NOTE: If you have imported a model before, do not import it again. You can select it from the list of available models.
You can import original and the Open Model Zoo models. Click Import Model under the list of available models.
To import a model, follow the steps below:
NOTE: To learn about the conversion logic, see the Model Optimizer documentation.
Once you import a model, you are directed back to the Create Configuration page, where you can select the imported model and proceed to select a dataset.
DL Workbench supports the following frameworks whether uploaded from a local folder or imported from the Open Model Zoo.
Framework | Original Models | Open Model Zoo |
---|---|---|
OpenVINO™ | ✔ | ✔ |
Caffe* | ✔ | ✔ |
MXNet* | ✔ | ✔ |
ONNX* | ✔ | ✔ |
TensorFlow* | ✔ | ✔ |
PyTorch* | ✔ |
NOTE: Internet connection is required for this option. If you are behind a corporate proxy, set environment variables during the Install from Docker Hub* step.
To use a model from the Open Model Zoo, go to the Open Model Zoo tab. Select a model and click Import:
TIP: Precision of models from the Open Model Zoo can be changed in the Conversion step.
Once you click Import, the Convert Model to IR section opens.
NOTE: To learn about the import process, see the Model Downloader documentation.
In the Original Model tab, you can upload original models stored in files on your operating system. The uploading process depends on the framework of your model.
To import an OpenVINO™ model, select the framework in the drop-down list, upload an .xml
file and a .bin
file, provide the name, and click Import. Since the model is already in the IR format, select the imported model and proceed to select a dataset.
NOTE: To learn more about Caffe* models, refer to the article.
To import a Caffe model, select the framework in the drop-down list, upload an [.prototxt
] file and a .caffemodel
file, and provide the name.
Once you click Import, the tool analyzes your model and opens the Convert Model to IR form with prepopulated conversion settings fields, which you can change.
NOTE: To learn more about MXNet models, refer to the article.
To import an MXNet model, select the framework in the drop-down list, upload an .json
file and a .params
file, and provide the name.
Once you click Import, the tool analyzes your model and opens the Convert Model to IR form with prepopulated conversion settings fields, which you can change.
NOTE: To learn more about ONNX models, refer to the article.
To import an ONNX model, select the framework in the drop-down list, upload an .onnx
file, and provide the name.
Once you click Import, the tool analyzes your model and opens the Convert Model to IR form with prepopulated conversion settings fields, which you can change.
NOTE: To learn about the difference between frozen and nob-frozen TensorFlow models, refer to the Freeze Tensorflow models and serve on web.
To import a frozen TensorFlow model, select the framework in the drop-down list, upload a .pb
or .pbtxt
file, provide the name, and make sure the Is Frozen Model box is checked.
You can also check the Use TensorFlow Object Detection API box and upload a pipeline configuration file.
To import a non-frozen TensorFlow model, select the framework in the drop-down list, provide the name, and uncheck the Is Frozen Model box. Then select input files type: Checkpoint or MetaGraph.
If you select the Checkpoint file type, provide the following files:
.pb
or .pbtxt
file.checkpoint
fileIf you select the MetaGraph file type, provide the following files:
.meta
file.index
fileRegardless of a file type, you can also check the Use TensorFlow Object Detection API box and upload a pipeline configuration file.
Once you click Import, the tool analyzes your model and opens the Convert Model to IR form with prepopulated conversion settings fields, which you can change.
You are automatically directed to the Convert Model to IR step after uploading a model. Besides general framework-agnostic parameters, you might need to specify framework-specific or advanced framework-agnostic parameters.
NOTE: For details on converting process, refer to Converting a Model to Intermediate Representation.
Refer to the table below to learn about parameters shared by all frameworks.
Parameter | Values | Description |
---|---|---|
Batch number | 1-256 | How many images at a time are propagated to a neural network |
Precision | FP16 or FP32 | Precision of a converted model |
Original color space | RGB or BGR | Color space of an original model |
When importing TensorFlow* models, provide an additional pipeline configuration file and choose a model conversion configuration file. For details, refer to Import Frozen TensorFlow SSD MobileNet v2 COCO Tutorial.
NOTE: Input layers are required for MXNet* models.
You can use default input layers or cut a model by specifying the layers you want to consider as input ones. To change default input layers, provide information about the layers:
TIP: To add more than one layer, click the Add Input button. To remove layers, click on the red remove sign next to the name of a layer.
NOTE: To learn more about means and scales, see Converting a Model Using General Conversion Parameters.
You can use default output layers or cut a model by specifying the layers you want to consider as output ones. To change default input layers, check the Override Outputs box and provide the name of a layer.
TIP: To add more than one layer, click the Add Output button. To remove layers, click on the red remove sign next to the name of a layer.
To find out the names of layers, view text files with a model in an editor or visualize the model in the Netron neural network viewer.
NOTE: For more information on preprocessing parameters, refer to Converting a Model Using General Conversion Parameters.
To convert a Caffe* or an ONNX* model, provide the following information in the General Parameters section:
You can also use default values by checking the box in the Advanced Parameters section.
NOTE: For details on converting Caffe and ONNX models, refer to Converting a Caffe* Model and Converting an ONNX* Model.
Once you click Convert, you are directed back to the Create Configuration page, where you can select the imported model and proceed to select a dataset.
To convert an MXNet model, provide the following information in the General Parameters section:
In the same section, specify framework-specific parameters by checking the boxes:
Checking the Use Default Values box in the Advanced Parameters section allows you to choose to use inputs and/or outputs.
Once you click Convert, you are directed back to the Create Configuration page, where you can select the imported model and proceed to select a dataset.
NOTE: For details on converting MXNet models, refer to Converting an MXNet* Model.
To convert a TensorFlow model, provide the following information in the General Parameters section:
Checking the Use Default Values box in the Advanced Parameters section allows you to choose to use inputs and/or outputs.
Once you click Convert, you are directed back to the Create Configuration page, where you can select the imported model and proceed to select a dataset.
NOTE: For details on converting TensorFlow models, refer to Converting a TensorFlow* Model.
To convert a model from the Open Model Zoo, provide the precision in the General Parameters section.
Once you click Convert, you are directed back to the Create Configuration page, where you can select the imported model and proceed to select a dataset.
NOTE: If you are behind a corporate proxy, set environment variables during the Install from Docker* Hub step.