-
Notifications
You must be signed in to change notification settings - Fork 24
/
Copy pathtrain_h36m.py
229 lines (205 loc) · 10.6 KB
/
train_h36m.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
import numpy as np
import os
import tensorflow as tf
from helper import train_helper as th
from helper import config
from model_runner.klstm.kfl_QRf import Model as kfl_QRf
from model_runner.klstm.kfl_QRFf import Model as kfl_QRFf
from model_runner.klstm.kfl_K import Model as kfl_K
from model_runner.lstm.tf_lstm import Model as lstm
from helper import dt_utils as dut
from helper import utils as ut
# gpu_config = tf.ConfigProto(
# device_count={'GPU':0}
# )
gpu_config = tf.ConfigProto()
# gpu_config.de
gpu_config.gpu_options.per_process_gpu_memory_fraction = 0.45
def test_data(sess,params,X,Y,index_list,S_list,R_L_list,F_list,e, pre_test,n_batches):
is_test=1
dic_state=ut.get_state_list(params)
I= np.asarray([np.diag([1.0]*params['n_output']) for i in range(params["batch_size"])],dtype=np.float32)
params["reset_state"]=-1 #Never reset
state_reset_counter_lst=[0 for i in range(batch_size)]
total_loss=0.0
total_pred_loss=0.0
total_meas_loss=0.0
total_n_count=0.0
for minibatch_index in xrange(n_batches):
state_reset_counter_lst=[s+1 for s in state_reset_counter_lst]
# print state_reset_counter_lst
(dic_state,x,y,r,f,_,state_reset_counter_lst,_)= \
th.prepare_batch(is_test,index_list, minibatch_index, batch_size,
S_list, dic_state, params, Y, X, R_L_list,F_list,state_reset_counter_lst)
feed=th.get_feed(tracker,params,r,x,y,I,dic_state, is_training=0)
states,final_output,final_pred_output,final_meas_output,y =sess.run([tracker.states,tracker.final_output,tracker.final_pred_output,tracker.final_meas_output,tracker.y], feed)
for k in states.keys():
dic_state[k] = states[k]
if params["normalise_data"]==3 or params["normalise_data"]==2:
final_output=ut.unNormalizeData(final_output,params["y_men"],params["y_std"])
final_pred_output=ut.unNormalizeData(final_pred_output,params["y_men"],params["y_std"])
final_meas_output=ut.unNormalizeData(final_meas_output,params["x_men"],params["x_std"])
y=ut.unNormalizeData(y,params["y_men"],params["y_std"])
if params["normalise_data"]==4:
final_output=ut.unNormalizeData(final_output,params["x_men"],params["x_std"])
final_pred_output=ut.unNormalizeData(final_pred_output,params["x_men"],params["x_std"])
final_meas_output=ut.unNormalizeData(final_meas_output,params["x_men"],params["x_std"])
y=ut.unNormalizeData(y,params["x_men"],params["x_std"])
test_loss,n_count=ut.get_loss(params,gt=y,est=final_output,r=r)
test_pred_loss,n_count=ut.get_loss(params,gt=y,est=final_pred_output,r=r)
test_meas_loss,n_count=ut.get_loss(params,gt=y,est=final_meas_output,r=r)
total_loss+=test_loss*n_count
total_pred_loss+=test_pred_loss*n_count
total_meas_loss+=test_meas_loss*n_count
total_n_count+=n_count
# if (minibatch_index%show_every==0):
# print pre_test+" test batch loss: (%i / %i / %i) %f"%(e,minibatch_index,n_train_batches,test_loss)
total_loss=total_loss/total_n_count
total_pred_loss=total_pred_loss/total_n_count
total_meas_loss=total_meas_loss/total_n_count
s =pre_test+' Loss --> epoch %i | error %f, %f, %f'%(e,total_loss,total_pred_loss,total_meas_loss)
ut.log_write(s,params)
return total_loss
def train(tracker,params):
I= np.asarray([np.diag([1.0]*params['n_output']) for i in range(params["batch_size"])],dtype=np.float32)
batch_size=params["batch_size"]
num_epochs=1000
decay_rate=0.9
show_every=100
deca_start=3
pre_best_loss=10000
with tf.Session(config=gpu_config) as sess:
tf.global_variables_initializer().run()
saver = tf.train.Saver()
