androidMotionLogger/julia/net.jl

"""
Author: Sebastian Vendt, University of Ulm



"""

using ArgParse
s = ArgParseSettings()
@add_arg_table s begin
    "--gpu"
        help = "set, if you want to train on the GPU"
		action = :store_true
	"--eval"
		help = "set, if you want to validate instead of test after training"
		action = :store_true
    "--epochs" 
		help = "Number of epochs"
		arg_type = Int64
		default = 40
	"--logmsg"
		help = "additional message describing the training log"
		arg_type = String
		default = ""
	"--csv"
		help = "set, if you additionally want a csv output of the learning process"
		action = :store_true
	"--runD"
		help = "set, if you want to run the default config"
		action = :store_true
end
parsed_args = parse_args(ARGS, s)

using Flux, Statistics
using Flux: onecold
using BSON
using Dates
using Printf
using NNlib
using FeedbackNets
include("./dataManager.jl")
include("./verbose.jl")
using .dataManager: make_batch
using .verbose
using Logging
using Random
import LinearAlgebra: norm
norm(x::TrackedArray{T}) where T = sqrt(sum(abs2.(x)) + eps(T)) 

######################
# PARAMETERS
######################
const batch_size = 100
momentum = 0.9f0
const lambda = 0.0005f0
const delta = 0.00001
learning_rate = 0.1f0
validate = parsed_args["eval"]
const epochs = parsed_args["epochs"]
const decay_rate = 0.1f0
const decay_step = 40
const usegpu = parsed_args["gpu"]
const printout_interval = 2
const time_format = "HH:MM:SS"
const date_format = "dd_mm_yyyy"
data_size = (60, 6) # resulting in a 300ms frame

# DEFAULT ARCHITECTURE
channels = 1
features = [32, 64, 128] # needs to find the relation between the axis which represents the screen position 
kernel = [(3,1), (3,1), (3,6)]  # convolute only horizontally, last should convolute all 6 rows together to map relations between the channels  
pooldims = [(2,1), (2,1)]# (30,6) -> (15,6)
# formula for calculating output dimensions of convolution: 
# dim1 = ((dim1 - Filtersize + 2 * padding) / stride) + 1
inputDense = [1664, 600, 300] # prod((data_size .÷ pooldims[1] .÷ pooldims[2]) .- kernel[3] .+ 1) * features[3]
dropout_rate = 0.3f0

# random search values
rs_momentum = [0.9, 0.92, 0.94, 0.96, 0.98, 0.99]
rs_features = [[32, 64, 128], [64, 64, 64], [32, 32, 32], [96, 192, 192]]
rs_dropout_rate = [0.1, 0.3, 0.4, 0.6, 0.8]
rs_kernel = [[(3,1), (3,1), (3,6)], [(5,1), (5,1), (3,6)], [(7,1), (7,1), (3,6)], [(3,1), (3,1), (2,6)], [(5,1), (5,1), (2,6)], [(7,1), (7,1), (2,6)], 
				[(7,1), (5,1), (2,6)], [(7,1), (5,1), (3,6)], [(5,1), (3,1), (2,6)], [(5,1), (3,1), (3,6)]]
rs_pooldims = [[(2,1), (2,1)], [(3,1), (3,1)]]
rs_learning_rate = [1, 0.3, 0.1, 0.03, 0.01, 0.003, 0.001]

dataset_folderpath = "../MATLAB/TrainingData/"
dataset_name = "2019_09_09_1658"

const model_save_location = "../trainedModels/"
const log_save_location = "./logs/"

if usegpu
    using CuArrays
end

debug_str = ""
log_msg = parsed_args["logmsg"]
csv_out = parsed_args["csv"]
runD = parsed_args["runD"]
io = nothing
io_csv = nothing
@debug begin
	global debug_str
	debug_str = "DEBUG_"
	"------DEBUGGING ACTIVATED------"
end

function adapt_learnrate(epoch_idx)
    return learning_rate * decay_rate^(epoch_idx / decay_step)
end

