Training definition is your manifest providing various details about the training you'd be running. On the right panel, click on the Add training definition link and create a New training definition.
1.Provide details about the experiment name and description.
2.Upload your training model code by dropping off the zip file found here in the section marked to upload the model file. In our case we will be using the code from here which is a sample MNIST program. The code contains of a python file which is basically a python program from tensorflow to train a model on MNIST dataset using convolution.
3.Select the framework as tensorflow 1.5 from the dropdown.
4.Copy the execution command as
python3 convolutional_network.py --trainImagesFile ${DATA_DIR}/train-images-idx3-ubyte.gz --trainLabelsFile ${DATA_DIR}/train-labels-idx1-ubyte.gz --testImagesFile ${DATA_DIR}/t10k-images-idx3-ubyte.gz --testLabelsFile ${DATA_DIR}/t10k-labels-idx1-ubyte.gz --learningRate 0.001 --trainingIters 200000The execution command is your entry point and is what triggers your program.
5.On the Training definition attributes section on the right panel, select Compute plan as 1 x NVIDIA® Tesla® K80 (2 GPU) . This specifies the GPUs that you will be able to use for your training. Running on a more powerful hardware is as easy as sleting a different compute configuration.
6.Since you are running a non distributed single learner training, select the number of nodes to be 1.
7.We will not be doing Hyperparameter optimization method as a part of this lab, so select the option as none.
8.Click on create button, this will create the training definition and will bring you back to the create experiment page.
9.Click on create and run button and this should start the training.
10.Hurray!! You have a training running. Follow the direction here to see the progress of the training.