I'd like to encapsulate my calc function and all its parameters inside an object, but vectorize the execution for millions of objects much like how numpy would do it. Any suggestions?
the calculation is still basic arithmetic which numpy should be able to vectorize.
Example code:
import numpy as np
myarray = np.random.rand(3, 10000000)
############################# This works fine: FAST ###################################
def calc(a,b,c):
return (a+b/c)**b/a
res1 = calc(*myarray) #0.7 seconds
############################# What I'd like to do (unsuccessfully): SLOW ###################################
class MyClass():
__slots__ = ['a','b','c']
def __init__(self, a,b,c):
self.a, self.b, self.c = a,b,c
def calc(self):
return (self.a + self.b / self.c) ** self.b / self.a
def classCalc(myClass:MyClass):
return myClass.calc()
vectorizedClassCalc = np.vectorize(classCalc)
myobjects = np.array([MyClass(*args) for args in myarray.transpose()])
res2 = vectorizedClassCalc(myobjects) #8 seconds no different from a list comprehension
res3 = [obj.calc() for obj in myobjects] #7.5 seconds
perhaps pandas has additional features?
Related
I am trying to get a GLCM implementation running in a custom Keras Layer in a reasonable fast time. So far I took the _glcm_loop from skimage-implementation, reduced it to what I needed and put it into a basic layer, like this:
import numpy as np
import tensorflow as tf
from time import time
from tensorflow import keras
from tensorflow.keras.preprocessing import image
from tensorflow.keras import layers
from skimage.feature import *
from numpy import array
from math import sin, cos
from time import time
import matplotlib.pyplot as plt
class GLCMLayer(keras.layers.Layer):
def __init__(self, greylevels=32, angles=[0], distances=[1], name=None, **kwargs):
self.greylevels = greylevels
self.angles = angles
self.distances = distances
super(GLCMLayer, self).__init__(name=name, **kwargs)
def _glcm_loop(self, image, distances, angles, levels, out):
rows = image.shape[0]
cols = image.shape[1]
for a_idx in range(len(angles)):
angle = angles[a_idx]
for d_idx in range(len(distances)):
distance = distances[d_idx]
offset_row = round(sin(angle) * distance)
offset_col = round(cos(angle) * distance)
start_row = max(0, -offset_row)
end_row = min(rows, rows - offset_row)
start_col = max(0, -offset_col)
end_col = min(cols, cols - offset_col)
for r in range(start_row, end_row):
for c in range(start_col, end_col):
i = image[r, c]
row = r + offset_row
col = c + offset_col
j = image[row, col]
out[i, j, d_idx, a_idx] += 1
def call(self, inputs):
P = np.zeros((self.greylevels, self.greylevels, len(self.distances), len(self.angles)), dtype=np.uint32, order='C')
self._glcm_loop(inputs, self.distances, self.angles, self.greylevels, P)
return P
def get_config(self):
config = {
'angle': self.angle,
'distance': self.distance,
'greylevel': self.greylevel,
}
base_config = super(GLCMLayer, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
My execution code looks like this:
def quant(img, greylevels):
return array(img)//(256//greylevels)
if __name__ == "__main__":
source_file = "<some sour file>"
img_raw = image.load_img(source_file, target_size=(150,150), color_mode="grayscale")
img = quant(img_raw, 32)
layer = GLCMLayer()
start = time()
aug = layer(img)
tf.print(time()-start)
This is my first step to create it as a preprocessing layer. The second step then will be to modify it to run it also as hidden layer inside a model. That is why I didn't put it to a complete model yet, but I feel like there will be additional changes required when doing so.
For some reason the execution time is about 15-20 seconds long. Executing the code on the CPU without the layer takes about 0.0009 seconds. Obviously, something is going wrong here.
I am fairly new to tf and keras, so I fear I am missing something in the way on how to use the framework. In order to resolve it, I read about (which doesn't mean I understood):
do not use np-functions inside tensorflow, but tf-functions instead,
use tf.Variable,
use tf.Data,
unfolding is not possible in some way (whatever that means)
I tried a little here and there, but couldn't get them running, instead finding various different exceptions. So my questions are:
What is the correct way to use tf-functions in a GLCM to get the best performance on the GPU?
What do I need to take care of when using the layer in a complete model?
From that point on, I should hopefully be able to then implement the GLCM properties.
Any help is greatly appreciated.
(Disclaimer: I assume that there is a lot of other stuff not optimal yet, if anything comes to your mind just add it.)
