viking714 / Chord-Recognition

Recognize chords in songs using Transformer || AI CUP 2020 - Chord Recognition Competition (9th place) / 和弦辨識競賽 (第九名)

Chord-Recognition

---

中文版 README

AI CUP 2020 - Chord Recognition Competition Website

---

Environment

• Hardware：
  • CPU：i7-9700
  • GPU：RTX 2070
• Software：
  • Cuda：cuda_11.0.2_451.48_win10
  • cuDNN：cudnn-11.0-windows-x64-v8.0.4.30
  • TensorFlow：tensorflow-2.4.0

Development Flow

1. answerAnalyze.py:

   Observe training data through CMD.

2. visualization.py:

   Observe training data much more conveniently by matplotlib.

3. score.py:

   Comprehend the principle of scoring method & implement the scoring program.

4. main.py → processData.py:

   Preprocess the training data into the input data for models.

5. model.py:

   Start building the fisrt version model.

6. mapping.py:

   Including all required mapping dictionaries of input data Y for the first version model. These mapping dictionaries concludes 544 possible input data Ys in training data -- "CE200".

7. model.py → trainModel.py → oneFrameModel.py:

   Modularize models & rename it to one-frame-input predicting model.

8. multiFrameModel.py:

   Start building the second version model, which can use multiple frames as input data X.

9. Improve processData.py:

   Divide input data X during preprocessing, which significantly improved the accuracy.

p.s. So far, I always input "random 40% part of one song" to predict "answers of the rest 60% part of the song". The oneFrameModel's accuracy is about 80%, and the multiFrameModel can achieve 99.9%. I then realized that this is an incorrect predicting method in the next step.

10. multiFrameModel-2.py → splitDataModel.py:

    Extends the conception of multiple-frame input & dividing input data X. Changing predicting pattern from "60% for a song -> answers of the rest 40%" to "60% songs -> the rest 40% songs". Ex: There are 20 songs in CE200_sample, so I will choose 8 songs as training data, and the rest 12 songs are for validation data. After changing, the accuracy drops to 45%.

11. Improve mapping.py:

    Adopt new mapping dictionary which only contains scoring data Ys. Accuracy raise from 45% to 50%.

12. fasterReadingModel.py：

    Adjust the time of preprocessing to the instant before inputting data X. This significantly reduce the time for data reading & the GPU's RAM usage.

Result

• Got the 9th place.
• The 1st place used the model in this paper.
• The second place used the same model with me, but more preprocess & Post-processing.
• Review: It seems like that information in data are much more important than the architecture of the model. Maybe I should do feature extraction on my own in the next time.