Use npy as a default format for prediction instead of json. (aws#63)

Author: nadiaya

src/sagemaker/pytorch/model.py

@@ -15,7 +15,7 @@
 from sagemaker.fw_utils import create_image_uri
 from sagemaker.model import FrameworkModel, MODEL_SERVER_WORKERS_PARAM_NAME
 from sagemaker.pytorch.defaults import PYTORCH_VERSION, PYTHON_VERSION
-from sagemaker.predictor import RealTimePredictor, json_serializer, json_deserializer
+from sagemaker.predictor import RealTimePredictor, npy_serializer, numpy_deserializer
 from sagemaker.utils import name_from_image
 
 
@@ -34,7 +34,7 @@ def __init__(self, endpoint_name, sagemaker_session=None):
                 Amazon SageMaker APIs and any other AWS services needed. If not specified, the estimator creates one
                 using the default AWS configuration chain.
         """
-        super(PyTorchPredictor, self).__init__(endpoint_name, sagemaker_session, json_serializer, json_deserializer)
+        super(PyTorchPredictor, self).__init__(endpoint_name, sagemaker_session, npy_serializer, numpy_deserializer)
 
 
 class PyTorchModel(FrameworkModel):

tests/integ/test_pytorch_train.py

@@ -43,14 +43,14 @@ def test_sync_fit_deploy(pytorch_training_job, sagemaker_session):
     with timeout(minutes=20):
         estimator = PyTorch.attach(pytorch_training_job, sagemaker_session=sagemaker_session)
         predictor = estimator.deploy(1, 'ml.c4.xlarge', endpoint_name=endpoint_name)
-        data = numpy.zeros(shape=(1, 1, 28, 28))
+        data = numpy.zeros(shape=(1, 1, 28, 28), dtype=numpy.float32)
         predictor.predict(data)
 
         batch_size = 100
-        data = numpy.random.rand(batch_size, 1, 28, 28)
+        data = numpy.random.rand(batch_size, 1, 28, 28).astype(numpy.float32)
         output = predictor.predict(data)
 
-        assert numpy.asarray(output).shape == (batch_size, 10)
+        assert output.shape == (batch_size, 10)
 
 
 def test_deploy_model(pytorch_training_job, sagemaker_session):
@@ -63,10 +63,10 @@ def test_deploy_model(pytorch_training_job, sagemaker_session):
         predictor = model.deploy(1, 'ml.m4.xlarge', endpoint_name=endpoint_name)
 
         batch_size = 100
-        data = numpy.random.rand(batch_size, 1, 28, 28)
+        data = numpy.random.rand(batch_size, 1, 28, 28).astype(numpy.float32)
         output = predictor.predict(data)
 
-        assert numpy.asarray(output).shape == (batch_size, 10)
+        assert output.shape == (batch_size, 10)
 
 
 def test_async_fit_deploy(sagemaker_session, pytorch_full_version):
@@ -92,10 +92,10 @@ def test_async_fit_deploy(sagemaker_session, pytorch_full_version):
             predictor = estimator.deploy(1, instance_type, endpoint_name=endpoint_name)
 
             batch_size = 100
-            data = numpy.random.rand(batch_size, 1, 28, 28)
+            data = numpy.random.rand(batch_size, 1, 28, 28).astype(numpy.float32)
             output = predictor.predict(data)
 
-            assert numpy.asarray(output).shape == (batch_size, 10)
+            assert output.shape == (batch_size, 10)
 
 
 # TODO(nadiaya): Run against local mode when errors will be propagated