mmpretrain/tests/test_models/test_backbones/test_mobileone.py

338 lines
12 KiB
Python

# Copyright (c) OpenMMLab. All rights reserved.
import os
import tempfile
import pytest
import torch
from mmengine.runner import load_checkpoint, save_checkpoint
from torch import nn
from torch.nn.modules import GroupNorm
from torch.nn.modules.batchnorm import _BatchNorm
from mmcls.models.backbones import MobileOne
from mmcls.models.backbones.mobileone import MobileOneBlock
from mmcls.models.utils import SELayer
def check_norm_state(modules, train_state):
"""Check if norm layer is in correct train state."""
for mod in modules:
if isinstance(mod, _BatchNorm):
if mod.training != train_state:
return False
return True
def is_norm(modules):
"""Check if is one of the norms."""
if isinstance(modules, (GroupNorm, _BatchNorm)):
return True
return False
def is_mobileone_block(modules):
if isinstance(modules, MobileOneBlock):
return True
return False
def test_mobileoneblock():
# Test MobileOneBlock with kernel_size 3
block = MobileOneBlock(5, 10, 3, 1, stride=1, groups=5)
block.eval()
x = torch.randn(1, 5, 16, 16)
y = block(x)
assert block.branch_norm is None
assert not hasattr(block, 'branch_reparam')
assert hasattr(block, 'branch_scale')
assert hasattr(block, 'branch_conv_list')
assert hasattr(block, 'branch_norm')
assert block.branch_conv_list[0].conv.kernel_size == (3, 3)
assert block.branch_conv_list[0].conv.groups == 5
assert block.se_cfg is None
assert y.shape == torch.Size((1, 10, 16, 16))
block.switch_to_deploy()
assert hasattr(block, 'branch_reparam')
assert block.branch_reparam.kernel_size == (3, 3)
assert block.branch_reparam.groups == 5
assert block.deploy is True
y_deploy = block(x)
assert y_deploy.shape == torch.Size((1, 10, 16, 16))
assert torch.allclose(y, y_deploy, atol=1e-5, rtol=1e-4)
# Test MobileOneBlock with num_con = 4
block = MobileOneBlock(5, 10, 3, 4, stride=1, groups=5)
block.eval()
x = torch.randn(1, 5, 16, 16)
y = block(x)
assert block.branch_norm is None
assert not hasattr(block, 'branch_reparam')
assert hasattr(block, 'branch_scale')
assert hasattr(block, 'branch_conv_list')
assert hasattr(block, 'branch_norm')
assert block.branch_conv_list[0].conv.kernel_size == (3, 3)
assert block.branch_conv_list[0].conv.groups == 5
assert len(block.branch_conv_list) == 4
assert block.se_cfg is None
assert y.shape == torch.Size((1, 10, 16, 16))
block.switch_to_deploy()
assert hasattr(block, 'branch_reparam')
assert block.branch_reparam.kernel_size == (3, 3)
assert block.branch_reparam.groups == 5
assert block.deploy is True
y_deploy = block(x)
assert y_deploy.shape == torch.Size((1, 10, 16, 16))
assert torch.allclose(y, y_deploy, atol=1e-5, rtol=1e-4)
# Test MobileOneBlock with kernel_size 1
block = MobileOneBlock(5, 10, 1, 1, stride=1, padding=0)
block.eval()
x = torch.randn(1, 5, 16, 16)
y = block(x)
assert block.branch_norm is None
assert not hasattr(block, 'branch_reparam')
assert hasattr(block, 'branch_scale')
assert hasattr(block, 'branch_conv_list')
assert hasattr(block, 'branch_norm')
assert block.branch_conv_list[0].conv.kernel_size == (1, 1)
assert block.branch_conv_list[0].conv.groups == 1
assert len(block.branch_conv_list) == 1
assert block.se_cfg is None
assert y.shape == torch.Size((1, 10, 16, 16))
block.switch_to_deploy()
assert hasattr(block, 'branch_reparam')
assert block.branch_reparam.kernel_size == (1, 1)
assert block.branch_reparam.groups == 1
assert block.deploy is True
y_deploy = block(x)
assert y_deploy.shape == torch.Size((1, 10, 16, 16))
assert torch.allclose(y, y_deploy, atol=1e-5, rtol=1e-4)
