Web: https://www.meetup.com/Tel-Aviv-Deep-Learning-Bootcamp/events/241762893/
Notebooks: On GitHub
Shlomo Kashani
Numerai provides payments based on the number of correctly predictted labels (LOGG_LOSS) in a data-set which changes every week.
Two data-sets are provided: numerai_training_data.csv and numerai_tournament_data.csv
This tutorial was written in order to demonstrate a fully working example of a PyTorch NN on a real world use case, namely a Binary Classification problem on the NumerAI data set. If you are interested in the sk-learn version of this problem please refer to: https://github.com/QuantScientist/deep-ml-meetups/tree/master/hacking-kaggle/python/numer-ai
For the scientific foundation behind Binary Classification and Logistic Regression, refer to: https://github.com/QuantScientist/Deep-Learning-Boot-Camp/tree/master/Data-Science-Interviews-Book
Every step, from reading the CSV into numpy arrays, converting to GPU based tensors, training and validation, are meant to aid newcomers in their first steps in PyTorch.
Additionally, commonly used Kaggle metrics such as ROC_AUC and LOG_LOSS are logged and plotted both for the training set as well as for the validation set.
Thus, the NN architecture is naive and by no means optimized. Hopefully, I will improve it over time and I am working on a second CNN based version of the same problem.
In [1]:
# !pip install http://download.pytorch.org/whl/cu75/torch-0.2.0.post1-cp27-cp27mu-manylinux1_x86_64.whl
# !pip install --upgrade torch
# !pip install torchvision
# ! pip install cv2
# import cv2
# !pip install pycuda
%reset -f
# %%timeit
import torch
from torch.autograd import Variable
import numpy as np
import pandas
import numpy as np
import pandas as pd
from sklearn import cross_validation
from sklearn import metrics
from sklearn.metrics import roc_auc_score, log_loss, roc_auc_score, roc_curve, auc
import matplotlib.pyplot as plt
from sklearn import cross_validation
from sklearn import metrics
from sklearn.metrics import roc_auc_score, log_loss, roc_auc_score, roc_curve, auc
from sklearn.cross_validation import StratifiedKFold, ShuffleSplit, cross_val_score, train_test_split
import logging
import numpy
import numpy as np
from __future__ import print_function
from __future__ import division
import math
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
import pandas as pd
import os
import torch
from torch.utils.data.dataset import Dataset
from torch.utils.data import DataLoader
from torchvision import transforms
from torch import nn
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
from sklearn.preprocessing import MultiLabelBinarizer
import time
from sklearn.preprocessing import PolynomialFeatures
import pandas as pd
import numpy as np
import scipy
%matplotlib inline
from pylab import rcParams
rcParams['figure.figsize'] = (6, 6) # setting default size of plots
import tensorflow as tf
print("tensorflow:" + tf.__version__)
!set "KERAS_BACKEND=tensorflow"
import torch
import sys
print('__Python VERSION:', sys.version)
print('__pyTorch VERSION:', torch.__version__)
print('__CUDA VERSION')
from subprocess import call
print('__CUDNN VERSION:', torch.backends.cudnn.version())
print('__Number CUDA Devices:', torch.cuda.device_count())
print('__Devices')
print("OS: ", sys.platform)
print("Python: ", sys.version)
print("PyTorch: ", torch.__version__)
print("Numpy: ", np.__version__)
handler=logging.basicConfig(level=logging.INFO)
lgr = logging.getLogger(__name__)
%matplotlib inline
# !pip install psutil
import psutil
def cpuStats():
print(sys.version)
print(psutil.cpu_percent())
print(psutil.virtual_memory()) # physical memory usage
pid = os.getpid()
py = psutil.Process(pid)
