-
Notifications
You must be signed in to change notification settings - Fork 2
/
eval_diffcut_openvoc.py
197 lines (157 loc) · 8.4 KB
/
eval_diffcut_openvoc.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
import os
import argparse
from typing import Literal
import logging
import torch
import torch.nn.functional as F
import pandas as pd
from scipy.ndimage import median_filter
from torchmetrics import JaccardIndex
from dataloader.iterator import DataIterator, get_fine_to_coarse, load_imdb
from diffcut.recursive_normalized_cut import DiffCut
from tools.ldm import LdmExtractor
from tools.pamr import PAMR
from tools.utils import MaskPooling
from tools.clip_classifier import CLIP
from tools.clip_classifier import open_vocabulary, get_classification_logits
from detectron2.data import MetadataCatalog
from data.metadata import datasets
import logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
class Benchmark_Segmentation:
def __init__(self,
model_name: Literal["SSD-1B", "SSD-vega", "SD1.4"] = "SSD-1B",
dataset_name: Literal["COCO-Object", "VOC20", "Context"] = "VOC20",
step: int = 50,
img_size: int = 1024,
refinement: bool = False,
alpha: int = 10,
):
refining = "pamr" if refinement else "no_pamr"
self.root_path = f'./OpenVocabulary_Evaluation/{dataset_name}/{refining}'
self.folder_path = os.path.join(self.root_path)
if not os.path.exists(self.folder_path):
os.makedirs(self.folder_path)
self.img_size = img_size
self.step = step
self.refinement = refinement
self.dataset_name = dataset_name
self.alpha = alpha
self.diffcut = DiffCut()
self.mask_pooling = MaskPooling()
if dataset_name == "COCO-Object":
file_list = load_imdb("./dataloader/coco/val2017/Coco164kFull_Stuff_Coarse_7.txt")
root = "./datasets/coco"
dataset = DataIterator(dataset_name, root, "val", file_list, self.img_size)
fine_to_coarse_map = get_fine_to_coarse("COCO-Object", "./dataloader/coco/coco_object_mapping.pickle")
self.test_metadata = MetadataCatalog.get("openvocab_coco_2017_val_panoptic")
self.test_metadata.thing_classes.insert(0, 'background')
self.N_CLASS = 81 # 80 classes + background
elif dataset_name == "VOC20":
root = "./datasets/pascal_voc_d2"
dataset = DataIterator(dataset_name, root, "validation", None, self.img_size)
fine_to_coarse_map = get_fine_to_coarse("VOC20")
self.test_metadata = MetadataCatalog.get("openvocab_pascal21_sem_seg_val")
self.N_CLASS = 21 # 20 classes + background
elif dataset_name == "Context":
root = "./datasets/pascal_ctx_d2"
dataset = DataIterator(dataset_name, root, "validation", None, self.img_size)
fine_to_coarse_map = get_fine_to_coarse("Context")
self.test_metadata = MetadataCatalog.get("openvocab_pascal_ctx59_sem_seg_val")
self.test_metadata.stuff_classes.insert(0, 'background')
self.N_CLASS = 60 # 59 classes + background
self.dataset = dataset
self.fine_to_coarse_map = fine_to_coarse_map
self.extractor = LdmExtractor(model_name=model_name)
with torch.no_grad():
self.clip_backbone = CLIP(model_name="convnext_large_d_320", pretrained="laion2b_s29b_b131k_ft_soup").to("cuda")
self.clip_backbone.clip_model.transformer.batch_first = False
self.ov = open_vocabulary(self.clip_backbone, self.test_metadata, self.test_metadata)
self.text_classifier, self.num_templates = self.ov.get_text_classifier()
def get_features(self, images):
features = self.extractor(images, step=self.step, img_size=self.img_size)
return features
def pamr(self, labels, image):
masks = torch.cat([1. * (labels == label) for label in torch.unique(labels)], dim=1)
labels = PAMR(num_iter=15, dilations=[1, 2, 4, 8, 12, 24, 32, 64])(image, masks)
labels = 1. * torch.argmax(labels, dim=1)
