-
Notifications
You must be signed in to change notification settings - Fork 0
/
calc_val_loss_table.py
293 lines (256 loc) · 13.1 KB
/
calc_val_loss_table.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
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
from tqdm import tqdm
import torch
import torch.nn as nn
import numpy as np
from train import build_model
import encoders
import os
import gymnasium as gym
from simple_env import SimpleEnv
from grid_world import GridWorld
import json
from stable_baselines3 import PPO
from decoder import *
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def get_random_policy(obs_size, a_size, hidden_size=64):
model = nn.Sequential(nn.Linear(obs_size, hidden_size),
nn.ReLU(),
nn.Linear(hidden_size, hidden_size),
nn.ReLU(),
nn.Linear(hidden_size, a_size)
)
return model
# TODO summarize this function somewhere (3/3)
def get_environment(env_name):
if env_name == "SimpleEnv":
return SimpleEnv()
if env_name == "GridWorld":
return GridWorld()
else:
env = gym.make(env_name)
return env
def get_transitions_dir(env_name, random_fraction):
dir_path = os.path.dirname(os.path.realpath(__file__))
val_path = os.path.join(dir_path, "val_transitions")
# Make directory for env if not existing
env_path = os.path.join(val_path, env_name)
# Make directory for random fraction if not existing
fraction_str = "expert_" + str(int(100*(1-random_fraction))) + "-" + "random_" + str(int(100*random_fraction))
fraction_path = os.path.join(env_path, fraction_str)
return fraction_path
def gather_context(length, batch_size, feature_size, env_name, action_gather_type="random"):
"""
Gather Transition of episodes and add to context used for prediction.
x_context (shape: seq_len, batch, num_features) contains observation and action
y_context (shape: seq_len, batch, num_features) contains new observation and reward
x_mean (shape: batch, num_features) is mean of x per column (first seq_len - 1 elements)
x_std (shape: batch, num_features) is std deviation of x per column (first seq_len - 1 elements)
y_mean (shape: batch, num_features) is mean of x per column (first seq_len - 1 elements)
y_std (shape: batch, num_features) is std deviation of y per column (first seq_len - 1 elements)
"""
x_context = torch.zeros((length, batch_size, feature_size))
y_context = torch.zeros((length, batch_size, feature_size))
for b in range(batch_size):
env = get_environment(env_name)
observation, info = env.reset()
if action_gather_type == "Random Policy":
if type(env.action_space) is gym.spaces.Discrete: # Discrete case
a_shape = env.action_space.n
else: # Continous
a_shape = env.action_space.shape[0]
policy = get_random_policy(env.observation_space.shape[0], a_shape)
elif action_gather_type == "Expert Policy" or action_gather_type == "mixture" or action_gather_type == "pExpert":
policy = PPO.load("val_transitions/expert_policies/PPO_" + env_name + ".zip")
repeats = 100
for k in range(length):
if action_gather_type == "random":
action = env.action_space.sample()
elif action_gather_type == "repeats":
if repeats >= 3:
action = env.action_space.sample()
repeats = 0
else:
repeats += 1
elif action_gather_type == "Random Policy":
with torch.no_grad():
if type(env.action_space) is gym.spaces.Discrete: # discrete case
action = torch.argmax(policy(torch.tensor(observation, dtype=torch.float)))
else: # Continous case
action = policy(torch.tensor(observation, dtype=torch.float))
action = action.numpy()
elif action_gather_type == "Expert Policy":
action, _ = policy.predict(observation)
elif action_gather_type == "mixture":
if k < length // 3:
action, _ = policy.predict(observation)
elif k < 2 * (length // 3):
eps = np.random.rand()
if eps < 0.5:
action = env.action_space.sample()
else:
action, _ = policy.predict(observation)
else:
action = env.action_space.sample()
elif action_gather_type == "pExpert":
eps = np.random.rand()
if eps < 0.5:
action = env.action_space.sample()
else:
action, _ = policy.predict(observation)
else:
raise NotImplementedError("No Valid Context gathering method used!")
