feat: Added Python accuracy tests using Nox

Signed-off-by: Anurag Dixit <anuragd@nvidia.com>

docker/dist-accuracy-test.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+set -o nounset
+set -o errexit
+set -o pipefail
+set -e
+
+post=${1:-""}
+
+# fetch bazel executable
+BAZEL_VERSION=4.2.1
+ARCH=$(uname -m)
+if [[ "$ARCH" == "aarch64" ]]; then ARCH="arm64"; fi
+wget -q https://github.com/bazelbuild/bazel/releases/download/${BAZEL_VERSION}/bazel-${BAZEL_VERSION}-linux-${ARCH} -O /usr/bin/bazel
+chmod a+x /usr/bin/bazel
+export NVIDIA_TF32_OVERRIDE=0
+
+cd /opt/pytorch/torch_tensorrt
+cp cp /opt/pytorch/torch_tensorrt/docker/WORKSPACE.docker  /opt/pytorch/torch_tensorrt/WORKSPACE
+
+pip install --user --upgrade nox
+nox 

noxfile.py

@@ -0,0 +1,132 @@
+import nox
+import os
+
+# Use system installed Python packages
+PYT_PATH='/opt/conda/lib/python3.8/site-packages'
+
+# Root directory for torch_tensorrt. Set according to docker container by default
+TOP_DIR='/opt/pytorch/torch_tensorrt'
+
+# Download the dataset
+@nox.session(python=["3"], reuse_venv=True)
+def download_datasets(session):
+    session.chdir(os.path.join(TOP_DIR, 'examples/int8/training/vgg16'))
+    session.run_always('wget', 'https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz')
+    session.run_always('tar', '-xvzf', 'cifar-10-binary.tar.gz')
+    session.run_always('mkdir', '-p', 
+                        os.path.join(TOP_DIR, 'tests/accuracy/datasets/data'))
+    session.run_always('cp', '-rpf', 
+                        os.path.join(TOP_DIR, 'examples/int8/training/vgg16/cifar-10-batches-bin'),
+                        os.path.join(TOP_DIR, 'tests/accuracy/datasets/data/cidar-10-batches-bin'),
+                        external=True)
+
+# Download the model
+@nox.session(python=["3"], reuse_venv=True)
+def download_models(session):
+    session.install('timm')
+    session.chdir('tests/modules')
+    session.run_always('python', 
+                        'hub.py',
+                        env={'PYTHONPATH': PYT_PATH})
+
+# Train the model
+@nox.session(python=["3"], reuse_venv=True)
+def train_model(session):
+    session.chdir(os.path.join(TOP_DIR, 'examples/int8/training/vgg16'))
+    session.run_always('python', 
+                        'main.py', 
+                        '--lr', '0.01', 
+                        '--batch-size', '128', 
+                        '--drop-ratio', '0.15', 
+                        '--ckpt-dir', 'vgg16_ckpts', 
+                        '--epochs', '25',
+                        env={'PYTHONPATH': PYT_PATH})
+
+    # Export model
+    session.run_always('python',
+                        'export_ckpt.py',
+                        'vgg16_ckpts/ckpt_epoch25.pth',
+                        env={'PYTHONPATH': PYT_PATH})
+
+# Finetune the model
+@nox.session(python=["3"], reuse_venv=True)
+def finetune_model(session):
+    # Install pytorch-quantization dependency
+    session.install('pytorch-quantization', '--extra-index-url', 'https://pypi.ngc.nvidia.com')
+
+    session.chdir(os.path.join(TOP_DIR, 'examples/int8/training/vgg16'))
+    session.run_always('python', 
+                        'finetune_qat.py', 
+                        '--lr', '0.01', 
+                        '--batch-size', '128', 
+                        '--drop-ratio', '0.15', 
+                        '--ckpt-dir', 'vgg16_ckpts', 
+                        '--start-from', '25', 
+                        '--epochs', '26',
+                        env={'PYTHONPATH': PYT_PATH})
+
+    # Export model 
+    session.run_always('python',
+                        'export_qat.py',
+                        'vgg16_ckpts/ckpt_epoch26.pth',
+                        env={'PYTHONPATH': PYT_PATH})
+
+# Run PTQ tests
+@nox.session(python=["3"], reuse_venv=True)
+def ptq_test(session):
+    session.chdir(os.path.join(TOP_DIR, 'tests/py'))
+    session.run_always('cp', '-rf', 
+                        os.path.join(TOP_DIR, 'examples/int8/training/vgg16', 'trained_vgg16.jit.pt'), 
+                        '.',
+                        external=True)
+    tests = [
+        'test_ptq_dataloader_calibrator.py',
+        'test_ptq_to_backend.py',
+        'test_ptq_trt_calibrator.py'
+        ]
+    for test in tests:
+        session.run_always('python', test,
+                        env={'PYTHONPATH': PYT_PATH})
+
+# Run QAT tests
+@nox.session(python=["3"], reuse_venv=True)
+def qat_test(session):
+    session.chdir(os.path.join(TOP_DIR, 'tests/py'))
+    session.run_always('cp', '-rf', 
+                        os.path.join(TOP_DIR, 'examples/int8/training/vgg16', 'trained_vgg16_qat.jit.pt'), 
+                        '.',
+                        external=True)
+
+    session.run_always('python',
+                        'test_qat_trt_accuracy.py',
+                        env={'PYTHONPATH': PYT_PATH})
+
+# Run Python API tests
+@nox.session(python=["3"], reuse_venv=True)
+def api_test(session):
+    session.chdir(os.path.join(TOP_DIR, 'tests/py'))
+    tests = [
+            "test_api.py",
+            "test_to_backend_api.py"
+            ]
+    for test in tests:
+        session.run_always('python',
+                            test,
+                            env={'PYTHONPATH': PYT_PATH})
+
+# Clean up
+@nox.session(reuse_venv=True)
+def cleanup(session):
+    target = [
+        'examples/int8/training/vgg16/*.jit.pt',
+        'examples/int8/training/vgg16/vgg16_ckpts',
+        'examples/int8/training/vgg16/cifar-10-*',
+        'examples/int8/training/vgg16/data',
+        'tests/modules/*.jit.pt',
+        'tests/py/*.jit.pt'
+    ]
+
+    target = ' '.join(x for x in [os.path.join(TOP_DIR, i) for i in target])
+    session.run_always('bash', '-c',
+                        str('rm -rf ') + target,
+                        external=True)

