如何充分利用开源项目
Transfer Learning is the process of taking a network pre-trained on a dataset and utilizing it to recognize the image, object detection, image segmentation, semantic segmentation, and many more. We can utilize the robust, discriminative filters learned by the state-of-the-art network on challenging datasets such as Imagenet/COCO and then apply these networks for some other tasks.
转移学习是对数据集进行预训练的网络,然后利用它来识别图像,对象检测,图像分割,语义分割等等的过程。 我们可以在诸如Imagenet / COCO之类的具有挑战性的数据集上利用最先进的网络学到的强大,有区别的过滤器,然后将这些网络应用于其他一些任务。
In today's article, we are going to talk about five of the open-source Transfer Learning projects to enhance your skills in the field of data science.
在今天的文章中,我们将讨论五个开源的Transfer Learning项目,以提高您在数据科学领域的技能 。
Note: This article is to just give a glimpse of some of the not-so-famous but really good open-source projects which you can use in your projects. To read more about each of them I recommend following the link given along the project.
注意:本文只是简要介绍一些您可以在项目中使用的不太著名但非常好的开源项目。 要阅读有关它们的更多信息,我建议您遵循项目中提供的链接。
![Image for post](https://miro.medium.com/max/9999/1*NCV871qd9xWiRlWbKlDN_Q.jpeg)
Having a good theoretical knowledge is amazing but implementing them in code in a real-time machine learning project is a completely different thing. You might get different and unexpected results based on different problems and datasets. So as a Bonus,I am also adding the links to the various courses which has helped me a lot in my journey to learn Data science and ML.I am personally a fan of DataCamp, I started from it and I am still learning through DataCamp and keep doing new courses. They seriously have some exciting courses. Do check them out.
拥有良好的理论知识是惊人的,但是在实时机器学习项目中以代码实现它们是完全不同的。 根据不同的问题和数据集,您可能会得到不同且出乎意料的结果。 因此,作为奖励,我还添加了指向各种课程的链接,这些链接对我学习数据科学和ML的旅程大有帮助。我个人是 DataCamp的爱好者 ,我从中开始,但仍在通过 DataCamp 学习 并继续学习新课程。 他们认真地开设了一些令人兴奋的课程。 请检查一下。
P.S: I am still using DataCamp and keep doing courses in my free time. I actually insist the readers to try out any of the above courses as per their interest, to get started and build a good foundation in Machine learning and Data Science. The best thing about these courses by DataCamp is that they explain it in a very elegant and different manner with a balanced focus on practical and well as conceptual knowledge and at the end, there is always a Case study. This is what I love the most about them. These courses are truly worth your time and money. These courses would surely help you also understand and implement transfer learning, machine learning in a better way and also implement it in Python or R. I am damn sure you will love it and I am claiming this from my personal opinion and experience.
PS:我仍在使用 DataCamp, 并在 业余时间 继续 上课 。 我实际上是坚持要求读者根据自己的兴趣尝试上述任何课程,以开始并为机器学习和数据科学打下良好的基础。 DataCamp 开设的这些课程的最好之 处 在于,他们以非常优雅且与众不同的方式 对课程进行了 解释,同时重点关注实践和概念知识,最后始终进行案例研究。 这就是我最喜欢他们的地方。 这些课程确实值得您花费时间和金钱。 这些课程肯定会帮助您更好地理解和实施迁移学习,机器学习,并且还可以在Python或R中实现它。我该死的您一定会喜欢它,并且我从我个人的观点和经验中主张这一点。
Also, I have noticed that DataCamp is giving unlimited access to all the courses for free for one week starting from 1st of September through 8th September 2020, 12 PM EST. So this would literally be the best time to grab some yearly subscriptions(which I have) which basically has unlimited access to all the courses and other things on DataCamp and make fruitful use of your time sitting at home during this Pandemic. So go for it folks and Happy learning
此外,我注意到, DataCamp 自2020年9月1日至美国东部时间12 PM,为期一周免费无限制地访问所有课程。 因此,从字面上看,这将是获取一些年度订阅(我拥有)的最佳时间,该订阅基本上可以无限制地访问 DataCamp 上的所有课程和其他内容, 并可以在这次大流行期间充分利用您在家里的时间。 因此,请亲朋好友学习愉快
Coming back to the topic -
回到主题-
1,密集深度 (1.Densedepth)
Densedepth is a high-quality Monocular Depth Estimation project built on top of Transfer Learning and NYU Depth V2 and KITTI dataset, implemented by Ibraheem Alhashim and Peter Wonka.
d ensedepth是建立在迁移学习和纽约大学的深度V2和KITTI数据集之上的高品质单目深度估计项目 ,通过实施易卜拉欣Alhashim 和Peter Wonka 。
Densedepth aims to apply state of the art approach to 3D images and produce high-quality dense maps from the images with the best-generalized performance using Transfer Learning and standard Neural network architecture.
