TF.Keras model.predict is slower than straight Numpy?
Question:
Thanks, everyone for trying to help me understand the issue below. I have updated the question and produced a CPU-only run and GPU-only of the run. In general, it also appears that in either case a direct numpy calculation hundreds of times faster than the model. predict(). Hopefully, this clarifies that this does not appear to be a CPU vs GPU issue (if it is, I would love an explanation).
Let’s create a trained model with keras.
import tensorflow as tf
(X,Y),(Xt,Yt) = tf.keras.datasets.mnist.load_data()
model = tf.keras.models.Sequential([
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(1000,'relu'),
tf.keras.layers.Dense(100,'relu'),
tf.keras.layers.Dense(10,'softmax'),
])
model.compile('adam','sparse_categorical_crossentropy')
model.fit(X,Y,epochs=20,batch_size=1024)
Now let’s re-create the model.predict function using numpy.
import numpy as np
W = model.get_weights()
def predict(X):
X = X.reshape((X.shape[0],-1)) #Flatten
X = X @ W[0] + W[1] #Dense
X[X<0] = 0 #Relu
X = X @ W[2] + W[3] #Dense
X[X<0] = 0 #Relu
X = X @ W[4] + W[5] #Dense
X = np.exp(X)/np.exp(X).sum(1)[...,None] #Softmax
return X
We can easily verify these are the same function (module machine errors in implementation).
print(model.predict(X[:100]).argmax(1))
print(predict(X[:100]).argmax(1))
We can also test out how fast these functions run. Using ipython:
%timeit model.predict(X[:10]).argmax(1) # 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # 1000 loops takes 356 µs
I get that predict runs about 10,000 times faster than model. predict at low batches and reduces to around 100 times faster at larger batches. Regardless, why is predict so much faster? In fact, predict isn’t even optimized, we could use numba, or even straight re-write predict in C code and compile it.
Thinking in terms of deployment purposes, why would manually extracting the weights from the model and re-writing the function be thousands of times faster than what keras does internally? This also means that writing a script to utilize a .h5 file or similar, maybe much slower than manually re-writing the prediction function. In general, is this true?
Ipython Output (CPU):
Python 3.8.5 (default, Sep 3 2020, 21:29:08) [MSC v.1916 64 bit (AMD64)]
Type 'copyright', 'credits' or 'license' for more information
IPython 7.19.0 -- An enhanced Interactive Python. Type '?' for help.
PyDev console: using IPython 7.19.0
Python 3.8.5 (default, Sep 3 2020, 21:29:08) [MSC v.1916 64 bit (AMD64)] on win32
import os
os.environ["CUDA_VISIBLE_DEVICES"]="-1"
import tensorflow as tf
(X,Y),(Xt,Yt) = tf.keras.datasets.mnist.load_data()
model = tf.keras.models.Sequential([
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(1000,'relu'),
tf.keras.layers.Dense(100,'relu'),
tf.keras.layers.Dense(10,'softmax'),
])
model.compile('adam','sparse_categorical_crossentropy')
model.fit(X,Y,epochs=20,batch_size=1024)
2021-04-19 15:10:58.323137: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library cudart64_110.dll
2021-04-19 15:11:01.990590: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library nvcuda.dll
2021-04-19 15:11:02.039285: E tensorflow/stream_executor/cuda/cuda_driver.cc:328] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected
2021-04-19 15:11:02.042553: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:169] retrieving CUDA diagnostic information for host: DESKTOP-G0U8S3P
2021-04-19 15:11:02.043134: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:176] hostname: DESKTOP-G0U8S3P
2021-04-19 15:11:02.128834: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:127] None of the MLIR optimization passes are enabled (registered 2)
Epoch 1/20
59/59 [==============================] - 4s 60ms/step - loss: 35.3708
Epoch 2/20
59/59 [==============================] - 3s 58ms/step - loss: 0.8671
Epoch 3/20
59/59 [==============================] - 3s 56ms/step - loss: 0.5641
Epoch 4/20
59/59 [==============================] - 3s 56ms/step - loss: 0.4359
Epoch 5/20
59/59 [==============================] - 3s 56ms/step - loss: 0.3447
Epoch 6/20
59/59 [==============================] - 3s 56ms/step - loss: 0.2891
Epoch 7/20
59/59 [==============================] - 3s 56ms/step - loss: 0.2371
Epoch 8/20
59/59 [==============================] - 3s 57ms/step - loss: 0.1977
Epoch 9/20
59/59 [==============================] - 3s 57ms/step - loss: 0.1713
Epoch 10/20
59/59 [==============================] - 3s 57ms/step - loss: 0.1381
Epoch 11/20
59/59 [==============================] - 4s 61ms/step - loss: 0.1203
Epoch 12/20
59/59 [==============================] - 3s 57ms/step - loss: 0.1095
Epoch 13/20
59/59 [==============================] - 3s 56ms/step - loss: 0.0877
Epoch 14/20
59/59 [==============================] - 3s 57ms/step - loss: 0.0793
Epoch 15/20
59/59 [==============================] - 3s 56ms/step - loss: 0.0727
Epoch 16/20
59/59 [==============================] - 3s 56ms/step - loss: 0.0702
Epoch 17/20
59/59 [==============================] - 3s 56ms/step - loss: 0.0701
Epoch 18/20
59/59 [==============================] - 3s 57ms/step - loss: 0.0631
Epoch 19/20
59/59 [==============================] - 3s 56ms/step - loss: 0.0539
Epoch 20/20
59/59 [==============================] - 3s 58ms/step - loss: 0.0493
Out[3]: <tensorflow.python.keras.callbacks.History at 0x143069fdf40>
import numpy as np
W = model.get_weights()
def predict(X):
X = X.reshape((X.shape[0],-1)) #Flatten
X = X @ W[0] + W[1] #Dense
X[X<0] = 0 #Relu
X = X @ W[2] + W[3] #Dense
X[X<0] = 0 #Relu
X = X @ W[4] + W[5] #Dense
X = np.exp(X)/np.exp(X).sum(1)[...,None] #Softmax
return X
%timeit model.predict(X[:10]).argmax(1) # 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # 1000 loops takes 356 µs
52.8 ms ± 2.13 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
640 µs ± 10.9 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
Ipython Output (GPU):
Python 3.7.7 (default, Mar 26 2020, 15:48:22)
Type 'copyright', 'credits' or 'license' for more information
IPython 7.4.0 -- An enhanced Interactive Python. Type '?' for help.
