transformers/tests/test_modeling_speech_to_text_2.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 Speech2Text model. """

import unittest

from transformers import Speech2Text2Config
from transformers.testing_utils import is_torch_available, require_torch, 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.models.speech_to_text_2.modeling_speech_to_text_2 import (
        Speech2Text2Decoder,
        Speech2Text2ForCausalLM,
    )


@require_torch
class Speech2Text2StandaloneDecoderModelTester:
    def __init__(
        self,
        parent,
        vocab_size=99,
        batch_size=13,
        d_model=16,
        decoder_seq_length=7,
        is_training=True,
        is_decoder=True,
        use_attention_mask=True,
        use_cache=False,
        use_labels=True,
        decoder_start_token_id=2,
        decoder_ffn_dim=32,
        decoder_layers=4,
        decoder_attention_heads=4,
        max_position_embeddings=30,
        pad_token_id=0,
        bos_token_id=1,
        eos_token_id=2,
        scope=None,
    ):
        self.parent = parent
        self.batch_size = batch_size
        self.decoder_seq_length = decoder_seq_length
        # For common tests
        self.seq_length = self.decoder_seq_length
        self.is_training = is_training
        self.use_attention_mask = use_attention_mask
        self.use_labels = use_labels

        self.vocab_size = vocab_size
        self.d_model = d_model
        self.hidden_size = d_model
        self.num_hidden_layers = decoder_layers
        self.decoder_layers = decoder_layers
        self.decoder_ffn_dim = decoder_ffn_dim
        self.decoder_attention_heads = decoder_attention_heads
        self.num_attention_heads = decoder_attention_heads
        self.eos_token_id = eos_token_id
        self.bos_token_id = bos_token_id
        self.pad_token_id = pad_token_id
        self.decoder_start_token_id = decoder_start_token_id
        self.use_cache = use_cache
        self.max_position_embeddings = max_position_embeddings

        self.scope = None
        self.decoder_key_length = decoder_seq_length
        self.base_model_out_len = 2
        self.decoder_attention_idx = 1

    def prepare_config_and_inputs(self):
        input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)

        attention_mask = None
        if self.use_attention_mask:
            attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)

        lm_labels = None
        if self.use_labels:
            lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)

        config = Speech2Text2Config(
            vocab_size=self.vocab_size,
            d_model=self.d_model,
            decoder_layers=self.decoder_layers,
            decoder_ffn_dim=self.decoder_ffn_dim,
            decoder_attention_heads=self.decoder_attention_heads,
            eos_token_id=self.eos_token_id,
            bos_token_id=self.bos_token_id,
            use_cache=self.use_cache,
            pad_token_id=self.pad_token_id,
            decoder_start_token_id=self.decoder_start_token_id,
            max_position_embeddings=self.max_position_embeddings,
        )

        return (
            config,
            input_ids,
            attention_mask,
            lm_labels,
        )

    def create_and_check_decoder_model_past(
        self,
        config,
        input_ids,
        attention_mask,
        lm_labels,
    ):
        config.use_cache = True
        model = Speech2Text2Decoder(config=config).to(torch_device).eval()
        input_ids = input_ids[:2]

        input_ids[input_ids == 0] += 1
        # first forward pass
        outputs = model(input_ids, use_cache=True)
        outputs_use_cache_conf = model(input_ids)
        outputs_no_past = model(input_ids, use_cache=False)

        self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
        self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)

        past_key_values = outputs["past_key_values"]

        # create hypothetical next token and extent to next_input_ids
        next_tokens = ids_tensor((2, 1), config.vocab_size - 1) + 1

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

        print(next_input_ids)

        output_from_no_past = model(next_input_ids)["last_hidden_state"]
        output_from_past = model(next_tokens, 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[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()

        # test that outputs are equal for slice
        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)

    def prepare_config_and_inputs_for_common(self):
        config_and_inputs = self.prepare_config_and_inputs()
        (
            config,
            input_ids,
            attention_mask,
            lm_labels,
        ) = config_and_inputs

        inputs_dict = {
            "input_ids": input_ids,
            "attention_mask": attention_mask,
        }
        return config, inputs_dict


@require_torch
class Speech2Text2StandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
    all_model_classes = (Speech2Text2Decoder, Speech2Text2ForCausalLM) if is_torch_available() else ()
    all_generative_model_classes = (Speech2Text2ForCausalLM,) if is_torch_available() else ()
    test_pruning = False

    def setUp(
        self,
    ):
        self.model_tester = Speech2Text2StandaloneDecoderModelTester(self, is_training=False)
        self.config_tester = ConfigTester(self, config_class=Speech2Text2Config)

    # not implemented currently
    def test_inputs_embeds(self):
        pass

    # speech2text2 has no base model
    def test_save_load_fast_init_from_base(self):
        pass

    # speech2text2 has no base model
    def test_save_load_fast_init_to_base(self):
        pass

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

    def test_decoder_model_past(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)

    # decoder cannot keep gradients
    def test_retain_grad_hidden_states_attentions(self):
        return