- Learn to manipulate MIDI file and represent symbolic music as tokens.
- Learn to train a transformer-based model for symbolic music generation.
Train a transformer-based model from scratch to generate 32 bars symbolic music.
- You may randomly select the prompt, treat it as a priming sequence for Transformer decoder to generate a continuation
- For example: [Bar_None, Position_1/16, Tempo_class, Tempo_value]
- Model: You can use any transformer-based model.
- Either 1-stage generation or 2-stage generation is fine.
TA will provide 3 midi files (8 bars) as prompt, you need to generate their continuation for 24 bars. (Total: 8+24 bars)
- You can use checkpoint in Task 1, don’t need to train another model.
- 1747 pop music piano performances (mid or midi file) transcribed from youtube audio.
- Single track, 4 minutes in duration, totaling 108 hours.
- 4/4 time signature (four beats per bar).
Objective metrics for generation results
- Pitch-Class Histogram Entropy (H) : measures erraticity of pitch usage in shorter timescales (e.g., 1 or 4 bars). (H4 is required)
- Grooving Pattern Similarity (GS) : measures consistency of rhythm across the entire piece. (required)
- Structureness Indicator (SI) : detects presence of repeated structures within a specified range of timescale. (optional)
# Clone the repo:
git clone https://github.com/PANpinchi/DeepMIR_HW3_PANpinchi.git
# Move into the root directory:
cd DeepMIR_HW3_PANpinchi# Create a virtual conda environment:
conda create -n deepmir_hw3 python=3.10
# Activate the environment:
conda activate deepmir_hw3
# Install PyTorch, TorchVision, and Torchaudio with CUDA 11.3
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.3 -c pytorch
# Download GPT2RGA.py
wget "https://raw.githubusercontent.com/asigalov61/tegridy-tools/main/tegridy-tools/GPT2RGA.py"
# Install additional dependencies from requirements.txt:
pip install -r requirements.txtRun the commands below to download the pre-trained model.
# The pre-trained GPT-2 model with Chord
gdown --folder https://drive.google.com/drive/folders/1pUOR4So048nR02UIk4no-KxgkU00oBn4?usp=drive_link
# The pre-trained GPT-2 model without Chord
gdown --folder https://drive.google.com/drive/folders/1EdvkETVFvjD-dDXKI94imOfndh06S0Yu?usp=drive_linkNote: The best_model.pth files should be placed in the setting_1 and setting_2 folders respectively.
You need to execute the following commands to download, unzip and preprocess Pop1K7 dataset.
# Download the Pop1K7 dataset
wget -O Pop1K7.zip "https://zenodo.org/records/13167761/files/Pop1K7.zip?download=1"
# Unzip
unzip Pop1K7.zip
# Preprocess
python movefile.py# Train a GPT-2 model with Chord.
python main.py --checkpoint_folder setting_1 --batch_size 8 --use_chord
# Train a GPT-2 model without Chord.
python main.py --checkpoint_folder setting_2 --batch_size 8# Settting 1: Test a GPT-2 model with Chord.
python generate_32_bars.py --use_chord --saved_dir setting_1 --top_k 50 --temperature 1.0 --output_dir outputs_1
# Settting 2: Test a GPT-2 model with Chord.
python generate_32_bars.py --use_chord --saved_dir setting_1 --top_k 50 --temperature 2.0 --output_dir outputs_2
# Settting 3: Test a GPT-2 model with Chord.
python generate_32_bars.py --use_chord --saved_dir setting_1 --top_k 100 --temperature 1.0 --output_dir outputs_3
# Settting 4: Test a GPT-2 model without Chord.
python generate_32_bars.py --saved_dir setting_2 --top_k 50 --temperature 1.0 --output_dir outputs_4
# Settting 5: Test a GPT-2 model without Chord.
python generate_32_bars.py --saved_dir setting_2 --top_k 50 --temperature 2.0 --output_dir outputs_5
