forked from xujinfan/Reinforced-Continual-Learning
-
Notifications
You must be signed in to change notification settings - Fork 0
/
RCL.py
203 lines (189 loc) · 11.3 KB
/
RCL.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
# -*- coding: utf-8 -*-
"""
Created on Fri Jan 12 19:27:34 2018
@author: Jason
"""
import warnings
warnings.filterwarnings("ignore")
import tensorflow as tf
import os
os.environ['TF_CPP_MIN_LOG_LEVEL']='3'
import numpy as np
from evaluate import evaluate
from policy_gradient import Controller
import argparse
import datetime
import time
import pickle
class RCL:
def __init__(self,args):
self.args = args
self.num_tasks = args.n_tasks
self.epochs = args.n_epochs
self.batch_size = args.batch_size
self.lr = args.lr
self.data_path = args.data_path
self.max_trials = args.max_trials
self.penalty = args.penalty
self.task_list = self.create_mnist_task()
self.evaluates = evaluate(task_list=self.task_list, args = args)
self.train()
def create_session(self):
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
return sess
def create_mnist_task(self):
data = pickle.load(open(self.data_path, "rb"))
return data
def train(self):
self.best_params={}
self.result_process = []
for task_id in range(0,self.num_tasks):
print("DEBUG : task{}/{} start task loop".format(task_id, self.num_tasks) )
self.best_params[task_id] = [0,0]
print("DEBUG : task{}/{} end reset best_params".format(task_id, self.num_tasks) )
if task_id == 0:
print("DEBUG : task{}/{} IF".format(task_id, self.num_tasks) )
with tf.Graph().as_default() as g:
with tf.name_scope("before"):
inputs = tf.placeholder(shape=(None, 784), dtype=tf.float32)
y = tf.placeholder(shape=(None, 10), dtype=tf.float32)
w1 = tf.Variable(tf.truncated_normal(shape=(784,312), stddev=0.01))
b1 = tf.Variable(tf.constant(0.1, shape=(312,)))
w2 = tf.Variable(tf.truncated_normal(shape=(312,128), stddev=0.01))
b2 = tf.Variable(tf.constant(0.1, shape=(128,)))
w3 = tf.Variable(tf.truncated_normal(shape=(128,10), stddev=0.01))
b3 = tf.Variable(tf.constant(0.1, shape=(10,)))
output1 = tf.nn.relu(tf.nn.xw_plus_b(inputs,w1,b1,name="output1"))
output2 = tf.nn.relu(tf.nn.xw_plus_b(output1,w2,b2,name="output2"))
output3 = tf.nn.xw_plus_b(output2,w3,b3,name="output3")
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=output3)) + \
0.0001*(tf.nn.l2_loss(w1) + tf.nn.l2_loss(w2) + tf.nn.l2_loss(w3))
if self.args.optimizer=="adam":
optimizer = tf.train.AdamOptimizer(learning_rate=self.args.lr)
elif self.args.optimizer=="rmsprop":
optimizer = tf.train.RMSPropOptimizer(learning_rate=self.lr)
elif self.args.optimizer=="sgd":
optimizer = tf.train.GradientDescentOptimizer(learning_rate=self.lr)
else:
raise Exception("please choose one optimizer")
train_step = optimizer.minimize(loss)
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(y,axis=1),tf.argmax(output3,axis=1)),tf.float32))
sess = self.create_session()
sess.run(tf.global_variables_initializer())
l = len(self.task_list[0][1])
for epoch in range(self.epochs):
print("task {}/{} epoch {} run for {} IF ".format(task_id, self.num_tasks, epoch, l ) )
flag = 0
for _ in range(l//self.batch_size+1):
batch_xs, batch_ys = (self.task_list[task_id][0][flag:flag+self.batch_size],self.task_list[task_id][1][flag:flag+self.batch_size])
flag += self.batch_size
sess.run(train_step,feed_dict={inputs:batch_xs, y:batch_ys})
print("task {}/{} epoch {} train_step {}/{}".format(task_id, self.num_tasks, epoch, flag, l) )
accuracy_test = sess.run(accuracy, feed_dict={inputs:self.task_list[task_id][4], y:self.task_list[task_id][5]})
print("task {}/{} test accuracy: {} IF".format(task_id, self.num_tasks, accuracy_test) )
self.vars = sess.run([w1,b1,w2,b2,w3,b3])
self.best_params[task_id] = [accuracy_test,self.vars]
else:
print("DEBUG : task{}/{} start ELSE".format(task_id, self.num_tasks) )
print("DEBUG : task{}/{} reset_default_graph ELSE".format(task_id, self.num_tasks) )
tf.reset_default_graph()
print("DEBUG : task{}/{} reset_default_graph ELSE".format(task_id, self.num_tasks) )
print("DEBUG : task{}/{} start Controller ELSE".format(task_id, self.num_tasks) )
controller = Controller(self.args)
print("DEBUG : taks{}/{} ELSE finish Controller ".format(task_id, self.num_tasks) )
results = []
print("DEBUG : taks{}/{} ELSE result Controller ".format(task_id, self.num_tasks) )
best_reward = 0
print("DEBUG : taks{}/{} ELSE start trial loop ".format(task_id, self.num_tasks) )
for trial in range(self.max_trials):
actions = controller.get_actions()
print("task {}/{} trial {}/{} *********actions for {} ELSE ".format(task_id, self.num_tasks, trial, self.max_trials, actions) )
