-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathNP_or_ANP_train_2d.py
167 lines (151 loc) · 7.87 KB
/
NP_or_ANP_train_2d.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
from data.Image_data_sampler import ImageReader
from module.NP import NeuralProcess as NP
from module.utils import compute_loss, comput_kl_loss, to_numpy, load_plot_data, img_mask_to_np_input, generate_mask, np_input_to_img
import torch
from torch.utils.tensorboard import SummaryWriter
import numpy
from tqdm import tqdm
import time
import matplotlib.pyplot as plt
import os
from torchmeta.datasets import MiniImagenet
from torchmeta.transforms import ClassSplitter
from torchmeta.utils.data import BatchMetaDataLoader
from torchvision.transforms import Compose, Resize, ToTensor
def validation(data_test, model):
total_ll = 0
model.eval()
for i, (img, _) in tqdm(enumerate(data_test)):
context_mask, target_mask = generate_mask(img)
x_context, y_context, x_target, y_target = img_mask_to_np_input(img, context_mask, target_mask, \
include_context=False)
(mean, var), _, _ = model(x_context.to(device), y_context.to(device), x_target.to(device))
loss = compute_loss(mean, var, y_target.to(device))
total_ll += -loss.item()
return total_ll / (i+1)
def validation_meta(data_test, model, VAL_ITER=10):
total_ll = 0
model.eval()
for i, batch in tqdm(enumerate(data_test)):
if i == VAL_ITER :
break
img, _ = batch["test"]
bs, n_shot, c, w, h = img.shape
img = torch.reshape(img, (bs*n_shot, c, w, h))
context_mask, target_mask = generate_mask(img)
x_context, y_context, x_target, y_target = img_mask_to_np_input(img, context_mask, target_mask, \
include_context=False)
(mean, var), _, _ = model(x_context.to(device), y_context.to(device), x_target.to(device))
loss = compute_loss(mean, var, y_target.to(device))
total_ll += -loss.item()
return total_ll / (i + 1)
def save_plot(epoch, data, model, imgsize):
imgsize = list(imgsize)
imgsize[0] = 1 # change batch_size = 1
ax, fig = plt.subplots()
(x_context, y_context) , x_target = data.query
(mean, var), _, _ = model(x_context.to(device), y_context.to(device), x_target.to(device))
# recover prediction
mean_recover = np_input_to_img(x_target, mean.cpu(), imgsize)
var_recover = np_input_to_img(x_target, var.cpu(), imgsize)
# recover x_target
target_recover = np_input_to_img(x_target, data.y_target, imgsize)
# recover x_context
context_recover = np_input_to_img(x_context, y_context, imgsize)
mean_recover += context_recover # fill in context data with prediction
raw_image = target_recover + context_recover # recover raw image
img_recover = torch.cat([raw_image, mean_recover, var_recover], dim=3)
plt.title("epoch:%d" % epoch)
# last squeeze for removing color channel for gray scale image
plt.imshow(to_numpy(img_recover.squeeze(0).permute(1,2,0).squeeze()))
model_path = "saved_fig/" + MODELNAME
kernel_path = os.path.join(model_path, kernel)
if not os.path.exists(model_path):
os.mkdir(model_path)
if not os.path.exists(kernel_path):
os.mkdir(kernel_path)
plt.savefig("saved_fig/"+MODELNAME+"/"+kernel+"/"+"%03d"%(epoch)+".png")
plt.close()
return fig
def main_nonmeta():
TRAINING_ITERATIONS = int(200)
BEST_LOSS = -numpy.inf
MODELNAME = 'NP' # 'NP' or 'ANP'
kernel = 'MNIST' # EQ/ period / MNIST/ SVHN / celebA
# set up tensorboard
time_stamp = time.strftime("%m-%d-%Y_%H:%M:%S", time.localtime())
writer = SummaryWriter('runs/' + kernel + '_' + MODELNAME + '_' + time_stamp)
