torch.manual_seed(SEED)
job_name_ewc_seed = job_name_ewc + '/seed_{}'.format(i)
e = {}
task_order = np.random.permutation(num_tasks)
for task_id in range(num_tasks):
e[task_id] = e_unshuffled[task_order[task_id]]
for ewc_lambda in lambda_range:
baseline_ewc = {}
for task_id in range(num_tasks):
baseline_ewc[task_id] = MLPBaseline(e[task_id].spec, reg_coef=1e-3, batch_size=64, epochs=2, learn_rate=1e-3, use_gpu=True)
policy_ewc = LinearPolicy(e[0].spec, seed=SEED)
agent_ewc = NPGEWC(e, policy_ewc, baseline_ewc, ewc_lambda=ewc_lambda, scaled_lambda=False, normalized_step_size=0.1, seed=SEED, save_logs=True)
# agent = BatchREINFORCE(e, policy, baseline, learn_rate=0.0001, seed=SEED, save_logs=True)
job_name_ewc_seed_lambda = job_name_ewc_seed + '/lambda{}'.format(ewc_lambda)
for task_id in range(num_tasks):
ts = timer.time()
train_agent(job_name=job_name_ewc_seed_lambda,
agent=agent_ewc,
seed=SEED,
niter=50,
gamma=0.995,
gae_lambda=0.97,
num_cpu=num_cpu,
sample_mode='trajectories',
num_traj=50,
job_name_ewc_seed = job_name_ewc + '/seed_{}'.format(i)
f = open(
job_name_ewc_seed + '/iterations/task_{}/'.format(t) +
'policy_0.pickle', 'rb')
policy_mtl = pickle.load(f)
f.close()
f = open(
job_name_ewc_seed + '/iterations/task_{}/'.format(t) +
'baseline_0.pickle', 'rb')
baseline_mtl[t] = pickle.load(f)
f.close()
agent_mtl = NPGEWC(e,
policy_mtl,
baseline_mtl,
ewc_lambda=1e-4,
scaled_lambda=False,
normalized_step_size=0.01,
seed=SEED,
save_logs=True)
mean_test_perf = agent_mtl.test_tasks(test_rollouts=10,
num_cpu=num_cpu,
task_ids=np.array([t]))
forward_transfer_results = {
**forward_transfer_results,
**mean_test_perf
}
result_file = open(job_name_ewc_seed + '/start_results.txt', 'w')
result_file.write(str(forward_transfer_results))