Uniformize model processors (#31368)

* add initial design for uniform processors + align model * add uniform processors for altclip + chinese_clip * add uniform processors for blip + blip2 * fix mutable default 👀 * add configuration test * handle structured kwargs w defaults + add test * protect torch-specific test * fix style * fix * rebase * update processor to generic kwargs + test * fix style * add sensible kwargs merge * update test * fix assertEqual * move kwargs merging to processing common * rework kwargs for type hinting * just get Unpack from extensions * run-slow[align] * handle kwargs passed as nested dict * add from_pretrained test for nested kwargs handling * [run-slow]align * update documentation + imports * update audio inputs * protect audio types, silly * try removing imports * make things simpler * simplerer * move out kwargs test to common mixin * [run-slow]align * skip tests for old processors * [run-slow]align, clip * !$#@!! protect imports, darn it * [run-slow]align, clip * [run-slow]align, clip * update common processor testing * add altclip * add chinese_clip * add pad_size * [run-slow]align, clip, chinese_clip, altclip * remove duplicated tests * fix * add blip, blip2, bridgetower Added tests for bridgetower which override common. Also modified common tests to force center cropping if existing * fix * update doc * improve documentation for default values * add model_max_length testing This parameter depends on tokenizers received. * Raise if kwargs are specified in two places * fix * removed copied from * match defaults * force padding * fix tokenizer test * clean defaults * move tests to common * add missing import * fix * adapt bridgetower tests to shortest edge * uniformize donut processor + tests * add wav2vec2 * extend common testing to audio processors * add testing + bert version * propagate common kwargs to different modalities * BC order of arguments * check py version * revert kwargs merging * add draft overlap test * update * fix blip2 and wav2vec due to updates * fix copies * ensure overlapping kwargs do not disappear * replace .pop by .get to handle duplicated kwargs * fix copies * fix missing import * add clearly wav2vec2_bert to uniformized models * fix copies * increase number of features * fix style * [run-slow] blip, blip2, bridgetower, donut, wav2vec2, wav2vec2_bert * [run-slow] blip, blip_2, bridgetower, donut, wav2vec2, wav2vec2_bert * fix concatenation * [run-slow] blip, blip_2, bridgetower, donut, wav2vec2, wav2vec2_bert * Update tests/test_processing_common.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * 🧹 * address comments * clean up + tests * [run-slow] instructblip, blip, blip_2, bridgetower, donut, wav2vec2, wav2vec2_bert --------- Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
2024-10-02 10:41:08 +02:00
parent 2292be6c1b
commit 50290cf7a0
18 changed files with 769 additions and 273 deletions
--- a/tests/models/bridgetower/test_processing_bridgetower.py
+++ b/tests/models/bridgetower/test_processing_bridgetower.py
@@ -0,0 +1,218 @@
+# Copyright 2023 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 shutil
+import tempfile
+import unittest
+
+import numpy as np
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import (
+        AutoProcessor,
+        BridgeTowerImageProcessor,
+        BridgeTowerProcessor,
+        RobertaTokenizerFast,
+    )
+
+
+@require_vision
+class Blip2ProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = BridgeTowerProcessor
+
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+
+        image_processor = BridgeTowerImageProcessor()
+        tokenizer = RobertaTokenizerFast.from_pretrained("BridgeTower/bridgetower-large-itm-mlm-itc")
+
+        processor = BridgeTowerProcessor(image_processor, tokenizer)
+
+        processor.save_pretrained(self.tmpdirname)
+
+    def get_tokenizer(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
+
+    def get_image_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def prepare_image_inputs(self):
+        """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
+        or a list of PyTorch tensors if one specifies torchify=True.
