image_classifier

Eager Mode example using torchvision image classifiers:

• TorchVision Image Classification Models : Download a pre-trained model state_dict for computer vision model that classifies images from the following :

  • Image Classification with AlexNet - https://download.pytorch.org/models/alexnet-owt-4df8aa71.pth
  • Image Classification with DenseNet161 - https://download.pytorch.org/models/densenet161-8d451a50.pth
  • Image Classification with ResNet18 - https://download.pytorch.org/models/resnet18-5c106cde.pth
  • Image Classification with SqueezeNet 1_1 - https://download.pytorch.org/models/squeezenet1_1-f364aa15.pth
  • Image Classification with VGG11 - https://download.pytorch.org/models/vgg11-bbd30ac9.pth

• Create a model archive file based on selected model or use the sample provided with above examples.

• You can use the index_to_name.json file for mapping predicted index to class or use a custom one.

• Create a torch model archive file using the above provided syntax command.

Sample commands to create a DenseNet161 eager mode model archive, register it on TorchServe and run image prediction

wget https://download.pytorch.org/models/densenet161-8d451a50.pth
torch-model-archiver --model-name densenet161 --version 1.0 --model-file examples/image_classifier/densenet_161/model.py --serialized-file densenet161-8d451a50.pth --handler image_classifier --extra-files examples/image_classifier/index_to_name.json
mkdir model_store
mv densenet161.mar model_store/
torchserve --start --model-store model_store --models densenet161=densenet161.mar
curl http://127.0.0.1:8080/predictions/densenet161 -T examples/image_classifier/kitten.jpg

TorchScript example using DenseNet161 image classifier:

• Save the Densenet161 model in as an executable script module or a traced script:

  • Save model using scripting

  #scripted mode
  from torchvision import models
  import torch
  model = models.densenet161(pretrained=True)
  sm = torch.jit.script(model)
  sm.save("densenet161.pt")

  • Save model using tracing

  #traced mode
  from torchvision import models
  import torch
  model = models.densenet161(pretrained=True)
  model.eval()
  example_input = torch.rand(1, 3, 224, 224)
  traced_script_module = torch.jit.trace(model, example_input)
  traced_script_module.save("densenet161.pt")

• Use following commands to register Densenet161 torchscript model on TorchServe and run image prediction

  torch-model-archiver --model-name densenet161_ts --version 1.0  --serialized-file densenet161.pt --extra-files examples/image_classifier/index_to_name.json --handler image_classifier
  mkdir model_store
  mv densenet161_ts.mar model_store/
  torchserve --start --model-store model_store --models densenet161=densenet161_ts.mar
  curl http://127.0.0.1:8080/predictions/densenet161 -T examples/image_classifier/kitten.jpg

TorchScript example using custom model and custom handler:

Following example demonstrates how to create and serve a custom NN model with custom handler archives in TorchServe :

• Digit recognition with MNIST