[MMSIG] Add new configuration files for StyleGAN2 (#2057)

* 1st * debug * 20230710 调整 * 调整代码，整合模型，避免editors import 过多class * 调整代码，整合模型，避免editors import 过多class * 支持 DeblurGANv2 inference demo 示例 python mmagic\demo\mmagic_inference_demo.py --model-name deblurganv2 --model-comfig ../configs/deblurganv2/deblurganv2_fpn_inception.py --model-ckpt 权重文件路径 --img 测试图片路径 --device cpu --result-out-dir ./out.png * 支持 DeblurGANv2 inference fix CI test * 支持 DeblurGANv2 inference fix CI test * 支持 DeblurGANv2 inference fix CI test * 支持 DeblurGANv2 inference update model-index * 支持 DeblurGANv2 inference Fix CI Test * 支持 DeblurGANv2 inference CI test fix and update readme.md * 支持 DeblurGANv2 inference fix CI test and update readme.md * 支持 DeblurGANv2 inference Fix CI test * 支持 DeblurGANv2 inference yapf 修正 * 支持 DeblurGANv2 inference 代码调整，参数名保持一致 * 支持 DeblurGANv2 inference doc string coverage * 支持 DeblurGANv2 inference add some doc string * 支持 DeblurGANv2 inference add some doc string * 支持 DeblurGANv2 inference * 支持 DeblurGANv2 inference * 支持 DeblurGANv2 inference * 支持 DeblurGANv2 inference * 支持 DeblurGANv2 inference add unit test * 支持 DeblurGANv2 inference add unit test * 支持 DeblurGANv2 inference add unit test * 支持 DeblurGANv2 inference fix unit test * Update .gitignore Co-authored-by: Yanhong Zeng <[email protected]> * Update .gitignore Co-authored-by: Yanhong Zeng <[email protected]> * Update .gitignore Co-authored-by: Yanhong Zeng <[email protected]> * Update .gitignore Co-authored-by: Yanhong Zeng <[email protected]> * Update .gitignore Co-authored-by: Yanhong Zeng <[email protected]> * Update .gitignore Co-authored-by: Yanhong Zeng <[email protected]> * Update .gitignore Co-authored-by: Yanhong Zeng <[email protected]> * Update configs/deblurganv2/README.md Co-authored-by: Yanhong Zeng <[email protected]> * 支持 DeblurGANv2 inference move the implementation of loss function to mmagic/models/losses add quick start to readme * 支持 DeblurGANv2 inference fix unit test * 支持 DeblurGANv2 inference re run unit test * Update configs/deblurganv2/deblurganv2_fpn-inception_1xb1_gopro.py Co-authored-by: Yanhong Zeng <[email protected]> * Update configs/deblurganv2/deblurganv2_fpn-inception_1xb1_gopro.py Co-authored-by: Yanhong Zeng <[email protected]> * Update configs/deblurganv2/deblurganv2_fpn-inception_1xb1_gopro.py Co-authored-by: Yanhong Zeng <[email protected]> * Update configs/deblurganv2/deblurganv2_fpn-inception_1xb1_gopro.py Co-authored-by: Yanhong Zeng <[email protected]> * Update configs/deblurganv2/deblurganv2_fpn-mobilenet_1xb1_gopro.py Co-authored-by: Yanhong Zeng <[email protected]> * Update configs/deblurganv2/deblurganv2_fpn-mobilenet_1xb1_gopro.py Co-authored-by: Yanhong Zeng <[email protected]> * Update configs/deblurganv2/deblurganv2_fpn-mobilenet_1xb1_gopro.py Co-authored-by: Yanhong Zeng <[email protected]> * Update configs/deblurganv2/deblurganv2_fpn-mobilenet_1xb1_gopro.py Co-authored-by: Yanhong Zeng <[email protected]> * 支持 DeblurGANv2 inference fix some url and path add README_zh-CN add Deblurring task into mmagic/apis/inferencers/__init__.py * Adding support for FastComposer 支持 FastComposer * Adding support for FastComposer Add some doc string and fix bugs * Adding support for FastComposer Fixed a bug * Adding support for FastComposer fix a bug * Adding support for FastComposer fix some bugs * Adding support for FastComposer fix some bugs * Adding support for FastComposer fix a bug * Adding support