def train_supervised(policy):
step = 50000
policy.train()
for i in range(step):
structure = random.choice(list(train_dict.keys()))
objectiveSeq = train_dict[structure][
0] # TODO select randomly insr=tead of 1st
N = len(objectiveSeq)
structureTemp, energy = RNA.fold(
rnalib.sequence_to_string(objectiveSeq))
if structureTemp != structure:
print('error 4534')
exit()
env = rnalib.RNAEnvironment(goal=structure, max_steps=1000)
optimum_actions = getBestActions(env.sequence, objectiveSeq)
# print(optimum_actions)
if i % 1000 == 0:
print('i=' + str(i) + '\t' + structure + '\tobj= ' +
str(objectiveSeq) + '\tinitSeq=' + str(env.sequence))
for a_tup in optimum_actions:
rnalib.update(actionTuple2Array(N, a_tup), policy, env)
reward = env.step(a_tup)
###########
if i % 10000 == 0:
a = array2actionTuple(
policy.get_action(np.expand_dims(env.state.T, axis=0)))
print('train_action = ' + str(a_tup) + '\t policy = ' + str(a))
def TestTabularPolicy(policy, max_steps=500, showStat=True):
structure = random.choice(list(train_dict.keys()))
objectiveSeq = train_dict[structure][0]
N = len(objectiveSeq)
env = rnalib.RNAEnvironment(goal=structure, max_steps=100)
# optimum_actions = getBestActions(env.sequence, objectiveSeq)
# print(env.terminated)
if showStat:
print('objective structure: ' + structure)
while not env.terminated:
a = epsilonGreedyAction(env.state, policy, epsilon=0.0)
prevState = env.state
r = env.step(a)
if showStat:
print('step: ' + str(env.count) + '\tseq = ' + str(env.sequence) +
'\t policy Actn = ' + str(a) + '\t objSeQ: ' +
str(objectiveSeq))
if r == 5:
if showStat:
print('Success')
flag = 0
while (env.state == prevState).all() and not env.terminated:
a = epsilonGreedyAction(env.state, policy, epsilon=1.0)
r = env.step(a)
flag = 1
if flag == 1:
if showStat:
print('step: ' + str(env.count) + '\tseq = ' +
str(env.sequence) + '\t RANDOM Actn = ' + str(a) +
'\t objSeQ: ' + str(objectiveSeq))
if r == 5:
if showStat:
print('***SUCCESS***')
return env.count
def train_DynaQ(alpha=0.5,
gamma=1.0,
numEpisodes=500,
maxPolicySize=10 * 1024 * 1024):
policy = {}
for e in range(numEpisodes):
structure = random.choice(list(train_dict.keys()))
N = len(structure)
env = rnalib.RNAEnvironment(goal=structure, max_steps=10000)
exp_replay = []
S = env.state
a = epsilonGreedyAction(S, policy, epsilon=0.6)
while not env.terminated:
r = env.step(a)
Sprime = env.state
aprime = epsilonGreedyAction(Sprime, policy, epsilon=0.6)
exp_replay.append((S, a, r, Sprime, aprime))
S = Sprime
a = aprime
for j in range(100):
for (S, a, r, Sprime,
aprime) in random.sample(exp_replay, len(exp_replay)):
policy[(list_to_tuple(S), a)] = policy[
(list_to_tuple(S), a)] + alpha * (r + gamma * policy[
(list_to_tuple(Sprime), aprime)] - policy[
(list_to_tuple(S), a)])
print(
str(e) + '\tpolicy Size: ' +
str(sys.getsizeof(policy) / (1024 * 1024)) + ' Mb')
if sys.getsizeof(policy) > maxPolicySize:
break
# print(policy)
# s = []
# testStrucks = random.sample(list(train_dict.keys()),min(100, len(train_dict.keys())))
# def findRNA(structure, policy, max_steps = 1000):
# env = rnalib.RNAEnvironment(goal = structure, max_steps = max_steps)
# while not env.terminated:
# a = epsilonGreedyAction(env.state, policy, epsilon = 0.0)
# prevState = env.state
# r = env.step(a)
# if r == 5:
# return 1
# while (env.state == prevState).all() and not env.terminated:
# a = epsilonGreedyAction(env.state, policy, epsilon = 1.0)
# r = env.step(a)
# if r == 5:
# return 1
# return 0
# count = 0
# for i in range(100):
# print(i,count)
# st = testStrucks[i%len(testStrucks)]
# count +=findRNA(st, policy)
# print('************')
# print(count)
return policy
def TestRandomPolicy():
structure = random.choice(list(train_dict.keys()))
objectiveSeq = train_dict[structure][0]
N = len(objectiveSeq)
env = rnalib.RNAEnvironment(goal = structure, max_steps = 10000)
max_step = 10e4
steps = 0
while not env.terminated and steps < max_step:
a = (random.randint(0,N-1),random.randint(0,3))
r = env.step(a)
steps += 1
return steps
def findRNA(structure, policy=None, useRandomPolicy=False):
env = rnalib.RNAEnvironment(goal=structure, max_steps=1000)
while not env.terminated:
if useRandomPolicy:
a = (random.randint(0, len(structure) - 1), random.randint(0, 3))
else:
if policy is None:
print('Need RNAPolicy if not random')
a = array2actionTuple(
policy.get_action(np.expand_dims(env.state.T, axis=0)))
prevState = env.state
r = env.step(a)
if r == 5:
return 1
if (env.state == prevState).all():
a = (random.randint(0, len(structure) - 1), random.randint(0, 3))
r = env.step(a)
if r == 5:
return 1
return 0
def DQNPolicy(random_policy=False):
counter,total_steps = 0,0
for puzzle in output:
if puzzle[0] == 'A' or puzzle[0] == 'U' or puzzle[0] == 'G' or puzzle[0] == 'C':
structure, _ = RNA.fold(puzzle)
else:
structure = puzzle
env = rnalib.RNAEnvironment(goal=structure,max_steps=1000)
if not random_policy:
policy = rnalib.RNA_BiLSTM_Policy(hidden_size= 15, num_layers= 4)
policy = torch.load('DQN_policy')
# print(len(structure))
steps = 0
start_time = time.time()
end_time = start_time + 30*60
f = 0
while not env.terminated and time.time() < end_time:
if random_policy:
a = (random.randint(0,len(structure)-1),random.randint(0,3))
else:
a = array2actionTuple(policy.get_action(np.expand_dims(env.state.T, axis=0)))
r = env.step(a)
steps += 1
if r == 5:
print('SUCCESS',structure,len(structure),steps)
counter += 1
total_steps += steps
f = 1
break
if f == 0:
print('FAILED',structure,len(structure),steps)
total_steps += steps
print("TOTAL SOLVED : ",counter, "AVERAGE :", total_steps/100 )
def TestPolicy(policy, max_steps=500, showStat=True):
structure = random.choice(list(train_dict.keys()))
objectiveSeq = train_dict[structure][0]
N = len(objectiveSeq)
env = rnalib.RNAEnvironment(goal=structure, max_steps=max_steps)
if showStat:
print('objective structure: ' + structure)
while not env.terminated:
a = array2actionTuple(
policy.get_action(np.expand_dims(env.state.T, axis=0)))
prevState = env.state
r = env.step(a)
if showStat:
print('step: ' + str(env.count) + '\tseq = ' + str(env.sequence) +
'\t policy Actn = ' + str(a) + '\t objSeQ: ' +
str(objectiveSeq))
if r == 5:
if showStat:
print('***SUCCESS***')
return env.count
break
flag = 0
while (env.state == prevState).all() and not env.terminated:
a = epsilonGreedyActionDQN(env.state, policy, epsilon=1.0)
r = env.step(a)
flag = 1
if flag == 1:
if showStat:
print('step: ' + str(env.count) + '\tseq = ' +
str(env.sequence) + '\t RANDOM Actn = ' + str(a) +
'\t objSeQ: ' + str(objectiveSeq))
if r == 5:
if showStat:
print('***SUCCESS***')
return env.count
return -1
def train_DQN(policy):
numEpisodes = 50
epsilon = 0.3
miniBatchSz = 50
alpha = 0.3
gamma = 1.0
frac = 0.3
policy.train()
for e in range(numEpisodes):
structure = random.choice(list(train_dict.keys()))
N = len(structure)
env = rnalib.RNAEnvironment(goal=structure, max_steps=10000)
exp_replay = []
print('STRUCTURE: ' + structure)
num_success = 0
while len(exp_replay) <= 500 or num_success <= miniBatchSz * frac:
S = env.state
a = epsilonGreedyActionDQN(S, policy,
epsilon=epsilon) #actionTuple
while not env.terminated:
r = env.step(a)
Sprime = env.state
aprime = epsilonGreedyActionDQN(Sprime,
policy,
epsilon=epsilon)
Y = r if env.terminated else r + gamma * np.max(
policy.get_action(np.expand_dims(Sprime.T, axis=0)))
if env.terminated:
# print('reward = 1 on DQNtrain')
exp_replay.append((S, a, Y, None))
num_success += 1
break
exp_replay.append((S, a, Y, Sprime))
S = Sprime
a = aprime
env.reset()
# print('##################### env reset exp_replay len = ' + str(len(exp_replay))+' num_success'+str(num_success))
# if True:
# print('---------')
# for (_,a,y,Sprime) in exp_replay[:5]:
# print(a,y,(Sprime is None))
# tempS = None
# for (tempS,a,y,Sprime) in [e for e in exp_replay if e[-1] is None][:5]:
# print(a,y,(Sprime is None))
# print(policy.get_action(np.expand_dims(tempS.T, axis=0)))
# print(policy.get_action(np.expand_dims(random.choice(exp_replay)[0].T, axis=0)))
# # exit()
print('training from exp_replay len = ' + str(len(exp_replay)) +
' \tEpisode: ' + str(e))
losses = []
for j in range(1000):
## force x% of minibatch to be terminal states
miniBatch = random.sample(exp_replay,
int(miniBatchSz * (1 - frac)))
miniBatch += random.sample(
[e for e in exp_replay if e[-1] is None],
int(miniBatchSz * frac))
# print(miniBatch)
losses.append(rnalib.updateParam(miniBatch, model=policy,
lr=alpha))
if j % 200 == 0:
print(losses[-1])
for b in range(4):
if a[i, b] == a[idx, base]:
base_list.append((i, b))
return random.choice(base_list)
# Attempt to solve a single puzzle. The target structure in dot bracket notation
# should be supplied as a command line argument.
puzzle = sys.argv[1]
if puzzle[0] == 'A' or puzzle[0] == 'U' or puzzle[0] == 'G' or puzzle[0] == 'C':
structure, _ = RNA.fold(puzzle)
else:
structure = puzzle
env = rnalib.RNAEnvironment(goal=structure, max_steps=1000)
policy = rnalib.RNA_BiLSTM_Policy(hidden_size=15, num_layers=4)
policy = torch.load('DQN_policy')
print(len(structure))
print(puzzle)
print(structure)
steps = 0
start_time = time.time()
end_time = start_time + 30 * 60
f = 0
while not env.terminated and time.time() < end_time:
a = array2actionTuple(
policy.get_action(np.expand_dims(env.state.T, axis=0)))
# a = (random.randint(0,len(structure)-1),random.randint(0,3))