-
Notifications
You must be signed in to change notification settings - Fork 4
Supported Flags for Retraining
Debayan Deb edited this page Oct 25, 2017
·
6 revisions
This page details all the supported flags for retraining.
| FLAG | Help |
|---|---|
| image_dir | Relative path to the root folder of your labeled dataset (Default '') |
| output_graph | Path where the trained frozen graph is saved (Default 'tmp/output_graph.pb') |
| intermediate_output_graphs_dir | Path where the intermediate frozen graphs are saved (Default '') |
| intermediate_store_frequency | Number of steps to store intermediate frozen graphs. If 0, then no storing. (Default 0) |
| output_labels | Path to save the trained graph's output labels (Default 'tmp/output_labels.txt') |
| summaries_dir | Path to save the summary logs for Tensorboard Visualization (Default 'tmp/retrain_logs') |
| model_dir | Path to save the frozen ImageNet pre-trained (Inception_v3 or Mobilenet) models (Default 'tmp/imagenet') |
| bottleneck_dir | Path to cache bottleneck layer values as text files (Default 'tmp/bottleneck') |
| FLAG | Help |
|---|---|
| testing_percentage | Percentage of images per class to use as test set (Default 10) |
| validation_percentage | Percentage of images per class to use as validation set (Default 10) |
| flip_left_right | Whether to randomly flip half of the training images horizontally (Default False) |
| random_scale | Percentage to randomly scale up the size of the training images (Default 0) |
| random_crop | Percentage determining the margin to randomly crop off the training images (Default 0) |
| random_brightness | Percentage determining how much to multiply the training image pixels up or down by (Default 0) |
| FLAGS | Help |
|---|---|
| learning_rate | Learning rate parameter (Default 0.01) |
| how_many_training_steps | Number of training steps to run until finish (Default 4000) |
| eval_step_interval | Frequency of evaluating training results (Default 10) |
| train_batch_size | Number of images to train at a time (Default 100) |
| test_batch_size | Number of images to test on. The test set is evaluated only once after training ceases. -1 implies entire test dataset is evaluated at a time, giving more stable results across runs. (Default -1) |
| validation_batch_size | Number of images to use in evaluation batch. The validation set is more often than testing set and gives an early indication on the progress of training. -1 implies entire test dataset is evaluated at a time, giving more stable results across runs, but might be slower on larger datasets. (Default -1) |
| print_misclassified_test_images | Whether to print out a list of all misclassified test images (Default False) |
| center_loss | Use center loss along-side cross-entropy (Softmax) loss function (Default False) |
| center_loss_alpha | Amount to update the center by (Default 0) |
| center_loss_factor | Ratio of the affect of center loss on total loss (Default 0) |
| FLAGS | Help |
|---|---|
| final_tensor_name | Name of the output classification layer in the retrained graph |
| architecture | Which model architecture to use. inception_v3 is the most accurate, but also the slowest. For faster or smaller models, chose a MobileNet with the form mobilenet_<parameter size>_<input_size>[_quantized]. For example, mobilenet_1.0_224 will pick a model that is 17 MB in size and takes 224 pixel input images, while mobilenet_0.25_128_quantized will choose a much less accurate, but smaller and faster network that's 920 KB on disk and takes 128x128 images. |