rollback

Dockerfile

@@ -1,30 +0,0 @@
-FROM ubuntu
-
-# Based on https://huggingface.co/spaces/radames/nginx-gradio-reverse-proxy/blob/main/Dockerfile
-USER root
-
-RUN apt-get -y update && apt-get -y install nginx
-RUN mkdir -p /var/cache/nginx \
-             /var/log/nginx \
-             /var/lib/nginx
-RUN touch /var/run/nginx.pid
-
-RUN chown -R 1000:1000 /var/cache/nginx \
-                       /var/log/nginx \
-                       /var/lib/nginx \
-                       /var/run/nginx.pid
-
-RUN useradd -m -u 1000 user
-
-USER user
-ENV HOME=/home/user
-
-RUN mkdir $HOME/app
-WORKDIR $HOME/app
-
-# Copy nginx configuration
-COPY --chown=user nginx.conf /etc/nginx/sites-available/default
-COPY --chown=user . .
-
-CMD ["bash", "run.sh"]
-

README.md

@@ -1,10 +1,9 @@
----
 title: Whisper JAX
 emoji: ⚡️
 colorFrom: yellow
 colorTo: indigo
-sdk: docker
+sdk: gradio
+sdk_version: 3.27.0
+app_file: app.py
 pinned: false
----
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,246 @@
+import base64
+import math
+import os
+import time
+from multiprocessing import Pool
+
+import gradio as gr
+import numpy as np
+import pytube
+import requests
+from processing_whisper import WhisperPrePostProcessor
+from transformers.models.whisper.tokenization_whisper import TO_LANGUAGE_CODE
+from transformers.pipelines.audio_utils import ffmpeg_read
+
+
+title = "Whisper JAX: The Fastest Whisper API ⚡️"
+
+description = """Whisper JAX is an optimised implementation of the [Whisper model](https://huggingface.co/openai/whisper-large-v2) by OpenAI. It runs on JAX with a TPU v4-8 in the backend. Compared to PyTorch on an A100 GPU, it is over [**70x faster**](https://github.com/sanchit-gandhi/whisper-jax#benchmarks), making it the fastest Whisper API available.
+
+Note that using microphone or audio file requires the audio input to be transferred from the Gradio demo to the TPU, which for large audio files can be slow. We recommend using YouTube where possible, since this directly downloads the audio file to the TPU, skipping the file transfer step.
+"""
+
+API_URL = os.getenv("API_URL")
+API_URL_FROM_FEATURES = os.getenv("API_URL_FROM_FEATURES")
+
+article = "Whisper large-v2 model by OpenAI. Backend running JAX on a TPU v4-8 through the generous support of the [TRC](https://sites.research.google/trc/about/) programme. Whisper JAX [code](https://github.com/sanchit-gandhi/whisper-jax) and Gradio demo by 🤗 Hugging Face."
+
+language_names = sorted(TO_LANGUAGE_CODE.keys())
+CHUNK_LENGTH_S = 30
+BATCH_SIZE = 16
+NUM_PROC = 16
+FILE_LIMIT_MB = 1000
+
+
+def query(payload):
+    response = requests.post(API_URL, json=payload)
+    return response.json(), response.status_code
+
+
+def inference(inputs, task=None, return_timestamps=False):
+    payload = {"inputs": inputs, "task": task, "return_timestamps": return_timestamps}
+
+    data, status_code = query(payload)
+
+    if status_code == 200:
+        text = data["text"]
+    else:
+        text = data["detail"]
+
+    timestamps = data.get("chunks")
+    if timestamps is not None:
+        timestamps = [
+            f"[{format_timestamp(chunk['timestamp'][0])} -> {format_timestamp(chunk['timestamp'][1])}] {chunk['text']}"
+            for chunk in timestamps
+        ]
+        text = "\n".join(str(feature) for feature in timestamps)
+    return text
+
+
+def chunked_query(payload):
+    response = requests.post(API_URL_FROM_FEATURES, json=payload)
+    return response.json()
+
+
+def forward(batch, task=None, return_timestamps=False):
+    feature_shape = batch["input_features"].shape
+    batch["input_features"] = base64.b64encode(batch["input_features"].tobytes()).decode()
+    outputs = chunked_query(
+        {"batch": batch, "task": task, "return_timestamps": return_timestamps, "feature_shape": feature_shape}