for FastComposer fix a bug * Adding support for FastComposer change for minimum version cpu check * Adding support for FastComposer avoid a windows CI bug which complains not enough memory. * Adding support for FastComposer rerun circleci check * Adding support for FastComposer add example code which run without gradio to readme add config of clip for running unittest without using "clip_vit_url = 'openai/clip-vit-large-patch14' " * Adding support for FastComposer rerun checks of build cu102 * Adding support for FastComposer a small change * Adding support for FastComposer some small changes * Adding support for FastComposer add some simple instructions to demo/README.md * Adding support for FastComposer resolve conflicts * Adding support for FastComposer rerun checks * Adding support for FastComposer add device for running with cuda by default * Adding support for Consistency Models * Adding support for Consistency Models * Update README.md * Adding support for Consistency Models * Adding support for Consistency Models mdformat debug * Adding support for Consistency Models * Adding support for Consistency Models add some doc string * Adding support for Consistency Models re run circle check * Adding support for Consistency Models rerun circle check * [FIX] Check circle ci memory add function teardown_module to test_fastcomposer * Adding support for Consistency Models rerun ci check * Add new configuration files for StyleGAN2 * Revert "Add new configuration files for StyleGAN2" This reverts commit a043b22. * Add new configuration files for StyleGAN2 * fix config-vaildate error * fix a bug * delete code of consistency model * delete code which in another pr * delete code which in another pr * Add new configuration files for StyleGAN2 rerun ci check * ci check memory --------- Co-authored-by: Yanhong Zeng <[email protected]> Co-authored-by: rangoliu <[email protected]>
open-mmlab · Dec 11, 2023 · cd183d9 · cd183d9
1 parent 73d7b01
commit cd183d9
Show file tree

Hide file tree

Showing 10 changed files with 619 additions and 0 deletions.
diff --git a/mmagic/configs/styleganv2/stylegan2_c2_8xb4_800kiters_ffhq_256x256.py b/mmagic/configs/styleganv2/stylegan2_c2_8xb4_800kiters_ffhq_256x256.py
@@ -0,0 +1,89 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.config import read_base
+from torch.optim import Adam
+
+from mmagic.engine import VisualizationHook
+from mmagic.evaluation import (FrechetInceptionDistance, PerceptualPathLength,
+                               PrecisionAndRecall)
+from mmagic.models import BaseGAN
+
+with read_base():
+    from .._base_.datasets.ffhq_flip import *  # noqa: F403,F405
+    from .._base_.gen_default_runtime import *  # noqa: F403,F405
+    from .._base_.models.base_styleganv2 import *  # noqa: F403,F405
+
+# reg params
+d_reg_interval = 16
+g_reg_interval = 4
+
+g_reg_ratio = g_reg_interval / (g_reg_interval + 1)
+d_reg_ratio = d_reg_interval / (d_reg_interval + 1)
+
+ema_half_life = 10.  # G_smoothing_kimg
+
+model.update(
+    generator=dict(out_size=256),
+    discriminator=dict(in_size=256),
+    ema_config=dict(
+        type=ExponentialMovingAverage,
+        interval=1,
+        momentum=1. - (0.5**(32. / (ema_half_life * 1000.)))),
+    loss_config=dict(
+        r1_loss_weight=10. / 2. * d_reg_interval,
+        r1_interval=d_reg_interval,
+        norm_mode='HWC',
+        g_reg_interval=g_reg_interval,
+        g_reg_weight=2. * g_reg_interval,
