Hello,
I've experienced these two issues when using crepe and spleeter libraries together, specifically when trying to call crepe.predict method after calling separate method of spleeter.
Here is a code in Google Colab:

!pip install crepe
!pip install spleeter

from spleeter.separator import Separator
from spleeter.audio.adapter import AudioAdapter
separator = Separator('spleeter:4stems')
audio_loader = AudioAdapter.default()
srate = 44100
audio1, _ = audio_loader.load('audio1.wav', sample_rate = srate)
audio2, _ = audio_loader.load('audio2.wav', sample_rate = srate)
separation = separator.separate(audio1)

import crepe
times1, freqs1, confidence1, activation1 = crepe.predict(audio1, srate, viterbi=True)

The output is following:

---

RuntimeError Traceback (most recent call last)

...

----> 2 times1, freqs1, confidence1, activation1 = crepe.predict(audio1, srate, viterbi=True)

4 frames

/usr/local/lib/python3.10/dist-packages/tensorflow/python/framework/ops.py in _check_not_finalized(self)
3321 """
3322 if self._finalized:
-> 3323 raise RuntimeError("Graph is finalized and cannot be modified.")
3324
3325 def _add_op(self, op, op_name):

RuntimeError: Graph is finalized and cannot be modified.

To avoid this, we can write the following commands instead of the last string:

import tensorflow as tf
with tf.Graph().as_default():
    times1, freqs1, confidence1, activation1 = crepe.predict(audio1, srate, viterbi=True)

In this case the prediction will be performed without errors.

But if we try to do another prediction after that, for example:

with tf.Graph().as_default():
    times2, freqs2, confidence2, activation2 = crepe.predict(audio2, srate, viterbi=True)

then we'll get the following error:

---

InvalidArgumentError Traceback (most recent call last)

in <cell line: 1>()
1 with tf.Graph().as_default():
----> 2 times2, freqs2, confidence2, activation2 = crepe.predict(audio2, srate, viterbi=True)

8 frames

/usr/local/lib/python3.10/dist-packages/tensorflow/python/client/session.py in init(self, session, callable_options)
1469 compat.as_bytes(callable_options.SerializeToString()))
1470 try:
-> 1471 self._handle = tf_session.TF_SessionMakeCallable(
1472 session._session, options_ptr)
1473 finally:

InvalidArgumentError: Tensor input:0, specified in either feed_devices or fetch_devices was not found in the Graph

To avoid this, we must call predict method for both files inside the same "with" section:

with tf.Graph().as_default():
    times1, freqs1, confidence1, activation1 = crepe.predict(audio1, srate, viterbi=True)
    times2, freqs2, confidence2, activation2 = crepe.predict(audio2, srate, viterbi=True)

But sometimes it is impossible.

The following versions were used:

• python 3.10
• crepe 0.0.15
• spleeter 2.4.0
• tensorflow 2.9.3