huangfu/InteractionImitation
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

removing gail-trpo since performance is about the same as gail, adding experiment evaluation script, updating metric averaging to work

Browse Source
This commit is contained in:
Arec Jamgochian
2022-02-27 23:03:28 -08:00
parent e110570092
commit e37911eeff
5 changed files with 133 additions and 329 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
src/eval_main.py
View File

@@ -15,7 +15,7 @@ from src.options import envs as options_envs2
from src.safe_options.policy import SetMaskedDiscretePolicy
from src.safe_options import options as options_envs3
from src.util.wrappers import IntersimpleTimeLimit
import os
from typing import Optional, List, Dict, Tuple
import torch
import numpy as np
@@ -44,29 +44,29 @@ def load_policy(method:str,
policy = SetPolicy(env.action_space.shape[-1])
policy.load_state_dict(torch.load(policy_file, map_location=ml))
policy.eval()
elif method == 'gail':
elif method == 'gail-trpo':
policy = SetPolicy(env.action_space.shape[-1])
policy(torch.zeros(env.observation_space.shape))
policy = ReparamPolicy(policy)
policy.load_state_dict(torch.load(policy_file, map_location=ml))
policy.eval()
elif method == 'gail-ppo':
elif method == 'gail':
policy = SetPolicy(env.action_space.shape[-1])
policy.load_state_dict(torch.load(policy_file, map_location=ml))
policy.eval()
elif method == 'rail':
raise NotImplementedError
elif method == 'ogail':
elif method == 'hail-trpo':
policy = SetDiscretePolicy(env.action_space.n)
policy(torch.zeros(env.observation_space.shape))
policy = ReparamPolicy(policy)
policy.load_state_dict(torch.load(policy_file, map_location=ml))
policy.eval()
elif method == 'ogail-ppo':
elif method == 'hail':
policy = SetDiscretePolicy(env.action_space.n)
policy.load_state_dict(torch.load(policy_file, map_location=ml))
policy.eval()
elif method == 'sgail':
elif method == 'shail-trpo':
policy = SetMaskedDiscretePolicy(env.action_space.n)
policy(
torch.zeros(env.observation_space['observation'].shape),
@@ -75,7 +75,7 @@ def load_policy(method:str,
policy = ReparamSafePolicy(policy)
policy.load_state_dict(torch.load(policy_file, map_location=ml))
policy.eval()
elif method == 'sgail-ppo':
elif method == 'shail':
policy = SetMaskedDiscretePolicy(env.action_space.n)
policy.load_state_dict(torch.load(policy_file, map_location=ml))
policy.eval()
@@ -375,6 +375,9 @@ def eval_main(
policy_file (str): path to saved policy
env (str): environment class
method (str): method (expert, bc, gail, rail, hgail, hrail)
Returns:
outbase (str): string to outbase
"""
print(f'#############################################################################')
print(f'Evaluating {method} from file {policy_file} on {env} at locations {locations}')
@@ -383,9 +386,17 @@ def eval_main(
# set seed
np.random.seed(seed)
torch.manual_seed(seed)
pfilename = policy_file.split('/')[-1].split('.')[0]
locstr = 'loc_'+'_'.join([f'r{ro}t{tr}' for (ro,tr) in locations])
outbase = f'out/{method}/{locstr}/{pfilename}_seed{seed}'
if policy_file == '':
method_path = method
name_base = method
else:
path_items = policy_file.split('/')
name_base = path_items[-1].split('.')[0]
method_path = ('/').join(path_items[1:-1])
outfolder = os.path.join('out',method_path,locstr)
filebase = name_base + f'_tseed{seed}'
outbase = os.path.join(outfolder,filebase)
# load expert metrics
expert_metrics = generate_expert_metrics(locations)
@@ -404,6 +415,8 @@ def eval_main(
save_metrics(smetrics, outbase+'_summary.pkl')
cmetrics = comparison_metrics(policy_metrics, expert_metrics, outbase=outbase)
save_metrics(cmetrics, outbase+'_comparison.pkl')
return outbase
if __name__=='__main__':
import fire

31
src/evaluation/utils.py
View File

@@ -34,13 +34,20 @@ def load_metrics(filestr:str):
metrics = pickle.load(f)
return metrics
def average_metrics(metric_list:List[Dict[str,float]]):
def average_metrics(metric_list:List[Dict[str,float]], verbose:bool=True) ->Dict[str, tuple]:
"""
Average all the metrics in the list
Args:
metric_list (list of dicts): list of metric dicts which each map a string to a float
verbose (bool): whether to print avg metrics
Returns:
average_metrics (Dict[str, tuple])
"""
average_metrics = {}
if len(metric_list) == 0:
return average_metrics
keys = list(metric_list[0].keys())
N = len(metric_list)
master_dict = {key:[] for key in keys}
@@ -50,29 +57,39 @@ def average_metrics(metric_list:List[Dict[str,float]]):
master_dict[key] = np.array(master_dict[key])
mu = np.nanmean(master_dict[key])
std2 = np.nanstd(master_dict[key])*2
print(f'{key}: {mu} \pm {std2}')
if verbose:
print(f'{key}: {mu} \pm {std2}')
average_metrics[key] = (mu, std2)
return average_metrics
def load_and_average(path:str):
def load_and_average(path:str, verbose:bool=True):
"""
Load and average all metric files in a particular folder
Args:
path (str)
verbose (bool): whether to print avg metrics
Returns:
avg_metrics (Dict[str, tuple])
"""
assert os.path.isdir(path)
# summary metrics
summary_files = [os.path.join(path,f) for f in os.listdir(path) if f.endswith('summary.pkl')]
print(*summary_files, sep='\n')
if verbose:
print(*summary_files, sep='\n')
all_summary_metrics = [load_metrics(f) for f in summary_files]
average_metrics(all_summary_metrics)
avg_metrics = average_metrics(all_summary_metrics, verbose=verbose)
# comparison metrics
comp_files = [os.path.join(path,f) for f in os.listdir(path) if f.endswith('comparison.pkl')]
print(*comp_files, sep='\n')
if verbose:
print(*comp_files, sep='\n')
all_comp_metrics = [load_metrics(f) for f in comp_files]
average_metrics(all_comp_metrics)
comp_avg = average_metrics(all_comp_metrics, verbose=verbose)
avg_metrics.update(comp_avg)
return avg_metrics
if __name__=='__main__':
import fire