# Test MobileOneBlock with stride = 2
block = MobileOneBlock(10, 10, 3, 4, stride=2, groups=10)
x = torch.randn(1, 10, 16, 16)
block.eval()
y = block(x)
assert block.branch_norm is None
assert not hasattr(block, 'branch_reparam')
assert hasattr(block, 'branch_scale')
assert hasattr(block, 'branch_conv_list')
assert hasattr(block, 'branch_norm')
assert block.branch_conv_list[0].conv.kernel_size == (3, 3)
assert block.branch_conv_list[0].conv.groups == 10
assert len(block.branch_conv_list) == 4
assert block.se_cfg is None
assert y.shape == torch.Size((1, 10, 8, 8))
block.switch_to_deploy()
assert hasattr(block, 'branch_reparam')
assert block.branch_reparam.kernel_size == (3, 3)
assert block.branch_reparam.groups == 10
assert block.deploy is True
y_deploy = block(x)
assert y_deploy.shape == torch.Size((1, 10, 8, 8))
assert torch.allclose(y, y_deploy, atol=1e-5, rtol=1e-4)
# # Test MobileOneBlock with padding == dilation == 2
block = MobileOneBlock(
10, 10, 3, 4, stride=1, groups=10, padding=2, dilation=2)
x = torch.randn(1, 10, 16, 16)
block.eval()
y = block(x)
assert not hasattr(block, 'branch_reparam')
assert hasattr(block, 'branch_scale')
assert hasattr(block, 'branch_conv_list')
assert hasattr(block, 'branch_norm')
assert block.branch_conv_list[0].conv.kernel_size == (3, 3)
assert block.branch_conv_list[0].conv.groups == 10
assert len(block.branch_conv_list) == 4
assert block.se_cfg is None
assert y.shape == torch.Size((1, 10, 16, 16))
block.switch_to_deploy()
assert hasattr(block, 'branch_reparam')
assert block.branch_reparam.kernel_size == (3, 3)
assert block.branch_reparam.groups == 10
assert block.deploy is True
y_deploy = block(x)
assert y_deploy.shape == torch.Size((1, 10, 16, 16))
assert torch.allclose(y, y_deploy, atol=1e-5, rtol=1e-4)
# Test MobileOneBlock with se
se_cfg = dict(ratio=4, divisor=1)
block = MobileOneBlock(32, 32, 3, 4, stride=1, se_cfg=se_cfg, groups=32)
x = torch.randn(1, 32, 16, 16)
block.eval()
y = block(x)
assert not hasattr(block, 'branch_reparam')
assert hasattr(block, 'branch_scale')
assert hasattr(block, 'branch_conv_list')
assert hasattr(block, 'branch_norm')
assert block.branch_conv_list[0].conv.kernel_size == (3, 3)
assert block.branch_conv_list[0].conv.groups == 32
assert len(block.branch_conv_list) == 4
assert isinstance(block.se, SELayer)
assert y.shape == torch.Size((1, 32, 16, 16))
block.switch_to_deploy()
assert hasattr(block, 'branch_reparam')
assert block.branch_reparam.kernel_size == (3, 3)
assert block.branch_reparam.groups == 32
assert block.deploy is True
y_deploy = block(x)
assert y_deploy.shape == torch.Size((1, 32, 16, 16))
assert torch.allclose(y, y_deploy, atol=1e-5, rtol=1e-4)
# Test MobileOneBlock with deploy == True
se_cfg = dict(ratio=4, divisor=1)
block = MobileOneBlock(
32, 32, 3, 4, stride=1, se_cfg=se_cfg, groups=32, deploy=True)
x = torch.randn(1, 32, 16, 16)
block.eval()
assert hasattr(block, 'branch_reparam')
assert block.branch_reparam.kernel_size == (3, 3)
assert block.branch_reparam.groups == 32
assert isinstance(block.se, SELayer)
assert block.deploy is True
y = block(x)
assert y.shape == torch.Size((1, 32, 16, 16))
def test_mobileone_backbone():
with pytest.raises(TypeError):
# arch must be str or dict
MobileOne(arch=[4, 6, 16, 1])
with pytest.raises(AssertionError):
# arch must in arch_settings
MobileOne(arch='S3')
with pytest.raises(KeyError):
arch = dict(num_blocks=[2, 4, 14, 1])
MobileOne(arch=arch)
# Test len(arch['num_blocks']) == len(arch['width_factor'])
with pytest.raises(AssertionError):
arch = dict(
