Update TS Archiver to v0.3.1 / Integ tests (#95)

* Update Archiver to v0.3.1 / Integ tests

* Fix Unit tests

* Integ test error : .dockercfg: /home/ubuntu is not defined

* Install git ?

* Git install

* Fix path

* Fix unit tests

* Ignore GPU tests

* Disable more tests

* Update Dockerfile.dlc.cpu

* Update Dockerfile.pytorch

* Update Dockerfile.dlc.gpu

* Update buildspec.yml

* Update Dockerfile for Model Archiver 0.3.1

* Update boto>=1.10.44

Co-authored-by: Ubuntu <[email protected]>

Author: dk19y

setup.py

@@ -54,7 +54,7 @@ def read(fname):
     # different packages for different variants
     install_requires=['numpy', 'retrying', 'sagemaker-inference>=1.3.1'],
     extras_require={
-        'test': ['boto3==1.10.32', 'coverage==4.5.3', 'docker-compose==1.23.2', 'flake8==3.7.7', 'Flask==1.1.1',
+        'test': ['boto3>=1.10.44', 'coverage==4.5.3', 'docker-compose==1.23.2', 'flake8==3.7.7', 'Flask==1.1.1',
                  'mock==2.0.0', 'pytest==4.4.0', 'pytest-cov==2.7.1', 'pytest-xdist==1.28.0', 'PyYAML==3.10',
                  'sagemaker==1.56.3', 'sagemaker-containers>=2.5.4', 'six==1.12.0', 'requests==2.20.0',
                  'requests_mock==1.6.0', 'torch==1.6.0', 'torchvision==0.7.0', 'tox==3.7.0']

src/sagemaker_pytorch_serving_container/torchserve.py

@@ -42,7 +42,6 @@
 )
 DEFAULT_TS_MODEL_DIRECTORY = os.path.join(os.getcwd(), ".sagemaker", "ts", "models")
 DEFAULT_TS_MODEL_NAME = "model"
-DEFAULT_TS_MODEL_SERIALIZED_FILE = "model.pth"
 DEFAULT_TS_CODE_DIR = "code"
 DEFAULT_HANDLER_SERVICE = "sagemaker_pytorch_serving_container.handler_service"
 
@@ -117,14 +116,12 @@ def _adapt_to_ts_format(handler_service):
         DEFAULT_TS_MODEL_NAME,
         "--handler",
         handler_service,
-        "--serialized-file",
-        os.path.join(environment.model_dir, DEFAULT_TS_MODEL_SERIALIZED_FILE),
         "--export-path",
         DEFAULT_TS_MODEL_DIRECTORY,
-        "--extra-files",
-        os.path.join(environment.model_dir, DEFAULT_TS_CODE_DIR, environment.Environment().module_name + ".py"),
         "--version",
         "1",
+        "--extra-files",
+        os.path.join(environment.model_dir)
     ]
 
     logger.info(model_archiver_cmd)

test/container/1.6.0/Dockerfile.dlc.cpu

@@ -1,6 +1,8 @@
 ARG region
 FROM 763104351884.dkr.ecr.$region.amazonaws.com/pytorch-inference:1.6.0-cpu-py3
 
+RUN pip install --upgrade torch-model-archiver==0.3.1
+
 COPY dist/sagemaker_pytorch_inference-*.tar.gz /sagemaker_pytorch_inference.tar.gz
 RUN pip install --upgrade --no-cache-dir /sagemaker_pytorch_inference.tar.gz && \
     rm /sagemaker_pytorch_inference.tar.gz

test/container/1.6.0/Dockerfile.dlc.gpu

@@ -1,6 +1,8 @@
 ARG region
 FROM 763104351884.dkr.ecr.$region.amazonaws.com/pytorch-inference:1.6.0-gpu-py3
 
+RUN pip install --upgrade torch-model-archiver==0.3.1
+
 COPY dist/sagemaker_pytorch_inference-*.tar.gz /sagemaker_pytorch_inference.tar.gz
 RUN pip install --upgrade --no-cache-dir /sagemaker_pytorch_inference.tar.gz && \
     rm /sagemaker_pytorch_inference.tar.gz

test/container/1.6.0/Dockerfile.pytorch

@@ -3,7 +3,8 @@ FROM pytorch/pytorch:1.6.0-cuda10.1-cudnn7-runtime
 LABEL com.amazonaws.sagemaker.capabilities.accept-bind-to-port=true
 LABEL com.amazonaws.sagemaker.capabilities.multi-models=true
 
-ARG TS_VERSION=0.1.1
+ARG TS_VERSION=0.3.1
+ARG TS_ARCHIVER_VERSION=0.3.1
 
 ENV SAGEMAKER_SERVING_MODULE sagemaker_pytorch_serving_container.serving:main
 ENV TEMP=/home/model-server/tmp
@@ -25,7 +26,7 @@ RUN conda install -c conda-forge opencv==4.0.1 \
     && ln -s /opt/conda/bin/pip /usr/local/bin/pip3
 
 RUN pip install torchserve==$TS_VERSION \
- && pip install torch-model-archiver==$TS_VERSION
+ && pip install torch-model-archiver==$TS_ARCHIVER_VERSION
 
