-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval_randomlang.py
233 lines (208 loc) · 10.1 KB
/
eval_randomlang.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
import numpy as np
#from sklearn.metrics import mutual_info_score
from sklearn.feature_selection import mutual_info_classif
import csv
from message_dataset import MessageDataset
from population import Population
#from model_utils import MessageClassifier, MessageMLP, find_lengths, add_eos_to_messages
from stats import AverageMeterSet, StatTracker
from tqdm import tqdm
import click
import os
import sys
from pathlib import Path
## Hyperparameters!
@click.command()
@click.option("--load-name", default=None)
@click.option("--start-gen", default=0)
@click.option("--end-gen", default=0)
@click.option("--cuda", is_flag=True)
@click.option("--generalize", is_flag=True)
@click.option("--csv-general-eval-file", default="langperm_eval_results.csv")
# learn hps
@click.option("--n-epochs-init", default=10, help="nb of epochs to run initial generation")
@click.option("--n-epochs-teach", default=10, help="nb of epochs to run for child-teacher play")
@click.option("--n-epochs-self", default=10, help="nb of epochs to run for selfplay")
@click.option("--n-epochs-comm", default=10, help="nb of epochs to run for community play")
@click.option("--lr", default=1e-3)
@click.option("--batch-size", default=32)
@click.option("--teacher", is_flag=True)
@click.option("--selfplay", is_flag=True)
@click.option("--ce-loss", is_flag=True)
@click.option("--class-weight", default=1)
@click.option("--similarity-weight", default=0)
@click.option("--distillation-weight", default=0)
# game_hps
@click.option("--n-distractors", default=3)
@click.option("--n-games", default=32000)
@click.option("--images", default="random")
@click.option("--perspective", is_flag=True)
@click.option("--proportion", default=1.0)
@click.option("--color-balance", is_flag=True)
# world hps
@click.option("--n-pairs", default=1)
@click.option("--population-size", default=2)
@click.option("--n-generations", default=2)
@click.option("--seed", default=0) # required
@click.option("--data-path", default="../game_data/clevr_1_10000/")
@click.option("--experiments-base", default="experiments")
# vision hp
@click.option("--encoding-size", default=1024)
@click.option("--compression-size", default=256, type=int)
# agent hps
@click.option("--hidden-size", default=128)
@click.option("--hidden-mlp", default=128)
@click.option("--emb-size", default=64)
@click.option("--message-length", default=7)
@click.option("--vocab-size", default=60)
def run(n_epochs_init, n_epochs_teach, n_epochs_self, n_epochs_comm, lr, batch_size, teacher, selfplay, ce_loss, class_weight, similarity_weight, distillation_weight, n_distractors, n_games, data_path, images, proportion, color_balance, n_pairs, population_size, n_generations,seed, load_name, cuda, encoding_size, compression_size, hidden_size, hidden_mlp, emb_size, message_length, vocab_size, generalize, perspective, csv_general_eval_file, start_gen, end_gen, experiments_base):
data_path = Path(data_path)
for gen in range(start_gen, (end_gen+1)):
learn_hps = {"n_epochs_init": n_epochs_init,
"n_epochs_teach": n_epochs_teach,
"n_epochs_self": n_epochs_self,
"n_epochs_comm": n_epochs_comm,
"lr": lr,
"batch_size": batch_size,
"teacher": teacher,
"selfplay": selfplay,
"ce_loss": ce_loss,
"class_weight": class_weight,
"similarity_weight": similarity_weight,
"distillation_weight": distillation_weight}
game_hps = {"n_distractors": n_distractors,
"n_games": n_games,
"data_path": data_path,
"images": images,
"perspective": perspective,
"proportion": proportion,
"color_balance": color_balance}
world_hps = {"n_pairs": n_pairs,
"population_size": population_size,
"n_generations": n_generations,
"seed": seed,
"experiment_name": load_name,
"load_name": load_name,
"load_gen": gen,
"cuda": cuda,
"generalize": generalize,
"csv_general_eval_file": csv_general_eval_file,
"experiments_base": experiments_base}
vision_hps = {"encoding_size": encoding_size,
"compression_size": compression_size}
agent_hps = {"hidden_size": hidden_size,
"hidden_mlp": hidden_mlp,
"emb_size": emb_size,
"message_length": message_length,
"vocab_size": vocab_size}
world = World(world_hps, learn_hps, game_hps, agent_hps, vision_hps)
