Afrinetwork7
commited on
Commit
•
d63d47a
1
Parent(s):
b3a33de
Update app.py
Browse files
app.py
CHANGED
@@ -15,9 +15,9 @@ from whisper_jax import FlaxWhisperPipline
|
|
15 |
app = FastAPI(title="Whisper JAX: The Fastest Whisper API ⚡️")
|
16 |
|
17 |
logger = logging.getLogger("whisper-jax-app")
|
18 |
-
logger.setLevel(logging.
|
19 |
ch = logging.StreamHandler()
|
20 |
-
ch.setLevel(logging.
|
21 |
formatter = logging.Formatter("%(asctime)s;%(levelname)s;%(message)s", "%Y-%m-%d %H:%M:%S")
|
22 |
ch.setFormatter(formatter)
|
23 |
logger.addHandler(ch)
|
@@ -37,7 +37,7 @@ stride_left = stride_right = round(stride_length_s * pipeline.feature_extractor.
|
|
37 |
step = chunk_len - stride_left - stride_right
|
38 |
|
39 |
# do a pre-compile step so that the first user to use the demo isn't hit with a long transcription time
|
40 |
-
logger.
|
41 |
start = time.time()
|
42 |
random_inputs = {
|
43 |
"input_features": np.ones(
|
@@ -46,11 +46,11 @@ random_inputs = {
|
|
46 |
}
|
47 |
random_timestamps = pipeline.forward(random_inputs, batch_size=BATCH_SIZE, return_timestamps=True)
|
48 |
compile_time = time.time() - start
|
49 |
-
logger.
|
50 |
|
51 |
@app.post("/transcribe_audio")
|
52 |
async def transcribe_chunked_audio(audio_file: UploadFile, task: str = "transcribe", return_timestamps: bool = False):
|
53 |
-
logger.
|
54 |
if not audio_file:
|
55 |
logger.warning("No audio file")
|
56 |
raise HTTPException(status_code=400, detail="No audio file submitted!")
|
@@ -59,30 +59,60 @@ async def transcribe_chunked_audio(audio_file: UploadFile, task: str = "transcri
|
|
59 |
logger.warning("Max file size exceeded")
|
60 |
raise HTTPException(status_code=400, detail=f"File size exceeds file size limit. Got file of size {file_size_mb:.2f}MB for a limit of {FILE_LIMIT_MB}MB.")
|
61 |
|
62 |
-
|
63 |
-
|
|
|
|
|
|
|
|
|
64 |
|
65 |
inputs = ffmpeg_read(inputs, pipeline.feature_extractor.sampling_rate)
|
66 |
inputs = {"array": inputs, "sampling_rate": pipeline.feature_extractor.sampling_rate}
|
67 |
-
logger.
|
68 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
return {"text": text, "runtime": runtime}
|
70 |
|
71 |
@app.post("/transcribe_youtube")
|
72 |
async def transcribe_youtube(yt_url: str = Form(...), task: str = "transcribe", return_timestamps: bool = False):
|
73 |
-
logger.
|
74 |
-
|
|
|
|
|
|
|
|
|
|
|
75 |
with tempfile.TemporaryDirectory() as tmpdirname:
|
76 |
filepath = os.path.join(tmpdirname, "video.mp4")
|
77 |
-
|
|
|
|
|
|
|
|
|
|
|
78 |
|
79 |
-
|
80 |
-
|
|
|
|
|
|
|
|
|
81 |
|
82 |
inputs = ffmpeg_read(inputs, pipeline.feature_extractor.sampling_rate)
|
83 |
inputs = {"array": inputs, "sampling_rate": pipeline.feature_extractor.sampling_rate}
|
84 |
-
logger.
|
85 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
return {"html_embed": html_embed_str, "text": text, "runtime": runtime}
|
87 |
|
88 |
def tqdm_generate(inputs: dict, task: str, return_timestamps: bool):
|
@@ -94,15 +124,19 @@ def tqdm_generate(inputs: dict, task: str, return_timestamps: bool):
|
|
94 |
dataloader = pipeline.preprocess_batch(inputs, chunk_length_s=CHUNK_LENGTH_S, batch_size=BATCH_SIZE)
|
95 |
model_outputs = []
|
96 |
start_time = time.time()
|
97 |
-
logger.
|
98 |
# iterate over our chunked audio samples - always predict timestamps to reduce hallucinations
|
99 |
for batch in dataloader:
|
100 |
model_outputs.append(pipeline.forward(batch, batch_size=BATCH_SIZE, task=task, return_timestamps=True))
|
101 |
runtime = time.time() - start_time
|
102 |
-
logger.
|
103 |
|
104 |
-
logger.
