export_onnx.py

"""
Export BSG-BAT PyTorch checkpoints to ONNX.

The model (code/supervised.py `Net`) is a plain 6-conv + 3-fc CNN.
Input  : float32 [batch, 1, 512, 128]  (1 channel, ntime=512, nfreq=128)
         a log10 mel spectrogram of 384 kHz mono audio
         (n_fft=1024, hop=768, n_mels=128, fmin=9000, fmax=150000), per-segment normalized.
Output : float32 [batch, 22]  logits (21 European bat species + Background).
         probability = sigmoid(logit)  (multi-label, BCEWithLogits training).

Usage: python export_onnx.py <model_in.pt> <model_out.onnx>
"""
import sys
import torch
import torch.nn as nn
import torch.nn.functional as F

NTIME = 512
NFREQ = 128
NCLASSES = 22


class Net(nn.Module):
    # Copied verbatim from bsgbat/code/supervised.py so we do not depend on its sys.path.
    def __init__(self, ntime=NTIME, nfreq=NFREQ, nclasses=NCLASSES):
        super().__init__()
        self.conv1 = nn.Conv2d(1, 32, 3, padding=1)
        self.conv2 = nn.Conv2d(32, 64, 3, padding=1)
        self.conv3 = nn.Conv2d(64, 128, 3, padding=1)
        self.conv4 = nn.Conv2d(128, 256, 3, padding=1)
        self.conv5 = nn.Conv2d(256, 512, 3, padding=1)
        self.conv6 = nn.Conv2d(512, 512, 3, padding=1)
        n_maxpool = 6
        nt, nr = ntime, nfreq
        for _ in range(n_maxpool):
            nt //= 2
            nr //= 2
        nr *= 4
        self.fc1 = nn.Linear(512 * nt * nr, 512)
        self.fc2 = nn.Linear(512, 128)
        self.fc3 = nn.Linear(128, nclasses)

    def forward(self, x):
        x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
        x = F.max_pool2d(F.relu(self.conv2(x)), 2)
        x = F.max_pool2d(F.relu(self.conv3(x)), 2)
        x = F.max_pool2d(F.relu(self.conv4(x)), (2, 2))
        x = F.max_pool2d(F.relu(self.conv5(x)), (2, 1))
        x = F.max_pool2d(F.relu(self.conv6(x)), (2, 1))
        x = torch.flatten(x, 1)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x


def main(argv):
    model_in, model_out = argv[0], argv[1]
    device = torch.device("cpu")
    model = Net(nclasses=NCLASSES)
    state = torch.load(model_in, map_location=device)
    model.load_state_dict(state)
    model.eval()

    dummy = torch.randn(1, 1, NTIME, NFREQ, dtype=torch.float32)

    torch.onnx.export(
        model,
        dummy,
        model_out,
        input_names=["spectrogram"],
        output_names=["logits"],
        dynamic_axes={"spectrogram": {0: "batch"}, "logits": {0: "batch"}},
        opset_version=17,
        do_constant_folding=True,
    )
    print(f"exported {model_in} -> {model_out}")

    # Numerical parity check PyTorch vs onnxruntime.
    import numpy as np
    import onnxruntime as ort

    with torch.no_grad():
        ref = model(dummy).numpy()
    sess = ort.InferenceSession(model_out, providers=["CPUExecutionProvider"])
    got = sess.run(["logits"], {"spectrogram": dummy.numpy()})[0]
    max_abs = float(np.max(np.abs(ref - got)))
    print(f"output shape {got.shape}, max abs diff PyTorch vs ORT = {max_abs:.3e}")
    assert got.shape == (1, NCLASSES), got.shape
    assert max_abs < 1e-4, max_abs
    print("PARITY_OK")


if __name__ == "__main__":
    main(sys.argv[1:])