/
evaluate.py
90 lines (81 loc) · 3.51 KB
/
evaluate.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import argparse
import pickle
import torch
import gym
import envs, gym_fetch_stack
import numpy as np
import wandb
# pre process the inputs
def preproc_inputs(grip, obj, g, x_norm):
# concatenate the stuffs
outs = x_norm.normalize(grip, obj, g)
return [torch.tensor(x, dtype=torch.float32) for x in outs]
@torch.no_grad()
def main(args):
api = wandb.Api()
if args.run_tag:
assert len(args.run_path) == 1
runs = api.runs(path=args.run_path[0], filters={"tags": {"$in": [args.run_tag]}})
else:
runs = []
for run_path in args.run_path:
runs.append(api.run(run_path))
# load the model param
for run in runs:
if run._state != "finished":
print("Run not finished, skipping.")
continue
try:
run.file('models/best.pt').download(root='/tmp', replace=True)
x_norm, actor_network = torch.load('/tmp/models/best.pt', map_location=lambda storage, loc: storage)
except:
run.file('models/latest.pt').download(root='/tmp', replace=True)
x_norm, actor_network = torch.load('/tmp/models/latest.pt', map_location=lambda storage, loc: storage)
actor_network.eval()
print(f"Evaluating {run.config['actor']['_target_']} in {args.env_name}")
for env_name in args.env_name:
if not args.overwrite and f"eval_return/{env_name}" in run.summary:
print(f"Run {run.id} already has eval for {env_name}.")
continue
env = gym.make(env_name)
results, rets, timesteps = [], [], []
# start to do the demo
for ep_num in range(args.num_eps):
done, ret, t = False, 0, 0
observation = env.reset()
while not done:
grip, obj = observation['gripper_arr'][None], observation['object_arr'][None]
g = observation['desired_goal_arr'][None]
inputs = preproc_inputs(grip, obj, g, x_norm)
with torch.no_grad():
pi = actor_network(*inputs)
action = pi.detach().numpy().squeeze()
# put actions into the environment
observation_new, reward, done, info = env.step(action)
observation = observation_new
ret += reward
done = done or info['is_success']
t += 1
results.append(info['is_success'])
rets.append(ret)
timesteps.append(t)
print(f"{env_name} avg SR ({len(results)} eps): {np.mean(results):0.4f}.")
run.summary[f"eval_success/{env_name}"] = np.mean(results)
run.summary[f"eval_return/{env_name}"] = np.mean(rets)
run.summary[f"eval_length/{env_name}"] = np.mean(timesteps)
run.summary.update()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--env_name', type=str, nargs='+')
parser.add_argument('--run_tag', type=str)
parser.add_argument('--run_path', type=str, nargs='+')
parser.add_argument('--num_eps', type=int, default=200)
parser.add_argument('--render', action='store_true')
parser.add_argument('--overwrite', action='store_true')
args = parser.parse_args()
main(args)