def test_metrify(): import numpy as np d = np.array(10) assert metrify(d) == 10 d = np.array(10.0) assert metrify(d) == 10.0 d = np.array([10.0, 2]) assert metrify(d) == [10.0, 2] d = np.array([10.0, 2]) assert metrify(d) == [10.0, 2]
def train_maml(*, n_tasks: int, tasks: MetaRLTasks, maml: E_MAML): if not G.inner_alg.startswith("BC"): path_gen = path_gen_fn(env=tasks.envs, policy=maml.runner.policy, start_reset=G.reset_on_start) next(path_gen) meta_path_gen = path_gen_fn(env=tasks.envs, policy=maml.meta_runner.policy, start_reset=G.reset_on_start) next(meta_path_gen) if G.load_from_checkpoint: # todo: add variable to checkpoint # todo: set the epoch_ind starting point here. logger.load_variables(G.load_from_checkpoint) if G.meta_sgd: assert maml.alpha is not None, "Coding Mistake if meta_sgd is trueful but maml.alpha is None." max_episode_length = tasks.spec.max_episode_steps sess = tf.get_default_session() epoch_ind, prefix = G.epoch_init - 1, "" while epoch_ind < G.epoch_init + G.n_epochs: logger.flush() logger.split() is_bc_test = (prefix != "test/" and G.eval_interval and epoch_ind % G.eval_interval == 0) prefix = "test/" if is_bc_test else "" epoch_ind += 0 if is_bc_test else 1 if G.meta_sgd: alpha_lr = sess.run(maml.alpha) # only used in the runner. logger.log(metrics={f"alpha_{i}/{stem(t.name, 2)}": a for i, a_ in enumerate(alpha_lr) for t, a in zip(maml.runner.trainables, a_)}, silent=True) else: alpha_lr = G.alpha.send(epoch_ind) if isinstance(G.alpha, Schedule) else np.array(G.alpha) logger.log(alpha=metrify(alpha_lr), epoch=epoch_ind, silent=True) beta_lr = G.beta.send(epoch_ind) if isinstance(G.beta, Schedule) else np.array(G.beta) clip_range = G.clip_range.send(epoch_ind) if isinstance(G.clip_range, Schedule) else np.array(G.clip_range) logger.log(beta=metrify(beta_lr), clip_range=metrify(clip_range), epoch=epoch_ind, silent=True) batch_timesteps = G.batch_timesteps.send(epoch_ind) \ if isinstance(G.batch_timesteps, Schedule) else G.batch_timesteps # Compute updates for each task in the batch # 0. save value of variables # 1. sample # 2. gradient descent # 3. repeat step 1., 2. until all gradient steps are exhausted. batch_data = defaultdict(list) maml.save_weight_cache() load_ops = [] if DEBUG.no_weight_reset else [maml.cache.load] if G.checkpoint_interval and epoch_ind % G.checkpoint_interval == 0 \ and not is_bc_test and epoch_ind >= G.start_checkpoint_after_epoch: cp_path = f"checkpoints/variables_{epoch_ind:04d}.pkl" logger.log_line(f'saving checkpoint {cp_path}') # note: of course I don't know that are all of the trainables at the moment. logger.save_variables(tf.trainable_variables(), path=cp_path) feed_dict = {} for task_ind in range(n_tasks if is_bc_test else G.n_tasks): graph_branch = maml.graphs[0] if G.n_graphs == 1 else maml.graphs[task_ind] if G.n_graphs == 1: gradient_sum_op = maml.gradient_sum.set_op if task_ind == 0 else maml.gradient_sum.add_op print(f"task_ind {task_ind}...") if not DEBUG.no_task_resample: if not is_bc_test: print(f'L250: sampling task') tasks.sample() elif task_ind < n_tasks: task_spec = dict(index=task_ind % n_tasks) print(f'L254: sampling task {task_spec}') tasks.sample(**task_spec) else: raise RuntimeError('should never hit here.') for k in range(G.n_grad_steps + 1): # 0 - 10 <== last one being the maml policy. _is_new = False # for imitation inner loss, we still sample trajectory for evaluation purposes, but # replace it with the demonstration data for learning if k < G.n_grad_steps: if G.inner_alg.startswith("BC"): p = p if G.single_sampling and k > 0 else \ bc.sample_demonstration_data(tasks.task_spec, key=("eval" if is_bc_test else None)) else: p, _is_new = path_gen.send(batch_timesteps), True elif k == G.n_grad_steps: if G.meta_alg.startswith("BC"): # note: use meta bc samples. p = bc.sample_demonstration_data(tasks.task_spec, key="meta") else: p, _is_new = meta_path_gen.send(batch_timesteps), True else: raise Exception('Implementation error. Should never reach this line.') if k in G.eval_grad_steps: _ = path_gen if k < G.n_grad_steps else meta_path_gen p_eval = p if _is_new else _.send(G.eval_timesteps) # reporting on new trajectory samples avg_r = p_eval['ep_info']['reward'] if G.normalize_env else np.mean(p_eval['rewards']) episode_r = avg_r * max_episode_length # default horizon for HalfCheetah if episode_r < G.term_reward_threshold: # todo: make this batch-based instead of on single episode logger.log_line("episode reward is too low: ", episode_r, "terminating training.", flush=True) raise RuntimeError('AVERAGE REWARD TOO LOW. Terminating the experiment.') batch_data[prefix + f"grad_{k}_step_reward"].append(avg_r if Reporting.report_mean else episode_r) if k in G.eval_grad_steps: logger.log_key_value(prefix + f"task_{task_ind}_grad_{k}_reward", episode_r, silent=True) _p = {k: v for k, v in p.items() if k != "ep_info"} if k < G.n_grad_steps: # note: under meta-SGD mode, the runner needs the k^th learning rate. _lr = alpha_lr[k] if G.meta_sgd else alpha_lr # clip_range is not used in BC mode. but still passed in. runner_feed_dict = \ path_to_feed_dict(inputs=maml.runner.inputs, paths=_p, lr=_lr, baseline=G.baseline, gamma=G.gamma, use_gae=G.use_gae, lam=G.lam, horizon=max_episode_length, clip_range=clip_range) # todo: optimize `maml.meta_runner` if k >= G.n_grad_steps. loss, *_, __ = maml.runner.optim.run_optimize(feed_dict=runner_feed_dict) runner_feed_dict.clear() for key, value in zip(maml.runner.model.reports.keys(), [loss, *_]): batch_data[prefix + f"grad_{k}_step_{key}"].append(value) logger.log_key_value(prefix + f"task_{task_ind}_grad_{k}_{key}", value, silent=True) if loss > G.term_loss_threshold: # todo: make this batch-based instead of on single episode logger.log_line(prefix + "episode loss blew up:", loss, "terminating training.", flush=True) raise RuntimeError('loss is TOO HIGH. Terminating the experiment.') # done: has bug when using fixed learning rate. Needs the learning rate as input. feed_dict.update( # do NOT pass in the learning rate because the graph already includes those. path_to_feed_dict(inputs=graph_branch.workers[k].inputs, paths=_p, lr=None if G.meta_sgd else alpha_lr, # but do with fixed alpha horizon=max_episode_length, baseline=G.baseline, gamma=G.gamma, use_gae=G.use_gae, lam=G.lam, clip_range=clip_range)) elif k == G.n_grad_steps: yield_keys = dict( movie=epoch_ind >= G.start_movie_after_epoch and epoch_ind % G.record_movie_interval == 0, eval=is_bc_test ) if np.fromiter(yield_keys.values(), bool).any(): yield yield_keys, epoch_ind, tasks.task_spec if is_bc_test: if load_ops: # we need to reset the weights. Otherwise the world would be on fire. tf.get_default_session().run(load_ops) continue # do NOT meta learn from test samples. # we don't treat the meta_input the same way even though we could. This is more clear to read. # note: feed in the learning rate only later. feed_dict.update( # do NOT need learning rate path_to_feed_dict(inputs=graph_branch.meta.inputs, paths=_p, horizon=max_episode_length, baseline=G.baseline, gamma=G.gamma, use_gae=G.use_gae, lam=G.lam, clip_range=clip_range)) if G.n_graphs == 1: # load from checkpoint before computing the meta gradient\nrun gradient sum operation if load_ops: tf.get_default_session().run(load_ops) # note: meta reporting should be run here. Not supported for simplicity. (need to reduce across # note: tasks, and can not be done outside individual task graphs. if G.meta_sgd is None: # note: copied from train_supervised_maml, not tested feed_dict[maml.alpha] = alpha_lr tf.get_default_session().run(gradient_sum_op, feed_dict) feed_dict.clear() if load_ops: tf.get_default_session().run(load_ops) if is_bc_test: continue # do NOT meta learn from test samples. # note: copied from