def generate_values_old(maze, model,**kwargs):
value_map = maze.empty_map
EC = kwargs.get('EC', None)
pcs = kwargs.get('pcs', None)
if EC is not None:
EC_pol_map = maze.make_map(maze.grid, pol=True)
MF_pol_map = maze.make_map(maze.grid, pol=True)
for loc in maze.useable:
if model.input_type == 'vector':
state = Variable(torch.FloatTensor(pcs.activity(loc)))
policy, value = sample_select_action(model,state)[1:3]
elif model.input_type == 'frame':
state = Variable(torch.FloatTensor(sg.get_frame(maze,agtlocation=loc)))
if isinstance (model.hidden[-1], nn.Linear):
policy, value, lin_act = sample_select_action(model,state, getlin=True)[1:4]
else:
policy, value = sample_select_action(model,state)[1:3]
value_map[loc[1]][loc[0]] = value
if EC is not None:
if model.input_type == 'vector':
EC_pol = EC.recall_mem(tuple(state.data[0]))
elif model.input_type == 'frame':
EC_pol = EC.recall_mem(tuple(lin_act.view(-1)))
EC_pol_map[loc[1]][loc[0]] = tuple(EC_pol.data[0])
MF_pol_map[loc[1]][loc[0]] = tuple(policy)
if EC == None:
return value_map
else:
return EC_pol_map, MF_pol_map
def generate_values(maze, model): value_map = maze.empty_map for loc in maze.useable: state = Variable(torch.FloatTensor(sg.get_frame(maze,agtlocation=loc))) policy, value = sample_select_action(model,state)[1:3] value_map[loc[1]][loc[0]] = value return value_map