-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataLayer.py
300 lines (265 loc) · 16.9 KB
/
dataLayer.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
"""
dataLayer generates pure future batch_y. e.g:
batch_x from 0-9
batch_y from 10-19
Another dataLayer2 generates one-step shifted future batch_y. e.g:
batch_x from 0-9
batch_y from 1-10
"""
import numpy as np
import pdb
from config import cfg
class DataLayer(object):
def __init__(self, datadb, random=False, is_test=False):
"""Set the datadb to be used by this layer during training."""
self.is_test = is_test
# Also set a random flag
self._random = random
if cfg.use_mixed_dataset:
self._video_ind = np.random.choice([s for s in range(37)])
self._get_train_test_user_id()
else:
self._video_ind = np.random.choice([s for s in range(9) if s!=cfg.test_video_ind])
if is_test:
self._video_ind = cfg.test_video_ind
self.running_length = cfg.running_length
self.predict_len = cfg.predict_len
self.predict_step = cfg.predict_step
self.fps = 30
self._init_2(datadb)
def _init_2(self,datadb):
if cfg.process_in_seconds:
#process in seconds
self.running_length = cfg.running_length*self.fps
self.predict_len = cfg.predict_len*self.fps
self.predict_step = cfg.predict_step*self.fps #during testing
# select a new video, shuffle
if cfg.use_cos_sin:
self.per_video_db = {}
self.per_video_db['cos'] = datadb[self._video_ind]['cos_yaw']
self.per_video_db['sin'] = datadb[self._video_ind]['sin_yaw']
self._num_db = int(self.per_video_db['cos'].shape[1]/self.running_length)
self._num_user = int(self.per_video_db['cos'].shape[0])
elif cfg.use_xyz:
self.data_dim = 3
self.per_video_db = np.stack((datadb[self._video_ind]['x'],datadb[self._video_ind]['y'],datadb[self._video_ind]['z']),axis=-1)
self.per_video_db = np.array(self.per_video_db)
if cfg.use_overlapping_chunks:
self._num_db = int((self.per_video_db.shape[1]-2*self.running_length)/cfg.data_chunk_stride)+1
else:
self._num_db = int(self.per_video_db.shape[1]/self.running_length)
self._num_user = self.per_video_db.shape[0]
elif cfg.use_yaw_pitch_roll:
self.data_dim = 2
# self.per_video_db = datadb[self._video_ind]['raw_yaw']
self.per_video_db = np.stack((datadb[self._video_ind]['raw_yaw'],datadb[self._video_ind]['raw_pitch']),axis=-1)
self.per_video_db = np.array(self.per_video_db)
self._num_db = int(self.per_video_db.shape[1]/self.running_length)
self._num_user = self.per_video_db.shape[0]
elif cfg.use_phi_theta:
self.data_dim = 2
self.per_video_db = np.stack((datadb[self._video_ind]['theta'],datadb[self._video_ind]['phi']),axis=-1)
self.per_video_db = np.array(self.per_video_db)
self._num_db = int(self.per_video_db.shape[1]/self.running_length)
self._num_user = self.per_video_db.shape[0]
# self._insert_end_of_stroke()
self._shuffle_datadb_inds()
def _get_train_test_user_id(self):
if self.is_test:
self.user_ind = np.load('./data/merged_dataset_user_idx/test_video'+str(self._video_ind)+'.npy')
else:
self.user_ind = np.load('./data/merged_dataset_user_idx/train_video'+str(self._video_ind)+'.npy')
def _shuffle_datadb_inds(self):
"""Randomly permute the training datadb."""
