-
Notifications
You must be signed in to change notification settings - Fork 0
/
MCTS_chemistry.py
348 lines (288 loc) · 12.3 KB
/
MCTS_chemistry.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
346
347
348
import pickle
import os
import random
import json
import csv
import numpy as np
import torch
from Node import Node
from ChemModel import translator
from schemes import chemistry
import time
from sampling import sampling_node
def num2ord(num):
if num % 10 == 1:
ord_str = str(num) + 'st'
elif num % 10 == 2:
ord_str = str(num) + 'nd'
elif num % 10 == 3:
ord_str = str(num) + 'rd'
else:
ord_str = str(num) + 'th'
return ord_str
class MCTS:
def __init__(self, search_space, tree_height, arch_code_len):
assert type(search_space) == type([])
assert len(search_space) >= 1
assert type(search_space[0]) == type([])
self.search_space = search_space
self.ARCH_CODE_LEN = arch_code_len
self.ROOT = None
self.Cp = 0.5
self.nodes = []
self.samples = {}
self.TASK_QUEUE = []
self.DISPATCHED_JOB = {}
self.energy_list = []
self.JOB_COUNTER = 0
self.TOTAL_SEND = 0
self.TOTAL_RECV = 0
self.ITERATION = 0
self.MAX_MAEINV = 0
self.MAX_SAMPNUM = 0
self.sample_nodes = []
self.tree_height = tree_height
# initialize a full tree
total_nodes = 2**tree_height - 1
for i in range(1, total_nodes + 1):
is_good_kid = False
if (i-1) > 0 and (i-1) % 2 == 0:
is_good_kid = False
if (i-1) > 0 and (i-1) % 2 == 1:
is_good_kid = True
parent_id = i // 2 - 1
if parent_id == -1:
self.nodes.append(Node(None, is_good_kid, self.ARCH_CODE_LEN, True))
else:
self.nodes.append(Node(self.nodes[parent_id], is_good_kid, self.ARCH_CODE_LEN, False))
self.ROOT = self.nodes[0]
self.CURT = self.ROOT
self.init_train()
def init_train(self):
for i in range(0, 3000):
net = random.choice(self.search_space)
self.search_space.remove(net)
self.TASK_QUEUE.append(net)
self.sample_nodes.append('random')
print("\ncollect " + str(len(self.TASK_QUEUE)) + " nets for initializing MCTS")
def dump_all_states(self, num_samples):
node_path = 'states/mcts_agent'
with open(node_path+'_'+str(num_samples), 'wb') as outfile:
pickle.dump(self, outfile)
def reset_node_data(self):
for i in self.nodes:
i.clear_data()
def populate_training_data(self):
self.reset_node_data()
for k, v in self.samples.items():
self.ROOT.put_in_bag(json.loads(k), v)
def populate_validation_data(self):
for k, v in validation.items():
self.ROOT.validation[k] = v
self.ROOT.bag = self.ROOT.validation.copy()
def populate_prediction_data(self):
# self.reset_node_data()
for k in self.search_space:
self.ROOT.put_in_bag(k, 0.0)
def train_nodes(self):
for i in self.nodes:
i.train()
def predict_nodes(self, method = None, dataset =None):
for i in self.nodes:
if dataset:
i.predict_validation()
else:
i.predict(method)
def node_performance(self):
for i in self.nodes:
if i.is_leaf == False:
i.f1.append(i.kids[0].get_performance())
def check_leaf_bags(self):
counter = 0
for i in self.nodes:
if i.is_leaf is True:
counter += len(i.bag)
assert counter == len(self.search_space)
def reset_to_root(self):
self.CURT = self.ROOT
def print_tree(self):
print('\n'+'-'*100)
for i in self.nodes:
print(i)
print('-'*100)
def select(self):
self.reset_to_root()
curt_node = self.ROOT
self.ROOT.counter += 1
while curt_node.is_leaf == False:
UCT = []
for i in curt_node.kids:
UCT.append(i.get_uct(self.Cp))
curt_node = curt_node.kids[np.random.choice(np.argwhere(UCT == np.amax(UCT)).reshape(-1), 1)[0]]
self.nodes[curt_node.id].counter += 1
return curt_node
def evaluate_jobs(self):
while len(self.TASK_QUEUE) > 0:
job = self.TASK_QUEUE.pop()
sample_node = self.sample_nodes.pop()
try:
# print("\nget job from QUEUE:", job)
job_str = json.dumps(job)
report = {'energy': dataset.get(job_str)}
# design = translator(job)
# print("translated to:\n{}".format(design))
# print("\nstart training:")
# if job_str in dataset:
# report = {'energy': dataset.get(job_str)}
# # print(report)
# else:
# report = chemistry(design)
maeinv = -1 * report['energy']
self.DISPATCHED_JOB[job_str] = maeinv
self.samples[job_str] = maeinv
self.energy_list.append(-1 * maeinv)
with open('results.csv', 'a+', newline='') as res:
writer = csv.writer(res)
metrics = report['energy']
writer.writerow([len(self.samples), job_str, sample_node, metrics])
# print("\nresults of current model saved")
# save all models and reports
# torch.save(best_model.state_dict(), 'models/model_weights_'+str(len(self.samples))+'.pth')
# with open('reports/report_'+str(len(self.samples)), 'wb') as file:
# pickle.dump(report, file)
# if maeinv > self.MAX_MAEINV:
# self.MAX_MAEINV = maeinv
# self.MAX_SAMPNUM = len(self.samples)
# torch.save(best_model.state_dict(), 'model_weights.pth')
# with open('report', 'wb') as file:
# pickle.dump(report, file)
# print("better model saved")
# print("current min_mae: {}({} sample)".format(1/self.MAX_MAEINV, num2ord(self.MAX_SAMPNUM)))
# print("current number of samples: {}".format(len(self.samples)))
except Exception as e:
print(e)
self.TASK_QUEUE.append(job)
self.sample_nodes.append(sample_node)
print("current queue length:", len(self.TASK_QUEUE))
def search(self):
if len(self.ROOT.validation) == 0:
self.populate_validation_data()
self.predict_nodes('mean')
self.reset_node_data()
while len(self.search_space) > 0 and self.ITERATION < 20:
if self.ITERATION > 0:
self.dump_all_states(len(self.samples))
print("\niteration:", self.ITERATION)
# evaluate jobs:
print("\nevaluate jobs...")
