unnir

A comparison of Theano with other deep learning frameworks, highlighting a series of low-level design choices in no particular order.

Overview

Differentiation

Symbolic: Theano, CGT; Automatic: Torch, MXNet

Symbolic and automatic differentiation are often confused or used interchangeably, although their implementations are significantly different.

unnir

unnir

num_ensembles = 7
y_pred = 0.0
for i in tqdm(range(num_ensembles)):
    model = CatBoostClassifier(random_seed = i+2, gradient_iterations = i+1 ,leaf_estimation_method ='Newton', learning_rate=0.057, l2_leaf_reg = 23, depth=6, od_pval=0.0000001, iterations = 920, loss_function='Logloss')
    fit_model = model.fit(train_data)
    y_pred +=  fit_model.predict_proba(test_data)[:,1]
y_pred /= num_ensembles
gc.collect()

unnir

#!/bin/bash

# install CUDA Toolkit v8.0
# instructions from https://developer.nvidia.com/cuda-downloads (linux -> x86_64 -> Ubuntu -> 16.04 -> deb (network))
CUDA_REPO_PKG="cuda-repo-ubuntu1604_8.0.61-1_amd64.deb"
wget http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/${CUDA_REPO_PKG}
sudo dpkg -i ${CUDA_REPO_PKG}
sudo apt-get update
sudo apt-get -y install cuda

unnir

#!/usr/bin/env bash

wget https://repo.continuum.io/archive/Anaconda3-5.0.1-Linux-x86_64.sh
bash Anaconda3-5.0.1-Linux-x86_64.sh 

unnir

#!/usr/bin/env python
"""
Example of using Keras to implement a 1D convolutional neural network (CNN) for timeseries prediction.
"""

from __future__ import print_function, division

import numpy as np
from keras.layers import Convolution1D, Dense, MaxPooling1D, Flatten
from keras.models import Sequential

unnir

#!/bin/bash

CUDA_REPO_PKG="cuda-repo-ubuntu1604_8.0.61-1_amd64.deb"
wget http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/${CUDA_REPO_PKG}
sudo dpkg -i ${CUDA_REPO_PKG}
sudo apt-get update
sudo apt-get -y install cuda-8-0
#nvidia-smi
echo 'export PATH=$PATH:$CUDA_HOME/bin' >> ~/.bashrc
echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_HOME/lib64' >> ~/.bashrc

unnir