Raspbery Pico 2 demo , trying to use quantization to int8 instead and memory utilization doubled from 120KB to 220KB. Why is the memory consumption on the allocator increasing

[landazur]

Hi guys

I am trying to follow along the raspberry pi pico2 demo from last week (https://github.com/pytorch/executorch/tree/main/examples/raspberry_pi/pico2) but with the modification that I want to quantized the model to int8

First below you can see that I first try to get the libquantized_ops_aot_lib.so with the steps provided in the links that are in the README (https://github.com/pytorch/executorch/blob/main/examples/raspberry_pi/pico2/README.md) for quantization

Then I create a new script export_mlp_mnist_int8.py that exports a file called balanced_tiny_mlp_mnist_quantized.pte that I will use to compile with build_firmware_pico.sh. This will be quantized model compiled into a pico2 .uf2 file

Below you are going to see that I can build the original demo and execute on my pico2. However to prove my point I changed the allocator size to 120KB instead of the 200KB that is shown in the git hub main.cpp file (https://github.com/pytorch/executorch/blob/main/examples/raspberry_pi/pico2/main.cpp#L358)

I execute the DEMO with the unquantized model balanced_tiny_mlp_mnist.pte and everything works proving that there is only need for 120KB for allocator size

Then I try to compile the balanced_tiny_mlp_mnist_quantized.pte still with the same allocator size of 120KB and does not load the model because of insufficient
memory error 33 in decimal, 0x21 hex
/// Could not allocate the requested memory.
MemoryAllocationFailed = 0x21,

Then I increase the allocator size to 280KB and recompile the main.cpp and now it works

Why did the memory allocator utilization increased and doubled? I thought with quantization both the ROM size and RAM size will reduce

Below are all my steps. Thanks for helping me understand why the memory usage increase and maybe I am doing some wrong on the quantization steps below

cd pico/executorch

python3 -m venv .pico3 && source .pico3/bin/activate

./install_executorch.sh

examples/arm/setup.sh --i-agree-to-the-contained-eula
source examples/arm/ethos-u-scratch/setup_path.sh

export PATH="~/pico/executorch/examples/arm/ethos-u-scratch/arm-gnu-toolchain-13.3.rel1-x86_64-arm-none-eabi/bin/:$PATH"

TO BUILD the so library for obtaining the quantized model, I run examples/xnnpack/quantization/test_quantize.sh cmake add
This produces cmake-out/kernels/quantized/libquantized_ops_aot_lib.so

Get the .pte file with python examples/raspberry_pi/pico2/export_mlp_mnist.py

Get the .pte file for the quantized model
THE MODEL IS QUANTIZED with this new script I created export_mlp_mnist_int8.py.
It uses torch.ops.load_library("cmake-out/kernels/quantized/libquantized_ops_aot_lib.so")

	import torch
	import torch.nn as nn
	import torch.nn.functional as F
	import time

	from executorch.exir import EdgeCompileConfig, to_edge
	import logging

	from executorch.backends.xnnpack.quantizer.xnnpack_quantizer import (
		get_symmetric_quantization_config,
		XNNPACKQuantizer,
	)

	from torchao.quantization.pt2e.quantize_pt2e import convert_pt2e, prepare_pt2e

	from torch.export import export
	from enum import Enum
	from executorch.extension.export_util.utils import export_to_edge, save_pte_program
	from executorch.exir.capture._config import ExecutorchBackendConfig
	class QuantType(Enum):
		NONE = 1
		# Used for Operations that don't have weights
		STATIC_PER_TENSOR = 2
		# Used best for CNN/RNN Models with Conv layers
		STATIC_PER_CHANNEL = 3
		# Used for Linear Layers and Transformer Based Models
		DYNAMIC_PER_CHANNEL = 4


	# Constants
	INPUT_SIZE = 784  # 28*28 for MNIST
	HIDDEN1_SIZE = 32
	HIDDEN2_SIZE = 16
	OUTPUT_SIZE = 10
	IMAGE_SIZE = 28


	class TinyMLPMNIST(nn.Module):
		"""A small MLP for MNIST digit classification."""

		def __init__(self):
			super().__init__()
			self.fc1 = nn.Linear(INPUT_SIZE, HIDDEN1_SIZE)
			self.fc2 = nn.Linear(HIDDEN1_SIZE, HIDDEN2_SIZE)
			self.fc3 = nn.Linear(HIDDEN2_SIZE, OUTPUT_SIZE)

		def forward(self, x):
			"""Forward pass through the network."""
			x = x.reshape(x.size(0), -1)
			x = F.relu(self.fc1(x))
			x = F.relu(self.fc2(x))
			x = self.fc3(x)
			return x


	def create_balanced_model():
		"""
		Create a balanced MLP model for MNIST digit classification.