In [1]: import tensorflow as tf
...:
...: (X,Y),(Xt,Yt) = tf.keras.datasets.mnist.load_data()
...:
...: model = tf.keras.models.Sequential([
...: tf.keras.layers.Flatten(),
...: tf.keras.layers.Dense(1000,'relu'),
...: tf.keras.layers.Dense(100,'relu'),
...: tf.keras.layers.Dense(10,'softmax'),
...: ])
...: model.compile('adam','sparse_categorical_crossentropy')
...: model.fit(X,Y,epochs=20,batch_size=1024)
2020-07-01 15:50:46.008518: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcuda.so.1
2020-07-01 15:50:46.054495: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1618] Found device 0 with properties:
name: GeForce RTX 2080 Ti major: 7 minor: 5 memoryClockRate(GHz): 1.545
pciBusID: 0000:05:00.0
2020-07-01 15:50:46.059582: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.0
2020-07-01 15:50:46.114562: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0
2020-07-01 15:50:46.142058: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcufft.so.10.0
2020-07-01 15:50:46.152899: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcurand.so.10.0
2020-07-01 15:50:46.217725: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcusolver.so.10.0
2020-07-01 15:50:46.260758: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcusparse.so.10.0
2020-07-01 15:50:46.374328: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudnn.so.7
2020-07-01 15:50:46.376747: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1746] Adding visible gpu devices: 0
2020-07-01 15:50:46.377688: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX FMA
2020-07-01 15:50:46.433422: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 4018875000 Hz
2020-07-01 15:50:46.434383: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x563e4d0d71c0 executing computations on platform Host. Devices:
2020-07-01 15:50:46.435119: I tensorflow/compiler/xla/service/service.cc:175] StreamExecutor device (0): Host, Default Version
2020-07-01 15:50:46.596077: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x563e4a9379f0 executing computations on platform CUDA. Devices:
2020-07-01 15:50:46.596119: I tensorflow/compiler/xla/service/service.cc:175] StreamExecutor device (0): GeForce RTX 2080 Ti, Compute Capability 7.5
2020-07-01 15:50:46.597894: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1618] Found device 0 with properties:
name: GeForce RTX 2080 Ti major: 7 minor: 5 memoryClockRate(GHz): 1.545
pciBusID: 0000:05:00.0
2020-07-01 15:50:46.597961: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.0
2020-07-01 15:50:46.597988: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0
2020-07-01 15:50:46.598014: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcufft.so.10.0
2020-07-01 15:50:46.598040: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcurand.so.10.0
2020-07-01 15:50:46.598065: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcusolver.so.10.0
2020-07-01 15:50:46.598090: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcusparse.so.10.0
2020-07-01 15:50:46.598115: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudnn.so.7
2020-07-01 15:50:46.599766: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1746] Adding visible gpu devices: 0
2020-07-01 15:50:46.600611: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.0
2020-07-01 15:50:46.603713: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1159] Device interconnect StreamExecutor with strength 1 edge matrix:
2020-07-01 15:50:46.603751: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1165] 0
2020-07-01 15:50:46.603763: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1178] 0: N
2020-07-01 15:50:46.605917: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1304] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 10311 MB memory) -> physical GPU (device: 0, name: GeForce RTX 2080 Ti, pci bus id: 0000:05:00.0, compute capability: 7.5)
Train on 60000 samples
Epoch 1/20
2020-07-01 15:50:49.995091: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0
60000/60000 [==============================] - 2s 26us/sample - loss: 9.9370
Epoch 2/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.6094
Epoch 3/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.3672
Epoch 4/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.2720
Epoch 5/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.2196
Epoch 6/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.1673
Epoch 7/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.1367
Epoch 8/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.1082
Epoch 9/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0895
Epoch 10/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0781
Epoch 11/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0666
Epoch 12/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0537
Epoch 13/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0459
Epoch 14/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0412
Epoch 15/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0401
Epoch 16/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0318
Epoch 17/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0275
Epoch 18/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0237
Epoch 19/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0212
Epoch 20/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0199
Out[1]: <tensorflow.python.keras.callbacks.History at 0x7f7c9000b550>
In [2]: import numpy as np
...:
...: W = model.get_weights()
...:
...: def predict(X):
...: X = X.reshape((X.shape[0],-1)) #Flatten
...: X = X @ W[0] + W[1] #Dense
...: X[X<0] = 0 #Relu
...: X = X @ W[2] + W[3] #Dense
...: X[X<0] = 0 #Relu
...: X = X @ W[4] + W[5] #Dense
...: X = np.exp(X)/np.exp(X).sum(1)[...,None] #Softmax
...: return X
...:
In [3]: print(model.predict(X[:100]).argmax(1))
...: print(predict(X[:100]).argmax(1))
[5 0 4 1 9 2 1 3 1 4 3 5 3 6 1 7 2 8 6 9 4 0 9 1 1 2 4 3 2 7 3 8 6 9 0 5 6
0 7 6 1 8 7 9 3 9 8 5 9 3 3 0 7 4 9 8 0 9 4 1 4 4 6 0 4 5 6 1 0 0 1 7 1 6
3 0 2 1 1 7 5 0 2 6 7 8 3 9 0 4 6 7 4 6 8 0 7 8 3 1]
/home/bobbyocean/anaconda3/bin/ipython3:12: RuntimeWarning: overflow encountered in exp
/home/bobbyocean/anaconda3/bin/ipython3:12: RuntimeWarning: invalid value encountered in true_divide
[5 0 4 1 9 2 1 3 1 4 3 5 3 6 1 7 2 8 6 9 4 0 9 1 1 2 4 3 2 7 3 8 6 9 0 5 6
0 7 6 1 8 7 9 3 9 8 5 9 3 3 0 7 4 9 8 0 9 4 1 4 4 6 0 4 5 6 1 0 0 1 7 1 6
3 0 2 1 1 7 5 0 2 6 7 8 3 9 0 4 6 7 4 6 8 0 7 8 3 1]
In [4]: %timeit model.predict(X[:10]).argmax(1)
37.7 ms ± 806 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [5]: %timeit predict(X[:10]).argmax(1)
361 µs ± 13.8 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
Answers:
As others have pointed out, the Tensorflow binary in question was compiled for GPU optimization: while GPUs are excellent at intensive number-crunching because of the extremely high number of computing cores they have, they’re painfully slow when it comes to moving data back and forth.
