deep learning
关注
分享
扫一扫
seamanj
这个作者很懒，什么都没留下…
展开
-
Visualizing and Understanding Convolutional Networks
代码请见:https://github.com/amir-saniyan/ZFNet # Input: 224x224x3. with tf.name_scope('input'): self.X = tf.placeholder(dtype=tf.float32, shape=[N...原创 2019-05-29 01:30:32 · 266 阅读 · 0 评论
-
drop out, learning rate in nn
use different initial learning rates, says: 1e-3, 1e-4, 1e-5, if 1e-5 is the best one, that means your network is too complicate. you may want reduce to the layers.原创 2020-01-15 21:23:50 · 306 阅读 · 0 评论 -
lenet5 in tensorflow and visualization in tensorboard
# Copyright 2015 The TensorFlow Authors. All Rights Reserved.## Licensed under the Apache License, Version 2.0 (the "License");# you may not use this file except in compliance with the License.# Y...原创 2019-10-16 23:07:11 · 187 阅读 · 0 评论 -
ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design
class ShuffleNetV2(nn.Module): def __init__(self, net_size): super(ShuffleNetV2, self).__init__() out_channels = configs[net_size]['out_channels'] num_blocks = configs[ne...原创 2019-05-29 19:53:03 · 743 阅读 · 0 评论 -
Aggregated Residual Transformations for Deep Neural Networks
/home/seamanj/Software/anaconda3/bin/python3.7 /home/seamanj/Workspace/AlexNet/main.pyDownloading: "https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth" to /home/seamanj/.cache/torch/ch...原创 2019-05-29 20:00:17 · 286 阅读 · 0 评论 -
Squeeze-and-Excitation Networks
'''SENet in PyTorch.SENet is the winner of ImageNet-2017. The paper is not released yet.'''import torchimport torch.nn as nnimport torch.nn.functional as Fclass BasicBlock(nn.Module): de...原创 2019-05-29 20:07:43 · 299 阅读 · 0 评论 -
Xception: Deep Learning with Depthwise Separable Convolutions
代码请见:https://github.com/Cadene/pretrained-models.pytorch/blob/master/pretrainedmodels/models/xception.py"""Ported to pytorch thanks to [tstandley](https://github.com/tstandley/Xception-PyTorch)@a...原创 2019-05-29 20:20:39 · 745 阅读 · 0 评论 -
Rethinking the Inception Architecture for Computer Vision
/home/seamanj/Software/anaconda3/bin/python3.7 /home/seamanj/Workspace/AlexNet/main.pyDownloading: "https://download.pytorch.org/models/inception_v3_google-1a9a5a14.pth" to /home/seamanj/.cache/t...原创 2019-05-29 20:34:11 · 857 阅读 · 0 评论 -
Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
from __future__ import print_function, division, absolute_importimport torchimport torch.nn as nnimport torch.nn.functional as Fimport torch.utils.model_zoo as model_zooimport osimport sys__a...原创 2019-05-29 20:41:06 · 648 阅读 · 0 评论 -
SQUEEZENET: ALEXNET-LEVEL ACCURACY WITH 50X FEWER PARAMETERS AND 0.5MB MODEL SIZE
/home/seamanj/Software/anaconda3/bin/python3.7 /home/seamanj/Workspace/AlexNet/main.pyDownloading: "https://download.pytorch.org/models/squeezenet1_0-a815701f.pth" to /home/seamanj/.cache/torch/ch...原创 2019-05-29 20:12:50 · 336 阅读 · 0 评论 -
CondenseNet: An Efficient DenseNet using Learned Group Convolutions
具体代码请见:https://github.com/ShichenLiu/CondenseNetclass CondenseNet(nn.Module): def __init__(self, args): super(CondenseNet, self).__init__() self.stages = args.stages se...原创 2019-05-29 20:55:10 · 473 阅读 · 0 评论 -
理解crossentropy loss
input 为 (N,C)维, 其中N为样本个数, C为分类数, 其中值为每个样本在某类上的得分, 注意这里不是概率, 可以为正也可以为负, 也可以大于1, 后面会经过softmax的处理, 统一变成概率的分布target为N维, 其值为truth groud 的分类值这个Loss就是正确得分(后面会转成概率)在所有得分所占的比重来确定的, 如果正确的概率为1, 其他都为0的话, 那么 根...原创 2019-06-15 02:51:32 · 5254 阅读 · 0 评论 -
