ps/ClimbingBoardGPT

Fork 0

Files

Pawel 528ea9058f Polish + more info

2026-05-25 15:27:10 -04:00

21 KiB

Raw Blame History

ClimbingBoardGPT

ClimbingBoardGPT is a unified transformer-style modeling project for climbing-board routes on:

Tension Board 2 Mirror (12ftx12ft)
Kilter Board Original (16ftx12ft)

The project treats climbing-board problems as symbolic sequences of board-aware hold-role tokens. It supports:

joint route tokenization for TB2 and Kilter,
transformer-based grade prediction,
GPT-style route generation conditioned on board, wall angle, and target grade,
calibrated board-background visualization,
command-line demo scripts for generation and grade prediction,
interactive FastAPI webapp with board-image overlay and click-to-build route prediction.

This repo is the transformer/GPT follow-up project to Tension-Board-2-Analysis and Kilter-Board-Analysis.

Core idea

A route is represented as a sequence like:

<BOS> <BOARD_TB2> <ANGLE_40> <GRADE_V6>
<TB2_p344_start> <TB2_p369_middle> <TB2_p603_finish>
<EOS>

or:

<BOS> <BOARD_KILTER> <ANGLE_40> <GRADE_V6>
<KILTER_p1084_start> <KILTER_p1231_middle> <KILTER_p1395_finish>
<EOS>

Hold tokens are board-namespaced, so a TB2 placement ID and a Kilter placement ID never collide.

For grade prediction, the grade token is removed:

<CLS> <BOARD_TB2> <ANGLE_40>
<TB2_p344_start> <TB2_p369_middle> <TB2_p603_finish>
<EOS>

The model then predicts the climb difficulty from the board, angle, and hold-role tokens.

How generation and grading work

The project uses one shared vocabulary across both boards. Every climb is converted into a short symbolic sequence: board token, angle token, optional grade token, and one token per hold/role pair. Hold tokens also carry board identity, so the model can learn TB2 and Kilter patterns together without mixing placement IDs.

The grade predictor is a transformer encoder. For this task the grade token is removed and <BOS> is replaced with <CLS>. The model reads the board, angle, hold roles, and learned coordinate features for each hold token, then regresses a continuous difficulty value. That numeric prediction is mapped back into a grouped V-grade for demos and evaluation.

At inference time, grade prediction is:

parse a frames string into (placement_id, role_id) pairs,
canonicalize the route order using role, height, and horizontal position,
convert the route to model tokens such as <CLS> <BOARD_TB2> <ANGLE_40> <TB2_p344_start> ... <EOS>,
encode those tokens as integer IDs and pad/truncate to the model's max sequence length,
add three coordinate features for each token: normalized x, normalized y, and whether the token is a hold,
run the transformer encoder and read the final <CLS> representation,
pass the route through a neural network to get a continuous difficulty prediction,
map that prediction into the grouped V-grade scale.

The route generator is a small GPT-style causal transformer. It starts from a prompt such as:

<BOS> <BOARD_KILTER> <ANGLE_40> <GRADE_V6>

Then it samples the next token repeatedly until <EOS> or a maximum length is reached. At each step:

the current sequence is cropped to the model's context window,
the causal transformer predicts logits for the next token,
forbidden tokens such as <PAD>, <UNK>, <BOS>, <CLS>, and <MASK> are masked out,
logits are divided by the sampling temperature,
optional top-k filtering keeps only the k most likely next tokens,
softmax turns the filtered logits into probabilities,
torch.multinomial samples one next token from that probability distribution,
the sampled token is appended to the sequence.

Lower temperature makes the distribution sharper and more conservative. Higher temperature flattens it and makes unusual tokens more likely. Top-k prevents very low-probability tokens from being sampled at all. The sampled hold-role tokens are converted back into a frames string such as p1084r12p1231r13....

Generation is checked after sampling rather than hard-constrained during decoding. The helper code removes duplicate placements, checks that all holds belong to the requested board, requires starts and finishes, and the webapp retries a few times when valid_only is enabled. The trained grade predictor can also score generated climbs as a critic, which is how the evaluation measures whether generated routes are close to the requested grade.

Quantitative results from the executed notebooks

These numbers come from the executed four-notebook run included with the project. They should be treated as the current benchmark for this checkpoint/data snapshot; rerun the pipeline if the raw databases, tokenization, model sizes, or train/validation/test split change.

Dataset and tokenization scale

The unified tokenizer builds one shared corpus across TB2 and Kilter.

