Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_pruner.py 8.04 KB
Newer Older
J-shang's avatar
[Compression] evaluator - step2 (#4992)  
J-shang committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172 
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT license.

from __future__ import annotations

import pytest

import torch
import torch.nn.functional as F

import nni
from nni.compression.pytorch.pruning import (
    LevelPruner,
    L1NormPruner,
    L2NormPruner,
    SlimPruner,
    FPGMPruner,
    ActivationAPoZRankPruner,
    ActivationMeanRankPruner,
    TaylorFOWeightPruner,
    ADMMPruner,
    MovementPruner
)

from ..assets.device import device
from ..assets.simple_mnist import (
    create_lighting_evaluator,
    create_pytorch_evaluator,
    training_model
)
from ..assets.common import create_model, validate_masks, validate_dependency_aware


@pytest.mark.parametrize('model_type', ['lightning', 'pytorch'])
def test_level_pruner(model_type: str):
    model, config_list, dummy_input = create_model(model_type)

    pruner = LevelPruner(model=model, config_list=config_list)

    _, masks = pruner.compress()
    model(dummy_input)
    pruner._unwrap_model()
    validate_masks(masks, model, config_list)


@pytest.mark.parametrize('model_type', ['lightning', 'pytorch'])
@pytest.mark.parametrize('pruning_type', ['l1', 'l2', 'fpgm'])
@pytest.mark.parametrize('mode', ['normal', 'dependency_aware'])
def test_norm_pruner(model_type: str, pruning_type: str, mode: str):
    model, config_list, dummy_input = create_model(model_type)

    if pruning_type == 'l1':
        pruner = L1NormPruner(model=model, config_list=config_list, mode=mode, dummy_input=dummy_input)
    elif pruning_type == 'l2':
        pruner = L2NormPruner(model=model, config_list=config_list, mode=mode, dummy_input=dummy_input)
    elif pruning_type == 'fpgm':
        pruner = FPGMPruner(model=model, config_list=config_list, mode=mode, dummy_input=dummy_input)
    else:
        raise ValueError(f'wrong norm: {pruning_type}')

    _, masks = pruner.compress()
    model(dummy_input)
    pruner._unwrap_model()
    validate_masks(masks, model, config_list)
    if mode == 'dependency_aware':
        validate_dependency_aware(model_type, masks)


@pytest.mark.parametrize('model_type', ['lightning', 'pytorch'])
@pytest.mark.parametrize('using_evaluator', [True, False])
@pytest.mark.parametrize('mode', ['global', 'normal'])
def test_slim_pruner(model_type: str, using_evaluator: bool, mode: str):
    model, _, dummy_input = create_model(model_type)
    config_list = [{'op_types': ['BatchNorm2d'], 'total_sparsity': 0.5}]

    if using_evaluator:
        evaluator = create_lighting_evaluator() if model_type == 'lightning' else create_pytorch_evaluator(model)
        pruner = SlimPruner(model=model, config_list=config_list, evaluator=evaluator, training_epochs=1, scale=0.0001, mode=mode)
    else:
        model = model.to(device)
        dummy_input = dummy_input.to(device)
        optimizer = nni.trace(torch.optim.SGD)(model.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4)
        pruner = SlimPruner(model=model, config_list=config_list, trainer=training_model, traced_optimizer=optimizer,
                            criterion=F.nll_loss, training_epochs=1, scale=0.0001, mode=mode)

    _, masks = pruner.compress()
    model(dummy_input)
    pruner._unwrap_model()
    validate_masks(masks, model, config_list, is_global=(mode == 'global'))