self.evaluate_jobs()
print("\nfinished all jobs in task queue")
# assemble the training data:
print("\npopulate training data...")
self.populate_training_data()
print("finished")
# training the tree
print("\ntrain classifiers in nodes...")
if torch.cuda.is_available():
print("using cuda device")
else:
print("using cpu device")
start = time.time()
self.train_nodes()
print("finished")
end = time.time()
print("Running time: %s seconds" % (end - start))
# self.print_tree()
# clear the data in nodes
self.reset_node_data()
# print("\npopulate validation data...")
# self.ROOT.bag = self.ROOT.validation.copy()
# self.predict_nodes()
# self.node_performance()
# self.reset_node_data()
self.predict_nodes(None, 'validation')
self.node_performance()
self.reset_node_data()
print("\npopulate prediction data...")
self.populate_prediction_data()
print("finished")
print("\npredict and partition nets in search space...")
self.predict_nodes()
self.check_leaf_bags()
print("finished")
self.print_tree()
# sampling nodes
nodes = list(range(16)) #[0, 1, 2, 3, 12, 13, 14, 15]
sampling_node(self, nodes, dataset, self.ITERATION)
for i in range(0, 50):
# select
shift = 1
target_bin = self.select()
id = target_bin.id
sampled_arch = target_bin.sample_arch()
# NOTED: the sampled arch can be None
if sampled_arch is not None:
# TODO: back-propogate an architecture
# push the arch into task queue
# print("\nselected node " + str(target_bin.id-15) + " in leaf layer")
# print("sampled arch:", sampled_arch)
if json.dumps(sampled_arch) not in self.DISPATCHED_JOB:
self.TASK_QUEUE.append(sampled_arch)
self.search_space.remove(sampled_arch)
self.sample_nodes.append(target_bin.id-15)
else:
# trail 1: pick a network from the left leaf
for n in self.nodes:
if n.is_leaf == True:
sampled_arch = n.sample_arch()
if sampled_arch is not None:
print("\nselected node" + str(n.id-15) + " in leaf layer")
# print("sampled arch:", sampled_arch)
if json.dumps(sampled_arch) not in self.DISPATCHED_JOB:
self.TASK_QUEUE.append(sampled_arch)
self.search_space.remove(sampled_arch)
self.sample_nodes.append(n.id-15)
break
else:
continue
self.ITERATION += 1
if __name__ == '__main__':
# set random seed
random.seed(42)
np.random.seed(42)
torch.random.manual_seed(42)
with open('search_space_OH', 'rb') as file:
search_space = pickle.load(file)
arch_code_len = len(search_space[0])
print("\nthe length of architecture codes:", arch_code_len)
print("total architectures:", len(search_space))
with open('data/OH_dataset', 'rb') as file:
dataset = pickle.load(file)
# with open('data/chemistry_validation', 'rb') as file:
# validation = pickle.load(file)
validation = dict(list(dataset.items())[-10000:])
if os.path.isfile('results.csv') == False:
with open('results.csv', 'w+', newline='') as res:
writer = csv.writer(res)
writer.writerow(['sample_id', 'arch_code', 'sample_node', 'Energy'])
# agent = MCTS(search_space, dataset, 5, arch_code_len)
# agent.search()
state_path = 'states'
if os.path.exists(state_path) == False:
os.makedirs(state_path)
files = os.listdir(state_path)
if files:
files.sort(key=lambda x: os.path.getmtime(os.path.join(state_path, x)))
node_path = os.path.join(state_path, files[-1])
# node_path = 'states/mcts_agent_4000'
with open(node_path, 'rb') as json_data:
agent = pickle.load(json_data)
print("\nresume searching,", agent.ITERATION, "iterations completed before")
print("=====>loads:", len(agent.nodes), "nodes")
print("=====>loads:", len(agent.samples), "samples")
print("=====>loads:", len(agent.DISPATCHED_JOB), "dispatched jobs")
print("=====>loads:", len(agent.TASK_QUEUE), "task_queue jobs from node:", agent.sample_nodes[0])
agent.search()
else:
agent = MCTS(search_space, 5, arch_code_len)
agent.search()