Quantity	Value
Total route/angle entries	321,085
TB2 entries	42,596
Kilter entries	278,489
Placement metadata rows	1,139
Shared vocabulary size	4,438 tokens
Special tokens	6
Board tokens	2
Angle tokens	12
Grade tokens	16
Hold-role tokens	4,402
Grade-predictor max sequence length	398
GPT-generator max sequence length	399

The train/validation/test split used in the executed notebooks was:

Board	Train	Validation	Test
TB2	33,719	4,430	4,447
Kilter	223,112	27,555	27,822
Total	256,831	31,985	32,269

Grade prediction performance

The grade predictor is a transformer encoder trained jointly on both boards. It receives board, angle, hold-role tokens, and coordinate features, but does not receive the grade token.

Metric	Overall	TB2	Kilter
MAE	1.481	1.420	1.490
RMSE	1.941	1.845	1.956
R²	0.768	0.800	0.763
Exact grouped V-grade	36.0%	37.3%	35.8%
Within ±1 V-grade	79.3%	80.0%	79.2%
Within ±2 V-grades	94.8%	95.5%	94.7%

The model has about 1.17M parameters. In the executed run, early stopping selected epoch 8 with validation MAE ≈ 1.480.

Route generator training

The route generator is a GPT-style causal transformer trained on grade-conditioned route sequences.

Quantity	Value
Model size	~1.41M parameters
Best validation loss	3.187
Best validation perplexity	24.2
Evaluation sample size	400 generated routes
Overall basic validity	91.5%
Overall strict validity	91.5%

During the generator evaluation run, routes were sampled across both boards, common angles, and target grades V1–V8.

Generated-route evaluation

Generated routes are evaluated by structural validity, novelty against real climbs, geometric features, and grade consistency using the trained grade predictor as a critic.

Metric	TB2	Kilter
Generated routes evaluated	200	200
Basic validity	89.0%	94.0%
Strict validity	89.0%	94.0%
Mean novelty distance	0.656	0.634
Median novelty distance	0.667	0.652
Mean generated hold count	11.11	12.90
Mean route height	130.76	142.32
Mean route width	61.66	74.94
Mean hand-reach distance	50.41	57.53

Grade consistency of generated climbs, measured by the trained grade predictor:

Metric	Overall	TB2	Kilter
Exact requested V-grade	28.2%	29.5%	27.0%
Within ±1 V-grade	70.8%	68.5%	73.0%
Within ±2 V-grades	92.0%	90.5%	93.5%
Mean V-grade error	—	-0.18	-0.30

Interpretation: the generator is usually structurally valid and usually close to the requested grade according to the critic, but exact grade control remains imperfect. That is expected: this is a small GPT-style model trained on symbolic route data, not a production setter.

Repository layout

ClimbingBoardGPT/
├── configs/
│   ├── tb2.json
│   └── kilter.json
├── data/
│   ├── raw/
│   │   ├── tb2.db
│   │   └── kilter.db
│   └── processed/
├── images/
│   ├── tb2_board_12x12_composite.png
│   └── kilter-original-16x12_composite.png
├── models/
│   ├── joint_transformer_grade_predictor.pth
│   └── joint_route_gpt_generator.pth
├── notebooks/
│   ├── 01_unified_route_tokenization.ipynb
│   ├── 02_joint_transformer_grade_prediction.ipynb
│   ├── 03_joint_route_generator.ipynb
│   └── 04_generated_route_evaluation.ipynb
├── scripts/
│   ├── 01_tokenize_routes.py
│   ├── 02_train_grade_predictor.py
│   ├── 03_train_route_generator.py
│   ├── 04_evaluate_generated_routes.py
│   ├── demo_generate_and_visualize.py
│   ├── demo_generate_tb2.py
│   ├── demo_generate_kilter.py
│   ├── demo_predict_grade.py
│   ├── demo_predict_tb2.py
│   └── demo_predict_kilter.py
├── src/climbingboardgpt/
├── webapp/
│   ├── app.py
│   ├── app.css
│   ├── app.js
│   ├── index.html
│   └── Dockerfile
├── docker-compose.webapp.yml
├── LICENSE
├── README.md
├── requirements.txt
└── pyproject.toml

Setup

Create and activate a virtual environment:

python -m venv .venv
source .venv/bin/activate

Install the package:

pip install -r requirements.txt
pip install -e .

For CPU-only demo use on a small VPS, the scripts support:

--torch-threads 1

This caps PyTorch CPU thread usage.

Data expected by the full training pipeline

The full tokenization/training pipeline expects raw board databases at:

data/raw/tb2.db
data/raw/kilter.db

These databases can be downloaded with the BoardLib CLI commands recorded in the board config files. After that import step, the project treats them simply as source board data.