When a model is executing on a graphics card, all the necessary data has to be burst over to the GPU — it has no access to the host system’s RAM (nor does the host system have access to the video memory). Once the GPU is finished processing, any results have to be shipped back over to the host system.
All of this moving around of data takes a great deal of time; moreover, a Tensorflow binary compiled to execute with a GPU/CUDA does not, to my knowledge, include any of the standard optimizations for executing on a CPU (like using faster extended instruction sets, e.g. AVX, AVX2, etc).
So you’re comparing a highly CPU-optimized scientific library which can process data without even having to go back to RAM half the time (CPU registers and cache storage on the chip); to code that has to collect every last bit it’s going to need before having to ship all of that data to the graphics card, and back. I’m also leaving out all of the data manipulations that goes on under the hood of Tensorflow. It works on its own data structures, after all.
Eager execution is also another layer of inefficiency, I suppose.
As for best practices in deploying Keras models, I’m of the opinion that it’s like everything else in software: premature optimization is the root of all evil. If you don’t need it to be fast and lean, then let it be slow, modular, reusable, and intuitive. But hey, if you need or want the efficiency, then power to you. Keras is designed for rapid development and research, not production code.
In short, the answer is that it’s for the same reason C++ is faster than Python (because the Python interpreter has so much more overhead, as does Tensorflow).
We observe that the main issue is the cause of the Eager Execution mode. We give shallow look at your code and corresponding results as per CPU and GPU bases. It is true that numpy doesn’t operate on GPU, so unlike tf-gpu, it doesn’t encounter any data shifting overhead.
But also it’s quite noticeable how much fast computation is done by your defined predict method with np compare to model. predict with tf. keras, whereas the input test set is 10 samples only. However, We’re not giving any deep analysis, like one piece of art here you may love to read.
My Setup is as follows. I’m using the Colab environment and checking with both CPU and GPU mode.
TensorFlow 1.15.2
Keras 2.3.1
Numpy 1.19.5
TensorFlow 2.4.1
Keras 2.4.0
Numpy 1.19.5
TF 1.15.2 – CPU
%tensorflow_version 1.x
import os
os.environ["CUDA_VISIBLE_DEVICES"]="-1"
import tensorflow as tf
from tensorflow.python.client import device_lib
print(tf.__version__)
print('A: ', tf.test.is_built_with_cuda)
print('B: ', tf.test.gpu_device_name())
local_device_protos = device_lib.list_local_devices()
([x.name for x in local_device_protos if x.device_type == 'GPU'],
[x.name for x in local_device_protos if x.device_type == 'CPU'])
TensorFlow 1.x selected.
1.15.2
A: <function is_built_with_cuda at 0x7f122d58dcb0>
B:
([], ['/device:CPU:0'])
Now, running your code.
import tensorflow as tf
import keras
print(tf.executing_eagerly()) # False
(X,Y),(Xt,Yt) = keras.datasets.mnist.load_data()
model = keras.models.Sequential([])
model.compile
model.fit
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
1000 loops, best of 5: 1.07 ms per loop
1000 loops, best of 5: 1.48 ms per loop
We can see that the execution times are comparable with old keras. Now, let’s test with GPU also.
TF 1.15.2 – GPU
%tensorflow_version 1.x
import os
os.environ["CUDA_VISIBLE_DEVICES"]="0"
import tensorflow as tf
from tensorflow.python.client import device_lib
print(tf.__version__)
print('A: ', tf.test.is_built_with_cuda)
print('B: ', tf.test.gpu_device_name())
local_device_protos = device_lib.list_local_devices()
([x.name for x in local_device_protos if x.device_type == 'GPU'],
[x.name for x in local_device_protos if x.device_type == 'CPU'])
1.15.2
A: <function is_built_with_cuda at 0x7f0b5ad46830>
B: /device:GPU:0
(['/device:GPU:0'], ['/device:CPU:0'])
...
...
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
1000 loops, best of 5: 1.02 ms per loop
1000 loops, best of 5: 1.44 ms per loop
Now, the execution time is also comparable here with old keras and no eager mode. Let’s now see the new tf. keras with eager mode first and then we observe without eager mode.