		The model is pre-initialized with specific weights to recognize
		digits 0, 1, 4, and 7 through hand-crafted feature detectors.

		Returns:
			torch.nn.Module: A TinyMLPMNIST model with balanced weights
		"""
		model = TinyMLPMNIST()

		with torch.no_grad():
			# Zero everything first
			for param in model.parameters():
				param.fill_(0.0)

			# Feature 0: Vertical lines (for 1, 4, 7)
			for row in range(IMAGE_SIZE):
				# Middle column
				model.fc1.weight[0, row * IMAGE_SIZE + 14] = 2.0

			# Feature 1: Top horizontal (for 7, 4)
			model.fc1.weight[1, 0:84] = 2.0  # Top 3 rows

			# Feature 2: Bottom horizontal (for 1, 4)
			model.fc1.weight[2, 25 * IMAGE_SIZE :] = 2.0  # Bottom 3 rows

			# Feature 3: STRONGER Oval detector for 0
			# Top and bottom curves
			model.fc1.weight[3, 1 * IMAGE_SIZE + 8 : 1 * IMAGE_SIZE + 20] = 2.0
			model.fc1.weight[3, 26 * IMAGE_SIZE + 8 : 26 * IMAGE_SIZE + 20] = 2.0
			# Left and right sides
			for row in range(4, 24):
				model.fc1.weight[3, row * IMAGE_SIZE + 7] = 2.0  # Left
				model.fc1.weight[3, row * IMAGE_SIZE + 20] = 2.0  # Right
			# Anti-middle (hollow center)
			for row in range(10, 18):
				model.fc1.weight[3, row * IMAGE_SIZE + 14] = -1.5

			# Feature 4: Middle horizontal (for 4) - make it STRONGER
			model.fc1.weight[4, 13 * IMAGE_SIZE : 15 * IMAGE_SIZE] = 3.0

			# Second layer: More decisive detection
			# Digit 0 detector: STRONG oval requirement
			model.fc2.weight[0, 3] = 5.0  # Strong oval requirement
			model.fc2.weight[0, 0] = -2.0  # Anti-vertical
			model.fc2.weight[0, 4] = -3.0  # Anti-middle horizontal

			# Digit 1 detector: vertical + bottom - others
			model.fc2.weight[1, 0] = 3.0  # Vertical
			model.fc2.weight[1, 2] = 2.0  # Bottom
			model.fc2.weight[1, 1] = -1.0  # Anti-top
			model.fc2.weight[1, 3] = -2.0  # Anti-oval

			# Digit 4 detector: REQUIRES middle horizontal
			model.fc2.weight[2, 0] = 2.0  # Vertical
			model.fc2.weight[2, 1] = 1.0  # Top
			model.fc2.weight[2, 4] = 4.0  # STRONG middle requirement
			model.fc2.weight[2, 3] = -2.0  # Anti-oval

			# Digit 7 detector: top + some vertical - bottom
			model.fc2.weight[3, 1] = 3.0  # Top
			model.fc2.weight[3, 0] = 1.0  # Some vertical
			model.fc2.weight[3, 2] = -2.0  # Anti-bottom

			# Output layer
			model.fc3.weight[0, 0] = 5.0  # Digit 0
			model.fc3.weight[1, 1] = 5.0  # Digit 1
			model.fc3.weight[4, 2] = 5.0  # Digit 4
			model.fc3.weight[7, 3] = 5.0  # Digit 7

			# Bias against other digits
			for digit in [2, 3, 5, 6, 8, 9]:
				model.fc3.bias[digit] = -3.0

		return model


	def test_comprehensive(model):
		"""
		Test model with clear digit patterns.