x_context[k, b, :observation.shape[0]] = torch.tensor(observation)
a_shape = 1 if type(action) is int or len(action.shape) == 0 else action.shape[0]
x_context[k, b, -a_shape:] = torch.tensor(action)
observation, reward, terminated, truncated, info = env.step(action)
y_context[k, b, :observation.shape[0]] = torch.tensor(observation)
y_context[k, b, -1] = torch.tensor(reward)
if terminated or truncated:
repeats = 100
observation, info = env.reset()
x_mean = torch.mean(x_context[:1000, :, :], dim=0)
x_std = torch.std(x_context[:1000, :, :], dim=0)
x_context = torch.nan_to_num((x_context - x_mean) / x_std, nan=0)
y_mean = torch.mean(y_context[:1000, :, :], dim=0)
y_std = torch.std(y_context[:1000, :, :], dim=0)
y_context = torch.nan_to_num((y_context - y_mean) / y_std, nan=0)
perm = torch.randperm(length)
x_context = x_context[perm]
y_context = y_context[perm]
return x_context.to(device), y_context.to(device), x_mean.to(device), x_std.to(device), y_mean.to(device), y_std.to(device)
def val_loss_table(model, debug_truncation=False):
# Number of states plus number of action -> maximum size
num_features = 14
number_context_batches = 8
summary_dict = {}
env_collection = ["GridWorld", "CartPole-v1", "Pendulum-v1", "SimpleEnv", "Reacher-v4", "MountainCar-v0"]
states_list = []
action_list = []
error_list = []
random_fractions = [1.0, 0.5, 0.0]
for environment in env_collection:
summary_dict[environment] = {}
x, y, x_means, x_stds, y_means, y_stds = gather_context(1001, number_context_batches, num_features, environment,
action_gather_type="pExpert")
train_len = 1000
#with open('simpleenv_context_expert.json', 'w') as fp:
# json.dump(((x[:train_len] * y_stds) + y_means)[:, :, :].view(-1, 2).cpu().tolist(), fp, indent=4)
train_x = x[:train_len]
train_y = y[:train_len]
test_x = x[:]
for fraction in random_fractions:
# Logging of validations
losses = []
axis_losses = []
context_losses = []
# get dir of env + random action fraction
transition_dir = get_transitions_dir(environment, fraction)
# load transitions for this setting
states = np.load(os.path.join(transition_dir, "states.npy"))
actions = np.load(os.path.join(transition_dir, "actions.npy"))
next_states = np.load(os.path.join(transition_dir, "next_states.npy"))
rewards = np.load(os.path.join(transition_dir, "rewards.npy"))
dones = np.load(os.path.join(transition_dir, "dones.npy"))
# for each transition predict using #batch_size different context
for i, transition in enumerate(tqdm(zip(states, actions, next_states, rewards, dones), total=states.shape[0])):
s, a, ns, r, d = transition
if debug_truncation and i > 5:
break
new_x = torch.zeros_like(test_x[1000, 0])
#states_list.append(s.tolist())
new_x[:s.shape[0]] = torch.tensor(s)
# scalar actions are not saved as 1-d arrays
action_shape = 1 if len(a.shape) == 0 else a.shape[0]
#action_list.append(a.tolist())
new_x[-action_shape:] = torch.tensor(a)
norm_x = torch.nan_to_num((new_x - x_means) / x_stds, nan=0)
test_x[1000] = norm_x
new_y = torch.zeros_like(y[1000, 0])
new_y[:ns.shape[0]] = torch.tensor(ns)
new_y[-1] = torch.tensor(r)
norm_y = torch.nan_to_num((new_y - y_means) / y_stds, nan=0)
y[1000] = norm_y
with torch.no_grad():
# predicting new logits
logits = model(train_x, train_y, test_x)
# from normalized values to true values
y_pred = (logits * y_stds) + y_means
y_target = (y * y_stds) + y_means
