train2.sh

#!/usr/bin/env bash

# Variables to modify ---------------------------------------------------------
MODEL_NAME=inception_resnet_v2
MODEL_ARG_SCOPE=InceptionResnetV2
PRETRAINED_MODEL_FILE=pretrained/inception_resnet_v2_2016_08_30.ckpt

NUM_TRAIN_RUNS=5
NUM_CLONES=2
BATCH_SIZE=32

# -----------------------------------------------------------------------------

# Loop over each dataset
DATASET_NAME=ImageCLEFmed2016partial-ara-autocrop
DATASET_DIR=datasets/${DATASET_NAME}
TRAIN_DIR_BASE=models/${MODEL_NAME}/${DATASET_NAME}

for((CURRENT_TRAINING_RUN=0; CURRENT_TRAINING_RUN<${NUM_TRAIN_RUNS}; CURRENT_TRAINING_RUN++))
do
    # Set training dir for current run
    TRAIN_DIR=${TRAIN_DIR_BASE}/run_${CURRENT_TRAINING_RUN}
    LOG_FILE=${TRAIN_DIR}/log.txt

    # Check if run was already trained successfully
    if test -e ${TRAIN_DIR}/DONE;then
        continue
    fi

    # Create train dir for log file
    mkdir -p ${TRAIN_DIR}

    # Fine-tune new layer for 10 epochs
    python slim/train_image_classifier.py \
        --train_dir=${TRAIN_DIR} \
        --dataset_dir=${DATASET_DIR} \
        --dataset_name=ImageCLEFmed2016 \
        --dataset_split_name=train \
        --preprocessing_name=ImageCLEFmed2016 \
        --model_name=${MODEL_NAME} \
        --batch_size=${BATCH_SIZE} \
        --num_clones=${NUM_CLONES} \
        --max_number_of_steps=1298 \
        --learning_rate=0.0001 \
        --learning_rate_decay_type=fixed \
        --weight_decay=0.0001 \
        --save_summaries_secs=30 \
        --save_interval_secs=1800 \
        --checkpoint_path=${PRETRAINED_MODEL_FILE} \
        --checkpoint_exclude_scopes=${MODEL_ARG_SCOPE}/Logits,${MODEL_ARG_SCOPE}/AuxLogits \
        --trainable_scopes=${MODEL_ARG_SCOPE}/Logits,${MODEL_ARG_SCOPE}/AuxLogits \
    2>&1 | tee ${LOG_FILE}

    python slim/eval_image_classifier.py \
        --checkpoint_path=${TRAIN_DIR} \
        --eval_dir=${TRAIN_DIR} \
        --dataset_dir=${DATASET_DIR} \
        --dataset_name=ImageCLEFmed2016 \
        --dataset_split_name=validation \
        --preprocessing_name=ImageCLEFmed2016 \
        --model_name=${MODEL_NAME} \
    2>&1 | tee -a ${LOG_FILE}

    # Fine-tune all layers for 20 epochs
    python slim/train_image_classifier.py \
        --train_dir=${TRAIN_DIR}/all \
        --dataset_dir=${DATASET_DIR} \
        --dataset_name=ImageCLEFmed2016 \
        --dataset_split_name=train \
        --preprocessing_name=ImageCLEFmed2016 \
        --model_name=${MODEL_NAME} \
        --batch_size=${BATCH_SIZE} \
        --num_clones=${NUM_CLONES} \
        --max_number_of_steps=2595 \
        --learning_rate=0.01 \
        --end_learning_rate=0.0001 \
        --learning_rate_decay_type=polynomial \
        --polynomial_decay_power=2.0 \
        --decay_steps=2000 \
        --weight_decay=0.0001 \
        --save_summaries_secs=30 \
        --save_interval_secs=1800 \
        --checkpoint_path=${TRAIN_DIR} \
    2>&1 | tee -a ${LOG_FILE}

    python slim/eval_image_classifier.py \
        --checkpoint_path=${TRAIN_DIR}/all \
        --eval_dir=${TRAIN_DIR}/all \
        --dataset_dir=${DATASET_DIR} \
        --dataset_name=ImageCLEFmed2016 \
        --dataset_split_name=validation \
        --preprocessing_name=ImageCLEFmed2016 \
        --model_name=${MODEL_NAME} \
    2>&1 | tee -a ${LOG_FILE}

    # Append accuracy to file
    grep "eval/Accuracy" ${LOG_FILE} \
    | tail --lines=1 \
    | grep -o "0\.[[:digit:]]*" \
    >> ${TRAIN_DIR_BASE}/results.txt

    # Create empty file to flag this run as done
    touch ${TRAIN_DIR}/DONE

done