-
Notifications
You must be signed in to change notification settings - Fork 1
/
analysis_classes.py
226 lines (191 loc) · 6.89 KB
/
analysis_classes.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
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
from gym_derk.envs import DerkEnv
import time
import torch
from torch import nn
import numpy as np
import copy
import time
import os
from tqdm import tqdm
from torch_truncnorm.TruncatedNormal import TruncatedNormal
import random
from derk_PPO_LSTM import lstm_agent
import matplotlib.pyplot as plt
import json
OBS_KEYS = [
"Hitpoints",
"Ability0Ready",
"FriendStatueDistance",
"FriendStatueAngle",
"Friend1Distance",
"Friend1Angle",
"Friend2Distance",
"Friend2Angle",
"EnemyStatueDistance",
"EnemyStatueAngle",
"Enemy1Distance",
"Enemy1Angle",
"Enemy2Distance",
"Enemy2Angle",
"Enemy3Distance",
"Enemy3Angle",
"HasFocus",
"FocusRelativeRotation",
"FocusFacingUs",
"FocusFocusingBack",
"FocusHitpoints",
"Ability1Ready",
"Ability2Ready",
"FocusDazed",
"FocusCrippled",
"HeightFront1",
"HeightFront5",
"HeightBack2",
"PositionLeftRight",
"PositionUpDown",
"Stuck",
"UnusedSense31",
"HasTalons",
"HasBloodClaws",
"HasCleavers ",
"HasCripplers",
"HasHealingGland",
"HasVampireGland",
"HasFrogLegs",
"HasPistol",
"HasMagnum",
"HasBlaster",
"HasParalyzingDart",
"HasIronBubblegum",
"HasHeliumBubblegum",
"HasShell",
"HasTrombone",
"FocusHasTalons",
"FocusHasBloodClaws",
"FocusHasCleavers",
"FocusHasCripplers",
"FocusHasHealingGland",
"FocusHasVampireGland",
"FocusHasFrogLegs",
"FocusHasPistol",
"FocusHasMagnum",
"FocusHasBlaster",
"FocusHasParalyzingDart",
"FocusHasIronBubblegum",
"FocusHasHeliumBubblegum",
"FocusHasShell",
"FocusHasTrombone",
"UnusedExtraSense30",
"UnusedExtraSense31"]
class analysis_agent(lstm_agent):
def __init__(self, lstm_size, device, activation = nn.Tanh()):
super().__init__(lstm_size, device, activation)
self.device = device
def analyze(self, obs, act):
act = torch.Tensor(act).to(self.device).flatten(end_dim = 1) #flatten actions
obs = torch.Tensor(obs).to(self.device)
obs.requires_grad = True
continuous_means, continuous_stds, discrete_output, value, _ = self(obs, None, act)
normal_dists = torch.distributions.Normal(continuous_means, continuous_stds)
entropy = normal_dists.entropy()
for i in range(len(discrete_output)):
discrete_probs = nn.functional.log_softmax(discrete_output[i], dim=1).exp()
discrete_dist = torch.distributions.Categorical(discrete_probs)
entropy = torch.cat((entropy, discrete_dist.entropy().unsqueeze(1)), axis=1)
self.optimizer.zero_grad()
loss = continuous_means.mean()
#loss = entropy[1]
loss.mean().backward(retain_graph = True)
'''
time_obs_grad = torch.abs(obs.grad).reshape(obs.grad.shape[0] // 150, 150, 64).mean(axis=0)
print(time_obs_grad.shape)
mean_time_obs_grad = (time_obs_grad / time_obs_grad.sum(axis=1).unsqueeze(-1)) * 100
plt.plot(mean_time_obs_grad[:,0].detach().cpu().numpy())
plt.show()
'''
obs_grad = torch.abs(obs.grad)
totals = torch.sum(obs_grad, axis = 2)
percent_obs_grad = (obs_grad / totals.unsqueeze(-1)) * 100
avg_percent_obs_grad = percent_obs_grad.mean(axis = 0).mean(axis = 0)
return avg_percent_obs_grad.cpu().detach().numpy().tolist()
device = "cuda:0"
ITERATIONS = 1000000
league = []
league_analysis = []
root_dir = "checkpoints/TEST_LEAGUE_AGENTS"
