test_modeling_tf_blenderbot.py 9.82 KB
Newer Older
1
# coding=utf-8
2
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
3
4
5
6
7
8
9
10
11
12
13
14
#
# 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.
Sylvain Gugger's avatar
Sylvain Gugger committed
15

16

17
18
import unittest

19
from transformers import BlenderbotConfig, BlenderbotTokenizer, is_tf_available
20
from transformers.testing_utils import require_tf, require_tokenizers, slow, tooslow
21
from transformers.utils import cached_property
22

Yih-Dar's avatar
Yih-Dar committed
23
24
from ...test_configuration_common import ConfigTester
from ...test_modeling_tf_common import TFModelTesterMixin, ids_tensor
25
from ...test_pipeline_mixin import PipelineTesterMixin
26
27


28
29
30
if is_tf_available():
    import tensorflow as tf

31
    from transformers import TFAutoModelForSeq2SeqLM, TFBlenderbotForConditionalGeneration, TFBlenderbotModel
32
33


34
35
@require_tf
class TFBlenderbotModelTester:
36
    config_cls = BlenderbotConfig
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
    config_updates = {}
    hidden_act = "gelu"

    def __init__(
        self,
        parent,
        batch_size=13,
        seq_length=7,
        is_training=True,
        use_labels=False,
        vocab_size=99,
        hidden_size=32,
        num_hidden_layers=5,
        num_attention_heads=4,
        intermediate_size=37,
        hidden_dropout_prob=0.1,
        attention_probs_dropout_prob=0.1,
        max_position_embeddings=20,
        eos_token_id=2,
        pad_token_id=1,
        bos_token_id=0,
    ):
        self.parent = parent
        self.batch_size = batch_size
        self.seq_length = seq_length
        self.is_training = is_training
        self.use_labels = use_labels
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.intermediate_size = intermediate_size

        self.hidden_dropout_prob = hidden_dropout_prob
        self.attention_probs_dropout_prob = attention_probs_dropout_prob
        self.max_position_embeddings = max_position_embeddings
        self.eos_token_id = eos_token_id
        self.pad_token_id = pad_token_id
        self.bos_token_id = bos_token_id

    def prepare_config_and_inputs_for_common(self):
        input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size)
        eos_tensor = tf.expand_dims(tf.constant([self.eos_token_id] * self.batch_size), 1)
        input_ids = tf.concat([input_ids, eos_tensor], axis=1)

        decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)

        config = self.config_cls(
            vocab_size=self.vocab_size,
            d_model=self.hidden_size,
            encoder_layers=self.num_hidden_layers,
            decoder_layers=self.num_hidden_layers,
            encoder_attention_heads=self.num_attention_heads,
            decoder_attention_heads=self.num_attention_heads,
            encoder_ffn_dim=self.intermediate_size,
            decoder_ffn_dim=self.intermediate_size,
            dropout=self.hidden_dropout_prob,
            attention_dropout=self.attention_probs_dropout_prob,
            max_position_embeddings=self.max_position_embeddings,
            eos_token_ids=[2],
            bos_token_id=self.bos_token_id,
            pad_token_id=self.pad_token_id,
            decoder_start_token_id=self.pad_token_id,
            **self.config_updates,
        )
        inputs_dict = prepare_blenderbot_inputs_dict(config, input_ids, decoder_input_ids)
        return config, inputs_dict

    def check_decoder_model_past_large_inputs(self, config, inputs_dict):
        model = TFBlenderbotModel(config=config).get_decoder()
        input_ids = inputs_dict["input_ids"]

        input_ids = input_ids[:1, :]
        attention_mask = inputs_dict["attention_mask"][:1, :]
111
        head_mask = inputs_dict["head_mask"]
112
113
114
        self.batch_size = 1

        # first forward pass
115
        outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
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

        output, past_key_values = outputs.to_tuple()

        # create hypothetical next token and extent to next_input_ids
        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
        next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)

        # append to next input_ids and
        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
        next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)

        output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
        output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[0]

        self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])

        # select random slice
        random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
        output_from_past_slice = output_from_past[:, :, random_slice_idx]