		Args:
			model: The PyTorch model to test
		"""

		# Create clearer test patterns
		def create_digit_1():
			"""Create a test pattern for digit 1."""
			digit = torch.zeros(1, IMAGE_SIZE, IMAGE_SIZE)
			# Thick vertical line in middle
			digit[0, 2:26, 13:16] = 1.0  # Thick vertical line
			# Top part (like handwritten 1)
			digit[0, 2:5, 11:14] = 1.0
			# Bottom base
			digit[0, 24:27, 10:19] = 1.0
			return digit

		def create_digit_7():
			"""Create a test pattern for digit 7."""
			digit = torch.zeros(1, IMAGE_SIZE, IMAGE_SIZE)
			# Top horizontal line
			digit[0, 1:4, 3:26] = 1.0
			# Diagonal line
			for i in range(23):
				row = 4 + i
				col = 23 - i
				if 0 <= row < IMAGE_SIZE and 0 <= col < IMAGE_SIZE:
					digit[0, row, col - 1 : col + 2] = 1.0  # Thick diagonal
			return digit

		def create_digit_0():
			"""Create a test pattern for digit 0."""
			digit = torch.zeros(1, IMAGE_SIZE, IMAGE_SIZE)
			# Oval shape
			for row in range(3, 25):
				for col in range(8, 20):
					condition1 = ((row - 14) ** 2 / 11**2 + (col - 14) ** 2 / 6**2) <= 1
					condition2 = ((row - 14) ** 2 / 8**2 + (col - 14) ** 2 / 3**2) > 1
					if condition1 and condition2:
						digit[0, row, col] = 1.0
			return digit

		patterns = {
			"Digit 1": create_digit_1(),
			"Digit 7": create_digit_7(),
			"Digit 0": create_digit_0(),
		}

		print("🧪 Testing with clear patterns:")
		model.eval()
		with torch.no_grad():
			for name, pattern in patterns.items():
				output = model(pattern)
				pred = output.argmax().item()
				confidence = F.softmax(output, dim=1)[0, pred].item()
				print(f"   {name} → predicted: {pred} (confidence: {confidence:.3f})")

				# Show top 3 predictions
				top3 = output.topk(3, dim=1)
				predictions = [
					(top3.indices[0, i].item(), top3.values[0, i].item()) for i in range(3)
				]
				print(f"      Top 3: {predictions}")


	def quantize(
		model, example_inputs, quant_type: QuantType = QuantType.STATIC_PER_TENSOR
	):
		"""This is the official recommended flow for quantization in pytorch 2.0 export"""
		logging.info(f"Original model: {model}")
		quantizer = XNNPACKQuantizer()
		# if we set is_per_channel to True, we also need to add out_variant of quantize_per_channel/dequantize_per_channel
		is_per_channel = (
			quant_type == QuantType.STATIC_PER_CHANNEL
			or quant_type == QuantType.DYNAMIC_PER_CHANNEL
		)
		is_dynamic = quant_type == QuantType.DYNAMIC_PER_CHANNEL
		operator_config = get_symmetric_quantization_config(
			is_per_channel=True,
			is_dynamic=is_dynamic,
		)
		quantizer.set_global(operator_config)
		m = prepare_pt2e(model, quantizer)
		# calibration
		m(*example_inputs)
		m = convert_pt2e(m)
		logging.info(f"Quantized model: {m}")
		# make sure we can export to flat buffer
		return m



	def main():
		"""Main function to create, test, and export the model."""
		print("🔥 Creating balanced MLP MNIST model...")

		model = create_balanced_model()

		# Test the model
		test_comprehensive(model)

		# Export
		example_inputs = torch.randn(1, IMAGE_SIZE, IMAGE_SIZE)
		param_count = sum(p.numel() for p in model.parameters())
		print(f"\n📊 Model parameters: {param_count:,}")

		print("📦 Exporting...")
		# pre-autograd export. eventually this will become torch.export
		with torch.no_grad():
			model = torch.export.export(model, (example_inputs,), strict=True).module()

		print("📦 Quantizing...")
		start = time.perf_counter()
		quantized_model = quantize(model, example_inputs)
		end = time.perf_counter()
		print(f"Quantize time: {end - start}s")

		print("📦 Exporting Quantized...")
		start = time.perf_counter()
		edge_compile_config = EdgeCompileConfig(_check_ir_validity=False)
		edge_m = export_to_edge(
			quantized_model, (example_inputs,), edge_compile_config=edge_compile_config
		)
		end = time.perf_counter()
		print(f"Export time: {end - start}s")