TF 2.4.1 – CPU
Eagerly
import os
os.environ["CUDA_VISIBLE_DEVICES"]="-1"
import tensorflow as tf
from tensorflow.python.client import device_lib
print(tf.__version__)
print('A: ', tf.test.is_built_with_cuda)
print('B: ', tf.test.gpu_device_name())
local_device_protos = device_lib.list_local_devices()
([x.name for x in local_device_protos if x.device_type == 'GPU'],
[x.name for x in local_device_protos if x.device_type == 'CPU'])
2.4.1
A: <function is_built_with_cuda at 0x7fed85de3560>
B:
([], ['/device:CPU:0'])
Now, running the code with eager mode.
import tensorflow as tf
import keras
print(tf.executing_eagerly()) # True
(X,Y),(Xt,Yt) = keras.datasets.mnist.load_data()
model = keras.models.Sequential([ ])
model.compile
model.fit
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
10 loops, best of 5: 28 ms per loop
1000 loops, best of 5: 1.73 ms per loop
Disable Eagerly
Now, if we disable the eager mode and run the same code as follows then we will get:
import tensorflow as tf
import keras
# # Disables eager execution
tf.compat.v1.disable_eager_execution()
# or,
# Disables eager execution of tf.functions.
# tf.config.run_functions_eagerly(False)
print(tf.executing_eagerly())
False
(X,Y),(Xt,Yt) = keras.datasets.mnist.load_data()
model = keras.models.Sequential([])
model.compile
model.fit
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
1000 loops, best of 5: 1.37 ms per loop
1000 loops, best of 5: 1.57 ms per loop
Now, we can see the execution times are comparable for disabling the eager mode in new tf. keras. Now, let’s test with GPU mode also.
TF 2.4.1 – GPU
Eagerly
import os
os.environ["CUDA_VISIBLE_DEVICES"]="0"
import tensorflow as tf
from tensorflow.python.client import device_lib
print(tf.__version__)
print('A: ', tf.test.is_built_with_cuda)
print('B: ', tf.test.gpu_device_name())
local_device_protos = device_lib.list_local_devices()
([x.name for x in local_device_protos if x.device_type == 'GPU'],
[x.name for x in local_device_protos if x.device_type == 'CPU'])
2.4.1
A: <function is_built_with_cuda at 0x7f16ad88f680>
B: /device:GPU:0
(['/device:GPU:0'], ['/device:CPU:0'])
import tensorflow as tf
import keras
print(tf.executing_eagerly()) # True
(X,Y),(Xt,Yt) = keras.datasets.mnist.load_data()
model = keras.models.Sequential([ ])
model.compile
model.fit
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
10 loops, best of 5: 26.3 ms per loop
1000 loops, best of 5: 1.48 ms per loop
Disable Eagerly
And lastly again, if we disable the eager mode and run the same code as follows, we will get:
# Disables eager execution
tf.compat.v1.disable_eager_execution()
# or,
# Disables eager execution of tf.functions.
# tf.config.run_functions_eagerly(False)
print(tf.executing_eagerly()) # False
(X,Y),(Xt,Yt) = keras.datasets.mnist.load_data()
model = keras.models.Sequential([ ])
model.compile
model.fit
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
1000 loops, best of 5: 1.12 ms per loop
1000 loops, best of 5: 1.45 ms per loop
And like before, the execution times are comparable with the non-eager mode in new tf. keras. That’s why, the Eager mode is the root cause of the slower performance of tf. keras than straight numpy.
Another answer is more useful in terms of "how to make tf keras predict faster", but I think the following can help more in terms of "what is it doing that takes so much time"? Even with eager mode disabled, you might be curious to see how execution looks like (e.g. with or without providing batch_size, etc.).
To answer this question you may find tracing profiler to be useful. Tracing execution is adding a lot of overhead (especially for the places that have a bunch of very lightweight python calls), but overall should give you quite a bit of insight into what part of python code is being executed, because, well, it just logs exactly what is happening. You can try pytracing since it produces files that Chrome browser visualizes nicely on it’s built-in chrome://tracing page. To use it e.g. in google colab you can do the following:
First, intstall pytracing:
!pip install pytracing
Then generate trace:
from pytracing import TraceProfiler
tp = TraceProfiler(output=open('/root/trace.out', 'wt'))
with tp.traced():
for i in range(2):
model.predict(X[:1000], batch_size=1000)
Then download trace:
from google.colab import files
files.download('/root/trace.out')
After this in Chrome browser open chrome://tracing page, click "Load" button, and select trace.out file, you’ve downloaded.
You’ll see something like the following – you can click on any element, see the full name of the python function and file it is from + wall time it took (again, all of this is higher than in normal runs due to tracing overhead):
You can see how disabling/enabling eager execution or changing the batch size will change the output and can see for yourself what is taking the most time. From what I currently see (in non-eagerly mode + call like model.predict(X[:1000], batch_size=1000)) quite a bit of time is spent on:
Standardizing your data (whatever it means): ~2.5ms (including tracing overhead!):
/usr/local/lib/python3.7/dist-packages/tensorflow/python/keras/engine/training_v1.py:2336:_standardize_user_data
Preparing callbacks (even though we didn’t set any): ~2ms (including tracing overhead)
/usr/local/lib/python3.7/dist-packages/tensorflow/python/keras/callbacks.py:133:configure_callbacks
As to the statement that numpy version is not optimized – I wouldn’t agree. The numpy implementation here is quite optimized – python is not making any pure python calls in it (execution of predict only results call to functions in C – I couldn’t believe it at first but it seems to be the case), so overhead from Python is really minimal. You might gain a little bit by optimizing the way you do ReLU and eliminating extra allocation/deallocation, but that could only result in a very minor perf improvement.
Instead of using model.predict(input), simply try model(input)
To have a fast predict function on small batches, a simple solution is to do this:
predict = lambda x: tf.function(model(x))
# Use predict(x) instead of model.predict(x)
If you want to use numpy arrays, you can do this:
predict_np = lambda x: predict(tf.convert_to_tensor(x))
I have a speed up of x5 and more on CPU. The performance is equivalent to the performance achieved with eager execution disabled, but here it doesn’t affect the whole code.
