ml_fpga

Important files:

Note all files are testbenched

Neural Net Files:

mnist_nn_small.sv

multiplier.sv

neuron.sv

relu.sv

neural_network.sv

iris_nn.sv

Serial and Top Files:

top_digits_1.sv

serial_version_2.sv (includes all 5 modules)

Early versions of serial:

uart_read.sv

uart_write.sv

gather_weights.sv

Python files:

For neural net:

(insert here)

For serial

send_data.py, arduino file

Instructions

Send data using send_data.py (weights, biases, x)

Watch the final_output update in the ILA (or on seven_seg)!

Old Version Instructions

#How to run the whole thing

1. upload write_to_python_nov_11.ino to the arduino.
2. generate bistream using top level: top_level_read_in.sv
3. upload bitfile to fpga
4. connect gnd on fpga to gnd on arduino; connect jb[0] to pin 2 on arduino.
5. press btnc (tells fpga to get ready to receive weights. kind of hacky right now.)
6. press btnd (wired to the nn's reset right now)
7. run wso_example_weights.py. watch the seven seg update. the digits on the display are [bias 6, weight 14, x_in, final_output]. (also the digits only go up to f, so bias 6 shows up as 0 because it is 16.)

#Other notes

1. wso.py has hardcoded weights right now. I ran them by doing 8 times the actual weight (bc we have 3 decimals). But I'm pretty sure the two's complement is done wrong for some of the negatives.
2. x_in is hardcoded in the top level

project_1/project_1.srcs/sources_1/ has sources project_1/project_1.srcs/sim_1/ has testbenches