我想在CNN中使用自定义过滤器。过滤器的大小为5 * 5，每个条目都是三个变量的函数：theta，Lambda，psi。有两个卷积层，然后是两个完全连接的层。我在MNIST数据集上测试了过滤器。但是，当我在GPU上运行它时，我遇到了错误： RuntimeError：预期的后端CUDA和dtype Float，但是却获得了后端CPU和dtype的Float。我猜这可能是由于我如何生成过滤器框，但我找不到确切的错误地方。基本上，我修改了此example code，仅使用自定义过滤器修改了网络结构。培训和测试保持不变。我在这里附上我的代码。谢谢！
from __future__ import print_function
import argparse
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
from torchvision import datasets, transforms
class Net(nn.Module):
def __init__(self, kernel_size, in_channels, channel1, channel2):
super(Net, self).__init__()
self.theta1, self.Lambda1, self.psi1, self.bias1 = self.generate_parameters(channel1, in_channels)
self.filter1 = self.whole_filter(in_channels, channel1, kernel_size, self.theta1, self.Lambda1, self.psi1).cuda()
self.theta2, self.Lambda2, self.psi2, self.bias2 = self.generate_parameters(channel2, channel1)
self.filter2 = self.whole_filter(channel1, channel2, kernel_size, self.theta2, self.Lambda2, self.psi2).cuda()
self.fc1 = nn.Linear(4*4*50, 500)
self.fc2 = nn.Linear(500, 10)
def forward(self, x):
x = F.conv2d(x, self.filter1, bias=self.bias1)
x = F.max_pool2d(x, 2, 2)
x = F.conv2d(x, self.filter2, bias=self.bias2)
x = F.max_pool2d(x, 2, 2)
x = x.view(-1, 4*4*50)
x = F.relu(self.fc1(x))
x = self.fc2(x)
return F.log_softmax(x, dim=1)
def generate_parameters(self, dim_out, dim_in):
theta = nn.Parameter(torch.randn(dim_out, dim_in))
Lambda = nn.Parameter(torch.randn(dim_out, dim_in))
psi = nn.Parameter(torch.randn(dim_out, dim_in))
bias = nn.Parameter(torch.randn(dim_out))
return theta, Lambda, psi, bias
def whole_filter(self, in_channels, out_channels, kernel_size, theta_column, Lambda_column, psi_column):
result = torch.zeros(out_channels, in_channels, kernel_size, kernel_size) # \text{out\_channels} , \frac{\text{in\_channels}}{\text{groups}} , kH , kW
for i in range(out_channels):
result[i] = self.one_filter(in_channels, kernel_size, theta_column[i], Lambda_column[i], psi_column[i])
return result
def one_filter(self, in_channels, kernel_size, theta, Lambda, psi):
result = torch.zeros(in_channels, kernel_size, kernel_size)
for i in range(in_channels):
result[i] = self.filter_fn(theta[i], Lambda[i], psi[i], kernel_size)
return result
def filter_fn(self, theta, Lambda, psi, kernel_size):
# Bounding box
half_size = (kernel_size - 1) // 2
ymin, xmin = -half_size, -half_size
ymax, xmax = half_size, half_size
(y, x) = np.meshgrid(np.arange(ymin, ymax + 1), np.arange(xmin, xmax + 1))
y = torch.FloatTensor(y)
x = torch.FloatTensor(x)
# Rotation
x_theta = x * torch.cos(theta) + y * torch.sin(theta)
y_theta = -x * torch.sin(theta) + y * torch.cos(theta)
box = torch.cos(y_theta) * torch.sin(2 * np.pi / Lambda * x_theta + psi)
return box
def train(args, model, device, train_loader, optimizer, epoch):
model.train()
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
output = model(data)
loss = F.nll_loss(output, target)
loss.backward(retain_graph=True)
optimizer.step()
if batch_idx % args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.item()))
def test(args, model, device, test_loader):
model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))
def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--batch-size', type=int, default=64, metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs', type=int, default=10, metavar='N',
help='number of epochs to train (default: 10)')
parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda', action='store_true', default=False,
help='disables CUDA training')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--log-interval', type=int, default=10, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--save-model', action='store_true', default=False,
help='For Saving the current Model')
args = parser.parse_args()
use_cuda = not args.no_cuda and torch.cuda.is_available()
torch.manual_seed(args.seed)
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=False, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.test_batch_size, shuffle=True, **kwargs)
model = Net(5, 1, 20, 50).to(device)
optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)
for param in model.parameters():
print(type(param.data), param.size())
for epoch in range(1, args.epochs + 1):
train(args, model, device, train_loader, optimizer, epoch)
test(args, model, device, test_loader)
if (args.save_model):
torch.save(model.state_dict(),"mnist_cnn.pt")
if __name__ == '__main__':
main()
答案 0 :(得分：0)
用以下内容替换filter_fn
import math
def filter_fn(self, theta, Lambda, psi, kernel_size, device):
# Bounding box
half_size = (kernel_size - 1) // 2
ymin, xmin = -half_size, -half_size
ymax, xmax = half_size, half_size
(x,y) = torch.meshgrid([torch.arange(xmin, xmax + 1).to(device), torch.arange(ymin, ymax + 1).to(device)])
# Rotation
x_theta = x * torch.cos(theta) + y * torch.sin(theta)
y_theta = -x * torch.sin(theta) + y * torch.cos(theta)
box = torch.cos(y_theta) * torch.sin(2 * math.pi / Lambda * x_theta + psi)
return box