weiaicunzai/bag_of_tricks_for_image_classification_with_convolutional_neural_networks

experiments on Paper <Bag of Tricks for Image Classification with Convolutional Neural Networks> and other useful tricks to improve CNN acc

bag_of_tricks_for_image_classification_with_convolutional_neural_networks's Issues

Forget covert BGR to RGB using cv2?

ToCVImage just set an image as uint8, but not to convert BGR2RGB because we usually use RGB. And the normalize is RGB data.

Forget covert BGR to RGB using cv2?

What's the influence of normalization(i.e. the value of std and mean)? And, how to set these?

Tricks also apply for ResNets?

Hi,

Thanks for your implementation and repo.

I tested on ResNet for CIFAR100 and seems Label Smoothing, No-bias-decay does not improve the result.

BTW, I disable all the image augmentation for both train and test, only use normalize and purely test on above mentioned tricks.

If you also tested on ResNet, say ResNet56, and have different results, please let me know.

Really thanks.

resnet-C structure cannot decrease computational cost

Paper says 'replace a 7x7 convolution with three 3x3 convolutions'. I am confused about it.
ex:
input: 224x224 channel=3
7x7 conv pad=3 stride=2 channel=64-> output dimension: 112x112
Its flops is: 7x7x3x64x112x112

If replace 7x7 conv with three 3x3 conv, there are two methods:

1. 'the first and second convolutions have their output channel of 32 and a stride of 2, with the last convolution, uses a 64-output channel', If this is the case, the three 3x3 conv output dimension is 64x64, It is not according to with 7x7 conv output dimension. (Form figure2.b, three 3x3 conv + maxpool similar to 7x7 conv + maxpool)

2. the first 3x3 conv have a stride of 2, the second 3x3 conv have a stride of 1, It has the same output dimension as 7x7 conv, but their flops are 3x3x3x32x112x112 + 3x3x32x32x112x112 + 3x3x32x64x112x112, the flops are bigger than original 7x7 conv.

Can you give me some tutorials about it?

update:
From Table 5, the flops of ResNet-50-B is 0.3 G bigger than original ResNet-50, so it may use the above setting 2, so Is the author's description wrong?

usage: train.py [-h] -net NET [-w W] [-b B] [-lr LR] [-e E] [-warm WARM] [-gpus GPUS [GPUS ...]] train.py: error: the following arguments are required: -net

usage: train.py [-h] -net NET [-w W] [-b B] [-lr LR] [-e E] [-warm WARM]
[-gpus GPUS [GPUS ...]]
train.py: error: the following arguments are required: -net

Errors when I run train.py

File "train.py", line 165, in
predicts = net(images)
File "/home//.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home//.local/lib/python3.6/site-packages/torch/nn/parallel/data_parallel.py", line 159, in forward
return self.module(*inputs[0], **kwargs[0])
File "/home//.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home//LJY/Bag_of_Tricks_for_Image_Classification_with_Convolutional_Neural_Networks-master/models/vgg.py", line 53, in forward
x = self.classifier(x)
File "/home//.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home//.local/lib/python3.6/site-packages/torch/nn/modules/container.py", line 117, in forward
input = module(input)
File "/home//.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home//.local/lib/python3.6/site-packages/torch/nn/modules/linear.py", line 93, in forward
return F.linear(input, self.weight, self.bias)
File "/home//.local/lib/python3.6/site-packages/torch/nn/functional.py", line 1690, in linear
ret = torch.addmm(bias, input, weight.t())
RuntimeError: mat1 dim 1 must match mat2 dim 0

About learning rate

I wonder how much the value of the base learning rate you use when batch size is 256? I have tried 0.1 ( as the figure 3 of the paper suggests ) , but got a bad consequence.
I find that in your code the default value of the learning rate is 0.04, does it work?

how to use resnet?

how to use resnet?

question

Do you use "xavier init" in code? But I did not find it.Thank you !