The paper titled "RNC: Efficient RAM-aware NAS and Compilation for DNNs on Resource-Constrained Edge Devices" presents a novel framework aimed at optimizing deep neural networks (DNNs) specifically for deployment on edge devices with limited computational resources.
install Python (== 3.9) dependencies. Note: you need to use specific version of the library. See requirements.txt
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
conda create -n rnc python=3.9
conda activate rnc
pip install -r requirments.txtThis project requires PyTorch and torchvision. However, the specific versions you need depend on your CUDA installation. Please install the appropriate versions for your system.
The following options can be passed as command-line arguments:
--exp-id: Experiment ID (default: 'dev')--model: Model for packing (options: resnet, mbv3_small, squeezenet; default: "mbv3_small")
** Extra model can be defined in rram_nas_comp_pack/box_converter/models/ with pytorch for compilation**
The hardware configuration is defined in SimConfig.ini. Notably, the Xbar_Size parameter is crucial for defining the dimensions of the crossbar array:
- Xbar_Size:
128, 128(Row, Column)
** The quantization bit of weight is fixed to 8-bit **
Related options:
--num-crossbars: Number of crossbars (default: 10)--depth-split-factor: Factor for depthwise layer split (default: 1)--layer-threshold: Threshold for layer conflict in crossbar (default: 1)--xbar-max-load: Max number of layers in crossbar (default: 6)--no-pack-heuristic: Disable packing heuristic (default: False)--find-solution: Enable solution finding with minimun number of crossbar (default: False)
To perform packing, you can call the script test_packing.sh.
This script automates the packing process and generates following files in the log/packing/[exp_id]/ directory:
logs.log:
- Experiment Configuration: Initial parameters such as experiment ID, model type, number of crossbars, etc.
- Packing Statistics: Information on the number of boxes, crossbar constraints, utilized crossbars, leftover area, and utilization percentage.
- Latency Measurements: Recorded latency values for each evaluation, including ideal latency and latency ratios compared to both packed and baseline models.
- Timing Information: Breakdown of time taken for partitioning, packing, evaluation, and total execution time.
partition.json:
- idx: Index of the layer in the model.
- multiply: Number of times the layer is duplicated.
- box_raw: 2D box representation of the layer.
- boxes: List of layer boxes after partition.
- child_boxes: List of layer boxes after duplication with boxes.
pack.csv:
- Layer Index: Index of the layer in the model.
- Crossbar Index: Index of the crossbar where the layer is placed.
- Location (X,Y): Coordinates of the layer on the crossbar.
- Max X, Y: Maximum x and y coordinates of the layer, indicating its size.
sim.log:
- Simulation Log: Detailed log of the simulation process, including layer indices, crossbar indices, and timing information.
Related options:
--duplicate: Enable duplication (default: False)
Margin factor for duplication is automatically tuned by the algorithm from 0.95 with step of 0.05.
To perform packing and duplication, you can call the script test_packing_duplicate.sh.
This script automates the packing process and generates extra files in the log/packing/[exp_id]/ directory:
partition_optimized.json: simular to partition.json, but after duplication.
Related options:
--num-crossbars: Number of crossbars (default: 10)--num-sample: Number of samples to evaluate in cycle/latency evaluation (default: 8)--evolute-method: Evolution method (options: ndga; default: "ndga")--resume: Resume evolution from checkpoint (default: False)
The pretrained ofa MobileNetV3 small should be downloaded from here and put in ofa/checkpoint/mbv3_small_best.pth.tar.
To perform the evolution for RRAM-NAS, you can call the script run_evolution.sh.
This script automates the packing process and generates following files in the log/evolution/[exp_id]/ directory:
logs.log:
- Experiment Configuration: Initial parameters such as experiment ID, model type, number of crossbars, etc.
- Evolution Statistics: Information on the number of boxes, crossbar constraints, utilized crossbars, leftover area, and utilization percentage.
checkpoint:
- Checkpoint: Checkpoint of the evolution process.
evolute.csv:
- Generation: Generation of the evolution iternation.
This project builds upon the
We thank the authors for their valuable work.