Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[REVIEW] Adding RAPIDS <-> DLFrameworks Jupyter Notebook #266

Open
wants to merge 4 commits into
base: branch-0.12
Choose a base branch
from
Open

[REVIEW] Adding RAPIDS <-> DLFrameworks Jupyter Notebook #266

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3b8b3ed
Adding RAPIDS <-> DLFrameworks Jupyter Notebook
awthomp Jan 17, 2020
4ad1f9d
Update README to include interoperability notebook
awthomp Jan 17, 2020
5071b58
Updating numba and cupy to TF examples with fixed tfdlpack-gpu
awthomp Jan 19, 2020
1ad9492
fix typo and add pytorch <-> tensorflow
awthomp Jan 19, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,6 +97,7 @@ Below is a listing of the notebooks in this repository. Each row will tell you
|cuspatial -> accelerate_geospatial_processing | [accelerate_geospatial_processing](blog_notebooks/cuspatial/trajectory_clustering.ipynb) | This is the notebook for blog [cuSpatial Accelerates Geospatial and Spatiotemporal Processing](https://medium.com/rapids-ai/releasing-cuspatial-to-accelerate-geospatial-and-spatiotemporal-processing-b686d8b32a9) by Milind Naphade, Jianting Zhang, Shuo Wang, Thomson Comer, Josh Paterson, Keith Kraus, Mark Harris, and Sujit Biswas. This notebook showcases cuSpatial benchmarking of directed Hausdorff distance for computing trajectory clustering on a large dataset. | SG | Trajectories Data and target_intersection.png |
| randomforest | [fruits_rf_notebook](blog_notebooks/randomforest/fruits_rf_notebook.ipynb) | This is the notebook for blog [GPU-accelerated Random Forest]() by Vishal Mehta, Myrto Papadopoulou, Thejaswi Rao. This notebook showcases how to use GPU accelerated Random Forest Classification in cuML. The fruit dataset used is Self generated and used as an example in the [Blog](https://medium.com/rapids-ai/accelerating-random-forests-up-to-45x-using-cuml-dfb782a31bea) | SG | Self Generated
| mortgage_deep_learning | [mortgage_e2e_deep_learning](blog_notebooks/mortgage_deep_learning/mortgage_e2e_deep_learning.ipynb) | **Archive Only.** This end to end notebook for the blog, [Using RAPIDS with PyTorch](https://medium.com/rapids-ai/using-rapids-with-pytorch-e602da018285), by Even Oldridge, combines the RAPIDS GPU data processing with a PyTorch deep learning neural network to predict mortgage loan delinquency. | MG | [RAPIDS-IPUMS Mortgage Dataset](https://rapidsai.github.io/demos/datasets/mortgage-data)
| interoperability | [gpu_interop_dlframeworks](blog_notebooks/interoperability/gpu_interop_dlframeworks.ipynb) | Demonstration of passing data between GPU enabled libraries and DL frameworks by Adam Thompson | SG | Self Generated |

