Tensorflow

Compatibility (Tensorflow, Python version, Compiler, Build tools, cuDNN and CUDA)

• see List
• Check the CUDA version (paths might differ slightly depending on the cuda version): cat /usr/local/cuda/version.txt
• and cuDNN version: grep CUDNN_MAJOR -A 2 /usr/local/cuda/include/cudnn.h

Check GPUs

Tensorflow 2

• tf.config.list_physical_devices('GPU')
• print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))

Porting (Tensorflow 1 to Tensorflow 2)

General

• List of Corresponding Commands

GraphDef

Error: module 'tensorflow' has no attribute 'GraphDef'

• source
  • Yeah, the syntax has changed in T2.0. Here’s the correct piece:

tf.compat.v1.GraphDef()   # -> instead of tf.GraphDef()
tf.compat.v2.io.gfile.GFile()   # -> instead of tf.gfile.GFile()

Session

• tf.compat.v1.Session() instead of tf.Session()

Tensorboard

• tensorboard --logdir=dir_with_eventfiles in bash terminal
  • this will start the tensorboard server and then tensorboard can be viewed e.g. in firefox

tfrecords

Inspecting the first image in a tfrecords file

import tensorflow as tf
raw_dataset = tf.data.TFRecordDataset("path-to-file")

for raw_record in raw_dataset.take(1):
    example = tf.train.Example()
    example.ParseFromString(raw_record.numpy())
    print(example)

• this also shows the other key-value pairs such as labels, bounding box coordinates, etc.

Memory Formats

• overview

NCHW

• data format for activations
• “NCHW” also describes the order in which the tensor values are laid out in memory
• source:
  • NCHW stands for: batch N, channels C, depth D, height H, width W
  • It is a way to store multidimensional arrays / data frames / matrix into memory, which can be considered as a 1-D array.

NHWC

Tensorflow API

tf.where

segmentation_map = tf.where(
                                    condition=tf.reduce_all(tf.equal(image, color), axis=-1),
                                    x=tf.cast(class_id, tf.uint8),
                                    y=segmentation_map
                                    )

Achtung: Reihenfolge von image und color bei tf.equal() ist wichtig !