def error(self, input_x, labels):
"""
error value
"""
labels = one_hot(labels, self.n_classes)
predicts = self.predict(input_x)
error = predicts != labels
return np.mean(error)
def loss(self, input_x, labels):
"""
loss function
"""
labels = one_hot(labels, self.n_classes)
for layer in self.layers:
input_x = layer.forward(input_x)
return self.cost_fun.loss(labels, input_x)
def gradient_descent(self, input_x, input_y):
"""
grandient descent
"""
input_a = input_x
for layer in self.layers:
input_a = layer.forward(input_a)
input_grad = self.cost_fun.grad(one_hot(input_y, self.n_classes),
input_a)
for layer in reversed(self.layers):
input_grad = layer.backward(input_grad)
def fixInputs(df):
# turn smoking marker into 2 vars (assume missing are former-smokers)
# never_smoke captures never-smokers (1) vs ever-smokers (2 & 3)
# still_smoke captures current-smokers (3) vs not-current-smokers (1 & 2)
df['never_smoke'] = df['smoke_3cat'].values == 1
df['still_smoke'] = df['smoke_3cat'].values == 3
##################################
### blood pressure calculations
# mean artierial pressure (MAP)
df['map'] = (df['sbp'] + 2 * df['dbp']) / 3
# pulse pressure (PP)
df['pp'] = df['sbp'] - df['dbp']
#df['bp_ratio'] = df['dbp']/df['sbp']
# clinical cvd implies subclinical cvd
df.loc[df['sub_clinicalcvd'] == 1, 'sub_subclinicalcvd'] = 1
# scale count variable
#df['n_agents'] = df['n_agents']/np.max(df['n_agents'].values)
# artifact of way data was recorded: if value was calculated to be less than 2, it was marked nan due to inaccuracy
df.loc[np.logical_and(df['umalcr'].values ==
np.nan, df['screat'].values != np.nan),
'umalcr'] = 1.5
###############################################################################################
### lipid calculations
### chr = cholesterol, trr = triglycerides, ldl/hdl = low/high density lipoproteins
### already checked that trr, chr, hdl are nans at the same time (implying same panel)
df['non_hdl'] = df['chr'] - df['hdl']
df['ldl'] = df['non_hdl'] - 0.2 * df['trr']
df['trr_hdl'] = df['trr'] / df['hdl']
df['chr_hdl'] = df['chr'] / df['hdl']
## assume all others are hispanics?
#df.loc[df['race4'].values == 'OTHER', 'race4'] = 'HISPANIC'
df = one_hot(df, 'race4')
return df.drop(columns=[
'newsiteid', 'race4', 'sbptertile', 'smoke_3cat', 'race_black',
'sub_senior', 'inclusionfrs', 'sub_cvd'
] + ['noagents', 'sub_ckd'])
def play_from(
board, positions, moves, values, lock
): #Possible improvement could be placing our data into tuples?? post-processing required tho
import chess
import chess.uci
import numpy as np
from helper import conversion, one_hot
stockfish = chess.uci.popen_engine("stockfish")
info_handler = chess.uci.InfoHandler()
stockfish.info_handlers.append(info_handler)
stockfish.uci()
while not board.is_game_over():
stockfish.position(board)
stock_move = stockfish.go(
movetime=250) #Set stockfish to 5 moves per second
#Generate position
turn = board.turn
k_castle = board.has_kingside_castling_rights(turn)
q_castle = board.has_queenside_castling_rights(turn)
kb_castle = board.has_kingside_castling_rights(not turn)
qb_castle = board.has_queenside_castling_rights(not turn)
check = board.is_check()
position = conversion(str(board), turn, k_castle, q_castle, check,
kb_castle,
qb_castle) #Append position to training data
mate = info_handler.info["score"][1].mate
#Acquire lock so that all data is appended to our lists in the correct order!
lock.acquire()
positions.append(position)
moves.append(one_hot(
stock_move[0])) #Add stockfish's choice to training data
if mate is None:
value = np.tanh(info_handler.info["score"][1].cp / 300)
values.append(value)
else:
values.append(mate / np.abs(mate))
lock.release()
board.push(stock_move[0])