train_supervised_maml, not tested if G.meta_sgd is None: feed_dict[maml.alpha] = alpha_lr if G.n_graphs == 1: assert G.meta_n_grad_steps == 1, "ERROR: Can only run 1 meta gradient step with a single graph." # note: remove meta reporting b/c meta report should be in each task in this case. tf.get_default_session().run(maml.meta_update_ops[0], {maml.beta: beta_lr}) else: assert feed_dict, "ERROR: It is likely that you jumped here from L:178." feed_dict[maml.beta] = beta_lr for i in range(G.meta_n_grad_steps): update_op = maml.meta_update_ops[0 if G.reuse_meta_optimizer else i] *reports, _ = tf.get_default_session().run(maml.meta_reporting + [update_op], feed_dict) if i not in (0, G.meta_n_grad_steps - 1): continue for key, v in zip(maml.meta_reporting_keys, reports): logger.log_key_value(prefix + f"grad_{G.n_grad_steps + i}_step_{key}", v, silent=True) feed_dict.clear() tf.get_default_session().run(maml.cache.save) # Now compute the meta gradients. # note: runner shares variables with the MAML graph. Reload from state_dict # note: if max_grad_step is the same as n_grad_steps then no need here. dt = logger.split() logger.log_line('Timer Starts...' if dt is None else f'{dt:0.2f} sec/epoch') logger.log(dt_epoch=dt or np.nan, epoch=epoch_ind) for key, arr in batch_data.items(): reduced = np.array(arr).mean() logger.log_key_value(key, reduced) logger.flush()
def train_supervised_maml(*, k_tasks=1, maml: E_MAML): # env used for evaluation purposes only. if G.meta_sgd: assert maml.alpha is not None, "Coding Mistake if meta_sgd is trueful but maml.alpha is None." assert G.n_tasks >= k_tasks, f"Is this intended? You probably want to have " \ f"meta-batch({G.n_tasks}) >= k_tasks({k_tasks})." sess = tf.get_default_session() epoch_ind, pref = -1, "" while epoch_ind < G.n_epochs: # for epoch_ind in range(G.n_epochs + 1): logger.flush() logger.split() is_bc_test = (pref != "test/" and G.eval_interval and epoch_ind % G.eval_interval == 0) pref = "test/" if is_bc_test else "" epoch_ind += 0 if is_bc_test else 1 if G.meta_sgd: alpha_lr = sess.run(maml.alpha) # only used in the runner. logger.log(metrics={f"alpha_{i}/{stem(t.name, 2)}": a for i, a_ in enumerate(alpha_lr) for t, a in zip(maml.runner.trainables, a_)}, silent=True) else: alpha_lr = G.alpha.send(epoch_ind) if isinstance(G.alpha, Schedule) else np.array(G.alpha) logger.log(alpha=metrify(alpha_lr), epoch=epoch_ind, silent=True) beta_lr = G.beta.send(epoch_ind) if isinstance(G.beta, Schedule) else np.array(G.beta) logger.log(beta=metrify(beta_lr), epoch=epoch_ind, silent=True) if G.checkpoint_interval and epoch_ind % G.checkpoint_interval == 0: yield "pre-update-checkpoint", epoch_ind # Compute updates for each task in the batch # 0. save value of variables # 1. sample # 2. gradient descent # 3. repeat step 1., 2. until all gradient steps are exhausted. batch_data = defaultdict(list) maml.save_weight_cache() load_ops = [] if DEBUG.no_weight_reset else [maml.cache.load] feed_dict = {} for task_ind in range(k_tasks if is_bc_test else G.n_tasks): graph_branch = maml.graphs[0] if G.n_graphs == 1 else maml.graphs[task_ind] if G.n_graphs == 1: gradient_sum_op = maml.gradient_sum.set_op if task_ind == 0 else maml.gradient_sum.add_op """ In BC mode, we don't have an environment. The sampling is handled here then fed to the sampler. > task_spec = dict(index=0) Here we make the testing more efficient. """ if not DEBUG.no_task_resample: if not is_bc_test: task_spec = dict(index=np.random.randint(0, k_tasks)) elif task_ind < k_tasks: task_spec = dict(index=task_ind % k_tasks) else: raise RuntimeError('should never hit here.') for k in range(G.n_grad_steps + 1): # 0 - 10 <== last one being the maml policy. # for imitation inner loss, we still sample trajectory for evaluation purposes, but # replace it with the demonstration data for learning if k < G.n_grad_steps: p = p if G.single_sampling