Post a Contest

modify tensorflow network to use binary instead of onehot encoding

Status: Closed
Prize: $20
Entries Received: 2
Winner: msavinash1139

Winner

Contest Brief

I need a simple change made to this sample code (attached). It was downloaded from https://github.com/atramos/titonn

Specifically I want to use a 7-bit binary encoding instead of the current one-hot encoding for each encoded character. I don't know how to do this in TF so you need to figure this out for me. For example the letter 'A' should be encoded as seven fields with value 1,0,0,0,0,0,1 (0b1000001 = 65 = equals 'A' in ASCII). In the current code only 6 values are allowed (a, b, c, d, e, f) but after this change, the entire 7-bit ASCII character set will be allowed.

This means also a different error function has to be used instead of softmax_cross_entropy_with_logits, but I don't know in advance what the correct function is. I will know the correct one when I see it in a working solution. Please discuss your reasoning in the comments so that I can be confident you have the right solution.

Contest will become #guaranteed after I see at least one serious submission.

Recommended Skills

Winner

Post a Contest Like This

Top entries from this contest

msavinash1139 India

0
dellkj1996 Algeria

0

Public Clarification Board

dellkj1996
- 3 years ago
Can you check entries and suggestions please
- 3 years ago
anubhavkumarc
- 3 years ago
If it's just about representing every character, why don't you just use sparse one hot encoding? You can represent the entire charset with that while using the predefined loss function.

I will try working on the binary encoding loss and submit it later though
- 3 years ago
1. Contest Holder
  - 3 years ago
  I am not familiar with sparse one-hot encoding. Does it reduce memory utilization?
  - 3 years ago
2. anubhavkumarc
  - 3 years ago
  It does because it uses sparse structures rather than a full blown array. One hot encoding has arrays which are mostly empty with only one position with the data, and sparse vectors can take advantage of that for it to be represented with just the position of that vector and the size of the array
  - 3 years ago