update data parallel integ test to use local mnist dataset

Author: ChoiByungWook

tests/data/smdistributed_dataparallel/mnist_pt.py

@@ -13,6 +13,7 @@
 from __future__ import print_function
 
 import argparse
+import os
 import time
 import torch
 import torch.nn as nn
@@ -150,8 +151,8 @@ def main():
     parser.add_argument(
         "--data-path",
         type=str,
-        default="/tmp/data",
-        help="Path for downloading " "the MNIST dataset",
+        default=os.environ["SM_CHANNEL_TRAINING"],
+        help="Path for downloading the MNIST dataset",
     )
 
     args = parser.parse_args()
@@ -186,7 +187,7 @@ def main():
         train_dataset = datasets.MNIST(
             data_path,
             train=True,
-            download=True,
+            download=False,  # True sets a dependency on an external site for our tests.
             transform=transforms.Compose(
                 [transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]
             ),

tests/integ/test_smdataparallel_pt.py

@@ -21,7 +21,7 @@
 
 from sagemaker.pytorch import PyTorch
 from tests.integ import timeout
-
+from tests.integ.test_pytorch import _upload_training_data
 
 smdataparallel_dir = os.path.join(
     os.path.dirname(__file__), "..", "data", "smdistributed_dataparallel"
@@ -51,4 +51,4 @@ def test_smdataparallel_pt_mnist(
     )
 
     with timeout.timeout(minutes=integ.TRAINING_DEFAULT_TIMEOUT_MINUTES):
-        estimator.fit(job_name=job_name)
+        estimator.fit({"training": _upload_training_data(estimator)}, job_name=job_name)