F1_Score

gistfile1.txt

import numpy as np
import pandas as pd
import bottleneck as bn

# Read And Count
trainDF = pd.read_csv('train.csv',sep=',',header=0)
lenTrain = len(trainDF)
valDF = pd.read_csv('submission.csv',sep=',',header=0)
lenVal = len(valDF)
trainFold = pd.read_csv('train_5fold_20181219.csv',sep=',',header=0)

# Fake Result, Put True Result in trainPred and valPred
trainPred = pd.DataFrame(np.random.uniform(size=(lenTrain,29)))
trainOut = pd.DataFrame(np.zeros_like(trainPred.values))
valPred = pd.DataFrame(np.random.uniform(size=(lenVal,29)))
valOut = pd.DataFrame(np.zeros_like(valPred.values))
trainPred[0] = trainDF.iloc[:,0]
trainOut[0] = trainDF.iloc[:,0]
valPred[0] = valDF.iloc[:,0]
valOut[0] = valDF.iloc[:,0]

# F1 Score Thresholder and Estimator

def Schumacher(s):
    sv = s.values
    sortindex = np.argsort(-sv)
    R = sum(sv)
    runningA = 0
    lastF = 0
    for i in range(len(s)):
        P = i + 1
        runningA = runningA + sv[sortindex[i]]
        runningF = runningA / (P + R)
        if runningF > lastF:
            lastF = runningF
        else:
            break
    K = i
    givePred = [0 for i in range(len(sv))]
    for idx in sortindex[:K]:
        givePred[idx] = 1
    return (np.array(givePred), lastF)
    
# Giving Train Estimator, For Specific FoldID
FoldID = 4
sumtotalF1Estimate = 0.0
sumtotalF1 = 0.0
for i in range(28):
    seqPred = trainPred[trainFold['inpool']==FoldID][i+1]
    seqGT = trainFold[trainFold['inpool']==FoldID]['Target'].map(lambda x: int(str(i) in x.split(' ')))
    # print seqPred, seqGT
    outDetail, F = Schumacher(seqPred)
    for idx,line_index in enumerate(trainFold[trainFold['inpool']==FoldID].index):
        trainOut.ix[line_index, i+1] = outDetail[idx]
    sumtotalF1Estimate  += F
    sumtotalF1 += (outDetail * seqGT.values).sum() * 1.0 / ( outDetail.sum()  + seqGT.sum() )
    print sumtotalF1Estimate,sumtotalF1
    
# Giving Val Estimator
sumtotalF1Estimate = 0.0
for i in range(28):
    seqPred = valPred[i+1]
    # print seqPred, seqGT
    outDetail, F = Schumacher(seqPred)
    valOut[i+1] = outDetail
    sumtotalF1Estimate  += F
    print sumtotalF1Estimate
    
# valOut is for Submission


################ReWeight###########
for i in range(28):
    trainDF[str(i)] = trainDF.Target.map(lambda x: int(str(i) in x.split(' ')))
value_counts = trainDF.sum(axis=0).iloc[2:]

# For class who had least total weight, adjust its weight to resample_bottomline * largest total weight
resample_bottomline = 0.25
value_max, value_min = value_counts.max(), value_counts.min()

raw_weight = value_counts.map(lambda x: (x-value_min)*(1 - resample_bottomline) + value_max * resample_bottomline) / value_counts

## Use adj_weight
adj_weight = raw_weight/raw_weight.mean()