|
105 |
-
|
|
|
|
|
|
|
|
|
106 |
text = post_processed["text"]
|
107 |
if return_timestamps:
|
108 |
timestamps = post_processed.get("chunks")
|
@@ -111,7 +145,7 @@ def tqdm_generate(inputs: dict, task: str, return_timestamps: bool):
|
|
111 |
for chunk in timestamps
|
112 |
]
|
113 |
text = "\n".join(str(feature) for feature in timestamps)
|
114 |
-
logger.
|
115 |
return text, runtime
|
116 |
|
117 |
def _return_yt_html_embed(yt_url):
|
@@ -125,8 +159,10 @@ def _return_yt_html_embed(yt_url):
|
|
125 |
def download_yt_audio(yt_url, filename):
|
126 |
info_loader = youtube_dl.YoutubeDL()
|
127 |
try:
|
|
|
128 |
info = info_loader.extract_info(yt_url, download=False)
|
129 |
except youtube_dl.utils.DownloadError as err:
|
|
|
130 |
raise HTTPException(status_code=400, detail=str(err))
|
131 |
|
132 |
file_length = info["duration_string"]
|
@@ -146,8 +182,10 @@ def download_yt_audio(yt_url, filename):
|
|
146 |
ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}
|
147 |
with youtube_dl.YoutubeDL(ydl_opts) as ydl:
|
148 |
try:
|
|
|
149 |
ydl.download([yt_url])
|
150 |
except youtube_dl.utils.ExtractorError as err:
|
|
|
151 |
raise HTTPException(status_code=400, detail=str(err))
|
152 |
|
153 |
def format_timestamp(seconds: float, always_include_hours: bool = False, decimal_marker: str = "."):
|
|
|
15 |
app = FastAPI(title="Whisper JAX: The Fastest Whisper API ⚡️")
|
16 |
|
17 |
logger = logging.getLogger("whisper-jax-app")
|
18 |
+
logger.setLevel(logging.DEBUG)
|
19 |
ch = logging.StreamHandler()
|
20 |
+
ch.setLevel(logging.DEBUG)
|
21 |
formatter = logging.Formatter("%(asctime)s;%(levelname)s;%(message)s", "%Y-%m-%d %H:%M:%S")
|
22 |
ch.setFormatter(formatter)
|
23 |
logger.addHandler(ch)
|
|
|
37 |
step = chunk_len - stride_left - stride_right
|
38 |
|
39 |
# do a pre-compile step so that the first user to use the demo isn't hit with a long transcription time
|
40 |
+
logger.debug("Compiling forward call...")
|
41 |
start = time.time()
|
42 |
random_inputs = {
|
43 |
"input_features": np.ones(
|
|
|
46 |
}
|
47 |
random_timestamps = pipeline.forward(random_inputs, batch_size=BATCH_SIZE, return_timestamps=True)
|
48 |
compile_time = time.time() - start
|
49 |
+
logger.debug(f"Compiled in {compile_time}s")
|
50 |
|
51 |
@app.post("/transcribe_audio")
|
52 |
async def transcribe_chunked_audio(audio_file: UploadFile, task: str = "transcribe", return_timestamps: bool = False):
|
53 |
+
logger.debug("Loading audio file...")
|
54 |
if not audio_file:
|
55 |
logger.warning("No audio file")
|
56 |
raise HTTPException(status_code=400, detail="No audio file submitted!")
|
|
|
59 |
logger.warning("Max file size exceeded")
|
60 |
raise HTTPException(status_code=400, detail=f"File size exceeds file size limit. Got file of size {file_size_mb:.2f}MB for a limit of {FILE_LIMIT_MB}MB.")
|
61 |
|
62 |
+
try:
|
63 |
+
with open(audio_file.filename, "rb") as f:
|
64 |
+
inputs = f.read()
|
65 |
+
except Exception as e:
|
66 |
+
logger.error("Error reading audio file:", exc_info=True)
|
67 |
+
raise HTTPException(status_code=500, detail="Error reading audio file")
|
68 |
|
69 |
inputs = ffmpeg_read(inputs, pipeline.feature_extractor.sampling_rate)
|
70 |
inputs = {"array": inputs, "sampling_rate": pipeline.feature_extractor.sampling_rate}
|
71 |
+
logger.debug("Done loading audio file")
|
72 |
+
|
73 |
+
try:
|
74 |
+
text, runtime = tqdm_generate(inputs, task=task, return_timestamps=return_timestamps)
|
75 |
+
except Exception as e:
|
76 |
+
logger.error("Error transcribing audio:", exc_info=True)
|
77 |
+
raise HTTPException(status_code=500, detail="Error transcribing audio")
|
78 |
+
|
79 |
return {"text": text, "runtime": runtime}
|
80 |
|
81 |
@app.post("/transcribe_youtube")
|
82 |
async def transcribe_youtube(yt_url: str = Form(...), task: str = "transcribe", return_timestamps: bool = False):
|
83 |
+
logger.debug("Loading YouTube file...")
