Wewe ni kupoteza GPU nguvu—Fix yako TensorFlow Input Pipeline leo

Maelezo ya Maudhui Mtazamo wa rasilimali ya Muundo wa Data ya Mafunzo ya loop Njia ya ufanisi Njia ya uadilifu Maelezo ya Prefetching Uchimbaji wa data parallel Mabadiliko ya data ya parallel ya caching Mipango ya Vectorizing Kupunguza kipimo cha kumbukumbu Mtazamo wa GPUs na TPUs zinaweza kupunguza muda unaohitajika kutekeleza hatua moja ya mafunzo. Kufikia utendaji wa kiwango cha juu unahitaji pipeline ya kuingia yenye ufanisi ambayo inatoa data kwa hatua inayofuata kabla ya hatua ya sasa imekamilika. API husaidia kujenga miundombinu ya ufumbuzi ya kiwango cha juu na yenye ufanisi. Makala hii inaonyesha jinsi ya kutumia API ya kujenga vifaa vya juu vya TensorFlow. tf.data tf.data Kabla ya kuendelea, angalia mwongozo wa kujifunza jinsi ya kutumia Moto wa moto. Ujenzi wa vifaa vya TensorFlow tf.data rasilimali ya Ujenzi wa vifaa vya TensorFlow Tf.data.Dataset ya API Uchambuzi wa utendaji wa tf.data na Profiler ya TF Muundo wa import tensorflow as tf import time 2024-08-15 02:29:59.860833: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:485] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered 2024-08-15 02:29:59.882136: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:8454] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered 2024-08-15 02:29:59.888640: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1452] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered Katika mwongozo huu, utakuwa iterate juu ya dataset na kupima utendaji. Kutoa viwango vya utendaji vinavyoweza kurejeshwa inaweza kuwa vigumu. Sababu tofauti zinazosababisha kurejeshwa ni pamoja na: Kiwango cha sasa cha CPU Traffic ya mtandao Mfumo wa kutosha, kama vile Cache Ili kupata benchmark inaweza kurejeshwa, utaunda mfano wa kiufundi. Data ya Kuanza na kufafanua darasa la urithi kutoka wito wa Kwa data hii: tf.data.Dataset ArtificialDataset Kuunda sampuli za num_samples (mfano ni 3) Inapumzika kwa muda fulani kabla ya kipengele cha kwanza kuiga kufungua faili Inapumzika kwa muda kabla ya kuzalisha kila kipengele ili kuiga data ya kusoma kutoka kwa faili class ArtificialDataset(tf.data.Dataset): def _generator(num_samples): # Opening the file time.sleep(0.03) for sample_idx in range(num_samples): # Reading data (line, record) from the file time.sleep(0.015) yield (sample_idx,) def __new__(cls, num_samples=3): return tf.data.Dataset.from_generator( cls._generator, output_signature = tf.TensorSpec(shape = (1,), dtype = tf.int64), args=(num_samples,) ) Takwimu hizi ni sawa na moja, kuongeza muda maalum katika mwanzo wa na kati ya kila sampuli. tf.data.Dataset.range Mafunzo ya loop Kisha, kuandika buffer ya mafunzo ambayo inathiri muda gani unachukua kutafakari juu ya dataset. def benchmark(dataset, num_epochs=2): start_time = time.perf_counter() for epoch_num in range(num_epochs): for sample in dataset: # Performing a training step time.sleep(0.01) print("Execution time:", time.perf_counter() - start_time) Ufanisi wa ufanisi Ili kuonyesha jinsi utendaji unaweza kuboreshwa, utaongeza utendaji wa . ArtificialDataset Njia ya uadilifu Kuanza na pipeline naivu kwa kutumia hakuna mashaka, iterating juu ya dataset as-is. benchmark(ArtificialDataset()) WARNING: All log messages before absl::InitializeLog() is called are written to STDERR I0000 00:00:1723689002.526086 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.529717 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.533395 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.537215 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.548874 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.552105 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.555507 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.559016 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.562268 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.565421 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.568842 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689002.572296 