Errata notes and general comments on TensorFlow_Tut_3_Object_Detection_Walk-through.docx

[mgedda]

Hello, Thanks a million for this great walk-through! I could not find any contact info to you so I thought I would post some errata notes/comments here instead.

TensorFlow_Tut_3_Object_Detection_Walk-through.docx - page 12, section 13:

• The directory training_data is referred to as only data in the paths.
• And later in section 14 it is again referred to as (repository_location)​​\data.

TensorFlow_Tut_3_Object_Detection_Walk-through.docx - page 12, section 14:

• The method class_text_to_int is really named classAsTextToClassAsInt in 2_generate_tfrecords.py

TensorFlow_Tut_3_Object_Detection_Walk-through.docx - page 13, section 14:

• ​(repository_location)​​\data should be ​(repository_location)​​\training_data
• “test.tfrecord” should be “eval.tfrecord“

TensorFlow_Tut_3_Object_Detection_Walk-through.docx - page 14, section 17:

• Here the reader is instructed to "Open (repository_location)\​​label_map.pbtxt in Notepad or a similar text editor and skim the contents", which is the same file we were instructed to inspect in section 16. I guess this should refer to (repository_location)\ssd_inception_v2_coco.config instead.
• Same again for “Comparing ​(repository_location)​​\label_map.pbtxt …” at the bottom of the page.

TensorFlow_Tut_3_Object_Detection_Walk-through.docx - page 15, section 17:

• The instructions say that the file ssd_inception_v2_coco.config in the (repository_location) has relative links to the directory ssd_inception_v2_coco_2017_11_17 but the file is updated to link to the directory ssd_inception:v2_coco_2018_01_28 (e.g., fine_tune_checkpoint: "ssd_inception_v2_coco_2018_01_28/model.ckp"). This typo occurs at three positions on page 15.

General comments

• When running the the object_detection_tutorial.ipynb in Jupyter notebook I needed to comment out the tensorflow version check for it to work even though I was using tensorflow v1.10.0.
• If the reader is using a conda environment then he/she needs to install nb_conda for the Jupyter notebook to work with conda environments. Might be a good point to add. (see https://stackoverflow.com/questions/37085665/in-which-conda-environment-is-jupyter-executing)
• If the reader is using a conda environment then he/she can use pip install -e [directory] to install the tensorflow/models/research and tensorflow/models/research/slim packages into the virtual environment instead of adding paths to PYTHONPATH. Then just change the imports from "from utils import ..." to "from object_detection.utils import". I found this a lot easier than adding the paths to PYTHONPATH when working in a virtual environment.

I might add more comments to this issue once I have finished the training and inference parts of the walk-through.