A visualization tool that takes parsed neuro-network info as input, and generates DOT code that can visualize the model via a DOT visualizer (e.g. https://dreampuf.github.io/GraphvizOnline/#).
An example of the resulting graph can be found in the example directory.
pip install graphvizpython3 visualizer.py prof.csv dep.csv optional_file_suffixWhere prof.csv contains the following profiling info:
- Layer name
- Average time consumption (in second)
- Output size (in MB)
- Average memory consumption
and dep.csv defines the inter-layer dependency of the model:
- Source
- Destination
This visualizer generates a result_DOT_code_<suffix>.txt.
- In
prof.csv, measure the data for every layer only once. - In
dep.csv, measure the dependency for every output, but only before it is used in another layer. - When recording the dependency, remember to include concat operations.
The profile, dependency and sample partitioning files for those models are stored in the paper directory. To generate DOT code for those models, copy the corresponding csv files to the root directory of this project, and run the visualizer.
You can also copy the relevant csv files from running the optimizer in the obj-detection-optimizer repo
A coloring tool that visualizes the partition of a neuro-network by different colors. Please use a DOT visualizer as above-mentioned to color the graph.
python3 colorer.py prof.csv dep.csv part.csv optional_file_suffix Where part.csv contains a partition of a model:
- Layer name
- device ID
The colorer supports up to eight (8) devices. For partitions that require more devices, please add more colors to line 51 in colorer.py:
colors = ["red", "blue", "black", "yellow", "green", "purple", "white", "orange"]There are several plotters located in child directories. See the corresponding readme files in those directories for more info.