File size: 7,690 Bytes
ae37ff6 |
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 torch as th
import torch.nn.functional as F
import torch.nn as nn
import lightning as ltn
import argparse
import lightning.pytorch as pl
from torch import Tensor
from torch import nn
from lightning.pytorch.callbacks.early_stopping import EarlyStopping
parser = argparse.ArgumentParser()
parser.add_argument("-n", "--n_epochs", type=int, default=200, help="number of epochs of training")
parser.add_argument("-b", "--batch", type=int, default=256, help="batch size of training")
parser.add_argument("-m", "--model", type=str, default='mnist0', help="model to execute")
opt = parser.parse_args()
if th.cuda.is_available():
accelerator = 'gpu'
elif th.backends.mps.is_available():
accelerator = 'cpu'
else:
accelerator = 'cpu'
class OptAEGV1(nn.Module):
def __init__(self, points=11):
super().__init__()
self.points = points
self.iscale = nn.Parameter(th.normal(0, 1, (1, 1, 1, 1)))
self.oscale = nn.Parameter(th.normal(0, 1, (1, 1, 1, 1)))
self.theta = th.linspace(-th.pi, th.pi, points)
self.velocity = th.linspace(0, th.e, points)
self.weight = nn.Parameter(th.normal(0, 1, (points, points)))
@th.compile
def integral(self, param, index):
return th.sum(param[index].view(-1, 1) * th.softmax(self.weight, dim=1)[index, :], dim=1)
@th.compile
def interplot(self, param, index):
lmt = param.size(0) - 1
p0 = index.floor().long()
p1 = p0 + 1
pos = index - p0
p0 = p0.clamp(0, lmt)
p1 = p1.clamp(0, lmt)
v0 = self.integral(param, p0)
v1 = self.integral(param, p1)
return (1 - pos) * v0 + pos * v1
@th.compile
def forward(self, data: Tensor) -> Tensor:
if self.theta.device != data.device:
self.theta = self.theta.to(data.device)
self.velocity = self.velocity.to(data.device)
shape = data.size()
data = (data - data.mean()) / data.std() * self.iscale
data = data.flatten(0)
theta = self.interplot(self.theta, th.sigmoid(data) * (self.points - 1))
ds = self.interplot(self.velocity, th.abs(th.tanh(data) * (self.points - 1)))
dx = ds * th.cos(theta)
dy = ds * th.sin(theta)
data = data * th.exp(dy) + dx
data = (data - data.mean()) / data.std() * self.oscale
return data.view(*shape)
class MNISTModel(ltn.LightningModule):
def __init__(self):
super().__init__()
self.learning_rate = 1e-3
self.counter = 0
self.labeled_loss = 0
self.labeled_correct = 0
def configure_optimizers(self):
optimizer = th.optim.Adam(self.parameters(), lr=self.learning_rate)
scheduler = th.optim.lr_scheduler.CosineAnnealingLR(optimizer, 37)
return [optimizer], [scheduler]
def training_step(self, train_batch, batch_idx):
x, y = train_batch
x = x.view(-1, 1, 28, 28)
z = self.forward(x)
loss = F.nll_loss(z, y)
self.log('train_loss', loss, prog_bar=True)
return loss
def validation_step(self, val_batch, batch_idx):
x, y = val_batch
x = x.view(-1, 1, 28, 28)
z = self.forward(x)
loss = F.nll_loss(z, y)
self.log('val_loss', loss, prog_bar=True)
pred = z.data.max(1, keepdim=True)[1]
correct = pred.eq(y.data.view_as(pred)).sum() / y.size()[0]
self.log('correct_rate', correct, prog_bar=True)
self.labeled_loss += loss.item() * y.size()[0]
self.labeled_correct += correct.item() * y.size()[0]
self.counter += y.size()[0]