|
84 |
+
try:
|
85 |
+
html_embed_str = _return_yt_html_embed(yt_url)
|
86 |
+
except Exception as e:
|
87 |
+
logger.error("Error generating YouTube HTML embed:", exc_info=True)
|
88 |
+
raise HTTPException(status_code=500, detail="Error generating YouTube HTML embed")
|
89 |
+
|
90 |
with tempfile.TemporaryDirectory() as tmpdirname:
|
91 |
filepath = os.path.join(tmpdirname, "video.mp4")
|
92 |
+
try:
|
93 |
+
logger.debug("Downloading YouTube audio...")
|
94 |
+
download_yt_audio(yt_url, filepath)
|
95 |
+
except Exception as e:
|
96 |
+
logger.error("Error downloading YouTube audio:", exc_info=True)
|
97 |
+
raise HTTPException(status_code=500, detail="Error downloading YouTube audio")
|
98 |
|
99 |
+
try:
|
100 |
+
with open(filepath, "rb") as f:
|
101 |
+
inputs = f.read()
|
102 |
+
except Exception as e:
|
103 |
+
logger.error("Error reading downloaded audio file:", exc_info=True)
|
104 |
+
raise HTTPException(status_code=500, detail="Error reading downloaded audio file")
|
105 |
|
106 |
inputs = ffmpeg_read(inputs, pipeline.feature_extractor.sampling_rate)
|
107 |
inputs = {"array": inputs, "sampling_rate": pipeline.feature_extractor.sampling_rate}
|
108 |
+
logger.debug("Done loading YouTube file")
|
109 |
+
|
110 |
+
try:
|
111 |
+
text, runtime = tqdm_generate(inputs, task=task, return_timestamps=return_timestamps)
|
112 |
+
except Exception as e:
|
113 |
+
logger.error("Error transcribing YouTube audio:", exc_info=True)
|
114 |
+
raise HTTPException(status_code=500, detail="Error transcribing YouTube audio")
|
115 |
+
|
116 |
return {"html_embed": html_embed_str, "text": text, "runtime": runtime}
|
117 |
|
118 |
def tqdm_generate(inputs: dict, task: str, return_timestamps: bool):
|
|
|
124 |
dataloader = pipeline.preprocess_batch(inputs, chunk_length_s=CHUNK_LENGTH_S, batch_size=BATCH_SIZE)
|
125 |
model_outputs = []
|
126 |
start_time = time.time()
|
127 |
+
logger.debug("Transcribing...")
|
128 |
# iterate over our chunked audio samples - always predict timestamps to reduce hallucinations
|
129 |
for batch in dataloader:
|
130 |
model_outputs.append(pipeline.forward(batch, batch_size=BATCH_SIZE, task=task, return_timestamps=True))
|
131 |
runtime = time.time() - start_time
|
132 |
+
logger.debug("Done transcription")
|
133 |
|
134 |
+
logger.debug("Post-processing...")
|
135 |
+
try:
|
136 |
+
post_processed = pipeline.postprocess(model_outputs, return_timestamps=True)
|
137 |
+
except Exception as e:
|
138 |
+
logger.error("Error post-processing transcription:", exc_info=True)
|
139 |
+
raise HTTPException(status_code=500, detail="Error post-processing transcription")
|
140 |
text = post_processed["text"]
|
141 |
if return_timestamps:
|
142 |
timestamps = post_processed.get("chunks")
|
|
|
145 |
for chunk in timestamps
|
146 |
]
|
147 |
text = "\n".join(str(feature) for feature in timestamps)
|
148 |
+
logger.debug("Done post-processing")
|
149 |
return text, runtime
|
150 |
|
151 |
def _return_yt_html_embed(yt_url):
|
|
|
159 |
def download_yt_audio(yt_url, filename):
|
160 |
info_loader = youtube_dl.YoutubeDL()
|
161 |
try:
|
162 |
+
logger.debug(f"Extracting info for YouTube URL: {yt_url}")
|
163 |
info = info_loader.extract_info(yt_url, download=False)
|
164 |
except youtube_dl.utils.DownloadError as err:
|
165 |
+
logger.error("Error extracting YouTube info:", exc_info=True)
|
166 |
raise HTTPException(status_code=400, detail=str(err))
|
167 |
|
168 |
file_length = info["duration_string"]
|
|
|
182 |
ydl_opts = {"outtmpl": filename, "format": "worstvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best"}
|
183 |
with youtube_dl.YoutubeDL(ydl_opts) as ydl:
|
184 |
try:
|
185 |
+
logger.debug(f"Downloading YouTube audio to {filename}")
|
186 |
ydl.download([yt_url])
|
187 |
except youtube_dl.utils.ExtractorError as err:
|
188 |
+
logger.error("Error downloading YouTube audio:", exc_info=True)
|
189 |
raise HTTPException(status_code=400, detail=str(err))
|
190 |
|
191 |
def format_timestamp(seconds: float, always_include_hours: bool = False, decimal_marker: str = "."):
|