yayatang / pytorch_adapters

pytorch based adapters for datallop platform

Torch Dataloop Model Adapters

These are pytorch model adapters examples.

1. ResNet50 (resnet_adapter.py)

Full Model Management documentation here.
Jupyter notebooks with examples on inference and training here

Deployment

Add the model to your project:

import json
import dtlpy as dl

project_name = 'My Model'
project = dl.projects.get(project_name=project_name)

codebase = dl.GitCodebase(git_url='https://github.com/dataloop-ai/pytorch_adapters',
                          git_tag='master')

model = project.models.create(model_name='ResNet',
                              description='Dataloop ResNet implemented in pytorch',
                              output_type=dl.AnnotationType.CLASSIFICATION,
                              codebase=codebase,
                              tags=['torch'],
                              default_configuration={
                                  'weights_filename': 'model.pth',
                                  'input_size': 256,
                              },
                              default_runtime=dl.KubernetesRuntime(),
                              entry_point='resnet_adapter.py')

Create the pretrained snapshot (ImageNet):

project = dl.projects.get(project_name)

resnet_ver = '50'  # 18 etc..
bucket = dl.buckets.create(dl.BucketType.GCS,
                           gcs_project_name='viewo-main',
                           gcs_bucket_name='model-mgmt-snapshots',
                           gcs_prefix='ResNet{}'.format(resnet_ver))

# load the imagenet label mapping into the snapshot definitions

with open('imagenet_labels.json', 'r') as f:
    labels = json.load(f)

snapshot = model.snapshots.create(snapshot_name='pretrained-resnet{}'.format(resnet_ver),
                                  description='resnet{} pretrained on imagenet'.format(resnet_ver),
                                  tags=['pretrained', 'imagenet'],
                                  dataset_id=None,
                                  status='trained',
                                  configuration={'weights_filename': 'model.pth',
                                                 'id_to_label_map': labels,
                                  project_id=project.id,
                                  bucket=bucket,
                                  labels=list(labels.values())
                                  )

Clone and Edit

Fork this repo to change and add your special touch into the model code.

Contributions

Help us get better! We welcome any contribution and suggestion to this repo.
Open an issue for bug/features requests.