Skip to content

Latest commit

 

History

History
115 lines (81 loc) · 2.78 KB

my-models.md

File metadata and controls

115 lines (81 loc) · 2.78 KB
Raw

In this section you can view and manage your models.

Models are added to this list in a few ways:

  • From the the Models list in the Explore section

  • Training will automatically save the last checkpoint

  • You can manually select a checkpoint after training

  • Manual upload

Models list

At the "Neural networks" page you will see all the models in your current workspace.

You can view the following fields:

  1. Creation date

  2. Model name

  3. Associated NN Architecture (Plugin)

  4. Storage - where the model weights are stored: Supervisely Server or some Node from your Cluster

Managing your models

Click the "three dots" icon to see all available actions for the model.

 

  1. Goto task - redirects to the task page and filters all tasks. User will be able to get all necessary information about training: logs, charts, checkpoints, etc.

  2. Edit - change the model name

  3. Upload to Supervisely - if the model is stored on some Node from Cluster, user can upload this model to the Supervisely Server. That allows using this model on other computational Nodes. Read mode on NN lifecycle

  4. Archive

  5. Config - training configuration (it is useful to see all training parameters including classes the model predicts).

Here is the example of the entire training configuration for our custom YOLO V3.

{
  "settings": {
    "lr": 0.0001,
    "epochs": 15,
    "batch_size": {
      "train": 12
    },
    "input_size": {
      "width": 256,
      "height": 256
    },
    "bn_momentum": 0.07,
    "gpu_devices": [
      0,
      1,
      2,
      3
    ],
    "data_workers": {
      "train": 3
    },
    "dataset_tags": {
      "train": "train"
    },
    "subdivisions": {
      "train": 1
    },
    "print_every_iter": 10,
    "weights_init_type": "transfer_learning",
    "enable_augmentations": false
  },
  "out_classes": [
    {
      "color": "#82789c",
      "shape": "rectangle",
      "title": "kiwi_bbox"
    },
    {
      "color": "#57c0cb",
      "shape": "rectangle",
      "title": "lemon_bbox"
    }
  ],
  "class_title_to_idx": {
    "kiwi_bbox": 0,
    "lemon_bbox": 1
  }
}

  • settings - training parameters

  • out_classes - classes model predicts

  • class_title_to_idx - mapping from class title to class index

Downloading models

If the model is stored on some Node from you Cluster, you have to upload it to the Supervisely Server (manage models). After that the "Download" button will be available. As a result user will download .tar archive with model weights.

Starting training / inference

From this page you can start training or inference. Please read corresponding chapters to get more information.