memoryUse = py.memory_info()[0] / 2. ** 30 # memory use in GB...I think
print('memory GB:', memoryUse)
cpuStats()
In [2]:
# %%timeit
use_cuda = torch.cuda.is_available()
# use_cuda = False
FloatTensor = torch.cuda.FloatTensor if use_cuda else torch.FloatTensor
LongTensor = torch.cuda.LongTensor if use_cuda else torch.LongTensor
Tensor = FloatTensor
lgr.info("USE CUDA=" + str (use_cuda))
# ! watch -n 0.1 'ps f -o user,pgrp,pid,pcpu,pmem,start,time,command -p `lsof -n -w -t /dev/nvidia*`'
# sudo apt-get install dstat #install dstat
# sudo pip install nvidia-ml-py #install Python NVIDIA Management Library
# wget https://raw.githubusercontent.com/datumbox/dstat/master/plugins/dstat_nvidia_gpu.py
# sudo mv dstat_nvidia_gpu.py /usr/share/dstat/ #move file to the plugins directory of dstat
In [3]:
# Data params
TARGET_VAR= 'target'
TOURNAMENT_DATA_CSV = 'numerai_tournament_data.csv'
TRAINING_DATA_CSV = 'numerai_training_data.csv'
BASE_FOLDER = 'numerai/'
# fix seed
seed=17*19
np.random.seed(seed)
torch.manual_seed(seed)
if use_cuda:
torch.cuda.manual_seed(seed)
In [4]:
# %%timeit
df_train = pd.read_csv(BASE_FOLDER + TRAINING_DATA_CSV)
df_train.head(5)
Out[4]:
In [5]:
from sklearn.preprocessing import LabelEncoder
from sklearn.pipeline import Pipeline
from collections import defaultdict
# def genBasicFeatures(inDF):
# print('Generating basic features ...')
# df_copy=inDF.copy(deep=True)
# magicNumber=21
# feature_cols = list(inDF.columns)
# inDF['x_mean'] = np.mean(df_copy.ix[:, 0:magicNumber], axis=1)
# inDF['x_median'] = np.median(df_copy.ix[:, 0:magicNumber], axis=1)
# inDF['x_std'] = np.std(df_copy.ix[:, 0:magicNumber], axis=1)
# inDF['x_skew'] = scipy.stats.skew(df_copy.ix[:, 0:magicNumber], axis=1)
# inDF['x_kurt'] = scipy.stats.kurtosis(df_copy.ix[:, 0:magicNumber], axis=1)
# inDF['x_var'] = np.var(df_copy.ix[:, 0:magicNumber], axis=1)
# inDF['x_max'] = np.max(df_copy.ix[:, 0:magicNumber], axis=1)
# inDF['x_min'] = np.min(df_copy.ix[:, 0:magicNumber], axis=1)
# return inDF
def addPolyFeatures(inDF, deg=2):
print('Generating poly features ...')
df_copy=inDF.copy(deep=True)
poly=PolynomialFeatures(degree=deg)
p_testX = poly.fit(df_copy)
# AttributeError: 'PolynomialFeatures' object has no attribute 'get_feature_names'
target_feature_names = ['x'.join(['{}^{}'.format(pair[0],pair[1]) for pair in tuple if pair[1]!=0]) for tuple in [zip(df_copy.columns,p) for p in poly.powers_]]
df_copy = pd.DataFrame(p_testX.transform(df_copy),columns=target_feature_names)
return df_copy
def oneHOT(inDF):
d = defaultdict(LabelEncoder)
X_df=inDF.copy(deep=True)
# Encoding the variable
X_df = X_df.apply(lambda x: d['era'].fit_transform(x))
return X_df
In [6]:
from sklearn import preprocessing
# Train, Validation, Test Split
def loadDataSplit():
df_train = pd.read_csv(BASE_FOLDER + TRAINING_DATA_CSV)
# TOURNAMENT_DATA_CSV has both validation and test data provided by NumerAI
df_test_valid = pd.read_csv(BASE_FOLDER + TOURNAMENT_DATA_CSV)
answers_1_SINGLE = df_train[TARGET_VAR]
df_train.drop(TARGET_VAR, axis=1,inplace=True)
df_train.drop('id', axis=1,inplace=True)
df_train.drop('era', axis=1,inplace=True)
df_train.drop('data_type', axis=1,inplace=True)
# df_train=oneHOT(df_train)
df_train.to_csv(BASE_FOLDER + TRAINING_DATA_CSV + 'clean.csv', header=False, index = False)
df_train= pd.read_csv(BASE_FOLDER + TRAINING_DATA_CSV + 'clean.csv', header=None, dtype=np.float32)
df_train = pd.concat([df_train, answers_1_SINGLE], axis=1)
feature_cols = list(df_train.columns[:-1])
# print (feature_cols)
target_col = df_train.columns[-1]
trainX, trainY = df_train[feature_cols], df_train[target_col]
