-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
287 lines (231 loc) · 9.65 KB
/
utils.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
import os
from termcolor import colored
import pickle
from time import time
from enum import Enum
import numpy as np
import torch
from collections import namedtuple, defaultdict
from tensorboardX import SummaryWriter
import IPython
def make_path(path):
prefix = ''
if path[0] == '/':
prefix = '/'
path = path[1:]
dirs = path.split("/")
dirs = ['{}{}'.format(prefix, "/".join(dirs[:i+1])) for i in range(len(dirs))]
for _dir in dirs:
if not os.path.isdir(_dir):
os.makedirs(_dir)
def print_c(txt, color="white"):
print(colored(txt, color))
def save_pkl(obj, path):
dir = "".join(path.split('/')[:-1])
make_path(dir)
with open(path, 'wb') as f:
pickle.dump(obj, f)
print_c('[i] Successfully pickled {}.'.format(path), 'green')
def load_pkl(path):
with open(path, 'rb') as f:
obj = pickle.load(f)
return obj
def count_parameters(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
class HistoryStoreCache(object):
def __init__(self, capacity):
self.capacity = capacity
self.reset()
def push(self, stuff):
if len(self.memory) < self.capacity:
self.memory.append(stuff)
else:
self.memory = self.memory[1:] + [stuff]
def replace_last(self, stuff):
self.memory[-1] = stuff
def __call__(self):
return self.memory
def reset(self):
self.memory = []
def __len__(self):
return len(self.memory)
def avg(self):
return np.mean(self.memory)
# Textworld specific
def get_points(score, wt):
possible_points = get_points_from_wt(wt)
percentage = float(score) / possible_points * 100
return score, possible_points, percentage
def get_points_from_wt(wt):
pts = []
dropped_things = []
for cmd in wt:
cmd = cmd.replace('pork chop', 'pork c_op')
if 'drop' in cmd:
dropped_things.append(cmd.replace('drop ', ''))
pts.append(0.)
continue
if 'take' in cmd and not 'knife' in cmd and not any([thing in cmd for thing in dropped_things]):
pts.append(1.)
continue
if 'cook' in cmd:
pts.append(1.)
continue
if 'slice' in cmd or 'dice' in cmd or 'chop' in cmd:
pts.append(1.)
continue
if cmd in ['prepare meal', 'eat meal']:
pts.append(1.)
continue
pts.append(0.)
return int(sum(pts))
class StatisticsTracker:
def __init__(self, tb_dir=None):
self._init_statistics()
self.writer = None
self.last_log_time = -1
self.stats_episode = defaultdict(list)
if tb_dir is not None and 'None' not in tb_dir and tb_dir != '':
make_path(tb_dir)
self.writer = SummaryWriter(tb_dir)
def _init_statistics(self):
len_cap_dict = {'short': 20,
'interm': 50,
'long': 100}
self.statistics_last_k_episodes = {
statistic: {length: HistoryStoreCache(capacity=capacity) for length, capacity in len_cap_dict.items()} for statistic in
['loss', 'percentage', 'accuracy', 'score', 'reward', 'policy', 'value', 'entropy', 'confidence']
}
def stats_episode_append(self, **kwargs):
for key, value in kwargs.items():
self.stats_episode[key].append(value)
def stats_episode_clear(self):
self.stats_episode = defaultdict(list)
def flush_episode_statistics(self, possible_points=5., episode_no=0, eta=0., steps=100., **kwargs):
'''
At the end of the episode, this method is invoked and takes all the statistics from the episode stats and
puts them in the according 'last_k_episode' stats. Then it writes the result to Tensorboard.