+    )
+    outputs["tokens"] = np.asarray(outputs["tokens"])
+    return outputs
+
+
+def identity(batch):
+    return batch
+
+
+# Copied from https://github.com/openai/whisper/blob/c09a7ae299c4c34c5839a76380ae407e7d785914/whisper/utils.py#L50
+def format_timestamp(seconds: float, always_include_hours: bool = False, decimal_marker: str = "."):
+    if seconds is not None:
+        milliseconds = round(seconds * 1000.0)
+
+        hours = milliseconds // 3_600_000
+        milliseconds -= hours * 3_600_000
+
+        minutes = milliseconds // 60_000
+        milliseconds -= minutes * 60_000
+
+        seconds = milliseconds // 1_000
+        milliseconds -= seconds * 1_000
+
+        hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
+        return f"{hours_marker}{minutes:02d}:{seconds:02d}{decimal_marker}{milliseconds:03d}"
+    else:
+        # we have a malformed timestamp so just return it as is
+        return seconds
+
+
+if __name__ == "__main__":
+    processor = WhisperPrePostProcessor.from_pretrained("openai/whisper-large-v2")
+    stride_length_s = CHUNK_LENGTH_S / 6
+    chunk_len = round(CHUNK_LENGTH_S * processor.feature_extractor.sampling_rate)
+    stride_left = stride_right = round(stride_length_s * processor.feature_extractor.sampling_rate)
+    step = chunk_len - stride_left - stride_right
+    pool = Pool(NUM_PROC)
+
+    def tqdm_generate(inputs: dict, task: str, return_timestamps: bool, progress: gr.Progress):
+        inputs_len = inputs["array"].shape[0]
+        all_chunk_start_idx = np.arange(0, inputs_len, step)
+        num_samples = len(all_chunk_start_idx)
+        num_batches = math.ceil(num_samples / BATCH_SIZE)
+        dummy_batches = list(
+            range(num_batches)
+        )  # Gradio progress bar not compatible with generator, see https://github.com/gradio-app/gradio/issues/3841
+
+        dataloader = processor.preprocess_batch(inputs, chunk_length_s=CHUNK_LENGTH_S, batch_size=BATCH_SIZE)
+        progress(0, desc="Pre-processing audio file...")
+        dataloader = pool.map(identity, dataloader)
+
+        model_outputs = []
+        start_time = time.time()
+        # iterate over our chunked audio samples
+        for batch, _ in zip(dataloader, progress.tqdm(dummy_batches, desc="Transcribing...")):
+            model_outputs.append(forward(batch, task=task, return_timestamps=return_timestamps))
+        runtime = time.time() - start_time
+
+        post_processed = processor.postprocess(model_outputs, return_timestamps=return_timestamps)
+        text = post_processed["text"]
+        timestamps = post_processed.get("chunks")
+        if timestamps is not None:
+            timestamps = [
+                f"[{format_timestamp(chunk['timestamp'][0])} -> {format_timestamp(chunk['timestamp'][1])}] {chunk['text']}"
+                for chunk in timestamps
+            ]
+            text = "\n".join(str(feature) for feature in timestamps)
+        return text, runtime
+
+    def transcribe_chunked_audio(inputs, task, return_timestamps, progress=gr.Progress()):
+        progress(0, desc="Loading audio file...")
+        file_size_mb = os.stat(inputs).st_size / (1024 * 1024)
+        if file_size_mb > FILE_LIMIT_MB:
+            raise gr.Error(
+                f"File size exceeds file size limit. Got file of size {file_size_mb:.2f}MB for a limit of {FILE_LIMIT_MB}MB."
+            )
+
+        with open(inputs, "rb") as f:
+            inputs = f.read()
+
+        inputs = ffmpeg_read(inputs, processor.feature_extractor.sampling_rate)
+        inputs = {"array": inputs, "sampling_rate": processor.feature_extractor.sampling_rate}
+        text, runtime = tqdm_generate(inputs, task=task, return_timestamps=return_timestamps, progress=progress)
