-
Notifications
You must be signed in to change notification settings - Fork 48
/
Copy pathqrdqn.py
174 lines (148 loc) · 6.19 KB
/
qrdqn.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
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
import os, sys
import matplotlib.pyplot as plt
import numpy as np
from scipy.stats import zscore, mode
import torch
from torch import nn
from rastermap import Rastermap , utils
from rl_zoo3.utils import ALGOS
from stable_baselines3.common.preprocessing import preprocess_obs
from stable_baselines3.common.env_util import make_atari_env
from stable_baselines3.common.vec_env import VecFrameStack
# setup model with learning_rate OFF
custom_objects = {
"learning_rate": 0.0,
"lr_schedule": lambda _: 0.0,
"clip_range": lambda _: 0.0,
"optimize_memory_usage": False
}
vis_cnn = {}
def hook_fn_cnn(m, i, o):
vis_cnn[m] = o.detach().clone().cpu()
vis_mlp = {}
def hook_fn_mlp(m, i, o):
vis_mlp[m] = o.detach().clone().cpu()
def get_all_layers(model, hook_fn):
for name, layer in model._modules.items():
if isinstance(layer, nn.Sequential):
get_all_layers(layer, hook_fn)
else:
layer.register_forward_hook(hook_fn)
def get_all_activations_qrdqn(cnn_net, mlp_net, obs_torch):
ilayer = np.zeros((0,), "int")
cnn = np.zeros((0,), np.float32)
mlp = np.zeros((0,), np.float32)
out = cnn_net(obs_torch)
out2 = mlp_net(out)
i = 0
for layer_name in vis_cnn.keys():
if "ReLU()" in str(layer_name):
act = vis_cnn[layer_name].flatten().detach().numpy()
cnn = np.concatenate((cnn, act), axis=0)
ilayer = np.concatenate((ilayer, i*np.ones(len(act), "int")), axis=0)
i += 1
mlp = out2.cpu().flatten().detach().numpy()
ilayer = np.concatenate((ilayer, i*np.ones(len(mlp), "int")), axis=0)
return cnn, mlp, ilayer
def run_qrdqn(model_folder, root, env_id, n_seeds=10, device=torch.device("cuda")):
algo = "qrdqn"
log_path = os.path.join(model_folder, algo, env_id+"_1")
model_path = os.path.join(log_path, f"{env_id}.zip")
print(f"using Atari env {env_id}")
env = make_atari_env(env_id, n_envs=1, seed=0)
# Frame-stacking with 4 frames
env = VecFrameStack(env, n_stack=4)
model = ALGOS[algo].load(model_path, env=env, custom_objects=custom_objects)
cnn_net = model.policy.quantile_net.features_extractor
print(cnn_net)
mlp_net = model.policy.quantile_net.quantile_net
print(mlp_net)
get_all_layers(cnn_net, hook_fn_cnn)
spks, eps_len, actions = [], [], []
for seed in range(n_seeds):
print(f"seed {seed}")
env = make_atari_env(env_id, n_envs=1, seed=seed)
# Frame-stacking with 4 frames
env = VecFrameStack(env, n_stack=4)
obs = env.reset()
deterministic = False
state = None
episode_reward = 0.0
ep_len = 0
n_iterations = 4000
# don't show game
render = False
obs_all, actions_all, states_all = [], [], []
iteration = 0
for _ in range(n_iterations):
action, state = model.predict(obs, state=state, deterministic=deterministic)
obs_torch = torch.from_numpy(obs.transpose(0,3,1,2).astype(np.float32).copy()).to(device)
preprocessed_obs = preprocess_obs(obs_torch,
model.policy.observation_space,
normalize_images=model.policy.normalize_images)
obs_all.append(env.get_images()[0])
#actions = torch.from_numpy(action[np.newaxis,:]).to(device)
#out = cnn_net(preprocessed_obs)
cnn, mlp, ilayer = get_all_activations_qrdqn(cnn_net, mlp_net, preprocessed_obs)
if iteration==0:
activations = np.zeros((len(cnn)+len(mlp), n_iterations), "float32")
print(activations.shape, ilayer.max()+1)
activations[:len(cnn), iteration] = cnn
activations[len(cnn):, iteration] = mlp
obs, reward, done, infos = env.step(action)
if render:
env.render("human")
episode_reward += reward[0]
ep_len += 1
iteration += 1
actions_all.append(action)
states_all.append(state)
if infos is not None:
episode_infos = infos[0].get("episode")
if episode_infos is not None:
print(f"Atari Episode Score: {episode_infos['r']:.2f}")
print("Atari Episode Length", episode_infos["l"])
break;
activations = activations[:, 4:iteration]
env.close()
spks.append(activations)
actions.append(np.array(actions_all))
eps_len.append(activations.shape[1])
obs = np.stack(tuple(obs_all), axis=0).squeeze()[4:]
spks = np.concatenate(tuple(spks), axis=1)
actions = np.concatenate(tuple(actions), axis=0)
eps_len = np.array(eps_len)
np.savez(os.path.join(root, "simulations/", f"qrdqn_{env_id}.npz"),
spks=spks, ilayer=ilayer, obs=obs, actions=actions,
eps_len=eps_len)
def sort_spks(root, env_id):
dat = np.load(os.path.join(root, "simulations/", f"qrdqn_{env_id}.npz"))
spks = dat["spks"]
obs = dat["obs"]
ilayer = dat["ilayer"]
eps_len = dat["eps_len"]
x_std = spks.std(axis=1)
igood = x_std > 1e-3
print(igood.mean())
S = zscore(spks[igood], axis=1)
rm_model = Rastermap(time_lag_window=10, locality=0.75).fit(S[:,:])
isort = rm_model.isort
# show last episode
bin_size = 50
X_embedding = zscore(utils.bin1d(S[:,-eps_len[-1]:][isort], bin_size, axis=0), axis=1)
#if env_id=="EnduroNoFrameskip-v4":
# X_embedding = X_embedding[:,780:]
# obs = obs[780:]
nn, nt = X_embedding.shape
if env_id=="EnduroNoFrameskip-v4":
nt = nt-900
iframes = np.linspace(780 + nt*0.1, 780 + nt*0.9, 4).astype("int")
else:
iframes = np.linspace(nt*0.1, nt*0.9, 4).astype("int")
print(iframes)
emb_layer = mode(ilayer[igood][isort][:nn*bin_size].reshape(-1, bin_size), axis=1, keepdims=False).mode
ex_frames = obs[iframes]
print(ex_frames.shape)
np.savez(os.path.join(root, "simulations/", f"qrdqn_{env_id}_results.npz"),
X_embedding=X_embedding, emb_layer=emb_layer,
ex_frames=ex_frames, iframes=iframes)