- Collect graph runtime speedup, time taken by optimizer, number of graphs explored (in candidate set).
examples/get_time.sh
shows how to run the benchmarks and get these results - To run GPU for ~2min each time before evaluating full graph runtime. This is to make sure GPU reaches a stable temperature, such that the measured results won't have a large variance due to temperature.
- Convert synthesized rules into format that can be taken by tamago. Run
python verify/read_rules.py graph_subst.pb
to convert the rules.
TASO optimizes the computation graphs of DNN models using automatically generated and verified graph transformations. For an arbitrary DNN model, TASO uses the auto-generated graph transformations to build a large search space of potential computation graphs that are equivalent to the original DNN model. TASO employs a cost-based search algorithm to explore the space, and automatically discovers highly optimized computation graphs. TASO outperforms the graph optimizers in existing deep learning frameworks by up to 3x.
See instructions to install TASO from source. We also provide prebuilt docker images with all dependencies pre-installed.
TASO can directly optimize any pre-trained DNN models in ONNX, TensorFlow, and PyTorch graph formats. TASO also provides a Python interface for optimizing arbitrary DNN architectures. TASO supports exporting the optimized computation graphs to ONNX, which can be directly used as inputs by most existing deep learning frameworks.
TASO can directly optimize pre-trained ONNX models, and this can be done in just a few lines of Python code. The following code snippet shows how to load a pre-trained DNN model from ONNX, optimize the model, and save the optimized model into a ONNX file.
import taso
import onnx
old_model = taso.load_onnx("/path/to/load/onnx/model")
taso_graph = taso.optimize(old_model)
new_model = taso.export_onnx(taso_graph)
onnx.save(new_model, "/path/to/save/new/onnx/model")
The optimized model has the same accuracy as the original and can be directly used by existing deep learning frameworks.
Some original and TASO-optimized ONNX files are available in the onnx
folder.
TASO can optimize TensorFlow models by converting the model to ONNX using tf2onnx.
- First, install
tf2onnx
from PyPi as follows or from source.
pip install -U tf2onnx
- Second, convert a TensorFlow model to ONNX using
tf2onnx
.
python -m tf2onnx.convert \
--saved-model /path/to/tensorflow/saved/model \
--output /path/to/onnx/model/file
- Third, use TASO to optimize the model in ONNX by following the above instructions.
PyTorch has built-in support for ONNX as a part of the torch.onnx package. TASO can directly optimize PyTorch models in the ONNX format.
TASO can also optimize arbitrary DNN architectures using the TASO Python interface.
The following code snippet builds the left-most DNN graph depicted in the figure. TASO automatically performs a series of non-trivial transformations, and eventually discovers the right-most DNN graph, which is 1.3x faster on a V100 GPU. More DNN examples are available in the examples
folder.
import taso
import onnx
#Build DNN model
graph = taso.new_graph()
input = graph.new_input(dims=(1,128,56,56))
w1 = graph.new_weight(dims=(128,128,3,3))
w2 = graph.new_weight(dims=(128,128,1,1))
w3 = graph.new_weight(dims=(128,128,3,3))
left = graph.conv2d(input=input, weight=w1, strides=(1,1), padding="SAME", activation="RELU")
left = graph.conv2d(input=input, weight=w3, strides=(1,1), padding="SAME")
right = graph.conv2d(input=input, weight=w2, strides=(1,1), padding="SAME", activation="RELU")
output = graph.add(left, right)
output = graph.relu(output)
#Optimize DNN model
new_graph = taso.optimize(graph)
onnx_model = taso.export_onnx(new_graph)
onnx.save(onnx_model, "/path/to/save/new/onnx/model")
-
Zhihao Jia, Oded Padon, James Thomas, Todd Warszawski, Matei Zaharia, and Alex Aiken. TASO: Optimizing Deep Learning Computation with Automated Generation of Graph Substitutions. In Proceedings of the Symposium on Operating Systems Principles (SOSP), Ontario, Canada, October 2019.
-
Zhihao Jia, James Thomas, Todd Warszawski, Mingyu Gao, Matei Zaharia, and Alex Aiken. Optimizing DNN Computation with Relaxed Graph Substitutions. In Proceedings of the Conference on Systems and Machine Learning (SysML), Palo Alto, CA, April 2019.