Thanks, everyone for trying to help me understand the issue below. I have updated the question and produced a CPU-only run and GPU-only of the run. In general, it also appears that in either case a direct numpy calculation hundreds of times faster than the model. predict(). Hopefully, this clarifies that this does not appear to be a CPU vs GPU issue (if it is, I would love an explanation).
Let’s create a trained model with keras.
import tensorflow as tf
(X,Y),(Xt,Yt) = tf.keras.datasets.mnist.load_data()
model = tf.keras.models.Sequential([
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(1000,'relu'),
tf.keras.layers.Dense(100,'relu'),
tf.keras.layers.Dense(10,'softmax'),
])
model.compile('adam','sparse_categorical_crossentropy')
model.fit(X,Y,epochs=20,batch_size=1024)
Now let’s re-create the model.predict function using numpy.
import numpy as np
W = model.get_weights()
def predict(X):
X = X.reshape((X.shape[0],-1)) #Flatten
X = X @ W[0] + W[1] #Dense
X[X<0] = 0 #Relu
X = X @ W[2] + W[3] #Dense
X[X<0] = 0 #Relu
X = X @ W[4] + W[5] #Dense
X = np.exp(X)/np.exp(X).sum(1)[...,None] #Softmax
return X
We can easily verify these are the same function (module machine errors in implementation).
print(model.predict(X[:100]).argmax(1))
print(predict(X[:100]).argmax(1))
We can also test out how fast these functions run. Using ipython:
%timeit model.predict(X[:10]).argmax(1) # 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # 1000 loops takes 356 µs
I get that predict runs about 10,000 times faster than model. predict at low batches and reduces to around 100 times faster at larger batches. Regardless, why is predict so much faster? In fact, predict isn’t even optimized, we could use numba, or even straight re-write predict in C code and compile it.
Thinking in terms of deployment purposes, why would manually extracting the weights from the model and re-writing the function be thousands of times faster than what keras does internally? This also means that writing a script to utilize a .h5 file or similar, maybe much slower than manually re-writing the prediction function. In general, is this true?
Ipython Output (CPU):
Python 3.8.5 (default, Sep 3 2020, 21:29:08) [MSC v.1916 64 bit (AMD64)]
Type 'copyright', 'credits' or 'license' for more information
IPython 7.19.0 -- An enhanced Interactive Python. Type '?' for help.
PyDev console: using IPython 7.19.0
Python 3.8.5 (default, Sep 3 2020, 21:29:08) [MSC v.1916 64 bit (AMD64)] on win32
import os
os.environ["CUDA_VISIBLE_DEVICES"]="-1"
import tensorflow as tf
(X,Y),(Xt,Yt) = tf.keras.datasets.mnist.load_data()
model = tf.keras.models.Sequential([
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(1000,'relu'),
tf.keras.layers.Dense(100,'relu'),
tf.keras.layers.Dense(10,'softmax'),
])
model.compile('adam','sparse_categorical_crossentropy')
model.fit(X,Y,epochs=20,batch_size=1024)
2021-04-19 15:10:58.323137: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library cudart64_110.dll
2021-04-19 15:11:01.990590: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library nvcuda.dll
2021-04-19 15:11:02.039285: E tensorflow/stream_executor/cuda/cuda_driver.cc:328] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected
2021-04-19 15:11:02.042553: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:169] retrieving CUDA diagnostic information for host: DESKTOP-G0U8S3P
2021-04-19 15:11:02.043134: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:176] hostname: DESKTOP-G0U8S3P
2021-04-19 15:11:02.128834: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:127] None of the MLIR optimization passes are enabled (registered 2)
Epoch 1/20
59/59 [==============================] - 4s 60ms/step - loss: 35.3708
Epoch 2/20
59/59 [==============================] - 3s 58ms/step - loss: 0.8671
Epoch 3/20
59/59 [==============================] - 3s 56ms/step - loss: 0.5641
Epoch 4/20
59/59 [==============================] - 3s 56ms/step - loss: 0.4359
Epoch 5/20
59/59 [==============================] - 3s 56ms/step - loss: 0.3447
Epoch 6/20
59/59 [==============================] - 3s 56ms/step - loss: 0.2891
Epoch 7/20
59/59 [==============================] - 3s 56ms/step - loss: 0.2371
Epoch 8/20
59/59 [==============================] - 3s 57ms/step - loss: 0.1977
Epoch 9/20
59/59 [==============================] - 3s 57ms/step - loss: 0.1713
Epoch 10/20
59/59 [==============================] - 3s 57ms/step - loss: 0.1381
Epoch 11/20
59/59 [==============================] - 4s 61ms/step - loss: 0.1203
Epoch 12/20
59/59 [==============================] - 3s 57ms/step - loss: 0.1095
Epoch 13/20
59/59 [==============================] - 3s 56ms/step - loss: 0.0877
Epoch 14/20
59/59 [==============================] - 3s 57ms/step - loss: 0.0793
Epoch 15/20
59/59 [==============================] - 3s 56ms/step - loss: 0.0727
Epoch 16/20
59/59 [==============================] - 3s 56ms/step - loss: 0.0702
Epoch 17/20
59/59 [==============================] - 3s 56ms/step - loss: 0.0701
Epoch 18/20
59/59 [==============================] - 3s 57ms/step - loss: 0.0631
Epoch 19/20
59/59 [==============================] - 3s 56ms/step - loss: 0.0539
Epoch 20/20
59/59 [==============================] - 3s 58ms/step - loss: 0.0493
Out[3]: <tensorflow.python.keras.callbacks.History at 0x143069fdf40>
import numpy as np
W = model.get_weights()
def predict(X):
X = X.reshape((X.shape[0],-1)) #Flatten
X = X @ W[0] + W[1] #Dense
X[X<0] = 0 #Relu
X = X @ W[2] + W[3] #Dense
X[X<0] = 0 #Relu
X = X @ W[4] + W[5] #Dense
X = np.exp(X)/np.exp(X).sum(1)[...,None] #Softmax
return X
%timeit model.predict(X[:10]).argmax(1) # 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # 1000 loops takes 356 µs
52.8 ms ± 2.13 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
640 µs ± 10.9 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
Ipython Output (GPU):
Python 3.7.7 (default, Mar 26 2020, 15:48:22)
Type 'copyright', 'credits' or 'license' for more information
IPython 7.4.0 -- An enhanced Interactive Python. Type '?' for help.
