Judo Clipper: Baseline Model

← Back to the Judo Clipper case study

Brief Summary of Model

The model is an LSTM with a feed-forward classification head. Its purpose is to determine whether a given input clip of a judo match contains at least one throw attempt.

The LSTM receives a sequence of 210 concatenated player pose vectors, each containing 68 features, with one vector for each frame in a 7-second clip. It produces one raw logit representing whether the complete clip contains a throw attempt. During evaluation or inference, the logit is converted into a probability and then into a binary prediction: throw attempt = 1 and no throw attempt = 0.

Architecture and Configurations

The baseline model consists of a two-layer LSTM followed by a feed-forward classification head. The classification head projects the LSTM’s final hidden-state representation into one raw logit for the complete clip.

Hyperparameter	Value	Description
LSTM Hidden Size	128	Balances representational capacity and overfitting
LSTM Layers	2	Captures hierarchical temporal features
Bidirectional	False	Processes sequences strictly forward in time
Classifier Hidden Size	64	Intermediate projection before final logits
Dropout Rate	0.30	Regularization for dense classification layers

Training Configurations

Parameter	Value	Details
Epochs	50	Total passes over the training data
Batch Size	32	Number of sequences per batch
Learning Rate	0.001	Base step size for the optimizer
Weight Decay	0.0001	L2 regularization to penalize large weights

Dataset Split

The dataset was divided into stratified training, validation, and test splits using scikit-learn with a random seed of 42. The resulting split assignments were frozen and persisted in splits/dataset_v1_stratified_seed_42.csv so that every experiment uses exactly the same samples.

The proportions were:

Split	Fraction (%)	Manifest File / Strategy	Details / Purpose
Train	80%	splits/dataset_v1_stratified_seed_42.csv (Seed: 42)	Model training and parameter updates
Validation	10%	splits/dataset_v1_stratified_seed_42.csv (Seed: 42)	Hyperparameter tuning & threshold calibration
Test	10%	splits/dataset_v1_stratified_seed_42.csv (Seed: 42)	Unbiased final performance evaluation

Given that the dataset contained 2,163 clips, the raw counts of the split were:

Class	Training	Validation	Test	Total
No throw	1,112	139	139	1,390
Throw attempt	618	77	78	773
Overall total	1,730	216	217	2,163

Results

Results of Training

Diagram of Training Losses

Results of Evaluation Using the Validation Dataset Split

Evaluation Policy

The model returns raw logits. During evaluation, sigmoid converts logits into probabilities. A threshold of 0.50 was used as the initial reference, after which the validation threshold was selected by maximising attempt F1 according to the predefined tie-breaking policy. This selected a threshold of 0.53.

Results

Model & Checkpoint Configuration

Parameter	Value
Split	validation
Checkpoint	best
Checkpoint Epoch	42
Checkpoint Validation Loss	0.4941
Classification Threshold	0.5300 (validation-selected)

Dataset Counts

Metric	Count
Total Samples	216
Actual Attempts	77
Actual No-Attempts	139

Confusion Matrix

Breakdown

Metric	Count
True Positives (TP)	61
True Negatives (TN)	105
False Positives (FP)	34
False Negatives (FN)	16

2×2 Matrix View

	Predicted: Attempt	Predicted: No-Attempt
Actual: Attempt	61 (TP)	16 (FN)
Actual: No-Attempt	34 (FP)	105 (TN)

Classification Metrics

Overall Performance

• Accuracy: 0.7685 (76.85%)
• Macro F1: 0.7585

Per-Class Metrics

Class	Precision	Recall	F1-Score
Attempt	0.6421	0.7922	0.7093
No-Attempt	0.8678	0.7554	0.8077

Analysis of Results

Overall, the validation metrics and the shape of the training and validation loss curves indicate that the LSTM provides a credible baseline, although controlled architecture and hyperparameter experiments may improve its performance.

Analysis of Training Curve

There are 3 main points of interest regarding the curve of training losses:

• The losses showed little apparent improvement during approximately the first 20–23 epochs, after which optimisation began producing clearer reductions.
• The erratic spike after the 42nd epoch
• The model doesn’t seem to be overfitting (the training and validation curves are moving together)

These findings could suggest that the model may be suffering from exploding gradients.

Analysis of Evaluation on Validation Data

The false positives are the main concern given the intended use for this model and 34 is slightly higher than desired for a baseline. However, the true positive count, true negative count, and macro F1 score are promising.

The F1 score for the no attempt category being slightly higher than for the attempt category was expected and isn’t currently of concern.

Next Steps

There are 2 relatively obvious next steps:

• Implementing gradient clipping to address the potential exploding gradient issue.
• Trying a bidirectional LSTM (as the model is processing the entire clip in post anyway)

These should both be tested as their own experiments and evaluated against the baseline.