How do I run the bytecover code correctly on a local dataset?

[cordutie]

I've been trying test the repo on a small local dataset and downloaded the weights. (sidenote: It's odd that the size of the downloaded weights was shown as 10000 x 8XXX which doesn't match the suggested weights size in the documentation)

Attached is the code we are trying to run. Any ideas on what is going wrong? There's a lot of guesswork involved in the process..

file_path = 'orfium-bytecover.pt'

Load the model

loaded_weights = torch.load(file_path, map_location=device)

Instantiate the model with the same parameters as in the loaded weights

model = Resnet50(
Bottleneck,
emb_dim=loaded_weights['fc.weight'].shape[1],
num_channels=1,
num_classes=loaded_weights['fc.weight'].shape[0],
sr=22050,
hop_lenght=512,
n_bins=84,
bins_per_octave=12,
window="hann",
compress_ratio=20,
tempo_factors=None,
)

weights loading

model.load_state_dict(loaded_weights)

#audio files into a batch to use forward pass
audio_data, sr_original = librosa.load("songs/videogames_1.mp3", sr=None)
x_1 = librosa.resample(y=audio_data, orig_sr=sr_original, target_sr=22050)[0:220503]
audio_data, sr_original = librosa.load("songs/videogames_2.mp3", sr=None)
x_2 = librosa.resample(y=audio_data, orig_sr=sr_original, target_sr=22050)[0:220503]
x_batch = torch.stack([torch.tensor(x_1), torch.tensor(x_2)], dim=0)

model is applied

y_batch = model(x_batch)

similarity computation

f_c_batch = y_batch['f_c'].detach().numpy()
y_1 = f_c_batch[0,:]
y_2 = f_c_batch[1,:]
print(np.linalg.norm(y_1- y_2)) # <--- this number is way too big and cosine similarity is not better