Image Classification Hyperparameters

Hyperparameters are parameters that are set before a machine learning model begins learning. The following hyperparameters are supported by the Amazon SageMaker built-in Image Classification algorithm. See Tune an Image Classification Model for information on image classification hyperparameter tuning.

Parameter Name	Description
num_classes	Number of output classes. This parameter defines the dimensions of the network output and is typically set to the number of classes in the dataset. Besides multi-class classification, multi-label classification is supported too. Please refer to Input/Output Interface for the Image Classification Algorithm for details on how to work with multi-label classification with augmented manifest files. Required Valid values: positive integer
num_training_samples	Number of training examples in the input dataset. If there is a mismatch between this value and the number of samples in the training set, then the behavior of the lr_scheduler_step parameter is undefined and distributed training accuracy might be affected. Required Valid values: positive integer
augmentation_type	Data augmentation type. The input images can be augmented in multiple ways as specified below. crop: Randomly crop the image and flip the image horizontally crop_color: In addition to ‘crop’, three random values in the range [-36, 36], [-50, 50], and [-50, 50] are added to the corresponding Hue-Saturation-Lightness channels respectively crop_color_transform: In addition to crop_color, random transformations, including rotation, shear, and aspect ratio variations are applied to the image. The maximum angle of rotation is 10 degrees, the maximum shear ratio is 0.1, and the maximum aspect changing ratio is 0.25. Optional Valid values: crop, crop_color, or crop_color_transform. Default value: no default value
beta_1	The beta1 for adam, that is the exponential decay rate for the first moment estimates. Optional Valid values: float. Range in [0, 1]. Default value: 0.9
beta_2	The beta2 for adam, that is the exponential decay rate for the second moment estimates. Optional Valid values: float. Range in [0, 1]. Default value: 0.999
checkpoint_frequency	Period to store model parameters (in number of epochs). Note that all checkpoint files are saved as part of the final model file "model.tar.gz" and uploaded to S3 to the specified model location. This increases the size of the model file proportionally to the number of checkpoints saved during training. Optional Valid values: positive integer no greater than epochs. Default value: no default value (Save checkpoint at the epoch that has the best validation accuracy)
early_stopping	True to use early stopping logic during training. False not to use it. Optional Valid values: True or False Default value: False
early_stopping_min_epochs	The minimum number of epochs that must be run before the early stopping logic can be invoked. It is used only when early_stopping = True. Optional Valid values: positive integer Default value: 10
early_stopping_patience	The number of epochs to wait before ending training if no improvement is made in the relevant metric. It is used only when early_stopping = True. Optional Valid values: positive integer Default value: 5
early_stopping_tolerance	Relative tolerance to measure an improvement in accuracy validation metric. If the ratio of the improvement in accuracy divided by the previous best accuracy is smaller than the early_stopping_tolerance value set, early stopping considers there is no improvement. It is used only when early_stopping = True. Optional Valid values: 0 ≤ float ≤ 1 Default value: 0.0
epochs	Number of training epochs. Optional Valid values: positive integer Default value: 30
eps	The epsilon for adam and rmsprop. It is usually set to a small value to avoid division by 0. Optional Valid values: float. Range in [0, 1]. Default value: 1e-8
gamma	The gamma for rmsprop, the decay factor for the moving average of the squared gradient. Optional Valid values: float. Range in [0, 1]. Default value: 0.9
image_shape	The input image dimensions, which is the same size as the input layer of the network. The format is defined as 'num_channels, height, width'. The image dimension can take on any value as the network can handle varied dimensions of the input. However, there may be memory constraints if a larger image dimension is used. Pretrained models can only use a fixed height x width image size of 224 x 224. Typical image dimensions for image classification are '3, 224, 224'. This is similar to the ImageNet dataset. Optional Valid values: string Default value: ‘3, 224, 224’
