-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset_utils.py
187 lines (168 loc) · 7.41 KB
/
dataset_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
import torch
import numpy as np
from dataset import StateActionEffectDataset
import argparse
import os
import zipfile
import wandb
def metrics_by_name(dataset_name):
train = StateActionEffectDataset(dataset_name, "train")
val = StateActionEffectDataset(dataset_name, "val")
total = StateActionEffectDataset(dataset_name, "")
path = os.path.join("data", dataset_name)
mtext = ("====TRAIN SET=====\n")
mtext += metrics(train)
mtext +=("====VAL SET=====\n")
mtext +=metrics(val)
mtext +=("====TOTAL=====\n")
mtext +=metrics(total)
f = open(os.path.join(path, "metrics.txt"), "w+")
f.write(mtext)
f.close()
def metrics(dataset):
metrics = ""
metrics += (f"{len(dataset)} samples \n")
count = 0
for e in dataset.effect:
if torch.any(e[:,2].abs().max() > 0.1):
count += 1
metrics += (f"{count / len(dataset) * 100}% resulted in ds\n")
metrics += (f"{count } samples\n")
count = 0
for e in dataset.effect:
if np.count_nonzero(e[:,2] > 0.1) == 2:
count += 1
metrics +=(f"{count / len(dataset) * 100:.2f}% multi object(2) \n")
metrics += (f"{count } samples\n")
count = 0
for e in dataset.effect:
if np.count_nonzero(e[:,2] > 0.1) > 2:
count += 1
metrics +=(f"{count / len(dataset) * 100:.2f}% multi object(3) movement\n")
metrics +=(f"{count } samples\n")
count = 0
for e in dataset.effect:
if np.count_nonzero(e[:,2] < -0.1) > 0 or np.count_nonzero(e[:,2 + 9] < -0.1) > 1:
count += 1
metrics += (f"{count / len(dataset) * 100:.2f}% towers falling down\n")
metrics +=(f"{count } samples\n")
count = 0
for sample in dataset:
target = torch.argmax(sample["action"][:,4])
if sample["effect"][target][2] < 0.1:
if torch.any((((sample["action"][:,0]) != -1).view((-1,1)) * sample["effect"])[:,2] > 0.2):
count += 1
metrics += (f"{count / len(dataset) * 100:.2f}% mistargets\n")
metrics +=(f"{count } samples\n")
num_objects = {}
for mask in dataset.mask:
mask = mask.item()
if mask in num_objects.keys():
num_objects[mask] += 1
else:
num_objects[mask] = 1
metrics += str(num_objects)
return metrics
def merge_datasets(args):
path = os.path.join("data", args.large)
l_state = torch.load(os.path.join(path, "state.pt"))
l_action = torch.load(os.path.join(path, "action.pt"))
l_effect = torch.load(os.path.join(path, "effect.pt"))
l_mask = torch.load(os.path.join(path, "mask.pt"))
l_post_state = torch.load(os.path.join(path, "post_state.pt"))
path = os.path.join("data", args.small)
s_state = torch.load(os.path.join(path, "state.pt"))
s_action = torch.load(os.path.join(path, "action.pt"))
s_effect = torch.load(os.path.join(path, "effect.pt"))
s_mask = torch.load(os.path.join(path, "mask.pt"))
s_post_state = torch.load(os.path.join(path, "post_state.pt"))
len_small = s_state.shape[0]
len_large = l_state.shape[0]
while s_state.shape[1] != l_state.shape[1]:
if s_state.shape[1] < l_state.shape[1]:
s_state = torch.cat((s_state, torch.zeros((s_state.shape[0],1 ,s_state.shape[2]))), dim = 1)
s_post_state = torch.cat((s_post_state, torch.zeros((s_state.shape[0],1 ,s_state.shape[2]))), dim = 1)
s_effect = torch.cat((s_effect, torch.zeros((s_effect.shape[0],1 ,s_effect.shape[2]))), dim = 1)
else:
l_state = torch.cat((l_state, torch.zeros((l_state.shape[0],1 ,l_state.shape[2]))), dim = 1)
l_post_state = torch.cat((l_post_state, torch.zeros((l_state.shape[0],1 ,l_state.shape[2]))), dim = 1)
