_argument('--log_file', type=str, default="log.txt")
parser.add_argument('--cluttered_scenes', action='store_true') # for icubworld dataset
parser.set_defaults(cluttered_scenes=False)
parser.add_argument('--refine_patch_maps', action='store_true')
parser.set_defaults(refine_patch_maps=False)
parser.add_argument('--verbose', action='store_true')
parser.set_defaults(verbose=False)
args = parser.parse_args()
main(args)