num_blocks=[2, 4, 14, 1],
width_factor=[0.75, 0.75, 0.75],
num_conv_branches=[1, 1, 1, 1],
num_se_blocks=[0, 0, 5, 1])
MobileOne(arch=arch)
# Test max(out_indices) < len(arch['num_blocks'])
with pytest.raises(AssertionError):
MobileOne('s0', out_indices=dict())
# Test out_indices not type of int or Sequence
with pytest.raises(AssertionError):
MobileOne('s0', out_indices=(5, ))
# Test MobileOne norm state
model = MobileOne('s0')
model.train()
assert check_norm_state(model.modules(), True)
# Test MobileOne with first stage frozen
frozen_stages = 1
model = MobileOne('s0', frozen_stages=frozen_stages)
model.train()
for param in model.stage0.parameters():
assert param.requires_grad is False
for i in range(0, frozen_stages):
stage_name = model.stages[i]
stage = model.__getattr__(stage_name)
for mod in stage:
if isinstance(mod, _BatchNorm):
assert mod.training is False
for param in stage.parameters():
assert param.requires_grad is False
# Test MobileOne with norm_eval
model = MobileOne('s0', norm_eval=True)
model.train()
assert check_norm_state(model.modules(), False)
# Test MobileOne forward with layer 3 forward
model = MobileOne('s0', out_indices=(3, ))
model.init_weights()
model.train()
for m in model.modules():
if is_norm(m):
assert isinstance(m, _BatchNorm)
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert isinstance(feat, tuple)
assert len(feat) == 1
assert isinstance(feat[0], torch.Tensor)
assert feat[0].shape == torch.Size((1, 1024, 7, 7))
# Test MobileOne forward
arch_settings = {
's0': dict(out_channels=[48, 128, 256, 1024], ),
's1': dict(out_channels=[96, 192, 512, 1280]),
's2': dict(out_channels=[96, 256, 640, 2048]),
's3': dict(out_channels=[128, 320, 768, 2048], ),
's4': dict(out_channels=[192, 448, 896, 2048], )
}
choose_models = ['s0', 's1', 's4']
# Test RepVGG model forward
for model_name, model_arch in arch_settings.items():
if model_name not in choose_models:
continue
model = MobileOne(model_name, out_indices=(0, 1, 2, 3))
model.init_weights()
# Test Norm
for m in model.modules():
if is_norm(m):
assert isinstance(m, _BatchNorm)
model.train()
imgs = torch.randn(1, 3, 224, 224)
feat = model(imgs)
assert feat[0].shape == torch.Size(
(1, model_arch['out_channels'][0], 56, 56))
assert feat[1].shape == torch.Size(
(1, model_arch['out_channels'][1], 28, 28))
assert feat[2].shape == torch.Size(
(1, model_arch['out_channels'][2], 14, 14))
assert feat[3].shape == torch.Size(
(1, model_arch['out_channels'][3], 7, 7))
# Test eval of "train" mode and "deploy" mode
gap = nn.AdaptiveAvgPool2d(output_size=(1))
fc = nn.Linear(model_arch['out_channels'][3], 10)
model.eval()
feat = model(imgs)
pred = fc(gap(feat[3]).flatten(1))
model.switch_to_deploy()
for m in model.modules():
if isinstance(m, MobileOneBlock):
assert m.deploy is True
feat_deploy = model(imgs)
pred_deploy = fc(gap(feat_deploy[3]).flatten(1))
for i in range(4):
torch.allclose(feat[i], feat_deploy[i])
torch.allclose(pred, pred_deploy)
def test_load_deploy_mobileone():
# Test output before and load from deploy checkpoint
model = MobileOne('s0', out_indices=(0, 1, 2, 3))
inputs = torch.randn((1, 3, 224, 224))
tmpdir = tempfile.gettempdir()
ckpt_path = os.path.join(tmpdir, 'ckpt.pth')
model.switch_to_deploy()
model.eval()
outputs = model(inputs)
model_deploy = MobileOne('s0', out_indices=(0, 1, 2, 3), deploy=True)
save_checkpoint(model.state_dict(), ckpt_path)
load_checkpoint(model_deploy, ckpt_path)
outputs_load = model_deploy(inputs)
for feat, feat_load in zip(outputs, outputs_load):
assert torch.allclose(feat, feat_load)
os.remove(ckpt_path)