		#torch.ops.load_library("cmake-out/kernels/quantized/libquantized_ops_lib.so")
		torch.ops.load_library("cmake-out/kernels/quantized/libquantized_ops_aot_lib.so")

		start = time.perf_counter()
		prog = edge_m.to_executorch(
			config=ExecutorchBackendConfig(extract_delegate_segments=False)
		)
		save_pte_program(prog, "balanced_tiny_mlp_mnist_quantized")
		end = time.perf_counter()
		print(f"Save time: {end - start}s")
		print("finished")

		#print("📦 Exporting...")
		#with torch.no_grad():
		#    exported_program = export(model, (example_input,))

		#print("⚙️ Converting to ExecuTorch...")
		#edge_config = EdgeCompileConfig(_check_ir_validity=False)
		#edge_manager = to_edge(exported_program, compile_config=edge_config)
		#et_program = edge_manager.to_executorch()

		# Save with error handling
		#filename = "balanced_tiny_mlp_mnist.pte"
		#print(f"💾 Saving {filename}...")
		#try:
		#    with open(filename, "wb") as f:
		#        f.write(et_program.buffer)

		#    model_size_kb = len(et_program.buffer) / 1024
		#    print("✅ Export complete!")
		#    print(f"📁 Model size: {model_size_kb:.1f} KB")
		#except IOError as e:
		#    print(f"❌ Failed to save model: {e}")


	if __name__ == "__main__":
		main()

NOW We have a quantized model balanced_tiny_mlp_mnist_quantized.pte. Look at the size difference
-rw-r--r-- 1 106216 Oct 19 10:33 balanced_tiny_mlp_mnist.pte
-rw-r--r-- 1 33104 Oct 21 20:09 balanced_tiny_mlp_mnist_quantized.pte

COMPILE THE DEMO with balanced_tiny_mlp_mnist.pte and change the memory allocator in the main.cpp to 120KB
static uint8_t method_allocator_pool[120 * 1024]; // it used to be 200KB - plenty for method metadata
static uint8_t activation_pool[120 * 1024]; // it used to be 200KB - plenty for activations

NOW WE START BUILDING THE RASPBERRY PI PICO FIRMWARE from the demo but with 120KB instead of 200KB. Notice it works

  examples/raspberry_pi/pico2/build_firmware_pico.sh --clean	  
  examples/raspberry_pi/pico2/build_firmware_pico.sh --model=balanced_tiny_mlp_mnist.pte

NOW FLASH

        Loading model data (106216 bytes)...
        ✅ Program loaded successfully
         Program info:
           Method count: 1
           Method 0 name: forward
         Loading method 'forward'...
        offset_bytes (100352) + size_bytes (3136) >= allocator size (122880) for memory_id 0
        offset_bytes (100352) + size_bytes (3136) >= allocator size (122880) for memory_id 0
        offset_bytes (0) + size_bytes (100352) >= allocator size (122880) for memory_id 0
        offset_bytes (103488) + size_bytes (128) >= allocator size (122880) for memory_id 0
        offset_bytes (100352) + size_bytes (128) >= allocator size (122880) for memory_id 0
        offset_bytes (0) + size_bytes (2048) >= allocator size (122880) for memory_id 0
        offset_bytes (103488) + size_bytes (64) >= allocator size (122880) for memory_id 0
        offset_bytes (100352) + size_bytes (64) >= allocator size (122880) for memory_id 0
        offset_bytes (0) + size_bytes (640) >= allocator size (122880) for memory_id 0
        offset_bytes (103488) + size_bytes (40) >= allocator size (122880) for memory_id 0
        ✅ Method 'forward' loaded successfully
         ExecuTorch MLP MNIST Demo (Neural network) on Pico2 (microcontroller) 
        離 Testing all supported digits: 
        ......
        Input stats: 159 white pixels out of 784 total
        Running neural network inference...
        Execute Start time
        Execute End
        ✅ Neural network results:
          Digit 0: 370.000
          Digit 1: 0.000
          Digit 2: -3.000
          Digit 3: -3.000
          Digit 4: 860.000
          Digit 5: -3.000
          Digit 6: -3.000
          Digit 7: 1640.000 ← PREDICTED
          Digit 8: -3.000
          Digit 9: -3.000
        
        🎯 PREDICTED: 7 (Expected: 7) ✅ CORRECT!
        
