# Extracting one image per class
indexes, labels = one_image_per_class(test_labels, len(class_names))
train_images = np.array([train_images[idx] for idx in indexes])
train_labels = np.array(labels)
# Extraction of a random image
image_index = random.randint(0, len(train_images) - 1)
#image_index = 1614
train_image = train_images[image_index, :, :, :]
train_label = int(train_labels[image_index])
train_image_4dim = np.reshape(train_image, (batch_size, 32, 32, 3))
# Convolutional features extraction
net_features = net.extract_features(train_image_4dim)
net_distribution = np.reshape(net(train_image_4dim).numpy(), (10, ))
# Environment initialization
environment = DyadicConvnetGymEnv(features=net_features,
image_class=train_label,
distribution=net_distribution,
max_steps=steps_per_episode)
num_actions = len(environment.actions)
environment = Environment.create(
environment=environment,
states=dict(
# 64 features + 3 positional coding
features=dict(type=float, shape=(67, )), ),
actions=dict(movement=dict(type=int, num_values=num_actions),
classification=dict(type=int,
num_values=len(class_names))),
max_episode_timesteps=steps_per_episode)
# Agent initialization
for i in range(1, int(last_episode / 1000) + 1):
old_episodes = i * 1000
print('Loading checkpoint. Last episode: %d' % old_episodes)
net.extract_features(image,
active_layers=layers,
last_layer=4))
idx += 1
test_images = copy.deepcopy(tmp)
del train_images, train_labels
del tmp
#########################################################################
# Environment initialization
environment = DyadicConvnetGymEnv(
dataset=test_images,
labels=test_labels,
images=RGB_images,
layers=layers,
max_steps=steps_per_episode,
visualize=visualize,
training=False,
class_penalty=class_penalty,
correct_class=correct_class,
illegal_mov=illegal_mov,
same_position=same_position,
non_classified=non_classified,
step_reward_multiplier=step_reward_multiplier)
num_actions = len(environment.actions)
environment = Environment.create(
environment=environment,
states=dict(features=dict(type=float, shape=(147, )), ),
actions=dict(type=int, num_values=num_actions + num_classes),
max_episode_timesteps=steps_per_episode)
dirs = ['models/RL/20210428-125328']
for directory in dirs:
check_dir = directory + '/checkpoints/'
old_epochs = 15
#threads = []
for size in num_agents:
environments = []
jedi_council = []
consensus = Consensus()
print('Testing {s} agents'.format(s=size))
for i in range(size):
environment = DyadicConvnetGymEnv(dataset=test_images,
labels=test_labels,
images=RGB_images,
layers=layers,
num_classes=10,
num_features=num_features,
max_steps=steps_per_episode,
visualize=visualize,
training=False,
class_penalty=class_penalty,
correct_class=correct_class,
illegal_mov=illegal_mov,
same_position=same_position,
non_classified=non_classified,
step_reward_multiplier=step_reward_multiplier
)
num_actions = len(environment.actions)
environment = Environment.create(environment=environment,
states=dict(
features=dict(type=float, shape=(num_features,)),
),
actions=dict(type=int, num_values=num_actions+num_classes),
max_episode_timesteps=steps_per_episode
)