获取预训练模型的网络输入尺寸

初学神经网络之时，常常会用到预训练的网络包。
例如

from torchvision import models
resnet = models.res18(pretrained=True)

但是初学者在使用的时候会犯难：我需要输入多大尺寸的图片呢？
解决方案：

方法一：读torchvision.models的说明文档

打开torchvision.models的网站：
https://pytorch.org/hub/research-models

搜索你需要的模型名称，得到resnet的网站：
https://pytorch.org/hub/pytorch_vision_resnet/
在里面看到resnet的指导文档：

import torch
model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet18', pretrained=True)
# or any of these variants
# model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet34', pretrained=True)
# model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet50', pretrained=True)
# model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet101', pretrained=True)
# model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet152', pretrained=True)
model.eval()

All pre-trained models expect input images normalized in the same way, i.e. mini-batches of 3-channel RGB images of shape (3 x H x W), where H and W are expected to be at least 224. The images have to be loaded in to a range of [0, 1] and then normalized using mean = [0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225].
Here’s a sample execution.

# sample execution (requires torchvision)
from PIL import Image
from torchvision import transforms
input_image = Image.open(filename)
preprocess = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
input_tensor = preprocess(input_image)
input_batch = input_tensor.unsqueeze(0) # create a mini-batch as expected by the model

# move the input and model to GPU for speed if available
if torch.cuda.is_available():
    input_batch = input_batch.to('cuda')
    model.to('cuda')

with torch.no_grad():
    output = model(input_batch)
# Tensor of shape 1000, with confidence scores over Imagenet's 1000 classes
print(output[0])
# The output has unnormalized scores. To get probabilities, you can run a softmax on it.
probabilities = torch.nn.functional.softmax(output[0], dim=0)
print(probabilities)

从中，可以了解到，需要将图片缩放到尺寸为[224, 224]。

方法二：读GitHub源代码

除了看torchvision的指导文档，也可以进入github看源码中的注释(View on GitHub)：

Github代码中包括如下内容：

class ResNet18_Weights(WeightsEnum):
    IMAGENET1K_V1 = Weights(
        url="https://download.pytorch.org/models/resnet18-f37072fd.pth",
        transforms=partial(ImageClassification, crop_size=224),
        meta={
    
            **_COMMON_META,
            "num_params": 11689512,
            "recipe": "https://github.com/pytorch/vision/tree/main/references/classification#resnet",
            "_metrics": {
    
                "ImageNet-1K": {
    
                    "[email protected]": 69.758,
                    "[email protected]": 89.078,
                }
            },
            "_docs": """These weights reproduce closely the results of the paper using a simple training recipe.""",
        },
    )
    DEFAULT = IMAGENET1K_V1

由其中的crop_size=224了解到应该缩放或者裁减为224大小。

方法三：暴力测试

尺寸设计为自己觉得可行的大小，在一定的区间内进行循环，如果尺寸不合适，Pytorch会报错说模型尺寸不相容。用except跳过这些异常，保留try成功的尺寸。

transfer_model = resnet18(pretrained=True)
print(transfer_model)

transfer_model.eval()
transfer_model = transfer_model.cuda()

batch_size = 16

ok_list = []
for length in range(1, 1000):
    x = torch.zeros(batch_size, 3, length, length, requires_grad=False).cuda()
    print('length = {}'.format(length))
    try:
        torch_out = transfer_model(x)
        print('length = {} OK!!!!'.format(length))
        ok_list.append(length)
        # break
    except:
        continue

print(ok_list)

硬试，试试输入尺寸是多少的时候，这个网络模型能跑通，跑的通再逐步测试。注意，由于神经网络里面有一些下采样的取整操作，所以尺寸在某些区间范围内，比如刚刚这段代码输出的结果是：

...
length = 998
length = 999
[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]

尺寸224只是可以接受的大小其中之一，也是设计者的初衷尺寸。