-
Notifications
You must be signed in to change notification settings - Fork 5
/
evaluate.py
152 lines (128 loc) · 6.46 KB
/
evaluate.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
import os
import sys
import torch
import random
import numpy as np
import argparse
from os import path as osp
from tqdm import tqdm
from copy import deepcopy
sys.path.append(osp.dirname(osp.dirname(osp.abspath(__file__))))
from data.data_module import DataModule
from models.model_utils import cosine_sim, linear_sim
from dp.exact_dp import dtw, drop_dtw, otam, NW, lcss
from dp.dp_utils import compute_all_costs
from models.nets import EmbeddingsMapping
from models.model_utils import load_last_checkpoint
from eval.metrics import framewise_accuracy, IoU, recall_crosstask
device = "cuda" if torch.cuda.is_available() else "cpu"
def framewise_eval(dataset, model, keep_p, gamma=1, config=None):
""" evaluate representations using framewise accuracy """
accuracy = {'dp': 0, 'simple': 0}
iou = {'dp': 0, 'simple': 0}
for i, sample in enumerate(tqdm(dataset)):
if sample['num_steps'] < 1:
continue
device = "cuda" if (model is not None and next(model.parameters()).is_cuda) else "cpu"
if model is not None:
frame_features = model.map_video(sample['frame_features'].to(device)).detach().cpu()
step_features = model.map_text(sample['step_features'].to(device)).detach().cpu()
else:
frame_features = sample['frame_features'].cpu()
step_features = sample['step_features'].cpu()
sample['frame_features'] = frame_features
sample['step_features'] = step_features
sim = (step_features @ frame_features.T)
if config.drop_cost == 'learn':
distractor = model.compute_distractors(step_features.mean(0).to(device)).detach().cpu()
else:
distractor = None
zx_costs, drop_costs, _ = compute_all_costs(
sample, distractor, gamma, drop_cost_type=config.drop_cost, keep_percentile=keep_p)
zx_costs, drop_costs = [t.detach().cpu().numpy() for t in [zx_costs, drop_costs]]
sim = sim.detach().cpu().numpy()
# defining matching and drop costs
if config.dp_algo in ['DropDTW', 'NW', 'LCSS']:
dp_fn_dict = {'DropDTW': drop_dtw, 'NW': NW, 'LCSS': lcss}
dp_fn = dp_fn_dict[config.dp_algo]
optimal_assignment = dp_fn(zx_costs, drop_costs, return_labels=True) - 1
elif config.dp_algo == 'OTAM':
_, path = otam(-sim)
optimal_assignment = np.zeros(sim.shape[1]) - 1
optimal_assignment[path[1]] = path[0]
else:
_, path = dtw(-sim)
_, uix = np.unique(path[1], return_index=True)
optimal_assignment = path[0][uix]
simple_assignment = np.argmax(sim, axis=0)
simple_assignment[drop_costs < zx_costs.min(0)] = -1
# get framewise accuracy for each vid
accuracy['simple'] += framewise_accuracy(
simple_assignment, sample, use_unlabeled=config.use_unlabeled)
accuracy['dp'] += framewise_accuracy(
optimal_assignment, sample, use_unlabeled=config.use_unlabeled)
iou['simple'] += IoU(simple_assignment, sample)
iou['dp'] += IoU(optimal_assignment, sample)
num_samples = len(dataset)
return [v / num_samples for v in
[accuracy['simple'], accuracy['dp'], iou['simple'], iou['dp']]]
def compute_all_metrics(dataset, model, gamma, config):
sim_matricies = []
distractors = []
for sample in dataset:
if sample['num_steps'] < 1:
continue
device = "cuda" if (model is not None and next(model.parameters()).is_cuda) else "cpu"
frame_features = sample['frame_features'].to(device)
step_features = sample['step_features'].to(device)
distractor = None
if model is not None:
frame_features = model.map_video(frame_features).detach().cpu()
step_features = model.map_text(step_features).detach().cpu()
if config.drop_cost == 'learn':
mean_step = sample['step_features'].mean(0).to(device)
distractor = model.compute_distractors(mean_step).detach().cpu()
# get pairwise similarity
if config.distance == 'inner':
text_clip_similarity = linear_sim(step_features, frame_features)
elif config.distance == 'cos':
text_clip_similarity = cosine_sim(step_features, frame_features)
sim_matricies.append(text_clip_similarity)
distractors.append(distractor)
keep_p = config.keep_percentile
keep_p = keep_p / 3 if config.dataset == 'CrossTask' else keep_p
keep_p = keep_p / 2 if config.dataset == 'YouCook2' else keep_p
accuracy_std, accuracy_dtw, iou_std, iou_dtw = framewise_eval(dataset, model, keep_p, gamma, config)
recall = recall_crosstask(dataset, model)
return accuracy_std * 100, iou_std * 100, accuracy_dtw * 100, iou_dtw * 100, recall
if __name__ == '__main__':
# read arguments
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', type=str, default='COIN', help='dataset')
parser.add_argument('--name', type=str, default='', help='model for evaluation, if nothing is given, evaluate pretrained features')
parser.add_argument('--distance', type=str, default='inner', help='distance type')
parser.add_argument('--dp_algo', type=str, default='DropDTW', choices=['DropDTW', 'OTAM', 'DTW', 'NW', 'LCSS'], help='distance type')
parser.add_argument('--drop_cost', type=str, default='logit', help='Whather do drop in drop-dtw')
parser.add_argument('--keep_percentile', type=float, default=0.3, help='If drop_cost is logits, the percentile to set the drop to')
parser.add_argument('--use_unlabeled', type=bool, default=True,
help='use unlabeled frames in comparison (useful to consider dropped steps)')
args = parser.parse_args()
print(args)
# fix random seed
torch.manual_seed(1)
random.seed(1)
dataset = DataModule(args.dataset, 1, 1).val_dataset
if args.name:
gamma = 30
model = EmbeddingsMapping(d=512, learnable_drop=(args.drop_cost == 'learn'),
video_layers=2, text_layers=0)
load_last_checkpoint(args.name, model, device, remove_name_preffix='model.')
model.to('cuda')
model.eval()
else:
model, gamma = None, 1
accuracy_std, iou_std, accuracy_dtw, iou_dtw, recall = compute_all_metrics(
dataset, model, gamma, args)
print(f"Recall is {recall:.1f}%")
print(f"DTW Accuracy is {accuracy_dtw:.1f}%")
print(f"DTW IoU is {iou_dtw:.1f}%")