lunny/keras
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Refactor layer benchmark (#333)

Browse Source 
* some benchmark setup

* fix something

* fix rnn benchmark

* Fix format

---------

Co-authored-by: chenmoneygithub <chenmoney@chenmoney-gpu-4.us-west1-a.c.keras-team-gcp.internal>
This commit is contained in:
Chen Qian 2023-06-12 16:02:14 -07:00 committed by Francois Chollet
parent dc8b81091c
commit 39bbd2a03d
12 changed files with 174 additions and 82 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
benchmarks/layer_benchmark/activation_benchmark.py
View File

@ -1,4 +1,4 @@
""" Benchmark activation layers.
"""Benchmark activation layers.
To run benchmarks, see the following command for an example, please change the
flag to your custom value:
@ -6,8 +6,8 @@ flag to your custom value:
```
python3 -m benchmarks.layer_benchmark.activation_benchmark \
--benchmark_name=benchmark_elu \
--num_samples=8192 \
--batch_size=1024 \
--num_samples=2048 \
--batch_size=256 \
--jit_compile=True
```
"""
@ -38,6 +38,10 @@ def benchmark_elu(
num_samples=num_samples,
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_prelu(
@ -58,6 +62,10 @@ def benchmark_prelu(
num_samples=num_samples,
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_relu(
@ -78,6 +86,10 @@ def benchmark_relu(
num_samples=num_samples,
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_leaky_relu(
@ -98,6 +110,10 @@ def benchmark_leaky_relu(
num_samples=num_samples,
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_softmax(
@ -118,6 +134,10 @@ def benchmark_softmax(
num_samples=num_samples,
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
BENCHMARK_NAMES = {
@ -136,7 +156,7 @@ def main(_):
jit_compile = FLAGS.jit_compile
if benchmark_name is None:
for name, benchmark_fn in BENCHMARK_NAMES:
for name, benchmark_fn in BENCHMARK_NAMES.items():
benchmark_fn(num_samples, batch_size, jit_compile)
return

8
benchmarks/layer_benchmark/attention_benchmark.py
View File

@ -1,4 +1,4 @@
""" Benchmark attention layers.
"""Benchmark attention layers.
To run benchmarks, see the following command for an example, please change the
flag to your custom value:
@ -6,8 +6,8 @@ flag to your custom value:
```
python3 -m benchmarks.layer_benchmark.attention_benchmark \
--benchmark_name=benchmark_attention \
--num_samples=8192 \
--batch_size=1024 \
--num_samples=2048 \
--batch_size=256 \
--jit_compile=True
```
"""
@ -115,7 +115,7 @@ def main(_):
jit_compile = FLAGS.jit_compile
if benchmark_name is None:
for name, benchmark_fn in BENCHMARK_NAMES:
for name, benchmark_fn in BENCHMARK_NAMES.items():
benchmark_fn(num_samples, batch_size, jit_compile)
return

19
benchmarks/layer_benchmark/base_benchmark.py
View File

@ -108,15 +108,28 @@ class LayerBenchmark:
input_shape,
flat_call_inputs=True,
jit_compile=True,
keras_core_layer=None,
tf_keras_layer=None,
):
self.layer_name = layer_name
self.input_shape = input_shape
_keras_core_layer_class = getattr(keras_core.layers, layer_name)
_tf_keras_layer_class = getattr(tf.keras.layers, layer_name)
self._keras_core_layer = _keras_core_layer_class(**init_args)
self._tf_keras_layer = _tf_keras_layer_class(**init_args)
if keras_core_layer is None:
# Sometimes you want to initialize the keras_core layer and tf_keras
# layer in a different way. For example, `Bidirectional` layer,
# which takes in `keras_core.layers.Layer` and
# `tf.keras.layer.Layer` separately.
self._keras_core_layer = _keras_core_layer_class(**init_args)
else:
self._keras_core_layer = keras_core_layer
if tf_keras_layer is None:
self._tf_keras_layer = _tf_keras_layer_class(**init_args)
else:
self._tf_keras_layer = tf_keras_layer
self.input_shape = input_shape
self._keras_core_model = self._build_keras_core_model(
input_shape, flat_call_inputs
)

