Video Caption
NTU 2017 Fall MLDS: Assignment 2
Describe the videos based on content. This task is to generate the description for the given video.
Results
Architecture: Seq2Seq
Simply use Seq2Seq is good enough to perform the task.
Experiment & Parameters
Some hints about training
- epoch: 200
- batch size: 64
- cell type: LSTM
- initalization: Orthogonal
- hidden size: 256
- layers: 2
- dropout: 0.5
- optimizer: RMSProp
- learning rate: 0.0001
- learning rate decay rate: 0.25
- decay epoch: 50
- loss function: Seqeunce Loss (cross entropy)
- attention: local (atten)
Dataset
MSVD : 1450 videos for training and 100 videos for testing Link: https://drive.google.com/file/d/0B18IKlS3niGFNlBoaHJTY3NXUkE/view