The project configs are:

configs/tb2.json
configs/kilter.json

They define board-specific details such as:

database path,
layout ID,
role IDs,
token prefix,
angle cutoff,
optional date / placement filters.

The demo scripts do not need the raw databases if the processed tokenization artifacts and trained model checkpoints already exist.

The interactive webapp also needs local demo assets:

data/processed/tokenized/token_metadata.csv
models/joint_transformer_grade_predictor.pth
models/joint_route_gpt_generator.pth
images/tb2_board_12x12_composite.png
images/kilter-original-16x12_composite.png

These files are ignored by git because they are generated or binary artifacts. Recreate them with the training pipeline, copy them from a previous run, or mount them into the Docker container as shown in docker-compose.webapp.yml.

Fast test pipeline

To verify that scripts 01 through 04 still work without retraining the full models, run the pipeline into a temporary output directory with a tiny data sample and tiny CPU-only models:

python scripts/01_tokenize_routes.py \
  --out-dir /tmp/cbgpt_smoke/tokenized \
  --max-routes-per-board 20

python scripts/02_train_grade_predictor.py \
  --tokenized-dir /tmp/cbgpt_smoke/tokenized \
  --out-dir /tmp/cbgpt_smoke/grade_prediction \
  --model-dir /tmp/cbgpt_smoke/models \
  --smoke-test

python scripts/03_train_route_generator.py \
  --tokenized-dir /tmp/cbgpt_smoke/tokenized \
  --out-dir /tmp/cbgpt_smoke/generation \
  --model-dir /tmp/cbgpt_smoke/models \
  --smoke-test \
  --generate-angles 40 \
  --generate-grades 6

python scripts/04_evaluate_generated_routes.py \
  --tokenized-dir /tmp/cbgpt_smoke/tokenized \
  --generated-dir /tmp/cbgpt_smoke/generation \
  --out-dir /tmp/cbgpt_smoke/evaluation \
  --grade-model-path /tmp/cbgpt_smoke/models/joint_transformer_grade_predictor.pth \
  --device cpu

The resulting metrics and generated climbs are not meaningful. This path is only a code-path check: it verifies database loading, tokenization, training loops, checkpoint saving/loading, generation, and evaluation without touching the normal data/processed or models outputs.

Full training pipeline

From the repository root:

python scripts/01_tokenize_routes.py --boards tb2,kilter
python scripts/02_train_grade_predictor.py
python scripts/03_train_route_generator.py
python scripts/04_evaluate_generated_routes.py

This produces the main processed artifacts and trained checkpoints.

Tokenization outputs

data/processed/tokenized/
├── route_sequences.csv
├── routes_tokenized.jsonl
├── token_vocab.json
├── token_metadata.csv
├── placement_metadata.csv
└── board_summary.csv

Grade-prediction outputs

data/processed/grade_prediction/
├── training_history.csv
├── test_predictions.csv
├── board_metrics.csv
└── overall_metrics.json

models/
└── joint_transformer_grade_predictor.pth

Route-generation outputs

data/processed/generation/
├── training_history.csv
└── generated_routes.csv

models/
└── joint_route_gpt_generator.pth

Generated-route evaluation outputs

data/processed/evaluation/
├── generated_route_evaluation.csv
└── top_generated_candidates.csv

Generate routes and visualize them

After training the route generator, or after placing a trained checkpoint at:

models/joint_route_gpt_generator.pth

you can generate and visualize climbs.

TB2

python scripts/demo_generate_tb2.py --angle 40 --grade 6 --n 4

Kilter

python scripts/demo_generate_kilter.py --angle 40 --grade 6 --n 4

Generic version

python scripts/demo_generate_and_visualize.py \
  --board tb2 \
  --angle 40 \
  --grade 6 \
  --n 4 \
  --temperature 0.9 \
  --top-k 50

Outputs are written to:

outputs/demo_routes/<board>/angle_<angle>/V<grade>/
├── generated_routes.csv
├── generated_route_001.png
├── generated_route_001.svg
├── generated_route_002.png
├── generated_route_002.svg
└── ...

Generated-route visualization

The visualization uses calibrated board backgrounds:

images/tb2_board_12x12_composite.png
images/kilter-original-16x12_composite.png

These are overlaid using product-size coordinate windows:

TB2:    x = [-68, 68],  y = [0, 144]
Kilter: x = [-24, 168], y = [0, 156]

These extents match the old visualization notebooks better than simply using the min/max of observed hold coordinates, because the hold coordinates are inset from the product boundary.