 COPY dist/sagemaker_pytorch_inference-*.tar.gz /sagemaker_pytorch_inference.tar.gz
 RUN pip install --no-cache-dir /sagemaker_pytorch_inference.tar.gz && \

test/integration/__init__.py

@@ -28,33 +28,33 @@
 model_cpu_dir = os.path.join(mnist_path, cpu_sub_dir)
 mnist_cpu_script = os.path.join(model_cpu_dir, code_sub_dir, 'mnist.py')
 model_cpu_tar = file_utils.make_tarfile(mnist_cpu_script,
-                                        os.path.join(model_cpu_dir, "model.pth"),
+                                        os.path.join(model_cpu_dir, "torch_model.pth"),
                                         model_cpu_dir,
                                         script_path="code")
 
 model_cpu_1d_dir = os.path.join(model_cpu_dir, '1d')
 mnist_1d_script = os.path.join(model_cpu_1d_dir, code_sub_dir, 'mnist_1d.py')
 model_cpu_1d_tar = file_utils.make_tarfile(mnist_1d_script,
-                                           os.path.join(model_cpu_1d_dir, "model.pth"),
+                                           os.path.join(model_cpu_1d_dir, "torch_model.pth"),
                                            model_cpu_1d_dir,
                                            script_path="code")
 
 model_gpu_dir = os.path.join(mnist_path, gpu_sub_dir)
 mnist_gpu_script = os.path.join(model_gpu_dir, code_sub_dir, 'mnist.py')
 model_gpu_tar = file_utils.make_tarfile(mnist_gpu_script,
-                                        os.path.join(model_gpu_dir, "model.pth"),
+                                        os.path.join(model_gpu_dir, "torch_model.pth"),
                                         model_gpu_dir,
                                         script_path="code")
 
 model_eia_dir = os.path.join(mnist_path, eia_sub_dir)
 mnist_eia_script = os.path.join(model_eia_dir, 'mnist.py')
 model_eia_tar = file_utils.make_tarfile(mnist_eia_script,
-                                        os.path.join(model_eia_dir, "model.pth"),
+                                        os.path.join(model_eia_dir, "torch_model.pth"),
                                         model_eia_dir)
 
 call_model_fn_once_script = os.path.join(model_cpu_dir, code_sub_dir, 'call_model_fn_once.py')
 call_model_fn_once_tar = file_utils.make_tarfile(call_model_fn_once_script,
-                                                 os.path.join(model_cpu_dir, "model.pth"),
+                                                 os.path.join(model_cpu_dir, "torch_model.pth"),
                                                  model_cpu_dir,
                                                  "model_call_model_fn_once.tar.gz",
                                                  script_path="code")

test/resources/mnist/model_cpu/1d/code/mnist_1d.py

@@ -42,6 +42,6 @@ def forward(self, x):
 
 def model_fn(model_dir):
     model = torch.nn.DataParallel(Net())
-    with open(os.path.join(model_dir, 'model.pth'), 'rb') as f:
+    with open(os.path.join(model_dir, 'torch_model.pth'), 'rb') as f:
         model.load_state_dict(torch.load(f))
     return model

test/resources/mnist/model_cpu/1d/model.pth renamed to test/resources/mnist/model_cpu/1d/torch_model.pth

@@ -52,6 +52,6 @@ def forward(self, x):
 def model_fn(model_dir):
     logger.info('model_fn')
     model = torch.nn.DataParallel(Net())
-    with open(os.path.join(model_dir, 'model.pth'), 'rb') as f:
+    with open(os.path.join(model_dir, 'torch_model.pth'), 'rb') as f:
         model.load_state_dict(torch.load(f))
     return model

test/resources/mnist/model_cpu/code/mnist.py

@@ -38,11 +38,11 @@ def model_fn(model_dir):
     logger.info('model_fn: Loading model with TorchScript from {}'.format(model_dir))
     # Scripted model is serialized with torch.jit.save().
     # No need to instantiate model definition then load state_dict
-    model = torch.jit.load('model.pth')
+    model = torch.jit.load('torch_model.pth')
     return model
 
 
 def save_model(model, model_dir):
     logger.info("Saving the model to {}.".format(model_dir))
-    path = os.path.join(model_dir, 'model.pth')
+    path = os.path.join(model_dir, 'torch_model.pth')
     torch.jit.save(model, path)

test/resources/mnist/model_cpu/model.pth renamed to test/resources/mnist/model_cpu/torch_model.pth

@@ -52,6 +52,6 @@ def forward(self, x):
 def model_fn(model_dir):
     logger.info('model_fn')
     model = torch.nn.DataParallel(Net())
-    with open(os.path.join(model_dir, 'model.pth'), 'rb') as f:
+    with open(os.path.join(model_dir, 'torch_model.pth'), 'rb') as f:
         model.load_state_dict(torch.load(f))
     return model

test/resources/mnist/model_eia/mnist.py

@@ -145,16 +145,12 @@ def test_adapt_to_ts_format(path_exists, make_dir, subprocess_check_call, set_py
         torchserve.DEFAULT_TS_MODEL_NAME,
         "--handler",
         handler_service,
-        "--serialized-file",
-        os.path.join(environment.model_dir, torchserve.DEFAULT_TS_MODEL_SERIALIZED_FILE),
         "--export-path",
         torchserve.DEFAULT_TS_MODEL_DIRECTORY,
-        "--extra-files",
-        os.path.join(environment.model_dir,
-                     torchserve.DEFAULT_TS_CODE_DIR,
-                     environment.Environment().module_name + ".py"),
         "--version",
         "1",
+        "--extra-files",
+        environment.model_dir
     ]
 
     subprocess_check_call.assert_called_once_with(model_archiver_cmd)