howto_tf

## make sure to do protobuf compilation and add pyton path before anything else

# From tensorflow/models/research
protoc object_detection/protos/*.proto --python_out=.

# From tensorflow/models/research
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim

## TRAINING TENSORFLOW
# Always reboot before new training
# From the tensorflow/models/research directory
python object_detection/train.py \
    --logtostderr \
    --pipeline_config_path=/path/to/model.config \
    --train_dir=/path/to/outputdirectory/train

## EVALUATION TENSORFLOW
# From the tensorflow/models/research directory
python object_detection/eval.py \
    --logtostderr \
    --pipeline_config_path=/path/to/model.config \
    --checkpoint_dir=/path/to/outputdirectory/train \
    --eval_dir=/path/to/outputdirectory/eval

## TENSORBOARD
tensorboard --logdir=/path/to/outputdirectory

## EXPORT MODEL TENSORFLOW
# From the tensorflow/models/research directory
python object_detection/export_inference_graph.py \
--input_type=image_tensor \
--pipeline_config_path=/path/to/model.config \
--trained_checkpoint_prefix /path/to/outputdirectory/train/model.ckpt-123 \
--output_directory=/path/to/outputdirectory/modelname

## GPU CPU OPTIONS
# For Evaluation Using only CPU: In object_detection/eval.py add:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
# For Training using two GPUS: in object_detection/train.py add:
os.environ['CUDA_VISIBLE_DEVICES'] = '0,1'

## CONFIG FILE OPTIONS #### 
# To check config options look at objectdetection/protos
in train_config:
batch_queue_capacity: 2 # default is 8
prefetch_queue_capacity: 2 # default is 10
keep_checkpoint_every_n_hours: 1
visualization_export_dir: "model/eval/visualization"

## TRAINING ON A LOCAL MACHINE WITH LIMITED RESOURCES ####
# Train on a single GPU and eval on CPU, In object_detection/trainer.py 
# add after "session_config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=False)" :
session_config.gpu_options.allow_growth = True
session_config.gpu_options.per_process_gpu_memory_fraction = 0.5 # last option
session_config.gpu_options.allocator_type ='BFC'

#### IF eval.py gets stuck ####
#### with WARNING:root:image 0 does not have groundtruth difficult flag specified ####
change eval_config.num_examples to the size of your val set
or change eval_input_reader.shuffle to true.

#### IF EXPORTTING A FROZEN MODEL FAILS ###
on line 72 in exporter.py
change "layout_optimizer=rewriter_config_pb2.RewriterConfig.ON" to "optimize_tensor_layout=True"

#### COULD BE NECESSARY IF IMPORT ERRORS OCCUR ####
#### DO THIS AS LAST OPTION, MAY CORRECT SOME ERRORS BUT LEAD TO NEW ####

# from tensorflow/models/research
python setup.py build
python setup.py install

# from tensorflow/models/research/slim
sudo pip2 install -e .