from __future__ import division

import pandas as pd

import os

os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" # see issue #152

os.environ["CUDA_VISIBLE_DEVICES"] = ""

from keras.layers import Input, Embedding, LSTM, Dense

from keras.models import Model

from keras.layers import multiply

from sklearn.metrics import roc_auc_score

from keras.utils import to_categorical

import numpy as np

from tqdm import tqdm

from keras import backend as K

import math

def norm(x, mean, sd):

return num/denom

def categorical_crossentropy(y_true, y_pred):

return K.categorical_crossentropy(y_true, y_pred,from_logits=True)

train = pd.read_csv("train.csv");

x = {i:[] for i in range(40)}

new_pr_buy_prob = np.zeros(2000)

for i in tqdm(range(2000)):

count = 0

new_count = 0

a_count = -20*np.ones(40)

for t in range(49):

a = np.zeros(40)

tr = train[(train["i"]==i)&(train["t"]==t)]

if len(tr)>0:

count += 1

bought_till = np.where(a_count>0)[0]

update_history = [w for w in bought_till if w not in list(tr["j"])]

new_bought = [w for w in list(tr["j"]) if w not in bought_till]

if len(new_bought)>0:

new_count += 1

a[np.array(list(tr["j"]))] = 1

a_count[np.array(list(tr["j"]))] = 1

if len(update_history)>0:

a_count[np.array(update_history)] += 1

else:

a_count[np.where(a_count)>0] += 1

new_pr_buy_prob[i] = new_count/count

for i in tqdm(range(2000)):

for j in range(40):

tr = train[(train["i"]==i)&(train["j"]==j)]

if len(tr)>1:

data = list(tr["t"])

for k in range(len(data)-1):

x[j].append(data[k+1]-data[k])

newly_bought = np.zeros((4,40))

for i in range(2000):

bought_before = []

for j in range(49):

tr = train[(train["i"]==i)&(train["j"]==j)]

if j>5:

new_items = [w for w in list(tr["j"]) if w not in bought_before]

for w in new_items:

newly_bought[j%4,w] += 1

bought_before = bought_before + list(tr["j"])

for i in range(4):

newly_bought[i] /= np.sum(newly_bought[i])

product_proba = np.zeros((49,40))

prob_till = np.zeros((4,40))

for i in range(49):

tr = train[train["t"]==i]

out = dict(tr["j"].value_counts())

out = [0 if n not in out.keys() else out[n] for n in range(40)]

prob_till[i%4] = prob_till[i%4] + np.array(out)

product_proba[i] = prob_till[i%4]/np.sum(prob_till[i%4])

avg_gap = np.array([-20 if len(x[i])==0 else np.mean(x[i]) for i in range(40)])

std_gap = np.array([2 if len(x[i])==0 else np.std(x[i]) for i in range(40)])

pr_def_prices = np.zeros(40)

pr_ch_price = np.zeros((49,40))

for i in range(40):

tr = train[(train["j"]==i)&(train["advertised"]==0)]

value = tr["price"].unique()

pr_def_prices[i] = value[0]

for t in range(49):

p = np.zeros(40)

for pr in list(train[(train["t"]==i)]["j"].unique()):

p[pr] = train[(train["t"]==i)&(train["j"]==pr)]["price"].unique()[0]

p[np.where(p)==0] = pr_def_prices[np.where(p)==0]

pr_ch_price[t] = p

pr_buying_history = np.zeros((2000,49,40))

pr_weekly_data = np.zeros((2000,49,40))

pr_price_diff = np.zeros((2000,49,40))

for i in tqdm(range(2000)):

a_count = -20*np.ones(40)

for t in range(49):

a = np.zeros(40)

tr = train[(train["i"]==i)&(train["t"]==t)]

if len(tr)>0:

bought_till = np.where(a_count>0)[0]

update_history = [w for w in bought_till if w not in list(tr["j"])]

a[np.array(list(tr["j"]))] = 1

a_count[np.array(list(tr["j"]))] = 1

if len(update_history)>0:

a_count[np.array(update_history)] += 1

else:

a_count[np.where(a_count)>0] += 1

non_bought = len(np.where(a_count<0)[0])

p = np.array([new_pr_buy_prob[i]*(newly_bought[t%4,n]) if a_count[n]<0 else (1-new_pr_buy_prob[i])*norm(a_count[n],avg_gap[n],std_gap[n]) for n in range(40)])

pr_buying_history[i,t,:] = p#np.multiply(p,product_proba[t])

pr_weekly_data[i,t,:] = a

pr_price_diff[i,t,:] = 1 - np.divide(pr_ch_price[t],pr_def_prices)

x_train,y_train = [],[]

x_test,y_test = [],[]

n_prev = 5

for i in tqdm(range(2000)):

for j in range(48):

#if j>20:

if j<40 :

data = np.concatenate((pr_weekly_data[i,j,:],pr_price_diff[i,j+1,:],pr_buying_history[i,j,:],product_proba[j]))

x_train.append(data)

y_train.append(pr_weekly_data[i,j+1,:])

else:

data = np.concatenate((pr_weekly_data[i,j,:],pr_price_diff[i,j+1,:],pr_buying_history[i,j,:],product_proba[j]))

x_test.append(data)

y_test.append(pr_weekly_data[i,j+1,:])

x_train,y_train = np.array(x_train),np.array(y_train)

x_test,y_test = np.array(x_test),np.array(y_test)

input_1 = Input(shape=(x_train.shape[1],))

out_1 = Dense(32,activation="relu")(input_1)

out_1 = Dense(32,activation="relu")(out_1)

out = Dense(32,activation="relu")(out_1)

out = Dense(40,activation="sigmoid")(out)

model = Model(input_1,out)

model.compile(loss="binary_crossentropy", optimizer='adam',metrics=["accuracy"])

for i in range(50):

model.fit(x_train,y_train,epochs=1,batch_size=32,validation_data=(x_test,y_test))

predicted = model.predict(x_test)

print(roc_auc_score(y_test.reshape((-1,1)),predicted.reshape((-1,1))))