All Projects → Chevalier1024 → Mpcnn

Chevalier1024 / Mpcnn

Multi-Perspective Sentence Similarity Modeling with Convolutional Neural Network复现

Programming Languages

python
139335 projects - #7 most used programming language

Projects that are alternatives of or similar to Mpcnn

Torchio
Medical image preprocessing and augmentation toolkit for deep learning
Stars: ✭ 708 (+3118.18%)
Mutual labels:  cnn
Srgan Tensorflow
Tensorflow implementation of the SRGAN algorithm for single image super-resolution
Stars: ✭ 754 (+3327.27%)
Mutual labels:  cnn
Exemplar Cnn
Discriminative Unsupervised Feature Learning with Convolutional Neural Networks
Stars: ✭ 16 (-27.27%)
Mutual labels:  cnn
Tensorflow cookbook
Code for Tensorflow Machine Learning Cookbook
Stars: ✭ 5,984 (+27100%)
Mutual labels:  cnn
Self driving pi car
A deep neural network based self-driving car, that combines Lego Mindstorms NXT with the computational power of a Raspberry Pi 3.
Stars: ✭ 744 (+3281.82%)
Mutual labels:  cnn
Tf cnnvis
CNN visualization tool in TensorFlow
Stars: ✭ 769 (+3395.45%)
Mutual labels:  cnn
Ultra Fast Lane Detection
Ultra Fast Structure-aware Deep Lane Detection (ECCV 2020)
Stars: ✭ 688 (+3027.27%)
Mutual labels:  cnn
Eda nlp
Data augmentation for NLP, presented at EMNLP 2019
Stars: ✭ 902 (+4000%)
Mutual labels:  cnn
Tensorflow Tutorial
TensorFlow and Deep Learning Tutorials
Stars: ✭ 748 (+3300%)
Mutual labels:  cnn
Ai sudoku
GUI based Smart Sudoku Solver that tries to extract a sudoku puzzle from a photo and solve it
Stars: ✭ 830 (+3672.73%)
Mutual labels:  cnn
Dlib face recognition from camera
Detect and recognize the faces from camera / 调用摄像头进行人脸识别,支持多张人脸同时识别
Stars: ✭ 719 (+3168.18%)
Mutual labels:  cnn
Lstm Char Cnn Tensorflow
in progress
Stars: ✭ 737 (+3250%)
Mutual labels:  cnn
Codes For Lane Detection
Learning Lightweight Lane Detection CNNs by Self Attention Distillation (ICCV 2019)
Stars: ✭ 801 (+3540.91%)
Mutual labels:  cnn
Pytorchinsight
a pytorch lib with state-of-the-art architectures, pretrained models and real-time updated results
Stars: ✭ 713 (+3140.91%)
Mutual labels:  cnn
Anime4kcpp
A high performance anime upscaler
Stars: ✭ 887 (+3931.82%)
Mutual labels:  cnn
Hand3d
Network estimating 3D Handpose from single color images
Stars: ✭ 690 (+3036.36%)
Mutual labels:  cnn
Sincnet
SincNet is a neural architecture for efficiently processing raw audio samples.
Stars: ✭ 764 (+3372.73%)
Mutual labels:  cnn
Deepcache
Cache design for CNN on mobile
Stars: ✭ 22 (+0%)
Mutual labels:  cnn
Tensorflow Tutorial
Some interesting TensorFlow tutorials for beginners.
Stars: ✭ 893 (+3959.09%)
Mutual labels:  cnn
Quickdraw
Implementation of Quickdraw - an online game developed by Google
Stars: ✭ 805 (+3559.09%)
Mutual labels:  cnn

MPCNN

Multi-Perspective Sentence Similarity Modeling with Convolutional Neural Network paper link:http://www.emnlp2015.org/proceedings/EMNLP/pdf/EMNLP181.pdf

论文分析: http://blog.csdn.net/liuchonge/article/details/62424805 http://blog.csdn.net/liuchonge/article/details/64128870 http://blog.csdn.net/liuchonge/article/details/64440110

glove file :http://nlp.stanford.edu/data/glove.6B.zip experiment on python3.5 and tensorflow-gpu1.4

引用代码:https://github.com/lc222/MPCNN-sentence-similarity-tensorflow

关于定位loss NAN的问题: 1.用tfdbg命令查找到计算欧式距离的时候有些输出为0,导致最后计算loss的时候输出为NAN。 2.利用tensorboard可视化每个层的输出及权重

如何解决loss NAN的问题: 1.调低学习率 2.梯度检验:手工计算的梯度和框架计算的梯度比较 3.如果cost function有log 函数,tf.clip_by_value(y,1e-4)将输入为0的去掉 4.梯度截断,效果不是很好

由于原来博主的代码存在loss NAN的问题,所以我对博主的代码做了以下修改: 1.计算相似度层中去掉了欧氏距离或者去掉tf.sqrt函数 2.每一卷积层加BN 3.所有可训练的变量加入到L2正则化中 4.activate function 都换成了 tanh

仍存在的问题: 1.加上attention layer仍然会出现loss NAN的问题

未实现的想法: 1.将欧式距离换为标准化欧氏距离 2.dropout设为0.8-0.9会不会更容易收敛

train.png是训练集的acc和loss曲线 valid.png是验证集的acc和loss曲线

Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].