training-scripts.txt

#########
#RMSProp#
#########
CUDA_VISIBLE_DEVICES=0 python train_image_classifier.py \
--train_dir=${TRAIN_DIR} \
--dataset_name=imagenet \
--dataset_split_name=train \
--dataset_dir=${DATASET_DIR} \
--model_name=inception_v3 \
--checkpoint_path=${CHECKPOINT_FILE} \
--clone_on_cpu=False \
--checkpoint_exclude_scopes=InceptionV3/Logits,InceptionV3/AuxLogits \
--max_number_of_steps=300 \
--save_checkpoints_steps=100 \
--save_summaries_steps=100 \
--batch_size=64 \
--learning_rate=0.001 \
--end_learning_rate=0.00001 \
--learning_rate_decay_type='exponential' \
--learning_rate_decay_factor=0.94 \
--num_epochs_per_decay=0.05 \
--optimizer='rmsprop'

#########################
#Newton Hessian Trick CG#
#########################

CUDA_VISIBLE_DEVICES=0 python train_image_classifier.py \
--train_dir=${TRAIN_DIR} \
--dataset_name=imagenet \
--dataset_split_name=train \
--dataset_dir=${DATASET_DIR} \
--model_name=inception_v3 \
--checkpoint_path=${CHECKPOINT_FILE} \
--clone_on_cpu=False \
--checkpoint_exclude_scopes=InceptionV3/Logits,InceptionV3/AuxLogits \
--max_number_of_steps=300 \
--save_checkpoints_steps=100 \
--save_summaries_steps=100 \
--batch_size=64 \
--learning_rate=0.001 \
--end_learning_rate=0.00001 \
--learning_rate_decay_type='exponential' \
--learning_rate_decay_factor=0.94 \
--num_epochs_per_decay=0.05 \
--optimizer='second' \
--eso_epsilon=0.01 \
--eso_cg_tol=0.00001 \
--eso_max_iter=20

#######################
# Just the last layer #
#######################
CUDA_VISIBLE_DEVICES=0 screen -d -m -L \
python ../train_image_classifier.py \
--train_dir=${TRAIN_DIR} \
--dataset_name=imagenet \
--dataset_split_name=train \
--dataset_dir=${DATASET_DIR} \
--clone_on_cpu=False \
--model_name=inception_v3 \
--checkpoint_path=${CHECKPOINT_FILE} \
--checkpoint_exclude_scopes=InceptionV3/Logits,InceptionV3/AuxLogits \
--trainable_scopes=InceptionV3/Logits,InceptionV3/AuxLogits \
--max_number_of_steps=100 \
--save_checkpoints_steps=300 \
--save_summaries_steps=100 \
--profiling_enabled=False \
--profile_every_n_steps=100 \
--profile_dir='profiles/InceptionV3' \
--batch_size=64 \
--learning_rate=0.001 \
--end_learning_rate=0.000001 \
--learning_rate_decay_type='exponential' \
--learning_rate_decay_factor=0.94 \
--num_epochs_per_decay=0.05 \
--optimizer='second' \
--eso_epsilon=0.01 \
--eso_cg_tol=0.00001 \
--eso_max_iter=20

# my default start learning rate: 0.001
# my default eso_epsilon: 0.01
# after 1k steps
# 1/10th of the default learning rates.
# learning rate decays every ca. 1.18k steps
# batch size 2*default
# default learning rate decay type
# default decay rate 0.94, now rate 0.74
# Store checkpoints and summaries every 100 steps.

###############################################
# Pretrained model checkpoint names.
# InceptionV3
PRETRAINED_TGZ='inception_v3_2016_08_28.tar.gz'
PRETRAINED_LINK="http://download.tensorflow.org/models/$PRETRAINED_TGZ"
#MobileNetV2
PRETRAINED_TGZ='mobilenet_v2_1.4_224.tgz'
PRETRAINED_LINK="https://storage.googleapis.com/$MODEL_NAME/checkpoints/$PRETRAINED_TGZ"