字幕表 動画を再生する 英語字幕をプリント Hi. I'm Megan [? Arau. ?] Today I will be showing you the Deep Learning SDK from Intel. Data scientists and software developers can use the Deep Learning SDK to simplify installation, easily prepare models using popular deep learning frameworks on Intel hardware, and optimize performance for training and inference on Intel architecture. The Intel deep learning SDK is a free set of tools to develop, train, and deploy deep learning solution. In today's presentation, you will see a live demonstration of the DL SDK training tool and learn how to visually set-up, and [? queue, ?] and train datasets using deep learning typologies like LeNet, AlexNet, and GoogLeNet on frameworks like [? Cafe ?] optimized for performance on Intel architecture. Now let's take a look at the training tool. You can download the DL SDK installer for Windows or Mac from the link provided in the description below, and install it using the instructions from the user guide. Launch the tool on your host machine using the IP address and port of the server on which the DL SDK is installed. Once launched, the home screen looks like this, and it provides the ability to upload images and create and train a model. Let's take a look at an example. I will use the MNIST dataset for the demonstration. Before uploading the data set, make sure that your [? guiding ?] structure has the labels at the highest level, with data corresponding to each label with them. Now the first step is to upload the raw data. Go to the upload step, create a supporter in the route plan provided, select your data, file and upload. Once uploaded, you can copy the source [? stat ?] for your data set from the history data. Now, let's create a data set. Proved a name for the data set, a brief description, and choose the percentage of data up from the raw data set that you would like to use for training and validation. The MNIST dataset has had over 60,000 items. You can experiment this by varying the percentages. The MNIST dataset images are all 28 by 28 grayscale. So make sure you select the right option in the image processing tab. If using color images for use with topologies like AlexNet and GoogLeNet, the processing options must be set accordingly. Check out the user guide to learn more about the other processing options. Now go to the database option stat and choose to database packet and the image encoding. Now you're ready to create the data set. Once successful you can visualize the number of data items in each label for the training and the validation data set. Let's now go to the model stat, select the right topology, and train the model. Leonard is one of the topologies that works best for handwritten digit recognition, so we select it. AlexNet and GoogLeNet work best for color images like [? C410 ?] and [INAUDIBLE]. You can also edit the built-in topologies using the new custom topology feature. Edit the topology file in the text field and save it. It's now ready to use to train model. Additional data transformation capabilities exist within the DL SDK to create a digital data set. We do not cover them here, but you can learn more from the use guide. Next, we move to the parameter selection tab. Based on the topology chosen, the hyper parameters are all, by default, set to the optimal values. But you can experiment by changing some of these values to see how the accuracy and the loss function change. Now, run the model. After all 50 [INAUDIBLE], we see that 100% training accuracy and approximately about 98% validation accuracy is achieved. You can validate the model by going to the testing tab, input a random digit that's not part of validation data set to see what the model spits out as potential [INAUDIBLE]. For each label category, you can also see the number of hits and misses. After training is complete, the set of all Cafe files of the train model can be downloaded and used on the deployment platform. For example, a smart camera or other mobile platform. The real-time data [? effect ?] to the model on the deployment platform can then predict an outcome based on the train model. Thanks for watching. You can learn more about DL SDK on IDZ and download the two from the link below. Make sure to like and share this video, and subscribe to the Intel software channel.