问题

I'm trying to run a pretrained pytorch InceptionV3 model on my own data.
我正在尝试在自己的数据上运行预训练的pytorch InceptionV3模型。

The data size is (N,N).
数据大小为(N,N)。

The batch size is 128.
批大小为128。

InceptionV3 requires (batch_size, 3, 229, 229) as input.
InceptionV3要求输入为(batch_size, 3, 229, 229)。

I tried to resize using the torch.transforms Resize function:
我尝试使用torch.transforms的Resize函数进行调整大小：

train_transforms = transforms.Compose([transforms.Resize((229, 229)),
                                       transforms.ToTensor()])

which did not work.
但是没有成功。

I also tried
我还尝试过：

train_transforms = transforms.Compose([transforms.Resize((3, 229, 229)),
                                       transforms.ToTensor()])

which also didn't work.
这也没有成功。

I'm already defining the batch size using a dataloader:
我已经在使用数据加载器定义了批大小：

train_loader = DataLoader(train_dataset, batch_size=128)

Also saw a suggestion here, but I do not understand how this number of neurons is calculated.
我也看到了一个建议，但我不理解如何计算神经元的数量。
https://github.com/pytorch/vision/issues/396

train_transforms = transforms.Compose([transforms.Resize((229, 229)),
                                       transforms.ToTensor()])

train_transforms = transforms.Compose([transforms.Resize((3, 229, 229)),
                                       transforms.ToTensor()])

train_loader = DataLoader(train_dataset, batch_size=128)

答案1

得分: 0

It seems that 299 is for evaluation, for training you need a bigger size (due to aux classifiers).

from torchvision import models
from torchvision import transforms
import numpy as np
from PIL import Image
from torch import nn

model = models.inception_v3(pretrained=True)
model.eval()
w = 299
inputs = np.random.randint(low=0, high=255, size=(512, 512, 3)).astype(np.uint8)
img = Image.fromarray(inputs)
train_transforms = transforms.Compose([transforms.Resize((w, w)),
                                        transforms.ToTensor()])

inputs = train_transforms(img)
print(inputs.shape)
model(inputs[None, ...])[0].shape

So in training, use a bigger size, for example, 350.

from torchvision import models
from torchvision import transforms
import numpy as np
from PIL import Image
from torch import nn

model = models.inception_v3(pretrained=True)
w = 350
inputs = np.random.randint(low=0, high=255, size=(512, 512, 3)).astype(np.uint8)
img = Image.fromarray(inputs)
train_transforms = transforms.Compose([transforms.Resize((w, w)),
                                       transforms.ToTensor()])

inputs = train_transforms(img)
print(inputs.shape)
model(inputs[None, ...])[0].shape

from torchvision import models
from torchvision import transforms
import numpy as np
from PIL import Image
from torch import nn

model = models.inception_v3(pretrained=True)
model.eval()
w = 299
inputs = np.random.randint(low=0, high=255, size=(512, 512, 3)).astype(np.uint8)
img = Image.fromarray(inputs)
train_transforms = transforms.Compose([transforms.Resize((w, w)),
                                        transforms.ToTensor()])

inputs = train_transforms(img)
print(inputs.shape)
model(inputs[None, ...])[0].shape

from torchvision import models
from torchvision import transforms
import numpy as np
from PIL import Image
from torch import nn

model = models.inception_v3(pretrained=True)
w = 350
inputs = np.random.randint(low=0, high=255, size=(512, 512, 3)).astype(np.uint8)
img = Image.fromarray(inputs)
train_transforms = transforms.Compose([transforms.Resize((w, w)),
                                       transforms.ToTensor()])

inputs = train_transforms(img)
print(inputs.shape)
model(inputs[None, ...])[0].shape