Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_multi_output_model.py 5.32 KB
Newer Older
aiss's avatar
push dsv0.8.2 version  
aiss 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 
'''Copyright The Microsoft DeepSpeed Team'''

import torch
import deepspeed
from pytest import approx
from unit.common import DistributedTest
from unit.multi_output_model import MultiOutputModel, multi_output_dataloader


class TestTwoOutputModel(DistributedTest):
    world_size = 1

    def test(self, tmpdir):
        grad_accumulation_steps = 2
        micro_batch_size = 1
        world_size = self.world_size
        config_dict = {
            "train_micro_batch_size_per_gpu": micro_batch_size,
            "gradient_accumulation_steps": grad_accumulation_steps,
            "train_batch_size": micro_batch_size * grad_accumulation_steps * world_size,
            "steps_per_print": 1,
            "optimizer": {
                "type": "Adam",
                "params": {
                    "lr": 0.00015
                }
            },
            "fp16": {
                "enabled": True
            }
        }

        hidden_dim = 10
        weight_value = 0.1

        model = MultiOutputModel(hidden_dim, weight_value)
        model, _, _, _ = deepspeed.initialize(config=config_dict,
                                              model=model,
                                              model_parameters=model.parameters())
        total_samples = 4
        data_loader = multi_output_dataloader(model=model,
                                              total_samples=total_samples,
                                              hidden_dim=hidden_dim,
                                              device=model.device,
                                              inputs=[1.0,
                                                      2.0],
                                              targets=[1,
                                                       2])
        for n, batch in enumerate(data_loader):
            assert len(batch) % 2 == 0, \
                 f"multi_output_dataloader failed to return even number of data samples (input+target)"

            midpoint = len(batch) // 2
            inputs, targets = batch[:midpoint], batch[midpoint:]
            loss_tuple = model(inputs, targets)

            expected_loss = torch.tensor(2.302734375,
                                         dtype=torch.half,
                                         device=model.device)
            for loss in loss_tuple:
                assert loss.shape == torch.Size([])
                assert loss.item() == approx(expected_loss.item())

            summed_loss = sum(loss_tuple)
            scaled_loss = model.backward(summed_loss)
            expected_scaled_loss = summed_loss.float() / grad_accumulation_steps
            assert scaled_loss.item() == approx(expected_scaled_loss.item())

            model.step()


class TestThreeOutputModel(DistributedTest):
    world_size = 1

    def test(self, tmpdir):
        grad_accumulation_steps = 3
        micro_batch_size = 1
        world_size = 1
        config_dict = {
            "train_micro_batch_size_per_gpu": micro_batch_size,
            "gradient_accumulation_steps": grad_accumulation_steps,
            "train_batch_size": micro_batch_size * grad_accumulation_steps * world_size,
            "steps_per_print": 1,
            "optimizer": {
                "type": "Adam",
                "params": {
                    "lr": 0.00015
                }
            },
            "fp16": {
                "enabled": True
            }
        }

        hidden_dim = 10
        weight_value = 0.1

        model = MultiOutputModel(hidden_dim, weight_value)
        model, _, _, _ = deepspeed.initialize(config=config_dict,
                                              model=model,
                                              model_parameters=model.parameters())

        total_samples = grad_accumulation_steps * micro_batch_size * 2
        data_loader = multi_output_dataloader(model=model,
                                              total_samples=total_samples,
                                              hidden_dim=hidden_dim,
                                              device=model.device,
                                              inputs=[1.0,
                                                      2.0,
                                                      3.0],
                                              targets=[1,
                                                       2,
                                                       3])
        for n, batch in enumerate(data_loader):
            assert len(batch) % 2 == 0, \
                 f"multi_output_dataloader failed to return even number of data samples (input+target)"

            midpoint = len(batch) // 2
            inputs, targets = batch[:midpoint], batch[midpoint:]
            loss_tuple = model(inputs, targets)
            assert len(loss_tuple) == 3

            expected_loss = torch.tensor(2.302734375,
                                         dtype=torch.half,
                                         device=model.device)

            for loss in loss_tuple:
                assert loss.shape == torch.Size([])
                assert loss.item() == approx(expected_loss.item())

            summed_loss = sum(loss_tuple)
            scaled_loss = model.backward(summed_loss)
            expected_scaled_loss = summed_loss.float() / grad_accumulation_steps
            assert scaled_loss.item() == approx(expected_scaled_loss.item())

            model.step()