Summary

mergennachin · mergennachin · commit 05a413448064 · 2025-07-11T08:57:53.000-04:00
Added 8 new representative models to ExecutorTorch examples:

EfficientNet-B4: Image classification with CNN architecture
DETR-ResNet50: Object detection using transformer decoder
SegFormer-ADE: Semantic segmentation transformer
Swin2SR: Super-resolution with Swin transformer
ALBERT: Lightweight BERT for NLP tasks
TrOCR: Optical character recognition transformer
Wav2Vec2: Cross-lingual speech representation learning
All models include XNNPACK backend support with appropriate quantization configurations and full CI integration.

Test plan:

Validate model export and execution with portable backend
Test XNNPACK delegation and quantization (with appropriate exclusions)
Integrate into CI workflows for automated testing
Verify all models perform their intended tasks accurately
diff --git a/.ci/scripts/test_model.sh b/.ci/scripts/test_model.sh
@@ -317,8 +317,9 @@ elif [[ "${BACKEND}" == *"xnnpack"* ]]; then
   echo "Testing ${MODEL_NAME} with xnnpack..."
   WITH_QUANTIZATION=true
   WITH_DELEGATION=true
-  if [[ "$MODEL_NAME" == "mobilebert" ]]; then
-    # TODO(T197452682)
+  if [[ "$MODEL_NAME" == "mobilebert" || "$MODEL_NAME" == "albert" ]]; then
+    # TODO(https://github.com/pytorch/executorch/issues/12341) 
+    # mobilebert, albert incompatible with XNNPACK quantization
     WITH_QUANTIZATION=false
   fi
   test_model_with_xnnpack "${WITH_QUANTIZATION}" "${WITH_DELEGATION}"
diff --git a/.github/workflows/trunk.yml b/.github/workflows/trunk.yml
@@ -63,7 +63,7 @@ jobs:
       contents: read
     strategy:
       matrix:
-        model: [linear, add, add_mul, ic3, ic4, mv2, mv3, resnet18, resnet50, vit, w2l, mobilebert, emformer_join, emformer_transcribe]
+        model: [linear, add, add_mul, ic3, ic4, mv2, mv3, resnet18, resnet50, vit, w2l, mobilebert, emformer_join, emformer_transcribe, efficientnet_b4, detr_resnet50, segformer_ade, albert, wav2vec2]
         backend: [portable, xnnpack-quantization-delegation]
         runner: [linux.arm64.2xlarge]
         include:
diff --git a/examples/models/__init__.py b/examples/models/__init__.py
@@ -37,6 +37,14 @@ class Model(str, Enum):
     EfficientSam = "efficient_sam"
     Qwen25 = "qwen2_5"
     Phi4Mini = "phi_4_mini"
+    EfficientNetB4 = "efficientnet_b4"
+    DetrResNet50 = "detr_resnet50"
+    SegformerADE = "segformer_ade"
+    Albert = "albert"
+    BiLSTM = "bilstm"
+    Swin2SR2x = "swin2sr_2x"
+    TrOCRHandwritten = "trocr_handwritten"
+    Wav2Vec2 = "wav2vec2"
 
     def __str__(self) -> str:
         return self.value
@@ -82,6 +90,13 @@ def __str__(self) -> str:
     str(Model.EfficientSam): ("efficient_sam", "EfficientSAM"),
     str(Model.Qwen25): ("qwen2_5", "Qwen2_5Model"),
     str(Model.Phi4Mini): ("phi_4_mini", "Phi4MiniModel"),
+    str(Model.EfficientNetB4): ("efficientnet_b4", "EfficientNetB4Model"),
+    str(Model.DetrResNet50): ("detr_resnet50", "DetrResNet50Model"),
+    str(Model.SegformerADE): ("segformer_ade", "SegformerADEModel"),
+    str(Model.Albert): ("albert", "AlbertModelExample"),
+    str(Model.Swin2SR2x): ("swin2sr_2x", "Swin2SR2xModel"),
+    str(Model.TrOCRHandwritten): ("trocr_handwritten", "TrOCRHandwrittenModel"),
+    str(Model.Wav2Vec2): ("wav2vec2", "Wav2Vec2Model"),
 }
 