# if params["model"] == "kfl_QRf":
# ckpt = tf.train.get_checkpoint_state(params["mfile"])
# if ckpt and ckpt.model_checkpoint_path:
# saver.restore(sess, ckpt.model_checkpoint_path)
# mfile = ckpt.model_checkpoint_path
# params["est_file"] = params["est_file"] + mfile.split('/')[-1].replace('.ckpt', '') + '/'
# print "Loaded Model: %s" % ckpt.model_checkpoint_path
# if params["model"] == "kfl_QRf":
# for var in tracker.tvars:
# path = '/mnt/Data1/hc/tt/cp/weights/' + var.name.replace('transitionF/','')
# if os.path.exists(path+'.npy'):
# val=np.load(path+'.npy')
# sess.run(tf.assign(var, val))
# print 'PreTrained LSTM model loaded...'
# sess.run(tracker.predict())
print 'Training model:'+params["model"]
noise_std = params['noise_std']
new_noise_std=0.0
for e in range(num_epochs):
if e>(deca_start-1):
sess.run(tf.assign(tracker.lr, params['lr'] * (decay_rate ** (e))))
else:
sess.run(tf.assign(tracker.lr, params['lr']))
total_train_loss=0
state_reset_counter_lst=[0 for i in range(batch_size)]
index_train_list_s=index_train_list
dic_state = ut.get_state_list(params)
# total_loss = test_data(sess, params, X_test, Y_test, index_test_list, S_Test_list, R_L_Test_list,
# F_list_test, e, 'Test Check', n_test_batches)
if params["shufle_data"]==1 and params['reset_state']==1:
index_train_list_s = ut.shufle_data(index_train_list)
for minibatch_index in xrange(n_train_batches):
is_test = 0
state_reset_counter_lst=[s+1 for s in state_reset_counter_lst]
(dic_state,x,y,r,f,_,state_reset_counter_lst,_)= \
th.prepare_batch(is_test,index_train_list_s, minibatch_index, batch_size,
S_Train_list, dic_state, params, Y_train, X_train, R_L_Train_list,F_list_train,state_reset_counter_lst)
if noise_std >0.0:
u_cnt= e*n_train_batches + minibatch_index
if u_cnt in params['noise_schedule']:
if u_cnt==params['noise_schedule'][0]:
new_noise_std=noise_std
else:
new_noise_std = noise_std * (u_cnt / (params['noise_schedule'][1]))
s = 'NOISE --> u_cnt %i | error %f' % (u_cnt, new_noise_std)
ut.log_write(s, params)
if new_noise_std>0.0:
noise=np.random.normal(0.0,new_noise_std,x.shape)
x=noise+x
feed = th.get_feed(tracker, params, r, x, y, I, dic_state, is_training=1)
train_loss,states,_ = sess.run([tracker.cost,tracker.states,tracker.train_op], feed)
for k in states.keys():
dic_state[k] = states[k]
total_train_loss+=train_loss
if (minibatch_index%show_every==0):
print "Training batch loss: (%i / %i / %i) %f"%(e,minibatch_index,n_train_batches,
train_loss)
total_train_loss=total_train_loss/n_train_batches
s='TRAIN --> epoch %i | error %f'%(e, total_train_loss)
ut.log_write(s,params)
pre_test = "TRAINING_Data"
total_loss = test_data(sess, params, X_train, Y_train, index_train_list, S_Train_list, R_L_Train_list,
F_list_train, e, pre_test, n_train_batches)
pre_test="TEST_Data"
total_loss= test_data(sess,params,X_test,Y_test,index_test_list,S_Test_list,R_L_Test_list,F_list_test,e, pre_test,n_test_batches)
base_cp_path = params["cp_file"] + "/"
lss_str = '%.5f' % total_loss
model_name = lss_str + "_" + str(e) + "_" + str(params["rn_id"]) + params["model"] + "_model.ckpt"
save_path = base_cp_path + model_name
saved_path = False
if pre_best_loss > total_loss:
pre_best_loss = total_loss
model_name = lss_str + "_" + str(e) + "_" + str(params["rn_id"]) + params["model"] + "_best_model.ckpt"
save_path = base_cp_path + model_name
saved_path = saver.save(sess, save_path)
else:
if e % 3.0 == 0:
saved_path = saver.save(sess, save_path)
if saved_path != "":
s = 'MODEL_Saved --> epoch %i | error %f path %s' % (e, total_loss, saved_path)
ut.log_write(s, params)
rnn_keep_prob_lst=[0.8]
rnn_input_prob_lst=[1.0]
seq_lst=[50]
reset_state=[5,100,20]
normalise_data_lst=[3]
params = config.get_params()
params["mfile"]='/mnt/Data1/hc/tt/cp/lstm_nostate1/cp/'
rnn_keep_prob=0.8
input_keep_prob=1.0
params['rnn_keep_prob']=rnn_keep_prob
params['input_keep_prob']=input_keep_prob
seq=50
res=5
with tf.Graph().as_default():
print "seq: ============== %s ============" % seq
print "reset_state: ============== %s ============" % res
print "rnn_keep_prob: ============== %s ============" % rnn_keep_prob
params['normalise_data'] = 4
params['reset_state']=res
params['seq_length']=seq
params["reload_data"] = 0
params = config.update_params(params)
params["model"] = "kfl_QRf"
if params["model"] == "lstm":
tracker = lstm(params=params)
elif params["model"] == "kfl_QRf":
tracker = kfl_QRf(params=params)
elif params["model"] == "kfl_Rf":
tracker = kfl_Rf(params=params)
elif params["model"] == "kfl_QRFf":
tracker = kfl_QRFf(params=params)
elif params["model"] == "kfl_K":
tracker = kfl_K(params=params)
params["rn_id"]="dobuleloss081500_nrm4_seq%i_res%i_keep%f_lr%f"%(seq,res,rnn_keep_prob,params["lr"])
params=config.update_params(params)
(params, X_train, Y_train, F_list_train, G_list_train, S_Train_list, R_L_Train_list,
X_test, Y_test, F_list_test, G_list_test, S_Test_list, R_L_Test_list) = \
dut.prepare_training_set(params)
show_every = 1
(index_train_list, S_Train_list) = dut.get_seq_indexes(params, S_Train_list)
(index_test_list, S_Test_list) = dut.get_seq_indexes(params, S_Test_list)
batch_size = params['batch_size']
n_train_batches = len(index_train_list)
n_train_batches /= batch_size
n_test_batches = len(index_test_list)
n_test_batches /= batch_size
params['training_size'] = len(X_train) * params['seq_length']
params['test_size'] = len(X_test) * params['seq_length']
ut.start_log(params)
ut.log_write("Model training started", params)
train(tracker,params)