class SSNES():
def __init__(self, x0, learning_rate_mult, popsize):
self.snes = SNES(x0, learning_rate_mult, popsize)
self.gen()
def gen(self):
self.asked = self.snes.ask()
self.scores = {n: [] for n in range(len(self.asked))}
def predict(self):
r = randint(0, len(self.asked) - 1)
asked = self.asked[r]
return asked, r
def fit(self, scores, r):
# sort them out
for i, score in enumerate(scores):
self.scores[r[i]].append(score)
told = []
for i in range(len(self.asked)):
told.append(np.array(self.scores[i]).mean())
self.snes.tell(self.asked, told)
self.gen()
def main():
print("Total number of weights to evolve is:", weights.shape)
all_examples_indices = list(range(x_train.shape[0]))
clf, _ = train_classifier(model, x_train, y_train)
y_pred = predict_classifier(model, clf, x_test)
print(y_test.shape, y_pred.shape)
test_accuracy = accuracy_score(y_test, y_pred)
print('Non-trained NN Test accuracy:', test_accuracy)
# print('Test MSE:', test_mse)
snes = SNES(weights, 1, POPULATION_SIZE)
for i in range(0, GENERATIONS):
start = timer()
asked = snes.ask()
# to be provided back to snes
told = []
# use a small number of training samples for speed purposes
subsample_indices = np.random.choice(all_examples_indices, size=SAMPLE_SIZE, replace=False)
# evaluate on another subset
subsample_indices_valid = np.random.choice(all_examples_indices, size=SAMPLE_SIZE + 1, replace=False)
# iterate over the population
for asked_j in asked:
# set nn weights
nnw.set_weights(asked_j)
# train the classifer and get back the predictions on the training data
clf, _ = train_classifier(model, x_train[subsample_indices], y_train[subsample_indices])
# calculate the predictions on a different set
y_pred = predict_classifier(model, clf, x_train[subsample_indices_valid])
score = accuracy_score(y_train[subsample_indices_valid], y_pred)
# clf, _ = train_classifier(model, x_train, y_train)
# y_pred = predict_classifier(model, clf, x_test)
# score = accuracy_score(y_test, y_pred)
# append to array of values that are to be returned
told.append(score)
snes.tell(asked, told)
end = timer()
print("It took", end - start, "seconds to complete generation", i + 1)
nnw.set_weights(snes.center)
clf, _ = train_classifier(model, x_train, y_train)
y_pred = predict_classifier(model, clf, x_test)
print(y_test.shape, y_pred.shape)
test_accuracy = accuracy_score(y_test, y_pred)
print('Test accuracy:', test_accuracy)
def train_agent(game_chosen,game,run,pop_size,generations):
"""
function to run the game
Args:
game_chosen (str): name of the game
game(str): folder name
run(int): folder name
pop_size: population size for the SNES algorithm
generations: number of generations for the SNES algorithm
"""
# start the desired game file
textPlayer = tP.TextPlayer(game_chosen)
# initialize a list to keep track of the game scores
track_scores = []
# set last score to zero
last_score = 0
# initialize the agent
agent = Neuroagent_agent.NeuroAgent()
# return the original weights of the neural network structure in the neuroevolution
initial_weights = agent.agent_return_weights()
state = textPlayer.run()
last_state = state
# pass variables to SNES
snes = SNES(initial_weights, 1, pop_size)
# start the timer
start = timer()
# iterate through number of generations
for i in range(0, generations):
asked = snes.ask()
told = []
j = 0
for asked_j in asked:
# use SNES to set the weights of the NN
last_state = state
action = agent.take_action(state,asked_j)
state = textPlayer.execute_command(action)
print('{0} >>> {1}'.format(action,state))
print('This is Population No. {0} of Generation no. {1}'.format(j,i))
if textPlayer.get_score() != None:
score, possible_score = textPlayer.get_score()
# if previous state is equal to current state, then agent reward gets deducted by value of -0.2 otherwise reward multiplies by a factor of * 1.2
reward = score - last_score
if last_state == state:
agent_reward = reward - 0.2
else:
agent_reward = reward * 1.2
told.append(agent_reward)
last_score = score
agent.update(reward)
accumulated_reward = agent.get_total_points_earned()
print ('Your overall score is {0} and you gained reward of {1} in the last action and agent reward of {2}'.format(accumulated_reward,reward,agent_reward))
track_scores.append((state,j,i,action,agent_reward,reward,accumulated_reward))
else:
#in case the game cant retrieve the score from the game engine, set the score for that generation to 0
score = 0
# if previous state is equal to current state, then agent reward gets deducted by value of -0.2 otherwise reward multiplies by a factor of * 1.2
reward = score - last_score
if last_state == state:
agent_reward = reward - 0.2
