REFERENCE MODEL

(c) FABU America Inc. This is emulator to give a one-to-one hardware output for the hardware accelerator. Framework organization is inspired from the open source implementation of Faster-RCNN from here. The framework internally uses Tensorflow to do major portion of the computations and then converts the results to a format similar to the hardware output.

Requirements: Software

1. Python 2.7.15 (recommended: Anaconda)
2. Python packages like: python-opencv, easydict, matplotlib, cv2 etc...
3. Tensorflow.

What's New

• PVANet Obstacle network support
• PVANet Traffic Light network support
• Traffic Lane network support
• New Layers for SoC (Deconv, Upsampling, etc)
• Proposal Layer similar to hardware
• Post processing API to visualize the RTL output
• Support for multi-size images
• more hacks :-D

Reference Model Architecture

```
RefModel
	|-- data
		|-- ob
			|-- raw
				|-- Images ... 
			|-- ref_data_from_caffe
				|-- blobs_obstacle_006.cpk
				|-- pre_obstacle_006.cpk
		|-- tl 
			|-- Images ... 
	|-- lib
		|-- __init__.py
		|-- Makefile
		|-- make.sh
		|-- setup.py
		|-- networks
			|-- __init__.py
			|-- factory.py
			|-- network.py
			|-- pvanet_ob.py
			|-- pvanet_tl.py
		|-- fast_rcnn
			|-- __init__.py 
			|-- bbox_transform.py
			|-- config.py
			|-- nms_wrapper.py
			|-- test.py
		|-- lp
			|-- __init__.py
			|-- lp_op.cc 
			|-- lp_op.py
			|-- lp_op_test.py
			|-- lp.so
			|-- make_ref_hw.sh
		|-- nms
			|-- __init__.py
			|-- cpu_nms.c
			|-- cpu_nms.pyx	
			|--	cpu_nms.so
			|--	gpu_nms.hpp
			|--	gpu_nms.pyx
			|--	nms_kernel.cu

			|-- py_cpu_nms.py
		|-- pucocotools
		|-- roi_pooling_layer
			|--	__init__.py
			|--	roi_pooling_op.cc
			|--	roi_pooling_op_gpu.h
			|--	roi_pooling_op_grad.pyc
			|--	roi_pooling_op.pyc
			|--	roi_pooling.so
			|--	roi_pooling_op_gpu.cu.cc
			|-- roi_pooling_op_grad.py	
			|--	roi_pooling_op.py
			|--	roi_pooling_op_test.py
		|-- rpn_msr				
			|--	generate_anchors.py
			|--	generate.py
			|--	__init__.py
			|--	nms_hw.py
			|--	proposal_layer_tf.py
			|--	sqrootLookUp.py
		|-- utils
			|-- bbox.c	
			|--	blob.py
			|--	box_grid.py
			|--	cython_bbox.so
			|-- fixed_point.py	
			|--	__init__.py
			|-- nms.c	
			|--	timer.py
			|--	bbox.pyx
			|--	cython_nms.so
			|--	nms.py
			|--	nms.pyx
	|-- models
		|--obstacle_model.cpk
		|--tl_model.cpk
	|-- 
		|-- demo.py
		|-- _init_paths.py
		|-- pvatest.py
		|-- refModel.py
	|-- README.md	
```

Setup and Run

1. Load appropriate python module:

   module unload python
   module load python\2.7.15
   

2. cd $cnn_emu
3. python refModel --network pvanet_ob --hw_sim 1 --image 000.jpg --resolution_mode 0

Running tensorboard:

1. Set cfg.ENABLE_TENSORBOARD = True
2. Run python refModel (with net and other parameters)
3. run in command prompt to launch tensorboard: tensorboard --logdir=output
4. open browser and go to http://decacs1:6006
5. Enjoy :-D