0. Intro
1. Installing the Tensorflow GPU version in Windows
1.5 Introduction to TF-Slim
2. Download, learn and evaluate slim models
3. Learning from my images (using caltech images)
4. Using learned models
5. Creating a Python Tkinter GUI application
According to the article released by Google Research, TF-Slim provides the Deep Learning CNN model for image classification such as AlexNet, VGG, ResNet, and Inception-V3. It provides not only a model network configuration, but also Datasets (ImageNet, CIFAR10, MNIST, FLOWER), model training and evoluating codes. It even offers pre-trained models using ImageNet and also provides Fine Tune functionality for new images
Wow! Then we will prepare a new image and the rest will be done by TF-Slim!
Yes, it is!
I've done it myself and we just have to prepare the image. Deep Learning is done by TF-Slim.
So before we do the actual exercise, let's look at TF-Slim more briefly.
There are four kinds of datasets, MNIST and Cifar10, which have been tried by anyone who is a beginner of Deep Learning. Of course you can also add other images. That's how I'll tell you in the next posting.
It also offers the latest Inception-ResNet-v2 model. This model has a Top-5 accuracy of 95.3%, which is 0.1% slightly higher than the Inception V4.
You can use the provided network for new learning, or you can learn new images on already created models. Because the model results are updated during learning, you can check them directly on the Tensor Board without adding any other setting into codes.
Codes is also provided to evaluate the accuracy of the learned model.
Next, I will try to use the basic TF-Slim code by practicing. After that, I will try to learn new images using TF-Slim.
1. Installing the Tensorflow GPU version in Windows
1.5 Introduction to TF-Slim
2. Download, learn and evaluate slim models
3. Learning from my images (using caltech images)
4. Using learned models
5. Creating a Python Tkinter GUI application
According to the article released by Google Research, TF-Slim provides the Deep Learning CNN model for image classification such as AlexNet, VGG, ResNet, and Inception-V3. It provides not only a model network configuration, but also Datasets (ImageNet, CIFAR10, MNIST, FLOWER), model training and evoluating codes. It even offers pre-trained models using ImageNet and also provides Fine Tune functionality for new images
Wow! Then we will prepare a new image and the rest will be done by TF-Slim!
Yes, it is!
I've done it myself and we just have to prepare the image. Deep Learning is done by TF-Slim.
So before we do the actual exercise, let's look at TF-Slim more briefly.
1) Providing datasets
There are four kinds of datasets, MNIST and Cifar10, which have been tried by anyone who is a beginner of Deep Learning. Of course you can also add other images. That's how I'll tell you in the next posting.
2) Providing TFRecord format conversion code
Except for ImageNet data, other datasets are downloaded directly from the provided code and automatically converted to label generation and TFRecord format.If you read a single piece of image and use it for learning, it takes a lot of time for FILE I / O and it takes a lot of learning time. The TFRecord format creates a large chunk of binary data by grouping the images together so that there is little time to enter the FILE I / O. A similar format is RecordIO.3) Pre-trained model provided
It also offers the latest Inception-ResNet-v2 model. This model has a Top-5 accuracy of 95.3%, which is 0.1% slightly higher than the Inception V4.
4) Provide Traning function
You can use the provided network for new learning, or you can learn new images on already created models. Because the model results are updated during learning, you can check them directly on the Tensor Board without adding any other setting into codes.
5) Model performance evaluation function
Codes is also provided to evaluate the accuracy of the learned model.
Next, I will try to use the basic TF-Slim code by practicing. After that, I will try to learn new images using TF-Slim.
댓글 없음:
댓글 쓰기