-
Notifications
You must be signed in to change notification settings - Fork 3
/
gtad_inference_fs.py
345 lines (303 loc) · 18.1 KB
/
gtad_inference_fs.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
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
import os
import math
import numpy as np
import pandas as pd
import torch.nn.parallel
from tqdm import tqdm
from gtad_lib import opts
from gtad_lib.models import GTAD
from gtad_lib.dataset_fs import VideoDataSet, VideoEpisodicDataSet
import torch.nn.functional as F
from gtad_lib.classifier_v2 import Classifier ## changed
from visdom_logger import VisdomLogger
from gtad_lib.util import AverageMeter, batch_intersectionAndUnionGPU, get_model_dir, main_process
from gtad_lib.util import find_free_port, setup, cleanup, to_one_hot, intersectionAndUnionGPU
from collections import defaultdict
if __name__ == '__main__':
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
opt = opts.parse_opt()
opt = vars(opt)
if not os.path.exists(opt['output'] + "/results1"):
os.makedirs(opt['output'] + "/results1")
data_path = opt["output"]+"/results1/"
for fls in os.listdir(data_path):
os.remove(os.path.join(data_path,fls))
model = GTAD(opt)
model = torch.nn.DataParallel(model, device_ids=[0]).cuda()
checkpoint = torch.load(opt["output"] + "/GTAD_best.pth.tar")
model.load_state_dict(checkpoint['state_dict'])
# print("bttle_dim", model.module.bottleneck_dim)
model.eval()
valid_mode = "Episodic" ## "Standard" or "Episodic"
if valid_mode == "Standard" :
test_loader = torch.utils.data.DataLoader(VideoDataSet(opt, subset="validation", mode='inference'),
batch_size=1, shuffle=False,
num_workers=8, pin_memory=True, drop_last=False)
else :
test_loader = torch.utils.data.DataLoader(VideoEpisodicDataSet(opt, subset="validation", mode='inference'),
batch_size=1, shuffle=False,
num_workers=8, pin_memory=True, drop_last=False)
tscale = opt["temporal_scale"]
print("Inference start")
if valid_mode == "Standard":
with torch.no_grad():
for idx, input_data in test_loader:
video_name = test_loader.dataset.video_list[idx[0]]
# offset = min(test_loader.dataset.data['indices'][idx[0]])
# video_name = video_name+'_{}'.format(math.floor(offset/250))
input_data = input_data.cuda()
# forward pass
confidence_map, start, end = model(input_data)
start_scores = start[0].detach().cpu().numpy()
end_scores = end[0].detach().cpu().numpy()
clr_confidence = (confidence_map[0][1]).detach().cpu().numpy()
reg_confidence = (confidence_map[0][0]).detach().cpu().numpy()
max_start = max(start_scores)
max_end = max(end_scores)
# use BMN post-processing to boost performance
start_bins = np.zeros(len(start_scores))
start_bins[0] = 1 # [1,0,0...,0,1]
for idx in range(1, tscale - 1):
if start_scores[idx] > start_scores[idx + 1] and start_scores[idx] > start_scores[idx - 1]:
start_bins[idx] = 1
elif start_scores[idx] > (0.5 * max_start):
start_bins[idx] = 1
end_bins = np.zeros(len(end_scores))
end_bins[-1] = 1
for idx in range(1, tscale - 1):
if end_scores[idx] > end_scores[idx + 1] and end_scores[idx] > end_scores[idx - 1]:
end_bins[idx] = 1
elif end_scores[idx] > (0.5 * max_end):
end_bins[idx] = 1
# enumerate sub-graphs as proposals
new_props = []
for idx in range(opt["max_duration"]):
for jdx in range(opt["temporal_scale"]):
start_index = jdx
end_index = start_index + idx+1
if end_index < opt["temporal_scale"] and start_bins[start_index] == 1 and end_bins[end_index] == 1:
xmin = start_index / opt['temporal_scale']
xmax = end_index / opt['temporal_scale']
clr_score = clr_confidence[idx, jdx]
reg_score = reg_confidence[idx, jdx]
new_props.append([xmin, xmax, clr_score, reg_score])
new_props = np.stack(new_props)
col_name = ["xmin", "xmax", "clr_score", "reg_socre"]
new_df = pd.DataFrame(new_props, columns=col_name)
new_df.to_csv(opt["output"]+"/results/" + video_name + ".csv", index=False)
else:
# with torch.no_grad():
## NO_GRAD SHOULD BE REMPOVED
## to do : we take features from sub_graph module of dimension [batch, c, 100, 100] c--> 9216
nb_episodes = opt["episode"]
# nb_episodes = 50
testing_mode = False
# model_feat = model.module.goi_align
# int. feat_dim : 100 x 100 x 9216
## 100 X 1
# model.module.extract_features(input_data)
# c = model.module.bottleneck_dim
# h = model.module.feature_res[0]
# w = model.module.feature_res[1]
h = model.module.features_dim[0]
