Optional evaluation/plotting dependencies are imported globally and can break inference setup

[tapstoop]

Some dependencies appear to be required only for evaluation metrics or plotting, but they are imported globally by modules used during inference.

This makes simple single-file inference fail unless optional packages are installed and pinned correctly.

Examples

ssr_eval

ssr_eval is used for evaluation metrics:

metrics = ssr_eval.metrics.AudioMetrics(rate=self.sampling_rate)

but it is imported globally:

import ssr_eval

This forces users to install ssr_eval even if they only want to run inference and do not need benchmark metrics.

moviepy

plotting_utils.py imports:

from moviepy.video.io.bindings import mplfig_to_npimage

This is required for plotting / logging utilities, but it can break inference if the installed MoviePy version is 2.x.

Suggested improvement

Use lazy imports for optional functionality.

For example:

def compute_eval_metrics(...):
    import ssr_eval
    ...

and:

def plot_spec_to_numpy(...):
    from moviepy.video.io.bindings import mplfig_to_npimage
    ...

This would allow simple inference to run with a smaller dependency set, simplify the installation and have less friction for users who only want audio restoration.