20
benchmarks/layer_benchmark/conv_benchmark.py
View File

@ -1,4 +1,4 @@
""" Benchmark conv layers.
"""Benchmark conv layers.
To run benchmarks, see the following command for an example, please change the
flag to your custom value:
@ -6,8 +6,8 @@ flag to your custom value:
```
python3 -m benchmarks.layer_benchmark.conv_benchmark \
--benchmark_name=benchmark_conv2D \
--num_samples=2000 \
--batch_size=20 \
--num_samples=2046 \
--batch_size=256 \
--jit_compile=True
```
"""
@ -28,13 +28,13 @@ def benchmark_conv1D(
):
layer_name = "Conv1D"
init_args = {
"filters": 16,
"filters": 64,
"kernel_size": 2,
}
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[256, 16],
input_shape=[1024, 256],
jit_compile=jit_compile,
)
@ -118,7 +118,7 @@ def benchmark_depthwise_conv1D(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[32, 4],
input_shape=[256, 64],
jit_compile=jit_compile,
)
@ -175,7 +175,7 @@ def benchmark_separable_conv1D(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[32, 4],
input_shape=[256, 64],
jit_compile=jit_compile,
)
@ -226,13 +226,13 @@ def benchmark_conv1D_transpose(
):
layer_name = "Conv1DTranspose"
init_args = {
"filters": 16,
"kernel_size": 2,
"filters": 32,
"kernel_size": 4,
}
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[32, 4],
input_shape=[256, 256],
jit_compile=jit_compile,
)

25
benchmarks/layer_benchmark/core_benchmark.py
View File

@ -1,4 +1,4 @@
""" Benchmark core layers.
"""Benchmark core layers.
To run benchmarks, see the following command for an example, please change the
flag to your custom value:
@ -6,8 +6,8 @@ flag to your custom value:
```
python3 -m benchmarks.layer_benchmark.core_benchmark \
--benchmark_name=benchmark_dense \
--num_samples=8192 \
--batch_size=1024 \
--num_samples=2048 \
--batch_size=256 \
--jit_compile=True
```
"""
@ -27,7 +27,7 @@ def benchmark_dense(
jit_compile=True,
):
layer_name = "Dense"
init_args = {"units": 128}
init_args = {"units": 256}
benchmark = LayerBenchmark(
layer_name,
init_args,
@ -54,12 +54,12 @@ def benchmark_einsum_dense(
layer_name = "EinsumDense"
init_args = {
"equation": "abc,cd->abd",
"output_shape": (None, 128),
"output_shape": (None, 256),
}
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[64, 32],
input_shape=[256, 256],
jit_compile=jit_compile,
)
@ -81,17 +81,19 @@ def benchmark_embedding(
):
layer_name = "Embedding"
init_args = {
"input_dim": 30,
"output_shape": 128,
"input_dim": 128,
"output_dim": 256,
}
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[64, 32],
input_shape=[
256,
],
jit_compile=jit_compile,
)
data = np.random.randint(30, size=(num_samples, 32))
data = [np.random.randint(30, size=(num_samples, 256))]
benchmark.benchmark_predict(
num_samples=num_samples,
batch_size=batch_size,
@ -108,6 +110,7 @@ def benchmark_embedding(
BENCHMARK_NAMES = {
"benchmark_dense": benchmark_dense,
"benchmark_einsum_dense": benchmark_einsum_dense,
"benchmark_embedding": benchmark_embedding,
}
@ -118,7 +121,7 @@ def main(_):
jit_compile = FLAGS.jit_compile
if benchmark_name is None:
for name, benchmark_fn in BENCHMARK_NAMES:
for name, benchmark_fn in BENCHMARK_NAMES.items():
benchmark_fn(num_samples, batch_size, jit_compile)
return

