Custom Training

Custom Training (also called "Advanced Training") is an advanced option recommended for users who are familiar with machine learning. Custom Training allows these users to have more granular control over pre-processing transforms and manual hyperparameter tuning.

Run Custom Training

1. Open the Project you want to train.
2. Click the downward-facing arrow next to Train and click Train with Custom Options.
   Train with Custom Options
3. Configure the settings and hyperparameters for custom model training.
   Configure Settings and Hyperparameters
4. After configuring the settings, click Train.
   Train the Model
5. LandingLens runs the Model Training process and creates a model. You can then save the model, just as you would for a Fast Train Model. For more information, go to Train Models.

Custom Model Settings

This section describes the settings you can configure to train a Custom Model. LandingLens applies some settings by default. Those defaults display on the Train Model page.

Data Transform

Data Transform allows you to resize or crop all the images in your dataset. This is an important feature because all training images need to be a standard input size for your model. Any Transforms applied here will be applied to the data you provide to your model for training and inference. It's important to be sure that the objects/regions you are trying to detect are clearly visible after applying these Transforms; otherwise, you can risk training poorly performing models.

If you add a Data Transform, you must add a Resize Transform first. This ensures that all data is consistently sized before cropping. Furthermore, resizing your data first prevents you from accidentally cropping outside the image boundaries of small images.

To add Data Transforms:

1. Click Add Transform.
   

   Add Transform
2. Click Resize or Crop accordingly.
   Resize or Crop 
3. Configure the settings.
4. Click Add to Pipeline to apply the change to all images in your dataset.
   Configure Settings and Add to Pipeline

Resize Images

When you use Custom Training, images are resized to the dimensions you provide. The maximum image dimensions are based on the Project Type, as described in the following table.

The maximum image size is measured by the area, not width and height. For example, the maximum image size for Object Detection is 1500x1500px. To get the area, multiply 1500 times 1500, so the equation looks like this:

1500 x 1500 = 2,250,000.

If you want to resize an Object Detection image, the area cannot exceed 2,250,000px. For example:

• A resize of 1000x600px will work because the area of this resize is 600,000px, which is fewer than 640,000px.
• A resize of 1500x1501px will not work because the area of this resize is 2,251,500px, which is greater than 2,250,000px.

Resize Settings

There are two types of Resize settings:

• Manual Resize: Enter the dimensions you want your images to be when you train the model. For example, if the original images are 1200x800 and the Resize is 512x512, all images will be resized into square images, even though the original images were in a 4:3 ratio. If you want to preserve the original aspect ratio, use Rescale with Padding.
  

  Manual Resize 
• Rescale with Padding: Enter the dimensions you want your images to be when you train the model. The image will be rescaled to that size and keep the original aspect ratio during resizing. This is done by adding padding of a constant pixel value of "0" to the image if the requested resize is not the same aspect ratio as the original image. If the resized image has the same aspect ratio, there will be no padded pixels.
  
  For example, say the original size of an image is 1200x800, and you rescale it to 512x512. Padding will be added around the rescaled image of 512x314 to make the height of the image 512. This is because the aspect ratio of the original image is 4:3, and the requested rescale is 1:1; thus, the extra space in the y dimension must be padded.
  
  
  Note:

  Rescale with Padding is set by default. You can adjust this setting as needed.
  

  Rescale with Padding

Resize: Large Image Support

If you are on LandingLens Enterprise, you can resize your images in Object Detection and Segmentation Projects to larger than the regular maximums. If you resize images to be larger than the following dimensions, the images are considered to be "large images":

• Object Detection: Over 1500x1500px
• Segmentation: Over 992x992px

With LandingLens Enterprise, you can resize your images up to these maximums:

• Object Detection: Up to 36MP
• Segmentation: Up to 36MP

If you resize your images to be "large images", you will only be able to deploy the model using LandingEdge. Cloud Deployment doesn't support models that were created with large images. Additionally, you won't be able to run the Predict tool on these models.

Crop

You can manually crop all images in your dataset. To do this, click and drag the sides of the box to create the crop you want to use for that image and click Add to Pipeline. Then use the left and right arrows to preview the crop of the other images in your dataset.

Crop Images

Data Augmentation

The Data Augmentation setting offers several pre-processing modifications that allow you to update all images in your dataset quickly. Here's how it works. First, you select a modification. Then enter the probability that the modification will take effect. For example, you can make it so there's a 50% chance that images will be rotated vertically.