def test_step(self, test_batch, batch_idx):
x, y = test_batch
x = x.view(-1, 1, 28, 28)
z = self(x)
pred = z.data.max(1, keepdim=True)[1]
correct = pred.eq(y.data.view_as(pred)).sum() / y.size()[0]
self.log('correct_rate', correct, prog_bar=True)
def on_save_checkpoint(self, checkpoint) -> None:
import glob, os
correct = self.labeled_correct / self.counter
loss = self.labeled_loss / self.counter
record = '%2.5f-%03d-%1.5f.ckpt' % (correct, checkpoint['epoch'], loss)
fname = 'best-%s' % record
with open(fname, 'bw') as f:
th.save(checkpoint, f)
for ix, ckpt in enumerate(sorted(glob.glob('best-*.ckpt'), reverse=True)):
if ix > 5:
os.unlink(ckpt)
self.counter = 0
self.labeled_loss = 0
self.labeled_correct = 0
print()
class MNIST_OptAEGV1(MNISTModel):
def __init__(self):
super().__init__()
self.pool = nn.MaxPool2d(2)
self.conv0 = nn.Conv2d(1, 2, kernel_size=7, padding=3, bias=False)
self.lnon0 = OptAEGV1()
self.conv1 = nn.Conv2d(2, 2, kernel_size=7, padding=3, bias=False)
self.lnon1 = OptAEGV1()
self.conv2 = nn.Conv2d(2, 2, kernel_size=7, padding=3, bias=False)
self.lnon2 = OptAEGV1()
self.conv3 = nn.Conv2d(2, 2, kernel_size=7, padding=3, bias=False)
self.lnon3 = OptAEGV1()
self.fc = nn.Linear(2 * 3 * 3, 10)
self.lnon4 = OptAEGV1()
def forward(self, x):
x = self.conv0(x)
x = self.lnon0(x)
x = self.pool(x)
x = self.conv1(x)
x = self.lnon1(x)
x = self.pool(x)
x = self.conv2(x)
x = self.lnon2(x)
x = self.pool(x)
x = th.flatten(x, 1)
x = self.fc(x)
x = self.lnon4(x)
x = F.log_softmax(x, dim=1)
return x
def test_best():
import glob
fname = sorted(glob.glob('best-*.ckpt'), reverse=True)[0]
with open(fname, 'rb') as f:
checkpoint = th.load(f)
model.load_state_dict(checkpoint['state_dict'], strict=False)
model.eval()
with th.no_grad():
counter, success = 0, 0
for test_batch in test_loader:
x, y = test_batch
x = x.view(-1, 1, 28, 28)
z = model(x)
pred = z.data.max(1, keepdim=True)[1]
correct = pred.eq(y.data.view_as(pred)).sum() / y.size()[0]
print('.', end='', flush=True)
if counter % 100 == 0:
print('')
success += correct.item()
counter += 1
print('')
print('Accuracy: %2.5f' % (success / counter))
th.save(model, 'mnist-optaeg-v1.pt')
if __name__ == '__main__':
print('loading data...')
from torch.utils.data import DataLoader
from torchvision.datasets import MNIST
from torchvision import transforms
mnist_train = MNIST('datasets', train=True, download=True, transform=transforms.Compose([
transforms.ToTensor(),
]))
mnist_test = MNIST('datasets', train=False, download=True, transform=transforms.Compose([
transforms.ToTensor(),
]))
train_loader = DataLoader(mnist_train, shuffle=True, batch_size=opt.batch, num_workers=8)
val_loader = DataLoader(mnist_test, batch_size=opt.batch, num_workers=8)
test_loader = DataLoader(mnist_test, batch_size=opt.batch, num_workers=8)
# training
print('construct trainer...')
trainer = pl.Trainer(accelerator=accelerator, precision=32, max_epochs=opt.n_epochs,
callbacks=[EarlyStopping(monitor="correct_rate", mode="max", patience=30)])
print('construct model...')
model = MNIST_OptAEGV1()
print('training...')
trainer.fit(model, train_loader, val_loader)
print('testing...')
test_best()
|