52
benchmarks/layer_benchmark/merge_benchmark.py
View File

@ -1,4 +1,4 @@
""" Benchmark merge layers.
"""Benchmark merge layers.
To run benchmarks, see the following command for an example, please change the
flag to your custom value:
@ -6,8 +6,8 @@ flag to your custom value:
```
python3 -m benchmarks.layer_benchmark.merge_benchmark \
--benchmark_name=benchmark_add \
--num_samples=8192 \
--batch_size=1024 \
--num_samples=2048 \
--batch_size=256 \
--jit_compile=True
```
"""
@ -40,6 +40,11 @@ def benchmark_add(
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_average(
num_samples,
@ -61,6 +66,11 @@ def benchmark_average(
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_concatenate(
num_samples,
@ -82,6 +92,11 @@ def benchmark_concatenate(
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_dot(
num_samples,
@ -89,7 +104,7 @@ def benchmark_dot(
jit_compile=True,
):
layer_name = "Dot"
init_args = {}
init_args = {"axes": [2, 1]}
benchmark = LayerBenchmark(
layer_name,
init_args,
@ -103,6 +118,11 @@ def benchmark_dot(
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_maximum(
num_samples,
@ -124,6 +144,11 @@ def benchmark_maximum(
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_minimum(
num_samples,
@ -145,6 +170,11 @@ def benchmark_minimum(
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_multiply(
num_samples,
@ -156,7 +186,7 @@ def benchmark_multiply(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[[256, 256], [256, 32]],
input_shape=[[256, 64], [256, 64]],
flat_call_inputs=False,
jit_compile=jit_compile,
)
@ -166,6 +196,11 @@ def benchmark_multiply(
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
def benchmark_subtract(
num_samples,
@ -187,6 +222,11 @@ def benchmark_subtract(
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
BENCHMARK_NAMES = {
"benchmark_add": benchmark_add,
@ -207,7 +247,7 @@ def main(_):
jit_compile = FLAGS.jit_compile
if benchmark_name is None:
for name, benchmark_fn in BENCHMARK_NAMES:
for name, benchmark_fn in BENCHMARK_NAMES.items():
benchmark_fn(num_samples, batch_size, jit_compile)
return

17
benchmarks/layer_benchmark/normalization_benchmark.py
View File

@ -1,4 +1,4 @@
""" Benchmark normalization layers.
"""Benchmark normalization layers.
To run benchmarks, see the following command for an example, please change the
flag to your custom value:
@ -6,8 +6,8 @@ flag to your custom value:
```
python3 -m benchmarks.layer_benchmark.normalization_benchmark \
--benchmark_name=benchmark_batch_normalization \
--num_samples=2400 \
--batch_size=300 \
--num_samples=2048 \
--batch_size=256 \
--jit_compile=True
```
"""
@ -82,7 +82,7 @@ def benchmark_layer_normalization(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[256, 256, 4],
input_shape=[256, 128, 4],
jit_compile=jit_compile,
)
@ -107,7 +107,7 @@ def benchmark_unit_normalization(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[256, 256, 4],
input_shape=[256, 128, 4],
jit_compile=jit_compile,
)
@ -116,6 +116,11 @@ def benchmark_unit_normalization(
batch_size=batch_size,
)
benchmark.benchmark_train(
num_samples=num_samples,
batch_size=batch_size,
)
BENCHMARK_NAMES = {
"benchmark_batch_normalization": benchmark_batch_normalization,
@ -132,7 +137,7 @@ def main(_):
jit_compile = FLAGS.jit_compile
if benchmark_name is None:
for name, benchmark_fn in BENCHMARK_NAMES:
for name, benchmark_fn in BENCHMARK_NAMES.items():
benchmark_fn(num_samples, batch_size, jit_compile)
return

