-
Notifications
You must be signed in to change notification settings - Fork 7
/
time_analysis.py
185 lines (145 loc) · 6.21 KB
/
time_analysis.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
import os
import argparse
import matplotlib
import pandas as pd
from sklearn.metrics import precision_recall_curve, average_precision_score
matplotlib.use('Agg')
import torch
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from matplotlib.backends.backend_pdf import PdfPages
from torch.autograd import Variable
from tqdm import tqdm
from dataset import MotionDataset
from utils import load_run
sns.set_style('whitegrid')
sns.set_context('notebook', font_scale=0.8)
def main_old(args):
dataset = MotionDataset('data/split1.pkl', fps=10)
actions = dataset.actions.keys()
run_info, model, loader = load_run(args.run_dir, args.data, keep_actions=actions)
params = run_info[-1]
out = os.path.join(args.run_dir, 'time-analysis.pdf')
with PdfPages(out) as pdf:
for i in tqdm(range(len(loader.dataset))):
x, annotations = loader.dataset[i]
if params['cuda']:
x = x.cuda()
x = Variable(x, volatile=True)
outs = model.steps_forward(x)
head = params.get('head', 'softmax')
if head == 'softmax':
outs = torch.nn.functional.softmax(outs, dim=1)
elif head == 'sigmoid':
outs = torch.nn.functional.sigmoid(outs)
confidences = outs.data.cpu().numpy() # [:, y]
n_samples, n_classes = outs.shape
time = np.arange(n_samples)
# time = torch.linspace(0, 1, seq_len).numpy()
groundtruth = np.zeros_like(confidences, dtype=int)
for a in annotations:
class_id = loader.dataset.action_id_to_ix[a['action_id']]
start = int(round(a['start_frame'] / loader.dataset.skip))
end = int(round((a['start_frame'] + a['duration']) / loader.dataset.skip))
groundtruth[start:end, class_id] = 1
cmap = plt.get_cmap('jet')
colors = cmap(np.linspace(0, 1.0, n_classes))
fig, axes = plt.subplots(3, 1)
for ax in axes:
ax.set_ylim([0, 1])
ax.set_color_cycle(colors)
(ax1, ax2, ax3) = axes
ax1.plot(time, confidences)
ax2.plot(time, groundtruth)
ax3.plot(time, confidences * groundtruth)
pdf.savefig()
# plt.savefig('time-analysis.pdf')
plt.close()
def main(args):
run_info, model, loader = load_run(args.run_dir, data=args.data)
params = run_info[-1]
out = os.path.join(args.run_dir, 'time-analysis.pdf')
labels = np.array([a.replace('hdm05_', '') for a in loader.dataset.action_descriptions])
best_f1s = []
targets = []
predictions = []
with PdfPages('/tmp/app.pdf') as pdf:
for i, (x, y) in enumerate(tqdm(loader)):
y = y.numpy().squeeze()
targets.append(y)
if params['cuda']:
x = x.cuda()
x = Variable(x, volatile=True)
logits = model.segment(x)
y_hat = torch.nn.functional.sigmoid(logits)
y_hat = y_hat.data.cpu().numpy().squeeze() # [:, y]
predictions.append(y_hat)
n_samples, n_classes = y_hat.shape
time = np.arange(n_samples)
# time = torch.linspace(0, 1, seq_len).numpy()
# np.savez('segmentation_outs_n_preds.npz', y=y, y_hat=y_hat)
# break
ap = average_precision_score(y, y_hat, average='micro')
p, r, t = precision_recall_curve(y.ravel(), y_hat.ravel())
t = np.insert(t, 0, 0)
f1 = 2 * (p * r) / (p + r)
best_f1, best_thr = max(zip(f1, t))
best_f1s.append(best_f1)
cmap = plt.get_cmap('jet')
colors = cmap(np.linspace(0, 1.0, n_classes))
fig, axes = plt.subplots(3, 1)
for ax in axes:
ax.set_ylim([0, 1.1])
ax.set_prop_cycle('color', colors)
(ax1, ax2, ax3) = axes
# (ax1, ax2) = axes
ax1.set_title('Prediction [AP={:.1%}, F1={:.1%} (thr={})]'.format(ap, best_f1, best_thr))
ax1.plot(time, y_hat, label=labels)
ax2.set_title('Groundtruth ({})'.format(loader.dataset.data[i]['seq_id']))
lines = ax2.plot(time, y)
ax3.set_title('Masked Prediction')
lines = ax3.plot(time, y_hat * y)
legends_ix = set(y.sum(axis=0).nonzero()[0].tolist() +
(y_hat > 0.2).sum(axis=0).nonzero()[0].tolist())
legends_ix = np.array(list(legends_ix))
lines = np.array(lines)
lines = lines[legends_ix]
legends = labels[legends_ix]
sns.despine()
lgd = ax2.legend(lines, legends, loc='center', ncol=6, bbox_to_anchor=(0.5, -0.42))
plt.tight_layout()
pdf.savefig(bbox_extra_artists=(lgd,), bbox_inches='tight')
# plt.savefig('time-analysis.pdf')
plt.close()
best_f1s = np.array(best_f1s)
order = np.argsort(best_f1s)[::-1] + 1
order = " ".join(map(str, order))
os.system('pdftk /tmp/app.pdf cat {} output {}'.format(order, out))
targets = np.concatenate(targets, axis=0)
predictions = np.concatenate(predictions, axis=0)
p, r, t = precision_recall_curve(targets.ravel(), predictions.ravel())
t = np.insert(t, 0, 0)
f1 = 2 * (p * r) / (p + r)
best_f1, best_thr = max(zip(f1, t))
print('Single Thr F1: {} {}'.format(best_f1, best_thr))
cat_f1s = []
cat_thr = []
for i in range(n_classes):
p, r, t = precision_recall_curve(targets[:, i], predictions[:, i])
f1 = 2 * (p * r) / (p + r)
t = np.insert(t, 0, 0)
b_f1, b_thr = max(zip(f1, t))
cat_f1s.append(b_f1)
cat_thr.append(b_thr)
data = pd.DataFrame(dict(BestF1=cat_f1s, Threshold=cat_thr), index=labels)
print(data)
support = targets.sum(axis=0)
avgF1 = (data['BestF1'].values * support).sum() / support.sum()
print(avgF1)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Time Analysis')
parser.add_argument('run_dir', help='folder of the model to use')
parser.add_argument('-d', '--data', help='data to analyze')
args = parser.parse_args()
main(args)