The image generation component enables training of generative adversarial network (GAN) models for high-quality image synthesis. It supports mainstream GAN architectures including Deep Convolutional GAN (DCGAN), Wasserstein GAN with Gradient Penalty (WGAN-GP), Least Squares GAN (LSGAN), Graph GAN (GGAN), Progressive Growing of GAN (PGGAN), and StyleGAN to generate diverse, high-fidelity images from training data.
Supported computing resources
Inputs and outputs
Input ports
-
Connect the Read File Data component to specify the Object Storage Service (OSS) path containing your training images.
-
Alternatively, configure the oss path to train data parameter directly to specify the OSS path where training images are stored.
Output port
The trained GAN model is saved to the OSS path specified by the oss path to output distributed parameter.
The following table describes the component parameters.
Configure the component
In Machine Learning Designer, add the image generation component to your pipeline and configure the parameters described in the following table.
|
Tab |
Parameter |
Required |
Description |
Default value |
|
|
Fields Setting |
oss path to train data |
No |
If you do not connect an upstream data source, manually specify the OSS path where training images are stored. |
None |
|
|
oss path to pretrained model |
No |
The OSS path to a pre-trained model for transfer learning. If left empty, training starts from scratch. The pre-trained model must be stored in the same OSS bucket as your training images. |
None |
||
|
oss path to output distributed |
Yes |
The OSS path where the trained model will be saved. The output path must be in the same OSS bucket as your training images. |
None |
||
|
Parameters Setting |
image generation network |
Yes |
The GAN architecture to use for training. Available options: DCGAN, WGAN-GP, LSGAN, GGAN, PGGAN, and StyleGAN. |
dcgan |
|
|
learning rate |
Yes |
The learning rate for model optimization. |
0.001 |
||
|
number of train iterations |
Yes |
The total number of training iterations. Each iteration processes one batch of images (determined by the batch size parameter). |
10000 |
||
|
visualization iterations |
Yes |
The interval at which the model generates sample images from random noise for visual inspection. Generated samples are saved to the training_samples folder in the path specified by oss path to output distributed. |
1000 |
||
|
batch size |
Yes |
The number of images processed in each training iteration. |
32 |
||
|
model save interval |
No |
The frequency at which model checkpoints are saved during training (measured in epochs). |
1 |
||
|
Tuning |
Select Resource Group |
Public Resource Group |
No |
Select the instance type and VPC for training. A GPU instance type is required for this algorithm. |
None |
|
Dedicated resource group |
No |
Specify the compute resources: CPU cores, memory, shared memory, and number of GPUs. |
None |
||
|
Maximum Running Duration (seconds) |
No |
The maximum execution time for the component. Training terminates if this duration is exceeded. |
None |
||