# load data set
dataset = ImageReader(dataset=kernel, batch_size=64, datapath='/share/scratch/xuesongwang/metadata/')
# load plot dataset for recording training progress
# plot_data = save_plot_data(dataset.valloader, kernel) # generate and save data for the first run
plot_data = load_plot_data(kernel)
np = NP(input_dim=2, latent_dim=128, output_dim=3 if kernel != 'MNIST' else 1, use_attention=MODELNAME == 'ANP').to(
device)
optim = torch.optim.Adam(np.parameters(), lr=3e-4, weight_decay=1e-5)
for epoch in tqdm(range(TRAINING_ITERATIONS)):
for i, (img, _) in tqdm(enumerate(dataset.trainloader)):
context_mask, target_mask = generate_mask(img)
x_context, y_context, x_target, y_target = img_mask_to_np_input(img, context_mask, target_mask, \
include_context=True)
(mean, var), prior, poster = np(x_context.to(device), y_context.to(device), x_target.to(device),
y_target.to(device))
nll_loss = compute_loss(mean, var, y_target.to(device))
kl_loss = comput_kl_loss(prior, poster)
loss = nll_loss + kl_loss
optim.zero_grad()
loss.backward()
optim.step()
writer.add_scalars("Image Log-likelihood", {"train/overall": -loss.item(),
"train/ll": -nll_loss.item(),
"train/kl": kl_loss.item()}, epoch)
val_loss = validation(dataset.valloader, np)
save_plot(epoch, plot_data, np, img.shape)
writer.add_scalars("Image Log-likelihood", {"val": val_loss}, epoch)
if val_loss > BEST_LOSS:
BEST_LOSS = val_loss
print("save module at epoch: %d, val log-likelihood: %.4f" % (epoch, val_loss))
torch.save(np.state_dict(), 'saved_model/' + kernel + '_' + MODELNAME + '.pt')
writer.close()
print("finished training: " + MODELNAME)
def main_meta():
data_name = 'miniImagenet'
MODELNAME = 'ANP' # 'NP' or 'ANP'
dataset = MiniImagenet("/share/scratch/xuesongwang/metadata/",
num_classes_per_task=2,
transform=Compose([Resize(32), ToTensor()]),
meta_train=True,
download=False)
dataset = ClassSplitter(dataset, shuffle=True, num_train_per_class=15, num_test_per_class=15)
dataloader = BatchMetaDataLoader(dataset, batch_size=4, num_workers=8)
np = NP(input_dim=2, latent_dim=128, output_dim=3, use_attention=MODELNAME == 'ANP').to(device)
optim = torch.optim.Adam(np.parameters(), lr=3e-4, weight_decay=1e-5)
TRAINING_ITERATIONS = int(500)
BEST_LOSS = -numpy.inf
for epoch in range(TRAINING_ITERATIONS):
for i, batch in tqdm(enumerate(dataloader)):
img, _ = batch["train"]
bs, n_shot, c, w, h = img.shape
img = torch.reshape(img, (bs * n_shot, c, w, h))
context_mask, target_mask = generate_mask(img)
x_context, y_context, x_target, y_target = img_mask_to_np_input(img, context_mask, target_mask, \
include_context=False)
(mean, var), prior, poster = np(x_context.to(device), y_context.to(device), x_target.to(device),
y_target.to(device))
nll_loss = compute_loss(mean, var, y_target.to(device))
kl_loss = comput_kl_loss(prior, poster)
loss = nll_loss + kl_loss
optim.zero_grad()
loss.backward()
optim.step()
val_loss = validation_meta(dataloader, np, VAL_ITER=50 if MODELNAME == 'ANP' else 10)
if val_loss > BEST_LOSS:
BEST_LOSS = val_loss
print("save module at epoch: %d, val log-likelihood: %.4f" % (epoch, val_loss))
torch.save(np.state_dict(), 'saved_model/' + data_name + '_' + MODELNAME + '.pt')
print("finished training " + MODELNAME +' '+ data_name+'!')
if __name__ == '__main__':
# define hyper parameters
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# main_nonmeta()
main_meta()