Here are some benefits of adding Data Augmentations:

• You can reduce the time spent collecting data by augmenting images to "create" images you would have otherwise had to capture.
• Augmentations can create variations of your images so your model can be more robust when you deploy it.
• It reduces the chance of overfitting because the model will be shown random versions of images during each epoch cycle to prevent models from simply "memorizing" the data.

This section describes the types of Data Augmentation LandingLens offers.

Data Augmentation Settings

Random Brightness

Apply Random Brightness for a chance that the images in your dataset will randomly brighten.

For example, say your camera took several images at night, and the objects in the image are dark. You want your model also to learn how images will look if they are taken during the day. You may be asking yourself why this is important. When you deploy or train a model with poorly lit images, it can perform poorly because when the lighting changes look "different" to the model, it may not recognize the area to detect. In this case, you can use the Random Brightness modification to improve the lighting. 

This modification offers many settings, as described in the table below.

Random Brightness 

Blur / Motion Blur / Gaussian Blur

Apply one of the blur modifications for a chance that the images in your dataset will randomly blur. LandingLens offers these types of blurs:

• Blur: Makes the entire image look out of focus. This is a heavy blur.
• Motion Blur: Makes the entire image look like it's in motion.
• Gaussian Blur: Makes the entire image look out of focus. This is a natural blur.
  

This modification offers many settings, as described in the table below.

Blur 

Hue Saturation Value

Apply Hue Saturation Value for a chance that the color intensity of the images will change. 

For example, some objects come in multiple colors, like a car. You may want to change the hue (color of the image), so the model trains on these various colors.

This modification offers many settings, as described in the table below.

Hue Saturation Value 

Random Contrast

Apply Random Contrast for a chance that the tone of the images in your dataset will randomly change. Random Contrast helps to simulate different camera settings or environments that a deployed model may encounter.

• A low-contrast image will look washed out or slightly dimmed.
• A high-contrast image will look somewhat overexposed, and the colors will merge into each other.

This modification offers many settings, as described in the table below.

Random Contrast 

Horizontal Flip / Vertical Flip

• Apply Horizontal Flip for a chance that images in your dataset will rotate 180 degrees left or right. This is useful for models to "learn" what objects would look like if they are mirrored.
  Note:

  Horizontal Flip is set by default. You can adjust this setting as needed. 
• Apply Vertical Flip for a chance that images in your dataset will rotate 180 degrees up or down. This is useful for models to "learn" what objects would look like if they were upside-down.

These modifications offer the Probability setting. This setting indicates the likelihood that images will randomly flip.

Probability of a Horizontal Flip 

RandAugment

Select RandAugment to randomly apply a set of augmentations to your dataset. The possible augmentations are described in the table below.

RandAugment Settings

The RandAugment feature offers many settings, as described in the table below.

RandAugment

Random Rotate

Apply Random Rotate for a chance that images in your dataset will randomly rotate.

This modification offers many settings, as described in the table below.

Random Rotate 

Preview Effects

After configuring an advanced setting, you can preview how the images will look. To do this, click Preview Transform Effect or Preview Augmentation Effect accordingly.

Preview Effects Buttons

Preview of the Effects Applies 

Note:

When you set Data Augmentations, you provide a range of how those Data Augmentations affect images. You can see other possibilities by clicking Reapply.

Hyperparameter

When a model trains, it goes through an extensive process in the background. If you want to have more granular control over how your model trains, you can configure the Hyperparameters. A Hyperparameter is a setting used to control the speed and quality of the learning process. 

Here are some important notes to keep in mind:

• The Hyperparameter settings are intended for users who are familiar with Machine Learning. If you are unfamiliar with Machine Learning, it is recommended to use the default settings.
• Hyperparameters should only be adjusted to be data-driven. 
• Repeatedly training a model with different Hyperparameters (called "Hyperparameter Sweep") is NOT recommended as this can be expensive.

Hyperparameters have a couple of settings, as described in the table below.

Hyperparameter 

Custom Training FAQ

Is Advanced Training the same thing as Custom Training?

Yes.

When should I use Custom Training?

If the performance results of your model are unsatisfactory after a Fasting Train, switch to Custom Training.

I'm unfamiliar with machine learning. Can I use Custom Training?

We recommend using Fast Training so that LandingLens can fine-tune your model settings for you.

What happens if I switch from Custom Training to Fast Training?

I applied Custom Training settings and trained a model. If I want to train a second model with identical settings, can I click the "Train" (Fast Train) button?

The "Train" button uses default configurations which are different than "Train with Custom Options". However, if you click "Train with Custom Options", your previous settings will be remembered.