+        return text, runtime
+
+    def _return_yt_html_embed(yt_url):
+        video_id = yt_url.split("?v=")[-1]
+        HTML_str = (
+            f'<center> <iframe width="500" height="320" src="https://www.youtube.com/embed/{video_id}"> </iframe>'
+            " </center>"
+        )
+        return HTML_str
+
+    def transcribe_youtube(yt_url, task, return_timestamps, progress=gr.Progress(), max_filesize=75.0):
+        progress(0, desc="Loading audio file...")
+        html_embed_str = _return_yt_html_embed(yt_url)
+        try:
+            yt = pytube.YouTube(yt_url)
+            stream = yt.streams.filter(only_audio=True)[0]
+        except KeyError:
+            raise gr.Error("An error occurred while loading the YouTube video. Please try again.")
+
+        if stream.filesize_mb > max_filesize:
+            raise gr.Error(f"Maximum YouTube file size is {max_filesize}MB, got {stream.filesize_mb:.2f}MB.")
+
+        stream.download(filename="audio.mp3")
+
+        with open("audio.mp3", "rb") as f:
+            inputs = f.read()
+
+        inputs = ffmpeg_read(inputs, processor.feature_extractor.sampling_rate)
+        inputs = {"array": inputs, "sampling_rate": processor.feature_extractor.sampling_rate}
+        text, runtime = tqdm_generate(inputs, task=task, return_timestamps=return_timestamps, progress=progress)
+        return html_embed_str, text, runtime
+
+    microphone_chunked = gr.Interface(
+        fn=transcribe_chunked_audio,
+        inputs=[
+            gr.inputs.Audio(source="microphone", optional=True, type="filepath"),
+            gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"),
+            gr.inputs.Checkbox(default=False, label="Return timestamps"),
+        ],
+        outputs=[
+            gr.outputs.Textbox(label="Transcription").style(show_copy_button=True),
+            gr.outputs.Textbox(label="Transcription Time (s)"),
+        ],
+        allow_flagging="never",
+        title=title,
+        description=description,
+        article=article,
+    )
+
+    audio_chunked = gr.Interface(
+        fn=transcribe_chunked_audio,
+        inputs=[
+            gr.inputs.Audio(source="upload", optional=True, label="Audio file", type="filepath"),
+            gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"),
+            gr.inputs.Checkbox(default=False, label="Return timestamps"),
+        ],
+        outputs=[
+            gr.outputs.Textbox(label="Transcription").style(show_copy_button=True),
+            gr.outputs.Textbox(label="Transcription Time (s)"),
+        ],
+        allow_flagging="never",
+        title=title,
+        description=description,
+        article=article,
+    )
+
+    youtube = gr.Interface(
+        fn=transcribe_youtube,
+        inputs=[
+            gr.inputs.Textbox(lines=1, placeholder="Paste the URL to a YouTube video here", label="YouTube URL"),
+            gr.inputs.Radio(["transcribe", "translate"], label="Task", default="transcribe"),
+            gr.inputs.Checkbox(default=False, label="Return timestamps"),
+        ],
+        outputs=[
+            gr.outputs.HTML(label="Video"),
+            gr.outputs.Textbox(label="Transcription").style(show_copy_button=True),
+            gr.outputs.Textbox(label="Transcription Time (s)"),
+        ],
+        allow_flagging="never",
+        title=title,
+        examples=[["https://www.youtube.com/watch?v=m8u-18Q0s7I", "transcribe", False]],
+        cache_examples=False,
+        description=description,
+        article=article,
+    )
+
+    demo = gr.Blocks()
+
+    with demo:
+        gr.TabbedInterface([microphone_chunked, audio_chunked, youtube], ["Microphone", "Audio File", "YouTube"])
+
+    demo.queue(concurrency_count=3, max_size=5)
+    demo.launch(show_api=False)
+

