The evaluation metric for this contest is simple classification accuracy.


## Submission Process

In this competition you will be submitting a TensorFlow Lite model file. The model must take one or more landmark frames as an input and return a float vector (the predicted probabilities of each sign class) as the output. Your model must be packaged into a `submission.zip` file and compatible with the [TensorFlow Lite Runtime v2.9.1](https://www.tensorflow.org/lite/guide/python#run_an_inference_using_tflite_runtime). You are welcome to train your model using the framework of your choice, as long as you convert the model checkpoint into the tflite format prior to submission.

Your model must also perform inference with less than 100 milliseconds of latency per video on average and use less than 40 MB of storage space. Expect to see approximately 40,000 videos in the test set. We allow an additional 10 minute buffer for loading the data and miscellaneous overhead.


Each video is loaded with the following function:

```
ROWS_PER_FRAME = 543  # number of landmarks per frame

def load_relevant_data_subset(pq_path):
    data_columns = ['x', 'y', 'z']
    data = pd.read_parquet(pq_path, columns=data_columns)
    n_frames = int(len(data) / ROWS_PER_FRAME)
    data = data.values.reshape(n_frames, ROWS_PER_FRAME, len(data_columns))
    return data.astype(np.float32)
```

Inference is performed (roughly) as follows, ignoring details like how we manage multiple videos:
```
import tflite_runtime.interpreter as tflite
interpreter = tflite.Interpreter(model_path)

found_signatures = list(interpreter.get_signature_list().keys())

if REQUIRED_SIGNATURE not in found_signatures:
    raise KernelEvalException('Required input signature not found.')

prediction_fn = interpreter.get_signature_runner("serving_default")
output = prediction_fn(inputs=frames)
sign = np.argmax(output["outputs"])
```