test_modeling_transfo_xl.py 8.19 KB
Newer Older
thomwolf's avatar
thomwolf committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
Aymeric Augustin's avatar
Aymeric Augustin committed
15
from __future__ import absolute_import, division, print_function
thomwolf's avatar
thomwolf committed
16
17

import random
18
import unittest
thomwolf's avatar
thomwolf committed
19

20
from transformers import is_torch_available
thomwolf's avatar
thomwolf committed
21

22
from .test_configuration_common import ConfigTester
23
from .test_modeling_common import ModelTesterMixin, ids_tensor
Aymeric Augustin's avatar
Aymeric Augustin committed
24
25
26
from .utils import CACHE_DIR, require_torch, slow, torch_device


27
if is_torch_available():
thomwolf's avatar
thomwolf committed
28
    import torch
29
    from transformers import TransfoXLConfig, TransfoXLModel, TransfoXLLMHeadModel
30
    from transformers.modeling_transfo_xl import TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
thomwolf's avatar
thomwolf committed
31

32
33

@require_torch
34
class TransfoXLModelTest(ModelTesterMixin, unittest.TestCase):
thomwolf's avatar
thomwolf committed
35

thomwolf's avatar
thomwolf committed
36
    all_model_classes = (TransfoXLModel, TransfoXLLMHeadModel) if is_torch_available() else ()
thomwolf's avatar
thomwolf committed
37
38
39
    test_pruning = False
    test_torchscript = False
    test_resize_embeddings = False
thomwolf's avatar
thomwolf committed
40

thomwolf's avatar
thomwolf committed
41
    class TransfoXLModelTester(object):
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
        def __init__(
            self,
            parent,
            batch_size=13,
            seq_length=7,
            mem_len=30,
            clamp_len=15,
            is_training=True,
            use_labels=True,
            vocab_size=99,
            cutoffs=[10, 50, 80],
            hidden_size=32,
            d_embed=32,
            num_attention_heads=4,
            d_head=8,
            d_inner=128,
            div_val=2,
            num_hidden_layers=5,
            scope=None,
            seed=1,
        ):
thomwolf's avatar
thomwolf committed
63
64
65
66
            self.parent = parent
            self.batch_size = batch_size
            self.seq_length = seq_length
            self.mem_len = mem_len
thomwolf's avatar
thomwolf committed
67
            self.key_length = seq_length + mem_len
thomwolf's avatar
thomwolf committed
68
69
70
71
72
            self.clamp_len = clamp_len
            self.is_training = is_training
            self.use_labels = use_labels
            self.vocab_size = vocab_size
            self.cutoffs = cutoffs
thomwolf's avatar
thomwolf committed
73
            self.hidden_size = hidden_size
thomwolf's avatar
thomwolf committed
74
            self.d_embed = d_embed
thomwolf's avatar
thomwolf committed
75
            self.num_attention_heads = num_attention_heads
thomwolf's avatar
thomwolf committed
76
77
78
            self.d_head = d_head
            self.d_inner = d_inner
            self.div_val = div_val
thomwolf's avatar
thomwolf committed
79
            self.num_hidden_layers = num_hidden_layers
thomwolf's avatar
thomwolf committed
80
81
82
83
            self.scope = scope
            self.seed = seed

        def prepare_config_and_inputs(self):
thomwolf's avatar
thomwolf committed
84
85
            input_ids_1 = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
            input_ids_2 = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
thomwolf's avatar
thomwolf committed
86
87
88

            lm_labels = None
            if self.use_labels:
thomwolf's avatar
thomwolf committed
89
                lm_labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
thomwolf's avatar
thomwolf committed
90
91

            config = TransfoXLConfig(
thomwolf's avatar
thomwolf committed
92
                vocab_size=self.vocab_size,
thomwolf's avatar
thomwolf committed
93
94
95
                mem_len=self.mem_len,
                clamp_len=self.clamp_len,
                cutoffs=self.cutoffs,
thomwolf's avatar
thomwolf committed
96
                d_model=self.hidden_size,
thomwolf's avatar
thomwolf committed
97
                d_embed=self.d_embed,
thomwolf's avatar
thomwolf committed
98
                n_head=self.num_attention_heads,
thomwolf's avatar
thomwolf committed
99
100
101
                d_head=self.d_head,
                d_inner=self.d_inner,
                div_val=self.div_val,
102
103
                n_layer=self.num_hidden_layers,
            )
thomwolf's avatar
thomwolf committed
104
105
106
107
108
109
110
111
112

            return (config, input_ids_1, input_ids_2, lm_labels)

        def set_seed(self):
            random.seed(self.seed)
            torch.manual_seed(self.seed)

        def create_transfo_xl_model(self, config, input_ids_1, input_ids_2, lm_labels):
            model = TransfoXLModel(config)
113
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
114
115
116
117
118
119
120
121
122
123
124
125
126
127
            model.eval()

            hidden_states_1, mems_1 = model(input_ids_1)
            hidden_states_2, mems_2 = model(input_ids_2, mems_1)
            outputs = {
                "hidden_states_1": hidden_states_1,
                "mems_1": mems_1,
                "hidden_states_2": hidden_states_2,
                "mems_2": mems_2,
            }
            return outputs