Deeper Depth Prediction with Fully Convolutional Residual Networks
from single RGB image infer the depth.原创 2019-06-15 15:00:04 · 377 阅读 · 0 评论 -
如何让NN分类输出[0,1]的值,并且相加等于1
只需要最后一层为softmax即可以lenet5为例import torch.nn as nnfrom collections import OrderedDictclass LeNet5(nn.Module): """ Input - 1x32x32 C1 - 6@28x28 (5x5 kernel) tanh S2 - 6@14x14 (2x...原创 2019-06-05 06:02:38 · 619 阅读 · 0 评论 -
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
代码网址： https://github.com/seamanj/faster-rcnn.pytorch原创 2019-06-17 05:16:48 · 162 阅读 · 0 评论 -
roipool and roialign difference
https://www.youtube.com/watch?v=XGi-Mz3do2s转载 2019-06-17 21:10:52 · 363 阅读 · 0 评论 -
A Hierarchical Deep Temporal Model for Group Activity Recognition
这里解释一下1-6,xtx_txt是LSTM cell 在t时间的输入然后iti_tit,ftf_tft, oto_tot, gtg_tgt分别是input gate, forget gate, output gate和 input modulation gate at time t.这堆ctc_tct是memory cell, 跟前一帧的memory cell的相关h...原创 2019-06-25 22:57:00 · 785 阅读 · 0 评论 -
PoseNet: A Convolutional Network for Real-Time 6-DOF Camera Relocalization
https://zhuanlan.zhihu.com/p/53044278https://zhuanlan.zhihu.com/p/45483149abstract: 训练了一个神经网络可以从单张图像端到端的回归出相机的位置和姿态，有别于SLAM，不再需要额外的工程操作或者图优化。室外精度达到2m，3°。室内精度达到0.5m，5°偏差。网络是23层，利用transfer learning fr...转载 2019-06-24 01:05:30 · 2032 阅读 · 0 评论 -
Scale-Adaptive Neural Dense Features: Learning via Hierarchical Context Aggregation
这篇文章主要讲给你一张图片, 怎样得到该图片中每个像素全局或者局部的feature,输入是HxWx3, 输出是HxWxn, n为feature的维数对于Table1来说, G,L, n就是3, 而GL,n就是6整个网络如图2那么如果 训练得到这个feature呢? 注意这句话A siamese network with two identical SAND branchesis tr...原创 2019-06-24 01:09:55 · 548 阅读 · 0 评论 -
CNN 总结 模型归类
部分模型实现代码:https://github.com/kuangliu/pytorch-cifar/tree/master/models原创 2019-05-29 19:42:43 · 229 阅读 · 0 评论 -
ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices
具体见: https://github.com/kuangliu/pytorch-cifar/tree/master/modelsclass ShuffleNet(nn.Module): def __init__(self, cfg): super(ShuffleNet, self).__init__() out_planes = cfg['out_p...原创 2019-05-29 19:34:54 · 251 阅读 · 0 评论 -
MobileNetV2: Inverted Residuals and Linear Bottlenecks
推荐阅读:https://blog.youkuaiyun.com/kangdi7547/article/details/8143157216->96->24, 为什么中间有96怎么确定, 由t来确定, 这里t为6, 所以16*6 =96MobileNetV2( (features): Sequential( (0): ConvBNReLU( (0): Conv2d(3...原创 2019-05-29 19:09:17 · 625 阅读 · 0 评论 -
Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation
完整代码:https://github.com/jfzhang95/pytorch-deeplab-xceptionclass DeepLab(nn.Module): def __init__(self, backbone='resnet', output_stride=16, num_classes=21, sync_bn=True, freeze...原创 2019-05-30 18:50:53 · 457 阅读 · 0 评论 -
YOLOv3: An Incremental Improvement
具体代码见: https://github.com/ultralytics/yolov3def create_modules(module_defs): """ Constructs module list of layer blocks from module configuration in module_defs """ hyperparams = modu...原创 2019-05-30 19:17:48 · 926 阅读 · 0 评论 -
Feature pyramid networks for object detection
具体分析请见: https://medium.com/@jonathan_hui/understanding-feature-pyramid-networks-for-object-detection-fpn-45b227b9106cFPN is not an object detector by itself. It is a feature detector that works with ...原创 2019-05-31 02:55:45 · 365 阅读 · 0 评论 -
Fully Convolutional Networks for Semantic Segmentation
具体实现代码见:https://github.com/wkentaro/pytorch-fcn/blob/master/torchfcn/models/fcn16s.pyclass FCN16s(nn.Module): pretrained_model = \ osp.expanduser('~/data/models/pytorch/fcn16s_from_caff...原创 2019-05-31 03:04:59 · 511 阅读 · 0 评论 -
Mask R-CNN
