-
Notifications
You must be signed in to change notification settings - Fork 13
/
train.py
185 lines (167 loc) · 7.68 KB
/
train.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
import torch
import torchvision
from torchvision import transforms
import os
import numpy as np
from utils import *
from model import *
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def train_coord_encoder(path):
coordenc = CoordinateEncoder().to(device)
coorddata = CoordinateData()
if os.path.exists(path):
coordenc.load_state_dict(torch.load(path, map_location=device))
MSE = torch.nn.MSELoss(reduce=False, size_average=False).to(device)
iter = int(1e6)
iter_save = 1e4
lr = 5e-4
optimizer = torch.optim.Adam(coordenc.parameters(), lr=lr)
for i in range(iter):
points, bitmap = coorddata.nextBatch()
pred = coordenc.forward(points.to(device))
loss = MSE(bitmap.to(device), pred)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (i + 1) % iter_save == 0:
lr *= 0.99
optimizer = torch.optim.Adam(coordenc.parameters(), lr=lr)
print('Coordinate encoder loss %f at iteration %d'
% (loss.cpu().detach().numpy(), i))
tensor2Image(pred[0, 0], 'bitmap.bmp')
tensor2Image(bitmap[0, 0], 'bitmap_gt.bmp')
torch.save(coordenc.state_dict(), path)
def train_generator(coordenc_path, gen_path, dataset_path):
tb = Threebody(dataset_path, batch_size=32)
sg = StrokeGenerator(coordenc_path=coordenc_path).to(device)
sg.coordEncoder.freeze()
if os.path.exists(gen_path):
sg.load_state_dict(torch.load(gen_path, map_location=device))
if not os.path.exists('./gen_output'):
os.mkdir('./gen_output')
MSE = torch.nn.MSELoss(reduce=False, size_average=False).to(device)
lr = 5e-5
optimizer = torch.optim.Adam(filter(
lambda p: p.requires_grad, sg.parameters()), lr=lr)
while tb.epoch < 5:
images, data, points = tb.nextBatch()
output = sg(torch.Tensor(data).to(device), torch.Tensor(points).to(device))
loss = MSE(torch.Tensor(images).to(device), output)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if tb.iteration % 400 == 0:
print('\n\rEpoch %d iteration %d loss %f'
% (tb.epoch, tb.iteration, loss.cpu().detach().numpy()), end='')
tensor2Image(output[0, 0, :, :], './gen_output/gen_%d%d.bmp' % (tb.epoch, tb.iteration))
tensor2Image(images[0, 0, :, :], './gen_output/img_%d%d.bmp' % (tb.epoch, tb.iteration))
if tb.iteration % 2000 == 0:
torch.save(sg.state_dict(), gen_path)
if tb.iteration % 5 == 0:
print('\rEpoch %d iteration %d loss %f'
% (tb.epoch, tb.iteration, loss.cpu().detach().numpy()), end='')
def train_agent_mnist(gen_path, mnist_agent_path, restrain=True):
size, channel = 256, 1
batch_size = 32
sg = StrokeGenerator().to(device)
sg.load_state_dict(torch.load(gen_path, map_location=device))
sg.freeze()
agent = Agent(size, channel)
if os.path.exists(mnist_agent_path):
agent.load_state_dict(torch.load(mnist_agent_path, map_location=device))
# MNIST preprocess
mnist_size = 28
mnist_resize = 120
brightness = 0.6
trans = transforms.Compose(transforms = [
transforms.Resize(mnist_resize),
transforms.Pad(int((256 - mnist_resize) / 2)),
transforms.ToTensor(),
lambda x: x * brightness
])
train_data = torchvision.datasets.MNIST(root='./dataset/mnist',train=True,
transform=trans, download=True)
data_loader = torch.utils.data.DataLoader(train_data,
batch_size=batch_size, shuffle=True)
