Unable to reproduce zeroshot classification results

Overview

I have attempted to reproduce the zeroshot classification results for ESC-50 outlined in the publication Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation.

In the paper, zeroshot classification accuracy (top-1) for the best model (K2C aug) is reported at 91.0%. I assume that this is the 630k-audioset-best.pt checkpoint.

• I am only able to report 60.2% top-1 accuracy for the ESC-50 dataset
• ESC-50 was downloaded from the google drive folder.
• I've found it quite difficult to follow the given evaluation and data preprocessing code, so I wrote my own.

Reproduce

I use the set of 50 unique captions in the test dataset, which are found in the text attr of each example's json file, e.g. "The sound of the crow".

Here's the loader for ESC-50:

class ESC50Dataset(Dataset):
    def __init__(
        self,
        path_to_esc50="./data/ESC50",
        split="test",
        audio_len = 480000,
    ):
        super().__init__()
        self.data_path = Path(path_to_esc50)
        self.audio_len = audio_len
       
        self.audio_files = sorted(glob.glob(str(self.data_path / split / "*.flac")))
        self.meta_files = sorted(glob.glob(str(self.data_path / split / "*.json")))
        assert len(self.audio_files) == len(self.meta_files), "Number of audio files and meta files must match"
        assert [osp.splitext(osp.basename(x))[0] for x in self.audio_files] == [osp.splitext(osp.basename(x))[0] for x in self.meta_files], "Audio files and meta files must have the same names"

        self.tags = []
        self.texts = []
        for f in self.meta_files:
            with open(f, 'r') as json_file:
                data = json.load(json_file)
                self.tags.append(data["tag"][0])
                self.texts.append(data["text"][0])

    def __getitem__(self, idx):
        x, _ = load_audio_torch(self.audio_files[idx], target_sr=48000, mono=True)
        x = random_slice(x, self.audio_len)
        return x, self.texts[idx]

    def __len__(self):
        return len(self.audio_files)

And the zeroshot retrieval script:

import os

import torch
import laion_clap

from data.loaders import ESC50Dataset

ckpt_path = "CLAP_checkpoints/laion_clap/"
model_params = {"ckpt": "630k-audioset-best.pt", "amodel": "HTSAT-tiny"}
model = laion_clap.CLAP_Module(enable_fusion=False, amodel=model_params["amodel"])
model.load_ckpt(os.path.join(ckpt_path, model_params["ckpt"]))

dataset = ESC50Dataset()
texts = list(set(dataset.texts)) # get the unique texts, e.g "The sound of the crow"

# get the text embeddings for each tag
z_text = torch.cat([torch.tensor(model.get_text_embedding([t, t])[0:1]) for t in texts])

z_audio = []
text_idxs = []
for item in dataset:
    x, text = item
    idx = texts.index(text) # get the index of this example's text
    text_idxs.append(idx)
    z_audio.append(torch.tensor(model.get_audio_embedding_from_data(x.numpy()))) # get its CLAP audio embedding
z_audio = torch.cat(z_audio)
sim = model.model.logit_scale_a.cpu() * z_audio @ z_text.T # compute pairwise dot products

# top-1 accuracy
acc = float(torch.sum(torch.argmax(sim, dim=1) == torch.tensor(text_idxs)) / len(sim))
print(f"Accuracy: {acc}")

Accuracy: 0.6025000214576721 

Hopefully I am missing something significant?

Start by running the provided ESC-50 reproduction using data.loaders.ESC50Dataset, the 630k-audioset-best.pt checkpoint, and the listed embedding script. Compare its preprocessing, captions, checkpoint settings, and reported accuracy with the publication and repository evaluation path. Done means identifying and documenting the cause of the 60.2% versus 91.0% discrepancy.