# TOURNAMENT_DATA_CSV has both validation and test data provided by NumerAI
# Validation set
df_validation_set=df_test_valid.loc[df_test_valid['data_type'] == 'validation']
df_validation_set=df_validation_set.copy(deep=True)
answers_1_SINGLE_validation = df_validation_set[TARGET_VAR]
df_validation_set.drop(TARGET_VAR, axis=1,inplace=True)
df_validation_set.drop('id', axis=1,inplace=True)
df_validation_set.drop('era', axis=1,inplace=True)
df_validation_set.drop('data_type', axis=1,inplace=True)
# df_validation_set=oneHOT(df_validation_set)
df_validation_set.to_csv(BASE_FOLDER + TRAINING_DATA_CSV + '-validation-clean.csv', header=False, index = False)
df_validation_set= pd.read_csv(BASE_FOLDER + TRAINING_DATA_CSV + '-validation-clean.csv', header=None, dtype=np.float32)
df_validation_set = pd.concat([df_validation_set, answers_1_SINGLE_validation], axis=1)
feature_cols = list(df_validation_set.columns[:-1])
target_col = df_validation_set.columns[-1]
valX, valY = df_validation_set[feature_cols], df_validation_set[target_col]
# Test set for submission (not labeled)
df_test_set = pd.read_csv(BASE_FOLDER + TOURNAMENT_DATA_CSV)
# df_test_set=df_test_set.loc[df_test_valid['data_type'] == 'live']
df_test_set=df_test_set.copy(deep=True)
df_test_set.drop(TARGET_VAR, axis=1,inplace=True)
tid_1_SINGLE = df_test_set['id']
df_test_set.drop('id', axis=1,inplace=True)
df_test_set.drop('era', axis=1,inplace=True)
df_test_set.drop('data_type', axis=1,inplace=True)
# df_test_set=oneHOT(df_validation_set)
feature_cols = list(df_test_set.columns) # must be run here, we dont want the ID
# print (feature_cols)
df_test_set = pd.concat([tid_1_SINGLE, df_test_set], axis=1)
testX = df_test_set[feature_cols].values
return trainX, trainY, valX, valY, testX, df_test_set
In [7]:
# %%timeit
trainX, trainY, valX, valY, testX, df_test_set = loadDataSplit()
min_max_scaler = preprocessing.MinMaxScaler()
# # Number of features for the input layer
N_FEATURES=trainX.shape[1]
print (trainX.shape)
print (trainY.shape)
print (valX.shape)
print (valY.shape)
print (testX.shape)
print (df_test_set.shape)
# print (trainX)
In [8]:
# seperate out the Categorical and Numerical features
import seaborn as sns
numerical_feature=trainX.dtypes[trainX.dtypes!= 'object'].index
categorical_feature=trainX.dtypes[trainX.dtypes== 'object'].index
print ("There are {} numeric and {} categorical columns in train data".format(numerical_feature.shape[0],categorical_feature.shape[0]))
corr=trainX[numerical_feature].corr()
sns.heatmap(corr)
Out[8]:
In [9]:
from pandas import *
import numpy as np
from scipy.stats.stats import pearsonr
import itertools
# from https://stackoverflow.com/questions/17778394/list-highest-correlation-pairs-from-a-large-correlation-matrix-in-pandas
def get_redundant_pairs(df):
'''Get diagonal and lower triangular pairs of correlation matrix'''
pairs_to_drop = set()
cols = df.columns
for i in range(0, df.shape[1]):
for j in range(0, i+1):
pairs_to_drop.add((cols[i], cols[j]))
return pairs_to_drop
def get_top_abs_correlations(df, n=5):
au_corr = df.corr().abs().unstack()
labels_to_drop = get_redundant_pairs(df)
au_corr = au_corr.drop(labels=labels_to_drop).sort_values(ascending=False)
return au_corr[0:n]
print("Top Absolute Correlations")
print(get_top_abs_correlations(trainX, 5))
In [10]:
# Convert the np arrays into the correct dimention and type
# Note that BCEloss requires Float in X as well as in y
def XnumpyToTensor(x_data_np):
x_data_np = np.array(x_data_np, dtype=np.float32)
print(x_data_np.shape)
print(type(x_data_np))
if use_cuda:
lgr.info ("Using the GPU")
X_tensor = Variable(torch.from_numpy(x_data_np).cuda()) # Note the conversion for pytorch
else:
lgr.info ("Using the CPU")
