# device = torch.device(c.CPU)
action_repeat = 1
num_frames = 1
hidden_state_dim = 128
memory_size = 1000000
max_total_steps = 1000000
experiment_setting = {
# Auxiliary Tasks
c.AUXILIARY_TASKS: {},
# Buffer
c.BUFFER_PREPROCESSING:
gt.Compose([gt.AsType()]),
c.BUFFER_SETTING: {
c.KWARGS: {
c.MEMORY_SIZE: memory_size,
c.OBS_DIM: (obs_dim, ),
c.H_STATE_DIM: (hidden_state_dim * 2, ),
c.ACTION_DIM: (action_dim, ),
c.REWARD_DIM: (1, ),
c.INFOS: {
c.MEAN: ((action_dim, ), np.float32),
c.VARIANCE: ((action_dim, ), np.float32),
c.ENTROPY: ((action_dim, ), np.float32),
c.LOG_PROB: ((1, ), np.float32),
c.VALUE: ((1, ), np.float32),
c.DISCOUNTING: ((1, ), np.float32)
},
obs_dim = int(np.product(raw_img_dim) + scalar_feature_dim)
latent_dim = 50
memory_size = max_total_steps = 1000000 // action_repeat
device = torch.device("cuda:0")
# device = torch.device(c.CPU)
experiment_setting = {
# Auxiliary Tasks
c.AUXILIARY_TASKS: {},
# Buffer
c.BUFFER_PREPROCESSING:
gt.Compose([
gt.Transpose((0, 3, 1, 2)),
gt.Reshape(),
]),
c.BUFFER_SETTING: {
c.KWARGS: {
c.MEMORY_SIZE: memory_size,
c.OBS_DIM: (obs_dim, ),
c.H_STATE_DIM: (1, ),
c.ACTION_DIM: (action_dim, ),
c.REWARD_DIM: (1, ),
c.INFOS: {
c.MEAN: ((action_dim, ), np.float32),
c.VARIANCE: ((action_dim, ), np.float32),
c.ENTROPY: ((action_dim, ), np.float32),
c.LOG_PROB: ((1, ), np.float32),
c.VALUE: ((1, ), np.float32),
c.DISCOUNTING: ((1, ), np.float32)
latent_dim = 50
memory_size = max_total_steps = 100000 // action_repeat
device = torch.device("cuda:0")
# device = torch.device(c.CPU)
experiment_setting = {
# Auxiliary Tasks
c.AUXILIARY_TASKS: {},
# Buffer
c.BUFFER_PREPROCESSING:
gt.Compose([
gt.Transpose((0, 3, 1, 2)),
it.NumPyCenterCrop(raw_img_dim, height=processed_h, width=processed_w),
gt.Reshape(),
]),
c.BUFFER_SETTING: {
c.KWARGS: {
c.MEMORY_SIZE: memory_size,
c.OBS_DIM: (obs_dim, ),
c.H_STATE_DIM: (1, ),
c.ACTION_DIM: (action_dim, ),
c.REWARD_DIM: (1, ),
c.INFOS: {
c.MEAN: ((action_dim, ), np.float32),
c.VARIANCE: ((action_dim, ), np.float32),
c.ENTROPY: ((action_dim, ), np.float32),
c.LOG_PROB: ((1, ), np.float32),
c.VALUE: ((1, ), np.float32),