RedEye is a vision sensor designed to execute early stages of a deep convolutional neural network (ConvNet) in the analog domain. This repo is a modification of Caffe for three main goals:
- Simulate and visualize the noise vs. energy tradeoffs of analog ConvNet processing
- Train and test RedEye-specific ConvNets
- Build RedEye-ConvNet programs to interface with RedEye hardware (PRIVATE).
redeye_sim Modifications to Caffe
- Gaussian Noise Layer: Parameterized with SNR, this layer simulates the signal infidelity of analog processing.
- Uniform Noise Layer: Parameterized with SNR, this layer uses uniform noise to simulate the signal infidelity of the quantization of the analog-to-digital converter.
- Quantization Layer: Parameterized with bits, this layer truncates data to simulate the quantization of the analog-to-digital converter.
- Weight Digitization: This represents the bit-resolution of the of the fixed-point model weights used in the mixed-signal multiply-accumulate unit.
- (Un)Gamma: Converts nonlinear input pixel values into linear luminance values.
- You can either choose to use the caffe source included in redeye_sim, or download and replace the
redeye_sim/caffedirectory with the latest caffe source. For the latter choice, execute
scriptsdirectory to add redeye_sim modifications.
- Copy the
make redeyecommand under the root redeye_sim directory.
- Download the ImageNet dataset. Execute
create_lmdb.shto generate lmdb files for train and validation.
- Download the pre-trained GoogLeNet caffemodel. Note: keep paths in
data.configconsistent when making data and run workflow.
Tuning and Testing
redeye_sim/simulation contains workflows to train the models (
tune.py) and generates statistics (
validate.py) for their performance. The workflows automatically inject noise layers to existing models, simulating analog noise behavior. Each workflow sweeps through a range of given noise SNRs, configured as follows:
- Configure datapath directories, batch sizes, and data layer noise in
configs/data.config. Recommended: use soft links for large data files.
- Configure parameter intervals and prototxt templates in
- Specify the sweeping interval of Gaussian noise “gaussian_intvl” and uniform noise “uniform_intvl”, which represents circuit noise and ADC noise, respectively.
- Specify the digitization bit-resolution “digitization_bits” of fixed-point weights.
- Specify the depth (number of convolution layers) “depth” of analog execution.
redeye_sim/simulation also contains a
workflow.py to procedurally train and test models in the SNR ranges. An example run of the complete workflow: under
python workflow.py --test_iter 500 --tune_iter 12000 --dest ../stats/test.csv --g 15,30 --q 15,25. This will sweep the SNR space and run fine tune for each noise-contaminated model for 12000 iterations, write results to
test.csv, and save the
test.png plot under
After too many runs, you can run
/tools/cleaner.sh to clean up the messy .prototxt’s and .caffemodel’s if you do not need them anymore. (Keep them if you want to save time running later)
Limitations & Assumptions
- The training data for simulation does not account for real-world image signal effects, e.g., over-exposure, white-balance.
- The simulation does not include circuit non-idealities, e.g., offset, gain, signal decay.
- Energy estimates for timing circuitry and signal propagation are not included.
Common Compilation Errors
If you are not using
configs, here are some common errors that may occur:
- Cannot find
Python.h: look in
- Cannot find
cblas.h: Modify the following lines in
BLAS_INCLUDE ?=in else statement, modify the path to your actual path that contains
cblas.h. This path can be found by unix command
find / -name cblas.h)
ifneq (,$(findstring version: 6,$(XCODE_CLT_VER))) BLAS_INCLUDE ?= /System/Library/Frameworks/Accelerate.framework/Versions/Current/Frameworks/vecLib.framework/Headers/ LDFLAGS += -framework Accelerate else BLAS_INCLUDE ?= /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/Headers/ #LDFLAGS += -framework vecLib LDFLAGS += -framework Accelerate endif
- Cannot load
image _caffe.so: Ensure that
libc++11. Ensure that
libc++11is correctly set in
NVCCFLAGS. These changes should be made in
If importing python results in a segfault, you are not using the right python for which caffe is installed. This is usually
This repo is distributed for free usage.