else:
agent_reward = reward * 1.2
told.append(agent_reward)
last_score = score
agent.update(reward)
accumulated_reward = agent.get_total_points_earned()
print ('Your overall score is {0} and you gained reward of {1} in the last action and agent reward of {2}'.format(accumulated_reward,reward,agent_reward))
track_scores.append((state,j,i,action,agent_reward,reward,accumulated_reward))
j += 1
snes.tell(asked,told)
save(i,track_scores,game,run)
tokens, labels = agent.agent_return_models()
save_tokens_labels_words_seen(i, tokens, labels,game,run)
# at end of training
# pass the final weights to the neural network structure
snes_centre_weights = snes.center
agent.pass_snes_centre_weight(snes_centre_weights)
word_seen = agent.agent_return_word_seen()
# save the scores
with open('./completed_runs/'+game+'/finalruns/run_'+str(run)+'/track_scores_'+game+'_pop50gen500.pkl', 'wb') as f:
pickle.dump(track_scores, f)
# save the words seen in the game
with open('./completed_runs/'+game+'/finalruns/run_'+str(run)+'/words_seen_'+game+'_pop50gen500.pkl', 'wb') as h:
pickle.dump(word_seen, h)
with open('./completed_runs/'+game+'/finalruns/run_'+str(run)+'/test_word_seen_vectors_'+game+'_pop50gen500.pkl', 'wb') as i:
pickle.dump(tokens, i)
with open('./completed_runs/'+game+'/finalruns/run_'+str(run)+'/test_word_seen_vocab_'+game+'_pop50gen500.pkl', 'wb') as k:
pickle.dump(labels, k)
# end the timer
end = timer()
print('The time taken to run the traning is {0}'.format(end - start))
textPlayer.quit()
# this is irrelevant for what we want to achieve
model.compile(loss="mse", optimizer="adam")
print("compilation is over")
nnw = NNWeightHelper(model)
weights = nnw.get_weights()
print("Total number of weights to evolve is:", weights.shape)
all_examples_indices = list(range(x_train.shape[0]))
clf, _ = train_classifier(model, x_train, y_train)
y_pred = predict_classifier(model, clf, x_test)
test_accuracy = accuracy_score(y_test, y_pred)
print('Non-trained NN Test accuracy:', test_accuracy)
# print('Test MSE:', test_mse)
snes = SNES(weights, 1, POPULATION_SIZE)
p = []
for i in range(0, GENERATIONS):
start = timer()
asked = snes.ask()
# to be provided back to snes
told = []
# use a small number of training samples for speed purposes
subsample_indices = np.random.choice(all_examples_indices,
size=SAMPLE_SIZE,
replace=False)
# evaluate on another subset
subsample_indices_valid = np.random.choice(all_examples_indices,
size=SAMPLE_SIZE + 1,
def __init__(self, x0, learning_rate_mult, popsize):
self.snes = SNES(x0, learning_rate_mult, popsize)
self.gen()
#weights = K.get_trainable_weights(model)
#dim = len(weights)
#steps = 2*SAMPLES/GENERATION**2/BATCH_SIZE #target generations for SNES to reach full adaptation
batch_size = int(len(source_adapt) / round(len(source_adapt) / BATCH_SIZE, 0))
steps = 2 * SAMPLES / GENERATION / batch_size
#batches = int(len(source_adapt)/batch_size) #number iterations for each epoch
batches = int(GENERATION * len(source_adapt) / batch_size)
batches -= batches % GENERATION
trgbatch_size = int(
len(target_adapt) / round(len(target_adapt) / batch_size, 0))
trg_batches = int(GENERATION * len(target_adapt) / trgbatch_size)
yd_fit = np.concatenate([np.tile(0, batch_size), np.tile(1, trgbatch_size)])
#epochs = int(EPOCHS*SAMPLES_PER_EPOCH/batch_size/batches)
epochs = int(EPOCHS * GENERATION * SAMPLES_PER_EPOCH / batch_size / batches)
snes = SNES(weights, LEARN_RATE, GENERATION)
#snes = SSNES(weights, 2, GENERATION)#np.zeros(dim), 1, GENERATION)
fitnesses = np.zeros(GENERATION)
losses = [[], [], [], []]
asked = snes.ask()
#asked = snes.asked
"""----------------------commence adaptation---------------------------- """
t0 = time.time()
step = 0
genloss = []
timeloss = []
trg_batch = 0
trg_start = 0
trgtrain_end = trgbatch_size
trgtest_start = trgtrain_end
trgtest_end = 2 * trgtrain_end
#hp.setrelheap()
track_scores = []
last_score = 0
# defining population size
pop_size = 10
# defining generation size
generations = 100
# init the agent
agent = vivan_agent.vivAgent()
initial_weights = agent.agent_return_weights()
state = textPlayer.run()
last_state = state
# print (state)
# pass variables to SNES
snes = SNES(initial_weights, 1, pop_size)
start = timer()
#while (counter < training_cycles):
for i in range(0, generations):
asked = snes.ask()
told = []
j = 0
for asked_j in asked:
# use SNES to set the weights of the NN
# only run the SNES after 1st round
# pass asked to function in vivan_wordfiles.py everytime
# init the neural network in vivan_wordfiles.py
# but it is run only when the def_results_for_words is called.
# how about reward?
