EasyCV/tests/models/selfsup/test_swav.py

78 lines
2.1 KiB
Python

# Copyright (c) Alibaba, Inc. and its affiliates.
import unittest
import torch
from torch import distributed as dist
from easycv.models.builder import build_model
from easycv.utils.test_util import pseudo_dist_init
_num_crops = [2, 6]
_base_model_cfg = model = dict(
type='SWAV',
pretrained=None,
train_preprocess=['randomGrayScale', 'gaussianBlur'],
backbone=dict(
type='ResNet',
depth=50,
in_channels=3,
out_indices=[4], # 0: conv-1, x: stage-x
norm_cfg=dict(type='SyncBN')),
# swav need mulit crop ,doesn't support vit based model
neck=dict(
type='NonLinearNeckSwav',
in_channels=2048,
hid_channels=2048,
out_channels=128,
with_avg_pool=False),
config=dict(
# multi crop setting
num_crops=_num_crops,
size_crops=[160, 96],
min_scale_crops=[0.14, 0.05],
max_scale_crops=[1, 0.14],
# swav setting
crops_for_assign=[0, 1],
epsilon=0.05,
nmb_prototypes=3000,
sinkhorn_iterations=3,
temperature=0.1,
# queue setting
queue_length=3840,
epoch_queue_starts=15))
class SWAVTest(unittest.TestCase):
def setUp(self):
print(('Testing %s.%s' % (type(self).__name__, self._testMethodName)))
def test_swav_train(self):
model = build_model(_base_model_cfg).cuda()
pseudo_dist_init()
model.train()
batch_size = 4
imgs = [torch.randn(batch_size, 3, 224, 224).cuda()] * 8
output = model(imgs, mode='train')
self.assertIn('loss', output)
self.assertEqual(output['loss'].shape, torch.Size([]))
dist.destroy_process_group()
def test_swav_extract(self):
model = build_model(_base_model_cfg).cuda()
pseudo_dist_init()
batch_size = 4
imgs = torch.randn(batch_size, 3, 224, 224).cuda()
output = model(imgs, mode='extract')
self.assertEqual(output['neck'].shape, torch.Size([4, 128]))
dist.destroy_process_group()
if __name__ == '__main__':
unittest.main()