test_modeling_imagegpt.py

# coding=utf-8
# Copyright 2021 The HuggingFace 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.


import copy
import inspect
import os
import tempfile
import unittest

from transformers import ImageGPTConfig
from transformers.file_utils import cached_property, is_torch_available, is_vision_available
from transformers.testing_utils import require_torch, require_vision, slow, torch_device

from .test_configuration_common import ConfigTester
from .test_generation_utils import GenerationTesterMixin
from .test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor, random_attention_mask


if is_torch_available():
    import torch

    from transformers import (
        IMAGEGPT_PRETRAINED_MODEL_ARCHIVE_LIST,
        ImageGPTForCausalImageModeling,
        ImageGPTForImageClassification,
        ImageGPTModel,
    )

if is_vision_available():
    from PIL import Image

    from transformers import ImageGPTFeatureExtractor


class ImageGPTModelTester:
    def __init__(
        self,
        parent,
        batch_size=14,
        seq_length=7,
        is_training=True,
        use_token_type_ids=True,
        use_input_mask=True,
        use_labels=True,
        use_mc_token_ids=True,
        vocab_size=99,
        hidden_size=32,
        num_hidden_layers=5,
        num_attention_heads=4,
        intermediate_size=37,
        hidden_act="gelu",
        hidden_dropout_prob=0.1,
        attention_probs_dropout_prob=0.1,
        max_position_embeddings=512,
        type_vocab_size=16,
        type_sequence_label_size=2,
        initializer_range=0.02,
        num_labels=3,
        num_choices=4,
        scope=None,
    ):
        self.parent = parent
        self.batch_size = batch_size
        self.seq_length = seq_length
        self.is_training = is_training
        self.use_token_type_ids = use_token_type_ids
        self.use_input_mask = use_input_mask
        self.use_labels = use_labels
        self.use_mc_token_ids = use_mc_token_ids
        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.type_vocab_size = type_vocab_size
        self.type_sequence_label_size = type_sequence_label_size
        self.initializer_range = initializer_range
        self.num_labels = num_labels
        self.num_choices = num_choices
        self.scope = None

    def get_large_model_config(self):
        return ImageGPTConfig.from_pretrained("imagegpt")

    def prepare_config_and_inputs(
        self, gradient_checkpointing=False, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False
    ):
        pixel_values = ids_tensor([self.batch_size, self.seq_length], self.vocab_size - 1)

        input_mask = None
        if self.use_input_mask:
            input_mask = random_attention_mask([self.batch_size, self.seq_length])

        token_type_ids = None
        if self.use_token_type_ids:
            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)

        mc_token_ids = None
        if self.use_mc_token_ids:
            mc_token_ids = ids_tensor([self.batch_size, self.num_choices], self.seq_length)

        sequence_labels = None
        token_labels = None
        choice_labels = None
        if self.use_labels:
            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
            choice_labels = ids_tensor([self.batch_size], self.num_choices)

        config = self.get_config(
            gradient_checkpointing=gradient_checkpointing,
            scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
            reorder_and_upcast_attn=reorder_and_upcast_attn,
        )

        head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)

        return (
            config,
            pixel_values,
            input_mask,
            head_mask,
            token_type_ids,
            mc_token_ids,
            sequence_labels,
            token_labels,
            choice_labels,
        )

    def get_config(
        self, gradient_checkpointing=False, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False
    ):
        return ImageGPTConfig(
            vocab_size=self.vocab_size,
            n_embd=self.hidden_size,
            n_layer=self.num_hidden_layers,
            n_head=self.num_attention_heads,
            n_inner=self.intermediate_size,
            activation_function=self.hidden_act,
            resid_pdrop=self.hidden_dropout_prob,
            attn_pdrop=self.attention_probs_dropout_prob,
            n_positions=self.max_position_embeddings,
            type_vocab_size=self.type_vocab_size,
            initializer_range=self.initializer_range,
            use_cache=True,
            gradient_checkpointing=gradient_checkpointing,
            scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
            reorder_and_upcast_attn=reorder_and_upcast_attn,
        )

    def prepare_config_and_inputs_for_decoder(self):
        (
            config,
            pixel_values,
            input_mask,
            head_mask,
            token_type_ids,
            mc_token_ids,
            sequence_labels,
            token_labels,
            choice_labels,
        ) = self.prepare_config_and_inputs()

        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)

        return (
            config,
            pixel_values,
            input_mask,
            head_mask,
            token_type_ids,
            sequence_labels,
            token_labels,
            choice_labels,
            encoder_hidden_states,
            encoder_attention_mask,
        )

    def create_and_check_imagegpt_model(self, config, pixel_values, input_mask, head_mask, token_type_ids, *args):
        model = ImageGPTModel(config=config)
        model.to(torch_device)
        model.eval()

        result = model(pixel_values, token_type_ids=token_type_ids, head_mask=head_mask)
        result = model(pixel_values, token_type_ids=token_type_ids)
        result = model(pixel_values)