+        pl_batch_shrink=2))
+
+train_cfg.update(max_iters=800002)
+
+optim_wrapper.update(
+    generator=dict(
+        optimizer=dict(
+            type=Adam, lr=0.002 * g_reg_ratio, betas=(0, 0.99**g_reg_ratio))),
+    discriminator=dict(
+        optimizer=dict(
+            type=Adam, lr=0.002 * d_reg_ratio, betas=(0, 0.99**d_reg_ratio))))
+
+batch_size = 4
+data_root = './data/ffhq/ffhq_imgs/ffhq_256'
+
+train_dataloader.update(
+    batch_size=batch_size, dataset=dict(data_root=data_root))
+
+val_dataloader.update(batch_size=batch_size, dataset=dict(data_root=data_root))
+
+test_dataloader.update(
+    batch_size=batch_size, dataset=dict(data_root=data_root))
+
+# VIS_HOOK
+custom_hooks = [
+    dict(
+        type=VisualizationHook,
+        interval=5000,
+        fixed_input=True,
+        vis_kwargs_list=dict(type=BaseGAN, name='fake_img'))
+]
+
+# METRICS
+metrics = [
+    dict(
+        type=FrechetInceptionDistance,
+        prefix='FID-50k',
+        fake_nums=50000,
+        real_nums=50000,
+        inception_style='StyleGAN',
+        sample_model='ema'),
+    dict(type=PrecisionAndRecall, fake_nums=50000, prefix='PR-50K'),
+    dict(type=PerceptualPathLength, fake_nums=50000, prefix='ppl-w')
+]
+# NOTE: config for save multi best checkpoints
+# default_hooks.update(
+#     checkpoint=dict(
+#         save_best=['FID-Full-50k/fid', 'IS-50k/is'],
+#         rule=['less', 'greater']))
+default_hooks.update(checkpoint=dict(save_best='FID-50k/fid'))
+
+val_evaluator.update(metrics=metrics)
+test_evaluator.update(metrics=metrics)
diff --git a/mmagic/configs/styleganv2/stylegan2_c2_8xb4_800kiters_lsun_cat_256x256.py b/mmagic/configs/styleganv2/stylegan2_c2_8xb4_800kiters_lsun_cat_256x256.py
@@ -0,0 +1,88 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.config import read_base
+from torch.optim import Adam
+
+from mmagic.engine import VisualizationHook
+from mmagic.evaluation import (FrechetInceptionDistance, PerceptualPathLength,
+                               PrecisionAndRecall)
+from mmagic.models import BaseGAN
+
+with read_base():
+    from .._base_.datasets.lsun_stylegan import *  # noqa: F403,F405
+    from .._base_.gen_default_runtime import *  # noqa: F403,F405
+    from .._base_.models.base_styleganv2 import *  # noqa: F403,F405
+
+# reg params
+d_reg_interval = 16
+g_reg_interval = 4
+
+g_reg_ratio = g_reg_interval / (g_reg_interval + 1)
+d_reg_ratio = d_reg_interval / (d_reg_interval + 1)
+
+ema_half_life = 10.  # G_smoothing_kimg
+
+model.update(
+    generator=dict(out_size=256),
+    discriminator=dict(in_size=256),
+    ema_config=dict(
+        type=ExponentialMovingAverage,
+        interval=1,
+        momentum=1. - (0.5**(32. / (ema_half_life * 1000.)))),
+    loss_config=dict(
+        r1_loss_weight=10. / 2. * d_reg_interval,
+        r1_interval=d_reg_interval,
+        norm_mode='HWC',
+        g_reg_interval=g_reg_interval,
+        g_reg_weight=2. * g_reg_interval,
+        pl_batch_shrink=2))
+
+train_cfg.update(max_iters=800002)
+
+optim_wrapper.update(
+    generator=dict(
+        optimizer=dict(
+            type=Adam, lr=0.002 * g_reg_ratio, betas=(0, 0.99**g_reg_ratio))),
+    discriminator=dict(
+        optimizer=dict(
+            type=Adam, lr=0.002 * d_reg_ratio, betas=(0, 0.99**d_reg_ratio))))
+
+batch_size = 4
+data_root = './data/lsun-cat'
+
+train_dataloader.update(
+    batch_size=batch_size, dataset=dict(data_root=data_root))
+
