init

transformers/tests/models/trocr/test_modeling_trocr.py

@@ -0,0 +1,188 @@
+# 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 TrOCR model."""
+
+import unittest
+
+from transformers import TrOCRConfig
+from transformers.testing_utils import is_torch_available, require_torch, torch_device
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, ids_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers.models.trocr.modeling_trocr import TrOCRDecoder, TrOCRForCausalLM
+
+
+@require_torch
+class TrOCRStandaloneDecoderModelTester:
+    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=2,
+        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 = TrOCRConfig(
+            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 = TrOCRDecoder(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)
+
+        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 TrOCRStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (TrOCRDecoder, TrOCRForCausalLM) if is_torch_available() else ()
+    pipeline_model_mapping = {"text-generation": TrOCRForCausalLM} if is_torch_available() else {}
+    fx_compatible = True
+    test_pruning = False
+
+    def setUp(self):
+        self.model_tester = TrOCRStandaloneDecoderModelTester(self, is_training=False)
+        self.config_tester = ConfigTester(self, config_class=TrOCRConfig)
+
+    @unittest.skip(reason="Not yet implemented")
+    def test_inputs_embeds(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)
+
+    @unittest.skip(reason="Decoder cannot keep gradients")
+    def test_retain_grad_hidden_states_attentions(self):
+        return
+
+    @unittest.skip(reason="The model doesn't support left padding")  # and it's not used enough to be worth fixing :)
+    def test_left_padding_compatibility(self):
+        pass

transformers/tests/models/trocr/test_processing_trocr.py

@@ -0,0 +1,133 @@
+import os
+import shutil
+import tempfile
+import unittest
+
+import pytest
+
+from transformers.models.xlm_roberta.tokenization_xlm_roberta import VOCAB_FILES_NAMES
+from transformers.testing_utils import (
+    require_sentencepiece,
+    require_tokenizers,
+    require_vision,
+)
+from transformers.utils import is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+if is_vision_available():
+    from transformers import TrOCRProcessor, ViTImageProcessor, XLMRobertaTokenizerFast
+
+
+@require_sentencepiece
+@require_tokenizers
+@require_vision
+class TrOCRProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    text_input_name = "labels"
+    processor_class = TrOCRProcessor
+
+    @classmethod
+    def setUpClass(cls):
+        cls.tmpdirname = tempfile.mkdtemp()
+
+        vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]", "want", "##want", "##ed", "wa", "un", "runn", "##ing", ",", "low", "lowest"]  # fmt: skip
+        cls.vocab_file = os.path.join(cls.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        with open(cls.vocab_file, "w", encoding="utf-8") as vocab_writer:
+            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
+
+        image_processor = ViTImageProcessor.from_pretrained("hf-internal-testing/tiny-random-vit")
+        tokenizer = XLMRobertaTokenizerFast.from_pretrained("FacebookAI/xlm-roberta-base")
+        processor = TrOCRProcessor(image_processor=image_processor, tokenizer=tokenizer)
+        processor.save_pretrained(cls.tmpdirname)
+
+    @classmethod
+    def tearDownClass(cls):
+        shutil.rmtree(cls.tmpdirname, ignore_errors=True)
+
+    def get_tokenizer(self, **kwargs):
+        return XLMRobertaTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_image_processor(self, **kwargs):
+        return ViTImageProcessor.from_pretrained(self.tmpdirname, **kwargs)
+
+    def test_save_load_pretrained_default(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            image_processor = self.get_image_processor()
+            tokenizer = self.get_tokenizer()
+            processor = TrOCRProcessor(image_processor=image_processor, tokenizer=tokenizer)
+
+            processor.save_pretrained(tmpdir)
+            processor = TrOCRProcessor.from_pretrained(tmpdir)
+
+        self.assertIsInstance(processor.tokenizer, XLMRobertaTokenizerFast)
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab())
+        self.assertIsInstance(processor.image_processor, ViTImageProcessor)
+        self.assertEqual(processor.image_processor.to_json_string(), image_processor.to_json_string())
+
+    def test_save_load_pretrained_additional_features(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            processor = TrOCRProcessor(tokenizer=self.get_tokenizer(), image_processor=self.get_image_processor())
+            processor.save_pretrained(tmpdir)
+            tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
+            image_processor_add_kwargs = self.get_image_processor(do_normalize=False, padding_value=1.0)
+
+            processor = TrOCRProcessor.from_pretrained(
+                tmpdir, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
+            )
+
+        self.assertIsInstance(processor.tokenizer, XLMRobertaTokenizerFast)
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
+
+        self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.image_processor, ViTImageProcessor)
+
+    def test_image_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        processor = TrOCRProcessor(tokenizer=tokenizer, image_processor=image_processor)
+        image_input = self.prepare_image_inputs()
+
+        input_feat_extract = image_processor(image_input, return_tensors="np")
+        input_processor = processor(images=image_input, return_tensors="np")
+
+        for key in input_feat_extract:
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        processor = TrOCRProcessor(tokenizer=tokenizer, image_processor=image_processor)
+        input_str = "lower newer"
+
+        encoded_processor = processor(text=input_str)
+        encoded_tok = tokenizer(input_str)
+
+        for key in encoded_tok:
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_processor_text(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        processor = TrOCRProcessor(tokenizer=tokenizer, image_processor=image_processor)
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(list(inputs.keys()), ["pixel_values", "labels"])
+
+        # test if it raises when no input is passed
+        with pytest.raises(ValueError):
+            processor()
+
+    def test_tokenizer_decode(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+        processor = TrOCRProcessor(tokenizer=tokenizer, image_processor=image_processor)
+        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
+
+        decoded_processor = processor.batch_decode(predicted_ids)
+        decoded_tok = tokenizer.batch_decode(predicted_ids)
+
+        self.assertListEqual(decoded_tok, decoded_processor)