        def check_transfo_xl_model_output(self, result):
            self.parent.assertListEqual(
128
129
                list(result["hidden_states_1"].size()), [self.batch_size, self.seq_length, self.hidden_size]
            )
130
            self.parent.assertListEqual(
131
132
                list(result["hidden_states_2"].size()), [self.batch_size, self.seq_length, self.hidden_size]
            )
thomwolf's avatar
thomwolf committed
133
134
            self.parent.assertListEqual(
                list(list(mem.size()) for mem in result["mems_1"]),
135
136
                [[self.mem_len, self.batch_size, self.hidden_size]] * self.num_hidden_layers,
            )
thomwolf's avatar
thomwolf committed
137
138
            self.parent.assertListEqual(
                list(list(mem.size()) for mem in result["mems_2"]),
139
140
                [[self.mem_len, self.batch_size, self.hidden_size]] * self.num_hidden_layers,
            )
thomwolf's avatar
thomwolf committed
141
142
143

        def create_transfo_xl_lm_head(self, config, input_ids_1, input_ids_2, lm_labels):
            model = TransfoXLLMHeadModel(config)
144
            model.to(torch_device)
thomwolf's avatar
thomwolf committed
145
146
            model.eval()

thomwolf's avatar
thomwolf committed
147
148
149
150
            lm_logits_1, mems_1 = model(input_ids_1)
            loss_1, _, mems_1 = model(input_ids_1, labels=lm_labels)
            lm_logits_2, mems_2 = model(input_ids_2, mems=mems_1)
            loss_2, _, mems_2 = model(input_ids_2, labels=lm_labels, mems=mems_1)
thomwolf's avatar
thomwolf committed
151
152
153

            outputs = {
                "loss_1": loss_1,
thomwolf's avatar
thomwolf committed
154
                "mems_1": mems_1,
thomwolf's avatar
thomwolf committed
155
156
                "lm_logits_1": lm_logits_1,
                "loss_2": loss_2,
thomwolf's avatar
thomwolf committed
157
                "mems_2": mems_2,
thomwolf's avatar
thomwolf committed
158
159
160
161
162
                "lm_logits_2": lm_logits_2,
            }
            return outputs

        def check_transfo_xl_lm_head_output(self, result):
163
            self.parent.assertListEqual(list(result["loss_1"].size()), [self.batch_size, self.seq_length])
thomwolf's avatar
thomwolf committed
164
            self.parent.assertListEqual(
165
166
                list(result["lm_logits_1"].size()), [self.batch_size, self.seq_length, self.vocab_size]
            )
thomwolf's avatar
thomwolf committed
167
            self.parent.assertListEqual(
thomwolf's avatar
thomwolf committed
168
                list(list(mem.size()) for mem in result["mems_1"]),
169
170
                [[self.mem_len, self.batch_size, self.hidden_size]] * self.num_hidden_layers,
            )
thomwolf's avatar
thomwolf committed
171

172
            self.parent.assertListEqual(list(result["loss_2"].size()), [self.batch_size, self.seq_length])
thomwolf's avatar
thomwolf committed
173
            self.parent.assertListEqual(
174
175
                list(result["lm_logits_2"].size()), [self.batch_size, self.seq_length, self.vocab_size]
            )
thomwolf's avatar
thomwolf committed
176
            self.parent.assertListEqual(
thomwolf's avatar
thomwolf committed
177
                list(list(mem.size()) for mem in result["mems_2"]),
178
179
                [[self.mem_len, self.batch_size, self.hidden_size]] * self.num_hidden_layers,
            )
thomwolf's avatar
thomwolf committed
180

thomwolf's avatar
thomwolf committed
181
182
183
        def prepare_config_and_inputs_for_common(self):
            config_and_inputs = self.prepare_config_and_inputs()
            (config, input_ids_1, input_ids_2, lm_labels) = config_and_inputs
184
            inputs_dict = {"input_ids": input_ids_1}
thomwolf's avatar
thomwolf committed
185
186
187
188
189
            return config, inputs_dict

    def setUp(self):
        self.model_tester = TransfoXLModelTest.TransfoXLModelTester(self)
        self.config_tester = ConfigTester(self, config_class=TransfoXLConfig, d_embed=37)
thomwolf's avatar
thomwolf committed
190

thomwolf's avatar
thomwolf committed
191
    def test_config(self):
thomwolf's avatar
thomwolf committed
192
193
194
195
196
197
198
199
200
201
202
203
204
        self.config_tester.run_common_tests()

    def test_transfo_xl_model(self):
        self.model_tester.set_seed()
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        output_result = self.model_tester.create_transfo_xl_model(*config_and_inputs)
        self.model_tester.check_transfo_xl_model_output(output_result)

    def test_transfo_xl_lm_head(self):
        self.model_tester.set_seed()
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        output_result = self.model_tester.create_transfo_xl_lm_head(*config_and_inputs)
        self.model_tester.check_transfo_xl_lm_head_output(output_result)
205

206
    @slow
207
    def test_model_from_pretrained(self):
208
        for model_name in list(TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
209
            model = TransfoXLModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
210
            self.assertIsNotNone(model)