In [1]: import tensorflow as tf
...:
...: (X,Y),(Xt,Yt) = tf.keras.datasets.mnist.load_data()
...:
...: model = tf.keras.models.Sequential([
...: tf.keras.layers.Flatten(),
...: tf.keras.layers.Dense(1000,'relu'),
...: tf.keras.layers.Dense(100,'relu'),
...: tf.keras.layers.Dense(10,'softmax'),
...: ])
...: model.compile('adam','sparse_categorical_crossentropy')
...: model.fit(X,Y,epochs=20,batch_size=1024)
2020-07-01 15:50:46.008518: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcuda.so.1
2020-07-01 15:50:46.054495: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1618] Found device 0 with properties:
name: GeForce RTX 2080 Ti major: 7 minor: 5 memoryClockRate(GHz): 1.545
pciBusID: 0000:05:00.0
2020-07-01 15:50:46.059582: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.0
2020-07-01 15:50:46.114562: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0
2020-07-01 15:50:46.142058: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcufft.so.10.0
2020-07-01 15:50:46.152899: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcurand.so.10.0
2020-07-01 15:50:46.217725: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcusolver.so.10.0
2020-07-01 15:50:46.260758: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcusparse.so.10.0
2020-07-01 15:50:46.374328: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudnn.so.7
2020-07-01 15:50:46.376747: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1746] Adding visible gpu devices: 0
2020-07-01 15:50:46.377688: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX FMA
2020-07-01 15:50:46.433422: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 4018875000 Hz
2020-07-01 15:50:46.434383: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x563e4d0d71c0 executing computations on platform Host. Devices:
2020-07-01 15:50:46.435119: I tensorflow/compiler/xla/service/service.cc:175] StreamExecutor device (0): Host, Default Version
2020-07-01 15:50:46.596077: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x563e4a9379f0 executing computations on platform CUDA. Devices:
2020-07-01 15:50:46.596119: I tensorflow/compiler/xla/service/service.cc:175] StreamExecutor device (0): GeForce RTX 2080 Ti, Compute Capability 7.5
2020-07-01 15:50:46.597894: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1618] Found device 0 with properties:
name: GeForce RTX 2080 Ti major: 7 minor: 5 memoryClockRate(GHz): 1.545
pciBusID: 0000:05:00.0
2020-07-01 15:50:46.597961: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.0
2020-07-01 15:50:46.597988: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0
2020-07-01 15:50:46.598014: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcufft.so.10.0
2020-07-01 15:50:46.598040: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcurand.so.10.0
2020-07-01 15:50:46.598065: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcusolver.so.10.0
2020-07-01 15:50:46.598090: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcusparse.so.10.0
2020-07-01 15:50:46.598115: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudnn.so.7
2020-07-01 15:50:46.599766: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1746] Adding visible gpu devices: 0
2020-07-01 15:50:46.600611: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.0
2020-07-01 15:50:46.603713: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1159] Device interconnect StreamExecutor with strength 1 edge matrix:
2020-07-01 15:50:46.603751: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1165] 0
2020-07-01 15:50:46.603763: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1178] 0: N
2020-07-01 15:50:46.605917: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1304] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 10311 MB memory) -> physical GPU (device: 0, name: GeForce RTX 2080 Ti, pci bus id: 0000:05:00.0, compute capability: 7.5)
Train on 60000 samples
Epoch 1/20
2020-07-01 15:50:49.995091: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0
60000/60000 [==============================] - 2s 26us/sample - loss: 9.9370
Epoch 2/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.6094
Epoch 3/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.3672
Epoch 4/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.2720
Epoch 5/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.2196
Epoch 6/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.1673
Epoch 7/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.1367
Epoch 8/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.1082
Epoch 9/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0895
Epoch 10/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0781
Epoch 11/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0666
Epoch 12/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0537
Epoch 13/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0459
Epoch 14/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0412
Epoch 15/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0401
Epoch 16/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0318
Epoch 17/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0275
Epoch 18/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0237
Epoch 19/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0212
Epoch 20/20
60000/60000 [==============================] - 0s 4us/sample - loss: 0.0199
Out[1]: <tensorflow.python.keras.callbacks.History at 0x7f7c9000b550>
In [2]: import numpy as np
...:
...: W = model.get_weights()
...:
...: def predict(X):
...: X = X.reshape((X.shape[0],-1)) #Flatten
...: X = X @ W[0] + W[1] #Dense
...: X[X<0] = 0 #Relu
...: X = X @ W[2] + W[3] #Dense
...: X[X<0] = 0 #Relu
...: X = X @ W[4] + W[5] #Dense
...: X = np.exp(X)/np.exp(X).sum(1)[...,None] #Softmax
...: return X
...:
In [3]: print(model.predict(X[:100]).argmax(1))
...: print(predict(X[:100]).argmax(1))
[5 0 4 1 9 2 1 3 1 4 3 5 3 6 1 7 2 8 6 9 4 0 9 1 1 2 4 3 2 7 3 8 6 9 0 5 6
0 7 6 1 8 7 9 3 9 8 5 9 3 3 0 7 4 9 8 0 9 4 1 4 4 6 0 4 5 6 1 0 0 1 7 1 6
3 0 2 1 1 7 5 0 2 6 7 8 3 9 0 4 6 7 4 6 8 0 7 8 3 1]
/home/bobbyocean/anaconda3/bin/ipython3:12: RuntimeWarning: overflow encountered in exp
/home/bobbyocean/anaconda3/bin/ipython3:12: RuntimeWarning: invalid value encountered in true_divide
[5 0 4 1 9 2 1 3 1 4 3 5 3 6 1 7 2 8 6 9 4 0 9 1 1 2 4 3 2 7 3 8 6 9 0 5 6
0 7 6 1 8 7 9 3 9 8 5 9 3 3 0 7 4 9 8 0 9 4 1 4 4 6 0 4 5 6 1 0 0 1 7 1 6
3 0 2 1 1 7 5 0 2 6 7 8 3 9 0 4 6 7 4 6 8 0 7 8 3 1]
In [4]: %timeit model.predict(X[:10]).argmax(1)
37.7 ms ± 806 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [5]: %timeit predict(X[:10]).argmax(1)
361 µs ± 13.8 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
As others have pointed out, the Tensorflow binary in question was compiled for GPU optimization: while GPUs are excellent at intensive number-crunching because of the extremely high number of computing cores they have, they’re painfully slow when it comes to moving data back and forth.
