sample_inference_script.py

from huggingface_hub import hf_hub_download
import pandas as pd
import numpy as np
import ai_edge_litert as litert

# Download model
repo_id = "sammlapp/perch2-tflite"
model_path = hf_hub_download(repo_id=repo_id, filename="Perch2.tflite")

# Download and read labels
labels_path = hf_hub_download(repo_id=repo_id, filename="perch2_class_labels.txt")
labels = pd.read_csv(labels_path, header=None).iloc[:, 0].tolist()

print("Model:", model_path)
print("Labels:", len(labels))

# path = "./Perch2.tflite"
interpreter = litert.interpreter.Interpreter(model_path=model_path)
interpreter.allocate_tensors()

sig = interpreter.get_signature_runner("serving_default")

input_details = interpreter.get_input_details()
signature_list = interpreter.get_signature_list()

input_name = signature_list["serving_default"]["inputs"][0]
output_names = signature_list["serving_default"]["outputs"]
input_shape = input_details[0]["shape"]
input_dtype = input_details[0]["dtype"]

sample_input = np.random.uniform(-1, 1, size=(10, 160000)).astype(input_dtype)

# Run via signature
result = sig(**{input_name: sample_input})

# Each output key in the signature dict will hold a numpy array (safe to access)
print("Output keys:", result.keys())
for k, v in result.items():
    print(k, v.shape)