test_modeling_blenderbot.py

# coding=utf-8
# Copyright 2021, The HuggingFace Inc. team. All rights reserved.
#
# 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.
""" Testing suite for the PyTorch Blenderbot model. """


import tempfile
import unittest

from transformers import is_torch_available
from transformers.file_utils import cached_property
from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device

from .test_configuration_common import ConfigTester
from .test_generation_utils import GenerationTesterMixin
from .test_modeling_common import ModelTesterMixin, ids_tensor


if is_torch_available():
    import torch

    from transformers import BlenderbotConfig, BlenderbotForConditionalGeneration, BlenderbotModel, BlenderbotTokenizer
    from transformers.models.blenderbot.modeling_blenderbot import BlenderbotDecoder, BlenderbotEncoder


def prepare_blenderbot_inputs_dict(
    config,
    input_ids,
    decoder_input_ids,
    attention_mask=None,
    decoder_attention_mask=None,
    head_mask=None,
    decoder_head_mask=None,
):
    if attention_mask is None:
        attention_mask = input_ids.ne(config.pad_token_id)
    if decoder_attention_mask is None:
        decoder_attention_mask = decoder_input_ids.ne(config.pad_token_id)
    if head_mask is None:
        head_mask = torch.ones(config.encoder_layers, config.encoder_attention_heads, device=torch_device)
    if decoder_head_mask is None:
        decoder_head_mask = torch.ones(config.decoder_layers, config.decoder_attention_heads, device=torch_device)
    return {
        "input_ids": input_ids,
        "decoder_input_ids": decoder_input_ids,
        "attention_mask": attention_mask,
        "decoder_attention_mask": attention_mask,
        "head_mask": head_mask,
        "decoder_head_mask": decoder_head_mask,
    }


@require_torch
class BlenderbotModelTester:
    def __init__(
        self,
        parent,
        batch_size=13,
        seq_length=7,
        is_training=True,
        use_labels=False,
        vocab_size=99,
        hidden_size=16,
        num_hidden_layers=2,
        num_attention_heads=4,
        intermediate_size=4,
        hidden_act="gelu",
        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_act = hidden_act
        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(self):
        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
            3,
        )
        input_ids[:, -1] = self.eos_token_id  # Eos Token

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

        config = BlenderbotConfig(
            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_id=self.eos_token_id,
            bos_token_id=self.bos_token_id,
            pad_token_id=self.pad_token_id,
        )
        inputs_dict = prepare_blenderbot_inputs_dict(config, input_ids, decoder_input_ids)
        return config, inputs_dict

    def prepare_config_and_inputs_for_common(self):
        config, inputs_dict = self.prepare_config_and_inputs()
        return config, inputs_dict

    def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict):
        model = BlenderbotModel(config=config).get_decoder().to(torch_device).eval()
        input_ids = inputs_dict["input_ids"]
        attention_mask = inputs_dict["attention_mask"]
        head_mask = inputs_dict["head_mask"]

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

        output, past_key_values = outputs.to_tuple()

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

        # append to next input_ids and
        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
        next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1)

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

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

        self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])

        # test that outputs are equal for slice
        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2))

    def check_encoder_decoder_model_standalone(self, config, inputs_dict):
        model = BlenderbotModel(config=config).to(torch_device).eval()
        outputs = model(**inputs_dict)

        encoder_last_hidden_state = outputs.encoder_last_hidden_state
        last_hidden_state = outputs.last_hidden_state

        with tempfile.TemporaryDirectory() as tmpdirname:
            encoder = model.get_encoder()
            encoder.save_pretrained(tmpdirname)
            encoder = BlenderbotEncoder.from_pretrained(tmpdirname).to(torch_device)

        encoder_last_hidden_state_2 = encoder(inputs_dict["input_ids"], attention_mask=inputs_dict["attention_mask"])[
            0
        ]

        self.parent.assertTrue((encoder_last_hidden_state_2 - encoder_last_hidden_state).abs().max().item() < 1e-3)

        with tempfile.TemporaryDirectory() as tmpdirname:
            decoder = model.get_decoder()
            decoder.save_pretrained(tmpdirname)
            decoder = BlenderbotDecoder.from_pretrained(tmpdirname).to(torch_device)

        last_hidden_state_2 = decoder(
            input_ids=inputs_dict["decoder_input_ids"],
            attention_mask=inputs_dict["decoder_attention_mask"],
            encoder_hidden_states=encoder_last_hidden_state,
            encoder_attention_mask=inputs_dict["attention_mask"],
        )[0]

        self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max().item() < 1e-3)


