How to export a fixed shape ONNX model?

[wangjl1993]

Great work!
I would like to export an ONNX model with a fixed shape (10 seconds input). However, I noticed that the input and output dimensions of the cache_decoder increase with each decoding step. Is there any way to solve this issue? My code like this:

```python
model_name = "moonshine/tiny"
model = moonshine.load_model(model_name)
output_dir = "onnx_models"
Path(output_dir).mkdir(parents=True, exist_ok=True)
convert_and_store(
    model.preprocessor.preprocess,
    input_signature=[tf.TensorSpec([1, 16_0000], dtype=tf.float32)],  # batch=1, seconds=10s
    output_file=f"{output_dir}/preprocess.onnx",
)

seq_len_spec = tf.TensorSpec([1], dtype=tf.int32)

convert_and_store(
    model.encoder.encoder,
    input_signature=[
        tf.TensorSpec([1, 415, model.dim], dtype=tf.float32),
        seq_len_spec,
    ],
    output_file=f"{output_dir}/encode.onnx",
)

input_spec = tf.TensorSpec([1, 1], dtype=tf.int32)
context_spec = tf.TensorSpec([1, 415, model.dim], dtype=tf.float32)
cache_spec = model.dec_n_layers * [
    tf.TensorSpec([1, 1, model.n_head, model.inner_dim // model.n_head], dtype=tf.float32,),
    tf.TensorSpec([1, 1, model.n_head, model.inner_dim // model.n_head], dtype=tf.float32,),
    tf.TensorSpec([1, 415, model.n_head, model.inner_dim // model.n_head],dtype=tf.float32),
    tf.TensorSpec([1, 415, model.n_head, model.inner_dim // model.n_head],dtype=tf.float32)
] 
print(model.dim, model.n_head, model.dec_n_layers, model.inner_dim, model.dec_n_layers)
convert_and_store(
    model.decoder.uncached_call,
    input_signature=[input_spec, context_spec, seq_len_spec],
    output_file=f"{output_dir}/uncached_decode.onnx",
)

convert_and_store(
    model.decoder.cached_call,
    input_signature=[input_spec, context_spec, seq_len_spec] + cache_spec,
    output_file=f"{output_dir}/cached_decode.onnx",
)

```

Thanks!