and k > 0 else \ bc.sample_demonstration_data(task_spec, key=("eval" if is_bc_test else None)) elif k == G.n_grad_steps: # note: use meta bc samples. p = bc.sample_demonstration_data(task_spec, key="meta") else: raise Exception('Implementation error. Should never reach this line.') _p = {k: v for k, v in p.items() if k != "ep_info"} if k < G.n_grad_steps: # note: under meta-SGD mode, the runner needs the k^th learning rate. _lr = alpha_lr[k] if G.meta_sgd else alpha_lr runner_feed_dict = \ path_to_feed_dict(inputs=maml.runner.inputs, paths=_p, lr=_lr) # todo: optimize `maml.meta_runner` if k >= G.n_grad_steps. loss, *_, __ = maml.runner.optim.run_optimize(feed_dict=runner_feed_dict) runner_feed_dict.clear() for key, value in zip(maml.runner.model.reports.keys(), [loss, *_]): batch_data[pref + f"grad_{k}_step_{key}"].append(value) logger.log_key_value(pref + f"task_{task_ind}_grad_{k}_{key}", value, silent=True) if loss > G.term_loss_threshold: # todo: make this batch-based instead of on single episode err = pref + "episode loss blew up:", loss, "terminating training." logger.log_line(colored(err, "red"), flush=True) raise RuntimeError('loss is TOO HIGH. Terminating the experiment.') # fixit: has bug when using fixed learning rate. Still needs to get learning rate from placeholder feed_dict.update(path_to_feed_dict(inputs=graph_branch.workers[k].inputs, paths=_p)) elif k == G.n_grad_steps: yield_keys = dict( movie=G.record_movie_interval and epoch_ind >= G.start_movie_after_epoch and epoch_ind % G.record_movie_interval == 0, eval=is_bc_test ) if np.fromiter(yield_keys.values(), bool).any(): yield yield_keys, epoch_ind, task_spec if is_bc_test: if load_ops: tf.get_default_session().run(load_ops) continue # do NOT meta learn from test samples. # we don't treat the meta_input the same way even though we could. This is more clear to read. # note: feed in the learning rate only later. feed_dict.update(path_to_feed_dict(inputs=graph_branch.meta.inputs, paths=_p)) if G.n_graphs == 1: # load from checkpoint before computing the meta gradient\nrun gradient sum operation if load_ops: tf.get_default_session().run(load_ops) # note: meta reporting should be run here. Not supported for simplicity. (need to reduce across # note: tasks, and can not be done outside individual task graphs. if G.meta_sgd is None: feed_dict[maml.alpha] = alpha_lr tf.get_default_session().run(gradient_sum_op, feed_dict) feed_dict.clear() if load_ops: tf.get_default_session().run(load_ops) if is_bc_test: continue # do NOT meta learn from test samples. if G.meta_sgd is None: feed_dict[maml.alpha] = alpha_lr if G.n_graphs == 1: assert G.meta_n_grad_steps == 1, "ERROR: Can only run 1 meta gradient step with a single graph." # note: remove meta reporting b/c meta report should be in each task in this case. tf.get_default_session().run(maml.meta_update_ops[0], {maml.beta: beta_lr}) else: assert feed_dict, "ERROR: It is likely that you jumped here from L:178." feed_dict[maml.beta] = beta_lr for i in range(G.meta_n_grad_steps): update_op = maml.meta_update_ops[0 if G.reuse_meta_optimizer else i] *reports, _ = tf.get_default_session().run(maml.meta_reporting + [update_op], feed_dict) if i not in (0, G.meta_n_grad_steps - 1): continue for key, v in zip(maml.meta_reporting_keys, reports): logger.log_key_value(pref + f"grad_{G.n_grad_steps + i}_step_{key}", v, silent=True) feed_dict.clear() tf.get_default_session().run(maml.cache.save) # Now compute the meta gradients. # note: runner shares variables with the MAML graph. Reload from state_dict # note: if max_grad_step is the same as n_grad_steps then no need here. dt = logger.split() logger.log_line('Timer Starts...' if dt is None else f'{dt:0.2f} sec/epoch') logger.log(dt_epoch=dt or np.nan, epoch=epoch_ind) for key, arr in batch_data.items(): reduced = np.array(arr).mean() logger.log_key_value(key, reduced)