# Settting 6: Test a GPT-2 model without Chord.
python generate_32_bars.py --saved_dir setting_2 --top_k 100 --temperature 1.0 --output_dir outputs_6# Compute metrics on real data
python eval_metrics.py --dict_path basic_event_dictionary.pkl --output_file_path Pop1K7/midi_analyzed --output_csv pop1k7.csv
# Compute metrics on Setting 1 results
python eval_metrics.py --dict_path basic_event_dictionary.pkl --output_file_path outputs_1 --output_csv outputs_1/outputs_1.csv
# Compute metrics on Setting 2 results
python eval_metrics.py --dict_path basic_event_dictionary.pkl --output_file_path outputs_2 --output_csv outputs_2/outputs_2.csv
# Compute metrics on Setting 3 results
python eval_metrics.py --dict_path basic_event_dictionary.pkl --output_file_path outputs_3 --output_csv outputs_3/outputs_3.csv
# Compute metrics on Setting 4 results
python eval_metrics.py --dict_path basic_event_dictionary.pkl --output_file_path outputs_4 --output_csv outputs_4/outputs_4.csv
# Compute metrics on Setting 5 results
python eval_metrics.py --dict_path basic_event_dictionary.pkl --output_file_path outputs_5 --output_csv outputs_5/outputs_5.csv
# Compute metrics on Setting 6 results
python eval_metrics.py --dict_path basic_event_dictionary.pkl --output_file_path outputs_6 --output_csv outputs_6/outputs_6.csv# Settting 1: Test a GPT-2 model without Chord.
python inference.py --use_chord --saved_dir setting_1 --inputs prompt_song/song_1.mid --top_k 50 --temperature 1.0 --output_dir outputs_1
python inference.py --use_chord --saved_dir setting_1 --inputs prompt_song/song_2.mid --top_k 50 --temperature 1.0 --output_dir outputs_1
python inference.py --use_chord --saved_dir setting_1 --inputs prompt_song/song_3.mid --top_k 50 --temperature 1.0 --output_dir outputs_1
# Settting 2: Test a GPT-2 model without Chord.
python inference.py --use_chord --saved_dir setting_1 --inputs prompt_song/song_1.mid --top_k 50 --temperature 2.0 --output_dir outputs_2
python inference.py --use_chord --saved_dir setting_1 --inputs prompt_song/song_2.mid --top_k 50 --temperature 2.0 --output_dir outputs_2
python inference.py --use_chord --saved_dir setting_1 --inputs prompt_song/song_3.mid --top_k 50 --temperature 2.0 --output_dir outputs_2
# Settting 3: Test a GPT-2 model without Chord.
python inference.py --use_chord --saved_dir setting_1 --inputs prompt_song/song_1.mid --top_k 100 --temperature 1.0 --output_dir outputs_3
python inference.py --use_chord --saved_dir setting_1 --inputs prompt_song/song_2.mid --top_k 100 --temperature 1.0 --output_dir outputs_3
python inference.py --use_chord --saved_dir setting_1 --inputs prompt_song/song_3.mid --top_k 100 --temperature 1.0 --output_dir outputs_3
# Settting 4: Test a GPT-2 model without Chord.
python inference.py --saved_dir setting_2 --inputs prompt_song/song_1.mid --top_k 50 --temperature 1.0 --output_dir outputs_4
python inference.py --saved_dir setting_2 --inputs prompt_song/song_2.mid --top_k 50 --temperature 1.0 --output_dir outputs_4
python inference.py --saved_dir setting_2 --inputs prompt_song/song_3.mid --top_k 50 --temperature 1.0 --output_dir outputs_4
# Settting 5: Test a GPT-2 model without Chord.
python inference.py --saved_dir setting_2 --inputs prompt_song/song_1.mid --top_k 50 --temperature 2.0 --output_dir outputs_5
python inference.py --saved_dir setting_2 --inputs prompt_song/song_2.mid --top_k 50 --temperature 2.0 --output_dir outputs_5
python inference.py --saved_dir setting_2 --inputs prompt_song/song_3.mid --top_k 50 --temperature 2.0 --output_dir outputs_5
# Settting 6: Test a GPT-2 model without Chord.
python inference.py --saved_dir setting_2 --inputs prompt_song/song_1.mid --top_k 100 --temperature 1.0 --output_dir outputs_6
python inference.py --saved_dir setting_2 --inputs prompt_song/song_2.mid --top_k 100 --temperature 1.0 --output_dir outputs_6
python inference.py --saved_dir setting_2 --inputs prompt_song/song_3.mid --top_k 100 --temperature 1.0 --output_dir outputs_6