@pytest.mark.parametrize('model_type', ['lightning', 'pytorch'])
@pytest.mark.parametrize('pruning_type', ['apoz', 'mean', 'taylor'])
@pytest.mark.parametrize('using_evaluator', [True, False])
@pytest.mark.parametrize('mode', ['normal', 'dependency_aware'])
def test_hook_based_pruner(model_type: str, pruning_type: str, using_evaluator: bool, mode: str):
    model, config_list, dummy_input = create_model(model_type)

    if using_evaluator:
        evaluator = create_lighting_evaluator() if model_type == 'lightning' else create_pytorch_evaluator(model)
        if pruning_type == 'apoz':
            pruner = ActivationAPoZRankPruner(model=model, config_list=config_list, evaluator=evaluator, training_steps=20,
                                              activation='relu', mode=mode, dummy_input=dummy_input)
        elif pruning_type == 'mean':
            pruner = ActivationMeanRankPruner(model=model, config_list=config_list, evaluator=evaluator, training_steps=20,
                                              activation='relu', mode=mode, dummy_input=dummy_input)
        elif pruning_type == 'taylor':
            pruner = TaylorFOWeightPruner(model=model, config_list=config_list, evaluator=evaluator, training_steps=20,
                                          mode=mode, dummy_input=dummy_input)
    else:
        model = model.to(device)
        dummy_input = dummy_input.to(device)
        optimizer = nni.trace(torch.optim.SGD)(model.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4)
        if pruning_type == 'apoz':
            pruner = ActivationAPoZRankPruner(model=model, config_list=config_list, trainer=training_model, traced_optimizer=optimizer,
                                              criterion=F.nll_loss, training_batches=20, activation='relu', mode=mode, dummy_input=dummy_input)
        elif pruning_type == 'mean':
            pruner = ActivationMeanRankPruner(model=model, config_list=config_list, trainer=training_model, traced_optimizer=optimizer,
                                              criterion=F.nll_loss, training_batches=20, activation='relu', mode=mode, dummy_input=dummy_input)
        elif pruning_type == 'taylor':
            pruner = TaylorFOWeightPruner(model=model, config_list=config_list, trainer=training_model, traced_optimizer=optimizer,
                                          criterion=F.nll_loss, training_batches=20, mode=mode, dummy_input=dummy_input)

    _, masks = pruner.compress()
    model(dummy_input)
    pruner._unwrap_model()
    validate_masks(masks, model, config_list)
    if mode == 'dependency_aware':
        validate_dependency_aware(model_type, masks)


@pytest.mark.parametrize('model_type', ['lightning', 'pytorch'])
@pytest.mark.parametrize('using_evaluator', [True, False])
@pytest.mark.parametrize('granularity', ['fine-grained', 'coarse-grained'])
def test_admm_pruner(model_type: str, using_evaluator: bool, granularity: str):
    model, config_list, dummy_input = create_model(model_type)

    if using_evaluator:
        evaluator = create_lighting_evaluator() if model_type == 'lightning' else create_pytorch_evaluator(model)
        pruner = ADMMPruner(model=model, config_list=config_list, evaluator=evaluator, iterations=2, training_epochs=1, granularity=granularity)
    else:
        model = model.to(device)
        dummy_input = dummy_input.to(device)
        optimizer = nni.trace(torch.optim.SGD)(model.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4)
        pruner = ADMMPruner(model=model, config_list=config_list, trainer=training_model, traced_optimizer=optimizer, criterion=F.nll_loss,
                            iterations=2, training_epochs=1, granularity=granularity)

    _, masks = pruner.compress()
    model(dummy_input)
    pruner._unwrap_model()
    validate_masks(masks, model, config_list)


@pytest.mark.parametrize('model_type', ['lightning', 'pytorch'])
@pytest.mark.parametrize('using_evaluator', [True, False])
def test_movement_pruner(model_type: str, using_evaluator: bool):
    model, config_list, dummy_input = create_model(model_type)

    if using_evaluator:
        evaluator = create_lighting_evaluator() if model_type == 'lightning' else create_pytorch_evaluator(model)
        pruner = MovementPruner(model=model, config_list=config_list, evaluator=evaluator, training_epochs=1, warm_up_step=10, cool_down_beginning_step=40)
    else:
        model = model.to(device)
        dummy_input = dummy_input.to(device)
        optimizer = nni.trace(torch.optim.SGD)(model.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4)
        pruner = MovementPruner(model=model, config_list=config_list, trainer=training_model, traced_optimizer=optimizer, criterion=F.nll_loss,
                                training_epochs=1, warm_up_step=10, cool_down_beginning_step=40)

    _, masks = pruner.compress()
    model(dummy_input)
    pruner._unwrap_model()
    validate_masks(masks, model, config_list)