Densedepth旨在将最先进的方法应用于3D图像,并使用Transfer Learning和标准的神经网络体系结构从图像中生成具有最佳概括性能的高质量密集地图。
Transfer learning has proved to be efficient between different tasks many of which are related to 3D reconstruction and using transfer learning, allows for a more modular architecture as we get to use the power of the pre-trained networks.
事实证明, 转移学习在不同的任务之间是有效的,其中许多任务与3D重建和使用转移学习有关,随着我们使用预训练网络的强大功能,可以采用更加模块化的体系结构。
A single straight forward encoder-decoder architecture is used to achieve a state of the art high-quality dense maps. Encoder part is performed by a pre-trained truncated DenseNet-169 architecture and the Decoder part is composed of basic blocks of Convolutional neural network layers.
使用单个直截了当的编码器/解码器体系结构来实现现有技术水平的高质量密集图 。 编码器 部分由预训练的截断DenseNet-169体系结构执行,而 解码器 部分由卷积神经网络层的基本块组成。
The main idea behind using Transfer Learning was to make use of the image encoders that are originally designed for the problem of image classification.
使用 转移学习 的主要思想是利用 最初为图像分类问题设计的图像编码器。
![Image for post](https://miro.medium.com/max/9999/1*hfFKSo2Q11A7B9mYzsBoSA.png)
Applications: It is used in scene understanding and scene reconstruction, GPS navigation systems, Image refocusing and augmented reality games.
应用范围:用于场景理解和场景重建,GPS导航系统,图像重新聚焦和增强现实游戏。
2.DeepDrive (2.DeepDrive)
DeepDrive is a free and open-source simulator that allows anyone with a computer to push the state of the art in self-driving AI cars. It is written in Python 3.6 and in the backend it involves Deep Reinforcement Learning(openai) and Transfers Learning(MobileNet V2). The model is trained on top of 100GB of 8.2 hours of driving of camera, depth, steering, throttle, and brake of an ‘oracle’ path following agent with rotated camera angles.
d eepDrive是一个自由和开源模拟器,允许任何人与计算机推入自驾车AI汽车的现有技术的状态 。 它是用Python 3.6编写的,在后端它涉及深度强化学习(openai) 和 转移学习(MobileNet V2) 。 该模型是在100GB的驱动力基础上训练的,该驱动程序具有8.2小时的摄像头驱动力,深度,转向,油门以及跟随具有旋转摄像头角度的特工的“ oracle”路径的制动。
The aim of this project was to experiment with self-driving AI cars and to drive the future of automotive perception.
该项目的目的是试验 自动驾驶人工智能汽车, 并推动汽车感知的未来。
![Image for post](https://miro.medium.com/max/9999/1*gi6p4e-1ZcLgTGVK_ablAQ.png)
3.时间序列分类的转移学习 (3. Transfer Learning for Time-Series Classification)
Using Transfer Learning for Time-Series Classification has proved that the accuracy of Deep Learning models can be improved if the model is fine-tuned from a pre-trained neural network instead of training the model from scratch.
ü唱迁移学习的时间序列分类已经证明, 深度学习模型的精度可以提高,如果该模型是微调从预训练神经网络,而不是从头开始训练模型。
The aim of using Transfer learning for time-series data was the observation that showed that subsequence learned by the learning shapelets approach in transfer learning was related to the filters(kernels) learned by the Convolutional neural networks.
对时间序列数据使用转移学习的目的是观察表明, 在转移学习中学习小波方法学习的子序列与卷积神经网络学习的过滤器(内核)有关。
Hence the method proposed in the paper Transfer Learning for Time-Series Classification can help you identify which dataset should be used for transfer learning given a new time-series classification problem using the Dynamic Time Warping algorithm as an inter datasets similarity measure to identify the relationships between the source dataset and the target datasets in our transfer learning framework.