# render program, could also be deployed locally at http://localhost:3000
renderer = Renderer('http://10.11.6.118:3000', size)
MSE = torch.nn.MSELoss(reduce=False, size_average=False).to(device)
mse = torch.nn.MSELoss(reduce=False, size_average=False).to(device)
lr = 1e-4
LAMBDA = 2e2
optimizer = torch.optim.Adam(agent.parameters(), lr=lr)
regularizer = torch.tensor([[1, 1, 1, 0.1]] * batch_size).to(device)
for epoch in range(10):
for i, (images, labels) in enumerate(data_loader):
images = images.to(device)
color_radius, action = agent(images)
if restrain:
color_radius = color_radius * regularizer
approx = sg(color_radius[:, 2:4], action)
penalty = mse(action[:, 0:15], action[:, 1:16])
loss = MSE(images, approx) + penalty * LAMBDA
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % 200 == 0:
tensor2Image(approx[0, 0], './agent_output/%d_approx.bmp' % i)
tensor2Image(images[0, 0], './agent_output/%d_mnist.bmp' % i)
data = color_radius[0].cpu().detach().numpy()
points = action[0].cpu().detach().numpy().tolist()
data = [1.0 - data[2]] * 3 + [data[3]]
image = renderer.render(data, points)
image.save('./agent_output/%d_render.png' % i)
f = open('./agent_output/%d_data.txt' % i, 'w')
f.write(str([data] + points))
f.close()
print('Iteration %d loss %f' % (i, loss.cpu().detach().numpy()))
print('Epoch ', epoch)
torch.save(agent.state_dict(), mnist_agent_path)
def train_recurrent_agent_mnist(gen_path, agent_path):
batch_size = 32
rnn_steps = 3
# MNIST preprocess
mnist_size = 28
mnist_resize = 120
brightness = 0.6
trans = transforms.Compose(transforms = [
transforms.Resize(mnist_resize),
transforms.Pad(int((256 - mnist_resize) / 2)),
transforms.ToTensor(),
lambda x: x * brightness
])
train_data = torchvision.datasets.MNIST(root='./dataset/mnist',train=True,
transform=trans, download=True)
data_loader = torch.utils.data.DataLoader(train_data,
batch_size=batch_size, shuffle=True)
# model definition and hyper-parameters
regularizer = torch.tensor([1, 1, 1, 0.1]).to(device)
# recurrent generator records and composite the canvas at each step
rg = RecurrentGenerator(1, gen_path).to(device)
# recurrent agent reads data from rg at each step
ra = RecurrentAgent(256, 1, rg, max_steps=rnn_steps).to(device)
if os.path.exists(agent_path):
ra.load_state_dict(torch.load(agent_path, map_location=device))
MSE = torch.nn.MSELoss(reduce=False, size_average=False).to(device)
lr = 1e-4
LAMBDA = 2e2
optimizer = torch.optim.Adam(filter(
lambda p: p.requires_grad, ra.parameters()), lr=lr)
for epoch in range(10):
for i, (images, labels) in enumerate(data_loader):
images = images.to(device)
approx, penalty = ra(images, regularizer=regularizer)
loss = MSE(images, approx) + LAMBDA * penalty
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % 200 == 0:
tensor2Image(approx[0, 0], './agent_output/%d_approx.bmp' % i)
tensor2Image(images[0, 0], './agent_output/%d_mnist.bmp' % i)
saveData(ra.steps, rnn_steps, './agent_output/%d_data.txt' % i)
print('\rIteration %d loss %f' % (i, loss.cpu().detach().numpy()), end='')
torch.save(ra.state_dict(), agent_path)
# train_coord_encoder('./model/coordenc.pkl')
# train_generator('./model/coordenc.pkl', './model/gen.pkl', './dataset/3')
# train_agent_mnist('./model/gen.pkl', './model/mnist_agent.pkl')
train_recurrent_agent_mnist('./model/gen.pkl', './model/recurrent_mnist_agent.pkl')