hongshaorou/dreamerv3-torch
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

merged action head into MLP and modified configs

Browse Source
This commit is contained in:
NM512
2024-01-05 10:26:48 +09:00
parent e0f2017e28
commit e0487f8206
5 changed files with 133 additions and 231 deletions
Download Patch File Download Diff File Expand all files Collapse all files

63
configs.yaml
View File

@@ -47,26 +47,25 @@ defaults:
dyn_temp_post: True
grad_heads: ['decoder', 'reward', 'cont']
units: 512
reward_layers: 2
cont_layers: 2
value_layers: 2
actor_layers: 2
act: 'SiLU'
norm: 'LayerNorm'
encoder:
{mlp_keys: '$^', cnn_keys: 'image', act: 'SiLU', norm: 'LayerNorm', cnn_depth: 32, kernel_size: 4, minres: 4, mlp_layers: 2, mlp_units: 512, symlog_inputs: True}
{mlp_keys: '$^', cnn_keys: 'image', act: 'SiLU', norm: True, cnn_depth: 32, kernel_size: 4, minres: 4, mlp_layers: 2, mlp_units: 512, symlog_inputs: True}
decoder:
{mlp_keys: '$^', cnn_keys: 'image', act: 'SiLU', norm: 'LayerNorm', cnn_depth: 32, kernel_size: 4, minres: 4, mlp_layers: 2, mlp_units: 512, cnn_sigmoid: False, image_dist: mse, vector_dist: symlog_mse}
value_head: 'symlog_disc'
reward_head: 'symlog_disc'
{mlp_keys: '$^', cnn_keys: 'image', act: 'SiLU', norm: True, cnn_depth: 32, kernel_size: 4, minres: 4, mlp_layers: 2, mlp_units: 512, cnn_sigmoid: False, image_dist: mse, vector_dist: symlog_mse, outscale: 1.0}
actor:
{layers: 2, dist: 'normal', entropy: 3e-4, unimix_ratio: 0.01, min_std: 0.1, max_std: 1.0, temp: 0.1, lr: 3e-5, eps: 1e-5, grad_clip: 100.0, outscale: 1.0}
critic:
{layers: 2, dist: 'symlog_disc', slow_target: True, slow_target_update: 1, slow_target_fraction: 0.02, lr: 3e-5, eps: 1e-5, grad_clip: 100.0, outscale: 0.0}
reward_head:
{layers: 2, dist: 'symlog_disc', scale: 1.0, outscale: 0.0}
cont_head:
{layers: 2, scale: 1.0, outscale: 1.0}
dyn_scale: 0.5
rep_scale: 0.1
kl_free: 1.0
cont_scale: 1.0
reward_scale: 1.0
weight_decay: 0.0
unimix_ratio: 0.01
action_unimix_ratio: 0.01
initial: 'learned'
# Training
@@ -77,15 +76,7 @@ defaults:
model_lr: 1e-4
opt_eps: 1e-8
grad_clip: 1000
value_lr: 3e-5
actor_lr: 3e-5
ac_opt_eps: 1e-5
value_grad_clip: 100
actor_grad_clip: 100
dataset_size: 1000000
slow_value_target: True
slow_target_update: 1
slow_target_fraction: 0.02
opt: 'adam'
# Behavior.
@@ -95,18 +86,10 @@ defaults:
imag_gradient: 'dynamics'
imag_gradient_mix: 0.0
imag_sample: True
actor_dist: 'normal'
actor_entropy: 3e-4
actor_state_entropy: 0.0
actor_init_std: 1.0
actor_min_std: 0.1
actor_max_std: 1.0
actor_temp: 0.1
expl_amount: 0.0
expl_amount: 0
eval_state_mean: False
collect_dyn_sample: True
behavior_stop_grad: True
value_decay: 0.0
future_entropy: False
# Exploration
@@ -150,13 +133,12 @@ crafter:
dyn_hidden: 1024
dyn_deter: 4096
units: 1024
reward_layers: 5
cont_layers: 5
value_layers: 5
actor_layers: 5
encoder: {mlp_keys: '$^', cnn_keys: 'image', cnn_depth: 96, mlp_layers: 5, mlp_units: 1024}
decoder: {mlp_keys: '$^', cnn_keys: 'image', cnn_depth: 96, mlp_layers: 5, mlp_units: 1024}
actor_dist: 'onehot'
actor: {layers: 5, dist: 'onehot'}
value: {layers: 5}
reward_head: {layers: 5}
cont_head: {layers: 5}
imag_gradient: 'reinforce'
atari100k:
@@ -166,7 +148,7 @@ atari100k:
train_ratio: 1024
video_pred_log: true
eval_episode_num: 100
actor_dist: 'onehot'
actor: {dist: 'onehot'}
imag_gradient: 'reinforce'
stickey: False
lives: unused
@@ -189,13 +171,12 @@ minecraft:
dyn_hidden: 1024
dyn_deter: 4096
units: 1024
reward_layers: 5
cont_layers: 5
value_layers: 5
actor_layers: 5
encoder: {mlp_keys: 'inventory|inventory_max|equipped|health|hunger|breath|reward', cnn_keys: 'image', cnn_depth: 96, mlp_layers: 5, mlp_units: 1024}
encoder: {mlp_keys: 'inventory|inventory_max|equipped|health|hunger|breath|obs_reward', cnn_keys: 'image', cnn_depth: 96, mlp_layers: 5, mlp_units: 1024}
decoder: {mlp_keys: 'inventory|inventory_max|equipped|health|hunger|breath', cnn_keys: 'image', cnn_depth: 96, mlp_layers: 5, mlp_units: 1024}
actor_dist: 'onehot'
actor: {layers: 5, dist: 'onehot'}
value: {layers: 5}
reward_head: {layers: 5}
cont_head: {layers: 5}
imag_gradient: 'reinforce'
break_speed: 100.0
time_limit: 36000
@@ -203,7 +184,7 @@ minecraft:
memorymaze:
steps: 1e8
action_repeat: 2
actor_dist: 'onehot'
actor: {dist: 'onehot'}
imag_gradient: 'reinforce'
task: 'memorymaze_9x9'