py-tensorflow-macos: Update to version 2.6.0, py-tensorflow-metal: Submission

[essandess]

Description

• Update to version 2.6.0
• Rename as py-tensorflow-macos per pypi
• Obsolete old github version tensorflow_macos
• Submit py-tensorflow-metal

cc: @cjones051073

Type(s)

• bugfix
• enhancement
• security fix

Tested on

macOS 11.6 20G165 x86_64
Xcode 13.0 13A233

Verification

Have you

• followed our Commit Message Guidelines?
• squashed and minimized your commits?
• checked that there aren't other open pull requests for the same change?
• referenced existing tickets on Trac with full URL?
• checked your Portfile with port lint?
• tried existing tests with sudo port test?
• tried a full install with sudo port -vst install?
• tested basic functionality of all binary files?

TensorFlow Metal CNN Benchmark

import numpy as np, os, pandas as pd, platform, time, warnings

import tensorflow as tf
print(tf.__version__)

from tensorflow.python.framework.ops import disable_eager_execution, enable_eager_execution
from tensorflow.python.keras import backend

from numpy import unique
from numpy import argmax
from tensorflow.keras.datasets.mnist import load_data
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Conv2D, MaxPool2D, Flatten, Dropout

from IPython.display import Image
from IPython.core.display import HTML 

print(tf.config.list_physical_devices('GPU'))
if tf.test.gpu_device_name():
    print('Default GPU Device: {}'.format(tf.test.gpu_device_name()))
else:
    print("Please install GPU version of TF")

import matplotlib as mpl, matplotlib.pyplot as plt
from matplotlib import cm
from matplotlib.colors import LightSource
from mpl_toolkits.mplot3d import Axes3D

# boldify
# Legible plot style defaults
# http://matplotlib.org/api/matplotlib_configuration_api.html
# http://matplotlib.org/users/customizing.html
mpl.rcParams['figure.figsize'] = (12.0, 8)
mpl.rc('font',**{'family': 'sans-serif', 'weight': 'bold', 'size': 14})
mpl.rc('axes',**{'titlesize': 20, 'titleweight': 'bold', 'labelsize': 16, 'labelweight': 'bold'})
mpl.rc('legend',**{'fontsize': 14})
mpl.rc('figure',**{'titlesize': 16, 'titleweight': 'bold'})
mpl.rc('lines',**{'linewidth': 2.5, 'markersize': 18, 'markeredgewidth': 0})
mpl.rc('mathtext',**{'fontset': 'custom', 'rm': 'sans:bold', 'bf': 'sans:bold', 'it': 'sans:italic', 'sf': 'sans:bold', 'default': 'it'})
# plt.rc('text',usetex=False) # [default] usetex should be False
mpl.rcParams['text.latex.preamble'] = r'\usepackage{amsmath,sfmath} \boldmath'

prop_cycle = plt.rcParams['axes.prop_cycle']
colors_default = prop_cycle.by_key()['color']

(trainX, trainy), (testX, testy) = tf.keras.datasets.mnist.load_data()
print('Train: X=%s, y=%s' % (trainX.shape, trainy.shape))
print('Test: X=%s, y=%s' % (testX.shape, testy.shape))

# a single grayscale channel
trainX = trainX[…, np.newaxis]
testX = testX[…, np.newaxis]

in_shape = trainX.shape[1:]
print(in_shape)

n_classes = len(unique(trainy))
print(n_classes)

# Normalize
trainX = trainX.astype('float32') / 255.0
testX = testX.astype('float32') / 255.0

# Create CNN Model
model = Sequential()
model.add(Conv2D(32, (3, 3), activation='relu', kernel_initializer='he_uniform', input_shape=in_shape))
model.add(MaxPool2D((2, 2), strides=(2,2)))
model.add(Conv2D(32, (2, 2), activation='relu', kernel_initializer='he_uniform'))
model.add(MaxPool2D((2, 2), strides=(2,2)))
model.add(Flatten())
model.add(Dense(500, activation='relu', kernel_initializer='he_uniform'))
# model.add(Dropout(0.5))
model.add(Dense(n_classes, activation='softmax'))
model.summary()

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

start_time = time.time()
model.fit(trainX, trainy, epochs=10, batch_size=2**12, verbose=1)
elapsed_time = time.time() - start_time

print(f"Time: {elapsed_time}")

# Evaluate
loss, acc = model.evaluate(testX, testy, verbose=0)
print('Accuracy: %.3f' % acc)


# Numpy comparison tests

# edges of all the image data
images_numpy = np.vstack([trainX, testX])

start_time = time.time()

images = tf.convert_to_tensor(images_numpy, dtype=np.float32)
grad_images = tf.image.sobel_edges(images)
grad_images_numpy = grad_images.numpy()

elapsed_time = time.time() - start_time

print(f"Time: {elapsed_time}")
print(grad_images_numpy.shape)

# Compare to `opencv`

import cv2

KERNEL_SIZE = 3

start_time = time.time()
gradx = np.array([gx for gx in map(lambda x: cv2.Sobel(x, cv2.CV_64F, 1, 0, ksize=KERNEL_SIZE), images_numpy)])
grady = np.array([gx for gx in map(lambda x: cv2.Sobel(x, cv2.CV_64F, 0, 1, ksize=KERNEL_SIZE), images_numpy)])
grad_images_numpy = np.stack([gradx, grady], axis=3)
elapsed_time = time.time() - start_time

print(f"Time: {elapsed_time}")
print(grad_images_numpy.shape)

[essandess]

Note this possible performance issue with the latest release of tensorflow-macos and tensorflow-metal: https://developer.apple.com/forums/thread/687654

Resolved by setting batch_size=2**12.

[essandess]

This is ready to merge to the best of my knowledge. I believe that the CI errors are timeouts.

[essandess]

@reneeotten Thanks again for all the great feedback and corrections. I now believe that TF’stensorflow-macos version is good to go and that the performance issues mentioned above are essentially user errors. Therefore, I’ll push a version of tensorflow_macos that deprecates those ports with the method you recommend.

[essandess]

I believe that all the CI errors are timeouts and that this PR is ready to merge.

[cjones051073]

Sorry to be late to the party here.. Having metal acceleration for TF is really nice. I don't see anything obvious here I object strongly to, but as @reneeotten needs to explicitly OK his requested changes lets wait until we hear back from him.

[essandess]

Sorry to be late to the party here.. Having metal acceleration for TF is really nice. I don't see anything obvious here I object strongly to, but as @reneeotten needs to explicitly OK his requested changes lets wait until we hear back from him.

👍🏻