kv_store	Weight update synchronization mode during distributed training. The weight updates can be updated either synchronously or asynchronously across machines. Synchronous updates typically provide better accuracy than asynchronous updates but can be slower. See distributed training in MXNet for more details. This parameter is not applicable to single machine training. dist_sync: The gradients are synchronized after every batch with all the workers. With dist_sync, batch-size now means the batch size used on each machine. So if there are n machines and we use batch size b, then dist_sync behaves like local with batch size n*b dist_async: Performs asynchronous updates. The weights are updated whenever gradients are received from any machine and the weight updates are atomic. However, the order is not guaranteed. Optional Valid values: dist_sync or dist_async Default value: no default value
learning_rate	Initial learning rate. Optional Valid values: float. Range in [0, 1]. Default value: 0.1
lr_scheduler_factor	The ratio to reduce learning rate used in conjunction with the lr_scheduler_step parameter, defined as lr_new = lr_old * lr_scheduler_factor. Optional Valid values: float. Range in [0, 1]. Default value: 0.1
lr_scheduler_step	The epochs at which to reduce the learning rate. As explained in the lr_scheduler_factor parameter, the learning rate is reduced by lr_scheduler_factor at these epochs. For example, if the value is set to "10, 20", then the learning rate is reduced by lr_scheduler_factor after 10th epoch and again by lr_scheduler_factor after 20th epoch. The epochs are delimited by ",". Optional Valid values: string Default value: no default value
mini_batch_size	The batch size for training. In a single-machine multi-GPU setting, each GPU handles mini_batch_size/num_gpu training samples. For the multi-machine training in dist_sync mode, the actual batch size is mini_batch_size*number of machines. See MXNet docs for more details. Optional Valid values: positive integer Default value: 32
momentum	The momentum for sgd and nag, ignored for other optimizers. Optional Valid values: float. Range in [0, 1]. Default value: 0.9
multi_label	Flag to use for multi-label classification where each sample can be assigned multiple labels. Average accuracy across all classes is logged. Optional Valid values: 0 or 1 Default value: 0
num_layers	Number of layers for the network. For data with large image size (for example, 224x224 - like ImageNet), we suggest selecting the number of layers from the set [18, 34, 50, 101, 152, 200]. For data with small image size (for example, 28x28 - like CIFAR), we suggest selecting the number of layers from the set [20, 32, 44, 56, 110]. The number of layers in each set is based on the ResNet paper. For transfer learning, the number of layers defines the architecture of base network and hence can only be selected from the set [18, 34, 50, 101, 152, 200]. Optional Valid values: positive integer in [18, 34, 50, 101, 152, 200] or [20, 32, 44, 56, 110] Default value: 152
optimizer	The optimizer type. For more details of the parameters for the optimizers, please refer to MXNet's API. Optional Valid values: One of sgd, adam, rmsprop, or nag. sgd: Stochastic gradient descent adam: Adaptive momentum estimation rmsprop: Root mean square propagation nag: Nesterov accelerated gradient Default value: sgd
precision_dtype	The precision of the weights used for training. The algorithm can use either single precision (float32) or half precision (float16) for the weights. Using half-precision for weights results in reduced memory consumption. Optional Valid values: float32 or float16 Default value: float32
resize	The number of pixels in the shortest side of an image after resizing it for training. If the parameter is not set, then the training data is used without resizing. Required when using image content types Optional when using the RecordIO content type Valid values: positive integer Default value: no default value
top_k	Reports the top-k accuracy during training. This parameter has to be greater than 1, since the top-1 training accuracy is the same as the regular training accuracy that has already been reported. Optional Valid values: positive integer larger than 1. Default value: no default value
use_pretrained_model	Flag to use pre-trained model for training. If set to 1, then the pretrained model with the corresponding number of layers is loaded and used for training. Only the top FC layer are reinitialized with random weights. Otherwise, the network is trained from scratch. Optional Valid values: 0 or 1 Default value: 0
use_weighted_loss	Flag to use weighted cross-entropy loss for multi-label classification (used only when multi_label = 1), where the weights are calculated based on the distribution of classes. Optional Valid values: 0 or 1 Default value: 0
weight_decay	The coefficient weight decay for sgd and nag, ignored for other optimizers. Optional Valid values: float. Range in [0, 1]. Default value: 0.0001