-
Notifications
You must be signed in to change notification settings - Fork 5
/
jointman.py
234 lines (187 loc) · 10.1 KB
/
jointman.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
234
import numpy as np
import argparse, json, math
import torch, torchvision
from torch.utils.data import Dataset, DataLoader
from torch import nn
import matplotlib.pyplot as plt
import flow, source, train, utils
parser = argparse.ArgumentParser(description="")
group = parser.add_argument_group("Target Parameters")
group.add_argument('-target', type=str, default='CIFAR', choices=['CIFAR', 'ImageNet32', 'ImageNet64', 'MNIST'], metavar='DATASET', help='Dataset choice.')
group = parser.add_argument_group("Architecture Parameters")
group.add_argument("-repeat", type=int, default=1, help="num of wavelet layers of each scale")
group.add_argument("-hchnl", type=int, default=12, help="intermediate channel dimension of Conv2d inside NICE inside NeuralWavelet")
group.add_argument("-nhidden", type=int, default=1, help="num of intermediate channel of Conv2d inside NICE inside NeuralWavelet")
group.add_argument("-nMixing", type=int, default=5, help="num of mixing distributions of last sub-priors")
group.add_argument("-simplePrior", action="store_true", help="if use simple version prior, no crossover net")
group.add_argument("-HUE", action="store_false", help="use YCbCr color scheme")
group.add_argument("-clamp", type=float, default=-1, help="clamp of last prior's mean")
group = parser.add_argument_group('Learning parameters')
group.add_argument("-epoch", type=int, default=400, help="num of epoches to train")
group.add_argument("-batch", type=int, default=200, help="batch size")
group.add_argument("-savePeriod", type=int, default=10, help="save after how many steps")
group.add_argument("-lr", type=float, default=0.001, help="learning rate")
group.add_argument("-decay", type=float, default=0.99, help="learning rate decay")
group = parser.add_argument_group("Etc")
group.add_argument("-folder", default=None, help="Path to save")
group.add_argument("-cuda", type=int, default=-1, help="Which device to use with -1 standing for CPU, number bigger than -1 is N.O. of GPU.")
group.add_argument("-load", action='store_true', help="If load or not")
args = parser.parse_args()
device = torch.device("cpu" if args.cuda < 0 else "cuda:" + str(args.cuda))
# Creating save folder
if args.folder is None:
rootFolder = './opt/default_easyMera_' + 'joinData' + "_YCC_" + str(args.HUE) + "_simplePrior_" + str(args.simplePrior) + "_repeat_" + str(args.repeat) + "_hchnl_" + str(args.hchnl) + "_nhidden_" + str(args.nhidden) + "_nMixing_" + str(args.nMixing) + "_clamp_" + str(args.clamp) + "/"
print("No specified saving path, using", rootFolder)
else:
rootFolder = args.folder
if rootFolder[-1] != '/':
rootFolder += '/'
utils.createWorkSpace(rootFolder)
if args.clamp < 0:
args.clamp = None
# Decoding parameters to mem, saving them to save folder.
if not args.load:
repeat = args.repeat
hchnl = args.hchnl
nhidden = args.nhidden
nMixing = args.nMixing
epoch = args.epoch
batch = args.batch
savePeriod = args.savePeriod
simplePrior = args.simplePrior
clamp = args.clamp
lr = args.lr
HUE = args.HUE
with open(rootFolder + "/parameter.json", "w") as f:
config = {'target': 'join', 'repeat': repeat, 'hchnl': hchnl, 'nhidden': nhidden, 'nMixing': nMixing, 'epoch': epoch, 'batch': batch, 'savePeriod': savePeriod, 'lr': lr, 'simplePrior': simplePrior, 'clamp': clamp, 'HUE': HUE}
json.dump(config, f)
else:
# load saved parameters, and decoding them to mem
with open(rootFolder + "/parameter.json", 'r') as f:
config = json.load(f)
locals().update(config)
if HUE:
lambd = lambda x: (x * 255).byte().to(torch.float32).to(device)
else:
lambd = lambda x: utils.rgb2ycc((x * 255).byte().float(), True).to(torch.float32).to(device)
#lambd = lambda x: (x * 255).byte().to(torch.float32).to(device)
trainsetTransform = torchvision.transforms.Compose([torchvision.transforms.ToTensor(), torchvision.transforms.Lambda(lambd)])
trainTarget0 = torchvision.datasets.CIFAR10(root='./data/cifar', train=True, download=True, transform=trainsetTransform)
testTarget0 = torchvision.datasets.CIFAR10(root='./data/cifar', train=False, download=True, transform=trainsetTransform)
targetTrainLoader0 = torch.utils.data.DataLoader(trainTarget0, batch_size=batch, shuffle=True)
targetTestLoader0 = torch.utils.data.DataLoader(testTarget0, batch_size=batch, shuffle=False)
trainTarget1 = utils.ImageNet(root='./data/ImageNet32', train=True, download=True, transform=trainsetTransform)
testTarget1 = utils.ImageNet(root='./data/ImageNet32', train=False, download=True, transform=trainsetTransform)
targetTrainLoader1 = torch.utils.data.DataLoader(trainTarget1, batch_size=batch, shuffle=True)
targetTestLoader1 = torch.utils.data.DataLoader(testTarget1, batch_size=batch, shuffle=False)
trainTarget2 = utils.ImageNet(root='./data/ImageNet64', train=True, download=True, transform=trainsetTransform, d64=True)
testTarget2 = utils.ImageNet(root='./data/ImageNet64', train=False, download=True, transform=trainsetTransform, d64=True)
targetTrainLoader2 = torch.utils.data.DataLoader(trainTarget2, batch_size=batch, shuffle=True)
targetTestLoader2 = torch.utils.data.DataLoader(testTarget2, batch_size=batch, shuffle=False)
class JointData(object):
def __init__(self, datas, sizes, batch):
self.datas = datas
self.iters = [iter(term) for term in datas]
self.sizes = np.ceil(np.array(sizes) / batch)
self.n = [0, 0, 0]
def __iter__(self):
self.iters = [iter(term) for term in self.datas]
self.n = [0, 0, 0]
return self
def __next__(self):
probs = self.sizes - np.array(self.n)
if np.allclose(probs, np.zeros(probs.shape)):
self.iters = [iter(term) for term in self.datas]
self.n = [0, 0, 0]
raise StopIteration
probs = probs / np.sum(probs)
no = np.argmax(np.random.multinomial(1, probs))
self.n[no] += 1
samples, labels = next(self.iters[no])
return samples, labels
'''
class JointData(Dataset):
def __init__(self, datas, sizes, batch):
self.datas = [term for term in datas]
self.sizes = np.ceil(np.array(sizes) / batch)
self.n = [0, 0, 0]
def __len__(self):
return np.sum(self.sizes)
def __getitem__(self, index):
assert index == np.sum(self.n)
probs = self.sizes - np.array(self.n)
probs = probs / np.sum(probs)
no = np.argmax(np.random.multinomial(1, probs))
self.n[no] += 1
samples, labels = next(self.datas[no])