# loss average over all axis
loss = torch.nn.functional.mse_loss(y_pred[1000:, :, :], y_target[1000:, :, :], reduction="none")
# Total mean loss
losses.append(loss.mean())
# loss per axis
#error_list.append(
# (y_target[1000:, :, :] - y_pred[1000:, :, :]).mean(axis=(0, 1)).cpu().tolist())
axis_losses.append(loss.mean(axis=(0, 1)))
# loss for each context to measure variance between context
context_losses.append(loss.mean(axis=(0, 2)))
single_setting_dict = {
# Overall loss of all axis context and steps
"loss": torch.tensor(losses).mean().item(),
# std of the losses
"loss_std": torch.tensor(losses).std().item(),
# losses per axis -> dim 1 of states .... and reward prediction (over all axis and steps)
"axis_loss": torch.cat(axis_losses).view(-1, num_features).mean(axis=0).tolist(),
# std for the per axis loss
"axis_loss_std": torch.cat(axis_losses).view(-1, num_features).std(dim=0).tolist(),
# Loss per context over all axis and all steps
"context_loss": torch.cat(context_losses).view(-1, number_context_batches).mean(axis=0).tolist(),
# std per context over all axis and steps
"context_loss_std": torch.cat(context_losses).view(-1, number_context_batches).std(dim=0).tolist(),
# std between per context loss -> variation of means of different contexts
"std_between_context": torch.cat(context_losses).view(-1, number_context_batches).mean(axis=0).std().tolist(),
}
summary_dict[environment][str(fraction)] = single_setting_dict
#plot_mc_error = {"state": states_list,
# "action": action_list,
# "error": error_list}
# with open('plotsimpleenverror_expert.json', 'w') as fp:
# json.dump(plot_mc_error, fp, indent=4)
return summary_dict
def get_model():
num_features = 14
# Loss to be used for evaluation of model
criterion = nn.MSELoss(reduction='none')
encoder_decoder_hps = {"decoder_activation": "sigmoid", "decoder_depth": 2, "decoder_res_connection": True,
"decoder_type": "cat", "decoder_use_bias": False, "decoder_width": 64,
"encoder_activation": "gelu", "encoder_depth": 3,
"encoder_res_connection": True, "encoder_type": "cat", "encoder_use_bias": True,
"encoder_width": 512}
if encoder_decoder_hps["encoder_type"] == "mlp":
gen_x = mlp_encoder_generator_generator(encoder_decoder_hps)
gen_y = gen_x
elif encoder_decoder_hps["encoder_type"] == "cat":
gen_x = cat_encoder_generator_generator(encoder_decoder_hps, target=False)
gen_y = cat_encoder_generator_generator(encoder_decoder_hps, target=True)
if encoder_decoder_hps["encoder_type"] == "mlp":
dec_model = mlp_decoder_generator_generator(encoder_decoder_hps)
elif encoder_decoder_hps["decoder_type"] == "cat":
dec_model = cat_decoder_generator_generator(encoder_decoder_hps)
decoder_dict = {"standard": (dec_model, 14)}
# building Transformer model and loading weights
hps = {'test': True}
pfn = build_model(
criterion=criterion,
encoder_generator=gen_x,
test_batch=None,
n_out=14,
emsize=512, nhead=8, nhid=1024, nlayers=6,
seq_len=1001,
y_encoder_generator=gen_y,
decoder_dict=decoder_dict,
extra_prior_kwargs_dict={'num_features': num_features, 'hyperparameters': hps},
)
print(
f"Using a Transformer with {sum(p.numel() for p in pfn.parameters()) / 1000 / 1000:.{2}f} M parameters"
)
pfn.load_state_dict(torch.load("saved_models/exp_seed_1.pt"))
pfn.eval()
print(device)
return pfn.to(device)
if __name__ == '__main__':
results = val_loss_table(get_model())
print(results)
with open('val_transitions/scores/validation_scores.json', 'w') as fp:
json.dump(results, fp, indent=4)