-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval.py
180 lines (156 loc) · 8.75 KB
/
eval.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
import os
import importlib
import torch
from tqdm import tqdm
from collections.abc import Sequence
import argparse
import yaml
from torch.utils.data import DataLoader
from utils.dataset import get_dataset, get_rel_dataset
from utils.metrics import calculate_metrics
from utils.logger import get_logger
logger = get_logger()
def predict(xb, xq, source_id_order, federated_dataset) -> Sequence[Sequence[str]]:
"""Predict the most related 3D models for each sub-dataset using Cosine Similarity.
Args:
xb: The embeddings of all 3D models, organized in the order of source_id_order.
xq: The embeddings of all sub-datasets, where each row represents an embedding of the textual query.
source_id_order: A list of ALL the source_id involved in the federated_dataset.
federated_dataset: A dict, whose key is the query_id, and value is the sub-dataset, which is a list of source_id.
Returns:
A dict of predictions, whose key is the query_id, and value is the list of source_id, sorted in reverse order by relevance DESC.
"""
result = {}
source_to_id = {source_id: i for i, source_id in enumerate(source_id_order)}
xb = xb.to(xq.device)
sim = xq @ xb.T # (nq, nb)
for i, query_id in tqdm(enumerate(federated_dataset.keys()), desc="searching..."):
related_id = [source_to_id[target_id] for target_id in federated_dataset[query_id]] # find all the index of 3D models in a certain sub-dataset
related_id = torch.tensor(related_id).to(sim.device)
related_ord = sim[i, related_id].argsort(descending=True)
result[query_id] = [source_id_order[related_id[j]] for j in related_ord.tolist()]
return result
def parse_args():
parser = argparse.ArgumentParser(description="Evaluate the performance of different feature extractors.")
parser.add_argument("--option", type=str, default="Uni3D", help="The feature extractor to be evaluated.")
parser.add_argument("--cross_modal", type=str, default="text_image_3D", help="The modalities that are used, seperated by '_'. \
The default value is '', representing features extracted using the 'encode' method in the feature extractor. \
If not '', use the modalities included. E.g., 'text_image' or 'text_image_3D'.")
parser.add_argument("--op", type=str, default="add", choices=["concat", "add"], help="The operation to be used to fuse different base embeddings.")
parser.add_argument("--angles", nargs="+", default=["all"], choices=["all", "diag_below", "diag_above", "right", "left", "back", "front", "above", "below"], help="The angles to be used to extract the embeddings if modalities including 'Image'.")
parser.add_argument("--batch_size", type=int, default=100, help="The batch size to be used to extract the embeddings.")
args = parser.parse_args()
logger.info(args)
return args
def get_embedding(option, modality, source_id_list, encode_fn, cache_dir, angle=None, batch_size=128, img_transform=None):
save_path = f'data/objaverse_{option}_{modality + (("_" + str(angle)) if angle is not None else "")}_embeddings.pt'
if os.path.exists(save_path):
return torch.load(save_path)
else:
embeddings = []
dataset = get_dataset(modality, source_id_list, cache_dir=cache_dir, angle=angle, img_transform=img_transform)
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=4, pin_memory=True, prefetch_factor=batch_size//4)
for batch in tqdm(dataloader, desc=f"Extracting {modality + (('_' + str(angle)) if angle is not None else '')} embeddings..."):
if modality == "image":
data = batch['image']
elif modality == "3D":
data = batch['pc']
elif modality == "text":
data = batch
embeddings.append(encode_fn(data))
embedding = torch.cat(embeddings, dim=0).cpu()
torch.save(embedding, save_path)
return embedding
def get_yaml():
file_yaml = 'config/config.yaml'
with open(file=file_yaml, mode='r', encoding='utf-8') as f:
yaml_data = yaml.safe_load(f)
logger.info(f"Loaded config from {file_yaml}")
logger.info(yaml_data)
return yaml_data
def parse_rel_dataset(rel_dataset):
