The project implements an image-based Deep Learning model to detect and classify tincans and bottles. The algorithm is based on Mobilenet SSD-Lite with post training quantization.
the whole model can be run with the command:
make clean all run
by default it will run using the input coming from the camera. It is possible to give some flags to the make file in order to change the this behaviour in particular:
make clean all run FROM_JTAG=0 IMAGE=path/to/image.ppm
will read from the jtag an image (indicated in the IMAGE Variable) and use it as input for the network.
make clean all run PERFORMANCE=0
will print on standard output the number of cicles and operations of each layer in the network, make clean all run MAP=0 will print on standard output the outputs of the network and then terminate the exeution.
There are three different quantized models that can be used in the tflite directory. Each one can be used by varing the DEPTH variable which can assume values {1(default),075,05}.
lastly
The utilis/memory_test.py script launches the Autotiler engine with different L2 memory configurations and plots the trend of several metrics to explore the optimal configuration:
python3 utils/memory_test.py
To visualize the perfomences of the different layers there is also the utilis/different_memory.py script which runs an inference with different L2 memory configurations and saves a series of plots with the number of operations for each layer , the operation/cycle performance and the execution time for memory configuration:
python3 utils/different_memory
the original model with the quantization script are included in the tflite_converter folder the script requires as parameters the full path to the frozen graph file and a name for the output. The script can apply full int8 quantization (default) or only quantization of the weights (with --no-full_8_bit). In the full int8 it requires also the path to a folder containing some images to calibrate the outputs.
python3 tensorflow_converter.py -m=./tflite_graph.pb -o=./fully_quantized -i=../images/
The tflite_test folder contains two Python scripts (test_tflite.py and comparison.py) the fisrt one can be used to test the tflite models:
python3 test_tflite.py -i=path/to/image
The compare script draws on the image the bounding boxes produced by the model with int8 activations(yellow) and float32 activations(black):
python3 comparison.py -m=./non_quantized.tflite -mq=../tflite_model/SSD_tin_can_bottle.tflite -i=path/to/image
if not specified the test image will be test_1_output.ppm located in the tflite_model folder.