X_tensor = Variable(torch.from_numpy(x_data_np)) # Note the conversion for pytorch
print(type(X_tensor.data)) # should be 'torch.cuda.FloatTensor'
print(x_data_np.shape)
print(type(x_data_np))
return X_tensor
# Convert the np arrays into the correct dimention and type
# Note that BCEloss requires Float in X as well as in y
def YnumpyToTensor(y_data_np):
y_data_np=y_data_np.reshape((y_data_np.shape[0],1)) # Must be reshaped for PyTorch!
print(y_data_np.shape)
print(type(y_data_np))
if use_cuda:
lgr.info ("Using the GPU")
# Y = Variable(torch.from_numpy(y_data_np).type(torch.LongTensor).cuda())
Y_tensor = Variable(torch.from_numpy(y_data_np)).type(torch.FloatTensor).cuda() # BCEloss requires Float
else:
lgr.info ("Using the CPU")
# Y = Variable(torch.squeeze (torch.from_numpy(y_data_np).type(torch.LongTensor))) #
Y_tensor = Variable(torch.from_numpy(y_data_np)).type(torch.FloatTensor) # BCEloss requires Float
print(type(Y_tensor.data)) # should be 'torch.cuda.FloatTensor'
print(y_data_np.shape)
print(type(y_data_np))
return Y_tensor
A multilayer perceptron is a logistic regressor where instead of feeding the input to the logistic regression you insert a intermediate layer, called the hidden layer, that has a nonlinear activation function (usually tanh or sigmoid) . One can use many such hidden layers making the architecture deep.
Here we define a simple MLP structure. We map the input feature vector to a higher space, then later gradually decrease the dimension, and in the end into a 1-dimension space. Because we are calculating the probability of each genre independently, after the final layer we need to use a sigmoid layer.
There are many ways to select the initial weights to a neural network architecture. A common initialization scheme is random initialization, which sets the biases and weights of all the nodes in each hidden layer randomly.
Before starting the training process, an initial value is assigned to each variable. This is done by pure randomness, using for example a uniform or Gaussian distribution. But if we start with weights that are too small, the signal could decrease so much that it is too small to be useful. On the other side, when the parameters are initialized with high values, the signal can end up to explode while propagating through the network.
In consequence, a good initialization can have a radical effect on how fast the network will learn useful patterns.For this purpose, some best practices have been developed. One famous example used is Xavier initialization. Its formulation is based on the number of input and output neurons and uses sampling from a uniform distribution with zero mean and all biases set to zero.
In effect (according to theory) initializing the weights of the network to values that would be closer to the optimal, and therefore require less epochs to train.
nninit.xavier_uniform(tensor, gain=1) - Fills tensor with values according to the method described in "Understanding the difficulty of training deep feedforward neural networks" - Glorot, X. and Bengio, Y., using a uniform distribution.nninit.xavier_normal(tensor, gain=1) - Fills tensor with values according to the method described in "Understanding the difficulty of training deep feedforward neural networks" - Glorot, X. and Bengio, Y., using a normal distribution.nninit.kaiming_uniform(tensor, gain=1) - Fills tensor with values according to the method described in "Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification" - He, K. et al. using a uniform distribution.nninit.kaiming_normal(tensor, gain=1) - Fills tensor with values according to the method described in ["Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification" - He, K. et al.]