+val_dataloader.update(batch_size=batch_size, dataset=dict(data_root=data_root))
+
+test_dataloader.update(
+    batch_size=batch_size, dataset=dict(data_root=data_root))
+
+# VIS_HOOK
+custom_hooks = [
+    dict(
+        type=VisualizationHook,
+        interval=5000,
+        fixed_input=True,
+        vis_kwargs_list=dict(type=BaseGAN, name='fake_img'))
+]
+
+# METRICS
+metrics = [
+    dict(
+        type=FrechetInceptionDistance,
+        prefix='FID-Full-50k',
+        fake_nums=50000,
+        inception_style='StyleGAN',
+        sample_model='ema'),
+    dict(type=PrecisionAndRecall, fake_nums=50000, prefix='PR-50K'),
+    dict(type=PerceptualPathLength, fake_nums=50000, prefix='ppl-w')
+]
+# NOTE: config for save multi best checkpoints
+# default_hooks.update(
+#     checkpoint=dict(
+#         save_best=['FID-Full-50k/fid', 'IS-50k/is'],
+#         rule=['less', 'greater']))
+default_hooks.update(checkpoint=dict(save_best='FID-Full-50k/fid'))
+
+val_evaluator.update(metrics=metrics)
+test_evaluator.update(metrics=metrics)
diff --git a/mmagic/configs/styleganv2/stylegan2_c2_8xb4_800kiters_lsun_church_256x256.py b/mmagic/configs/styleganv2/stylegan2_c2_8xb4_800kiters_lsun_church_256x256.py
@@ -0,0 +1,88 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.config import read_base
+from torch.optim import Adam
+
+from mmagic.engine import VisualizationHook
+from mmagic.evaluation import (FrechetInceptionDistance, PerceptualPathLength,
+                               PrecisionAndRecall)
+from mmagic.models import BaseGAN
+
+with read_base():
+    from .._base_.datasets.lsun_stylegan import *  # noqa: F403,F405
+    from .._base_.gen_default_runtime import *  # noqa: F403,F405
+    from .._base_.models.base_styleganv2 import *  # noqa: F403,F405
+
+# reg params
+d_reg_interval = 16
+g_reg_interval = 4
+
+g_reg_ratio = g_reg_interval / (g_reg_interval + 1)
+d_reg_ratio = d_reg_interval / (d_reg_interval + 1)
+
+ema_half_life = 10.  # G_smoothing_kimg
+
+model.update(
+    generator=dict(out_size=256),
+    discriminator=dict(in_size=256),
+    ema_config=dict(
+        type=ExponentialMovingAverage,
+        interval=1,
+        momentum=1. - (0.5**(32. / (ema_half_life * 1000.)))),
+    loss_config=dict(
+        r1_loss_weight=10. / 2. * d_reg_interval,
+        r1_interval=d_reg_interval,
+        norm_mode='HWC',
+        g_reg_interval=g_reg_interval,
+        g_reg_weight=2. * g_reg_interval,
+        pl_batch_shrink=2))
+
+train_cfg.update(max_iters=800002)
+
+optim_wrapper.update(
+    generator=dict(
+        optimizer=dict(
+            type=Adam, lr=0.002 * g_reg_ratio, betas=(0, 0.99**g_reg_ratio))),
+    discriminator=dict(
+        optimizer=dict(
+            type=Adam, lr=0.002 * d_reg_ratio, betas=(0, 0.99**d_reg_ratio))))
+
+batch_size = 4
+data_root = './data/lsun-church'
+
+train_dataloader.update(
+    batch_size=batch_size, dataset=dict(data_root=data_root))
+
+val_dataloader.update(batch_size=batch_size, dataset=dict(data_root=data_root))
+
+test_dataloader.update(
+    batch_size=batch_size, dataset=dict(data_root=data_root))
+
+# VIS_HOOK
+custom_hooks = [
+    dict(
+        type=VisualizationHook,
+        interval=5000,
+        fixed_input=True,
+        vis_kwargs_list=dict(type=BaseGAN, name='fake_img'))
+]
+
+# METRICS
+metrics = [
+    dict(
+        type=FrechetInceptionDistance,
+        prefix='FID-Full-50k',
+        fake_nums=50000,
+        inception_style='StyleGAN',
