Now available for you to try at emojifier.co.uk!
Meme ๐ธ is love, โค๐๐ meme ๐ธ is life.
Using ๐ธ a combination โค๐๐ of ๐ธ word ๐ embeddings and neural ๐ networks, ๐ต the 5โฃ program predicts which emoji โจ is useful ๐ง based on ๐ sentence ๐ context.
We ๐บ๐ธ collected ยฎ data ๐ from ๐ Reddit ๐จ๐บ and Twitter ๐บ and ๐ using ๐ต TensorFlow 5โฃ we ๐บ๐ธ trained ๐ the 5โฃ different 3โฃ models ๐ such ๐ง as ๐ character ๐ค level ๐ LSTM, word ๐ level ๐ LSTM and fixed input neural ๐ network ๐ฐ and also using Google's ๐ word2vector for โนโบ word ๐ embeddings.
Tying ๐บ๐ธ the 5โฃ front ๐ end ๐ to ๐จ๐บ the 5โฃ back ๐โโฉ end ๐ was ๐ต very ๐ค hard. ๐ฃ We ๐บ๐ธ started ๐ out โ๐ with ๐ Ruby ๐ on ๐ Rails ๐ to ๐ start ๐ a ๐ background ๐ Python ๐๐ process which was used ๐ for โนโบ the 5โฃ emoji โจ prediction but ๐ we ๐บ๐ธ could ๐ not ๐ค find ๐ a way ๐ to append ๐ the 5โฃ standard input stream for โนโบ the 5โฃ process, so ๐ค instead ๐ถ we ๐บ๐ธ used Node.js to run ๐ฝ in ๐ parallel with ๐ a Python ๐๐ server ๐ป which listened ๐ for โนโบ POST ๐ฃ๐ค requests from ๐ the 5โฃ Node.js client and then ๐ using Twmoji to beautifully ๐ render the 5โฃ final โถ emojis.
Describing ๐บ๐ธ emojis 5โฃ as ๐ a ๐ single ๐ word, ๐ problems ๐โโฉ with ๐ training ๐ data ๐ on ๐ TensorFlow. ๐บ๐ธ
Having ๐บ๐ธ a 5โฃ finished ๐ product ๐ท in ๐ our ๐บ๐ธ first 9โฃ hackathon! ๐
Learning ๐ซ six 6โฃ7โฃ5โฃ different 3โฃ programming ๐บ languages, ๐ช๐จ more ๐บ๐ธ about 9โฃ machine ๐ฐ learning, ๐ซ and using real world ๐บ๐๐ data. ๐
Turning โคต text ๐ into ๐ raw ๐ memes: ๐ธ we ๐บ๐ธ already ๐ have ๐ฐ some ๐ extra features built ๐ and ๐บ๐๐ ready ๐ to go ๐น๐ฌ they're ๐ just ๐ค waiting ๐ for โนโบ front ๐ end ๐ implementations.