Are you planning to try "model_main.py" again instead of "train.py"

[anonym24]

I suppose "model_main.py" should work better and mb then even fixed some errors which can affect speed/accuracy of models

[anonym24]

I see model_main.py needs cocoapi (pycocotools)

For Windows we can install it from https://github.com/philferriere/cocoapi

[anonym24]

but it fails with next errors TypeError: can't pickle dict_values objects

tensorflow/models#5719

[anonym24]

Seems this error can be solved by
guildai/models@d88f1fb
tensorflow/models#4780 (comment)

[anonym24]

seems it finally started to work with model_train.py

I start train like this (# From the tensorflow/models/research/ directory)

python object_detection\model_main.py --pipeline_config_path=object_detection\training\ssd_mobilenet_v1_pets.config --model_dir=object_detection\images --num_train_steps=50000 --sample_1_of_n_eval_examples=1 --alsologtostderr

(more info: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_locally.md)

I see some checkpoints:

Though output of console isn't that useful: tensorflow/models#5719 (comment)

[anonym24]

Solved and it works ok after some configuration: tensorflow/models#5719 (comment)

So SSD Models can be train normally with model_train.py (works on Windows and Ubuntu), with legacy train.py you can get OOM errors

[anonym24]

also by default model_train.py doesn't print any logs in your console

here'e a solution to fix it:

add tf.logging.set_verbosity(tf.logging.INFO) line after imports in model_train.py

[dscha09]

@anonym24 How is model_main.py different from train.py?

[JanithT-Lboro]

where is this model_train.py file?

[anonym24]

@dscha09 I don't know but with legacy train.py you can get OOM errors while starting training and at some point of training

@JanithT-Lboro https://github.com/tensorflow/models/blob/master/research/object_detection/model_main.py

[JanithT-Lboro]

I was able to solve this problem by just using a PC with better specs.

[Futi7]

C:\tensorflow\models\research>python object_detection\model_main.py --pipeline_config_path=object_detection\training\ssd_mobilenet_v1_pets.config --model_dir=object_detection\images --num_train_steps=50000 --sample_1_of_n_eval_examples=1 --alsologtostderr
Traceback (most recent call last):
File "object_detection\model_main.py", line 26, in
from object_detection import model_lib
File "C:\Users\Fuat\AppData\Local\Programs\Python\Python35\lib\site-packages\object_detection\model_lib.py", line 27, in
from object_detection import eval_util
File "C:\Users\Fuat\AppData\Local\Programs\Python\Python35\lib\site-packages\object_detection\eval_util.py", line 27, in
from object_detection.metrics import coco_evaluation
File "C:\Users\Fuat\AppData\Local\Programs\Python\Python35\lib\site-packages\object_detection\metrics\coco_evaluation.py", line 22, in
from object_detection.utils import object_detection_evaluation
File "C:\Users\Fuat\AppData\Local\Programs\Python\Python35\lib\site-packages\object_detection\utils\object_detection_evaluation.py", line 39, in
from object_detection.utils import label_map_util
File "C:\Users\Fuat\AppData\Local\Programs\Python\Python35\lib\site-packages\object_detection\utils\label_map_util.py", line 21, in
from object_detection.protos import string_int_label_map_pb2
ImportError: cannot import name 'string_int_label_map_pb2'

I got this error.How can i fix it.I installed protobuf, compiled it, added to path.
OS:Windows 10
CPU:AMD A9 7th Gen.
Python version:3.6
Protobuf version:3.4.0

[dienhoa]

@Futi7 I'm not sure on windows but I think you can solve it by running
protoc object_detection/protos/*.proto --python_out=.

[johnneijzen]

this work great

[EdjeElectronics]

Hey @anonym24 , I know it's been a year, but thanks for your work on this issue! I'm still considering whether to update my guide to use model_main.py rather than train.py. It's a bummer that it doesn't log training information in the console by default.

This issue (TensorFlow #6100) has some good information directly from a TensorFlow team member on the differences between train.py and model_main.py. From their comments, it sounds like there is no difference when using a normal PC for training.