federated_dataset_path, source_id_list_path, gt_dict_path, source_to_caption_path = "data/federated_dataset.pt", "data/source_id_list.pt", "data/gt_dict.pt", "data/source_to_caption.pt"
if os.path.exists(federated_dataset_path) and os.path.exists(source_id_list_path) and os.path.exists(gt_dict_path) and os.path.exists(source_to_caption_path):
return torch.load(federated_dataset_path), torch.load(source_id_list_path), torch.load(gt_dict_path), torch.load(source_to_caption_path)
else:
federated_dataset, source_id_list, gt_dict, source_to_caption = {}, set(), {}, {}
for query_id in range(len(rel_dataset['query_id'])):
federated_dataset[query_id] = rel_dataset['target_ids'][query_id]
gt_dict[query_id] = rel_dataset['GT_ids'][query_id]
source_to_caption[query_id] = rel_dataset['caption'][query_id]
for value in federated_dataset.values():
source_id_list.update(value)
source_id_list = list(source_id_list)
source_id_list.sort()
torch.save(federated_dataset, federated_dataset_path)
torch.save(source_id_list, source_id_list_path)
torch.save(gt_dict, gt_dict_path)
torch.save(source_to_caption, source_to_caption_path)
return federated_dataset, source_id_list, gt_dict, source_to_caption
def evaluate(args, config):
option = args.option.lower()
angles = args.angles
if "all" in angles:
angles = ["diag_below", "diag_above", "right", "left", "back", "front", "above", "below"]
op = args.op
batch_size = args.batch_size
cross_modal = args.cross_modal
cache_dir = config['general']['cache_dir']
rel_dataset = get_rel_dataset(cache_dir)
federated_dataset, source_id_list, gt_dict, source_to_caption = parse_rel_dataset(rel_dataset)
logger.info(f"Initializing {option}...")
module = importlib.import_module(f"feature_extractors.{option}_embedding_encoder")
encoder = getattr(module, f"{option.capitalize()}EmbeddingEncoder")(cache_dir, method_config=config.pop(option, None))
# encode base embeddings
embeddings = []
if "text" in cross_modal:
logger.info(f"Extracting text embeddings using {option}...")
embeddings.append(get_embedding(option, "text", source_id_list, encoder.encode_text, cache_dir=cache_dir, batch_size=1024))
if "image" in cross_modal:
img_transform = encoder.get_img_transform()
for angle in angles:
logger.info(f"Extracting image embeddings using {option} at angle {angle}...")
embeddings.append(get_embedding(option, "image", source_id_list, encoder.encode_image, cache_dir=cache_dir, angle=angle, batch_size=batch_size, img_transform=img_transform))
if "3D" in cross_modal:
logger.info(f"Extracting 3D embeddings using {option}...")
embeddings.append(get_embedding(option, "3D", source_id_list, encoder.encode_3D, cache_dir=cache_dir, batch_size=batch_size))
## fuse base embeddings
if len(embeddings) > 1:
if op == "concat":
embeddings = torch.cat(embeddings, dim=-1)
elif op == "add":
embeddings = sum(embeddings)
else:
raise ValueError(f"Unsupported operation: {op}")
embeddings /= embeddings.norm(dim=-1, keepdim=True)
else:
embeddings = embeddings[0]
# encode query embeddings
logger.info(f"Encoding query embeddings using {option}...")
xq = encoder.encode_query([source_to_caption[k] for k in federated_dataset.keys()])
if op == "concat":
xq = xq.repeat(1, embeddings.shape[-1] // xq.shape[-1]) # repeat to be aligned with the xb
xq /= xq.norm(dim=-1, keepdim=True)
pred_dict = predict(embeddings, xq, source_id_list, federated_dataset)
# # test val split
# import random
# random.seed(2)
# query_id_list = random.sample(list(gt_dict.keys()), 200)
# new_pred_dict = {}
# for query_id in query_id_list:
# new_pred_dict[query_id] = pred_dict[query_id]
# pred_dict = new_pred_dict
logger.info(f"option: {option}, all difficulties:")
log_message = "\tmAP: {},\tmnDCG: {},\tmFT: {},\tmST: {}".format(
calculate_metrics(pred_dict, gt_dict, metric="mAP"),
calculate_metrics(pred_dict, gt_dict, metric="nDCG"),
calculate_metrics(pred_dict, gt_dict, metric="FT"),
calculate_metrics(pred_dict, gt_dict, metric="ST")
)
logger.info(log_message)
def main():
args = parse_args()
config = get_yaml()
evaluate(args, config)
if __name__ == '__main__':
main()