20
benchmarks/layer_benchmark/pooling_benchmark.py
View File

@ -1,4 +1,4 @@
""" Benchmark pooling layers.
"""Benchmark pooling layers.
To run benchmarks, see the following command for an example, please change the
flag to your custom value:
@ -33,7 +33,7 @@ def benchmark_average_pooling1d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[256, 3],
input_shape=[1024, 256],
jit_compile=jit_compile,
)
@ -87,7 +87,7 @@ def benchmark_average_pooling3d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[64, 64, 64, 3],
input_shape=[64, 64, 32, 3],
jit_compile=jit_compile,
)
@ -114,7 +114,7 @@ def benchmark_max_pooling1d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[256, 3],
input_shape=[1024, 256],
jit_compile=jit_compile,
)
@ -168,7 +168,7 @@ def benchmark_max_pooling3d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[64, 64, 64, 3],
input_shape=[64, 64, 32, 3],
jit_compile=jit_compile,
)
@ -193,7 +193,7 @@ def benchmark_global_average_pooling1d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[256, 3],
input_shape=[1024, 256],
jit_compile=jit_compile,
)
@ -243,7 +243,7 @@ def benchmark_global_average_pooling3d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[64, 64, 64, 3],
input_shape=[64, 64, 32, 3],
jit_compile=jit_compile,
)
@ -268,7 +268,7 @@ def benchmark_global_max_pooling1d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[256, 3],
input_shape=[1024, 256],
jit_compile=jit_compile,
)
@ -318,7 +318,7 @@ def benchmark_global_max_pooling3d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[64, 64, 64, 3],
input_shape=[64, 64, 32, 3],
jit_compile=jit_compile,
)
@ -356,7 +356,7 @@ def main(_):
jit_compile = FLAGS.jit_compile
if benchmark_name is None:
for name, benchmark_fn in BENCHMARK_NAMES:
for name, benchmark_fn in BENCHMARK_NAMES.items():
benchmark_fn(num_samples, batch_size, jit_compile)
return

8
benchmarks/layer_benchmark/regularization_benchmark.py
View File

@ -1,4 +1,4 @@
""" Benchmark regularization layers.
"""Benchmark regularization layers.
To run benchmarks, see the following command for an example, please change the
flag to your custom value:
@ -52,7 +52,7 @@ def benchmark_gaussian_dropout(
batch_size,
jit_compile=True,
):
layer_name = "GaussionDropout"
layer_name = "GaussianDropout"
init_args = {
"rate": 0.5,
}
@ -79,7 +79,7 @@ def benchmark_gaussian_noise(
batch_size,
jit_compile=True,
):
layer_name = "GaussionNoise"
layer_name = "GaussianNoise"
init_args = {
"stddev": 0.5,
}
@ -199,7 +199,7 @@ def main(_):
jit_compile = FLAGS.jit_compile
if benchmark_name is None:
for name, benchmark_fn in BENCHMARK_NAMES:
for name, benchmark_fn in BENCHMARK_NAMES.items():
benchmark_fn(num_samples, batch_size, jit_compile)
return

12
benchmarks/layer_benchmark/reshaping_benchmark.py
View File

@ -1,4 +1,4 @@
""" Benchmark reshaping layers.
"""Benchmark reshaping layers.
To run benchmarks, see the following command for an example, please change the
flag to your custom value:
@ -30,7 +30,7 @@ def benchmark_cropping1d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[256, 3],
input_shape=[1024, 256],
jit_compile=jit_compile,
)
@ -127,7 +127,7 @@ def benchmark_permute(
):
layer_name = "Permute"
init_args = {
"dim": (2, 1),
"dims": (2, 1),
}
benchmark = LayerBenchmark(
layer_name,
@ -182,7 +182,7 @@ def benchmark_up_sampling2d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[256, 256, 3],
input_shape=[128, 128, 3],
jit_compile=jit_compile,
)
@ -207,7 +207,7 @@ def benchmark_up_sampling3d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[32, 32, 32, 3],
input_shape=[32, 16, 16, 3],
jit_compile=jit_compile,
)
@ -319,7 +319,7 @@ def main(_):
jit_compile = FLAGS.jit_compile
if benchmark_name is None:
for name, benchmark_fn in BENCHMARK_NAMES:
for name, benchmark_fn in BENCHMARK_NAMES.items():
benchmark_fn(num_samples, batch_size, jit_compile)
return