The role markers are:

Role	Marker
start	green circle
middle	blue circle
finish	red star
foot	small yellow square

Annotate holds

To label route holds by placement ID:

python scripts/demo_generate_tb2.py \
  --angle 40 \
  --grade 6 \
  --n 2 \
  --annotate

CPU- friendly run

python scripts/demo_generate_tb2.py \
  --angle 40 \
  --grade 6 \
  --n 2 \
  --torch-threads 1

Temperature and sampling

The --temperature argument controls generation randomness.

The model predicts probabilities for the next token. Temperature rescales those probabilities before sampling.

Temperature	Effect
`0.3`–`0.6`	conservative; picks safer/common tokens
`0.9`	balanced default
`1.0`	samples directly from the learned probabilities
`1.1`–`1.3`	more exploratory; can produce weirder climbs

Example:

python scripts/demo_generate_kilter.py \
  --angle 40 \
  --grade 6 \
  --n 4 \
  --temperature 0.6

Predict grade from board, angle, and frames string

After training the grade predictor, or after placing a trained checkpoint at:

models/joint_transformer_grade_predictor.pth

you can predict a grade directly from a frames string.

Generic

python scripts/demo_predict_grade.py \
  --board tb2 \
  --angle 40 \
  --frames 'p652r5p631r6p322r6p326r7'

TB2 wrapper

python scripts/demo_predict_tb2.py \
  --angle 40 \
  --frames 'p652r5p631r6p322r6p326r7'

Kilter wrapper

python scripts/demo_predict_kilter.py \
  --angle 40 \
  --frames 'p1127r12p1196r13p1216r13p1388r14'

Example output:

Board:        Tension Board 2 Mirror (tb2)
Angle:        40°
Frames:       p652r5p631r6p322r6p326r7
Predicted:    V6
Difficulty:   22.400

The Predicted line is the grouped V-grade. The Difficulty line is the model's continuous prediction on the source difficulty scale.

JSON output

python scripts/demo_predict_grade.py \
  --board kilter \
  --angle 40 \
  --frames 'p1127r12p1196r13p1216r13p1388r14' \
  --json

Show model tokens

python scripts/demo_predict_tb2.py \
  --angle 40 \
  --frames 'p652r5p631r6p322r6p326r7' \
  --show-tokens

Save a visualization of the input climb

python scripts/demo_predict_tb2.py \
  --angle 40 \
  --frames 'p652r5p631r6p322r6p326r7' \
  --visualize

This writes:

outputs/grade_predictions/<board>/angle_<angle>/
├── <name>.png
├── <name>.svg
└── <name>.json

Example with custom output name:

python scripts/demo_predict_kilter.py \
  --angle 40 \
  --frames 'p1127r12p1196r13p1216r13p1388r14' \
  --visualize \
  --output-name my_kilter_climb

Grade prediction in generated-route visualizations

If both checkpoints exist:

models/joint_route_gpt_generator.pth
models/joint_transformer_grade_predictor.pth

then the generation demo automatically scores each generated climb with the grade predictor.

Example:

python scripts/demo_generate_tb2.py --angle 40 --grade 6 --n 4

The terminal output includes something like:

predicted=V5 (difficulty=20.81, error=-1 V)

The visualization subtitle also includes:

predicted V5 (20.81) | error -1V

To disable this scoring:

python scripts/demo_generate_tb2.py \
  --angle 40 \
  --grade 6 \
  --n 4 \
  --no-grade-prediction

To use a non-default grade predictor:

python scripts/demo_generate_and_visualize.py \
  --board kilter \
  --angle 40 \
  --grade 6 \
  --grade-model-path models/joint_transformer_grade_predictor.pth

Important caveats

Generated climbs are machine-generated candidates, not guaranteed to be safe, good, or fun.

The grade predictor is a model-based estimate, not ground truth. Climbing grades are noisy and subjective, and board climbs can be highly style-dependent.

The route sequence is a canonical ordering of holds, not necessarily actual beta order. This is fine for symbolic modeling, but it should not be interpreted as the intended movement sequence.

The visualizations are calibrated to match the existing board images, but any change in image file, crop, or coordinate convention may require adjusting board extents in:

src/climbingboardgpt/visualization.py

Webapp demo

The repository includes a lightweight FastAPI webapp. It is inference-only:

loads the generator and grade predictor once at startup,
serves the TB2/Kilter board images as static assets,
returns hold coordinates and roles as JSON,
draws the climb overlay in the browser as SVG.

Run locally