w = model.module.features_dim[1]
c = 256 # 9216
batch_size_val = 2 # consumes gpu-ram
shot = opt["shot"]
norm_feat = True
n_runs=2
H= 100
W = 1
runtimes = torch.zeros(n_runs)
deltas_init = torch.zeros((n_runs, nb_episodes, batch_size_val))
deltas_final = torch.zeros((n_runs, nb_episodes, batch_size_val))
val_IoUs = np.zeros(n_runs)
val_losses = np.zeros(n_runs)
use_callback = False
# ========== Perform the runs ==========
for run in tqdm(range(n_runs)): ## epochs
loss_meter = AverageMeter()
iter_num = 0
cls_intersection = defaultdict(int) # Default value is 0
cls_union = defaultdict(int)
IoU = defaultdict(int)
# =============== episode = group of tasks ===============
# batch_size_val
# shot
for e in tqdm(range(nb_episodes)):
features_s = torch.zeros(batch_size_val, shot, c, h, w).to(device)
features_s_grad = torch.zeros(batch_size_val, shot, c, h, w).to(device)
features_q = torch.zeros(batch_size_val, 1, c, h, w).to(device)
features_q_grad = torch.zeros(batch_size_val, 1, c, h, w).to(device)
gt_s = torch.zeros(batch_size_val, shot, 100,
1).long().to(device)
gt_q = torch.zeros(batch_size_val, 1, 100,
1).long().to(device)
n_shots = torch.zeros(batch_size_val).to(device)
classes = [] # All classes considered in the tasks
iter_loader = iter(test_loader)
# =========== Generate tasks and extract features for each task ===============
# i=0
## batch_size_ means no of task per episode
for i in range(batch_size_val):
try:
idx, input_data, qry_img, q_label, spprt_imgs, s_label, subcls, video_name = iter_loader.next()
except:
iter_loader = iter(test_loader)
idx, input_data, qry_img, q_label, spprt_imgs, s_label, subcls, video_name = iter_loader.next()
# iter_loader = iter(test_loader)
# idx, input_data, qry_img, q_label, spprt_imgs, s_label, subcls, video_name = iter_loader.next()
# idx, input_data, qry_img, q_label, spprt_imgs, s_label, subcls = test_loader.next()
input_data = input_data.cuda()
qry_img = qry_img.cuda()
# print("vid_name", video_name)
q_label = q_label.cuda()
spprt_imgs = spprt_imgs.cuda()
s_label = s_label.cuda()
iter_num += 1
# print(spprt_imgs.size())
# print(len(subcls))
# print("qry_img", qry_img.size())
# print("support img", spprt_imgs.size())
# split = spprt_imgs.size(2)/shot
# f_s = model.module.extract_features(spprt_imgs[:,i*400:(i+1)*400,:])
f_s = model.module.extract_features(spprt_imgs.squeeze(0)) ## spprt_imgs --? [1,shot,400,100]
f_q = model.module.extract_features(qry_img)
# print("suppport", f_s.size())
# print("query", f_q.size())
shot = f_s.size(0)
n_shots[i] = shot
features_s[i,:shot] = f_s.detach()
features_s_grad[i,:shot] = f_s
# print("suppport", features_s.size())
features_q[i] = f_q.detach()
features_q_grad[i] = f_q
# gt_s[i,:shot] = s_label[:,i*100:(i+1)*100,:]
gt_s[i,:shot] = s_label
gt_q[i,0] = q_label
classes.append([class_.item() for class_ in subcls])
# print("class", classes)#
# =========== Normalize features along channel dimension ===============
## feat_dim : [task=batch, shot=1, c, h, w]
if norm_feat:
features_s = F.normalize(features_s, dim=1)
features_q = F.normalize(features_q, dim=1)
# =========== Create a callback is args.visdom_port != -1 ===============
callback = VisdomLogger(port=2) if use_callback else None
# =========== Initialize the classifier + prototypes + F/B parameter Π ===============
classifier = Classifier()
# n_task, shot, c, h, w = features_s.size()
# print("before", features_s.size())
# features_s = F.adaptive_avg_pool2d(features_s.view(n_task*shot,c, h, w),[10 , 10])
# features_q = F.adaptive_avg_pool2d(features_q.view(n_task,c, h, w),[10 , 10])
# b, c, h, w = features_s.size()
# features_s = features_s.view(n_task,shot,c, h, w)
# features_q = features_q.view(n_task,1,c, h, w)
# print("after", features_s.size())
classifier.init_prototypes(features_s, features_s_grad, features_q, features_q_grad, gt_s, gt_q, classes, callback)
batch_deltas = classifier.compute_FB_param(features_q=features_q, gt_q=gt_q)
# print("delta", batch_deltas)
deltas_init[run, e, :] = batch_deltas.cpu()
# =========== Perform RePRI inference ===============
batch_deltas = classifier.RePRI(features_s, features_q, gt_s, gt_q, classes, n_shots, callback) ## trained W,b and Transformer weights
print("Linear classifier in epoch "+ str(n_runs)+" trained")
deltas_final[run, e, :] = batch_deltas