        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
        self.parent.assertEqual(len(result.past_key_values), config.n_layer)

    def create_and_check_lm_head_model(self, config, pixel_values, input_mask, head_mask, token_type_ids, *args):
        model = ImageGPTForCausalImageModeling(config)
        model.to(torch_device)
        model.eval()

        labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size - 1)
        result = model(pixel_values, token_type_ids=token_type_ids, labels=labels)
        self.parent.assertEqual(result.loss.shape, ())
        # ImageGPTForCausalImageModeling doens't have tied input- and output embeddings
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size - 1))

    def create_and_check_imagegpt_for_image_classification(
        self, config, pixel_values, input_mask, head_mask, token_type_ids, mc_token_ids, sequence_labels, *args
    ):
        config.num_labels = self.num_labels
        model = ImageGPTForImageClassification(config)
        model.to(torch_device)
        model.eval()
        result = model(pixel_values, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))

    def prepare_config_and_inputs_for_common(self):
        config_and_inputs = self.prepare_config_and_inputs()

        (
            config,
            pixel_values,
            input_mask,
            head_mask,
            token_type_ids,
            mc_token_ids,
            sequence_labels,
            token_labels,
            choice_labels,
        ) = config_and_inputs

        inputs_dict = {
            "pixel_values": pixel_values,
            "token_type_ids": token_type_ids,
            "head_mask": head_mask,
        }

        return config, inputs_dict


@require_torch
class ImageGPTModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):

    all_model_classes = (
        (ImageGPTForCausalImageModeling, ImageGPTForImageClassification, ImageGPTModel) if is_torch_available() else ()
    )
    all_generative_model_classes = (ImageGPTForCausalImageModeling,) if is_torch_available() else ()
    test_missing_keys = False
    input_name = "pixel_values"

    # as ImageGPTForImageClassification isn't included in any auto mapping, we add labels here
    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
        inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)

        if return_labels:
            if model_class.__name__ == "ImageGPTForImageClassification":
                inputs_dict["labels"] = torch.zeros(
                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
                )

        return inputs_dict

    # we overwrite the _check_scores method of GenerationTesterMixin, as ImageGPTForCausalImageModeling doesn't have tied input- and output embeddings
    def _check_scores(self, batch_size, scores, length, config):
        expected_shape = (batch_size, config.vocab_size - 1)
        self.assertIsInstance(scores, tuple)
        self.assertEqual(len(scores), length)
        self.assertListEqual([iter_scores.shape for iter_scores in scores], [expected_shape] * len(scores))

    def setUp(self):
        self.model_tester = ImageGPTModelTester(self)
        self.config_tester = ConfigTester(self, config_class=ImageGPTConfig, n_embd=37)

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

    def test_imagegpt_model(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_imagegpt_model(*config_and_inputs)

    def test_imagegpt_causal_lm(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_lm_head_model(*config_and_inputs)

    def test_imagegpt_image_classification(self):
        config_and_inputs = self.model_tester.prepare_config_and_inputs()
        self.model_tester.create_and_check_imagegpt_for_image_classification(*config_and_inputs)

    @slow
    def test_model_from_pretrained(self):
        for model_name in IMAGEGPT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
            model = ImageGPTModel.from_pretrained(model_name)
            self.assertIsNotNone(model)

    def test_forward_signature(self):
        config, _ = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            model = model_class(config)
            signature = inspect.signature(model.forward)
            # signature.parameters is an OrderedDict => so arg_names order is deterministic
            arg_names = [*signature.parameters.keys()]

            expected_arg_names = ["pixel_values"]
            self.assertListEqual(arg_names[:1], expected_arg_names)

    def test_resize_tokens_embeddings(self):
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
        if not self.test_resize_embeddings:
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config)
            model.to(torch_device)

            if self.model_tester.is_training is False:
                model.eval()

            model_vocab_size = config.vocab_size
            # Retrieve the embeddings and clone theme
            model_embed = model.resize_token_embeddings(model_vocab_size)
            cloned_embeddings = model_embed.weight.clone()

            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
            model_embed = model.resize_token_embeddings(model_vocab_size + 10)
            self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
            # Check that it actually resizes the embeddings matrix
            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
            model_embed = model.resize_token_embeddings(model_vocab_size - 15)
            self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
            # Check that it actually resizes the embeddings matrix
            self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)

            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            # Input ids should be clamped to the maximum size of the vocabulary
            inputs_dict["pixel_values"].clamp_(max=model_vocab_size - 15 - 1)

            # Check that adding and removing tokens has not modified the first part of the embedding matrix.
            models_equal = True
            for p1, p2 in zip(cloned_embeddings, model_embed.weight):
                if p1.data.ne(p2.data).sum() > 0:
                    models_equal = False

            self.assertTrue(models_equal)

    def test_resize_embeddings_untied(self):
        (
            original_config,
            inputs_dict,
        ) = self.model_tester.prepare_config_and_inputs_for_common()
        if not self.test_resize_embeddings:
            return

        original_config.tie_word_embeddings = False

        # if model cannot untied embeddings -> leave test
        if original_config.tie_word_embeddings:
            return

        for model_class in self.all_model_classes:
            config = copy.deepcopy(original_config)
            model = model_class(config).to(torch_device)