In [19]:
# p is the probability of being dropped in PyTorch
# NN params
DROPOUT_PROB = 0.95
LR = 0.005
MOMENTUM= 0.9
dropout = torch.nn.Dropout(p=1 - (DROPOUT_PROB))
lgr.info(dropout)
hiddenLayer1Size=256
hiddenLayer2Size=int(hiddenLayer1Size/2)
linear1=torch.nn.Linear(N_FEATURES, hiddenLayer1Size, bias=True)
torch.nn.init.xavier_uniform(linear1.weight)
linear2=torch.nn.Linear(hiddenLayer1Size, hiddenLayer2Size)
torch.nn.init.xavier_uniform(linear2.weight)
linear3=torch.nn.Linear(hiddenLayer2Size, 1)
torch.nn.init.xavier_uniform(linear3.weight)
linear3=torch.nn.Linear(hiddenLayer2Size, 1)
torch.nn.init.xavier_uniform(linear3.weight)
sigmoid = torch.nn.Sigmoid()
tanh=torch.nn.Tanh()
relu=torch.nn.LeakyReLU()
net = torch.nn.Sequential(linear1,relu,
linear2,relu,
linear3,sigmoid
)
# net = torch.nn.Sequential(linear1,dropout,nn.BatchNorm1d(hiddenLayer1Size),relu,
# linear2,dropout,nn.BatchNorm1d(hiddenLayer2Size),relu,
# linear3,dropout,sigmoid
# )
lgr.info(net) # net architecture
In [20]:
# optimizer = torch.optim.SGD(net.parameters(), lr=0.02)
# optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
# optimizer = optim.SGD(net.parameters(), lr=LR, momentum=MOMENTUM, weight_decay=5e-3)
#L2 regularization can easily be added to the entire model via the optimizer
optimizer = torch.optim.Adam(net.parameters(), lr=LR,weight_decay=5e-5) # L2 regularization
loss_func=torch.nn.BCELoss() # Binary cross entropy: http://pytorch.org/docs/nn.html#bceloss
# http://andersonjo.github.io/artificial-intelligence/2017/01/07/Cost-Functions/
if use_cuda:
lgr.info ("Using the GPU")
net.cuda()
loss_func.cuda()
# cudnn.benchmark = True
lgr.info (optimizer)
lgr.info (loss_func)
In [21]:
import time
start_time = time.time()
epochs=100 # change to 1500 for better results
all_losses = []
X_tensor_train= XnumpyToTensor(trainX)
Y_tensor_train= YnumpyToTensor(trainY)
print(type(X_tensor_train.data), type(Y_tensor_train.data)) # should be 'torch.cuda.FloatTensor'
# From here onwards, we must only use PyTorch Tensors
for step in range(epochs):
out = net(X_tensor_train) # input x and predict based on x
cost = loss_func(out, Y_tensor_train) # must be (1. nn output, 2. target), the target label is NOT one-hotted
optimizer.zero_grad() # clear gradients for next train
cost.backward() # backpropagation, compute gradients
optimizer.step() # apply gradients
if step % 5 == 0:
loss = cost.data[0]
all_losses.append(loss)
print(step, cost.data.cpu().numpy())