# t1 = time.time()
# runtime += t1 - t0
### after trained W , pass it to train transformer using CE loss and
frozen_trans = classifier.TransformerRePRI(features_q,gt_q) ## added
logits_u = classifier.get_logits_frozen(features_q,frozen_trans) ## added
### after trained prototypes doing test time logits pass
# logits = classifier.get_logits(features_q) # [n_tasks, shot, h, w] # used transformer here to get updated feature values
logits = F.interpolate(logits_u,
size=(H, W),
mode='bilinear',
align_corners=True)
probas = classifier.get_probas(logits).detach()
# print(probas.size()) # [task=2, shot=1, class=2, h=100, w=1]
intersection, union, _ = batch_intersectionAndUnionGPU(probas, gt_q, 2, classes, video_name) # [n_tasks, shot, num_class]
intersection, union = intersection.cpu(), union.cpu() ### intersection_size : [50,1,2] --> has values of one hot
# print(intersection.size())
# ================== Log metrics ==================
one_hot_gt = to_one_hot(gt_q, 2) ## gt_q --> had value among 1 and 255 , one_hot_gt --> contains value among 1 and 0
# print(one_hot_gt)
valid_pixels = gt_q != 0
loss = classifier.get_ce(probas, valid_pixels, one_hot_gt, reduction='mean')
loss_meter.update(loss.item())
for i, task_classes in enumerate(classes):
# print("task_classes" , task_classes)
for j, class_ in enumerate(task_classes):
cls_intersection[class_] += intersection[i, 0, j + 1] # Do not count background
cls_union[class_] += union[i, 0, j + 1]
for class_ in cls_union:
IoU[class_] = cls_intersection[class_] / (cls_union[class_] + 1e-10)
# print("num",iter_num)
if (iter_num % 100 == 0):
mIoU = np.mean([IoU[i] for i in IoU])
print('Test: [{}/{}] '
'mIoU {:.4f} '
'Loss {loss_meter.val:.4f} ({loss_meter.avg:.4f}) '.format(iter_num,
1000,
mIoU,
loss_meter=loss_meter,
))
mIoU = np.mean(list(IoU.values()))
print('mIoU---Val result: mIoU {:.4f}.'.format(mIoU))
for class_ in cls_union:
print("Class {} : {:.4f}".format(class_, IoU[class_]))
val_IoUs[run] = mIoU
val_losses[run] = loss_meter.avg
if testing_mode:
classifier = Classifier()
for j in range(0,500):
idx, input_data, qry_img, q_label, spprt_imgs, s_label, subcls, video_name = iter_loader.next()
qry_img = qry_img.cuda()
# print("vid_name", video_name)
q_label = q_label.cuda()
f_q = model.module.extract_features(qry_img)
classifier.infer(f_q,q_label)
# for idx, input_data in test_loader:
# video_name = test_loader.dataset.video_list[idx[0]]
# # offset = min(test_loader.dataset.data['indices'][idx[0]])
# # video_name = video_name+'_{}'.format(math.floor(offset/250))
# input_data = input_data.cuda()
# # forward pass
# confidence_map, start, end = model(input_data)
# start_scores = start[0].detach().cpu().numpy()
# end_scores = end[0].detach().cpu().numpy()
# clr_confidence = (confidence_map[0][1]).detach().cpu().numpy()
# reg_confidence = (confidence_map[0][0]).detach().cpu().numpy()
# max_start = max(start_scores)
# max_end = max(end_scores)
# # use BMN post-processing to boost performance
# start_bins = np.zeros(len(start_scores))
# start_bins[0] = 1 # [1,0,0...,0,1]
# for idx in range(1, tscale - 1):
# if start_scores[idx] > start_scores[idx + 1] and start_scores[idx] > start_scores[idx - 1]:
# start_bins[idx] = 1
# elif start_scores[idx] > (0.5 * max_start):
# start_bins[idx] = 1
# end_bins = np.zeros(len(end_scores))
# end_bins[-1] = 1
# for idx in range(1, tscale - 1):
# if end_scores[idx] > end_scores[idx + 1] and end_scores[idx] > end_scores[idx - 1]:
# end_bins[idx] = 1
# elif end_scores[idx] > (0.5 * max_end):
# end_bins[idx] = 1
# # enumerate sub-graphs as proposals
# new_props = []
# for idx in range(opt["max_duration"]):
# for jdx in range(opt["temporal_scale"]):
# start_index = jdx
# end_index = start_index + idx+1
# if end_index < opt["temporal_scale"] and start_bins[start_index] == 1 and end_bins[end_index] == 1:
# xmin = start_index / opt['temporal_scale']
# xmax = end_index / opt['temporal_scale']
# clr_score = clr_confidence[idx, jdx]
# reg_score = reg_confidence[idx, jdx]
# new_props.append([xmin, xmax, clr_score, reg_score])
# new_props = np.stack(new_props)
# col_name = ["xmin", "xmax", "clr_score", "reg_socre"]
# new_df = pd.DataFrame(new_props, columns=col_name)
# new_df.to_csv(opt["output"]+"/results/" + video_name + ".csv", index=False)
print("Inference finished")