nginx.conf

@@ -1,23 +0,0 @@
-server {
-    listen 7860 default_server;
-    listen [::]:7860 default_server;
-    
-    root /usr/share/nginx/html;
-    index index.html index.htm;
-
-    server_name _;
-    location / {
-        proxy_pass https://API_URL;
-        proxy_set_header    Host    API_URL;
-        proxy_set_header    X-Real-IP   $remote_addr;
-        proxy_set_header    X-Forwarded-For $proxy_add_x_forwarded_for;
-        #proxy_set_header   X-Forwarded-Proto   $scheme;
-        proxy_set_header    X-Forwarded-Proto   http;
-        proxy_set_header    X-Forwarded-Ssl     off;
-        proxy_set_header    X-Url-Scheme        http;
-        proxy_buffering     off;
-        proxy_http_version  1.1;
-        proxy_set_header    Upgrade $http_upgrade;
-        proxy_set_header    Connection  "upgrade";
-    }
-}

packages.txt

@@ -0,0 +1,2 @@
+ffmpeg
+

processing_whisper.py

@@ -0,0 +1,146 @@
+import math
+
+import numpy as np
+from transformers import WhisperProcessor
+
+
+class WhisperPrePostProcessor(WhisperProcessor):
+    def chunk_iter_with_batch(self, inputs, chunk_len, stride_left, stride_right, batch_size):
+        inputs_len = inputs.shape[0]
+        step = chunk_len - stride_left - stride_right
+
+        all_chunk_start_idx = np.arange(0, inputs_len, step)
+        num_samples = len(all_chunk_start_idx)
+
+        num_batches = math.ceil(num_samples / batch_size)
+        batch_idx = np.array_split(np.arange(num_samples), num_batches)
+
+        for i, idx in enumerate(batch_idx):
+            chunk_start_idx = all_chunk_start_idx[idx]
+
+            chunk_end_idx = chunk_start_idx + chunk_len
+
+            chunks = [inputs[chunk_start:chunk_end] for chunk_start, chunk_end in zip(chunk_start_idx, chunk_end_idx)]
+            processed = self.feature_extractor(
+                chunks, sampling_rate=self.feature_extractor.sampling_rate, return_tensors="np"
+            )
+
+            _stride_left = np.where(chunk_start_idx == 0, 0, stride_left)
+            is_last = np.where(stride_right > 0, chunk_end_idx > inputs_len, chunk_end_idx >= inputs_len)
+            _stride_right = np.where(is_last, 0, stride_right)
+
+            chunk_lens = [chunk.shape[0] for chunk in chunks]
+            strides = [
+                (int(chunk_l), int(_stride_l), int(_stride_r))
+                for chunk_l, _stride_l, _stride_r in zip(chunk_lens, _stride_left, _stride_right)
+            ]
+
+            yield {"stride": strides, **processed}
+
+    def preprocess_batch(self, inputs, chunk_length_s=0, stride_length_s=None, batch_size=None):
+        stride = None
+        if isinstance(inputs, dict):
+            stride = inputs.pop("stride", None)
+            # Accepting `"array"` which is the key defined in `datasets` for
+            # better integration
+            if not ("sampling_rate" in inputs and ("raw" in inputs or "array" in inputs)):
+                raise ValueError(
+                    "When passing a dictionary to FlaxWhisperPipline, the dict needs to contain a "
+                    '"raw" or "array" key containing the numpy array representing the audio, and a "sampling_rate" key '
+                    "containing the sampling rate associated with the audio array."
+                )
+
+            _inputs = inputs.pop("raw", None)
+            if _inputs is None:
+                # Remove path which will not be used from `datasets`.
+                inputs.pop("path", None)
+                _inputs = inputs.pop("array", None)
+            in_sampling_rate = inputs.pop("sampling_rate")
+            inputs = _inputs
+
+            if in_sampling_rate != self.feature_extractor.sampling_rate:
+                try:
+                    import librosa
+                except ImportError as err:
+                    raise ImportError(
+                        "To support resampling audio files, please install 'librosa' and 'soundfile'."
