mmclassification/tests/test_models/test_backbones/test_mvit.py

186 lines
6.5 KiB
Python

# Copyright (c) OpenMMLab. All rights reserved.
from copy import deepcopy
from unittest import TestCase
import torch
from mmcls.models.backbones import MViT
class TestMViT(TestCase):
def setUp(self):
self.cfg = dict(arch='tiny', img_size=224, drop_path_rate=0.1)
def test_arch(self):
# Test invalid default arch
with self.assertRaisesRegex(AssertionError, 'not in default archs'):
cfg = deepcopy(self.cfg)
cfg['arch'] = 'unknown'
MViT(**cfg)
# Test invalid custom arch
with self.assertRaisesRegex(AssertionError, 'Custom arch needs'):
cfg = deepcopy(self.cfg)
cfg['arch'] = {
'embed_dims': 96,
'num_layers': 10,
}
MViT(**cfg)
# Test custom arch
cfg = deepcopy(self.cfg)
embed_dims = 96
num_layers = 10
num_heads = 1
downscale_indices = (2, 5, 7)
cfg['arch'] = {
'embed_dims': embed_dims,
'num_layers': num_layers,
'num_heads': num_heads,
'downscale_indices': downscale_indices
}
model = MViT(**cfg)
self.assertEqual(len(model.blocks), num_layers)
for i, block in enumerate(model.blocks):
if i in downscale_indices:
num_heads *= 2
embed_dims *= 2
self.assertEqual(block.out_dims, embed_dims)
self.assertEqual(block.attn.num_heads, num_heads)
def test_init_weights(self):
# test weight init cfg
cfg = deepcopy(self.cfg)
cfg['use_abs_pos_embed'] = True
cfg['init_cfg'] = [
dict(
type='Kaiming',
layer='Conv2d',
mode='fan_in',
nonlinearity='linear')
]
model = MViT(**cfg)
ori_weight = model.patch_embed.projection.weight.clone().detach()
# The pos_embed is all zero before initialize
self.assertTrue(torch.allclose(model.pos_embed, torch.tensor(0.)))
model.init_weights()
initialized_weight = model.patch_embed.projection.weight
self.assertFalse(torch.allclose(ori_weight, initialized_weight))
self.assertFalse(torch.allclose(model.pos_embed, torch.tensor(0.)))
self.assertFalse(
torch.allclose(model.blocks[0].attn.rel_pos_h, torch.tensor(0.)))
self.assertFalse(
torch.allclose(model.blocks[0].attn.rel_pos_w, torch.tensor(0.)))
# test rel_pos_zero_init
cfg = deepcopy(self.cfg)
cfg['rel_pos_zero_init'] = True
model = MViT(**cfg)
model.init_weights()
self.assertTrue(
torch.allclose(model.blocks[0].attn.rel_pos_h, torch.tensor(0.)))
self.assertTrue(
torch.allclose(model.blocks[0].attn.rel_pos_w, torch.tensor(0.)))
def test_forward(self):
imgs = torch.randn(1, 3, 224, 224)
cfg = deepcopy(self.cfg)
model = MViT(**cfg)
outs = model(imgs)
self.assertIsInstance(outs, tuple)
self.assertEqual(len(outs), 1)
feat = outs[-1]
self.assertEqual(feat.shape, (1, 768, 7, 7))
# test multiple output indices
cfg = deepcopy(self.cfg)
cfg['out_scales'] = (0, 1, 2, 3)
model = MViT(**cfg)
model.init_weights()
outs = model(imgs)
self.assertIsInstance(outs, tuple)
self.assertEqual(len(outs), 4)
for stride, out in zip([1, 2, 4, 8], outs):
self.assertEqual(out.shape,
(1, 96 * stride, 56 // stride, 56 // stride))
# test dim_mul_in_attention = False
cfg = deepcopy(self.cfg)
cfg['out_scales'] = (0, 1, 2, 3)
cfg['dim_mul_in_attention'] = False
model = MViT(**cfg)
outs = model(imgs)
self.assertIsInstance(outs, tuple)
self.assertEqual(len(outs), 4)
for dim_mul, stride, out in zip([2, 4, 8, 8], [1, 2, 4, 8], outs):
self.assertEqual(out.shape,
(1, 96 * dim_mul, 56 // stride, 56 // stride))
# test rel_pos_spatial = False
cfg = deepcopy(self.cfg)
cfg['out_scales'] = (0, 1, 2, 3)
cfg['rel_pos_spatial'] = False
cfg['img_size'] = None
model = MViT(**cfg)
outs = model(imgs)
self.assertIsInstance(outs, tuple)
self.assertEqual(len(outs), 4)
for stride, out in zip([1, 2, 4, 8], outs):
self.assertEqual(out.shape,
(1, 96 * stride, 56 // stride, 56 // stride))
# test residual_pooling = False
cfg = deepcopy(self.cfg)
cfg['out_scales'] = (0, 1, 2, 3)
cfg['residual_pooling'] = False
model = MViT(**cfg)
outs = model(imgs)
self.assertIsInstance(outs, tuple)
self.assertEqual(len(outs), 4)
for stride, out in zip([1, 2, 4, 8], outs):
self.assertEqual(out.shape,
(1, 96 * stride, 56 // stride, 56 // stride))
# test use_abs_pos_embed = True
cfg = deepcopy(self.cfg)
cfg['out_scales'] = (0, 1, 2, 3)
cfg['use_abs_pos_embed'] = True
model = MViT(**cfg)
model.init_weights()
outs = model(imgs)
self.assertIsInstance(outs, tuple)
self.assertEqual(len(outs), 4)
for stride, out in zip([1, 2, 4, 8], outs):
self.assertEqual(out.shape,
(1, 96 * stride, 56 // stride, 56 // stride))
# test dynamic inputs shape
cfg = deepcopy(self.cfg)
cfg['out_scales'] = (0, 1, 2, 3)
model = MViT(**cfg)
imgs = torch.randn(1, 3, 352, 260)
h_resolution = (352 + 2 * 3 - 7) // 4 + 1
w_resolution = (260 + 2 * 3 - 7) // 4 + 1
outs = model(imgs)
self.assertIsInstance(outs, tuple)
self.assertEqual(len(outs), 4)
expect_h = h_resolution
expect_w = w_resolution
for i, out in enumerate(outs):
self.assertEqual(out.shape, (1, 96 * 2**i, expect_h, expect_w))
expect_h = (expect_h + 2 * 1 - 3) // 2 + 1
expect_w = (expect_w + 2 * 1 - 3) // 2 + 1
def test_structure(self):
# test drop_path_rate decay
cfg = deepcopy(self.cfg)
cfg['drop_path_rate'] = 0.2
model = MViT(**cfg)
for i, block in enumerate(model.blocks):
expect_prob = 0.2 / (model.num_layers - 1) * i
if expect_prob > 0:
self.assertAlmostEqual(block.drop_path.drop_prob, expect_prob)