# If the random flag is set,
# then the database is shuffled according to system time
# Useful for the validation set
if self._random:
st0 = np.random.get_state()
millis = int(round(time.time() * 1000)) % 4294967295
np.random.seed(millis)
if cfg.shuffle_data:
self._perm = np.random.permutation(np.arange(self._num_db-1))
# Restore the random state
if self._random:
np.random.set_state(st0)
else:
self._perm = np.arange(self._num_db-1)
self._cur = 0
def reshape_data_in_seconds(self,batch_x,batch_y,others_future):
# change the shape from frame_length*3 to (frame_length/30)*(30*3)
batch_x = np.reshape(batch_x,(batch_x.shape[0],batch_x.shape[1],batch_x.shape[2]/self.fps,self.fps*self.data_dim))
batch_y = np.reshape(batch_y,(batch_y.shape[0],batch_y.shape[1]/self.fps,self.fps*self.data_dim))
others_future = np.reshape(others_future,(others_future.shape[0],others_future.shape[1],others_future.shape[2]/self.fps,self.fps*self.data_dim))
if cfg.own_history_only:
batch_x = batch_x[:,0,:,:]
return batch_x,batch_y,others_future
def _get_next_minibatch_inds(self,datadb,batch_size):
"""Return the index (starting from where) for the next minibatch."""
if self._cur + batch_size >= self._num_db:
# after going over one video, select another one
if cfg.use_more_video and (not self.is_test):
self._video_ind = np.random.choice([s for s in range(9) if s!=cfg.test_video_ind])
if cfg.use_mixed_dataset:
self._get_train_test_user_id()
# print('change to video',self._video_ind)
self._init_2(datadb)
db_index = self._perm[self._cur:self._cur + batch_size]
self._cur += batch_size
return db_index
def get_batch_CNN_format(self,db_index,target_uer_ID):
"""get a minibatch in n_usr*(2*t*fps)*3 format"""
this_x = self.per_video_db[:,db_index*cfg.data_chunk_stride:db_index*cfg.data_chunk_stride+2*self.running_length,:].copy()
# put target row in the bottom
target_row = this_x[target_uer_ID,:]
this_x = np.delete(this_x,target_uer_ID, axis=0)
this_x = np.vstack((this_x,target_row[np.newaxis,:]))
this_y = target_row[self.running_length:,:]
this_x[-1,self.running_length:]=0
return this_x[np.newaxis,:],this_y[np.newaxis,:]
def get_batch_convLSTM_format(self,db_index,target_uer_ID):
"""get a minibatch in timestamp*47*fps*3"""
this_x = self.per_video_db[:,db_index*cfg.data_chunk_stride:db_index*cfg.data_chunk_stride+self.running_length,:].copy()
this_x_future = self.per_video_db[:,db_index*cfg.data_chunk_stride+self.running_length:db_index*cfg.data_chunk_stride+self.running_length+self.predict_step,:].copy()
# delte target row
this_x = np.delete(this_x,target_uer_ID, axis=0)
target_row = this_x_future[target_uer_ID,:]
this_x_future = np.delete(this_x_future,target_uer_ID, axis=0)
# reshape into timestamp*47*fps*3"""
this_x_per_second_past = this_x.reshape(self._num_user-1,cfg.running_length,self.fps,self.data_dim)
this_x_per_second_past = this_x_per_second_past.transpose((1,0,2,3))
this_x_per_second_future = this_x_future.reshape(self._num_user-1,cfg.running_length,self.fps,self.data_dim)
this_x_per_second_future = this_x_per_second_future.transpose((1,0,2,3))
this_y = target_row
return this_x_per_second_past[np.newaxis,:],this_x_per_second_future[np.newaxis,:],this_y[np.newaxis,:]
def _get_target_user_id(self):
if cfg.fix_target_user:
_target_user = cfg.target_uer_ID
else:
if cfg.use_mixed_dataset:
_target_user = np.random.choice(self.user_ind)
else:
_target_user = np.random.randint(0,self._num_user)
return _target_user
def _get_next_minibatch(self,datadb,batch_size,format=None):