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.797478 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.799457 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.801473 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.803565 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.805613 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.807466 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.809382 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.811373 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.813299 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.815122 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.817022 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.819005 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.857728 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.860099 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.862063 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.864094 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.866073 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.867912 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.869940 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.871942 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.873924 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.876256 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.878556 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 I0000 00:00:1723689003.880974 112933 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355 Execution time: 0.28992818999995507 Chini ya kapu, hii ni jinsi wakati wako wa utekelezaji ulitumika: Mchoro unaonyesha kwamba kutekeleza hatua ya mafunzo inahusisha: Kufungua faili ikiwa haijafunguliwa Kufuta data ya kuingia kutoka kwa faili Kutumia data kwa ajili ya mafunzo Hata hivyo, katika utekelezaji mzuri wa synchronous kama hapa, wakati pipeline yako inachukua data, mfano wako unachunguzwa. kinyume chake, wakati mfano wako ni mafunzo, pipeline ya kuingia inachunguzwa. Muda wa hatua ya mafunzo ni hivyo jumla ya kufungua, kusoma na mafunzo ya muda. Sehemu zifuatazo zinajenga juu ya mtiririko huu wa kuingia, na kuonyesha mazoea bora ya kubuni mtiririko wa kuingia wa TensorFlow. Maelezo ya Prefetching Prefetching inashughulikia utekelezaji wa preprocessing na mfano wa hatua ya mafunzo. , pipa ya kuingia inasoma data kwa hatua Kufanya hivyo hupunguza muda wa hatua kwa kiwango cha juu (kwa kinyume na jumla) ya mafunzo na muda unachukua kuondoa data. s s+1 ya Moto huo unatoa mabadiliko. Inaweza kutumika kutenganisha muda ambapo data hutengenezwa kutoka wakati data hutumiwa. Hasa, mabadiliko hutumia thread ya nyuma na buffer ya ndani ili kurekebisha vipengele kutoka kwenye dataset ya kuingia kabla ya muda wanaohitajika. Idadi ya vipengele vya kurekebisha lazima iwe sawa (au uwezekano mkubwa zaidi) na idadi ya mikataba zinazotumiwa na hatua moja ya mafunzo. Unaweza au kurekebisha thamani hii kwa manually, au kuweka kwa ambayo itasaidia kwa haraka Runtime ili kurekebisha thamani kwa nguvu wakati wa runtime. tf.data tf.data.Dataset.prefetch tf.data.AUTOTUNE tf.data Kumbuka kwamba mabadiliko ya prefetch hutoa faida wakati wowote kuna fursa ya kuunganisha kazi ya "mfanyabiashara" na kazi ya "mfanyabiashara". benchmark( ArtificialDataset() .prefetch(tf.data.AUTOTUNE) ) Execution time: 0.282483695999872 Sasa, kama mstari wa muda wa utekelezaji wa data unaonyesha, wakati hatua ya mafunzo inafanya kazi kwa sampuli 0, mstari wa kuingia unasoma data kwa sampuli 1, na kadhalika. Uchimbaji wa data parallel Katika mazingira ya ulimwengu halisi, data ya kuingia inaweza kuhifadhiwa mbali (kwa mfano, kwenye Google Cloud Storage au HDFS). Njia ya mstari wa data ambayo inafanya kazi vizuri wakati wa kusoma data kwa ndani inaweza kuingia kwenye I/O wakati wa kusoma data kwa mbali kutokana na tofauti zifuatazo kati ya uhifadhi wa ndani na wa mbali: Time-to-first-byte: Kusoma byte ya kwanza ya faili kutoka storage mbali inaweza kuchukua amri ya ukubwa zaidi kuliko kutoka storage ya ndani. Kusoma uwezo: Wakati storage