hamori/notebooks/colab_finetune.ipynb

hamori — fine-tuning on personal chord corpus¶

This notebook fine-tunes a pre-trained ChordTransformer on your tokenized .pt files using Google Colab (GPU T4 recommended).

Prerequisites (done locally before uploading):

• python scripts/prepare_data.py --input-dir data/raw_user --output-dir data/processed/user
• python scripts/make_colab_zip.py --mode finetune
• Have checkpoints/pretrained.pt from a completed pre-training run.

Steps:

1. Check GPU
2. Upload hamori_colab_finetune.zip
3. Extract and install dependencies
4. Upload pretrained.pt checkpoint
5. Verify processed data
6. Run fine-tuning
7. Inspect results
8. Download checkpoint and logs

In [ ]:

# ── 1. GPU check ────────────────────────────────────────────────────────────
import torch
if torch.cuda.is_available():
    print(f"GPU: {torch.cuda.get_device_name(0)}")
    print(f"VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")
else:
    print("No GPU found — training will be slow on CPU.")
    print("Go to Runtime → Change runtime type → T4 GPU and re-run.")

In [ ]:

# ── 2. Upload hamori_colab_finetune.zip ──────────────────────────────────────
# Build it locally first:
#   python scripts/make_colab_zip.py --mode finetune
from google.colab import files
uploaded = files.upload()   # select hamori_colab_finetune.zip
print("Uploaded:", list(uploaded.keys()))

In [ ]:

# ── 3. Extract and install dependencies ─────────────────────────────────────
import zipfile, os

WORK_DIR = "/content/hamori"
os.makedirs(WORK_DIR, exist_ok=True)

zip_name = [k for k in uploaded if k.endswith(".zip")][0]
with zipfile.ZipFile(zip_name) as zf:
    zf.extractall(WORK_DIR)
    print(f"Extracted {len(zf.namelist())} files to {WORK_DIR}")

os.chdir(WORK_DIR)
print("Working directory:", os.getcwd())

In [ ]:

# Colab ships torch; only install the extra deps
!pip install -q pretty_midi mido music21 matplotlib

In [ ]:

# ── 4. Upload pretrained checkpoint ─────────────────────────────────────────
# Skip this cell if you built the zip with --include-checkpoint.
import os
from pathlib import Path
from google.colab import files

ckpt_path = Path("checkpoints/pretrained.pt")
if ckpt_path.exists():
    print(f"Checkpoint already present: {ckpt_path} ({ckpt_path.stat().st_size / 1e6:.1f} MB)")
else:
    print("Upload checkpoints/pretrained.pt from your local machine.")
    uploaded_ckpt = files.upload()   # select pretrained.pt
    ckpt_path.parent.mkdir(parents=True, exist_ok=True)
    src = list(uploaded_ckpt.keys())[0]
    os.rename(src, ckpt_path)
    print(f"Saved to {ckpt_path} ({ckpt_path.stat().st_size / 1e6:.1f} MB)")

In [ ]:

# ── 5. Verify processed user corpus ─────────────────────────────────────────
from pathlib import Path
train_pt = list(Path("data/processed/user/train").glob("*.pt"))
val_pt   = list(Path("data/processed/user/val").glob("*.pt"))
print(f"Train: {len(train_pt)} files")
print(f"Val:   {len(val_pt)} files")
if not train_pt:
    print()
    print("ERROR: no training data found.")
    print("Run locally first: python scripts/prepare_data.py ")
    print("  --input-dir data/raw_user --output-dir data/processed/user")
    print("Then rebuild the zip: python scripts/make_colab_zip.py --mode finetune")

In [ ]:

# ── 6. Fine-tune ─────────────────────────────────────────────────────────────
# Outputs:
#   checkpoints/finetuned.pt
#   checkpoints/finetuned.log.csv
#   checkpoints/finetuned_curves.png
#   checkpoints/finetuned.report.txt
!python scripts/train.py

In [ ]:

# ── 7a. Show report ───────────────────────────────────────────────────────────
from pathlib import Path
report = Path("checkpoints/finetuned.report.txt")
if report.exists():
    print(report.read_text(encoding="utf-8"))
else:
    print("Report not found — training may have failed.")

In [ ]:

# ── 7b. Show loss curves ─────────────────────────────────────────────────────
from IPython.display import Image
Image("checkpoints/finetuned_curves.png")

In [ ]:

# ── 8. Download results ───────────────────────────────────────────────────────
import shutil
from google.colab import files

shutil.make_archive("/content/finetune_results", "zip", WORK_DIR, "checkpoints")
files.download("/content/finetune_results.zip")