Add application file

.gitignore

@@ -0,0 +1,166 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+
+
+# Below are files that are specific to this project
+flagged/
+.vscode/
+.DS_Store

app.py

@@ -0,0 +1,80 @@
+from typing import Dict
+
+import gradio as gr
+import whisper
+from whisper.tokenizer import get_tokenizer
+
+import classify
+
+
+def zero_shot_classify(audio_path: str, class_names: str, model_name: str) -> Dict[str, float]:
+    class_names = class_names.split(",")
+    tokenizer = get_tokenizer(multilingual=".en" not in model_name)
+    model = whisper.load_model(model_name)
+
+    internal_lm_average_logprobs = classify.calculate_internal_lm_average_logprobs(
+        model=model,
+        class_names=class_names,
+        tokenizer=tokenizer,
+    )
+    audio_features = classify.calculate_audio_features(audio_path, model)
+    average_logprobs = classify.calculate_average_logprobs(
+        model=model,
+        audio_features=audio_features,
+        class_names=class_names,
+        tokenizer=tokenizer,
+    )
+    average_logprobs -= internal_lm_average_logprobs
+    scores = average_logprobs.softmax(-1).tolist()
+    return {class_name: score for class_name, score in zip(class_names, scores)}
+
+
+def main():
+    CLASS_NAMES = "[dog barking],[helicopter whirring],[laughing],[birds chirping],[clock ticking]"
+    AUDIO_PATHS = [
+        "./data/(dog)1-100032-A-0.wav",
+        "./data/(helicopter)1-181071-A-40.wav",
+        "./data/(laughing)1-1791-A-26.wav",
+        "./data/(chirping_birds)1-34495-A-14.wav",
+        "./data/(clock_tick)1-21934-A-38.wav",
+    ]
+    EXAMPLES = []
+    for audio_path in AUDIO_PATHS:
+        EXAMPLES.append([audio_path, CLASS_NAMES, "small"])
+
+    DESCRIPTION = """
+        <div style="text-align: center;">
+            <p>
+                This demo allows you to try out zero-shot audio classification using
+                [Whisper](https://github.com/openai/whisper).
+            </p>
+            <p>Github: [https://github.com/jumon/zac](https://github.com/jumon/zac)</p>
+            <p>
+                Example audio files are from the [ESC-50](https://github.com/karolpiczak/ESC-50)
+                dataset (CC BY-NC 3.0).
+            </p>
+        </div>
+    """
+
+    demo = gr.Interface(
+        fn=zero_shot_classify,
+        inputs=[
+            gr.Audio(source="upload", type="filepath", label="Audio File"),
+            gr.Textbox(lines=1, label="Candidate class names (comma-separated)"),
+            gr.Radio(
+                choices=["tiny", "base", "small", "medium", "large"],
+                value="small",
+                label="Model Name",
+            ),
+        ],
+        outputs="label",
+        examples=EXAMPLES,
+        title="Zero-shot Audio Classification using Whisper",
+        description=DESCRIPTION,
+    )
+
+    demo.launch()
+
+
+if __name__ == "__main__":
+    main()

classify.py

@@ -0,0 +1,66 @@
+from typing import List, Optional
+
+import torch
+import torch.nn.functional as F
+from whisper.audio import N_FRAMES, N_MELS, log_mel_spectrogram, pad_or_trim
+from whisper.model import Whisper
+from whisper.tokenizer import Tokenizer
+
+
+@torch.no_grad()
+def calculate_audio_features(audio_path: Optional[str], model: Whisper) -> torch.Tensor:
+    if audio_path is None:
+        segment = torch.zeros((N_MELS, N_FRAMES), dtype=torch.float32).to(model.device)
+    else:
+        mel = log_mel_spectrogram(audio_path)
+        segment = pad_or_trim(mel, N_FRAMES).to(model.device)
+    return model.embed_audio(segment.unsqueeze(0))
+
+
+@torch.no_grad()
+def calculate_average_logprobs(
+    model: Whisper,
+    audio_features: torch.Tensor,
+    class_names: List[str],
+    tokenizer: Tokenizer,
+) -> torch.Tensor:
+    initial_tokens = (
+        torch.tensor(tokenizer.sot_sequence_including_notimestamps).unsqueeze(0).to(model.device)
+    )
+    eot_token = torch.tensor([tokenizer.eot]).unsqueeze(0).to(model.device)
+
+    average_logprobs = torch.zeros(len(class_names))
+    for i, class_name in enumerate(class_names):
+        class_name_tokens = (
+            torch.tensor(tokenizer.encode(" " + class_name)).unsqueeze(0).to(model.device)
+        )
+        input_tokens = torch.cat([initial_tokens, class_name_tokens, eot_token], dim=1)
+
+        logits = model.logits(input_tokens, audio_features)  # (1, T, V)
+        logprobs = F.log_softmax(logits, dim=-1).squeeze(0)  # (T, V)
+        logprobs = logprobs[len(tokenizer.sot_sequence_including_notimestamps) - 1 : -1]  # (T', V)
+        logprobs = torch.gather(logprobs, dim=-1, index=class_name_tokens.view(-1, 1))  # (T', 1)
+        average_logprob = logprobs.mean().item()
+        average_logprobs[i] = average_logprob
+
+    return average_logprobs
+
+
+def calculate_internal_lm_average_logprobs(
+    model: Whisper,
+    class_names: List[str],
+    tokenizer: Tokenizer,
+    verbose: bool = False,
+) -> torch.Tensor:
+    audio_features_from_empty_input = calculate_audio_features(None, model)
+    average_logprobs = calculate_average_logprobs(
+        model=model,
+        audio_features=audio_features_from_empty_input,
+        class_names=class_names,
+        tokenizer=tokenizer,
+    )
+    if verbose:
+        print("Internal LM average log probabilities for each class:")
+        for i, class_name in enumerate(class_names):
+            print(f"  {class_name}: {average_logprobs[i]:.3f}")
+    return average_logprobs

packages.txt

@@ -0,0 +1 @@
+ffmpeg

requirements.txt

@@ -0,0 +1,2 @@
+git+https://github.com/openai/whisper.git
+tqdm