'''
if self.writer is None:
return
for key, stat_episode in self.stats_episode.items():
if isinstance(stat_episode, list) and len(stat_episode) > 0 and isinstance(stat_episode[0], torch.Tensor):
stat_episode = [val.item() for val in stat_episode]
episode_value = np.mean(stat_episode) if key not in ['score'] else np.sum(stat_episode)
try:
for length_key, history_cache in self.statistics_last_k_episodes[key].items():
history_cache.push(episode_value)
self.writer.add_scalar('{} ({})'.format(key, length_key), history_cache.avg(), episode_no)
except:
print('Flushing episode statistics did not work for {}'.format(key))
if key == 'score':
# compute percentage
for length_key, history_cache in self.statistics_last_k_episodes['percentage'].items():
history_cache.push((episode_value, possible_points))
percentage = np.sum(np.array(history_cache.memory)[:,0]) / np.sum(np.array(history_cache.memory)[:,1])
self.writer.add_scalar('Percentage ({})'.format(length_key), percentage, episode_no)
self.writer.add_scalar('Steps', steps, episode_no)
self.writer.add_scalar('TF probability', eta, episode_no)
if 'cmds' in kwargs:
# list of list of commands
for idx, command_list in enumerate(kwargs['cmds']):
points = np.sum(self.stats_episode['score'])
if 'points' in kwargs:
points = kwargs['points'][idx]
self.writer.add_text('commands', 'Score: {}/{} points. #### {}'.format(points,
possible_points,
", ".join(command_list[:-1])),
episode_no)
break
class Saver:
def __init__(self, model, ckpt_path='NOPATH', experiment_tag='NONAME', load_pretrained=False,
pretrained_model_path=None, device='cpu', save_frequency=600):
self.model = model
self.device = device
self.model_checkpoint_path = ckpt_path
self.experiment_tag = experiment_tag
self.last_save_time = time()
self.save_frequency = save_frequency
self.only_load = ckpt_path == 'NOPATH'
if load_pretrained and pretrained_model_path is not None:
self.pretrained_model_path = pretrained_model_path
self._load_from_checkpoint()
def save(self, epoch=None, episode=None):
if time() - self.last_save_time > self.save_frequency and not self.only_load:
self._save_checkpoint(epoch, episode)
self.last_save_time = time()
def _save_checkpoint(self, epoch=None, episode=None):
"""
Save the model checkpoint.
"""
if self.only_load:
return
make_path(self.model_checkpoint_path)
save_to = "{}/{}".format(self.model_checkpoint_path, self.experiment_tag)
if epoch is not None:
save_to += '_epoch{}'.format(epoch)
if episode is not None:
save_to += '_episode{}'.format(episode)
torch.save(self.model.state_dict(), save_to)
print("Saved model to '{}'".format(save_to))
self.last_save_time = time()
def _load_from_checkpoint(self):
load_from = self.pretrained_model_path
# print("Trying to load model from {}.".format(load_from))
try:
if self.device == 'cpu':
state_dict = torch.load(load_from, map_location='cpu')
else:
state_dict = torch.load(load_from, map_location=self.device)
self.model.load_state_dict(state_dict, strict=True)
self.model.to(self.device)
print("Loaded model from '{}'".format(load_from))
except:
print("Failed to load checkpoint {} ...".format(load_from))
IPython.embed()
Event = Enum('Event', 'NEWEPOCH STARTTRAINING NEWEPISODE SAVEMODEL TRACKSTATS')
class EventHandler:
Event = Event
def __init__(self):
self.handlers = {event_name:[] for event_name in [event.name for event in Event]}
def add(self, handler, event):
self.handlers[event.name].append(handler)
return self
def remove(self, handler, event):
self.handlers[event.name].remove(handler)
return self
def __call__(self, event, **kwargs):
for handler in self.handlers[event.name]:
handler(**kwargs)
class StepCounter:
def __init__(self, batch_size=1, max_nb_steps=100):
self.batch_size = batch_size
self.max_nb_steps = max_nb_steps
self.counter = {
'epoch': 0,
'episode': 0,
'global_steps': 0,
'steps': 0,
'steps_taken': np.array([0] * self.batch_size)
}
def __call__(self, key):
return self.counter[key]
def new_episode(self):
self.counter['episode'] += 1
self.counter['steps'] = 0
self.counter['steps_taken'] = np.array([0] * self.batch_size)
def step(self):
self.counter['steps'] += 1
self.counter['global_steps'] += 1
def new_epoch(self):
self.counter['epoch'] += 1
def increase_steps_taken(self, idx):
self.counter['steps_taken'][idx] += 1
def recompute_steps_taken(self, just_finished_mask):
self.counter['steps_taken'] = [self.counter['steps'] if jf else st for jf, st in zip(just_finished_mask, self.counter['steps_taken'])]
class flist(list):
def append(self, object_):
if isinstance(object_, list):
[super(flist, self).append(o) for o in object_]
else:
super(flist, self).append(object_)