# RuntimeError: can't convert CUDA tensor to numpy (it doesn't support GPU arrays).
# Use .cpu() to move the tensor to host memory first.
prediction = (net(X_tensor_train).data).float() # probabilities
# prediction = (net(X_tensor).data > 0.5).float() # zero or one
# print ("Pred:" + str (prediction)) # Pred:Variable containing: 0 or 1
# pred_y = prediction.data.numpy().squeeze()
pred_y = prediction.cpu().numpy().squeeze()
target_y = Y_tensor_train.cpu().data.numpy()
tu = (log_loss(target_y, pred_y),roc_auc_score(target_y,pred_y ))
print ('LOG_LOSS={}, ROC_AUC={} '.format(*tu))
end_time = time.time()
print ('{} {:6.3f} seconds'.format('GPU:', end_time-start_time))
%matplotlib inline
import matplotlib.pyplot as plt
plt.plot(all_losses)
plt.show()
false_positive_rate, true_positive_rate, thresholds = roc_curve(target_y,pred_y)
roc_auc = auc(false_positive_rate, true_positive_rate)
plt.title('LOG_LOSS=' + str(log_loss(target_y, pred_y)))
plt.plot(false_positive_rate, true_positive_rate, 'b', label='AUC = %0.6f' % roc_auc)
plt.legend(loc='lower right')
plt.plot([0, 1], [0, 1], 'r--')
plt.xlim([-0.1, 1.2])
plt.ylim([-0.1, 1.2])
plt.ylabel('True Positive Rate')
plt.xlabel('False Positive Rate')
plt.show()
In [23]:
net.eval()
# Validation data
print (valX.shape)
print (valY.shape)
X_tensor_val= XnumpyToTensor(valX)
Y_tensor_val= YnumpyToTensor(valY)
print(type(X_tensor_val.data), type(Y_tensor_val.data)) # should be 'torch.cuda.FloatTensor'
predicted_val = (net(X_tensor_val).data).float() # probabilities
# predicted_val = (net(X_tensor_val).data > 0.5).float() # zero or one
pred_y = predicted_val.cpu().numpy()
target_y = Y_tensor_val.cpu().data.numpy()
print (type(pred_y))
print (type(target_y))
tu = (log_loss(target_y, pred_y),roc_auc_score(target_y,pred_y ))
print ('\n')
print ('log_loss={} roc_auc={} '.format(*tu))
false_positive_rate, true_positive_rate, thresholds = roc_curve(target_y,pred_y)
roc_auc = auc(false_positive_rate, true_positive_rate)
plt.title('LOG_LOSS=' + str(log_loss(target_y, pred_y)))
plt.plot(false_positive_rate, true_positive_rate, 'b', label='AUC = %0.6f' % roc_auc)
plt.legend(loc='lower right')
plt.plot([0, 1], [0, 1], 'r--')
plt.xlim([-0.1, 1.2])
plt.ylim([-0.1, 1.2])
plt.ylabel('True Positive Rate')
plt.xlabel('False Positive Rate')
plt.show()
# print (pred_y)
In [56]:
# testX, df_test_set
# df[df.columns.difference(['b'])]
# trainX, trainY, valX, valY, testX, df_test_set = loadDataSplit()
print (df_test_set.shape)
columns = ['id', 'probability']
df_pred=pd.DataFrame(data=np.zeros((0,len(columns))), columns=columns)
# df_pred.id.astype(int)
for index, row in df_test_set.iterrows():
rwo_no_id=row.drop('id')
# print (rwo_no_id.values)
x_data_np = np.array(rwo_no_id.values, dtype=np.float32)
if use_cuda:
X_tensor_test = Variable(torch.from_numpy(x_data_np).cuda()) # Note the conversion for pytorch
else:
X_tensor_test = Variable(torch.from_numpy(x_data_np)) # Note the conversion for pytorch
X_tensor_test=X_tensor_test.view(1, trainX.shape[1]) # does not work with 1d tensors
predicted_val = (net(X_tensor_test).data).float() # probabilities
p_test = predicted_val.cpu().numpy().item() # otherwise we get an array, we need a single float
df_pred = df_pred.append({'id':row['id'], 'probability':p_test},ignore_index=True)
# df_pred = df_pred.append({'id':row['id'].astype(int), 'probability':p_test},ignore_index=True)
df_pred.head(5)
In [ ]:
# df_pred.id=df_pred.id.astype(int)
def savePred(df_pred, loss):
# csv_path = 'pred/p_{}_{}_{}.csv'.format(loss, name, (str(time.time())))
csv_path = 'pred/pred_{}_{}.csv'.format(loss, (str(time.time())))
df_pred.to_csv(csv_path, columns=('id', 'probability'), index=None)
print (csv_path)
savePred (df_pred, log_loss(target_y, pred_y))
In [ ]:
In [ ]: