File size: 10,525 Bytes
0874a62
 
 
c0146e7
 
0874a62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
import multiprocessing
import threading
import time
from vad import AbstractTranscription, TranscriptionConfig, get_audio_duration
from whisperContainer import WhisperCallback

from multiprocessing import Pool

from typing import Any, Dict, List
import os


class ParallelContext:
    def __init__(self, num_processes: int = None, auto_cleanup_timeout_seconds: float = None):
        self.num_processes = num_processes
        self.auto_cleanup_timeout_seconds = auto_cleanup_timeout_seconds
        self.lock = threading.Lock()

        self.ref_count = 0
        self.pool = None
        self.cleanup_timer = None

    def get_pool(self):
        # Initialize pool lazily
        if (self.pool is None):
            context = multiprocessing.get_context('spawn')
            self.pool = context.Pool(self.num_processes)

        self.ref_count = self.ref_count + 1

        if (self.auto_cleanup_timeout_seconds is not None):
            self._stop_auto_cleanup()

        return self.pool

    def return_pool(self, pool):
        if (self.pool == pool and self.ref_count > 0):
            self.ref_count = self.ref_count - 1

            if (self.ref_count == 0):
                if (self.auto_cleanup_timeout_seconds is not None):
                    self._start_auto_cleanup()

    def _start_auto_cleanup(self):
        if (self.cleanup_timer is not None):
            self.cleanup_timer.cancel()
        self.cleanup_timer = threading.Timer(self.auto_cleanup_timeout_seconds, self._execute_cleanup)
        self.cleanup_timer.start()

        print("Started auto cleanup of pool in " + str(self.auto_cleanup_timeout_seconds) + " seconds")

    def _stop_auto_cleanup(self):
        if (self.cleanup_timer is not None):
            self.cleanup_timer.cancel()
            self.cleanup_timer = None

            print("Stopped auto cleanup of pool")

    def _execute_cleanup(self):
        print("Executing cleanup of pool")

        if (self.ref_count == 0):
            self.close()

    def close(self):
        self._stop_auto_cleanup()

        if (self.pool is not None):
            print("Closing pool of " + str(self.num_processes) + " processes")
            self.pool.close()
            self.pool.join()
        self.pool = None

class ParallelTranscriptionConfig(TranscriptionConfig):
    def __init__(self, device_id: str, override_timestamps, initial_segment_index, copy: TranscriptionConfig = None):
        super().__init__(copy.non_speech_strategy, copy.segment_padding_left, copy.segment_padding_right, copy.max_silent_period, copy.max_merge_size, copy.max_prompt_window, initial_segment_index)
        self.device_id = device_id
        self.override_timestamps = override_timestamps

class ParallelTranscription(AbstractTranscription):
    # Silero VAD typically takes about 3 seconds per minute, so there's no need to split the chunks 
    # into smaller segments than 2 minute (min 6 seconds per CPU core)
    MIN_CPU_CHUNK_SIZE_SECONDS = 2 * 60

    def __init__(self, sampling_rate: int = 16000):
        super().__init__(sampling_rate=sampling_rate)

    def transcribe_parallel(self, transcription: AbstractTranscription, audio: str, whisperCallable: WhisperCallback, config: TranscriptionConfig, 
                            cpu_device_count: int, gpu_devices: List[str], cpu_parallel_context: ParallelContext = None, gpu_parallel_context: ParallelContext = None):
        total_duration = get_audio_duration(audio)

        # First, get the timestamps for the original audio
        if (cpu_device_count > 1 and not transcription.is_transcribe_timestamps_fast()):
            merged = self._get_merged_timestamps_parallel(transcription, audio, config, total_duration, cpu_device_count, cpu_parallel_context)
        else:
            timestamp_segments = transcription.get_transcribe_timestamps(audio, config, 0, total_duration)
            merged = transcription.get_merged_timestamps(timestamp_segments, config, total_duration)

        # We must make sure the whisper model is downloaded
        if (len(gpu_devices) > 1):
            whisperCallable.model_container.ensure_downloaded()

        # Split into a list for each device
        # TODO: Split by time instead of by number of chunks
        merged_split = list(self._split(merged, len(gpu_devices)))

        # Parameters that will be passed to the transcribe function
        parameters = []
        segment_index = config.initial_segment_index

        for i in range(len(gpu_devices)):
            # Note that device_segment_list can be empty. But we will still create a process for it,
            # as otherwise we run the risk of assigning the same device to multiple processes.
            device_segment_list = list(merged_split[i]) if i < len(merged_split) else []
            device_id = gpu_devices[i]

            print("Device " + str(device_id) + " (index " + str(i) + ") has " + str(len(device_segment_list)) + " segments")

