-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathpredictor.py
191 lines (158 loc) · 7.38 KB
/
predictor.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
import os
import argparse
import torch
import torch.nn as nn
import torch.utils.data as Data
import torchvision
from torch.autograd import Variable
from torchvision.utils import save_image
from torchvision import transforms
import numpy as np
from utils import CustomDataSetForNextFrame
from models import Predictor, CNN
from PIL import Image
from typing import List
import shutil
from torch.utils.tensorboard import SummaryWriter
parser = argparse.ArgumentParser()
parser.add_argument("--n_epochs", type=int, default=3000, help="number of epochs of training")
parser.add_argument("--batch_size", type=int, default=1024, help="size of the batches")
parser.add_argument("--lr", type=float, default=0.0002, help="adam: learning rate")
parser.add_argument("--latent_dim", type=int, default=64, help="dimensionality of the latent space")
parser.add_argument("--img_size", type=int, default=64, help="size of each image dimension")
parser.add_argument("--ngf", type=int, default=32, help="channels of generator")
parser.add_argument("--ndf", type=int, default=32, help="channels of discriminitor")
parser.add_argument("--n_classes", type=int, default=10, help="number of classes for dataset")
parser.add_argument("--action_dim", type=int, default=7, help="number of actions for dataset")
arg = parser.parse_args()
os.makedirs('predictorimages', exist_ok=True)
torch.manual_seed(1) # reproducible
cuda = True if torch.cuda.is_available() else False
FloatTensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
LongTensor = torch.cuda.LongTensor if cuda else torch.LongTensor
def train_predictor(
data_path: List,
env_name: List,
load_old_policy,
old_policy_path:str=None,
):
task_name = ''
for name in env_name:
task_name += name
task_name += '_'
root_path = 'trained_predictors'
log_dir = 'trained_predictors/log/' + task_name
if os.path.exists(log_dir):
shutil.rmtree(log_dir)
os.makedirs(log_dir, exist_ok=True)
else:
os.makedirs(log_dir, exist_ok=True)
os.makedirs(root_path+'/'+task_name+'/predicted_images', exist_ok=True)
writer = SummaryWriter(log_dir)
train_data = CustomDataSetForNextFrame(data_path, train=True)
train_loader = Data.DataLoader(dataset=train_data, batch_size=arg.batch_size, shuffle=True)
if load_old_policy == True:
predictor = Predictor()
predictor = torch.load(old_policy_path).cuda()
else:
predictor = Predictor().cuda()
optimizer = torch.optim.Adam(predictor.parameters(), lr=arg.lr)
loss_func = nn.MSELoss()
for epoch in range(arg.n_epochs):
for step, (x, action, next_img, task_index) in enumerate(train_loader):
task_index = task_index.reshape((task_index.shape[0]), 1)
task = torch.LongTensor(task_index).cuda()
task = torch.zeros(task.shape[0], arg.n_classes).cuda().scatter_(1, task, 1)
x, action, next_img = x.cuda(), action.cuda(), next_img.cuda()
action = torch.reshape(action, (action.shape[0], 1)).cuda()
action = torch.zeros(action.shape[0], arg.action_dim).cuda().scatter_(1, action, 1)
decoded = predictor(x, action, task)
loss = loss_func(decoded, next_img)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step % 10 == 0:
l = loss.data.cpu().numpy()
print('tasks: ', len(env_name), 'Epoch: ', epoch, '| train loss: %.4f' % l)
writer.add_scalar('loss', l, epoch*len(train_loader) + step)
if epoch % 100 == 0:
pair = torch.cat([next_img[0:5], decoded[0:5]], 0)
save_image(pair.data, root_path+'/'+task_name+'/predicted_images/'+"/%d.png" % epoch, nrow=5, normalize=True)
torch.save(predictor, root_path+'/'+task_name+'/predictor.pth')
def predict_images(
data_path,
policy_path,
predictor_path,
envs
):
env_names = []
for _, folders, _ in os.walk(data_path):
env_num = len(folders)
env_names = envs[0:env_num]
break
transform = transforms.Compose([
transforms.Resize(64),
transforms.ToTensor(),
]
)
policy = torch.load(policy_path).cuda()
policy.eval()
predictor = torch.load(predictor_path).cuda()
predictor.eval()
for i in range(1, env_num+1):
task = Variable(torch.LongTensor((i-1) * torch.LongTensor(np.ones([1, 1])))).cuda()
task_index = torch.zeros(task.shape[0], arg.n_classes).cuda().scatter_(1, task, 1)
for _, folders, _ in os.walk(data_path+'/'+env_names[i-1]+'_'):
rollout_num = len(folders)
break
for rollout in range(1, rollout_num+1):
folder_name = data_path+'/'+env_names[i-1]+'_/rollout'+str(rollout)
print(folder_name)
for _,_,c in os.walk(folder_name):
break
for fileitem in c:
if fileitem == '0.png':
pass
else:
os.remove(folder_name+'/'+fileitem)
print('generating rollout ', rollout, 'env name: ', env_names[i-1])
actions = []
for j in range(150): # max number of steps
img = Image.open(folder_name+'/'+str(j)+'.png').convert('RGB')
img = transform(img).cuda()
img = torch.unsqueeze(img, 0)
action = policy(img)
loss_to_stop = nn.CrossEntropyLoss()(action, torch.Tensor([0]).long().cuda())
print('step: ', j, 'loss: ', loss_to_stop.item())
action = torch.argmax(action)
print('step: ', j, 'action: ', action.data.cpu().numpy())
if action == 0:
action = torch.Tensor([0])[0].long().cuda()
actions.append(action.data.cpu().numpy())
actions = np.array(actions)
new_actions = []
for i in range(actions.shape[0]):
if actions[i] == 0:
new_actions.append([0, 0, 0, 0])
elif actions[i] == 1:
new_actions.append([2, 0, 0, 0])
elif actions[i] == 2:
new_actions.append([-2, 0, 0, 0])
elif actions[i] == 3:
new_actions.append([0, 2, 0, 0])
elif actions[i] == 4:
new_actions.append([0, -2, 0, 0])
elif actions[i] == 5:
new_actions.append([0, 0, 2, 0])
elif actions[i] == 6:
new_actions.append([0, 0, -2, 0])
actions = np.array(new_actions)
np.save(folder_name+'/'+ 'actions.npy', actions)
break
else:
actions.append(action.data.cpu().numpy())
action = torch.reshape(action, (1, 1)).cuda()
action = torch.zeros(action.shape[0], arg.action_dim).cuda().scatter_(1, action, 1)
next_img = predictor(img, action, task_index)
save_image(next_img.data, folder_name+'/'+str(j+1)+'.png', nrow=1, normalize=True)
img = next_img