Tensorflow tf.data.Dataset.cache 好像没有达到预期的效果
Tensorflow tf.data.Dataset.cache seems do not take the expected effect
我正在努力按照 Better performance with the tf.data API 指南提高我的模型训练性能。但是，我观察到使用 .cache() 的性能与没有 .cache().
的相同设置相比几乎相同甚至更差。
datafile_list = load_my_files()
RAW_BYTES = 403*4
BATCH_SIZE = 32
raw_dataset = tf.data.FixedLengthRecordDataset(filenames=datafile_list, record_bytes=RAW_BYTES, num_parallel_reads=10, buffer_size=1024*RAW_BYTES)
raw_dataset = raw_dataset.map(tf.autograph.experimental.do_not_convert(decode_and_prepare),
num_parallel_calls=tf.data.AUTOTUNE)
raw_dataset = raw_dataset.cache()
raw_dataset = raw_dataset.shuffle(buffer_size=1024)
raw_dataset = raw_dataset.batch(BATCH_SIZE)
raw_dataset = raw_dataset.prefetch(tf.data.AUTOTUNE)
datafile_list 中的数据有 9.92GB，相当适合系统可用的总物理 RAM (100GB)。系统交换已禁用。
通过使用数据集训练模型：
model = build_model()
model.fit(raw_dataset, epochs=5, verbose=2)
结果：
Epoch 1/5
206247/206247 - 126s - loss: 0.0043 - mae: 0.0494 - mse: 0.0043
Epoch 2/5
206247/206247 - 125s - loss: 0.0029 - mae: 0.0415 - mse: 0.0029
Epoch 3/5
206247/206247 - 129s - loss: 0.0027 - mae: 0.0397 - mse: 0.0027
Epoch 4/5
206247/206247 - 125s - loss: 0.0025 - mae: 0.0386 - mse: 0.0025
Epoch 5/5
206247/206247 - 125s - loss: 0.0024 - mae: 0.0379 - mse: 0.0024
这个结果令人沮丧。由 docs:
The first time the dataset is iterated over, its elements will be cached either in the specified file or in memory. Subsequent iterations will use the cached data.
来自this guide：
When iterating over this dataset, the second iteration will be much faster than the first one thanks to the caching.
但是，所有epoch 所用的时间几乎相同。此外，在训练期间 CPU 和 GPU 使用率都非常低（见下图）。
通过注释掉 raw_dataset = raw_dataset.cache() 行，结果没有显示任何显着差异：
Epoch 1/5
206067/206067 - 129s - loss: 0.0042 - mae: 0.0492 - mse: 0.0042
Epoch 2/5
206067/206067 - 127s - loss: 0.0028 - mae: 0.0412 - mse: 0.0028
Epoch 3/5
206067/206067 - 134s - loss: 0.0026 - mae: 0.0393 - mse: 0.0026
Epoch 4/5
206067/206067 - 127s - loss: 0.0024 - mae: 0.0383 - mse: 0.0024
Epoch 5/5
206067/206067 - 126s - loss: 0.0023 - mae: 0.0376 - mse: 0.0023
正如文档中所指出的，我的期望是使用缓存会导致更快的训练时间。我想知道我做错了什么。
附件
训练期间使用缓存的 GPU 使用率：
无缓存训练期间的 GPU 使用率：
使用缓存训练期间的系统统计信息（内存，CPU 等）：
无缓存训练期间的系统统计信息（内存，CPU 等）：
只是使用 Google Colab 的一个小观察。根据 docs:
Note: For the cache to be finalized, the input dataset must be iterated through in its entirety. Otherwise, subsequent iterations will not use cached data.
和
Note: cache will produce exactly the same elements during each
iteration through the dataset. If you wish to randomize the iteration
order, make sure to call shuffle after calling cache.
