Training a model
Training a neural network model is a quite complicated process where one has to choose an architecture of the network, tune the parameters, choose the optimization method and so on. Luckily here we have it done for you!
Interactive model training is the part where you choose the type of a model (currently, single label classification or multi label classification), and train it in one click. You will get a state-of-the-art model trained in just a few minutes! Moreover, if you are advanced user, you can adjust many of the training parameters yourself.
Note: object detection and image segmentation model training is only available as a custom project.
How to use the SentiSight.ai platform
When you have your images labeled, click Train on the menu and choose the type of your model. You will see a window with some information about your data:
- Total number of images that will be used to train the model
- Total number of distinct labels
- Count of each label in images (how many times the label occurred in the images)
On the top left corner you can choose to switch to Advanced view, which enables you to change more parameters for training session. We recommend using this view if you have some background on neural network training.
- In the basic view there are options to write the name of your model and set the train time. Note that you can check the remaining free training time in your User profile information
- In the advanced view you can specify the percentage of your images to be used for validation set, change learning rate and batch size. There is also some additional information about estimated training steps and pre-processing time for creating bottleneck features.
After you set the parameters press Start to start training your model. First, some image pre-processing steps will occur that do not count toward your user train time limit. After that your model will be placed on a queue for training. If there is no queue for training the model, the training process will start immediately after the image pre-processing. Choose Explore models/predict in the menu to watch the image pre-preprocessing and training progress.
The process is run in the background, so you may continue working: adding and labeling more images, setting up a new model, etc.
Learn more about the process
What training time should I choose for a model?
That depends on the number and the size of images. We recommend selecting at least 3 minutes for single-label classification and at least 5 minutes for multi-label classification. After the training, you can analyze the statistics and see the actual predictions on the images to see if they are satisfactory. If you would like to achieve a higher accuracy, try increasing the training time or increasing the number of training images. Also, please, note that longer training does not always provide a more accurate model. Therefore, we continuously test the model's performance during the training and keep only the best model (according to the validation error).
Which images the algorithm is trained on?
By default, the model is trained on all images that are properly labeled for the selected model type.
However if you wish to train only on a certain group of images, you may filter them yourself by using the panel on the lower left or you can simply select the images by using the mouse before training.