+        """
+
+        image_inputs = [np.random.randint(255, size=(3, 30, 400), dtype=np.uint8)]
+
+        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+
+        return image_inputs
+
+    # Some kwargs tests are overriden from common tests to handle shortest_edge
+    # and size_divisor behaviour
+
+    @require_torch
+    @require_vision
+    def test_image_processor_defaults_preserved_by_image_kwargs(self):
+        if "image_processor" not in self.processor_class.attributes:
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        image_processor = self.get_component(
+            "image_processor",
+            crop_size={"shortest_edge": 234, "longest_edge": 234},
+        )
+        tokenizer = self.get_component("tokenizer", max_length=117, padding="max_length")
+
+        processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+        self.assertEqual(len(inputs["pixel_values"][0][0]), 234)
+
+    @require_torch
+    @require_vision
+    def test_structured_kwargs_nested_from_dict(self):
+        if "image_processor" not in self.processor_class.attributes:
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+
+        image_processor = self.get_component("image_processor")
+        tokenizer = self.get_component("tokenizer")
+
+        processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
+        self.skip_processor_without_typed_kwargs(processor)
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        # Define the kwargs for each modality
+        all_kwargs = {
+            "common_kwargs": {"return_tensors": "pt"},
+            "images_kwargs": {
+                "crop_size": {"shortest_edge": 214},
+            },
+            "text_kwargs": {"padding": "max_length", "max_length": 76},
+        }
+
+        inputs = processor(text=input_str, images=image_input, **all_kwargs)
+        self.assertEqual(inputs["pixel_values"].shape[2], 214)
+
+        self.assertEqual(len(inputs["input_ids"][0]), 76)
+
+    @require_torch
+    @require_vision
+    def test_kwargs_overrides_default_image_processor_kwargs(self):
+        if "image_processor" not in self.processor_class.attributes:
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        image_processor = self.get_component("image_processor", crop_size={"shortest_edge": 234})
+        tokenizer = self.get_component("tokenizer", max_length=117)
+        if not tokenizer.pad_token:
+            tokenizer.pad_token = "[TEST_PAD]"
+        processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+        inputs = processor(text=input_str, images=image_input, crop_size={"shortest_edge": 224})
+        self.assertEqual(len(inputs["pixel_values"][0][0]), 224)
+
+    @require_torch
+    @require_vision
+    def test_unstructured_kwargs_batched(self):
+        if "image_processor" not in self.processor_class.attributes:
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        image_processor = self.get_component("image_processor")
+        tokenizer = self.get_component("tokenizer")
+        if not tokenizer.pad_token:
+            tokenizer.pad_token = "[TEST_PAD]"
+        processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        input_str = ["lower newer", "upper older longer string"]
+        image_input = self.prepare_image_inputs() * 2
+        inputs = processor(
+            text=input_str,
+            images=image_input,
+            return_tensors="pt",
+            crop_size={"shortest_edge": 214},
+            padding="longest",
+            max_length=76,
+        )
+        self.assertEqual(inputs["pixel_values"].shape[2], 214)
+
+        self.assertEqual(len(inputs["input_ids"][0]), 6)
+
+    @require_torch
+    @require_vision
+    def test_unstructured_kwargs(self):
+        if "image_processor" not in self.processor_class.attributes:
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        image_processor = self.get_component("image_processor")
+        tokenizer = self.get_component("tokenizer")
+        if not tokenizer.pad_token:
+            tokenizer.pad_token = "[TEST_PAD]"
+        processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+        inputs = processor(
+            text=input_str,
+            images=image_input,
+            return_tensors="pt",
+            crop_size={"shortest_edge": 214},
+            padding="max_length",
+            max_length=76,
+        )
+
+        self.assertEqual(inputs["pixel_values"].shape[2], 214)
+        self.assertEqual(len(inputs["input_ids"][0]), 76)
+
+    @require_torch
+    @require_vision
+    def test_structured_kwargs_nested(self):
+        if "image_processor" not in self.processor_class.attributes:
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        image_processor = self.get_component("image_processor")
+        tokenizer = self.get_component("tokenizer")
+        if not tokenizer.pad_token:
+            tokenizer.pad_token = "[TEST_PAD]"
+        processor = self.processor_class(tokenizer=tokenizer, image_processor=image_processor)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        # Define the kwargs for each modality
+        all_kwargs = {
+            "common_kwargs": {"return_tensors": "pt"},
+            "images_kwargs": {"crop_size": {"shortest_edge": 214}},
+            "text_kwargs": {"padding": "max_length", "max_length": 76},
+        }
+
+        inputs = processor(text=input_str, images=image_input, **all_kwargs)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        self.assertEqual(inputs["pixel_values"].shape[2], 214)
+
+        self.assertEqual(len(inputs["input_ids"][0]), 76)