 __all__ = [
diff --git a/examples/models/albert/__init__.py b/examples/models/albert/__init__.py
@@ -0,0 +1,11 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .model import AlbertModelExample
+
+__all__ = [
+    "AlbertModelExample",
+]
diff --git a/examples/models/albert/model.py b/examples/models/albert/model.py
@@ -0,0 +1,30 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+import torch
+
+from transformers import AlbertModel, AutoTokenizer  # @manual
+
+from ..model_base import EagerModelBase
+
+
+class AlbertModelExample(EagerModelBase):
+    def __init__(self):
+        pass
+
+    def get_eager_model(self) -> torch.nn.Module:
+        logging.info("Loading ALBERT model")
+        # pyre-ignore
+        model = AlbertModel.from_pretrained("albert-base-v2", return_dict=False)
+        model.eval()
+        logging.info("Loaded ALBERT model")
+        return model
+
+    def get_example_inputs(self):
+        tokenizer = AutoTokenizer.from_pretrained("albert-base-v2")
+        return (tokenizer("Hello, my dog is cute", return_tensors="pt")["input_ids"],)
diff --git a/examples/models/detr_resnet50/__init__.py b/examples/models/detr_resnet50/__init__.py
@@ -0,0 +1,11 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .model import DetrResNet50Model
+
+__all__ = [
+    "DetrResNet50Model",
+]
diff --git a/examples/models/detr_resnet50/model.py b/examples/models/detr_resnet50/model.py
@@ -0,0 +1,45 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+import torch
+from transformers import DetrForObjectDetection
+
+from ..model_base import EagerModelBase
+
+
+class DetrWrapper(torch.nn.Module):
+    """Wrapper for HuggingFace DETR model to make it torch.export compatible"""
+
+    def __init__(self, model_name="facebook/detr-resnet-50"):
+        super().__init__()
+        self.detr = DetrForObjectDetection.from_pretrained(model_name)
+        self.detr.eval()
+
+    def forward(self, pixel_values):
+        # pixel_values: [batch, 3, height, width] - RGB image
+        with torch.no_grad():
+            outputs = self.detr(pixel_values)
+        # Return logits and boxes for object detection
+        return outputs.logits, outputs.pred_boxes
+
+
+class DetrResNet50Model(EagerModelBase):
+    def __init__(self):
+        pass
+
+    def get_eager_model(self) -> torch.nn.Module:
+        logging.info("Loading DETR ResNet-50 model from HuggingFace")
+        model = DetrWrapper("facebook/detr-resnet-50")
+        model.eval()
+        logging.info("Loaded DETR ResNet-50 model")
+        return model
+
+    def get_example_inputs(self):
+        # DETR standard input size: 800x800 RGB image (can handle various sizes)
+        tensor_size = (1, 3, 800, 800)
+        return (torch.randn(tensor_size),)
diff --git a/examples/models/efficientnet_b4/__init__.py b/examples/models/efficientnet_b4/__init__.py
@@ -0,0 +1,11 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .model import EfficientNetB4Model
+
+__all__ = [
+    "EfficientNetB4Model",
+]
diff --git a/examples/models/efficientnet_b4/model.py b/examples/models/efficientnet_b4/model.py
@@ -0,0 +1,31 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+import torch
+
+from torchvision.models import efficientnet_b4  # @manual
+from torchvision.models.efficientnet import EfficientNet_B4_Weights
+
+from ..model_base import EagerModelBase
+
+
+class EfficientNetB4Model(EagerModelBase):
+    def __init__(self):
+        pass
+
+    def get_eager_model(self) -> torch.nn.Module:
+        logging.info("Loading EfficientNet-B4 model")
+        model = efficientnet_b4(weights=EfficientNet_B4_Weights.IMAGENET1K_V1)
+        model.eval()
+        logging.info("Loaded EfficientNet-B4 model")
+        return model
+
+    def get_example_inputs(self):
+        # EfficientNet-B4 uses 380x380 input size
+        tensor_size = (1, 3, 380, 380)
+        return (torch.randn(tensor_size),)
diff --git a/examples/models/segformer_ade/__init__.py b/examples/models/segformer_ade/__init__.py
@@ -0,0 +1,11 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .model import SegformerADEModel
+
+__all__ = [
+    "SegformerADEModel",
+]
diff --git a/examples/models/segformer_ade/model.py b/examples/models/segformer_ade/model.py
@@ -0,0 +1,44 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+import torch
+from transformers import SegformerForSemanticSegmentation
+
+from ..model_base import EagerModelBase
+
+
+class SegformerWrapper(torch.nn.Module):
+    """Wrapper for HuggingFace SegFormer model to make it torch.export compatible"""
+
+    def __init__(self, model_name="nvidia/segformer-b0-finetuned-ade-512-512"):
