-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathtest.py
170 lines (133 loc) · 6.66 KB
/
test.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
import os
import numpy as np
import torch
import torch.cuda
import torch.utils.data
import logging
import copy
import skimage.transform
import scipy.stats
from tqdm import tqdm
from utils.proposal import generate_anchors, proposal_layer
from utils.sphere import viewport_alignment
from flownet2.models import FlowNet2
from dataset.dataset_VQA_ODV import DS_VQA_ODV, VQA_ODV_Transform
from models import VP_net, VQ_net
def main(log_dir, batch_size, num_workers, flownet_ckpt, test_start_frame, test_interval):
arguments = copy.deepcopy(locals())
if not torch.cuda.is_available():
raise RuntimeError('At least 1 GPU is needed by FlowNet2.')
device_main = torch.device('cuda:0')
# For viewport alignment on 8K frame, more than 6 GB GPU memory is needed,
# and thus it needs a different GPU device or fallback to CPU
if torch.cuda.device_count() > 1:
device_alignment = torch.device('cuda:1')
else:
device_alignment = torch.device('cpu')
torch.backends.cudnn.benchmark = True
logger = logging.getLogger("test")
logger.handlers = []
logger.setLevel(logging.DEBUG)
ch = logging.StreamHandler()
ch.setLevel(logging.INFO)
logger.addHandler(ch)
logger.info("%s", repr(arguments))
bandwidth = 128
test_set = DS_VQA_ODV(root=os.path.join(log_dir, "VQA_ODV"), dataset_type='test', tr_te_file='tr_te_VQA_ODV.txt',
ds_list_file='VQA_ODV.txt', test_interval=test_interval, test_start_frame=test_start_frame,
transform=VQA_ODV_Transform(bandwidth=bandwidth, down_resolution=(1024, 2048), to_rgb=True))
anchor_shape = (16, 16)
anchors = torch.tensor(generate_anchors(np.array(anchor_shape)))
# Gaussian center bias
cb = np.load(os.path.join(log_dir, 'cb256.npy')).astype(np.float32)[np.newaxis, np.newaxis, ...]
cb = torch.tensor(cb).to(device_main)
# Mask for anchors
anchor_mask = np.load(os.path.join(log_dir, 'anchor_mask.npy')).astype(np.int64)
anchor_mask = torch.tensor(anchor_mask)
vpnet = VP_net.Model()
vpnet.to(device_main)
vpnet.load_state_dict(torch.load(os.path.join(log_dir, 'vp_state.pkl')))
logger.info("Successfully loaded VP-net pre-trained model.")
vqnet = VQ_net.Model()
vqnet.to(device_main)
vqnet.load_state_dict(torch.load(os.path.join(log_dir, 'vq_state.pkl')))
logger.info("Successfully loaded VQ-net pre-trained model.")
class FlowNetParams:
rgb_max = 255.0
fp16 = False
flownet = FlowNet2(args=FlowNetParams())
flownet.to(device_main)
if isinstance(flownet_ckpt, str):
flownet_ckpt = torch.load(flownet_ckpt)
flownet.load_state_dict(flownet_ckpt['state_dict'])
logger.info("Successfully loaded FlowNet2 pre-trained model.")
flownet.eval()
test_loader = torch.utils.data.DataLoader(test_set, batch_size=batch_size, num_workers=num_workers,
shuffle=False, pin_memory=True, drop_last=False)
pred = []
targets = []
vpnet.eval()
vqnet.eval()
for batch_idx, img_tuple in enumerate(tqdm(test_loader)):
with torch.no_grad():
img_s2, img_original, img_down, img_gap_s2, gap_down, ref_original, target = img_tuple
gap_down = gap_down.to(device_main)
img_down = img_down.to(device_main)
gap_down = gap_down.view((-1, *gap_down.shape[-3:]))
img_down = img_down.view((-1, *img_down.shape[-3:]))
# Optical flow
flow = torch.stack((gap_down, img_down), dim=0).permute(1, 2, 0, 3, 4)
flow = flownet(flow)
flow = flow.cpu().numpy().transpose((2, 3, 1, 0))
flow = skimage.transform.resize(flow, (bandwidth * 2, bandwidth * 2) + flow.shape[-2:], order=1,
anti_aliasing=True, mode='reflect', preserve_range=True).astype(np.float32)
flow_s2 = torch.tensor(flow.transpose((3, 2, 0, 1)))
flow_s2 = flow_s2.to(device_main)
# VP net
img_s2 = img_s2.to(device_main)
img_gap_s2 = img_gap_s2.to(device_main)
img_s2 = img_s2.view((-1, *img_s2.shape[-3:]))
img_gap_s2 = img_gap_s2.view((-1, *img_gap_s2.shape[-3:]))
vp_hm_weight, vp_hm_offset, _ = vpnet(img_s2, flow_s2, cb)
# Viewport softer NMS
hm_after_nms, hm_weight = proposal_layer(vp_hm_weight, vp_hm_offset, 20, 7.5, anchors.to(vp_hm_offset),
mask=anchor_mask)
# Viewport alignment
hm_after_nms = hm_after_nms.to(device_alignment)
img_original = img_original.to(device_alignment)
img_original = img_original.view((-1, *img_original.shape[-3:]))
img_viewport = viewport_alignment(img_original, hm_after_nms[:, 0], hm_after_nms[:, 1])
del img_original
img_viewport = img_viewport.to(device_main)
ref_original = ref_original.to(device_alignment)
ref_original = ref_original.view((-1, *ref_original.shape[-3:]))
ref_viewport = viewport_alignment(ref_original, hm_after_nms[:, 0], hm_after_nms[:, 1])
del ref_original
ref_viewport = ref_viewport.to(device_main)
# VQ net
vq_score, _ = vqnet(img_viewport, ref_viewport - img_viewport)
vq_score = vq_score.flatten()
vq_score = (vq_score * hm_weight).sum(dim=0, keepdim=True)
pred.append(float(vq_score))
target = target.mean(dim=1).reshape((-1,))
targets.append(target.numpy())
pred = np.array(pred)
targets = np.concatenate(targets, 0)
video_cnt = len(test_set.cum_frame_num)
pred = [pred[test_set.cum_frame_num_prev[i]:test_set.cum_frame_num[i]].mean() for i in range(video_cnt)]
targets = [targets[test_set.cum_frame_num_prev[i]:test_set.cum_frame_num[i]].mean() for i in range(video_cnt)]
np.savetxt(os.path.join(log_dir, 'test_pred_scores.txt'), np.array(pred))
np.savetxt(os.path.join(log_dir, 'test_targets.txt'), np.array(targets))
srocc, _ = scipy.stats.spearmanr(pred, targets)
logger.info("SROCC:{:.4}".format(srocc))
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--log_dir", type=str, required=True)
parser.add_argument("--flownet_ckpt", type=str, required=True)
parser.add_argument("--batch_size", type=int, default=1)
parser.add_argument("--num_workers", type=int, default=4)
parser.add_argument("--test_start_frame", type=int, default=21)
parser.add_argument("--test_interval", type=int, default=45)
args = parser.parse_args()
main(**args.__dict__)