## Conference Notebooks:

Expand Down
291 changes: 291 additions & 0 deletions blog_notebooks/interoperability/gpu_interop_dlframeworks.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,291 @@
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Sharing is Caring: GPU Interoperability and <3 of All Frameworks"
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"import cupy as cp\n",
"import numpy as np\n",
"from numba import cuda\n",
"\n",
"# PyTorch 1.4 supports direct __cuda_array_interface__ handoff.\n",
"import torch\n",
"\n",
"# RFC: https://github.com/tensorflow/community/pull/180\n",
"# !pip install tfdlpack-gpu\n",
"import tfdlpack"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Create GPU Arrays and Move to DL Frameworks with `__cuda_array_interface__`"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Frameworks that leverage the `__cuda_array_interface__` can be seamlessly transferred from compatiable libraries (CuPy, Numba, cuSignal, etc) directly, without using an intermediate Tensor format like [DLPack](https://github.com/dmlc/dlpack)\n",
"\n",
"**CuPy <-> PyTorch**"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"CuPy GPU Array Pointer: (140657360371712, False)\n",
"PyTorch GPU Tensor Pointer: (140657360371712, False)\n",
"CuPy GPU Pointer: (140657360371712, False)\n"
]
}
],
"source": [
"# CuPy - GPU Array (like NumPy!)\n",
"gpu_arr = cp.random.rand(10_000, 10_000)\n",
"\n",
"# Look at pointer\n",
"print('CuPy GPU Array Pointer: ', gpu_arr.__cuda_array_interface__['data'])\n",
"\n",
"# Migrate from CuPy to PyTorch\n",
"torch_arr = torch.as_tensor(gpu_arr, device='cuda')\n",
"\n",
"# Look at pointer -- it's the same as the CuPy array above!\n",
"print('PyTorch GPU Tensor Pointer: ', torch_arr.__cuda_array_interface__['data'])\n",
"\n",
"# Migrate from PyTorch to CuPy\n",
"cupy_arr = cp.asarray(torch_arr)\n",
"\n",
"# Look at pointer\n",
"print('CuPy GPU Pointer: ', cupy_arr.__cuda_array_interface__['data'])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"**Numba CUDA <-> PyTorch**"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Numba GPU Array Pointer: (140655749758976, False)\n",
"PyTorch GPU Tensor Pointer: (140655749758976, False)\n",
"Numba GPU Pointer: (140655749758976, False)\n"
]
}
],
"source": [
"# NumPy - CPU Array\n",
"cpu_arr = np.random.rand(10_000, 10_000)\n",
"\n",
"# Use Numba to move to GPU\n",
"numba_gpu_arr = cuda.to_device(cpu_arr)\n",
"\n",
"# Migrate from Numba, used for custom CUDA JIT kernels to PyTorch\n",
"torch_arr_numba = torch.as_tensor(numba_gpu_arr, device='cuda')\n",
"\n",
"# Migrate from PyTorch back to Numba\n",
"numba_arr_from_torch = cuda.to_device(torch_arr_numba)\n",
"\n",
"# Pointer love again\n",
"print('Numba GPU Array Pointer: ', numba_gpu_arr.__cuda_array_interface__['data'])\n",
"print('PyTorch GPU Tensor Pointer: ', torch_arr_numba.__cuda_array_interface__['data'])\n",
"print('Numba GPU Pointer: ', numba_arr_from_torch.__cuda_array_interface__['data'])"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Create GPU Arrays and Move to DL Frameworks with DLPack"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Not all major frameworks currently support the `__cuda_array_interface__`, cough, [TensorFlow](https://www.tensorflow.org/). We can use the aforementioned DLPack as a bridge between the GPU ecosystem and TensorFlow with `tfdlpack`. See [this RFC](https://github.com/tensorflow/community/pull/180) for more information.\n",
"\n",
"Optional: Allow GPU growth in TensorFlow or TF will take over the entire GPU."
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [],
"source": [
"!export TF_FORCE_GPU_ALLOW_GROWTH=false"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"**CuPy <-> TensorFlow**"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"/job:localhost/replica:0/task:0/device:GPU:0\n"
]
}
],
"source": [
"# CuPy - GPU Array (like NumPy!)\n",
"gpu_arr = cp.random.rand(10_000, 10_000)\n",
"\n",
"# Use CuPy's built in `toDlpack` function to move to a DLPack capsule\n",
"dlpack_arr = gpu_arr.toDlpack()\n",
"\n",
"# Use `tfdlpack` to migrate to TensorFlow\n",
"tf_tensor = tfdlpack.from_dlpack(dlpack_arr)\n",
"\n",
"# Confirm TF tensor is on GPU\n",
"print(tf_tensor.device)\n",
"\n",
"# Use `tfdlpack` to migrate back to CuPy\n",
"dlpack_capsule = tfdlpack.to_dlpack(tf_tensor)\n",
"cupy_arr = cp.fromDlpack(dlpack_capsule)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"**Numba CUDA <-> TensorFlow**"
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"/job:localhost/replica:0/task:0/device:GPU:0\n"
]
}
],
"source": [
"# Reset CUDA memory\n",
"cuda.close()\n",
"\n",
"# NumPy - CPU Array\n",
"cpu_arr = np.random.rand(10_000, 10_000)\n",
"\n",
"# Use Numba to move to GPU\n",
"numba_gpu_arr = cuda.to_device(cpu_arr)\n",
"\n",
"# Use CuPy's asarray function and toDlpack to create DLPack capsule. There are multiple other ways to do this (i.e. PyTorch Utils)\n",
"dlpack_arr = cp.asarray(numba_gpu_arr).toDlpack()\n",
"\n",
"# Migrate from Numba, used for custom CUDA JIT kernels to PyTorch\n",
"tf_tensor = tfdlpack.from_dlpack(dlpack_arr)\n",
"\n",
"# Confirm TF tensor is on GPU\n",
"print(tf_tensor.device)\n",
"\n",
"# Use `tfdlpack` to migrate back to Numba\n",
"dlpack_capsule = tfdlpack.to_dlpack(tf_tensor)\n",
"numba_arr = cuda.to_device(cp.fromDlpack(dlpack_capsule))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"**PyTorch <-> TensorFlow**"
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"/job:localhost/replica:0/task:0/device:GPU:0\n"
]
}
],
"source": [
"import torch\n",
"import tfdlpack\n",
"from torch.utils import dlpack as th_dlpack\n",
"\n",
"# Torch - GPU Array\n",
"gpu_arr = torch.rand(10_000, 10_000).cuda()\n",
"\n",
"# Use Torch's DLPack function to get DLPack Capsule\n",
"dlpack_arr = th_dlpack.to_dlpack(gpu_arr)\n",
"\n",
"# Use `tfdlpack` to migrate to TensorFlow\n",
"tf_tensor = tfdlpack.from_dlpack(dlpack_arr)\n",
"\n",
"# Confirm TF tensor is on GPU\n",
"print(tf_tensor.device)\n",
"\n",
"# Use `tfdlpack` to migrate back to PyTorch\n",
"dlpack_capsule = tfdlpack.to_dlpack(tf_tensor)\n",
"torch_arr = th_dlpack.from_dlpack(dlpack_capsule)"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.3"
}
},
"nbformat": 4,
"nbformat_minor": 4
}