Since the structure of each algorithms in baselines is different, we implementation the sample/load function in different ways.

baselines_sample.py

# code locate in baselines/gail

def sample(algo, load_model_path, policy_fn):
  
  assert algo in ['trpo', 'ppo', 'acktr', 'ddpg', 'a2c']
  if algo in ['trpo', 'ppo']:
    with tf.Session() as sess:
      # manually build graph 
      policy = policy_fn()
      # load model
      U.load_state(load_model_path)
    
  elif algo in ['acktr', 'ddpg', 'a2c']:
    policy = Model(policy_fn) # sess/graph declare inside
    policy.load(load_model_path)
    
  # sample expert 
  Sampler(algo, policy, sample_steps)
  
def Sampler(algo, policy, sample_steps):
  
  # start sampling code
  for _ in range(sample_steps):
    if algo == 'ddpg':
      actions, q = act(obs)
    elif algo == 'a2c' or 'acktr':
      actions, values, policy.states = policy.step(obs, policy.states, dones)
    elif algo == 'ppo' or 'trpo:
      actions, v = policy.act(stochastic, obs)
      
    obs, rw, dones, _ = env.step(act)
    

how about this?
Since the code in ppo, trpouse act() like this

def sample(algo, load_model_path, policy_fn):
  if algo in ['trpo', 'ppo']:
      U.make_session(num_cpu=1).__enter__()
      # manually build graph 
      policy = policy_fn()
      # load model
      U.load_state(load_model_path)