Add Mix transformer (#1780)

* Add MixTransformer

* fix testcase

* test changes and comments

* lint fix

* update config list

* modify testcase for 2 layers

Author: sachinprasadhs

keras_nlp/api/models/__init__.py

@@ -189,6 +189,12 @@
     MistralCausalLMPreprocessor,
 )
 from keras_nlp.src.models.mistral.mistral_tokenizer import MistralTokenizer
+from keras_nlp.src.models.mix_transformer.mix_transformer_backbone import (
+    MiTBackbone,
+)
+from keras_nlp.src.models.mix_transformer.mix_transformer_classifier import (
+    MiTImageClassifier,
+)
 from keras_nlp.src.models.opt.opt_backbone import OPTBackbone
 from keras_nlp.src.models.opt.opt_causal_lm import OPTCausalLM
 from keras_nlp.src.models.opt.opt_causal_lm_preprocessor import (

keras_nlp/src/models/mix_transformer/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2024 The KerasNLP Authors
+#
+# 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
+#
+#     https://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.

keras_nlp/src/models/mix_transformer/mix_transformer_backbone.py

@@ -0,0 +1,181 @@
+# Copyright 2024 The KerasNLP Authors
+#
+# 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
+#
+#     https://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 keras
+import numpy as np
+from keras import ops
+
+from keras_nlp.src.api_export import keras_nlp_export
+from keras_nlp.src.models.feature_pyramid_backbone import FeaturePyramidBackbone
+from keras_nlp.src.models.mix_transformer.mix_transformer_layers import (
+    HierarchicalTransformerEncoder,
+)
+from keras_nlp.src.models.mix_transformer.mix_transformer_layers import (
+    OverlappingPatchingAndEmbedding,
+)
+
+
+@keras_nlp_export("keras_nlp.models.MiTBackbone")
+class MiTBackbone(FeaturePyramidBackbone):
+    def __init__(
+        self,
+        depths,
+        num_layers,
+        blockwise_num_heads,
+        blockwise_sr_ratios,
+        end_value,
+        patch_sizes,
+        strides,
+        include_rescaling=True,
+        image_shape=(224, 224, 3),
+        hidden_dims=None,
+        **kwargs,
+    ):
+        """A Backbone implementing the MixTransformer.
+
+        This architecture to be used as a backbone for the SegFormer
+        architecture [SegFormer: Simple and Efficient Design for Semantic
+        Segmentation with Transformers](https://arxiv.org/abs/2105.15203)
+        [Based on the TensorFlow implementation from DeepVision](
+            https://github.com/DavidLandup0/deepvision/tree/main/deepvision/models/classification/mix_transformer)
+
+        Args:
+            depths: The number of transformer encoders to be used per layer in the
+                network.
+            num_layers: int. The number of Transformer layers.
+            blockwise_num_heads: list of integers, the number of heads to use
+            in the attention computation for each layer.
+            blockwise_sr_ratios: list of integers, the sequence reduction
+                ratio to perform for each layer on the sequence before key and
+                value projections. If set to > 1, a `Conv2D` layer is used to
+                reduce the length of the sequence.
+            end_value: The end value of the sequence.
+            include_rescaling: bool, whether to rescale the inputs. If set
+                to `True`, inputs will be passed through a `Rescaling(1/255.0)`
+                layer. Defaults to `True`.
+            image_shape: optional shape tuple, defaults to (224, 224, 3).
+            hidden_dims: the embedding dims per hierarchical layer, used as
+                the levels of the feature pyramid.
+            patch_sizes: list of integers, the patch_size to apply for each layer.
+            strides: list of integers, stride to apply for each layer.
+
+        Examples:
+
+        Using the class with a `backbone`:
+
+        ```python
+        images = np.ones(shape=(1, 96, 96, 3))
+        labels = np.zeros(shape=(1, 96, 96, 1))
+        backbone = keras_nlp.models.MiTBackbone.from_preset("mit_b0_imagenet")
+
+        # Evaluate model
+        model(images)
+
+        # Train model
+        model.compile(
+            optimizer="adam",
+            loss=keras.losses.BinaryCrossentropy(from_logits=False),
+            metrics=["accuracy"],
+        )
+        model.fit(images, labels, epochs=3)
+        ```
+        """
+        dpr = [x for x in np.linspace(0.0, end_value, sum(depths))]
+
+        # === Layers ===
+        cur = 0
+        patch_embedding_layers = []
+        transformer_blocks = []
+        layer_norms = []
+
+        for i in range(num_layers):
+            patch_embed_layer = OverlappingPatchingAndEmbedding(
+                project_dim=hidden_dims[i],
+                patch_size=patch_sizes[i],
+                stride=strides[i],
+                name=f"patch_and_embed_{i}",
+            )
+            patch_embedding_layers.append(patch_embed_layer)
+
+            transformer_block = [
+                HierarchicalTransformerEncoder(
+                    project_dim=hidden_dims[i],
+                    num_heads=blockwise_num_heads[i],
+                    sr_ratio=blockwise_sr_ratios[i],
+                    drop_prob=dpr[cur + k],
+                    name=f"hierarchical_encoder_{i}_{k}",
+                )
+                for k in range(depths[i])
+            ]
+            transformer_blocks.append(transformer_block)
+            cur += depths[i]
+            layer_norms.append(keras.layers.LayerNormalization())
+
+        # === Functional Model ===
+        image_input = keras.layers.Input(shape=image_shape)
+        x = image_input
+
+        if include_rescaling:
+            x = keras.layers.Rescaling(scale=1 / 255)(x)
+
+        pyramid_outputs = {}
+        for i in range(num_layers):
+            # Compute new height/width after the `proj`
+            # call in `OverlappingPatchingAndEmbedding`
+            stride = strides[i]
+            new_height, new_width = (
+                int(ops.shape(x)[1] / stride),
+                int(ops.shape(x)[2] / stride),
+            )
+
+            x = patch_embedding_layers[i](x)
+            for blk in transformer_blocks[i]:
+                x = blk(x)
+            x = layer_norms[i](x)
+            x = keras.layers.Reshape(
+                (new_height, new_width, -1), name=f"output_level_{i}"
+            )(x)
+            pyramid_outputs[f"P{i + 1}"] = x
+
+        super().__init__(inputs=image_input, outputs=x, **kwargs)
+
+        # === Config ===
+        self.depths = depths
+        self.include_rescaling = include_rescaling
+        self.image_shape = image_shape
+        self.hidden_dims = hidden_dims
+        self.pyramid_outputs = pyramid_outputs
+        self.num_layers = num_layers
+        self.blockwise_num_heads = blockwise_num_heads
+        self.blockwise_sr_ratios = blockwise_sr_ratios
+        self.end_value = end_value
+        self.patch_sizes = patch_sizes
+        self.strides = strides
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "depths": self.depths,
+                "include_rescaling": self.include_rescaling,
+                "hidden_dims": self.hidden_dims,
+                "image_shape": self.image_shape,
+                "num_layers": self.num_layers,
+                "blockwise_num_heads": self.blockwise_num_heads,
+                "blockwise_sr_ratios": self.blockwise_sr_ratios,
+                "end_value": self.end_value,
+                "patch_sizes": self.patch_sizes,
+                "strides": self.strides,
+            }
+        )
+        return config

