mmfewshot/tests/test_classification_models/test_classifiers/test_maml.py

# Copyright (c) OpenMMLab. All rights reserved.
import copy

import pytest
import torch

from mmfewshot.classification.models import CLASSIFIERS


@pytest.mark.parametrize('classifier', ['MAML'])
def test_image_classifier(classifier):
    model_cfg = dict(
        type=classifier,
        backbone=(dict(type='Conv4')),
        head=dict(type='LinearHead', num_classes=5, in_channels=1600))

    imgs_a = torch.randn(4, 3, 84, 84)
    imgs_b = torch.randn(4, 3, 84, 84)
    label = torch.LongTensor([0, 1, 2, 3])

    model_cfg_ = copy.deepcopy(model_cfg)
    model = CLASSIFIERS.build(model_cfg_)

    # test property
    assert not model.with_neck
    assert model.with_head

    assert model.device
    assert model.get_device()

    # test train_step
    outputs = model.train_step(
        {
            'support_data': {
                'img': imgs_a,
                'gt_label': label,
                'mode': 'train',
                'img_metas': [_ for _ in range(4)]
            },
            'query_data': {
                'img': imgs_b,
                'gt_label': label,
                'mode': 'train',
                'img_metas': [_ for _ in range(4)]
            }
        }, None)
    assert outputs['loss'].item() > 0
    assert outputs['num_samples'] == 4
    for weight in model.parameters():
        assert weight.fast is None

    # test first order
    model_cfg = dict(
        type=classifier,
        first_order=True,
        backbone=(dict(type='Conv4')),
        head=dict(type='LinearHead', num_classes=5, in_channels=1600))

    imgs_a = torch.randn(4, 3, 84, 84)
    imgs_b = torch.randn(4, 3, 84, 84)
    label = torch.LongTensor([0, 1, 2, 3])

    model_cfg_ = copy.deepcopy(model_cfg)
    model = CLASSIFIERS.build(model_cfg_)

    # test property
    assert not model.with_neck
    assert model.with_head

    assert model.device
    assert model.get_device()

    # test train_step
    outputs = model.train_step(
        {
            'support_data': {
                'img': imgs_a,
                'gt_label': label,
                'mode': 'train',
                'img_metas': [_ for _ in range(4)]
            },
            'query_data': {
                'img': imgs_b,
                'gt_label': label,
                'mode': 'train',
                'img_metas': [_ for _ in range(4)]
            }
        }, None)
    assert outputs['loss'].item() > 0
    assert outputs['num_samples'] == 4
    for weight in model.parameters():
        assert weight.fast is None

    # test val_step
    outputs = model.val_step(
        {
            'support_data': {
                'img': imgs_a,
                'gt_label': label,
                'mode': 'train',
                'img_metas': [_ for _ in range(4)]
            },
            'query_data': {
                'img': imgs_b,
                'gt_label': label,
                'mode': 'train',
                'img_metas': [_ for _ in range(4)]
            }
        }, None)
    assert outputs['loss'].item() > 0
    assert outputs['num_samples'] == 4
    for weight in model.parameters():
        assert weight.fast is None

    model.before_meta_test(dict(support={'num_inner_steps': 2}))
    model.before_forward_support()

    # test support step
    model(**{'img': imgs_a, 'gt_label': label, 'mode': 'support'})
    for weight in model.parameters():
        assert weight.fast is not None

    model.before_forward_query()
    # test query step
    outputs = model(**{'img': imgs_b, 'gt_label': label, 'mode': 'query'})
    assert isinstance(outputs, list)
    assert len(outputs) == 4
    assert outputs[0].shape[0] == 5