Cookies help us deliver our services. By using our services, you agree to our use of cookies. Learn more

crowdAI is shutting down - please read our blog post for more information

Dark Skies

Tensorflow solution for the Dark Skies Challenge

We are now ready to fine-tune a pre-trained Inception-v3 model on the darkskies-challenge data set. This requires two notable steps:
- Build the exact same model as pretrained Inception-v3 except we change the number of labels in the final classification layer.
- Restore all weights from the pre-trained Inception-v3 except for the final classification layer; this will get randomly initialized instead.

We can perform these two operations by specifying two flags: --pretrained_model_checkpoint_path and --fine_tune.

The first flag is a string that points to the path of a pre-trained Inception-v3 model. If this flag is specified, it will load the entire model from the checkpoint before the script begins training.

The second flag --fine_tune is a boolean that indicates whether the last classification layer should be randomly initialized or restored. You may set this flag to false if you wish to continue training a pre-trained model from a checkpoint. If you set this flag to true, you can train a new classification layer from scratch.

Putting this all together you can retrain a pre-trained Inception-v3 model on the darkskies-challenge data set with the following commands.

## Finetune 
# Build the training binary to run on a GPU. If you do not have a GPU, 
# then exclude '--config=cuda' 
bazel build -c opt --config=cuda inception/dk_train 

# Directory where to save the checkpoint and events files. 
# Directory where preprocessed TFRecord files reside. 
# Path to the downloaded Inception-v3 model. 
# Run the fine-tuning on the dark-challenge dataset starting from the pre-trained 
# inception-v3 model. 
bazel-bin/inception/dk_train --train_dir="${FINETUNE_DIR}" --data_dir="${DK_DATA_DIR}" --pretrained_model_checkpoint_path="${MODEL_PATH}" --fine_tune=True --initial_learning_rate=0.001 --batch_size=32 --input_queue_memory_factor=8 –num_gpus=1 --num_epochs_per_decay=20 --max_steps=1000000

Fine-tuning a model a separate data set requires significantly lowering the initial learning rate. We set the initial learning rate to 0.001.

Now the training is in progress, it constantly outputs to terminal screen:

2016-10-13 11:56:22.949164: step 0, loss = 3.11 (1.6 examples/sec; 20.053 sec/batch)
2016-10-13 11:56:48.508299: step 10, loss = 2.55 (46.3 examples/sec; 0.692 sec/batch)
2016-10-13 11:56:55.458712: step 20, loss = 2.49 (42.9 examples/sec; 0.746 sec/batch)
2016-10-13 11:57:02.557317: step 30, loss = 2.43 (45.7 examples/sec; 0.700 sec/batch)
2016-10-13 11:57:09.584892: step 40, loss = 2.39 (45.1 examples/sec; 0.710 sec/batch)
2016-10-13 11:57:16.581422: step 50, loss = 2.20 (45.7 examples/sec; 0.700 sec/batch)
2016-10-13 11:57:23.572435: step 60, loss = 1.51 (46.2 examples/sec; 0.693 sec/batch)
2016-10-13 11:57:30.571183: step 70, loss = 1.97 (45.6 examples/sec; 0.701 sec/batch)
2016-10-13 11:57:37.520570: step 80, loss = 1.80 (46.0 examples/sec; 0.696 sec/batch)
2016-10-13 11:57:44.490582: step 90, loss = 1.62 (45.9 examples/sec; 0.698 sec/batch)
2016-10-13 11:57:51.469971: step 100, loss = 1.51 (46.0 examples/sec; 0.696 sec/batch)

The loss should be decreased gradually. I trained this model for 45000 steps.