accuracy_val, accuracy_test = self.evaluates.evaluate_action(var_list = self.vars,
actions=actions, task_id = task_id)
results.append(accuracy_val)
print("task {}/{} trial {}/{}, test accuracy: {} ELSE".format(task_id, self.num_tasks, trial, self.max_trials, accuracy_test) )
reward = accuracy_val - self.penalty*sum(actions)
print(" reward: {} ELSE".format(reward) )
if reward > best_reward:
best_reward = reward
self.best_params[task_id] = (accuracy_test, self.evaluates.var_list)
print("DEBUG : taks{}/{} trial {}/{} ELSE done best_reward ".format(task_id, self.num_tasks, trial, self.max_trials) )
controller.train_controller(reward)
print("DEBUG : taks{}/{} trial {}/{} ELSE end trial internal loop for train_control ".format(task_id, self.num_tasks,trial, self.max_trials) )
print("DEBUG : taks{}/{} ELSE end trial loop ".format(task_id, self.num_tasks) )
#controller.close_session()
print("DEBUG : taks{}/{} ELSE end controller session ".format(task_id, self.num_tasks) )
self.result_process.append(results)
print("DEBUG : taks{}/{} ELSE end result append ".format(task_id, self.num_tasks) )
self.vars = self.best_params[task_id][1]
print("DEBUG : taks{}/{} ELSE end self vars ".format(task_id, self.num_tasks) )
print("DEBUG : taks{}/{} ELSE end loop ".format(task_id, self.num_tasks) )
if __name__ == "__main__":
print("DEBUG : RCL.py step1. start...")
parser = argparse.ArgumentParser(description='Reinforced Continual learning')
# model parameters
parser.add_argument('--n_tasks', type=int, default=10,
help='number of tasks')
parser.add_argument('--n_hiddens', type=str, default='312,218',
help='number of hidden neurons at each layer')
parser.add_argument('--n_layers', type=int, default=2,
help='number of hidden layers')
# optimizer parameters
parser.add_argument('--n_epochs', type=int, default=15,
help='Number of epochs per task')
parser.add_argument('--batch_size', type=int, default=32,
help='batch size')
parser.add_argument('--lr', type=float, default=1e-3,
help='SGD learning rate')
parser.add_argument('--max_trials', type=int, default=50,
help='max_trials')
# experiment parameters
parser.add_argument('--seed', type=int, default=0,
help='random seed')
parser.add_argument('--save_path', type=str, default='/mnt/dataset/lsun/Reinforced-Continual-Learning/results/',
help='save models at the end of training')
# data parameters
parser.add_argument('--data_path', default='/mnt/dataset/lsun/Reinforced-Continual-Learning/mnist_permutations.pkl',
help='path where data is located')
parser.add_argument('--state_space', type=int, default=30, help="the state space for search")
parser.add_argument('--actions_num', type=int, default=2, help="how many actions to dscide")
parser.add_argument('--hidden_size', type=int, default=100, help="the hidden size of RNN")
parser.add_argument('--num_layers', type=int, default=2, help="the layer of a RNN cell")
parser.add_argument('--cuda', type=bool, default=True, help="use GPU or not")
parser.add_argument('--bendmark', type=str, default='critic', help="the type of bendmark")
parser.add_argument('--penalty', type=float, default=0.0001, help="the type of bendmark")#0.0001
parser.add_argument('--optimizer', type=str, default="adam", help="the type of optimizer")#
parser.add_argument('--method', type=str, default='policy', help="method for generate actions")
args = parser.parse_args()
start = time.time()
print("DEBUG : RCL.py start RCL train")
jason = RCL(args)
print("DEBUG : RCL.py end RCL train")
end = time.time()
params = jason.best_params
print("DEBUG : RCL.py step2. argparse finished...")
if not os.path.exists(args.save_path):
os.makedirs(args.save_path)
print("DEBUG : RCL.py step3. mkdir finished...")
fname = "RCL_FC_" + args.data_path.split('/')[-1] + "_" + datetime.datetime.now().strftime("%Y-%m-%d_%H:%M:%S")
fname += '_' + str(args.lr) + str("_") + str(args.n_epochs) + '_' + str(args.max_trials) + '_' + str(args.batch_size) + \
'_' + args.bendmark + '_' + str(args.penalty) + '_' + args.optimizer + '_' + str(args.state_space) + '_' + \
str(end-start) + '_' + args.method
fname = os.path.join(args.save_path, fname)
print("DEBUG : RCL.py step4. fsave finished...")
f = open(fname + '.txt', 'w')
accuracy = []
for index,value in params.items():
print([_.shape for _ in value[1]], file=f)
accuracy.append(value[0])
print(accuracy,file=f)
f.close()
print(fname)
print("DEBUG : RCL.py step5. fclose for txt finished...")
name = fname + '.pkl'
f = open(name, 'wb')
pickle.dump(jason.result_process, f)
f.close()
print("DEBUG : RCL.py step6. fclose for pkl finished...")