Posts

Showing posts from June, 2020

Converting Caffe models to tensorflow 2.2 models in colab

Posting this quickly because I could not for the life of me figure out how to do it !mmconvert --srcFramework caffe --inputWeight /tmp/nst/resnext50.caffemodel --inputNetwork /tmp/nst/resnext50_deploy.prototxt --dstFramework tensorflow --outputModel /tmp/nst/out import tensorflow as tf imported = tf.saved_model.load('/tmp/nst/out') s = imported.signatures['serving_default'] print(s.inputs) # Tells you desired shape of inputs test_data = tf.zeros([1, 224, 224, 3]) s(tf.constant(test_data))

Full regression results

Image
This is a technical post for people interested in details. The punchline is: even if you include a wide variety of possible predictors of success, a simple least-squares model looking only at likes is quite similar in performance to more complicated models using followers, tags, etc. I don't think the information here is a very wide interest, so I'm just going to present results and link to my GitHub for those who want details. Feel free to ask questions in the comments if you would like more info. Lasso Ridge OLS OLS-Likes only 1 0 0 2854.06965 N/A Shares 0 11359.86 65.6032316 N/A Comments 0 222.1719 -10.255222 N/A Likes 4.800058 -422.25 4.2506603 5.79909414 Follower Count 0.071327 -15296 0.09265454 N/A Following Count 0 -21.6523 0.67554451 N/A Bio is empty 0 2395.795 2326.99936 N/A Any Hashtags 0 20106.24

How important is the first video you create on TikTok?

Image
Summary : There is a widespread belief that the first video one creates on TikTok 1) gets an automatic “boost” from the recommendation algorithm and 2) has extra importance in influencing future video success. I test these claims on a data set of 33,925 videos and find that both propositions are unsupported by the evidence. If you have been holding off on creating your first video because you want it to be perfect, you should just go ahead and make an imperfect one. Background I sometimes hear people argue that the first video one creates on TikTok is unusually important. For example : The first video that any user on Tik Tok posts gets an instant spike in impressions. As a new user, Tik Tok metaphorically “throws you a bone” and will in most cases show your video to more people right off the bat than it would a video by a repeat poster. This means that it’s vital to make your very first post on Tik Tok engaging. Unlike Youtube, for example, where you can feel free to po

What impact does the number of followers have on TikTok view counts?

Image
In a previous post , we estimated the impact of various metrics on the view counts for a video. This post extends that analysis to include information about the creator’s metrics, notably their follower count. Key results: Each additional follower increases the expected view count by 0.1 (i.e. ~10% of followers see each video) Other results are qualitatively similar (i.e. you need ~0.2 likes per view in order to have your video be promoted; share and comment counts are not predictive) Detailed results Please refer to our previous post for more information on the data set and methods used. Transformation Regression Type Shares Comments Likes Follower Count Intercept Alpha R^2 Normal Lasso 0.00 0.00 5.03 0.10 -1176.69 100000000.00 0.89 Normal Ridge -3.73 -10.95 5.17 0.10 -2334.26 100000000.00 0.89 Normal OLS

Are likes, shares, or comments more important for TikTok video success?

Image
Summary TikTok’s recommendation algorithm is an iterative process which repeatedly displays videos to larger and larger test audiences, until engagement from the test audience falls below some threshold. It is unknown how “engagement” is measured, but it is usually speculated to be some combination of likes, comments, and shares. Various people have hypothesized the weights of these metrics (e.g. one share = 10 likes), but, to our knowledge, no one has ever measured it. We therefore attempted to measure this on a representative sample of ~20,000 videos. Our research concludes that these metrics are highly collinear and no predictive value is found from comment and share count. (I.e. you should just optimize for likes, and don’t need to optimize for shares or comments .) You will need about 1/5 of your viewers to like your video in order for it to be promoted . Background and Motivation More background on the structure of the TikTok algorithm can be found here . Veed cont