It comes to a situation we need a train a bunch of data(about 22GiB), I make a test with two methods to generate random data and try to train it with Dask, however, the data generated by Numpy would raise an exception(msgpack: bytes object is too large) while the Dask.array one works. Did anybody know why?
from dask.distributed import Client
from dask_cuda import LocalCUDACluster
from dask import array as da
import numpy as np
import xgboost as xgb
import time
def main(client):
regressor = None
pre = None
n=3000
m=1000000
# numpy generated data will raise an exception
X = np.random.random((m, n))
y = np.random.random((m, 1))
X = da.from_array(X, chunks=(1000, n))
y = da.from_array(y, chunks=(1000, 1))
# data generated by dask.array works well
# X = da.random.random(size=(m, n), chunks=(1000, n))
# y = da.random.random(size=(m, 1), chunks=(1000, 1))
dtrain = xgb.dask.DaskDMatrix(client, X, y)
del X
del y
params = {'tree_method':'gpu_hist'}
watchlist = [(dtrain, 'train')]
start = time.time()
bst = xgb.dask.train(client, params, dtrain, num_boost_round=100, evals=watchlist)
print('consume:', time.time() - start)
if __name__ == '__main__':
with LocalCUDACluster(n_workers=4, device_memory_limit='12 GiB') as cluster:
with Client(cluster) as client:
main(client)
After making a few tests, I found out the reason, The da.random.random is a delay function as well(so it pass worker only the definition of random), in our situation, the msgpack limit the data size(4GiB) pass to each worker, so, in general, it wouldn't work for data size more than 4GiB directly communicate with Dask XGBoost(BTW, we can switch to parquet data and read it as dash.dataframe chunk data to bypass the limitation of msgpack)
the following commands proved my guess.
The following code fails when find_root is decorated with nb.jit. This is a toy example, but the idea is to have the ability to find the root of a scalar function (or potentially a multivariate function using root) for an array of values and store them in a numpy array.
Error message: TypingError: cannot determine Numba type of <class 'function'>
import numba as nb
import numpy as np
from scipy.optimize import root_scalar
a = 3.0
b = 1.0
c = -10.5
#nb.jit(nopython=True)
def f(x):
return a*x**2 + b*x + c
#nb.jit(nopython=True)
def fprime(x):
return 2*a*x + b
#nb.jit(nopython=True)
def fprime2(x):
return 2*a
#nb.jit(nopython=True) # <-- Commenting this line makes the code work but it is slow
def findroot(arr):
for i in range(len(arr)):
arr[i] = root_scalar(f, fprime=fprime, fprime2=fprime2, x0=0).root
if __name__ == '__main__':
arr = np.zeros(20, np.float)
import timeit
start = timeit.time.process_time()
findroot(arr)
end = timeit.time.process_time()
print(end - start)
SOLUTION BELOW:
Scenario:
I am trying to compute the jacobian of a user defined function many, many times in a loop. I am able to do this with TF 2's GradientTape as well as the older session based tf.gradients() method. The problem is that GradientTape is terribly slow (100x slower) than tf.gradients(). It has features i'd like to use (bath_jacobian, hessian support, etc), but if it's 100x slower then i can't use it.
The Question:
It's not clear to me if i'm simply misusing GradientTape, or if it will always be slower because it has to re-differentiate the provided function every time its called (my suspicion). I'm asking for tips to fix my use of GradientTape or a confirmation that it will always be fundamentally slower than tf.gradients by orders of magnitude.
Related Questions:
Repeated use of GradientTape for multiple Jacobian calculations - same scenario, unanswered
Does `GradientTape` need to re-differentiate each evaluation of a derivative? - same scenario, unanswered
using one GradientTape with global context - loosely related, having trouble applyng that solution to my scenario
Fully contained minimum example to compare GradientTape and tf.gradients():
import tensorflow as tf
from tensorflow.python.framework.ops import disable_eager_execution
import numpy as np
# from tensorflow.python.ops.parallel_for.gradients import jacobian, batch_jacobian
import timeit
class FunctionCaller(object):
def __init__(self, func, nX, dtype=tf.float64, useSessions=True):
if useSessions:
disable_eager_execution()
self.func = func
self.nX = nX
self.useSessions = useSessions
self.dtype = dtype
self.sess = tf.compat.v1.Session() if useSessions else None
if not useSessions:
return
#
# we are in session mode, so build the graph and take the batch-jacobian of the function's outputs
#
xTensor = tf.compat.v1.placeholder(dtype, shape=[None, nX])
# add function to graph and guarantee its output shape
func_tensor = tf.reshape(func(xTensor), [-1, nX])
# take the gradient for each output, one at a time, and stack the results back together
each_output = tf.unstack(func_tensor, nX, axis=1)
jac_x = tf.stack([tf.gradients(output, xTensor, unconnected_gradients='zero')[0]
for output in each_output], axis=1)
# record these tensors so we can use them later with session.run()
self.xTensor = xTensor
self.func_tensor = func_tensor
self.jac_func_tensor = jac_x
def jac(self, x_i):
if self.useSessions:
return self.sess.run(self.jac_func_tensor, {self.xTensor: x_i})
else:
return self._useGradientTape(x_i)