tests/README.md

@@ -1,6 +1,9 @@
 # Tests
 
-Right now there are two types of tests. Converter level tests and Module level tests.
+Currently, following tests are supported:
+1. Converter level tests
+2. Module level tests
+3. Accuracy tests
 
 The goal of Converter tests are to tests individual converters againsts specific subgraphs. The current tests in `core/conveters` are good examples on how to write these tests. In general every converter should have at least 1 test. More may be required if the operation has switches that change the behavior of the op.
 
@@ -20,6 +23,24 @@ bazel test //tests --compilation_mode=dbg --test_output=errors --jobs=4 --runs_p
 
 `--jobs=4` is useful and is sometimes required to prevent too many processes to use GPU memory and cause CUDA out of memory issues.
 
+Additionally, accuracy tests are supported for Python backend using Nox. Please refer [dist-accuracy-test.sh](../docker/dist-accuracy-test.sh) for reference.
+```
+# To run complete Python accuracy + API tests
+nox
+
+nox -l
+```
+
+Note: Supported Python tests
+```
+* download_datasets-3
+* download_models-3
+* train_model-3
+* finetune_model-3
+* ptq_test-3
+* qat_test-3
+* cleanup
+```
 ### Testing using pre-built Torch-TensorRT library
 
 Currently, the default strategy when we run all the tests (`bazel test //tests`) is to build the testing scripts along with the full Torch-TensorRT library (`libtorchtrt.so`) from scratch. This can lead to increased testing time and might not be needed incase you already have a pre-built Torch-TensorRT library that you want to link against.