plt.close()
rep_types = {
'unstructured': onehot,
'random': random,
'place_cell': place_cell,
'structured': sr,
'conv': convs
}
if rep == 'conv':
# saved weights
saved_network = torch.load(
'../../Data/agents/6a956906-c06c-47ef-aad1-3593fb9068d1.pt')
# load agent weights into new network
network = nets.shallow_ActorCritic(input_dims=600,
hidden_dims=400,
output_dims=4,
lr=5e-4)
new_state_dict = {}
for key in saved_network.keys():
if key[0:6] == 'output':
if key[7] == '0':
new_key = 'pol' + key[8:]
new_state_dict[new_key] = saved_network[key]
elif key[7] == '1':
new_key = 'val' + key[8:]
new_state_dict[new_key] = saved_network[key]
elif key[0:8] == 'hidden.5':
new_key = 'hidden' + key[8:]
new_state_dict[new_key] = saved_network[key]
network.load_state_dict(new_state_dict)
# set up parameters
env_name = 'gridworld:gridworld-v1' #v1 novel #v11 moved reward
network_id = None # '97b5f281-a60e-4738-895d-191a04edddd6'
actor = 'EC'
ntrials = 5000
# create environment
env = gym.make(env_name)
plt.close()
# generate network
if network_id == None:
# generate parameters for network from environment observation shape
params = nets.fc_params(env)
network = nets.ActorCritic(params)
else:
network = torch.load(f'./Data/agents/load_agents/{network_id}.pt')
memory = Memory.EpisodicMemory(cache_limit=400, entry_size=env.action_space.n)
agent = Agent(network, memory=memory)
if actor == 'MF':
agent.get_action = agent.MF_action
elif actor == 'EC':
agent.get_action = agent.EC_action
run = expt(agent, env)
run.run(NUM_TRIALS=ntrials, NUM_EVENTS=250)
directory = '../../Data/' # ../../Data if you are in Tests/CH2
env_name = f'gridworld:gridworld-v{version}'
# make gym environment
env = gym.make(env_name)
plt.close()
print(env.rewards)
num_trials = 10
num_events = 250
# load in trained network
net, state_dict = load_network(version, rep_type, directory)
# load weights to head_ac network from previously learned agent
AC_head_agent = nets.flat_ActorCritic(400,
output_dims=env.action_space.n,
lr=learning_rate)
top_layer_dict = {}
top_layer_dict['pol.weight'] = state_dict['output.0.weight']
top_layer_dict['pol.bias'] = state_dict['output.0.bias']
top_layer_dict['val.weight'] = state_dict['output.1.weight']
top_layer_dict['val.bias'] = state_dict['output.1.bias']
AC_head_agent.load_state_dict(top_layer_dict)
# get state inputs
h0, h1 = get_net_activity(env_name, rep_type, net)
state_reps, representation_name, = h1, f'h1_{rep_type}_latents'
memory = None #Memory.EpisodicMemory(cache_limit=cache_size_for_env, entry_size=env.action_space.n)
agent = Agent(AC_head_agent, memory=memory, state_representations=state_reps)
plt.close()
cache_size_for_env = int(len(env.useable) * (cache_size / 100))
print(env.rewards)
rep_types = {
'onehot': onehot,
'random': random,
'place_cell': place_cell,
'sr': sr,
'latent': latents
}
state_reps, representation_name, input_dims, _ = rep_types[rep_type](env)
# load weights to head_ac network from previously learned agent
AC_head_agent = nets.fc_ActorCritic([input_dims],
fc1_dims=200,
fc2_dims=200,
output_dims=env.action_space.n,
lr=learning_rate)
memory = None #Memory.EpisodicMemory(cache_limit=cache_size_for_env, entry_size=env.action_space.n)
agent = Agent(AC_head_agent, memory=memory, state_representations=state_reps)
run = expt(agent, env)
run.run(NUM_TRIALS=num_trials, NUM_EVENTS=num_events)
run.record_log(env_name,
representation_name,
num_trials,
num_events,
dir=directory,
file=write_to_file)
env = gym.make(env_name)
plt.close()
cache_size_for_env = int(len(env.useable) * (cache_size / 100))
print(env.rewards)
rep_types = {
'onehot': onehot,
'random': random,
'place_cell': place_cell,
'sr': sr,
'latent': latents
}
state_reps, representation_name, input_dims, _ = rep_types[rep_type](env)
# load weights to head_ac network from previously learned agent
AC_head_agent = nets.shallow_ActorCritic(input_dims,
hidden_dims=200,
output_dims=env.action_space.n,
lr=learning_rate)
memory = None #Memory.EpisodicMemory(cache_limit=cache_size_for_env, entry_size=env.action_space.n)
agent = Agent(AC_head_agent, memory=memory, state_representations=state_reps)
#from torchsummary import summary
#print(summary(AC_head_agent, (2,20,20)))
run = expt(agent, env)
run.run(NUM_TRIALS=num_trials, NUM_EVENTS=num_events)
run.record_log(env_name,
representation_name,
num_trials,
num_events,
dir=directory,
env = gym.make(env_name)
plt.close()
cache_size_for_env = int(len(env.useable) * (cache_size / 100))
print(env.rewards)
rep_types = {
'onehot': onehot,
'random': random,
'place_cell': place_cell,
'sr': sr,
'latent': latents
}
state_reps, representation_name, input_dims, _ = rep_types[rep_type](env)
# load weights to head_ac network from previously learned agent
AC_head_agent = nets.flat_ActorCritic(input_dims,
env.action_space.n,
lr=learning_rate)
if load_from != None:
AC_head_agent.load_state_dict(
torch.load(directory + f'agents/{load_from}.pt'))
print(f"weights loaded from {load_from}")
memory = Memory.EpisodicMemory(cache_limit=cache_size_for_env,
entry_size=env.action_space.n)
agent = Agent(AC_head_agent, memory=memory, state_representations=state_reps)
run = expt(agent, env)
run.run(NUM_TRIALS=num_trials, NUM_EVENTS=num_events)
run.record_log(env_name,
num_trials = 1000
num_events = 250
# make gym environment
env = gym.make(env_name)
plt.close()
cache_size_for_env = int(len(env.useable)*(cache_size/100))
print(env.rewards)
rep_types = {'onehot':onehot, 'random':random, 'place_cell':place_cell, 'sr':sr, 'latent':latents}
state_reps, representation_name, input_dims, _ = rep_types[rep_type](env)
# load weights to head_ac network from previously learned agent
AC_head_agent = nets.shallow_ActorCritic(input_dims, 200, env.action_space.n, lr=learning_rate)
if load_from != None:
AC_head_agent.load_state_dict(torch.load(directory+f'agents/{load_from}.pt'))
print(f"weights loaded from {load_from}")
memory = None#Memory.EpisodicMemory(cache_limit=cache_size_for_env, entry_size=env.action_space.n)
agent = Agent(AC_head_agent, memory=memory, state_representations=state_reps)
run = expt(agent, env)
run.run(NUM_TRIALS=num_trials, NUM_EVENTS=num_events)
print([(x, len(run.data[x])) for x in run.data.keys()])
run.record_log(env_name, representation_name,num_trials,num_events,dir=directory, file=write_to_file, load_from=load_from, extra=['randomwalk'])