l_effect = torch.cat((l_effect, torch.zeros((l_effect.shape[0],1 ,l_effect.shape[2]))), dim = 1)
divide_index = int(len_small/2)
action = torch.concat([s_action, l_action] , dim=0)
state = torch.concat([s_state, l_state], dim=0)
effect = torch.concat([s_effect, l_effect], dim=0)
mask = torch.concat([s_mask, l_mask], dim=0)
post_state = torch.concat([s_post_state, l_post_state], dim=0)
n_train_large = int(len(l_state) * 0.8)
n_train_small = int(len(s_state) * 0.8)
n_val_large = int(len(l_state) * 0.1)
n_val_small = int(len(s_state) * 0.1)
shuffle = torch.cat([
torch.arange(n_train_small),
torch.arange(n_train_large) + len(s_state),
torch.arange(n_val_small) + (n_train_small),
torch.arange(n_val_large) + (n_train_large)+ len(s_state),
torch.arange(n_val_small) + (n_train_small+n_val_small),
torch.arange(n_val_large) + (n_train_large+n_val_large)+ len(s_state)
])
action=action[shuffle]
state=state[shuffle]
effect=effect[shuffle]
mask=mask[shuffle]
post_state=post_state[shuffle]
args.o = os.path.join("./data", args.o)
if not os.path.exists(args.o):
os.makedirs(args.o)
torch.save(state, os.path.join(args.o, f"state.pt"))
torch.save(action, os.path.join(args.o, f"action.pt"))
torch.save(mask, os.path.join(args.o, f"mask.pt"))
torch.save(effect, os.path.join(args.o, f"effect.pt"))
torch.save(post_state, os.path.join(args.o, f"post_state.pt"))
def merge_rolls(args):
keys = ["action", "effect", "mask", "state", "post_state"]
output_folder = os.path.join("./data", args.o)
for key in keys:
field = torch.cat([torch.load(os.path.join(output_folder, f"{key}_{i}.pt")) for i in range(args.i)], dim=0)
torch.save(field, os.path.join(os.path.join(output_folder, f"{key}.pt")))
# for i in range(args.i):
# os.remove(os.path.join(output_folder, f"{key}_{i}.pt"))
metrics_by_name(args.o)
def upload_dataset_to_wandb(name, path):
with zipfile.ZipFile(f"{name}.zip", "w", zipfile.ZIP_DEFLATED) as zipf:
for file in os.listdir(path):
if file != ".DS_Store":
zipf.write(os.path.join(path, file), arcname=file)
wandb.init(project="acvite_exploration", entity="colorslab")
artifact = wandb.Artifact(name, type="dataset")
artifact.add_file(f"{name}.zip")
wandb.log_artifact(artifact)
os.remove(f"{name}.zip")
def get_dataset_from_wandb(name):
artifact = wandb.use_artifact(f"colorslab/active_exploration/{name}:latest", type="dataset")
artifact_dir = artifact.download()
archive = zipfile.ZipFile(os.path.join(artifact_dir, f"{name}.zip"), "r")
archive.extractall(os.path.join("data", name))
archive.close()
os.remove(os.path.join(artifact_dir, f"{name}.zip"))
if __name__ == "__main__":
parser = argparse.ArgumentParser("See dataset metrics.")
parser.add_argument("action", type=str)
parser.add_argument("-o", help="dataset name", type=str)
parser.add_argument("-s", "--small", help="smaller dataset", type=str)
parser.add_argument("-l", "--large", help="appended dataset", type=str)
parser.add_argument("-i", help="number of rolls", type=int)
args = parser.parse_args()
if args.action == "metrics":
metrics_by_name(args.o)
if args.action == "merge_datasets":
merge_datasets(args)
if args.action == "merge_rolls":
merge_rolls(args)
if args.action == "upload":
name = args.o
metrics_by_name(name)
path = os.path.join("./data", name)
upload_dataset_to_wandb(name, path)
if args.action == "download":
wandb.init(project="active_exploration", entity="colorslab")
get_dataset_from_wandb(args.o)