When a model is executing on a graphics card, all the necessary data has to be burst over to the GPU — it has no access to the host system’s RAM (nor does the host system have access to the video memory). Once the GPU is finished processing, any results have to be shipped back over to the host system.
All of this moving around of data takes a great deal of time; moreover, a Tensorflow binary compiled to execute with a GPU/CUDA does not, to my knowledge, include any of the standard optimizations for executing on a CPU (like using faster extended instruction sets, e.g. AVX, AVX2, etc).
So you’re comparing a highly CPU-optimized scientific library which can process data without even having to go back to RAM half the time (CPU registers and cache storage on the chip); to code that has to collect every last bit it’s going to need before having to ship all of that data to the graphics card, and back. I’m also leaving out all of the data manipulations that goes on under the hood of Tensorflow. It works on its own data structures, after all.
Eager execution is also another layer of inefficiency, I suppose.
As for best practices in deploying Keras models, I’m of the opinion that it’s like everything else in software: premature optimization is the root of all evil. If you don’t need it to be fast and lean, then let it be slow, modular, reusable, and intuitive. But hey, if you need or want the efficiency, then power to you. Keras is designed for rapid development and research, not production code.
In short, the answer is that it’s for the same reason C++ is faster than Python (because the Python interpreter has so much more overhead, as does Tensorflow).
We observe that the main issue is the cause of the Eager Execution mode. We give shallow look at your code and corresponding results as per CPU and GPU bases. It is true that numpy doesn’t operate on GPU, so unlike tf-gpu, it doesn’t encounter any data shifting overhead.
But also it’s quite noticeable how much fast computation is done by your defined predict method with np compare to model. predict with tf. keras, whereas the input test set is 10 samples only. However, We’re not giving any deep analysis, like one piece of art here you may love to read.
My Setup is as follows. I’m using the Colab environment and checking with both CPU and GPU mode.
TensorFlow 1.15.2
Keras 2.3.1
Numpy 1.19.5
TensorFlow 2.4.1
Keras 2.4.0
Numpy 1.19.5
TF 1.15.2 – CPU
%tensorflow_version 1.x
import os
os.environ["CUDA_VISIBLE_DEVICES"]="-1"
import tensorflow as tf
from tensorflow.python.client import device_lib
print(tf.__version__)
print('A: ', tf.test.is_built_with_cuda)
print('B: ', tf.test.gpu_device_name())
local_device_protos = device_lib.list_local_devices()
([x.name for x in local_device_protos if x.device_type == 'GPU'],
[x.name for x in local_device_protos if x.device_type == 'CPU'])
TensorFlow 1.x selected.
1.15.2
A: <function is_built_with_cuda at 0x7f122d58dcb0>
B:
([], ['/device:CPU:0'])
Now, running your code.
import tensorflow as tf
import keras
print(tf.executing_eagerly()) # False
(X,Y),(Xt,Yt) = keras.datasets.mnist.load_data()
model = keras.models.Sequential([])
model.compile
model.fit
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
1000 loops, best of 5: 1.07 ms per loop
1000 loops, best of 5: 1.48 ms per loop
We can see that the execution times are comparable with old keras. Now, let’s test with GPU also.
TF 1.15.2 – GPU
%tensorflow_version 1.x
import os
os.environ["CUDA_VISIBLE_DEVICES"]="0"
import tensorflow as tf
from tensorflow.python.client import device_lib
print(tf.__version__)
print('A: ', tf.test.is_built_with_cuda)
print('B: ', tf.test.gpu_device_name())
local_device_protos = device_lib.list_local_devices()
([x.name for x in local_device_protos if x.device_type == 'GPU'],
[x.name for x in local_device_protos if x.device_type == 'CPU'])
1.15.2
A: <function is_built_with_cuda at 0x7f0b5ad46830>
B: /device:GPU:0
(['/device:GPU:0'], ['/device:CPU:0'])
...
...
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
1000 loops, best of 5: 1.02 ms per loop
1000 loops, best of 5: 1.44 ms per loop
Now, the execution time is also comparable here with old keras and no eager mode. Let’s now see the new tf. keras with eager mode first and then we observe without eager mode.