因此,本文提出的方法 针对时间序列分类的转移学习可以使用动态时间规整算法 作为数据集之间的相似性度量来识别 源数据集 和 目标 之间的关系,从而在给定新的时间序列分类问题的情况下,帮助您确定应将哪个数据集用于转移学习迁移学习框架中的 数据集 。
To predict the best source dataset for a given target dataset, Dynamic Time Warping method to measure the inter dataset similarities has been implemented in the paper Transfer Learning for Time-Series Classification.
为了预测给定目标数据集的最佳源数据集, 已在论文 《时间序列分类的转移学习》中 实现了用于度量数据集间相似性的 动态时间规整方法 。
![Image for post](https://miro.medium.com/max/9999/1*b_HTy6IiIUpkApxOYFzawQ.png)
![Image for post](https://miro.medium.com/max/9999/1*LNYhdOrEj-EHcpyQLvDrxQ.png)
Example: First training the model on the ElectricDevices dataset and then fine-tune the same model on the OSULeaf dataset to improve the performance of model with TransferLearning.
示例 :首先在ElectricDevices数据集上训练模型,然后在OSULeaf数据集上微调相同的模型,以通过TransferLearning改善模型的性能。
4.动漫字符识别的转移学习 (4. Transfer Learning for Anime Character Recognition)
Recently, the Transfer Learning approach is proposed for face recognition using CNN and the results have verified that the given transfer learning method gives better recognition results as compared to other methods.
最近,提出了一种使用CNN进行面部识别的转移学习方法,结果证明,与其他方法相比,给定的转移学习方法具有更好的识别结果 。
Keeping this in mind, an experiment was done to try the performance of transfer learning further by using three different anime characters that have many similar features so see how well the transfer learning works in detecting the Anime characters in images.
牢记这一点,我们进行了一项实验,通过使用具有许多相似功能的三个不同的动漫角色来进一步尝试迁移学习的性能,因此请查看迁移学习在检测图像中的动漫人物方面的工作情况。
This project was structured in various steps:
该项目分为多个步骤:
detecting faces of anime characters from each image using lbpcascade_animeface
使用 lbpcascade_animeface 从每个图像检测动漫人物的 面Kong
resize the images to 96*96 pixels and then split the image to train and test images
将图像调整为 96 * 96像素 ,然后将其拆分以训练和测试图像
features extraction and preprocessing was performed prior to training
在训练之前进行 特征提取和预处理
train the model and evaluate the results on test and validation images
Transfer learning allows the Convolutional Neural Network to learn features from the VGG-16 model pre-trained with huge ImageNet weights to train the images from the face database and then the extracted features are fed as input to the fully connected layer and softmax activation function for classification.
转移学习 允许 卷积神经网络 从 预先训练有巨大ImageNet权重 的 VGG-16模型中 学习特征,以 从人脸数据库中训练图像,然后将提取的特征作为输入提供给全连接层和softmax激活函数,用于分类 。
![Image for post](https://miro.medium.com/max/9999/1*aaR4Ih6SRIMMH3hGH2VDJA.png)
![Image for post](https://miro.medium.com/max/9999/1*39TvRkv3FOTIhppfpFUyGA.png)
5,亚马逊森林计算机视觉 (5.Amazon Forest Computer Vision)
This was a project to understand the Amazon Forest from space using satellite data to track the human footprints in the Amazon rain forest. The task was to label satellite image chips with atmospheric conditions and various classes of land covers(see below image). The predictions produced will help the global community better understand where, how and why deforestation happened all over the world and how to deal with this problem.
这是一个使用卫星数据从太空了解亚马逊森林的项目,该数据可跟踪亚马逊雨林中的人类足迹 。 任务是给卫星图像芯片贴上大气条件和各种土地覆被标签 (见下图) 。 产生的预测将帮助国际社会更好地了解全世界毁林的地点,方式和原因,以及如何解决这一问题。
The model was fine-tuned with custom weights from Pycaffe on VGG16 and DenseNet121 models for best outputs using Pytorch and Keras for better understanding.
该模型是微调从Pycaffe上VGG16和DenseNet121模型定制权使用Pytorch和Keras为了更好的理解最好的输出。
![Image for post](https://miro.medium.com/max/9999/1*YumsQpfdMIjNtkMMAxI3vg.png)
如何充分利用开源项目