+        super().__init__()
+        self.segformer = SegformerForSemanticSegmentation.from_pretrained(model_name)
+        self.segformer.eval()
+
+    def forward(self, pixel_values):
+        # pixel_values: [batch, 3, height, width] - RGB image
+        with torch.no_grad():
+            outputs = self.segformer(pixel_values)
+        return outputs.logits
+
+
+class SegformerADEModel(EagerModelBase):
+    def __init__(self):
+        pass
+
+    def get_eager_model(self) -> torch.nn.Module:
+        logging.info("Loading SegFormer ADE model from HuggingFace")
+        model = SegformerWrapper("nvidia/segformer-b0-finetuned-ade-512-512")
+        model.eval()
+        logging.info("Loaded SegFormer ADE model")
+        return model
+
+    def get_example_inputs(self):
+        # SegFormer standard input size: 512x512 RGB image
+        tensor_size = (1, 3, 512, 512)
+        return (torch.randn(tensor_size),)
diff --git a/examples/models/swin2sr_2x/__init__.py b/examples/models/swin2sr_2x/__init__.py
@@ -0,0 +1,11 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .model import Swin2SR2xModel
+
+__all__ = [
+    "Swin2SR2xModel",
+]
diff --git a/examples/models/swin2sr_2x/model.py b/examples/models/swin2sr_2x/model.py
@@ -0,0 +1,44 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+import torch
+from transformers import Swin2SRForImageSuperResolution
+
+from ..model_base import EagerModelBase
+
+
+class Swin2SRWrapper(torch.nn.Module):
+    """Wrapper for HuggingFace Swin2SR model to make it torch.export compatible"""
+
+    def __init__(self, model_name="caidas/swin2SR-classical-sr-x2-64"):
+        super().__init__()
+        self.swin2sr = Swin2SRForImageSuperResolution.from_pretrained(model_name)
+        self.swin2sr.eval()
+
+    def forward(self, pixel_values):
+        # pixel_values: [batch, 3, height, width] - RGB image
+        with torch.no_grad():
+            outputs = self.swin2sr(pixel_values)
+        return outputs.reconstruction
+
+
+class Swin2SR2xModel(EagerModelBase):
+    def __init__(self):
+        pass
+
+    def get_eager_model(self) -> torch.nn.Module:
+        logging.info("Loading Swin2SR 2x model from HuggingFace")
+        model = Swin2SRWrapper("caidas/swin2SR-classical-sr-x2-64")
+        model.eval()
+        logging.info("Loaded Swin2SR 2x model")
+        return model
+
+    def get_example_inputs(self):
+        # Swin2SR input size: 64x64 RGB image for 2x super-resolution
+        tensor_size = (1, 3, 64, 64)
+        return (torch.randn(tensor_size),)
diff --git a/examples/models/trocr_handwritten/__init__.py b/examples/models/trocr_handwritten/__init__.py
@@ -0,0 +1,11 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .model import TrOCRHandwrittenModel
+
+__all__ = [
+    "TrOCRHandwrittenModel",
+]
diff --git a/examples/models/trocr_handwritten/model.py b/examples/models/trocr_handwritten/model.py
@@ -0,0 +1,45 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+import torch
+from transformers import VisionEncoderDecoderModel
+
+from ..model_base import EagerModelBase
+
+
+class TrOCRWrapper(torch.nn.Module):
+    """Wrapper for HuggingFace TrOCR model to make it torch.export compatible"""
+
+    def __init__(self, model_name="microsoft/trocr-base-handwritten"):
+        super().__init__()
+        self.trocr = VisionEncoderDecoderModel.from_pretrained(model_name)
+        self.trocr.eval()
+
+    def forward(self, pixel_values):
+        # pixel_values: [batch, 3, height, width] - RGB image
+        with torch.no_grad():
+            # Generate text from image
+            generated_ids = self.trocr.generate(pixel_values, max_length=50)
+        return generated_ids
+
+
+class TrOCRHandwrittenModel(EagerModelBase):
+    def __init__(self):
+        pass
+
+    def get_eager_model(self) -> torch.nn.Module:
+        logging.info("Loading TrOCR handwritten model from HuggingFace")
+        model = TrOCRWrapper("microsoft/trocr-base-handwritten")
+        model.eval()
+        logging.info("Loaded TrOCR handwritten model")
+        return model
+
+    def get_example_inputs(self):
+        # TrOCR input: 384x384 RGB text image
+        pixel_values = torch.randn(1, 3, 384, 384)
+        return (pixel_values,)
diff --git a/examples/models/wav2vec2/__init__.py b/examples/models/wav2vec2/__init__.py
@@ -0,0 +1,9 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .model import Wav2Vec2Model
+
+__all__ = ["Wav2Vec2Model"]
diff --git a/examples/models/wav2vec2/model.py b/examples/models/wav2vec2/model.py
diff --git a/examples/xnnpack/__init__.py b/examples/xnnpack/__init__.py
