Model prediction

This notebook downloads an example baseline model. All baseline models are available here although they are beta models and not recommended for research use.

# suppress warnings
import warnings
warnings.simplefilter('ignore')

#import modules from Opensoundscape
from opensoundscape.torch.predict import predict
from opensoundscape.datasets import SingleTargetAudioDataset
from opensoundscape.helpers import run_command
from opensoundscape.datasets import SplitterDataset
from opensoundscape.raven import lowercase_annotations

import torch
import torch.nn
import torchvision.models
import torch.utils.data

import yaml
import os.path
import pandas as pd
from pathlib import Path
from math import floor

Prepare model

Download model

Download the example model for Wood Thrush, Hylocichla mustelina.

def download_from_box(link, name):
    run_command(f"curl -L {link} -o ./{name}")

folder_name = "prediction_example"
folder_path = Path(folder_name)
if not folder_path.exists(): folder_path.mkdir()
model_filename = folder_path.joinpath("hylocichla-mustelina-epoch-4.model")
download_from_box(
    link = "https://pitt.box.com/shared/static/dslgslmag7y8ojqxv28mwhbnt7irpgeo.model",
    name = model_filename
)

Load model

The model must be loaded with the same specifications that it was created with: a combination of a resnet18 convolutional neural network and a Linear classifier. This model predicts two “classes”: the presence and absence of Wood Thrush.

num_classes = 2
model = torchvision.models.resnet18(pretrained=False)
model.fc = torch.nn.Linear(model.fc.in_features, num_classes)
model.load_state_dict(torch.load(model_filename))
#model.load_state_dict(torch.load("scolopax-minor-epoch-4.model"))

<All keys matched successfully>

Prepare prediction files

Download an example soundscape which contains Wood Thrush vocalizations.

Download data

data_filename = folder_path.joinpath("1min.wav")
download_from_box(
    link = "https://pitt.box.com/shared/static/z73eked7quh1t2pp93axzrrpq6wwydx0.wav",
    name = data_filename
)

Split data

The example soundscape must be split up into soundscapes of the same size as the ones the model was trained on. In this case, the soundscapes should be 5s long.

files_to_split = [data_filename]
split_directory = folder_path.joinpath("split_files")
if not split_directory.exists(): split_directory.mkdir()
dataset = SplitterDataset(
    files_to_split,
    overlap=0,
    duration=5,
    output_directory=split_directory,
    include_last_segment=True
)

dataloader = torch.utils.data.DataLoader(
    dataset,
    batch_size=1,
    shuffle=False,
    collate_fn=SplitterDataset.collate_fn,
)

results_csv = folder_path.joinpath("prediction_files.csv")
with open(results_csv, "w") as f:
    if False:
        f.write("Source,Annotations,Begin (s),End (s),Destination,Labels\n")
    else:
        f.write("Source,Begin (s),End (s),Destination\n")
    for idx, data in enumerate(dataloader):
        for output in data:
            f.write(f"{output}\n")

Create a Dataset

Create a dataset from these data. We create a dictionary that associates numeric labels with the class names: 1 is for predicting a Wood Thrush’s presence; 0 is for predicting a Wood Thrush’s absence.

files_to_analyze=list(split_directory.glob("*.wav"))
sample_df = pd.DataFrame(columns=['file'],data=files_to_analyze)

label_dict = {0:'absent', 1:'hylocichla-mustelina'}
test_dataset = SingleTargetAudioDataset(
    sample_df,
    filename_column = "file",
    label_dict = label_dict
)

Use model on prediction files

model.eval()
prediction_df = predict(model, test_dataset, label_dict=label_dict)
prediction_df

	absent	hylocichla-mustelina
prediction_example/split_files/bc645003351149f4a7e2c7109b22afc1.wav	0.816133	-0.903320
prediction_example/split_files/e36a0f200cdf42a23d49e78445121387.wav	1.480433	-0.927409
prediction_example/split_files/4940c91a1837410240042cf55ccad568.wav	1.940377	-1.725088
prediction_example/split_files/cfc05bd9e1b97eebdca3badc288de0cd.wav	2.629047	-1.988923
prediction_example/split_files/32747f95e81ee34c56ed177c4f7e7df5.wav	2.513747	-2.366485
prediction_example/split_files/369134205221b5a25fac0e264d0a1482.wav	2.351259	-1.628652
prediction_example/split_files/f3d6aeabe7725f649dc56d6db04aa83f.wav	1.570931	-1.124706
prediction_example/split_files/54534197c0768b6bb2a9305013e8c1af.wav	1.744635	-1.055664
prediction_example/split_files/e0c2d4aed1d79d4a6194be948d3292da.wav	1.315882	-1.407135
prediction_example/split_files/9d276a5dd54b631c4aa63da407a1225d.wav	1.766514	-1.096341
prediction_example/split_files/636f23557581b700f286b7db29d01b61.wav	0.273381	-0.397208
prediction_example/split_files/e55ba1b5a1316fcda5f3b4d73b2e36ee.wav	2.138355	-1.632506

This command “cleans up” by deleting all the downloaded files and results.

import shutil
shutil.rmtree(folder_path)