Lesson 8 in-class

rachel · February 28, 2017, 3:15am

Ah, that’s understandable, although I’m worried about the internet already being overloaded. We will share the slides and video afterwards for you to catch the extra details

rachel · February 28, 2017, 3:16am

@renjithmadhavan OpenCV is commonly used for computer vision, although it can definitely be a pain to configure

hamelsmu · February 28, 2017, 3:17am

With the high level APIs now in Tensorflow, what is going to be the difference b/w Keras API and TensorFlow API?

Kjeanclaude · February 28, 2017, 3:18am

Would we have an example on how to solve gradient saturation with the tensorflow tools ?

harveyslash · February 28, 2017, 3:18am

I also read that tensorflow officially supports keras. So what is the difference between keras and the high level api in TF ?

davecg · February 28, 2017, 3:19am

OpenCV is commonly used, although you might find a lot of overlap with scikit-image and scipy.ndimage (which are both simple conda installs).

benediktschifferer · February 28, 2017, 3:19am

about productionalizing deep learning application:

is there an idea for using deep learning on AWS lambda

btw: I would like to work on that as an project

benediktschifferer · February 28, 2017, 3:25am

does precompiling mean, that we will write tensorflow code/test it and when we train a big model, than we precompile the code and train our model?

alex_izvorski · February 28, 2017, 3:25am

All videos from TF Summit here: https://www.youtube.com/playlist?list=PLOU2XLYxmsIKGc_NBoIhTn2Qhraji53cv

prabu · February 28, 2017, 3:28am

What is the relationship between TF and Pytorch

kishore_p_v · February 28, 2017, 3:31am

Can we use TF and Torch together, like some part of code in one rest in other?

harveyslash · February 28, 2017, 3:36am

Doesnt the batch size also depend heavily on the video ram ?
Or does just upgrading RAM allow bigger batch sizes?

davecg · February 28, 2017, 3:38am

He’s talking about GPU RAM.

For keras with TF, this is helpful for larger batch sizes if you have multiple GPUs (splits batch across your GPUs for easy data parallelization):

github.com

kuza55/keras-extras/blob/master/utils/multi_gpu.py

from keras.layers import merge
from keras.layers.core import Lambda
from keras.models import Model

import tensorflow as tf

def make_parallel(model, gpu_count):
    def get_slice(data, idx, parts):
        shape = tf.shape(data)
        size = tf.concat([ shape[:1] // parts, shape[1:] ],axis=0)
        stride = tf.concat([ shape[:1] // parts, shape[1:]*0 ],axis=0)
        start = stride * idx
        return tf.slice(data, start, size)

    outputs_all = []
    for i in range(len(model.outputs)):
        outputs_all.append([])

    #Place a copy of the model on each GPU, each getting a slice of the batch
    for i in range(gpu_count):

This file has been truncated. show original

iNLyze · February 28, 2017, 3:38am

I can confirm NVMe drives are awesome. Just got a Samsung Evo 960. In conjunction with bcolz it is really amazing.

renjithmadhavan · February 28, 2017, 3:38am

I got this idea from part 1 … that deep learning can be used for any machine learning problem( including non-image data sets ) with very good results. Is this a true statement? Does this mean traditional ML methods such as decision tree, linear regression need not be used . Is there a problem where you think deep learning may not be a good idea ?

rachel · February 28, 2017, 3:39am

@iNLyze We will be talking about bcolz later in the course

jpuderer · February 28, 2017, 3:39am

I’m currently using a Google Cloud Engine instance. It’s not bad. Close to the same price as AWS, but they use Titan X Pascal GPUs.

rachel · February 28, 2017, 3:40am

Good question @renjithmadhavan ! We will be using deep learning on structured, time-series data in a later session

Edit: I meant class session in this course

geniusgeek · February 28, 2017, 3:40am

Its annoying to setup, but its nice though

sakiran · February 28, 2017, 3:44am

Where to find the current research papers?