huangfu/InteractionImitation
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fixing issues with lazylinear sequential, setting off a big run

Browse Source
This commit is contained in:
Arec Jamgochian
2022-02-25 17:48:56 -08:00
parent 2dfd7e3c2b
commit fa98601fa6
2 changed files with 19 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
sgail-ppo-options-setobs2.py
View File

@@ -108,9 +108,9 @@ def training_function(config):
value=value,
v_opt=v_opt,
v_iters=config['value']['iterations_per_epoch'],
epochs=2, # 200 FIXME
rollout_episodes=6, #60, FIXME
rollout_steps=6, #60, FIXME
epochs=200,
rollout_episodes=60,
rollout_steps=60,
gamma=0.99,
gae_lambda=0.9,
clip_ratio=config['policy']['clip_ratio'],
@@ -124,7 +124,6 @@ def training_function(config):
# save model
torch.save(policy.state_dict(), 'policy_final.pt')
analysis = tune.run(
training_function,
config={
@@ -135,7 +134,7 @@ analysis = tune.run(
'clip_ratio': 0.2, #tune.grid_search([0.2]),
'iterations_per_epoch': 100, #tune.grid_search([100]),
'hidden_layer_size': tune.grid_search([10, 20, 40]),
'n_hidden_layers': 1, #tune.grid_search([1, 2, 3]), #FIXME
'n_hidden_layers': tune.grid_search([2, 3, 4]),
'activation':tune.grid_search([torch.nn.LeakyReLU, torch.nn.Tanh]),
'option': tune.grid_search(list(range(len(option_list))))
},

32
src/core/policy.py
View File

@@ -36,15 +36,20 @@ class BasePolicy(nn.Module):
class Policy(BasePolicy):
def __init__(self, *args, **kwargs):
def __init__(self, *args, hidden_layer_size=50, n_hidden_layers=2, activation=nn.Tanh, **kwargs):
super().__init__(*args, **kwargs)
self.nn = nn.Sequential(
nn.LazyLinear(50),
nn.Tanh(),
nn.LazyLinear(50),
nn.Tanh(),
nn.LazyLinear(2 * self.action_dim),
)
layers = sum([[nn.LazyLinear(hidden_layer_size),
activation()] for _ in range(n_hidden_layers)],[])
self.nn = nn.Sequential(*layers, nn.LazyLinear(2 *self.action_dim))
# old
# self.nn = nn.Sequential(
# nn.LazyLinear(50),
# nn.Tanh(),
# nn.LazyLinear(50),
# nn.Tanh(),
# nn.LazyLinear(2 * self.action_dim),
#)
def forward(self, states):
return self.nn(states)
@@ -53,17 +58,10 @@ class DiscretePolicy(BasePolicy):
def __init__(self, *args, hidden_layer_size=50, n_hidden_layers=2, activation=nn.Tanh, **kwargs):
super().__init__(*args, **kwargs)
layers = [nn.LazyLinear(hidden_layer_size), activation()] * n_hidden_layers
layers = sum([[nn.LazyLinear(hidden_layer_size),
activation()] for _ in range(n_hidden_layers)],[])
self.nn = nn.Sequential(*layers, nn.LazyLinear(self.action_dim))
#self.nn = nn.Sequential(
# nn.LazyLinear(hidden_layer_size),
# nn.Tanh(),
# nn.LazyLinear(hidden_layer_size),
# nn.Tanh(),
# nn.LazyLinear(self.action_dim),
#)
def forward(self, states):
return self.nn(states)