Deep Learning Battle: Pytorch vs. Tensorflow - Jon Krohn

Deep Learning Battle: Pytorch vs. Tensorflow - Jon Krohn, Chief Data Scientist at Nebula

Explore the pros and cons of TensorFlow vs PyTorch for training and deploying deep-learning models

[ClaudioBrandy]

Jon & Harpreet, thank you for this very interesting and informative discussion!

[abubakarsadiqsuleiman]

thanks boss this content really help me clear my doubt ....I go for pytorch

[nightraver]

Very intersting!
Im coming in from Industry development of Vision inspection programming using Cognex & Keyence sensors & IDE.

Understanding of photography & optics IS important, depending on the criteria and I cant see how "Deep Learning" is going to save time initially when building a model. It could certainly be useful optimizing an existing model, but "overkill" is the nicest word I could use, probably "gratuitious waste of resources" would be better.

For example, if we get pattern-detection as:

1) scan reference image typically in greyscale white=255 black=0 greyscale=<255 && >0
, pattern-definition festures like lines=contiguous pixels that changed from light to dark

for non-reflective objects everything is peachy, reflective surfaces like chrome are going to be a nightmare if you keep trying to optimize pattern-recognition on a bunch of sensor captures with reflective flare.

You could spend 00s of 000s of hours running your oversold deep-learning model trying to extrapolate a pattern-match from crazy numbers of edge-cases,

or you read a bit on light reflectivity & dispersion, and have your light-source bounce off an egg-shell white bowl, instead of bouncing the light directly off the chrome object, that will get you a predictive model much faster & much lower compute.

What about a vision-inspection model for detecting cracks on shiny metal stampings, you discover alot of stamping markings are the same grayscale pixel color & shape of legitimate cracks so you're flooded with false positives.

Hey increase the compute and feed it millions of samples maybe the deep learning will figure out some pixel signature that detects the difference between a crack and a stamp-marking or scrape, which is not a defect.

Does the client have time for you to teach your deep-learning model for 1 year while it keeps false positiving good parts? and maybe letting some bad parts through costing the company alot of money in client-required sorting & human-inspection fees?

A basic understanding of optics might lead you to project a grid of lasers at the part, flare reflections are gone, scrapes are still flat unbroken metal, but cracks will show a pronounced broken line in the laser projection.

How many random forms of lighting, reflection, refraction, distance-measuring laser feedback, laser-grid projections <<ad nauseum>> are you gonna have to run through your deep learning model?

[jesterflint]

I like tensorflow just for the documentation and the youtube videos from Google.

[drm]

Please tell me, is opencv certification an good thing to get ?