            # Create a new config with the given device ID
            device_config = ParallelTranscriptionConfig(device_id, device_segment_list, segment_index, config)
            segment_index += len(device_segment_list)

            parameters.append([audio, whisperCallable, device_config]);

        merged = {
            'text': '',
            'segments': [],
            'language': None
        }

        created_context = False

        perf_start_gpu = time.perf_counter()

        # Spawn a separate process for each device
        try:
            if (gpu_parallel_context is None):
                gpu_parallel_context = ParallelContext(len(gpu_devices))
                created_context = True

            # Get a pool of processes
            pool = gpu_parallel_context.get_pool()

            # Run the transcription in parallel
            results = pool.starmap(self.transcribe, parameters)

            for result in results:
                # Merge the results
                if (result['text'] is not None):
                    merged['text'] += result['text']
                if (result['segments'] is not None):
                    merged['segments'].extend(result['segments'])
                if (result['language'] is not None):
                    merged['language'] = result['language']

        finally:
            # Return the pool to the context
            if (gpu_parallel_context is not None):
                gpu_parallel_context.return_pool(pool)
            # Always close the context if we created it
            if (created_context):
                gpu_parallel_context.close()

        perf_end_gpu = time.perf_counter()
        print("Parallel transcription took " + str(perf_end_gpu - perf_start_gpu) + " seconds")

        return merged

    def _get_merged_timestamps_parallel(self, transcription: AbstractTranscription, audio: str, config: TranscriptionConfig, total_duration: float, 
                                       cpu_device_count: int, cpu_parallel_context: ParallelContext = None):
        parameters = []

        chunk_size = max(total_duration / cpu_device_count, self.MIN_CPU_CHUNK_SIZE_SECONDS)
        chunk_start = 0
        cpu_device_id = 0

        perf_start_time = time.perf_counter()

        # Create chunks that will be processed on the CPU
        while (chunk_start < total_duration):
            chunk_end = min(chunk_start + chunk_size, total_duration)

            if (chunk_end - chunk_start < 1):
                # No need to process chunks that are less than 1 second
                break

            print("Parallel VAD: Executing chunk from " + str(chunk_start) + " to " + 
                    str(chunk_end) + " on CPU device " + str(cpu_device_id))
            parameters.append([audio, config, chunk_start, chunk_end]);

            cpu_device_id += 1
            chunk_start = chunk_end

        created_context = False

        # Spawn a separate process for each device
        try:
            if (cpu_parallel_context is None):
                cpu_parallel_context = ParallelContext(cpu_device_count)
                created_context = True

            # Get a pool of processes
            pool = cpu_parallel_context.get_pool()

            # Run the transcription in parallel. Note that transcription must be picklable.
            results = pool.starmap(transcription.get_transcribe_timestamps, parameters)

            timestamps = []

            # Flatten the results
            for result in results:
                timestamps.extend(result)

            merged = transcription.get_merged_timestamps(timestamps, config, total_duration)

            perf_end_time = time.perf_counter()
            print("Parallel VAD processing took {} seconds".format(perf_end_time - perf_start_time))
            return merged

        finally:
            # Return the pool to the context
            if (cpu_parallel_context is not None):
                cpu_parallel_context.return_pool(pool)
            # Always close the context if we created it
            if (created_context):
                cpu_parallel_context.close()

    def get_transcribe_timestamps(self, audio: str, config: ParallelTranscriptionConfig, start_time: float, duration: float):
        return []

    def get_merged_timestamps(self,  timestamps: List[Dict[str, Any]], config: ParallelTranscriptionConfig, total_duration: float):
        # Override timestamps that will be processed
        if (config.override_timestamps is not None):
            print("Using override timestamps of size " + str(len(config.override_timestamps)))
            return config.override_timestamps
        return super().get_merged_timestamps(timestamps, config, total_duration)

    def transcribe(self, audio: str, whisperCallable: WhisperCallback, config: ParallelTranscriptionConfig):
        # Override device ID the first time
        if (os.environ.get("INITIALIZED", None) is None):
            os.environ["INITIALIZED"] = "1"

            # Note that this may be None if the user didn't specify a device. In that case, Whisper will
            # just use the default GPU device.
            if (config.device_id is not None):
                print("Using device " + config.device_id)
                os.environ["CUDA_VISIBLE_DEVICES"] = config.device_id
        
        return super().transcribe(audio, whisperCallable, config)

    def _split(self, a, n):
        """Split a list into n approximately equal parts."""
        k, m = divmod(len(a), n)
        return (a[i*k+min(i, m):(i+1)*k+min(i+1, m)] for i in range(n))