for root, dirs, files in os.walk(root_dir):
for name in files: # Load in the agents stored at root_dir
if "best" in name:
temp = analysis_agent(1024, device)
temp.load_state_dict(torch.load(os.path.join(root, name)))
temp.name = name.replace("_best", "")
league.append(temp)
league_elos={}
with open(root_dir+"/elo_rankings.json") as f:
data=f.read()
data=data.replace("_best", '')
league_elos=json.loads(data)
teams_per_member = len(league)
arm_weapons = [["Pistol", "Magnum", "Blaster"], ["Talons", "BloodClaws", "Cleavers", "Cripplers"], ["Talons", "Blaster", "Cleavers"]]
misc_weapons=[["IronBubblegum", "HeliumBubblegum"], ["Shell", "Trombone"], ["FrogLegs", "HeliumBubblegum", "Shell"]]
tail_weapons=[["HealingGland"], ["VampireGland", "ParalyzingDart"], ["VampireGland", "ParalyzingDart"]]
max_arenas = 800
teams_per_member = max_arenas // (len(league) // 2) // 4
n_arenas = (len(league)*teams_per_member) // 2
random_configs = [{"slots": [random.choice(arm_weapons[i%3]), random.choice(misc_weapons[i%3]), random.choice(tail_weapons[i%3])]} for i in range(3 * n_arenas)]
env = DerkEnv(n_arenas = n_arenas, turbo_mode = True, home_team = random_configs, away_team = random_configs)
for i in range(1):
#randomize matchings between league members
scrambled_team_IDS = np.random.permutation(env.n_agents // 3)
league_agent_mappings = []
for i in range(len(league)):
member_matches = scrambled_team_IDS[teams_per_member*i:teams_per_member*(i+1)]
league_agent_mappings.append(np.concatenate([(member_matches * 3) + i for i in range(3)], axis = 0))
observation = [[] for i in range(len(league))]
action = [[] for i in range(len(league))]
reward = [[] for i in range(len(league))]
states = [None for i in range(len(league))]
observation_n = env.reset()
with torch.no_grad():
while True:
action_n = np.zeros((env.n_agents, 5))
for i in range(len(league)):
action_n[league_agent_mappings[i]], states[i] = league[i].get_action(observation_n[league_agent_mappings[i]], states[i])
#act in environment and observe the new obervation and reward (done tells you if episode is over)
observation_n, reward_n, done_n, _ = env.step(action_n)
#collect experience data to learn from
for i in range(len(league)):
observation[i].append(observation_n[league_agent_mappings[i]])
reward[i].append(reward_n[league_agent_mappings[i]])
action[i].append(action_n[league_agent_mappings[i]])
if all(done_n):
break
# reshapes all collected data to [episode num, timestep]
for i in range(len(league)):
observation[i] = np.swapaxes(np.array(observation[i]), 0, 1)
reward[i] = np.swapaxes(np.array(reward[i]), 0, 1)
action[i] = np.swapaxes(np.array(action[i]), 0, 1)
league_analysis.append(league[i].analyze(observation[i], action[i]))
for i, key in enumerate(OBS_KEYS):
x = np.zeros(len(league))
y = np.zeros(len(league))
for j in range(len(league)):
x[j] = league_elos[league[j].name]
y[j] = league_analysis[j][i]
plt.scatter(x, y)
plt.title(key + " Percent Gradient vs ELO")
plt.xlabel("ELO")
plt.ylabel("Percent of Gradient")
plt.savefig("analysis/" + key + ".png")
plt.close()
x=np.zeros(len(league))
y=np.zeros(len(league))
for j in range(len(league)):
x[j] = league_elos[league[j].name]
y[j] = int(league[j].name)
plt.scatter(x, y)
plt.title("Iteration vs ELO")
plt.xlabel("ELO")
plt.ylabel("Iteration")
plt.savefig("analysis/iteration.png")
plt.close()