print("DEVICE : "+str(world.device))
permprops = [0.2, 0.4, 0.6, 0.8, 1.0]
#permprops = [0.8]
for permprop in permprops:
world.randlang_eval(permprop)
class World(object):
def __init__(self, world_hps, learn_hps, game_hps, agent_hps, vision_hps):
self.n_pairs = world_hps["n_pairs"]
self.population_size = world_hps["population_size"]
self.n_generations = world_hps["n_generations"]
self.load_name = world_hps["load_name"]
self.gen = world_hps["load_gen"]
self.csv_general_eval_file = world_hps["csv_general_eval_file"]
self.learn_hps = learn_hps
self.game_hps = game_hps
self.world_hps = world_hps
self.agent_hps = agent_hps
self.vision_hps = vision_hps
self.output_dir = os.path.join(self.world_hps["experiments_base"], self.world_hps["experiment_name"])
self.stat_tracker = StatTracker(log_dir=os.path.join(self.output_dir, "tensorboard_log_results"))
if self.world_hps["cuda"]:
self.device = "cuda"
else:
self.device = "cpu"
def calculate_mutual_information(self, dataset):
labels = []
messages = []
batch_size = 200
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=True)
bar = tqdm(dataloader)
for idx, sample in enumerate(bar, start=1):
message = sample["message"]
messages.append(message)
label = sample["label"]
label = label.squeeze(1)
labels.append(label)
messages = torch.cat(messages).numpy()
labels = torch.cat(labels).numpy()
mi_mess_label = np.sum(mutual_info_classif(messages, labels))
return mi_mess_label
def randlang_eval(self, permprop):
print("\n----------------------------------------")
print("INITIALIZING POPULATION")
print("----------------------------------------\n")
pop = Population(self.population_size, self.n_generations, self.n_pairs,
self.learn_hps, self.agent_hps, self.vision_hps, self.game_hps,
self.world_hps, self.device, self.stat_tracker, self.csv_general_eval_file)
imgs_all = pop.test_set.imgs
labels_all = pop.test_set.labels
size = imgs_all.shape[0]
labels_all = labels_all[:, 1]
labels_all = np.resize(labels_all, (size, 1))
n_batches, last_batch_size = divmod(size, 200)
batches = []
for i in range(0, n_batches):
img_batch = torch.tensor(imgs_all[(i*200):(i*200+200)]).float()
label_batch = torch.tensor(labels_all[(i*200):(i*200+200)])
batches.append((img_batch,label_batch))
if last_batch_size > 0:
final_img_batch = torch.tensor(imgs_all[(n_batches*200):]).float()
final_label_batch = torch.tensor(labels_all[(n_batches*200):])
batches.append((final_img_batch,final_label_batch))
print("\n----------------------------------------")
print("EVALUATING MUTUAL INFORMATION")
print("----------------------------------------\n")
filename = os.path.join("experiments", str("mi_"+self.csv_general_eval_file))
idx = np.array(range(0,self.agent_hps["vocab_size"]))
perm_size = int(permprop * len(idx))
perm_id = np.random.choice(len(idx), size=perm_size, replace=False)
rand = idx[perm_id]
np.random.shuffle(rand)
for agent_id, agent in enumerate(pop.child_agents):
print("\n----------------------------------------")
print("PERMUTE EMBEDDINGS")
print("----------------------------------------\n")
idx[perm_id] = rand
input = torch.LongTensor([idx]).to(self.device)
permuted_embeddings = agent.shared_embedding(input).squeeze(0)
agent.shared_embedding.weight.data.copy_(permuted_embeddings)
messages = []
first_embeddings = []
labels = []
agent.eval()
print("\n----------------------------------------")
print("CREATING MESSAGE DATASET")
print("----------------------------------------\n")
for img_batch,label_batch in batches:
labels.append(label_batch)
with torch.no_grad():
tgt_img = img_batch.to(self.device)
message, first_embedding = agent.get_message_firstembedding(tgt_img)
first_embeddings.append(first_embedding)
first_embeddings = torch.cat(first_embeddings).float()
labels = torch.cat(labels).long()
embedding_dataset = MessageDataset(first_embeddings, labels)
mi_emb_label = self.calculate_mutual_information(embedding_dataset)
with open(filename,'a') as f:
writer = csv.writer(f)
writer.writerow([self.load_name, self.gen, permprop, agent_id, mi_emb_label])
print("\n----------------------------------------")
print("EVALUATING ACCURACY")
print("----------------------------------------\n")
for pair in pop.pairs:
pop.eval(pair, n_distractors=3, n_games=1000, images="shape", only_rewards=True, permprop=permprop)
if __name__ == "__main__":
run()