code : https://github.com/facebookresearch/maskrcnn-benchmark?utm_source=mybridge&utm_medium=blog&utm_campaign=read_more原创 2019-05-31 03:15:34 · 189 阅读 · 0 评论 -
Gradient-Based Learning Applied to Document Recognition
对应codeimport torch.nn as nnfrom collections import OrderedDictclass LeNet5(nn.Module): """ Input - 1x32x32 C1 - 6@28x28 (5x5 kernel) tanh S2 - 6@14x14 (2x2 kernel, stride 2) ...原创 2019-05-28 17:12:47 · 1744 阅读 · 0 评论 -
imagenet classification with deep convolutional neural networks
3.1这里用了ReLU, 而不是sigmoid或者是tanh, 在文章中称sigmoid或tanh为saturating nonlinearities. 为啥呢, 具体可见https://www.quora.com/Why-would-a-saturated-neuron-be-a-problem这里我挑重点The logistic has outputs between 0 and 1, ...原创 2019-05-28 20:11:21 · 259 阅读 · 0 评论 -
Learning to Segment Object Candidates
看下源代码class DeepMask(nn.Module): def __init__(self, config=default_config, context=True): super(DeepMask, self).__init__() self.config = config self.context = context # w...原创 2019-05-28 20:25:47 · 271 阅读 · 0 评论 -
R-CNN, Fast R-CNN, and Faster R-CNN Basics
regions with convolutional neural networks (R-CNN), combines rectangular region proposals with convolutional neural network features. R-CNN is a two-stage detection algorithm. The first stage identifi...转载 2019-06-12 16:48:36 · 184 阅读 · 0 评论 -
Learning to Refine Object Segments
源代码里面的版本看下源代码class SharpMask(nn.Module): def __init__(self, config=default_config, context=True): super(SharpMask, self).__init__() self.context = context # with context ...原创 2019-05-28 23:31:30 · 620 阅读 · 0 评论 -
torchvision datasets and models
AlexNetVGGResNetSqueezeNetDenseNetInception v3GoogLeNetShuffleNet v2MobileNet v2ResNeXt原创 2019-05-29 02:15:34 · 259 阅读 · 1 评论 -
Going deeper with convolutions
import torchvision.models as modelsif __name__ == '__main__': googlenet = models.googlenet(pretrained=True) print(googlenet)GoogLeNet( (conv1): BasicConv2d( (conv): Conv2d(3, 64, k...原创 2019-05-29 03:00:04 · 156 阅读 · 0 评论 -
VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION
VGG( (features): Sequential( (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): ReLU(inplace) (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1,...原创 2019-05-29 03:27:34 · 170 阅读 · 0 评论 -
Deep Residual Learning for Image Recognition
resnet也有各种变种ResNet( (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False) (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)...原创 2019-05-29 03:47:42 · 163 阅读 · 0 评论 -
Densely Connected Convolutional Networks
DenseNet( (features): Sequential( (conv0): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False) (norm0): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track...原创 2019-05-29 18:01:12 · 242 阅读 · 0 评论 -
Convolutional Neural Networks with Alternately Updated Clique
具体代码见:https://github.com/iboing/CliqueNetclass build_cliquenet(nn.Module): def __init__(self, input_channels, list_channels, list_layer_num, if_att): super(build_cliquenet, self).__init_...原创 2019-05-29 18:32:15 · 237 阅读 · 0 评论 -
MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
具体实现代码见:https://github.com/marvis/pytorch-mobilenetclass Net(nn.Module): def __init__(self): super(Net, self).__init__() def conv_bn(inp, oup, stride): return nn.Seq...原创 2019-05-29 18:42:48 · 199 阅读 · 0 评论 -
Dilated Residual Networks
完整代码请见: https://github.com/fyu/drnclass DRN(nn.Module): def __init__(self, block, layers, num_classes=1000, channels=(16, 32, 64, 128, 256, 512, 512, 512), ou...原创 2019-05-30 18:40:35 · 856 阅读 · 0 评论