# TODO different idea for the accuracy: draw circle around ground truth and if prediction lays within the circle count this as a hit 

function accuracy(model, x, y)
	y_hat = Tracker.data(model(x))
	return mean(mapslices(button_number, y_hat, dims=1) .== mapslices(button_number, y, dims=1))
end

function accuracy(model, dataset)
   acc = 0.0f0
   for (data, labels) in dataset
      acc += accuracy(model, data, labels)
   end
   return acc / length(dataset)
end

function button_number(X)
	return (X[1] * 1080) ÷ 360 + 3 * ((X[2] * 980) ÷ 245)
end

function loss(model, x, y) 
	# quadratic euclidean distance + parameternorm
	return Flux.mse(model(x), y) + lambda * sum(norm, params(model))
end

function loss(model, dataset)
	loss_val = 0.0f0
	for (data, labels) in dataset
		loss_val += Tracker.data(loss(model, data, labels))
	end
	return loss_val / length(dataset)
end

function load_dataset()
	train = make_batch(dataset_folderpath, "$(dataset_name)_TRAIN.mat", normalize_data=false, truncate_data=false)
	val = make_batch(dataset_folderpath, "$(dataset_name)_VAL.mat", normalize_data=false, truncate_data=false)
	test = make_batch(dataset_folderpath, "$(dataset_name)_TEST.mat", normalize_data=false, truncate_data=false)
	return (train, val, test)
end

function create_model()
	return Chain(
		Conv(kernel[1], channels=>features[1], relu, pad=map(x -> x ÷ 2, kernel[1])),
		MaxPool(pooldims[1], stride=pooldims[1]), 
		Conv(kernel[2], features[1]=>features[2], relu, pad=map(x -> x ÷ 2, kernel[2])),
		MaxPool(pooldims[2], stride=pooldims[2]),
		Conv(kernel[3], features[2]=>features[3], relu),
		# MaxPool(),
		flatten, 
		Dense(prod((data_size .÷ pooldims[1] .÷ pooldims[2]) .- kernel[3] .+ 1) * features[3], inputDense[2], relu),
		Dropout(dropout_rate),
		Dense(inputDense[2], inputDense[3], relu),
		Dropout(dropout_rate),
		Dense(inputDense[3], 2, σ), # coordinates between 0 and 1
	)
end

function log(model, epoch, use_testset)
	Flux.testmode!(model, true)
	
	if(epoch == 0) # evalutation phase 
		if(use_testset) @printf(io, "[%s] INIT Loss(test): f% Accuarcy: %f\n", Dates.format(now(), time_format), loss(model, test_set), accuracy(model, test_set)) 
		else @printf(io, "[%s] INIT Loss(val): %f Accuarcy: %f\n", Dates.format(now(), time_format), loss(model, validation_set), accuracy(model, validation_set)) end
	elseif(epoch == epochs)
        @printf(io, "[%s] Epoch %3d: Loss(train): %f Loss(val): %f\n", Dates.format(now(), time_format), epoch, loss(model, train_set), loss(model, validation_set))
		if(use_testset) 
		   @printf(io, "[%s] FINAL(%d) Loss(test): %f Accuarcy: %f\n", Dates.format(now(), time_format), epoch, loss(model, test_set), accuracy(model, test_set)) 
		else 
		   @printf(io, "[%s] FINAL(%d) Loss(val): %f Accuarcy: %f\n", Dates.format(now(), time_format), epoch, loss(model, validation_set), accuracy(model, validation_set)) 
	   end
	else # learning phase
		if (rem(epoch, printout_interval) == 0) 
			@printf(io, "[%s] Epoch %3d: Loss(train): %f Loss(val): %f\n", Dates.format(now(), time_format), epoch, loss(model, train_set), loss(model, validation_set)) 
		end
	end

