def __init__(self, env, agent, replay_buffer, arm):
self.env = env
self.agent = agent
self.replay_buffer = replay_buffer
self.obs = env.reset(arm)
self.done = False
self.arm = arm
self.observation_steps = 50 # 200, need to change average to match in train.py
self.stacked_frames = 0
self.episode = 0
self.video_record = True
self.frames_total = 0
self.state = self.env.getInputImage()
self.state, self.stacked_frames = stack_frames(stack_frames,
self.state, True)
self.score = []
self.video_array = []
self.distance = self.env.calcDistance()
self.solved = False
self.index = 0
if self.episode % cons.VIDEO_INTERVAL == 0:
self.video_record = True
else:
self.video_record = False
self.video_array.append(self.env.getVideoImage())
# Reset the graph
tf.reset_default_graph()
# Instantiate the DQNetwork
DQNetwork = DQNetwork(state_size, action_size, learning_rate)
# Instantiate memory
### create empty memory
memory = Memory(max_size=memory_size)
for i in range(pretrain_length):
# If it's the first step
if i == 0:
state = env.reset()
state, stacked_frames = stack_frames(stacked_frames, state, True, stack_size)
# Get the next_state, the rewards, done by taking a random action
choice = random.randint(1, len(possible_actions)) - 1
action = possible_actions[choice]
next_state, reward, done, _ = env.step(action)
# env.render()
# Stack the frames
next_state, stacked_frames = stack_frames(stacked_frames, next_state, False, stack_size)
# If the episode is finished (we're dead 3x)
if done:
# We finished the episode
next_state = np.zeros(state.shape)
temperature = -1
size = (512, 512)
time.sleep(2)
start_time = time.time()
while frames_total < cons.num_frames:
episode += 1
state = sim.get_input_image()
state, stacked_frames = stack_frames(stacked_frames, state, True,
cons.num_frames_stacked)
score = []
video_array = []
distance = sim.calc_distance()
solved = False
index = 0
if episode % cons.video_interval == 0:
video_record = True
else:
video_record = False
# last speaker for scoring
score_speakers = [speakers[i] for i in ids[2:]]
# ====== generate training and validating data ====== #
X_train = []
y_train = []
X_score = []
y_score = []
first_sample = None
for name in all_name:
features = feat_data[name]
# change the INPUT_FEATURE index to
# use different feature for training
x = features[INPUT_FEATURE]
num_frames, num_features = x.shape
# adding context window
x = stack_frames(x, frame_length=CONTEXT_LENGTH * 2 + 1)
# sequencing the image
x = np.reshape(x,
newshape=(num_frames, CONTEXT_LENGTH * 2 + 1, num_features))
y = [int(name.split('_')[0])] * len(x)
# add to appropriate set
if any(spk in name for spk in train_speakers):
X_train.append(x)
y_train += y
if first_sample is None:
first_sample = name
else:
X_score.append(x)
y_score += y
# ====== merge all array into a matrix ====== #
X_train = np.concatenate(X_train, axis=0)
X_train = []
y_train = []
X_score = []
y_score = []
first_sample = None
for name in all_name:
features = feat_data[name]
# change the INPUT_FEATURE index to
# use different feature for training
x = features[INPUT_FEATURE]
num_frames, num_features = x.shape
# adding context window
# TODO: replace `None` with appropriate `frame_length` based on
# `CONTEXT_LENGTH`, remember we have left context, right context
# and main frame in the middle
x = stack_frames(x, frame_length=None)
# sequencing the image
x = np.reshape(x,
newshape=(num_frames, CONTEXT_LENGTH * 2 + 1, num_features))
y = [int(name.split('_')[0])] * len(x)
# add to appropriate set
if any(spk in name for spk in train_speakers):
X_train.append(x)
y_train += y
if first_sample is None:
first_sample = name
else:
X_score.append(x)
y_score += y
# ====== merge all array into a matrix ====== #
X_train = np.concatenate(X_train, axis=0)
# def add(self, experience):
# self.buffer.append(experience)
# def sample(self, batch_size):
# buffer_size = len(self.buffer)
# index = np.random.choice(np.arange(buffer_size),
# size = batch_size,
# replace = False)
# return [self.buffer[i] for i in index]
memory = Memory(max_size = memory_size)
stacked_frames = deque([np.zeros((84,84), dtype=np.int) for i in range(stack_size)], maxlen=stack_size)
for i in range(pretrain_length):
if i == 0:
frame = env.reset()
stacked_frames, stacked_state = utils.stack_frames(stacked_frames, frame, is_new_episode)
action = np.random.randint(0, action_size)
new_frame, reward, done, _ = env.step(action)
stacked_frames, new_stacked_state = utils.stack_frames(stacked_frames, new_frame, is_new_episode)
memory.add((stacked_state, action, reward, new_stacked_state, done))
if done:
frame = env.reset()
is_new_episode = True
stacked_frames, stacked_state = utils.stack_frames(stacked_frames, frame, is_new_episode)
is_new_episode = False
else:
stacked_state = new_stacked_state.copy()
"""
This function will do the part