        # test that outputs are equal for slice
        tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)


def prepare_blenderbot_inputs_dict(
    config,
    input_ids,
    decoder_input_ids,
    attention_mask=None,
    decoder_attention_mask=None,
147
148
    head_mask=None,
    decoder_head_mask=None,
149
    cross_attn_head_mask=None,
150
151
152
153
154
155
156
157
158
159
160
):
    if attention_mask is None:
        attention_mask = tf.cast(tf.math.not_equal(input_ids, config.pad_token_id), tf.int8)
    if decoder_attention_mask is None:
        decoder_attention_mask = tf.concat(
            [
                tf.ones(decoder_input_ids[:, :1].shape, dtype=tf.int8),
                tf.cast(tf.math.not_equal(decoder_input_ids[:, 1:], config.pad_token_id), tf.int8),
            ],
            axis=-1,
        )
161
162
163
164
    if head_mask is None:
        head_mask = tf.ones((config.encoder_layers, config.encoder_attention_heads))
    if decoder_head_mask is None:
        decoder_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
165
166
    if cross_attn_head_mask is None:
        cross_attn_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
167
168
169
170
171
    return {
        "input_ids": input_ids,
        "decoder_input_ids": decoder_input_ids,
        "attention_mask": attention_mask,
        "decoder_attention_mask": decoder_attention_mask,
172
173
        "head_mask": head_mask,
        "decoder_head_mask": decoder_head_mask,
174
        "cross_attn_head_mask": cross_attn_head_mask,
175
    }
176
177
178


@require_tf
179
class TFBlenderbotModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
180
    all_model_classes = (TFBlenderbotForConditionalGeneration, TFBlenderbotModel) if is_tf_available() else ()
181
    all_generative_model_classes = (TFBlenderbotForConditionalGeneration,) if is_tf_available() else ()
182
183
184
185
186
187
    pipeline_model_mapping = (
        {
            "conversational": TFBlenderbotForConditionalGeneration,
            "feature-extraction": TFBlenderbotModel,
            "summarization": TFBlenderbotForConditionalGeneration,
            "text2text-generation": TFBlenderbotForConditionalGeneration,
Yih-Dar's avatar
Yih-Dar committed
188
            "translation": TFBlenderbotForConditionalGeneration,
189
190
191
192
        }
        if is_tf_available()
        else {}
    )
193
194
    is_encoder_decoder = True
    test_pruning = False
195
    test_onnx = False
196
197

    def setUp(self):
Julien Plu's avatar
Julien Plu committed
198
        self.model_tester = TFBlenderbotModelTester(self)
199
200
201
202
203
        self.config_tester = ConfigTester(self, config_class=BlenderbotConfig)

    def test_config(self):
        self.config_tester.run_common_tests()

204
205
206
    def test_decoder_model_past_large_inputs(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
        self.model_tester.check_decoder_model_past_large_inputs(*config_and_inputs)
207

208
209
210
211
212
213
    def test_model_common_attributes(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            assert isinstance(model.get_input_embeddings(), tf.keras.layers.Layer)
214
215
216
217
218
219
220
221
222
223
224
225
226

            if model_class in self.all_generative_model_classes:
                x = model.get_output_embeddings()
                assert isinstance(x, tf.keras.layers.Layer)
                name = model.get_bias()
                assert isinstance(name, dict)
                for k, v in name.items():
                    assert isinstance(v, tf.Variable)
            else:
                x = model.get_output_embeddings()
                assert x is None
                name = model.get_bias()
                assert name is None
227

228
    @tooslow
Julien Plu's avatar
Julien Plu committed
229
230
231
    def test_saved_model_creation(self):
        pass

232

233
@require_tokenizers
Lysandre Debut's avatar
Lysandre Debut committed
234
@require_tf
235
236
237
class TFBlenderbot400MIntegrationTests(unittest.TestCase):
    src_text = ["My friends are cool but they eat too many carbs."]
    model_name = "facebook/blenderbot-400M-distill"
238
239
240

    @cached_property
    def tokenizer(self):
241
        return BlenderbotTokenizer.from_pretrained(self.model_name)
242
243
244

    @cached_property
    def model(self):
245
        model = TFAutoModelForSeq2SeqLM.from_pretrained(self.model_name)
246
247
248
        return model

    @slow
249
    def test_generation_from_long_input(self):
250
251
252
253
254
        model_inputs = self.tokenizer(self.src_text, return_tensors="tf")
        generated_ids = self.model.generate(
            model_inputs.input_ids,
        )
        generated_words = self.tokenizer.batch_decode(generated_ids.numpy(), skip_special_tokens=True)[0]
255
256
257
        assert (
            generated_words
            == " That's unfortunate. Are they trying to lose weight or are they just trying to be healthier?"
258
        )