mbali kawaida hutoa bandwidth kubwa ya jumla, kusoma faili moja inaweza tu kuwa na uwezo wa kutumia sehemu ndogo ya bandwidth hii. Zaidi ya hayo, mara bajeti nyekundu zimefungwa kwenye kumbukumbu, inaweza pia kuwa muhimu kuwasha na / au kufuta data (kwa mfano. Hii juu ni sasa bila kujali kama data ni kuhifadhiwa ndani au mbali, lakini inaweza kuwa mbaya katika kesi ya mbali kama data si prefetched kwa ufanisi. Mpango wa Protobuf Ili kupunguza athari za uchimbaji wa data mbalimbali kwa ujumla, muundo inaweza kutumika kulinganisha hatua ya kupakia data, kuacha maudhui ya datasets nyingine (kama vile data file reader). idadi ya datasets kuunganisha inaweza kufafanuliwa na kwa upande mwingine, hata hivyo, kiwango cha ubadilishaji unaweza kufafanuliwa na Mstari wa chini: Roho inawakilishwa Mabadiliko ya, ya Msaada wa Mabadiliko , ambayo itatoa uamuzi juu ya kiwango gani cha parallelism kutumia kwa ya runtime. tf.data.Dataset.interleave cycle_length num_parallel_calls prefetch interleave tf.data.AUTOTUNE tf.data Mchakato wa Interleave Kwa sababu ya maneno ya dharura ya Mabadiliko hufanya interleave sampuli moja kutoka datasets mbili kwa mfululizo. tf.data.Dataset.interleave benchmark( tf.data.Dataset.range(2) .interleave(lambda _: ArtificialDataset()) ) Execution time: 0.4770545540004605 Mchakato huu wa utekelezaji wa data unaruhusu kuonyesha tabia ya kubadilisha, kukusanya sampuli mbadala kutoka kwa datasets mbili inapatikana. Hata hivyo, hakuna kuboresha utendaji unaohusika hapa. interleave Upatikanaji wa Interleave Sasa, matumizi ya Miongoni mwa hoja za Mabadiliko. hii inashughulikia seti nyingi za data kwa pamoja, kupunguza muda wa kusubiri faili kufunguliwa. num_parallel_calls interleave benchmark( tf.data.Dataset.range(2) .interleave( lambda _: ArtificialDataset(), num_parallel_calls=tf.data.AUTOTUNE ) ) Execution time: 0.37743708600009995 Wakati huu, kama mfululizo wa muda wa utekelezaji wa data unaonyesha, kusoma ya seti mbili za data ni parallelized, kupunguza muda wa usindikaji wa data duniani. Mabadiliko ya data ya parallel Wakati wa kuandaa data, vipengele vya kuingia vinaweza kuhitajika kabla ya usindikaji. Moto huo unatolewa kwa transformation, ambayo inatumia kazi iliyowekwa na mtumiaji kwa kila kipengele cha data ya input. Kwa sababu vipengele vya input ni huru kutoka kwa kila mmoja, pre-processing inaweza kuwa parallelized juu ya viungo vingi vya CPU. na ya Mabadiliko ya, ya Mabadiliko hayo yanatoa Kwa hiyo, unahitaji kuelezea kiwango cha parallelism. tf.data tf.data.Dataset.map prefetch interleave map num_parallel_calls Kuchagua thamani bora kwa ajili ya hoja inategemea vifaa vyako, sifa za data yako ya mafunzo (kama vile ukubwa na sura yake), gharama ya kazi ya ramani yako, na usindikaji mwingine unaofanyika kwenye CPU wakati huo huo. na ya Mabadiliko ya, ya Msaada wa Mabadiliko ambayo itatoa uamuzi juu ya kiwango gani cha parallelism kutumika kwa ya runtime. num_parallel_calls prefetch interleave map tf.data.AUTOTUNE tf.data def mapped_function(s): # Do some hard pre-processing tf.py_function(lambda: time.sleep(0.03), [], ()) return s Mipango ya sequential Kuanza kwa kutumia Mabadiliko yasiyo na parallelism kama mfano wa msingi. map benchmark( ArtificialDataset() .map(mapped_function) ) Execution time: 0.49222556600034295 Kama kwa ajili ya , hapa, kama mandhari inaonyesha, muda uliotumika kwa ajili ya kufungua, kusoma, kabla ya usindikaji (mapping) na mafunzo hatua pamoja kwa iteration moja. Njia ya Naive Ramani ya Pamoja Sasa, hutumia kazi hiyo ya kabla ya usindikaji lakini inatumia parallel juu ya sampuli nyingi. benchmark( ArtificialDataset() .map( mapped_function, num_parallel_calls=tf.data.AUTOTUNE ) ) Execution time: 0.36392719900049997 Kama maelezo ya data inaonyesha, hatua za kabla ya usindikaji zinachanganya, kupunguza muda wa jumla kwa iteration moja. ya caching ya Mabadiliko yanaweza kuhifadhi mkusanyiko wa data, katika kumbukumbu au kwenye uhifadhi wa ndani. Hii itaokoa shughuli fulani (kama vile kufungua faili na kusoma data) kutoka kwa kutekelezwa wakati wa kila