_target_user = np.zeros(batch_size).astype('int')
db_index = self._get_next_minibatch_inds(datadb,batch_size)
"""fetch a batch from tsinghua dataset"""
if format == 'CNN':
batch_x = np.zeros((1,self._num_user,self.running_length*2,self.data_dim))
batch_y = np.zeros((1,self.running_length,self.data_dim))
for ii in range(batch_size):
_target_user[ii] = self._get_target_user_id()
this_x,this_y = self.get_batch_CNN_format(db_index[ii],_target_user[ii])
batch_x = np.vstack((batch_x,this_x))
batch_y = np.vstack((batch_y,this_y))
batch_x = batch_x[1:,:,:,:]
batch_y = batch_y[1:,:,:]
return batch_x,batch_y
elif format == 'convLSTM':
batch_x_past = np.zeros((1,cfg.running_length,self._num_user-1,self.fps,self.data_dim))
batch_x_future = np.zeros((1,cfg.running_length,self._num_user-1,self.fps,self.data_dim))
batch_y = np.zeros((1,self.predict_step,self.data_dim))
for ii in range(batch_size):
_target_user[ii] = self._get_target_user_id()
this_x_per_second_past,this_x_per_second_future,this_y = self.get_batch_convLSTM_format(db_index[ii],_target_user[ii])
batch_x_past = np.vstack((batch_x_past,this_x_per_second_past))
batch_x_future = np.vstack((batch_x_future,this_x_per_second_future))
batch_y = np.vstack((batch_y,this_y))
batch_x_past = batch_x_past[1:,:,:,:,:]
batch_x_future = batch_x_future[1:,:,:,:,:]
batch_y = batch_y[1:,:,:]
return batch_x_past,batch_x_future,batch_y
else:
if cfg.use_cos_sin:
_target_user = np.random.randint(0,self._num_user)
temp_cos = self.per_video_db['cos'][:,db_index*cfg.data_chunk_stride:db_index*cfg.data_chunk_stride+self.running_length].copy()
temp_sin = self.per_video_db['sin'][:,db_index*cfg.data_chunk_stride:db_index*cfg.data_chunk_stride+self.running_length].copy()
batch_x = np.vstack((temp_cos[np.newaxis,:],temp_sin[np.newaxis,:]))
batch_x = batch_x.transpose((0,2,1)) #2*5000*48
batch_x[:,self.running_length-self.predict_len:self.running_length,_target_user] = 0
temp_cos = self.per_video_db['cos'][_target_user,db_index*cfg.data_chunk_stride:db_index*cfg.data_chunk_stride+self.running_length]
temp_sin = self.per_video_db['sin'][_target_user,db_index*cfg.data_chunk_stride:db_index*cfg.data_chunk_stride+self.running_length]
batch_y = np.vstack((temp_cos[np.newaxis,:],temp_sin[np.newaxis,:]))
else:
# fetch batch_x
if cfg.own_history_only:
batch_x = np.zeros((1,1,self.running_length,self.data_dim))
else:
if cfg.include_own_history:
batch_x = np.zeros((1,self._num_user,self.running_length,self.data_dim))
else:
batch_x = np.zeros((1,self._num_user-1,self.running_length,self.data_dim))
for ii in range(batch_size):
_target_user[ii] = self._get_target_user_id()
this_x = self.per_video_db[:,db_index[ii]*cfg.data_chunk_stride:db_index[ii]*cfg.data_chunk_stride+self.running_length,:].copy()
if cfg.stuff_zero:
this_x[_target_user[ii],self.running_length-self.predict_len:self.running_length,:] = 0
elif cfg.stuff_last:
this_x[_target_user[ii],self.running_length-self.predict_len:self.running_length,:] = this_x[_target_user[ii],self.running_length-self.predict_len,:]
target_row = this_x[_target_user[ii],:]
if cfg.own_history_only:
batch_x = np.vstack((batch_x,target_row[np.newaxis,np.newaxis,:,:]))
else:
if cfg.include_own_history:
# put target row in the bottom
this_x = np.delete(this_x,_target_user[ii], axis=0)
this_x = np.vstack((this_x,target_row[np.newaxis,:]))
else:
this_x = np.delete(this_x,_target_user[ii], axis=0)
batch_x = np.vstack((batch_x,this_x[np.newaxis,:]))
batch_x = batch_x[1:,:,:,:]
# fetch batch_y
if cfg.has_reconstruct_loss:
batch_y = np.zeros((1,cfg.self.running_length,self.data_dim))
others_future = np.zeros((1,self._num_user-1,cfg.self.running_length,self.data_dim))
else:
batch_y = np.zeros((1,self.predict_len,self.data_dim))
batch_y_further = np.zeros((1,self.predict_step,self.data_dim))#predict_step = multiple*predict_len
others_future = np.zeros((1,self._num_user-1,self.predict_len,self.data_dim))
others_future_further = np.zeros((1,self._num_user-1,self.predict_step,self.data_dim))
# rep_last_loc = np.zeros((1,self.predict_len))
for ii in range(batch_size):
if cfg.has_reconstruct_loss:
batch_y = np.vstack((batch_y,self.per_video_db[_target_user[ii],
db_index[ii]*cfg.data_chunk_stride:db_index[ii]*cfg.data_chunk_stride+self.running_length]
.reshape((-1,self.running_length,self.data_dim))))
else:
batch_y = np.vstack((batch_y,self.per_video_db[_target_user[ii],
db_index[ii]*cfg.data_chunk_stride+self.running_length:db_index[ii]*cfg.data_chunk_stride+self.running_length+self.predict_len]
.reshape((-1,self.predict_len,self.data_dim))))
y_further = self.per_video_db[_target_user[ii],
db_index[ii]*cfg.data_chunk_stride+self.running_length:db_index[ii]*cfg.data_chunk_stride+self.running_length+self.predict_step]
try:
batch_y_further = np.vstack((batch_y_further,y_further.reshape((-1,self.predict_step,self.data_dim))))
except: #don't have enough future data lasting for cfg.predict_step
if len(y_further)<self.predict_step:
# stack (1.123), bc the range should be [-1,1]
y_further = np.vstack((y_further,np.ones((self.predict_step-len(y_further),self.data_dim))*(1.123)))
batch_y_further = np.vstack((batch_y_further,y_further.reshape((-1,self.predict_step,self.data_dim))))
others_future_db = np.delete(self.per_video_db,_target_user[ii],axis=0)
if cfg.has_reconstruct_loss:
others_future_temp = others_future_db[:,db_index[ii]*cfg.data_chunk_stride:db_index[ii]*cfg.data_chunk_stride+self.running_length,:][np.newaxis,:]
else:
others_future_temp = others_future_db[:,db_index[ii]*cfg.data_chunk_stride+self.running_length:db_index[ii]*cfg.data_chunk_stride+self.running_length+self.predict_len][np.newaxis,:]
others_future_further_temp = others_future_db[:,db_index[ii]*cfg.data_chunk_stride+self.running_length:db_index[ii]*cfg.data_chunk_stride+self.running_length+self.predict_step][np.newaxis,:]
others_future = np.vstack((others_future,others_future_temp))
try:
others_future_further = np.vstack((others_future_further,others_future_further_temp))
except:
if others_future_further_temp.shape[2]<self.predict_step:
# stack (1.123), bc the range should be [-1,1]
others_future_further_temp = np.concatenate((others_future_further_temp,np.ones((1,self._num_user-1,self.predict_step-others_future_further_temp.shape[2],self.data_dim))*1.123),axis=2)
others_future_further = np.vstack((others_future_further,others_future_further_temp))
# last_val = np.array([self.per_video_db[_target_user[ii],db_index[ii]*cfg.data_chunk_stride+self.running_length-self.predict_len-1]]*self.predict_len).reshape((-1,self.predict_len))
# rep_last_loc = np.vstack((rep_last_loc,last_val))
batch_y = batch_y[1:,:,:]
batch_y_further = batch_y_further[1:,:,:]
others_future = others_future[1:,:,:]
others_future_further = others_future_further[1:,:,:]
# rep_last_loc = rep_last_loc[1:,:]
if cfg.process_in_seconds:
batch_x,batch_y,others_future = self.reshape_data_in_seconds(batch_x,batch_y,others_future)
else:
batch_x = batch_x[:,0,:,:]
return batch_x,batch_y,others_future,batch_y_further,db_index,others_future_further