def __init__(self, input_shape, z_size, variational=True, first_kernel=5, first_stride=2, second_kernel=5, second_stride=2):
self.input_shape = input_shape
self.z_size = z_size
encoder = self.Encoder(input_shape, z_size, first_kernel, first_stride, second_kernel, second_stride)
decoder = self.Decoder(z_size, encoder.z_shape, first_kernel, first_stride, second_kernel, second_stride)
BaseVAE.__init__(self, encoder, decoder, variational)
Storeable.__init__(self)
def __init__(self,
input_shape,
z_size,
variational=True,
first_kernel=5,
first_stride=2,
second_kernel=5,
second_stride=2):
self.vae = ConvVAE4Fixed(self, input_shape, z_size, variational,
first_kernel, first_stride, second_kernel,
second_stride)
self.shot_vae = ConvVAE4Fixed(self, input_shape, z_size, variational,
1, first_kernel, first_stride,
second_kernel, second_stride)
Storeable.__init__(self)
def __init__(self,
input_shape,
z_size,
input_channels=3,
output_channels=3,
first_kernel=5,
first_stride=2,
second_kernel=5,
second_stride=2):
self.input_shape = input_shape
self.z_size = z_size
encoder = self.Encoder(input_shape, z_size, input_channels,
first_kernel, first_stride, second_kernel,
second_stride)
decoder = self.Decoder(z_size, encoder.z_shape, output_channels,
first_kernel, first_stride, second_kernel,
second_stride)
BaseAE.__init__(self, encoder, decoder)
Storeable.__init__(self)
def test_rollout_gen(self):
env = gym.make('SpaceInvaders-v4')
models = config.basepath() / 'SpaceInvaders-v4' / 'models'
visualsfile = models / 'GM53H301W5YS38XH'
visuals = Storeable.load(str(visualsfile)).to(config.device())
controllerfile = models / 'best_model68'
controller = torch.load(str(controllerfile))
policy = VCPolicy(visuals, controller)
viewer = ImageViewer('screen', (420, 360), 'numpyRGB')
for screen, observation, reward, done, info, action in RolloutGen(
env, policy):
viewer.update(screen)
def test_gymsim(self):
env = gym.make('SpaceInvaders-v4')
models = config.basepath() / 'SpaceInvaders-v4' / 'models'
visualsfile = models / 'GM53H301W5YS38XH'
visuals = Storeable.load(str(visualsfile)).to(config.device())
controllerfile = models / 'best_model68'
controller = torch.load(str(controllerfile))
policy = VCPolicy(visuals, controller)
dataset = GymSimulatorDataset(env, policy, 3000)
loader = torch.utils.data.DataLoader(dataset=dataset,
batch_size=10,
shuffle=False,
drop_last=True)
for screen, observation, action, reward, done, _ in loader:
print(reward)
drop_last=False,
)
train = data_utils.DataLoader(
train,
batch_size=batch_size,
collate_fn=collate_action_observation,
drop_last=False,
)
test = data_utils.DataLoader(
test,
batch_size=batch_size,
collate_fn=collate_action_observation,
drop_last=False,
)
convolutions = Storeable.load('C:\data\models\GM53H301W5YS38XH').to(device)
convolutions.decoder.register_forward_hook(
UniImageViewer('decoded').view_output)
ground_truth_decoder = Storeable.load(
'C:\data\models\GM53H301W5YS38XH').to(device)
ground_truth_decoder.decoder.register_forward_hook(
UniImageViewer('ground_truth').view_input)
model = MDNRNN(i_size=i_size,
z_size=z_size,
hidden_size=32,
num_layers=3,
n_gaussians=3).to(device)
#model = Storeable.load(r'C:\data\runs\687\mdnrnn-i_size-6-z_size-16-hidden_size-256-num_layers-1-n_gaussians-5_10.md').to(device)
rollout.register_before_hook(frame)
def save(args):
action_encoder.save_session()
def action_encoder_frame(step):
episode = step.meta['episode']
file_path = config.basepath() / gym_environment / output_dir / str(episode)
step.meta['filename'] = str(file_path)
action_encoder.update(step)
rollout.register_step(action_encoder_frame)
rollout.register_end_session(save)
for i_episode in range(531, 1000):
rollout.rollout(policy, max_timesteps=3000, episode=i_episode)
if __name__ == '__main__':
gym_environment = 'SpaceInvaders-v4'
visuals = Storeable.load('.\modelzoo\GM53H301W5YS38XH').to(config.device())
controller = torch.load(r'.\modelzoo\best_model68')
policy = VCPolicy(visuals, controller)
output_dir = 'rl_raw_v2'
main(gym_environment, policy, output_dir)