labels = median_filter(labels.cpu().numpy(), 3).astype(int)
return labels
def associate_label(self, image, mask):
final_mask = torch.zeros_like(mask).to("cuda")
if self.dataset_name == "COCO-Object":
image = F.interpolate(image, size=(2048, 2048), mode='bilinear')
with torch.no_grad():
features = self.clip_backbone(image)
clip_features = features["clip_vis_dense"]
for i in torch.unique(mask):
cls_idx = 1.*(mask == i.item())
mask_embed = self.mask_pooling(clip_features, cls_idx)[0]
pooled_clip_feature = mask_embed.reshape(1, 1, -1)
with torch.no_grad():
pooled_clip_feature = self.clip_backbone.visual_prediction_forward_convnext(pooled_clip_feature)
out_vocab_cls_results = get_classification_logits(pooled_clip_feature, self.text_classifier, self.clip_backbone.clip_model.logit_scale, self.num_templates)
idx = torch.argmax(out_vocab_cls_results[..., :-1].softmax(-1)).item()
final_mask[mask==i] = idx
return final_mask
def evaluate(self,
tau: int = 0.5):
#Dataloader
validation_dataloader = torch.utils.data.DataLoader(self.dataset, batch_size=1)
jaccard = JaccardIndex(task="multiclass", num_classes=self.N_CLASS).to("cuda")
ALL = 0
IoU = 0
for i, batch in enumerate(validation_dataloader):
# Transfer to GPU
batch_size = batch["images"].shape[0]
images = batch["images"].to("cuda")
labels = self.fine_to_coarse_map(batch["labels"])
features = self.extractor(images, step=self.step, img_size=self.img_size)
for j in range(batch_size):
img_feat = features[j][None].to(torch.float32)
label_map = labels[j]
pred = self.diffcut.generate_masks(img_feat, tau, mask_size=(1024, 1024), alpha=self.alpha, img_size=self.img_size)
pred += 1
if self.refinement:
pred = torch.Tensor(pred).to("cuda")
pred = self.pamr(pred, images[j][None])[None]
pred = self.associate_label(images[j][None], torch.Tensor(pred).to("cuda"))
# Interpolate label_map on gpu
label_map = F.interpolate(torch.Tensor(label_map)[None].to("cuda"), size=(1024, 1024), mode='nearest-exact')
label_map += 1
IoU += jaccard(pred, label_map)
if torch.isnan(IoU):
IoU += 0
ALL += 1
# Print accuracy and mean IoU occasionally.
if (i+1) % 10 == 0:
mIoU = IoU / ALL
logging.info("mIoU:{}".format(mIoU))
# Print final mean IoU.
mIoU = IoU / ALL
logging.info("mIoU: %s", mIoU)
# Save results in a csv file.
new_row = [{'mIoU': mIoU}]
df = pd.DataFrame(new_row, columns = ['mIoU'])
file_path = self.folder_path + f'/{tau}'
if not os.path.exists(file_path):
os.makedirs(file_path)
df.to_csv(os.path.join(file_path, f'eval_alpha_{self.alpha}_tau_{tau}.csv'), index=False)
def parse_args():
parser = argparse.ArgumentParser("Segmentation Benchmark Script")
parser.add_argument("--model_name", type=str, default="SSD-1B", help="Model name")
parser.add_argument("--dataset_name", type=str, default="VOC20", choices=["COCO-Object", "VOC20", "Context"], help="dataset")
parser.add_argument("--step", type=int, default=50, help="Denoising timestep")
parser.add_argument("--img_size", type=int, default=1024, help="Size of input images")
parser.add_argument("--refinement", dest='refinement', default=True, action='store_true', help="Mask refinement with PAMR")
parser.add_argument("--tau", type=float, default=0.5, help="Threshold value for Recursive NCut")
parser.add_argument("--alpha", type=int, default=10, help="Affinity matrix exponent value")
args = parser.parse_args()
return args
if __name__ == "__main__":
args = parse_args()
logging.info(args)
sem_seg = Benchmark_Segmentation(model_name=args.model_name, dataset_name=args.dataset_name,
step=args.step, img_size=args.img_size, refinement=args.refinement,
alpha=args.alpha)
sem_seg.evaluate(tau=args.tau)