Built without use of any framework besides tensorflow's low level API (see tf.matmul,tf.add, etc). F1 Score is currently at 82% Built by @neelsankaran and I
Input Data: Protein Primary Structure | Shape: (5600(examples),700(amino acids in sequence),22(features in embeding))
Labels: Protein Secondary Structure | Shape: (5600(examples),700(amino acids in sequence),10(features in embeding))
Loss Function: Cross Entropy
Y^(0) | LSTM Updates A With Previous Relevant Data |
^ ^
| | Next Y^
| | ^
---------------------------------------- | |
| | | -----------------
| | | | |
| Y^ = tanH(conc*Wy'+By) | | | |
A(0) -> | nextA = tanH((In*Wx'+Bx)+(A*Wa'+Ba)) | ----> nextA ----> | NextTimeStep |....N=700 -> Calc Cost With All Y^'s |
| | | | |
| | | | |
| | | | |
---------------------------------------- ----------------- |
^ ^ |
| | |
| Next Input |
Input Data(0) | |
Gradient From Cost Flows Back |
<---------------------------------------------------------------------------------------------------------------------