            # if no output embeddings -> leave test
            if model.get_output_embeddings() is None:
                continue

            # Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
            model_vocab_size = config.vocab_size
            model.resize_token_embeddings(model_vocab_size + 10)
            self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
            output_embeds = model.get_output_embeddings()
            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size + 10)
            # Check bias if present
            if output_embeds.bias is not None:
                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size + 10)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

            # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
            model.resize_token_embeddings(model_vocab_size - 15)
            self.assertEqual(model.config.vocab_size, model_vocab_size - 15)
            # Check that it actually resizes the embeddings matrix
            output_embeds = model.get_output_embeddings()
            self.assertEqual(output_embeds.weight.shape[0], model_vocab_size - 15)
            # Check bias if present
            if output_embeds.bias is not None:
                self.assertEqual(output_embeds.bias.shape[0], model_vocab_size - 15)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            # Input ids should be clamped to the maximum size of the vocabulary
            inputs_dict["pixel_values"].clamp_(max=model_vocab_size - 15 - 1)
            # Check that the model can still do a forward pass successfully (every parameter should be resized)
            model(**self._prepare_for_class(inputs_dict, model_class))

    def test_inputs_embeds(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)
            model.to(torch_device)
            model.eval()

            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))

            pixel_values = inputs["pixel_values"]
            del inputs["pixel_values"]

            wte = model.get_input_embeddings()
            inputs["inputs_embeds"] = wte(pixel_values)

            with torch.no_grad():
                model(**inputs)[0]

    def _create_and_check_torchscript(self, config, inputs_dict):
        if not self.test_torchscript:
            return

        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
        configs_no_init.torchscript = True
        for model_class in self.all_model_classes:
            model = model_class(config=configs_no_init)
            model.to(torch_device)
            model.eval()
            inputs = self._prepare_for_class(inputs_dict, model_class)

            try:
                pixel_values = inputs["pixel_values"]
                traced_model = torch.jit.trace(model, pixel_values)
            except RuntimeError:
                self.fail("Couldn't trace module.")

            with tempfile.TemporaryDirectory() as tmp_dir_name:
                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")

                try:
                    torch.jit.save(traced_model, pt_file_name)
                except Exception:
                    self.fail("Couldn't save module.")

                try:
                    loaded_model = torch.jit.load(pt_file_name)
                except Exception:
                    self.fail("Couldn't load module.")

            model.to(torch_device)
            model.eval()

            loaded_model.to(torch_device)
            loaded_model.eval()

            model_state_dict = model.state_dict()
            loaded_model_state_dict = loaded_model.state_dict()

            non_persistent_buffers = {}
            for key in loaded_model_state_dict.keys():
                if key not in model_state_dict.keys():
                    non_persistent_buffers[key] = loaded_model_state_dict[key]

            loaded_model_state_dict = {
                key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers
            }

            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))

            model_buffers = list(model.buffers())
            for non_persistent_buffer in non_persistent_buffers.values():
                found_buffer = False
                for i, model_buffer in enumerate(model_buffers):
                    if torch.equal(non_persistent_buffer, model_buffer):
                        found_buffer = True
                        break

                self.assertTrue(found_buffer)
                model_buffers.pop(i)

            models_equal = True
            for layer_name, p1 in model_state_dict.items():
                if layer_name in loaded_model_state_dict:
                    p2 = loaded_model_state_dict[layer_name]
                    if p1.data.ne(p2.data).sum() > 0:
                        models_equal = False

            self.assertTrue(models_equal)


# We will verify our results on an image of cute cats
def prepare_img():
    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
    return image


@require_torch
@require_vision
class ImageGPTModelIntegrationTest(unittest.TestCase):
    @cached_property
    def default_feature_extractor(self):
        return ImageGPTFeatureExtractor.from_pretrained("openai/imagegpt-small") if is_vision_available() else None

    @slow
    def test_inference_causal_lm_head(self):
        model = ImageGPTForCausalImageModeling.from_pretrained("openai/imagegpt-small").to(torch_device)

        feature_extractor = self.default_feature_extractor
        image = prepare_img()
        inputs = feature_extractor(images=image, return_tensors="pt").to(torch_device)

        # forward pass
        outputs = model(**inputs)

        # verify the logits
        expected_shape = torch.Size((1, 1024, 512))
        self.assertEqual(outputs.logits.shape, expected_shape)

        expected_slice = torch.tensor(
            [[2.3445, 2.6889, 2.7313], [1.0530, 1.2416, 0.5699], [0.2205, 0.7749, 0.3953]]
        ).to(torch_device)

        self.assertTrue(torch.allclose(outputs.logits[0, :3, :3], expected_slice, atol=1e-4))