43
benchmarks/layer_benchmark/rnn_benchmark.py
View File

@ -1,4 +1,4 @@
""" Benchmark rnn layers.
"""Benchmark rnn layers.
To run benchmarks, see the following command for an example, please change the
flag to your custom value:
@ -12,6 +12,7 @@ python3 -m benchmarks.layer_benchmark.rnn_benchmark \
```
"""
import tensorflow as tf
from absl import app
from absl import flags
@ -62,7 +63,7 @@ def benchmark_conv_lstm2d(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[32, 64, 64, 3],
input_shape=[32, 32, 32, 3],
jit_compile=jit_compile,
)
@ -84,13 +85,13 @@ def benchmark_conv_lstm3d(
):
layer_name = "ConvLSTM3D"
init_args = {
"filters": 16,
"filters": 8,
"kernel_size": 2,
}
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[16, 32, 32, 16, 3],
input_shape=[8, 16, 16, 16, 3],
jit_compile=jit_compile,
)
@ -117,7 +118,7 @@ def benchmark_gru(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[32, 256],
input_shape=[256, 256],
jit_compile=jit_compile,
)
@ -144,7 +145,7 @@ def benchmark_lstm(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[32, 256],
input_shape=[256, 256],
jit_compile=jit_compile,
)
@ -171,7 +172,7 @@ def benchmark_simple_rnn(
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[32, 256],
input_shape=[256, 256],
jit_compile=jit_compile,
)
@ -192,14 +193,18 @@ def benchmark_bidirectional(
jit_compile=True,
):
layer_name = "Bidirectional"
init_args = {
"layer": keras_core.layers.LSTM(32),
}
init_args = {}
keras_core_layer = keras_core.layers.Bidirectional(
keras_core.layers.LSTM(32)
)
tf_keras_layer = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(32))
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[32, 256],
input_shape=[256, 256],
jit_compile=jit_compile,
keras_core_layer=keras_core_layer,
tf_keras_layer=tf_keras_layer,
)
benchmark.benchmark_predict(
@ -219,14 +224,20 @@ def benchmark_time_distributed(
jit_compile=True,
):
layer_name = "TimeDistributed"
init_args = {
"layer": keras_core.layers.Conv2D(64, (3, 3)),
}
init_args = {}
keras_core_layer = keras_core.layers.TimeDistributed(
keras_core.layers.Conv2D(16, (3, 3))
)
tf_keras_layer = tf.keras.layers.TimeDistributed(
tf.keras.layers.Conv2D(16, (3, 3))
)
benchmark = LayerBenchmark(
layer_name,
init_args,
input_shape=[10, 128, 128, 3],
input_shape=[10, 32, 32, 3],
jit_compile=jit_compile,
keras_core_layer=keras_core_layer,
tf_keras_layer=tf_keras_layer,
)
benchmark.benchmark_predict(
@ -259,7 +270,7 @@ def main(_):
jit_compile = FLAGS.jit_compile
if benchmark_name is None:
for name, benchmark_fn in BENCHMARK_NAMES:
for name, benchmark_fn in BENCHMARK_NAMES.items():
benchmark_fn(num_samples, batch_size, jit_compile)
return

4
keras_core/layers/reshaping/zero_padding1d.py
View File

@ -56,8 +56,8 @@ class ZeroPadding1D(Layer):
def compute_output_shape(self, input_shape):
output_shape = list(input_shape)
if input_shape[1] is not None:
input_shape[1] += self.padding[0] + self.padding[1]
if output_shape[1] is not None:
output_shape[1] += self.padding[0] + self.padding[1]
return tuple(output_shape)
def call(self, inputs):