我确实注意到在预先使用缓存和迭代数据集时存在一些差异。这是一个例子。
准备数据:
import random
import struct
import tensorflow as tf
import numpy as np
RAW_N = 2 + 20*20 + 1
bytess = random.sample(range(1, 5000), RAW_N*4)
with open('mydata.bin', 'wb') as f:
f.write(struct.pack('1612i', *bytess))
def decode_and_prepare(register):
register = tf.io.decode_raw(register, out_type=tf.float32)
inputs = register[2:402]
label = tf.random.uniform(()) + register[402:]
return inputs, label
raw_dataset = tf.data.FixedLengthRecordDataset(filenames=['/content/mydata.bin']*7000, record_bytes=RAW_N*4)
raw_dataset = raw_dataset.map(decode_and_prepare)
训练模型没有预先缓存和迭代:
total_data_entries = len(list(raw_dataset.map(lambda x, y: (x, y))))
train_ds = raw_dataset.shuffle(buffer_size=total_data_entries).batch(32).prefetch(tf.data.AUTOTUNE)
inputs = tf.keras.layers.Input((400,))
x = tf.keras.layers.Dense(200, activation='relu', kernel_initializer='normal')(inputs)
x = tf.keras.layers.Dense(100, activation='relu', kernel_initializer='normal')(x)
outputs = tf.keras.layers.Dense(1, kernel_initializer='normal')(x)
model = tf.keras.Model(inputs, outputs)
model.compile(optimizer='adam', loss='mse')
model.fit(train_ds, epochs=5)
Epoch 1/5
875/875 [==============================] - 4s 3ms/step - loss: 0.1425
Epoch 2/5
875/875 [==============================] - 4s 3ms/step - loss: 0.0841
Epoch 3/5
875/875 [==============================] - 4s 3ms/step - loss: 0.0840
Epoch 4/5
875/875 [==============================] - 4s 3ms/step - loss: 0.0840
Epoch 5/5
875/875 [==============================] - 4s 3ms/step - loss: 0.0840
<keras.callbacks.History at 0x7fc41be037d0>
训练模型有缓存但是没有迭代:
total_data_entries = len(list(raw_dataset.map(lambda x, y: (x, y))))
train_ds = raw_dataset.shuffle(buffer_size=total_data_entries).cache().batch(32).prefetch(tf.data.AUTOTUNE)
inputs = tf.keras.layers.Input((400,))
x = tf.keras.layers.Dense(200, activation='relu', kernel_initializer='normal')(inputs)
x = tf.keras.layers.Dense(100, activation='relu', kernel_initializer='normal')(x)
outputs = tf.keras.layers.Dense(1, kernel_initializer='normal')(x)
model = tf.keras.Model(inputs, outputs)
model.compile(optimizer='adam', loss='mse')
model.fit(train_ds, epochs=5)
Epoch 1/5
875/875 [==============================] - 4s 2ms/step - loss: 0.1428
Epoch 2/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0841
Epoch 3/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0840
Epoch 4/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0840
Epoch 5/5
875/875 [==============================] - 2s 3ms/step - loss: 0.0840
<keras.callbacks.History at 0x7fc41fa87810>
训练模型with缓存和迭代:
total_data_entries = len(list(raw_dataset.map(lambda x, y: (x, y))))
train_ds = raw_dataset.shuffle(buffer_size=total_data_entries).cache().batch(32).prefetch(tf.data.AUTOTUNE)
_ = list(train_ds.as_numpy_iterator()) # iterate dataset beforehand
inputs = tf.keras.layers.Input((400,))
x = tf.keras.layers.Dense(200, activation='relu', kernel_initializer='normal')(inputs)
x = tf.keras.layers.Dense(100, activation='relu', kernel_initializer='normal')(x)
outputs = tf.keras.layers.Dense(1, kernel_initializer='normal')(x)
model = tf.keras.Model(inputs, outputs)
model.compile(optimizer='adam', loss='mse')
model.fit(train_ds, epochs=5)
Epoch 1/5
875/875 [==============================] - 3s 3ms/step - loss: 0.1427
Epoch 2/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0841
Epoch 3/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0840
Epoch 4/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0840
Epoch 5/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0840
<keras.callbacks.History at 0x7fc41ac9c850>
结论：数据集的缓存和之前的迭代似乎对训练有影响，但在这个例子中只使用了 7000 个文件。
我正在努力按照 Better performance with the tf.data API 指南提高我的模型训练性能。但是，我观察到使用 .cache() 的性能与没有 .cache().