TF 2.4.1 – CPU
Eagerly
import os
os.environ["CUDA_VISIBLE_DEVICES"]="-1"
import tensorflow as tf
from tensorflow.python.client import device_lib
print(tf.__version__)
print('A: ', tf.test.is_built_with_cuda)
print('B: ', tf.test.gpu_device_name())
local_device_protos = device_lib.list_local_devices()
([x.name for x in local_device_protos if x.device_type == 'GPU'],
[x.name for x in local_device_protos if x.device_type == 'CPU'])
2.4.1
A: <function is_built_with_cuda at 0x7fed85de3560>
B:
([], ['/device:CPU:0'])
Now, running the code with eager mode.
import tensorflow as tf
import keras
print(tf.executing_eagerly()) # True
(X,Y),(Xt,Yt) = keras.datasets.mnist.load_data()
model = keras.models.Sequential([ ])
model.compile
model.fit
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
10 loops, best of 5: 28 ms per loop
1000 loops, best of 5: 1.73 ms per loop
Disable Eagerly
Now, if we disable the eager mode and run the same code as follows then we will get:
import tensorflow as tf
import keras
# # Disables eager execution
tf.compat.v1.disable_eager_execution()
# or,
# Disables eager execution of tf.functions.
# tf.config.run_functions_eagerly(False)
print(tf.executing_eagerly())
False
(X,Y),(Xt,Yt) = keras.datasets.mnist.load_data()
model = keras.models.Sequential([])
model.compile
model.fit
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
1000 loops, best of 5: 1.37 ms per loop
1000 loops, best of 5: 1.57 ms per loop
Now, we can see the execution times are comparable for disabling the eager mode in new tf. keras. Now, let’s test with GPU mode also.
TF 2.4.1 – GPU
Eagerly
import os
os.environ["CUDA_VISIBLE_DEVICES"]="0"
import tensorflow as tf
from tensorflow.python.client import device_lib
print(tf.__version__)
print('A: ', tf.test.is_built_with_cuda)
print('B: ', tf.test.gpu_device_name())
local_device_protos = device_lib.list_local_devices()
([x.name for x in local_device_protos if x.device_type == 'GPU'],
[x.name for x in local_device_protos if x.device_type == 'CPU'])
2.4.1
A: <function is_built_with_cuda at 0x7f16ad88f680>
B: /device:GPU:0
(['/device:GPU:0'], ['/device:CPU:0'])
import tensorflow as tf
import keras
print(tf.executing_eagerly()) # True
(X,Y),(Xt,Yt) = keras.datasets.mnist.load_data()
model = keras.models.Sequential([ ])
model.compile
model.fit
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
10 loops, best of 5: 26.3 ms per loop
1000 loops, best of 5: 1.48 ms per loop
Disable Eagerly
And lastly again, if we disable the eager mode and run the same code as follows, we will get:
# Disables eager execution
tf.compat.v1.disable_eager_execution()
# or,
# Disables eager execution of tf.functions.
# tf.config.run_functions_eagerly(False)
print(tf.executing_eagerly()) # False
(X,Y),(Xt,Yt) = keras.datasets.mnist.load_data()
model = keras.models.Sequential([ ])
model.compile
model.fit
%timeit model.predict(X[:10]).argmax(1) # yours: 10 loops takes 37.7 ms
%timeit predict(X[:10]).argmax(1) # yours: 1000 loops takes 356 µs
1000 loops, best of 5: 1.12 ms per loop
1000 loops, best of 5: 1.45 ms per loop
And like before, the execution times are comparable with the non-eager mode in new tf. keras. That’s why, the Eager mode is the root cause of the slower performance of tf. keras than straight numpy.
Another answer is more useful in terms of "how to make tf keras predict faster", but I think the following can help more in terms of "what is it doing that takes so much time"? Even with eager mode disabled, you might be curious to see how execution looks like (e.g. with or without providing batch_size, etc.).
To answer this question you may find tracing profiler to be useful. Tracing execution is adding a lot of overhead (especially for the places that have a bunch of very lightweight python calls), but overall should give you quite a bit of insight into what part of python code is being executed, because, well, it just logs exactly what is happening. You can try pytracing since it produces files that Chrome browser visualizes nicely on it’s built-in chrome://tracing page. To use it e.g. in google colab you can do the following:
First, intstall pytracing:
!pip install pytracing
Then generate trace:
from pytracing import TraceProfiler
tp = TraceProfiler(output=open('/root/trace.out', 'wt'))
with tp.traced():
for i in range(2):
model.predict(X[:1000], batch_size=1000)
Then download trace:
from google.colab import files
files.download('/root/trace.out')
After this in Chrome browser open chrome://tracing page, click "Load" button, and select trace.out file, you’ve downloaded.
You’ll see something like the following – you can click on any element, see the full name of the python function and file it is from + wall time it took (again, all of this is higher than in normal runs due to tracing overhead):
You can see how disabling/enabling eager execution or changing the batch size will change the output and can see for yourself what is taking the most time. From what I currently see (in non-eagerly mode + call like model.predict(X[:1000], batch_size=1000)) quite a bit of time is spent on:
Standardizing your data (whatever it means): ~2.5ms (including tracing overhead!):
/usr/local/lib/python3.7/dist-packages/tensorflow/python/keras/engine/training_v1.py:2336:_standardize_user_data
Preparing callbacks (even though we didn’t set any): ~2ms (including tracing overhead)
/usr/local/lib/python3.7/dist-packages/tensorflow/python/keras/callbacks.py:133:configure_callbacks
As to the statement that numpy version is not optimized – I wouldn’t agree. The numpy implementation here is quite optimized – python is not making any pure python calls in it (execution of predict only results call to functions in C – I couldn’t believe it at first but it seems to be the case), so overhead from Python is really minimal. You might gain a little bit by optimizing the way you do ReLU and eliminating extra allocation/deallocation, but that could only result in a very minor perf improvement.
Instead of using model.predict(input), simply try model(input)
To have a fast predict function on small batches, a simple solution is to do this:
predict = lambda x: tf.function(model(x))
# Use predict(x) instead of model.predict(x)
If you want to use numpy arrays, you can do this:
predict_np = lambda x: predict(tf.convert_to_tensor(x))
I have a speed up of x5 and more on CPU. The performance is equivalent to the performance achieved with eager execution disabled, but here it doesn’t affect the whole code.