+        sample_model='ema'),
+    dict(type=PrecisionAndRecall, fake_nums=50000, prefix='PR-50K'),
+    dict(type=PerceptualPathLength, fake_nums=50000, prefix='ppl-w')
+]
+# NOTE: config for save multi best checkpoints
+# default_hooks.update(
+#     checkpoint=dict(
+#         save_best=['FID-Full-50k/fid', 'IS-50k/is'],
+#         rule=['less', 'greater']))
+default_hooks.update(checkpoint=dict(save_best='FID-Full-50k/fid'))
+
+val_evaluator.update(metrics=metrics)
+test_evaluator.update(metrics=metrics)
diff --git a/mmagic/configs/styleganv2/stylegan2_c2_8xb4_800kiters_lsun_horse_256x256.py b/mmagic/configs/styleganv2/stylegan2_c2_8xb4_800kiters_lsun_horse_256x256.py
@@ -0,0 +1,88 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from mmengine.config import read_base
+from torch.optim import Adam
+
+from mmagic.engine import VisualizationHook
+from mmagic.evaluation import (FrechetInceptionDistance, PerceptualPathLength,
+                               PrecisionAndRecall)
+from mmagic.models import BaseGAN
+
+with read_base():
+    from .._base_.datasets.lsun_stylegan import *  # noqa: F403,F405
+    from .._base_.gen_default_runtime import *  # noqa: F403,F405
+    from .._base_.models.base_styleganv2 import *  # noqa: F403,F405
+
+# reg params
+d_reg_interval = 16
+g_reg_interval = 4
+
+g_reg_ratio = g_reg_interval / (g_reg_interval + 1)
+d_reg_ratio = d_reg_interval / (d_reg_interval + 1)
+
+ema_half_life = 10.  # G_smoothing_kimg
+
+model.update(
+    generator=dict(out_size=256),
+    discriminator=dict(in_size=256),
+    ema_config=dict(
+        type=ExponentialMovingAverage,
+        interval=1,
+        momentum=1. - (0.5**(32. / (ema_half_life * 1000.)))),
+    loss_config=dict(
+        r1_loss_weight=10. / 2. * d_reg_interval,
+        r1_interval=d_reg_interval,
+        norm_mode='HWC',
+        g_reg_interval=g_reg_interval,
+        g_reg_weight=2. * g_reg_interval,
+        pl_batch_shrink=2))
+
+train_cfg.update(max_iters=800002)
+
+optim_wrapper.update(
+    generator=dict(
+        optimizer=dict(
+            type=Adam, lr=0.002 * g_reg_ratio, betas=(0, 0.99**g_reg_ratio))),
+    discriminator=dict(
+        optimizer=dict(
+            type=Adam, lr=0.002 * d_reg_ratio, betas=(0, 0.99**d_reg_ratio))))
+
+batch_size = 4
+data_root = './data/lsun-horse'
+
+train_dataloader.update(
+    batch_size=batch_size, dataset=dict(data_root=data_root))
+
+val_dataloader.update(batch_size=batch_size, dataset=dict(data_root=data_root))
+
+test_dataloader.update(
+    batch_size=batch_size, dataset=dict(data_root=data_root))
+
+# VIS_HOOK
+custom_hooks = [
+    dict(
+        type=VisualizationHook,
+        interval=5000,
+        fixed_input=True,
+        vis_kwargs_list=dict(type=BaseGAN, name='fake_img'))
+]
+
+# METRICS
+metrics = [
+    dict(
+        type=FrechetInceptionDistance,
+        prefix='FID-Full-50k',
+        fake_nums=50000,
+        inception_style='StyleGAN',
+        sample_model='ema'),
+    dict(type=PrecisionAndRecall, fake_nums=50000, prefix='PR-50K'),
+    dict(type=PerceptualPathLength, fake_nums=50000, prefix='ppl-w')
+]
+# NOTE: config for save multi best checkpoints
+# default_hooks.update(
+#     checkpoint=dict(
+#         save_best=['FID-Full-50k/fid', 'IS-50k/is'],
+#         rule=['less', 'greater']))
+default_hooks.update(checkpoint=dict(save_best='FID-Full-50k/fid'))
+
+val_evaluator.update(metrics=metrics)
+test_evaluator.update(metrics=metrics)