if index == self.__len__() - 1:
self.n = [0, 0, 0]
return samples, labels
'''
joinTargetTrainLoader = JointData([targetTrainLoader0, targetTrainLoader1, targetTrainLoader2], [len(trainTarget0), len(trainTarget1), len(trainTarget2)], batch)
joinTargetTestLoader = JointData([targetTestLoader0, targetTestLoader1, targetTestLoader2], [len(testTarget0), len(testTarget1), len(testTarget2)], batch)
targetSize = [3, 64, 64]
dimensional = 2
channel = targetSize[0]
blockLength = targetSize[-1]
decimal = flow.ScalingNshifting(256, -128)
rounding = utils.roundingWidentityGradient
depth = int(math.log(targetSize[-1], 2))
def buildLayers(shapeList):
layers = []
for no, chn in enumerate(shapeList[:-1]):
if no != 0 and no != len(shapeList) - 2:
layers.append(torch.nn.Conv2d(chn, shapeList[no + 1], 1))
else:
layers.append(torch.nn.Conv2d(chn, shapeList[no + 1], 3, padding=1, padding_mode="replicate"))
if no != len(shapeList) - 2:
layers.append(torch.nn.ReLU(inplace=True))
return layers
layerList = []
shapeList = [targetSize[0] * 3] + [hchnl] * (nhidden + 1) + [targetSize[0]]
for i in range(4 * repeat):
layers = buildLayers(shapeList)
layerList.append(torch.nn.Sequential(*layers))
#layerList.append(torch.nn.Sequential(torch.nn.Conv2d(9, hchnl, 3, padding=1), torch.nn.ReLU(inplace=True), torch.nn.Conv2d(hchnl, hchnl, 1, padding=0), torch.nn.ReLU(inplace=True), torch.nn.Conv2d(hchnl, 3, 3, padding=1)))
torch.nn.init.zeros_(layerList[-1][-1].weight)
torch.nn.init.zeros_(layerList[-1][-1].bias)
layerList = layerList * depth
shapeList = [targetSize[0]] + [hchnl] * (nhidden + 1) + [targetSize[0] * 3]
if not simplePrior:
meanNNlist = []
scaleNNlist = []
layers = buildLayers(shapeList)
meanNNlist.append(torch.nn.Sequential(*layers))
#meanNNlist.append(torch.nn.Sequential(torch.nn.Conv2d(3, hchnl, 3, padding=1), torch.nn.ReLU(inplace=True), torch.nn.Conv2d(hchnl, hchnl, 1, padding=0), torch.nn.ReLU(inplace=True), torch.nn.Conv2d(hchnl, 9, 3, padding=1)))
layers = buildLayers(shapeList)
scaleNNlist.append(torch.nn.Sequential(*layers))
#scaleNNlist.append(torch.nn.Sequential(torch.nn.Conv2d(3, hchnl, 3, padding=1), torch.nn.ReLU(inplace=True), torch.nn.Conv2d(hchnl, hchnl, 1, padding=0), torch.nn.ReLU(inplace=True), torch.nn.Conv2d(hchnl, 9, 3, padding=1)))
torch.nn.init.zeros_(meanNNlist[-1][-1].weight)
torch.nn.init.zeros_(meanNNlist[-1][-1].bias)
torch.nn.init.zeros_(scaleNNlist[-1][-1].weight)
torch.nn.init.zeros_(scaleNNlist[-1][-1].bias)
meanNNlist = meanNNlist * depth
scaleNNlist = scaleNNlist * depth
else:
meanNNlist = None
scaleNNlist = None
# Building MERA model
f = flow.SimpleMERA(blockLength, layerList, meanNNlist, scaleNNlist, repeat, None, nMixing, decimal=decimal, rounding=utils.roundingWidentityGradient, clamp=clamp, compatible=True).to(device)
# Define plot function
def plotfn(f, train, test, LOSS, VALLOSS):
# loss plot
lossfig = plt.figure(figsize=(8, 5))
lossax = lossfig.add_subplot(111)
epoch = len(LOSS)
lossax.plot(np.arange(epoch), np.array(LOSS), 'go-', label="loss", markersize=2.5)
lossax.plot(np.arange(epoch), np.array(VALLOSS), 'ro-', label="val. loss", markersize=2.5)
lossax.set_xlim(0, epoch)
lossax.legend()
lossax.set_title("Loss Curve")
plt.savefig(rootFolder + 'pic/lossCurve.png', bbox_inches="tight", pad_inches=0)
plt.close()
# Training
f = train.forwardKLD(f, joinTargetTrainLoader, joinTargetTestLoader, epoch, lr, savePeriod, rootFolder, plotfn=plotfn, lr_decay=args.decay)
# Pasuse
import pdb
pdb.set_trace()