@require_torch
class BlenderbotModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
    all_model_classes = (BlenderbotModel, BlenderbotForConditionalGeneration) if is_torch_available() else ()
    all_generative_model_classes = (BlenderbotForConditionalGeneration,) if is_torch_available() else ()
    is_encoder_decoder = True
    test_pruning = False
    test_missing_keys = False

    def setUp(self):
        self.model_tester = BlenderbotModelTester(self)
        self.config_tester = ConfigTester(self, config_class=BlenderbotConfig)

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

    def test_save_load_strict(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs()
        for model_class in self.all_model_classes:
            model = model_class(config)

            with tempfile.TemporaryDirectory() as tmpdirname:
                model.save_pretrained(tmpdirname)
                model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True)
            self.assertEqual(info["missing_keys"], [])

    def test_decoder_model_past_with_large_inputs(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)

    def test_encoder_decoder_model_standalone(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
        self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs)

    def test_generate_fp16(self):
        config, input_dict = self.model_tester.prepare_config_and_inputs()
        input_ids = input_dict["input_ids"]
        attention_mask = input_ids.ne(1).to(torch_device)
        model = BlenderbotForConditionalGeneration(config).eval().to(torch_device)
        if torch_device == "cuda":
            model.half()
        model.generate(input_ids, attention_mask=attention_mask)
        model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)


def assert_tensors_close(a, b, atol=1e-12, prefix=""):
    """If tensors have different shapes, different values or a and b are not both tensors, raise a nice Assertion error."""
    if a is None and b is None:
        return True
    try:
        if torch.allclose(a, b, atol=atol):
            return True
        raise
    except Exception:
        pct_different = (torch.gt((a - b).abs(), atol)).float().mean().item()
        if a.numel() > 100:
            msg = f"tensor values are {pct_different:.1%} percent different."
        else:
            msg = f"{a} != {b}"
        if prefix:
            msg = prefix + ": " + msg
        raise AssertionError(msg)


@unittest.skipUnless(torch_device != "cpu", "3B test too slow on CPU.")
@require_torch
@require_sentencepiece
@require_tokenizers
class Blenderbot3BIntegrationTests(unittest.TestCase):
    ckpt = "facebook/blenderbot-3B"

    @cached_property
    def tokenizer(self):
        return BlenderbotTokenizer.from_pretrained(self.ckpt)

    @slow
    def test_generation_from_short_input_same_as_parlai_3B(self):
        FASTER_GEN_KWARGS = dict(num_beams=1, early_stopping=True, min_length=15, max_length=25)
        TOK_DECODE_KW = dict(skip_special_tokens=True, clean_up_tokenization_spaces=True)

        torch.cuda.empty_cache()
        model = BlenderbotForConditionalGeneration.from_pretrained(self.ckpt).half().to(torch_device)

        src_text = ["Sam"]
        model_inputs = self.tokenizer(src_text, return_tensors="pt").to(torch_device)

        generated_utterances = model.generate(**model_inputs, **FASTER_GEN_KWARGS)
        tgt_text = 'Sam is a great name. It means "sun" in Gaelic.'

        generated_txt = self.tokenizer.batch_decode(generated_utterances, **TOK_DECODE_KW)
        assert generated_txt[0].strip() == tgt_text

        src_text = "Social anxiety\nWow, I am never shy. Do you have anxiety?\nYes. I end up sweating and blushing and feel like i'm going to throw up.\nand why is that?"

        model_inputs = self.tokenizer([src_text], return_tensors="pt").to(torch_device)

        generated_ids = model.generate(**model_inputs, **FASTER_GEN_KWARGS)[0]
        reply = self.tokenizer.decode(generated_ids, **TOK_DECODE_KW)

        assert "I think it's because we are so worried about what people think of us." == reply.strip()
        del model