class Game_Master():
def __init__(self,target_size,vocab_size,bias,test_data,seed,episodes=1000):
self.data = test_data
self.vocab_size = vocab_size
self.target_size = target_size
self.speaker_bias = 1.0 - bias
self.listener_bias = bias
self.speaker = Speaker(512,self.target_size,self.vocab_size,self.speaker_bias,seed)
self.listener = Listener(512,self.vocab_size,self.target_size,self.listener_bias,seed)
self.episode_rewards = []
self.accuracy_history = deque(maxlen=100)
self.symbols_used = {}
self.synonym_array = np.zeros((target_size,vocab_size))
self.accuracy_record = []
self.episodes = 1000
def create_labels(self,num_labels,num_repeats):
arr = np.array([])
for i in range(num_labels):
n = np.repeat(i,num_repeats)
arr = np.concatenate((arr,n))
np.random.shuffle(arr)
return arr
def sample(self,label,test_or_train="X_train"):
train = self.data[test_or_train]
size = train[0].shape[0]
target_one_hot = to_categorical(label,num_classes=self.target_size)
target_image = 0
image_list = []
for i in range(self.target_size):
sampled_image = np.expand_dims(train[i][np.random.choice(size),:],0)
if i == label:
target_image = sampled_image
image_list.append(sampled_image)
return target_image, image_list, target_one_hot
def cross_entropy(self,predictions, targets, epsilon=1e-12):
predictions = np.clip(predictions, epsilon, 1. - epsilon)
N = predictions.shape[0]
ce = -np.sum(np.sum(targets*np.log(predictions+1e-9)))/N
return ce
def test(self):
y = self.create_labels(10,1)
one_hot = to_categorical(y)
it = 0
while it < y.shape[0]:
label = y[it]
target_image, image_list, target_one_hot = self.sample(label,"X_test")
speaker_action, speaker_probs = self.speaker.act(target_image)
speaker_action_class = np.expand_dims(to_categorical(speaker_action,num_classes=self.vocab_size),0)
image_list.append(speaker_action_class)
listener_action, listener_probs = self.listener.act(image_list)
if listener_action == label:
speaker_reward = 1.0
listener_reward = 1.0
else:
speaker_reward = 0.0
listener_reward = 0.0
self.synonym_array[int(label)][int(speaker_action)] += 1
it += 1
def play(self):
y = self.create_labels(10,100)
one_hot = to_categorical(y)
bar = tqdm(np.arange(self.episodes))
for i in bar:
it = 0
acc = np.zeros(self.target_size)
right = 0
while it < y.shape[0]:
label = y[it]
target_image, image_list, target_one_hot = self.sample(label)
speaker_action, speaker_probs = self.speaker.act(target_image)
speaker_action_class = np.expand_dims(to_categorical(speaker_action,num_classes=self.vocab_size),0)
image_list.append(speaker_action_class)
listener_action, listener_probs = self.listener.act(image_list)
if listener_action == label:
speaker_reward = 1.0
listener_reward = 1.0
right += 1
else:
speaker_reward = 0.0
listener_reward = 0.0
acc = np.vstack((acc,listener_probs))
self.speaker.save(target_image,speaker_action,speaker_probs,speaker_reward)
self.listener.save(image_list,speaker_action_class,listener_action,listener_probs,listener_reward)
it += 1
accuracy = right/y.shape[0]
total = np.sum(self.speaker.params["rewards"])
cross = self.cross_entropy(acc[1:,:],one_hot)
self.episode_rewards.append(total)
self.accuracy_history.append(accuracy)
self.accuracy_record.append(accuracy)
bar.set_description("Cross Entropy: " + str(cross) + ", Rolling Acc 100: " + str(np.mean(self.accuracy_history)) + ", Accuracy: " + str(accuracy))
self.speaker.train()
self.listener.train()
self.test()
