Environment
- ubuntu16.04
- python2.7
- tensorflow
Github
Preparing the datasets
- Prepare your own data and divide it into train(90%) and validation(10%) sets
- Put in "data_prepare/pic/train" and "data_prepare/pic/validation"
pic
--train
--class 0
--class 1
...
--validation
--class 0
--class 1
...Converting to TFRecord format
- Execute the "convert.sh" file
python data_convert.py -t pic/ \
--train-shards 2 \
--validation-shards 2 \
--num-threads 2 \
--dataset-name satellite- Will get 4 tf-record files and 1 label file
- Move these 5 files to "slim\satellite\data"
Modify the "slim\datasets\satellite.py" file
- FILE_PATTERN = ‘satellite%s_*.tfrecord’ （ not recommend to modify）
- SPLITS_TO_SIZES = {‘train’: 16, ‘validation’: 4} （Total number of train and validation set ）（The sum of all categories）
- _NUM_CLASSES = 2 （Number of categories）
- ‘image/format’: tf.FixedLenFeature((), tf.string, default_value=’jpg’) （Image format, here is jpg）
Download the pre-training model Mobilenet V2
- https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet
- After unzipping, copy the 3 ckpt files to "slim\satellite\pretrained"
Execute "model_train.sh" in the "slim/" folder to train:
python train_image_classifier.py \
--train_dir=satellite/train_log \
--dataset_name=satellite \
--train_image_size=500 \
--dataset_split_name=train \
--dataset_dir=satellite/data \
--model_name="mobilenet_v2_140" \
--checkpoint_path=satellite/pretrained/mobilenet_v2_1.4_224.ckpt \
--checkpoint_exclude_scopes=MobilenetV2/Logits,MobilenetV2/AuxLogits \
--trainable_scopes=MobilenetV2/Logits,MobilenetV2/AuxLogits \
--max_number_of_steps=1000 \
--batch_size=16 \
--learning_rate=0.001 \
--learning_rate_decay_type=fixed \
--log_every_n_steps=10 \
--optimizer=rmsprop \
--weight_decay=0.00004 \
--label_smoothing=0.1 \
--num_clones=1 \
--num_epochs_per_decay=2.5 \
--moving_average_decay=0.9999 \
--learning_rate_decay_factor=0.98 \
--preprocessing_name="inception_v2" Execute "model_test.sh" in the "slim/" folder to evaluate the model:
python eval_image_classifier.py \
--checkpoint_path=satellite/train_log \
--eval_dir=satellite/eval_log \
--dataset_name=satellite \
--dataset_split_name=validation \
--dataset_dir=satellite/data \
--model_name="mobilenet_v2_140" \
--batch_size=32 \
--num_preprocessing_threads=2 \
--eval_image_size=500-
Or execute "start_train_test.sh" at the project root directory to train and evaluate the model.
Export trained models
- Execute "export_inference_graph.py" in the "slim/" folder to exporting the Inference Graph
python export_inference_graph.py \
--alsologtostderr \
--model_name="mobilenet_v2_140" \
--image_size=500 \
--output_file=slim/satellite/export/mobilenet_v2_140_inf_graph.pb \
--dataset_name satellite- Execute "freeze_graph.py" at the project root directory to freezing the exported Graph
- Remember to modify the train steps "-****" in "--input_checkpoint"
python freeze_graph.py \
--input_graph slim/satellite/export/mobilenet_v2_140_inf_graph.pb \
--input_checkpoint slim/satellite/train_log/model.ckpt-****(8000) \
--input_binary true \
--output_node_names MobilenetV2/Predictions/Reshape_1 \
--output_graph slim/satellite/freeze/mobilenet_v2_140.pb
- Or execute "export_freeze.sh" at the project root directory to export and freeze Graph.