def train(output_file, params_file, cur_ind, model_path, config):
from phase2_utils import (
ReplayBuffer,
load_image_model,
)
# load pre-trained model
net = load_image_model(config['model_args']['backbone'],
model_path,
device=config['device'])
# load CMAES parameters, cur_id is the population member id i.e. [0, pop_size-1]
params = np.load(params_file)['params'][cur_ind]
# overwrite weights
di = net.state_dict()
st = 0
for i in range(4):
w_key = 'location_pred.{}.1.weight'.format(i)
b_key = 'location_pred.{}.1.bias'.format(i)
w_shape = di[w_key].shape
w_num = np.prod(w_shape)
b_shape = di[b_key].shape
b_num = np.prod(b_shape)
w = params[st:st + w_num]
b = params[st + w_num:st + w_num + b_num]
st += (w_num + b_num)
reshaped_w = np.reshape(w, w_shape)
di[w_key] = torch.from_numpy(reshaped_w)
di[b_key] = torch.from_numpy(b)
net.load_state_dict(di)
image_agent_kwargs = {'camera_args': config["agent_args"]['camera_args']}
ret = rollout(net,
image_agent_kwargs=image_agent_kwargs,
port=config['port'],
seed=config['seed'])
all_dists = []
total_fit = ret
all_dists.append(total_fit)
avg_dist = np.mean(all_dists)
print('Writing average distance..')
with open(output_file, 'w+') as w:
w.write(str(float(total_fit)))
with open('%s.rew' % output_file, 'w') as w:
for dist in all_dists:
w.write('%f\n' % dist)
def train(config):
import utils.bz_utils as bzu
bzu.log.init(config['log_dir'])
bzu.log.save_config(config)
teacher_config = bzu.log.load_config(config['teacher_args']['model_path'])
from phase2_utils import (CoordConverter, ReplayBuffer, LocationLoss,
load_birdview_model, load_image_model,
get_optimizer)
criterion = LocationLoss()
net = load_image_model(config['model_args']['backbone'],
config['phase1_ckpt'],
device=config['device'])
teacher_net = load_birdview_model(teacher_config['model_args']['backbone'],
config['teacher_args']['model_path'],
device=config['device'])
image_agent_kwargs = {'camera_args': config["agent_args"]['camera_args']}
coord_converter = CoordConverter(**config["agent_args"]['camera_args'])
replay_buffer = ReplayBuffer(**config["buffer_args"])
# optimizer = get_optimizer(net.parameters(), config["optimizer_args"]["lr"])
for episode in tqdm.tqdm(range(config['max_episode']), desc='Episode'):
rollout(replay_buffer,
coord_converter,
net,
teacher_net,
episode,
image_agent_kwargs=image_agent_kwargs,
port=config['port'])
# import pdb; pdb.set_trace()
_train(replay_buffer, net, teacher_net, criterion, coord_converter,
bzu.log, config, episode)
def train(result_file_path, valuation_file_path, params_file, gen, indiv_idx,
model_path, config, carla_process):
from phase2_utils import (
ReplayBuffer,
load_image_model,
)
# load pre-trained model
net = load_image_model(config['model_args']['backbone'],
model_path,
device=config['device'])
# load CMAES parameters, indiv_idx is the population member id i.e. [0, pop_size-1]
params = np.load(params_file)['params'][indiv_idx]
# overwrite weights
di = net.state_dict()
st = 0
for i in range(4):
w_key = 'location_pred.{}.1.weight'.format(i)
b_key = 'location_pred.{}.1.bias'.format(i)
w_shape = di[w_key].shape
w_num = np.prod(w_shape)
b_shape = di[b_key].shape
b_num = np.prod(b_shape)
w = params[st:st + w_num]
b = params[st + w_num:st + w_num + b_num]
st += (w_num + b_num)
reshaped_w = np.reshape(w, w_shape)
di[w_key] = torch.from_numpy(reshaped_w)
di[b_key] = torch.from_numpy(b)
net.load_state_dict(di)
image_agent_kwargs = {'camera_args': config["agent_args"]['camera_args']}
ret = rollout(net,
valuation_file_path,
indiv_idx,
episode_length=1000,
image_agent_kwargs=image_agent_kwargs,
port=config['port'],
env_seed=config['env_seed'])
all_dists = []
total_fit = ret
all_dists.append(total_fit)
avg_dist = np.mean(all_dists)
# write to results files by appending new line using format: generation individual_ID score \n
print('Writing to results files..')
with open(result_file_path, 'a+') as file:
file.write(
str(int(gen)) + ' ' + str(int(indiv_idx)) + ' ' +
str(float(total_fit)) + '\n')
file.close()
with open('%s.rew' % result_file_path, 'a+') as file:
for dist in all_dists:
file.write('%f %f %f\n' % (gen, indiv_idx, dist))
file.close()