Tracking

Stefan

01.11.

• Read about neural language models in
  • A Neural Probabilistic Language Model
    • Learning simultaneously: Distributed Word Vector Representations and a Statistical Language Model (given all previous words, what's the probability distribution for the next word?)
    • Shallow model, but can process large datasets
    • N previous Words are translated via 1-V-Mapping
  • Efficient Estimation of Word Representations in Vector Space
  • Read https://code.google.com/p/word2vec/

02.11.

• Read through distance.c in word2vec to understand word2vec binary format
• Write Python program to read word vectors with Joseph
• Set up virtual python environment on server

09.11.

• Basic network, and experiment infrastructure
• Write glue code connecting sentence files, sliding window and level db creation

12.11.

• Create databases, run first experiment on server
• Caffe Multi-class Precision and Recall

From here I forgot to do the daily tracking, below is a summary of the remaining project.

• Caffe Multi-class Precision and Recall
• Debug memory consumption in word2vec (float64 to float32)
• Develop configuration file infrastructure
• Extract Wikipedia files in plain text, develop filters to maintain good text quality
• Skeleton for web app
• Refactoring of training instance generation
• Debugging our net/architecture/code for issues regarding our low precision and recall
• Write program to generate all configurations for the experiments
• Adapt experiment.sh and training.sh files
• Continuously run and monitor the experiments over christmas
• Search for audio feature libraries to extract the pitch and energy levels, conversion of raw files and developed scripts for extraction
• Run experiment with train and test from Wikipedia
• Run experiments with info gain loss matrix to tackle class imbalance
• Refactoring of audio feature generation
• Audio deployment
• Develop executable fusion evaluation
• Double check evaluation results and chart plots
• Final presentation
• Paper

Joseph

• Read https://code.google.com/p/word2vec/
• Understand distance.c, implement Python program to read word vectors
• Explore how to use NLTK for POS tagging
• Write python script for demo show cases
• Documentation of demo script
• Net configuration script
• PlainText parser
• Unify python paths and working directory
• Help messages + README for python script folder
• Generator usage to reduce memory consumption
• Intermediate presentation
• Progress for training instance generation and parsing adapted
• Python path problems fixed
• Config file
• Initial LSTM testing
• Initial paper stub
• Parameter evaluation script improved and generalized
• Chart for harmonic mean, comparison of different features
• Help Ricarda with error fixing for demo model selection
• Backwards compatibility of config files
• Convenience download script for downloading all model files needed for demo from server
• Try batch processing for speed up of demo (unsuccessful - reverted)
• Initial fusion methods and class structure, evaluation of these done by Stefan
• Refactoring for improved readability and versatility of Javascript Code
• Improved and revised readme file with Ricarda
• Fix Acoustic Model Index Conversion
• Paper: Worked on introduction, conclusion, demo sections

Tanja

• Read https://code.google.com/p/word2vec/
• Raad some papers
• Write Python script to parse xml and ASR transcript files
• Write Python script to create basic training instances using a sliding window
• Write training instances to leveldb script
• Ensure valid train and test split
• Caffe Multi-class Precision and Recall
• Pipeline work
• Use POS-Tags as features
• Introduced a flag to turn on/off POS-Tagging
• Use parameters from config file
• Refactoring the input parser
• Web Demo
• Presentation
• Debugging our net/architecture/code for issues regarding our low precision and recall
• Several trainings to get the baseline
• Continously run and monitor the experiments over christmas
• Converting xml and txt files into line format. POS Tags can be preprocessed and written to disk.
• Main program gets only a config file as argument
• Refactoring of line parser
• Preprocessing of POS-Tagging: Write data files with POS tags
• Refactroing of sliding window: Punctuation pos can be at any position
• Debugging of Word2Vec: Use Float32 instead of Float64
• Parse ctm files for generating accoustic training instances.
• Parse pitch and energy files. Create pitch and energy features for the audio model.
• Include audio model into web demo
• Implement first basic fusion
• Major refactoring of web demo backend
• Implement evaluation of fusion
• Final presentation
• Paper

Ricarda

• Read https://code.google.com/p/word2vec/
• Read papers to get familiar with Deep Learning
• Read papers for lexical sentence boundary approaches
• Write Python script to parse xml and ASR transcript files
• Write Python script to create basic training instances using a sliding window
• Refactor script for creating trainings instances and work on pipeline to create instances
• Write python script for demo show cases
• Use POS-Tags as features
• LineParser implemented
• Generator usage to reduce memory consumption
• Script to collect experiment results
• Web demo model selection
• use balanced data for trainings/testing including flag in config
• json converter for prediciting results to demo
• show pos tags in demo together with joseph
• loading spinner in demo
• fix selection options in demo
• Demo writing results to file
• Demo choose input text from existing files
• Setup Demo on Server and fix availablility from outside
• Write guide for demo setup
• Paper