	Flux.testmode!(model, false)
end

function log_csv(model, epoch)
	Flux.testmode!(model, true)
	if(csv_out) @printf(io_csv, "%d, %f, %f\n", epoch, loss(model, train_set), loss(model, validation_set)) end
	Flux.testmode!(model, false)
end

function eval_model(model)
	Flux.testmode!(model, true)
	if (validate) return (loss(model, validation_set), accuracy(model, validation_set))
	else return (loss(model, test_set), accuracy(model, test_set)) end
end

function train_model()
	model = create_model()
	if (usegpu) model = gpu(model) end
	opt = Momentum(learning_rate, momentum)
	log(model, 0, !validate)
	Flux.testmode!(model, false) # bring model in training mode
	last_loss = loss(model, train_set)
    for i in 1:epochs
		flush(io)
        Flux.train!((x, y) -> loss(model, x, y), params(model), train_set, opt)
        opt.eta = adapt_learnrate(i)
		log_csv(model, i)
		log(model, i, !validate)
		
		# early stopping
		curr_loss = loss(model, train_set)
		if(abs(last_loss - curr_loss) < delta)
			@printf(io, "Early stopping with Loss(train) %f at epoch %d (Accuracy: %f)\n", curr_loss, i, accuracy(model, validation_set))
			return eval_model(model)
		end
		last_loss = curr_loss
    end
    return eval_model(model)
end

function random_search()
	rng = MersenneTwister()
	results = []
	for search in 1:800
		# create random set
		global momentum = rand(rng, rs_momentum)
		global features = rand(rng, rs_features)
		global dropout_rate = rand(rng, rs_dropout_rate)
		global kernel = rand(rng, rs_kernel)
		global pooldims = rand(rng, rs_pooldims)
		global learning_rate = rand(rng, rs_learning_rate)
		
		# printf configuration
		config1 = "momentum$(momentum), features=$(features), dropout_rate=$(dropout_rate)"
		config2 = "kernel=$(kernel), pooldims=$(pooldims), learning_rate=$(learning_rate)"
		@printf(io, "\nSearch %d of %d\n", search, 500)
		@printf(io, "%s\n", config1)
		@printf(io, "%s\n\n", config2)
		
		(loss, accuracy) = train_model()
		push!(results, (search, loss, accuracy))
	end
	return results
end


# logging framework 
fp = "$(log_save_location)$(debug_str)log_$(Dates.format(now(), date_format)).log"
io = open(fp, "a+")
global_logger(SimpleLogger(io)) # for debug outputs
@printf(Base.stdout, "Logging to File: %s\n", fp)
@printf(io, "\n--------[%s %s]--------\n", Dates.format(now(), date_format), Dates.format(now(), time_format))
@printf(io, "%s\n", log_msg)

# csv handling
if (csv_out)
	fp_csv = "$(log_save_location)$(debug_str)csv_$(Dates.format(now(), date_format)).csv"
	io_csv = open(fp_csv, "w+") # read, write, create, truncate
	@printf(io_csv, "epoch, loss(train), loss(val)\n")
end	

# dump configuration 
@debug begin
	for symbol in names(Main)
		var = "$(symbol) = $(eval(symbol))"
		@printf(io, "%s\n", var)
	end
	"--------End of VAR DUMP--------"
end
flush(io)
flush(Base.stdout)

train, validation, test = load_dataset()

if (usegpu)
	const train_set = gpu.(train)
	const validation_set = gpu.(validation)
	const test_set = gpu.(test)
end

if(!runD)
	results = random_search()
	BSON.@save "results.bson" results
	#TODO sort and print best 5-10 results 
	sort!(results, by = x -> x[2])
	# print results 
	@printf("Best results by Loss:\n")
	for idx in 1:5 
		@printf("#%d: Loss %f, accuracy %f in Search: %d\n", idx, results[idx][2], results[idx][3], results[idx][1])
	end
	
	sort!(results, by = x -> x[3], rev=true)
	@printf("Best results by Accuarcy:\n")
	for idx in 1:5 
		@printf("#%d: Accuarcy: %f, Loss %f in Search: %d\n", idx, results[idx][3], results[idx][2], results[idx][1])
	end
else 
	train_model() 
end