# THIS FUNCTION IS SUPER INEFFICIENT.
def _useGradientTape(self, x_i):
with tf.GradientTape(persistent=True) as g:
xTensor = tf.Variable(x_i, dtype=self.dtype) # is this my problem??? i recreate x every time?
y = tf.reshape(self.func(xTensor), [-1, self.nX])
jac_x_at_i = g.batch_jacobian(y, xTensor)
# del g
return jac_x_at_i.numpy()
def __del__(self):
if self.sess is not None:
self.sess.close()
def main():
#tf.function
def Xdot(x_i):
x_0, x_1, x_2 = tf.split(x_i, 3, axis=1)
return tf.concat([x_2 * tf.sin(x_2), x_2 * tf.cos(x_2), x_2], axis=1)
nT = 20
nX = 3
# create some trash data
x_i = np.arange(nT*nX).reshape([-1, nX])
nTrials = 100
# try the eager version first
caller_eager = FunctionCaller(Xdot, nX, useSessions=False)
start_time = timeit.default_timer()
for _ in range(nTrials):
jac_eager = caller_eager.jac(x_i)
elapsed = timeit.default_timer() - start_time
print("eager code took {} sec: {} sec/trial".format(elapsed, elapsed/nTrials))
# now try the sessions version
caller_sessions = FunctionCaller(Xdot, nX, useSessions=True)
start_time = timeit.default_timer()
caller_sessions.jac(x_i) # call it once to do its graph building stuff?
for _ in range(nTrials):
jac_session = caller_sessions.jac(x_i)
elapsed = timeit.default_timer() - start_time
print("session code took {} sec: {} sec/trial".format(elapsed, elapsed/nTrials))
residual = np.max(np.abs(jac_eager - jac_session))
print('residual between eager and session trials is {}'.format(residual))
if __name__ == "__main__":
main()
EDIT - SOLUTION:
xdurch0 pointed out below that I should wrap _useGradientTape() in a #tf.function - something I was unsuccessful with before for other reasons. Once I did that, I had to move xTensor's definition outside the #tf.function wrapper by making it a member variable and using tf.assign().
With all this done, I find that GradientTape (for this simple example) is now on the same order of magnitude as tf.gradints. When running enough trials (~1E5), it's twice as fast as tf.gradients. awesome!
import tensorflow as tf
from tensorflow.python.framework.ops import disable_eager_execution
import numpy as np
import timeit
class FunctionCaller(object):
def __init__(self, func, nT, nX, dtype=tf.float64, useSessions=True):
if useSessions:
disable_eager_execution()
self.func = func
self.nX = nX
self.useSessions = useSessions
self.dtype = dtype
self.sess = tf.compat.v1.Session() if useSessions else None
if not useSessions:
# you should be able to create without an initial value, but tf is demanding one
# despite what the docs say. bug?
# tf.Variable(initial_value=None, shape=[None, nX], validate_shape=False, dtype=self.dtype)
self.xTensor = tf.Variable([[0]*nX]*nT, dtype=self.dtype) # x needs to be properly sized once
return
#
# we are in session mode, so build the graph and take the batch-jacobian of the function's outputs
#
xTensor = tf.compat.v1.placeholder(dtype, shape=[None, nX])
# add function to graph and guarantee its output shape
func_tensor = tf.reshape(func(xTensor), [-1, nX])
# take the gradient for each output, one at a time, and stack the results back together
each_output = tf.unstack(func_tensor, nX, axis=1)
jac_x = tf.stack([tf.gradients(output, xTensor, unconnected_gradients='zero')[0]
for output in each_output], axis=1)
# record these tensors so we can use them later with session.run()
self.xTensor = xTensor
self.func_tensor = func_tensor
self.jac_func_tensor = jac_x
def jac(self, x_i):
if self.useSessions:
return self.sess.run(self.jac_func_tensor, {self.xTensor: x_i})
else:
return self._useGradientTape(x_i).numpy()
#tf.function # THIS IS CRUCIAL
def _useGradientTape(self, x_i):
with tf.GradientTape(persistent=True) as g:
self.xTensor.assign(x_i) # you need to create the variable once outside the graph
y = tf.reshape(self.func(self.xTensor), [-1, self.nX])
jac_x_at_i = g.batch_jacobian(y, self.xTensor)
# del g
return jac_x_at_i
def __del__(self):
if self.sess is not None:
self.sess.close()
def main():
#tf.function
def Xdot(x_i):
x_0, x_1, x_2 = tf.split(x_i, 3, axis=1)
return tf.concat([x_2 * tf.sin(x_2), x_2 * tf.cos(x_2), x_2], axis=1)
nT = 20
nX = 3
# create some trash data
x_i = np.random.random([nT, nX])
nTrials = 1000 # i find that nTrials<=1E3, eager is slower, it's faster for >=1E4, it's TWICE as fast for >=1E5
# try the eager version first
caller_eager = FunctionCaller(Xdot, nT, nX, useSessions=False)
start_time = timeit.default_timer()
for _ in range(nTrials):
jac_eager = caller_eager.jac(x_i)
elapsed = timeit.default_timer() - start_time
print("eager code took {} sec: {} sec/trial".format(elapsed, elapsed/nTrials))
# now try the sessions version
caller_sessions = FunctionCaller(Xdot, nT, nX, useSessions=True)
start_time = timeit.default_timer()
for _ in range(nTrials):
jac_session = caller_sessions.jac(x_i)
elapsed = timeit.default_timer() - start_time
print("session code took {} sec: {} sec/trial".format(elapsed, elapsed/nTrials))
residual = np.max(np.abs(jac_eager - jac_session))
print('residual between eager and session trials is {}'.format(residual))
if __name__ == "__main__":
main()
I am trying to speed up the conversion of select tfrecords to a series of python dictionaries. Here's what I have. Initially the CPU utilization spikes, but then goes to almost zero, suggesting my code is not working correctly.
My goal is to have 3 dictionaries saved and pickled. There are 14,000+ tfrecord files (2 gigs appx). At the current rate, it will take about 84 hours to run on a single process.
Are there any problems with my use of manage dicts
import glob
import tensorflow as tf
import cPickle
import numpy as np
from tqdm import tqdm
import collections
from multiprocessing import Process, Manager, Pool
def get_multihot_encoding(example_label):
enc = np.zeros(10)
for label in example_label:
if label in lookup.values():
index = lookup_inverted[label]
enc[index] = 1
return list(enc)
# Set-up MultiProcessing
manager = Manager()
audio_embeddings_dict = manager.dict()
audio_labels_dict = manager.dict()
audio_multihot_dict = manager.dict()
sess = tf.Session()
# The iterable which gets passed to the function
all_tfrecord_filenames = glob.glob('/Users/jeff/features/audioset_v1_embeddings/unbal_train/*.tfrecord')
def process_tfrecord(tfrecord):
for idx, example in enumerate(tf.python_io.tf_record_iterator(tfrecord)):
tf_example = tf.train.Example.FromString(example)
vid_id = tf_example.features.feature['video_id'].bytes_list.value[0].decode(encoding='UTF-8')
example_label = list(np.asarray(tf_example.features.feature['labels'].int64_list.value))
# Non zero intersect of 2 sets is True - only create dict entries if this is true!
if set(example_label) & label_filters:
print(set(example_label) & label_filters, " Is the intersection of the two")
tf_seq_example = tf.train.SequenceExample.FromString(example)
n_frames = len(tf_seq_example.feature_lists.feature_list['audio_embedding'].feature)
audio_frame = []
for i in range(n_frames):
audio_frame.append(tf.cast(tf.decode_raw(
tf_seq_example.feature_lists.feature_list['audio_embedding'].feature[i].bytes_list.value[0],tf.uint8)
,tf.float32).eval(session=sess))
audio_embeddings_dict[vid_id] = audio_frame
audio_labels_dict[vid_id] = example_label
audio_multihot_dict[vid_id] = get_multihot_encoding(example_label)
#print(get_multihot_encoding(example_label), "Is the encoded label")
if idx % 100 == 0:
print ("Saving dictionary at loop: {}".format(idx))
cPickle.dump(audio_embeddings_dict, open('audio_embeddings_dict_unbal_train_multi_{}.pkl'.format(idx), 'wb'))
cPickle.dump(audio_multihot_dict, open('audio_multihot_dict_bal_untrain_multi_{}.pkl'.format(idx), 'wb'))
cPickle.dump(audio_multihot_dict, open('audio_labels_unbal_dict_multi_{}.pkl'.format(idx), 'wb'))
pool = Pool(50)
result = pool.map(process_tfrecord, all_tfrecord_filenames)