+                    ) from err
+
+                inputs = librosa.resample(
+                    inputs, orig_sr=in_sampling_rate, target_sr=self.feature_extractor.sampling_rate
+                )
+                ratio = self.feature_extractor.sampling_rate / in_sampling_rate
+            else:
+                ratio = 1
+
+        if not isinstance(inputs, np.ndarray):
+            raise ValueError(f"We expect a numpy ndarray as input, got `{type(inputs)}`.")
+        if len(inputs.shape) != 1:
+            raise ValueError(
+                f"We expect a single channel audio input for the Flax Whisper API, got {len(inputs.shape)} channels."
+            )
+
+        if stride is not None:
+            if stride[0] + stride[1] > inputs.shape[0]:
+                raise ValueError("Stride is too large for input.")
+
+            # Stride needs to get the chunk length here, it's going to get
+            # swallowed by the `feature_extractor` later, and then batching
+            # can add extra data in the inputs, so we need to keep track
+            # of the original length in the stride so we can cut properly.
+            stride = (inputs.shape[0], int(round(stride[0] * ratio)), int(round(stride[1] * ratio)))
+
+        if chunk_length_s:
+            if stride_length_s is None:
+                stride_length_s = chunk_length_s / 6
+
+            if isinstance(stride_length_s, (int, float)):
+                stride_length_s = [stride_length_s, stride_length_s]
+
+            chunk_len = round(chunk_length_s * self.feature_extractor.sampling_rate)
+            stride_left = round(stride_length_s[0] * self.feature_extractor.sampling_rate)
+            stride_right = round(stride_length_s[1] * self.feature_extractor.sampling_rate)
+
+            if chunk_len < stride_left + stride_right:
+                raise ValueError("Chunk length must be superior to stride length.")
+
+            for item in self.chunk_iter_with_batch(
+                inputs,
+                chunk_len,
+                stride_left,
+                stride_right,
+                batch_size,
+            ):
+                yield item
+        else:
+            processed = self.feature_extractor(
+                inputs, sampling_rate=self.feature_extractor.sampling_rate, return_tensors="np"
+            )
+            if stride is not None:
+                processed["stride"] = stride
+            yield processed
+
+    def postprocess(self, model_outputs, return_timestamps=None, return_language=None):
+        # unpack the outputs from list(dict(list)) to list(dict)
+        model_outputs = [dict(zip(output, t)) for output in model_outputs for t in zip(*output.values())]
+
+        time_precision = self.feature_extractor.chunk_length / 1500  # max source positions = 1500
+        # Send the chunking back to seconds, it's easier to handle in whisper
+        sampling_rate = self.feature_extractor.sampling_rate
+        for output in model_outputs:
+            if "stride" in output:
+                chunk_len, stride_left, stride_right = output["stride"]
+                # Go back in seconds
+                chunk_len /= sampling_rate
+                stride_left /= sampling_rate
+                stride_right /= sampling_rate
+                output["stride"] = chunk_len, stride_left, stride_right
+
+        text, optional = self.tokenizer._decode_asr(
+            model_outputs,
+            return_timestamps=return_timestamps,
+            return_language=return_language,
+            time_precision=time_precision,
+        )
+        return {"text": text, **optional}
+

requirements.txt

@@ -0,0 +1,4 @@
+transformers
+pytube
+requests>=2.28.2
+

run.sh

@@ -1,5 +0,0 @@
-#!/bin/bash
-
-cat nginx.conf | sed "s|API_URL|${API_URL}|g" > /etc/nginx/sites-available/default
-service nginx start
-sleep infinity