keras_nlp/src/models/mix_transformer/mix_transformer_backbone_test.py

@@ -0,0 +1,75 @@
+# Copyright 2024 The KerasNLP Authors
+#
+# 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
+#
+#     https://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 numpy as np
+import pytest
+from keras import models
+
+from keras_nlp.src.models.mix_transformer.mix_transformer_backbone import (
+    MiTBackbone,
+)
+from keras_nlp.src.tests.test_case import TestCase
+
+
+class MiTBackboneTest(TestCase):
+    def setUp(self):
+        self.init_kwargs = {
+            "depths": [2, 2],
+            "include_rescaling": True,
+            "image_shape": (16, 16, 3),
+            "hidden_dims": [4, 8],
+            "num_layers": 2,
+            "blockwise_num_heads": [1, 2],
+            "blockwise_sr_ratios": [8, 4],
+            "end_value": 0.1,
+            "patch_sizes": [7, 3],
+            "strides": [4, 2],
+        }
+        self.input_size = 16
+        self.input_data = np.ones(
+            (2, self.input_size, self.input_size, 3), dtype="float32"
+        )
+
+    def test_backbone_basics(self):
+        self.run_backbone_test(
+            cls=MiTBackbone,
+            init_kwargs=self.init_kwargs,
+            input_data=self.input_data,
+            expected_output_shape=(2, 2, 2, 8),
+            run_quantization_check=False,
+            run_mixed_precision_check=False,
+        )
+
+    def test_pyramid_output_format(self):
+        init_kwargs = self.init_kwargs
+        backbone = MiTBackbone(**init_kwargs)
+        model = models.Model(backbone.inputs, backbone.pyramid_outputs)
+        output_data = model(self.input_data)
+
+        self.assertIsInstance(output_data, dict)
+        self.assertEqual(
+            list(output_data.keys()), list(backbone.pyramid_outputs.keys())
+        )
+        self.assertEqual(list(output_data.keys()), ["P1", "P2"])
+        for k, v in output_data.items():
+            size = self.input_size // (2 ** (int(k[1:]) + 1))
+            self.assertEqual(tuple(v.shape[:3]), (2, size, size))
+
+    @pytest.mark.large
+    def test_saved_model(self):
+        self.run_model_saving_test(
+            cls=MiTBackbone,
+            init_kwargs=self.init_kwargs,
+            input_data=self.input_data,
+        )