Migration of Distributed Training Settings¶

We have merged MMGeneration 1.x into MMagic. Here is migration of Distributed Training Settings about MMGeneration.

In 0.x version, MMGeneration uses DDPWrapper and DynamicRunner to train static and dynamic model (e.g., PGGAN and StyleGANv2) respectively. In 1.x version, we use MMSeparateDistributedDataParallel provided by MMEngine to implement distributed training.

The configuration differences are shown below:

Static Model in 0.x Version

Static Model in 1.x Version

# Use DDPWrapper
use_ddp_wrapper = True
find_unused_parameters = False

runner = dict(
    type='DynamicIterBasedRunner',
    is_dynamic_ddp=False)

model_wrapper_cfg = dict(
    type='MMSeparateDistributedDataParallel',
    broadcast_buffers=False,
    find_unused_parameters=False)

Dynamic Model in 0.x Version

Dynamic Model in 1.x Version

use_ddp_wrapper = False
find_unused_parameters = False

# Use DynamicRunner
runner = dict(
    type='DynamicIterBasedRunner',
    is_dynamic_ddp=True)

model_wrapper_cfg = dict(
    type='MMSeparateDistributedDataParallel',
    broadcast_buffers=False,
    find_unused_parameters=True) # set `find_unused_parameters` for dynamic models