datafile_list = load_my_files()
RAW_BYTES = 403*4
BATCH_SIZE = 32
raw_dataset = tf.data.FixedLengthRecordDataset(filenames=datafile_list, record_bytes=RAW_BYTES, num_parallel_reads=10, buffer_size=1024*RAW_BYTES)
raw_dataset = raw_dataset.map(tf.autograph.experimental.do_not_convert(decode_and_prepare),
num_parallel_calls=tf.data.AUTOTUNE)
raw_dataset = raw_dataset.cache()
raw_dataset = raw_dataset.shuffle(buffer_size=1024)
raw_dataset = raw_dataset.batch(BATCH_SIZE)
raw_dataset = raw_dataset.prefetch(tf.data.AUTOTUNE)
datafile_list 中的数据有 9.92GB，相当适合系统可用的总物理 RAM (100GB)。系统交换已禁用。
通过使用数据集训练模型：
model = build_model()
model.fit(raw_dataset, epochs=5, verbose=2)
结果：
Epoch 1/5
206247/206247 - 126s - loss: 0.0043 - mae: 0.0494 - mse: 0.0043
Epoch 2/5
206247/206247 - 125s - loss: 0.0029 - mae: 0.0415 - mse: 0.0029
Epoch 3/5
206247/206247 - 129s - loss: 0.0027 - mae: 0.0397 - mse: 0.0027
Epoch 4/5
206247/206247 - 125s - loss: 0.0025 - mae: 0.0386 - mse: 0.0025
Epoch 5/5
206247/206247 - 125s - loss: 0.0024 - mae: 0.0379 - mse: 0.0024
这个结果令人沮丧。由 docs:
The first time the dataset is iterated over, its elements will be cached either in the specified file or in memory. Subsequent iterations will use the cached data.
来自this guide：
When iterating over this dataset, the second iteration will be much faster than the first one thanks to the caching.
但是，所有epoch 所用的时间几乎相同。此外，在训练期间 CPU 和 GPU 使用率都非常低（见下图）。
通过注释掉 raw_dataset = raw_dataset.cache() 行，结果没有显示任何显着差异：
Epoch 1/5
206067/206067 - 129s - loss: 0.0042 - mae: 0.0492 - mse: 0.0042
Epoch 2/5
206067/206067 - 127s - loss: 0.0028 - mae: 0.0412 - mse: 0.0028
Epoch 3/5
206067/206067 - 134s - loss: 0.0026 - mae: 0.0393 - mse: 0.0026
Epoch 4/5
206067/206067 - 127s - loss: 0.0024 - mae: 0.0383 - mse: 0.0024
Epoch 5/5
206067/206067 - 126s - loss: 0.0023 - mae: 0.0376 - mse: 0.0023
正如文档中所指出的，我的期望是使用缓存会导致更快的训练时间。我想知道我做错了什么。
附件
训练期间使用缓存的 GPU 使用率：
无缓存训练期间的 GPU 使用率：
使用缓存训练期间的系统统计信息（内存，CPU 等）：
无缓存训练期间的系统统计信息（内存，CPU 等）：
只是使用 Google Colab 的一个小观察。根据 docs:
Note: For the cache to be finalized, the input dataset must be iterated through in its entirety. Otherwise, subsequent iterations will not use cached data.
和
Note: cache will produce exactly the same elements during each iteration through the dataset. If you wish to randomize the iteration order, make sure to call shuffle after calling cache.
