Cite sources, highlight modifications made for XNNPACK and Core ML exports, add performance metrics

Author: jakmro

examples/models/efficient_sam/README.md

@@ -30,6 +30,20 @@ cd executorch
 python -m examples.xnnpack.aot_compiler -m efficient_sam
 ```
 
+# Performance
+
+Tests were conducted on an Apple M1 Pro chip using the instructions for building and running Executorch with [Core ML](https://pytorch.org/executorch/main/build-run-coreml.html#runtime) and [XNNPACK](https://pytorch.org/executorch/main/tutorial-xnnpack-delegate-lowering.html#running-the-xnnpack-model-with-cmake) backends.
+
+| Backend Configuration  | Average Inference Time (seconds) |
+| ---------------------- | -------------------------------- |
+| Core ML (CPU, GPU, NE) | 34.8                             |
+| Core ML (CPU, GPU)     | 34.7                             |
+| Core ML (CPU, NE)      | 26.4                             |
+| Core ML (CPU)          | 22.8                             |
+| XNNPACK                | 4.1                              |
+
+All models were tested with `float32` precision.
+
 # Licensing
 
 The code in the `efficient_sam_core` directory is licensed under the [Apache License 2.0](./efficient_sam_core/LICENSE.txt).

examples/models/efficient_sam/efficient_sam_core/build_efficient_sam.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the license found in the
 # LICENSE file in the same directory.
 
+# Source: https://github.com/yformer/EfficientSAM/blob/main/efficient_sam/build_efficient_sam.py
+
 from .efficient_sam import build_efficient_sam
 
 

examples/models/efficient_sam/efficient_sam_core/efficient_sam.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the license found in the
 # LICENSE file in the same directory.
 
+# Source: https://github.com/yformer/EfficientSAM/blob/main/efficient_sam/efficient_sam.py
+
 from typing import List, Tuple
 
 import torch

examples/models/efficient_sam/efficient_sam_core/efficient_sam_decoder.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the license found in the
 # LICENSE file in the same directory.
 
+# Source: https://github.com/yformer/EfficientSAM/blob/main/efficient_sam/efficient_sam_decoder.py
+
 from typing import List, Tuple, Type
 
 import numpy as np
@@ -123,6 +125,7 @@ def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
         # outputs d_1 x ... x d_n x C shape
         return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
 
+    # TODO: Remove custom_cumsum implementation once issue #6201 is resolved
     def custom_cumsum(self, tensor: torch.Tensor, dim: int) -> torch.Tensor:
         """Custom cumulative sum."""
         tensor = tensor.transpose(dim, 0)
@@ -140,6 +143,7 @@ def forward(self, size: Tuple[int, int]) -> torch.Tensor:
         h, w = size
         device = self.positional_encoding_gaussian_matrix.device
         grid = torch.ones([h, w], device=device, dtype=torch.float32)
+        # Modification: Use custom_cumsum as a workaround for issue #6201
         y_embed = self.custom_cumsum(grid, dim=0) - 0.5
         x_embed = self.custom_cumsum(grid, dim=1) - 0.5
         y_embed = y_embed / h
@@ -158,6 +162,7 @@ def forward_with_coords(
         return self._pe_encoding(coords.to(torch.float))  # B x N x C
 
 
+# TODO: Remove CustomGroupNorm implementation once issue #6817 is resolved
 class CustomGroupNorm(nn.Module):
     def __init__(self, num_groups, num_channels, eps=1e-5, affine=True):
         """Custom Group Normalization."""
@@ -241,6 +246,7 @@ def __init__(
                         stride=2,
                     ),
                     (
+                        # Modification: Use CustomGroupNorm as a workaround for issue #6817
                         CustomGroupNorm(1, layer_dims)
                         if idx < len(upscaling_layer_dims) - 1
                         else nn.Identity()

examples/models/efficient_sam/efficient_sam_core/efficient_sam_encoder.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the license found in the
 # LICENSE file in the same directory.
 
+# Source: https://github.com/yformer/EfficientSAM/blob/main/efficient_sam/efficient_sam_encoder.py
+
 import math
 from typing import List, Type
 
@@ -165,6 +167,9 @@ def get_abs_pos(
 
     if size != h or size != w:
         new_abs_pos = F.interpolate(
+            # Modification: Change memory format to contiguous
+            # 1. Makes it exportable to ExecuTorch
+            # 2. XNNPACK backend only supports contiguous memory format for inputs
             abs_pos.reshape(1, size, size, -1).permute(0, 3, 1, 2).contiguous(),
             size=(h, w),
             mode="bicubic",

examples/models/efficient_sam/efficient_sam_core/mlp.py

@@ -1,3 +1,5 @@
+# Source: https://github.com/yformer/EfficientSAM/blob/main/efficient_sam/mlp.py
+
 from typing import Type
 
 from torch import nn

examples/models/efficient_sam/efficient_sam_core/two_way_transformer.py

@@ -1,3 +1,5 @@
+# Source: https://github.com/yformer/EfficientSAM/blob/main/efficient_sam/two_way_transformer.py
+
 import math
 from typing import Tuple, Type