else:
from urllib.request import urlretrieve
urlretrieve(snapshot_url, snapshot_path)
load(resolver, snapshot_path)
epsilon.set_value(0)
##one-turn response function
prev_memory_state = T.vector("prev_mem_state_input", dtype=floatX)
observation = T.vector("observation_input", dtype=floatX)
prev_memory_tensor = prev_memory_state.reshape([1, -1])
observation_tensor = observation.reshape([1, -1])
new_state_tensor, Qvalues_tensor, action_tensor, _ = agent.get_agent_reaction(
prev_memory_tensor, observation_tensor)
new_state, Qvalues, action = new_state_tensor[0], Qvalues_tensor[
0], action_tensor[0]
get_agent_reaction = theano.function([prev_memory_state, observation],
[new_state, Qvalues, action])
##auxilary transformer function
def response_to_observation(response, prev_action, is_alive=True):
"""creates a wikicat-format observation from"""
observation = np.zeros(env.observation_size, dtype=floatX)
observation[0] = response
##one-turn response function
prev_memory_state = T.vector("prev_mem_state_input",dtype=floatX)
observation = T.vector("observation_input",dtype=floatX)
prev_memory_tensor = prev_memory_state.reshape([1,-1])
prev_gru1 = prev_memory_tensor[:,:n_hid_1]
prev_gru2 = prev_memory_tensor[:,n_hid_1:]
observation_tensor = observation.reshape([1,-1])
action_channels,_, outputs, = agent.get_agent_reaction(
{
gru1:prev_gru1,
gru2:prev_gru2
},observation_tensor,)
action_tensor = action_channels[0]
Qvalues_tensor = outputs[0]
new_state_tensor = outputs[1] #concatenated memory
new_state, Qvalues, action = new_state_tensor[0],Qvalues_tensor[0],action_tensor[0]
get_agent_reaction = theano.function([prev_memory_state,observation], [new_state,Qvalues,action])
epsilon.set_value(0)
##one-turn response function
prev_memory_state = T.vector("prev_mem_state_input",dtype=floatX)
observation = T.vector("observation_input",dtype=floatX)
prev_memory_tensor = prev_memory_state.reshape([1,-1])
observation_tensor = observation.reshape([1,-1])
new_state_tensor, Qvalues_tensor, action_tensor,_ = agent.get_agent_reaction(prev_memory_tensor,observation_tensor)
new_state, Qvalues, action = new_state_tensor[0],Qvalues_tensor[0],action_tensor[0]
get_agent_reaction = theano.function([prev_memory_state,observation], [new_state,Qvalues,action])
##auxilary transformer function
def response_to_observation(response,prev_action,is_alive=True):
"""creates a wikicat-format observation from"""
observation = np.zeros(env.observation_size,dtype=floatX)
epsilon.set_value(0)
##one-turn response function
prev_memory_state = T.vector("prev_mem_state_input", dtype=floatX)
observation = T.vector("observation_input", dtype=floatX)
prev_memory_tensor = prev_memory_state.reshape([1, -1])
prev_gru1 = prev_memory_tensor[:, :n_hid_1]
prev_gru2 = prev_memory_tensor[:, n_hid_1:]
observation_tensor = observation.reshape([1, -1])
action_channels, _, outputs, = agent.get_agent_reaction(
{
gru1: prev_gru1,
gru2: prev_gru2
},
observation_tensor,
)
action_tensor = action_channels[0]
Qvalues_tensor = outputs[0]
new_state_tensor = outputs[1] #concatenated memory
new_state, Qvalues, action = new_state_tensor[0], Qvalues_tensor[
0], action_tensor[0]
get_agent_reaction = theano.function([prev_memory_state, observation],
[new_state, Qvalues, action])