taki0112/senet-tensorflow

Simple Tensorflow implementation of "Squeeze and Excitation Networks" using Cifar10 (ResNeXt, Inception-v4, Inception-resnet-v2)

SENet-Tensorflow

Simple Tensorflow implementation of Squeeze Excitation Networks using Cifar10

I implemented the following SENet

• ResNeXt paper
• Inception-v4, Inception-resnet-v2 paper

If you want to see the original author's code, please refer to this link

Requirements

• Tensorflow 1.x
• Python 3.x
• tflearn (If you are easy to use global average pooling, you should install tflearn)

Issue

Image_size

• In paper, experimented with ImageNet
• However, due to image size issues in Inception network, so I used zero padding for the Cifar10

input_x = tf.pad(input_x, [[0, 0], [32, 32], [32, 32], [0, 0]]) # size 32x32 -> 96x96

NOT ENOUGH GPU Memory

• If not enough GPU memory, Please edit the code

with tf.Session() as sess : NO
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess : OK

Idea

What is the "SE block" ?

def Squeeze_excitation_layer(self, input_x, out_dim, ratio, layer_name):
    with tf.name_scope(layer_name) :
        squeeze = Global_Average_Pooling(input_x)

        excitation = Fully_connected(squeeze, units=out_dim / ratio, layer_name=layer_name+'_fully_connected1')
        excitation = Relu(excitation)
        excitation = Fully_connected(excitation, units=out_dim, layer_name=layer_name+'_fully_connected2')
        excitation = Sigmoid(excitation)

        excitation = tf.reshape(excitation, [-1,1,1,out_dim])

        scale = input_x * excitation

        return scale

How apply ? (Inception, Residual)

 

How "Reduction ratio" should I set?

• original refers to ResNet-50

ImageNet Results

Benefits against Network Depth

Incorporation with Modern Architecture

Comparison with State-of-the-art

Cifar10 Results

Will be soon

Related works

• Densenet-Tensorflow
• ResNeXt-Tensorflow
• ResNet-Tensorflow

Reference

• Inception_korean

Author

Junho Kim

senet-tensorflow's Issues

Can it support Tensorflow 2.0 with Windows10?

Can it support Tensorflow 2.0 with Windows10?
THX!

Result

Epoch：【100】 train accuracy：0.9999 test accuracy：0.9487

new dataset

@taki0112 I want to train new dataset with this model, must I resize the image to 32*32?

how to run the classification with unseem image (se_inception_resnet_v2 )

Hi, does anyone how to run the model with a unseen image, here is my code, hv no idea about the tensor name of output layer. Thanks.

import tensorflow as tf
import numpy as np
import cv2

sess=tf.Session()
saver = tf.train.import_meta_graph('model/Inception_resnet_v2.ckpt.meta')
saver.restore(sess, tf.train.latest_checkpoint("model/"))
graph = tf.get_default_graph()

x = tf.placeholder(tf.float32, shape=[None, 32, 32, 3])
src = cv2.imread("./plane.jpg")
src = 2*(src/255.0)-1.0
dst = cv2.resize(src, (32, 32), interpolation=cv2.INTER_CUBIC)
picture = dst.reshape(1, 32, 32, 3)

feed_dict ={x: picture}

op_to_restore = graph.get_tensor_by_name("dense_44/bias/Momentum:1")
print (sess.run(op_to_restore,feed_dict))

training and test data normalization

In color_preprocessing function, should we use the same mean and standard deviation both in training and testing?

Is there pretained model for downloading?

Is there pretained model for downloading? Thanks!

can just add squeeze_abode _layer in residual connection?

So if I just want to put a layer in the residuum connection, squeeze_abode _layer ok?
`
if residual :
if a.shape[-1] != b.shape[-1]:
a = squeeze_abode _layer (a,class_number,ratio,layer_number)
b = b+a
else:
b = b+a

`
I am not sure how to set ratio and layer_number parameters

The generated graph is degenerate

The flow of tensors is not visible in the graph. The name and variable scoping is not done properly due to which, the tensorboard graph visualizer ends up creating a flat and degenerate graph.

in the resnext block, why the [1, 1] convolution has stride>1, where [3, 3] has stride 1, the downsample is realized by [3, 3] convolution with stride=2 in resnet

nan output

I use SE_Inception_v4.py to train the cifar10 data. The code can run normally, but the "train_loss", "train_acc", "test_loss" and "test_acc" are "nan". I try to output "batch_loss" from "_, batch_loss = sess.run([train, cost], feed_dict=train_feed_dict)", the result is "nan". Did anyone meet this problem?

Question:Tried to convert 'tensor' to a tensor and failed. Error: Argument must be a dense tensor: Last layer is: DenseLayer (srgan_g5_densenet_se/SE_block1_dense2) [16, 32] - got shape [19], but wanted [].

The error occurred in "excitation = tf.reshape(excitation,(-1,1,1,32)) ".

Thanks for your reply.

Test acc on cifar10

Could you please upload the results on cifar10? Did you reproduce the results in the paper?

training accuracy and test accuracy is always too low

I train the net in cifar-10 without change any parameter. But the train accuracy only reach 0.81 and the test is about 0.62

se should be added after inception block, right?

hi, I think squeeze_excitation_layer should be added after inception block ,not inception_resnet block
Am I right ?

ensorflow.python.framework.errors_impl.InvalidArgumentError: Reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero..'

I got this error when I retrain the network with my data set.

'tensorflow.python.framework.errors_impl.InvalidArgumentError: Reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero..'

I use a small data set of 525 and class_num is 6. How to modify the number of iterations and the 50,0000 value in the code? (In SE_Inception_v4.py)

I already put the new iteration value to a smaller value, say (iteration_n = len(train_x)/batch_size+ 1)
where batch_sze=128 and use this:

if pre_index + batch_size < batch_size*iteration_n: #in place of 50000
....
Any help is appreciated.

How to run the code?

About the structure of SENet

Hi，
Before see your paper, I have saw the channel-wise in 'SCA-CNN', I think the two structures are vary similar, could you tell me the difference between them?
Thank you very much!

Can you please suggest how much gpu memory should I use in order to complete the training? mine is a 4g gpu, and with that config set in sess, it ran out of resource every time I try. Thanks.