enzi. tf.data.Dataset.cache benchmark( ArtificialDataset() .map( # Apply time consuming operations before cache mapped_function ).cache( ), 5 ) Execution time: 0.3864543270001377 Hapa, utaratibu wa muda wa utekelezaji wa data unaonyesha kwamba wakati unavyohifadhi data, mabadiliko kabla ya moja (kama vile kufungua faili na kusoma data) hufanywa tu wakati wa kipindi cha kwanza. Mabadiliko ya cache cache Ikiwa kazi ya user-defined imepita kwenye Mabadiliko ni ya gharama kubwa, na Mabadiliko baada ya uongofu kwa muda mrefu kama dataset inayotokana bado inaweza kuingizwa katika kumbukumbu au uhifadhi wa ndani. Ikiwa kazi iliyoelezwa na mtumiaji huongeza nafasi inayohitajika kuhifadhi dataset zaidi ya uwezo wa cache, angalia baada ya kubadilisha au kuzingatia kabla ya usindikaji wa data yako kabla ya kazi yako ya mafunzo ili kupunguza matumizi ya rasilimali. map cache map cache Mipango ya Vectorizing Kuomba kazi ya user-defined iliyotolewa katika Mabadiliko haya yanahusiana na mipango na utekelezaji wa kazi iliyoelezwa na mtumiaji. Vectorize kazi iliyoelezwa na mtumiaji (yaani, kufanya kazi juu ya mfululizo wa input kwa wakati mmoja) na kutumia Mageuzi ya ya Mabadiliko ya map batch kabla ya map Ili kuonyesha mazoea haya mazuri, dataset yako ya kibinadamu hainafaa. Utaratibu wa mipangilio ni karibu 10 microseconds (10e-6 sekunde), kidogo kuliko maelfu ya milliseconds zilizotumika katika Kwa hiyo, athari zake ni vigumu kuona. ArtificialDataset Kwa mfano, matumizi ya msingi kazi na kufafanua mzunguko wa mafunzo kwa fomu yake rahisi zaidi. tf.data.Dataset.range fast_dataset = tf.data.Dataset.range(10000) def fast_benchmark(dataset, num_epochs=2): start_time = time.perf_counter() for _ in tf.data.Dataset.range(num_epochs): for _ in dataset: pass tf.print("Execution time:", time.perf_counter() - start_time) def increment(x): return x+1 Mipango ya Scalar fast_benchmark( fast_dataset # Apply function one item at a time .map(increment) # Batch .batch(256) ) Execution time: 0.23993100699954084 Mchoro hapo juu unaonyesha kile kinachotokea (na sampuli ndogo) kwa kutumia mbinu ya mapitio ya scalar. Inaonyesha kwamba kazi iliyochapishwa inatumika kwa kila sampuli. Ingawa kazi hii ni ya haraka sana, ina baadhi ya juu ambayo huathiri utendaji wa wakati. Mipango ya Vectorized fast_benchmark( fast_dataset .batch(256) # Apply function on a batch of items # The tf.Tensor.__add__ method already handle batches .map(increment) ) Execution time: 0.04984995199993136 Wakati huu, kazi iliyoandikwa inachukuliwa mara moja na inatumika kwa mfululizo wa sampuli. Kama mpango wa muda wa utekelezaji wa data unaonyesha, wakati kazi inaweza kuchukua muda zaidi kutekeleza, juu huonekana mara moja tu, kuboresha utendaji wa jumla wa muda. Kupunguza kipimo cha kumbukumbu Baadhi ya mabadiliko, ikiwa ni pamoja na ya na ya Kuanza kucheza katika TopSlotSite , kufuata 3 hatua rahisi kujiandikisha. kubadilisha kubadilisha ukubwa wa vipengele, kisha utaratibu wa mabadiliko ya ramani na mabadiliko ambayo vipengele buffer huathiri matumizi ya kumbukumbu. Kwa ujumla, kuchagua utaratibu ambayo ina matokeo ya chini ya kumbukumbu, isipokuwa utaratibu tofauti ni tamaa kwa utendaji. interleave prefetch shuffle map Kuhifadhi kompyuta ya sehemu Inashauriwa kuhifadhi data baada ya uongofu isipokuwa uongofu huu hufanya data kuwa kubwa sana ili kuingizwa katika kumbukumbu. Uongofu unaweza kupatikana ikiwa kazi yako iliyochapishwa inaweza kugawanywa katika sehemu mbili: sehemu ambayo inachukua muda mmoja na sehemu ambayo inachukua kumbukumbu. map dataset.map(time_consuming_mapping).cache().map(memory_consuming_mapping) Kwa njia hii, sehemu ambayo inachukua muda hufanywa tu wakati wa kipindi cha kwanza, na unakataa kutumia nafasi nyingi ya cache. Awali iliyochapishwa kwenye tovuti ya TensorFlow, makala hii inaonekana hapa chini ya kichwa kipya na inatolewa kwa leseni chini ya CC BY 4.0. Awali iliyochapishwa kwenye tovuti ya TensorFlow, makala hii inaonekana hapa chini ya kichwa kipya na inatolewa kwa leseni chini ya CC BY 4.0. Uwezo wa TensorFlow Uwezo wa TensorFlow