我确实注意到在预先使用缓存和迭代数据集时存在一些差异。这是一个例子。
准备数据:
import random
import struct
import tensorflow as tf
import numpy as np
RAW_N = 2 + 20*20 + 1
bytess = random.sample(range(1, 5000), RAW_N*4)
with open('mydata.bin', 'wb') as f:
f.write(struct.pack('1612i', *bytess))
def decode_and_prepare(register):
register = tf.io.decode_raw(register, out_type=tf.float32)
inputs = register[2:402]
label = tf.random.uniform(()) + register[402:]
return inputs, label
raw_dataset = tf.data.FixedLengthRecordDataset(filenames=['/content/mydata.bin']*7000, record_bytes=RAW_N*4)
raw_dataset = raw_dataset.map(decode_and_prepare)
训练模型没有预先缓存和迭代:
total_data_entries = len(list(raw_dataset.map(lambda x, y: (x, y))))
train_ds = raw_dataset.shuffle(buffer_size=total_data_entries).batch(32).prefetch(tf.data.AUTOTUNE)
inputs = tf.keras.layers.Input((400,))
x = tf.keras.layers.Dense(200, activation='relu', kernel_initializer='normal')(inputs)
x = tf.keras.layers.Dense(100, activation='relu', kernel_initializer='normal')(x)
outputs = tf.keras.layers.Dense(1, kernel_initializer='normal')(x)
model = tf.keras.Model(inputs, outputs)
model.compile(optimizer='adam', loss='mse')
model.fit(train_ds, epochs=5)
Epoch 1/5
875/875 [==============================] - 4s 3ms/step - loss: 0.1425
Epoch 2/5
875/875 [==============================] - 4s 3ms/step - loss: 0.0841
Epoch 3/5
875/875 [==============================] - 4s 3ms/step - loss: 0.0840
Epoch 4/5
875/875 [==============================] - 4s 3ms/step - loss: 0.0840
Epoch 5/5
875/875 [==============================] - 4s 3ms/step - loss: 0.0840
<keras.callbacks.History at 0x7fc41be037d0>
训练模型有缓存但是没有迭代:
total_data_entries = len(list(raw_dataset.map(lambda x, y: (x, y))))
train_ds = raw_dataset.shuffle(buffer_size=total_data_entries).cache().batch(32).prefetch(tf.data.AUTOTUNE)
inputs = tf.keras.layers.Input((400,))
x = tf.keras.layers.Dense(200, activation='relu', kernel_initializer='normal')(inputs)
x = tf.keras.layers.Dense(100, activation='relu', kernel_initializer='normal')(x)
outputs = tf.keras.layers.Dense(1, kernel_initializer='normal')(x)
model = tf.keras.Model(inputs, outputs)
model.compile(optimizer='adam', loss='mse')
model.fit(train_ds, epochs=5)
Epoch 1/5
875/875 [==============================] - 4s 2ms/step - loss: 0.1428
Epoch 2/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0841
Epoch 3/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0840
Epoch 4/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0840
Epoch 5/5
875/875 [==============================] - 2s 3ms/step - loss: 0.0840
<keras.callbacks.History at 0x7fc41fa87810>
训练模型with缓存和迭代:
total_data_entries = len(list(raw_dataset.map(lambda x, y: (x, y))))
train_ds = raw_dataset.shuffle(buffer_size=total_data_entries).cache().batch(32).prefetch(tf.data.AUTOTUNE)
_ = list(train_ds.as_numpy_iterator()) # iterate dataset beforehand
inputs = tf.keras.layers.Input((400,))
x = tf.keras.layers.Dense(200, activation='relu', kernel_initializer='normal')(inputs)
x = tf.keras.layers.Dense(100, activation='relu', kernel_initializer='normal')(x)
outputs = tf.keras.layers.Dense(1, kernel_initializer='normal')(x)
model = tf.keras.Model(inputs, outputs)
model.compile(optimizer='adam', loss='mse')
model.fit(train_ds, epochs=5)
Epoch 1/5
875/875 [==============================] - 3s 3ms/step - loss: 0.1427
Epoch 2/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0841
Epoch 3/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0840
Epoch 4/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0840
Epoch 5/5
875/875 [==============================] - 2s 2ms/step - loss: 0.0840
<keras.callbacks.History at 0x7fc41ac9c850>
结论：数据集的缓存和之前的迭代似乎对训练有影响，但在这个例子中只使用了 7000 个文件。