        ==================================================

NOW Build the quantized version with the same 120KB size for allocator
In build_firmware_pico.sh we add
-DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON
-DEXECUTORCH_BUILD_KERNELS_QUANTIZED_AOT=ON \

Then in we add the quantized libs in CMakeList.txt
-Wl,--whole-archive
${BAREMETAL_BUILD_DIR}/lib/libportable_ops_lib.a
${BAREMETAL_BUILD_DIR}/lib/libquantized_ops_lib.a
-Wl,--no-whole-archive
${BAREMETAL_BUILD_DIR}/lib/libportable_kernels.a
${BAREMETAL_BUILD_DIR}/lib/libquantized_kernels.a

COMPILE
examples/raspberry_pi/pico2/build_firmware_pico.sh --model=balanced_tiny_mlp_mnist_quantized.pte
[SERIAL/DIRECT] CONNECTED TO PORT COM15 (115200-8N1)

    Loading model data (33104 bytes)...
    ✅ Program loaded successfully
    📊 Program info:
       Method count: 1
       Method 0 name: forward
    🔄 Loading method 'forward'...
    offset_bytes (0) + size_bytes (3136) >= allocator size (122880) for memory_id 0
    offset_bytes **(100352) + size_bytes (100352) >= allocator size (122880)** for memory_id 0
    ❌ Failed to load method: error 33
       → **Not enough memory t**o load method
    Failed to load and prepare model    

CHANGE main.cpp allocator size to 280KB. and now it WORKS . BUT Why did the memory needs went up so much from an unquantized model to the quantized model

RECOMPILE the quantized model and it runs now
static uint8_t method_allocator_pool[120 * 1024];
static uint8_t activation_pool[280 * 1024];

        Loading model data (33104 bytes)...
        ✅ Program loaded successfully
         Program info:
           Method count: 1
           Method 0 name: forward
         Loading method 'forward'...
        offset_bytes (0) + size_bytes (3136) >= allocator size (286720) for memory_id 0
        offset_bytes (100352) + size_bytes (100352) >= allocator size (286720) for memory_id 0
        offset_bytes (203840) + size_bytes (2048) >= allocator size (286720) for memory_id 0
        offset_bytes (205888) + size_bytes (640) >= allocator size (286720) for memory_id 0
        offset_bytes (0) + size_bytes (3136) >= allocator size (286720) for memory_id 0
        offset_bytes (3136) + size_bytes (784) >= allocator size (286720) for memory_id 0
        offset_bytes (200704) + size_bytes (3136) >= allocator size (286720) for memory_id 0
        offset_bytes (0) + size_bytes (100352) >= allocator size (286720) for memory_id 0
        offset_bytes (100352) + size_bytes (128) >= allocator size (286720) for memory_id 0
        offset_bytes (0) + size_bytes (128) >= allocator size (286720) for memory_id 0
        offset_bytes (200704) + size_bytes (32) >= allocator size (286720) for memory_id 0
        offset_bytes (100352) + size_bytes (128) >= allocator size (286720) for memory_id 0
        offset_bytes (0) + size_bytes (2048) >= allocator size (286720) for memory_id 0
        offset_bytes (200704) + size_bytes (64) >= allocator size (286720) for memory_id 0
        offset_bytes (0) + size_bytes (64) >= allocator size (286720) for memory_id 0
        offset_bytes (200704) + size_bytes (16) >= allocator size (286720) for memory_id 0
        offset_bytes (100352) + size_bytes (64) >= allocator size (286720) for memory_id 0
        offset_bytes (0) + size_bytes (640) >= allocator size (286720) for memory_id 0
        offset_bytes (200704) + size_bytes (40) >= allocator size (286720) for memory_id 0
        offset_bytes (100352) + size_bytes (10) >= allocator size (286720) for memory_id 0
        offset_bytes (0) + size_bytes (40) >= allocator size (286720) for memory_id 0
        ✅ Method 'forward' loaded successfully
         ExecuTorch MLP MNIST Demo (Neural network) on Pico2 (microcontroller) 
        離 Testing all supported digits: 
        
        Input stats: 159 white pixels out of 784 total
        Running neural network inference...
        Execute Start time
        Execute End
        ✅ Neural network results:
          Digit 0: 367.468
          Digit 1: 0.000
          Digit 2: -2.763
          Digit 3: -2.763
          Digit 4: 436.540 ← PREDICTED
          Digit 5: -2.763
          Digit 6: -2.763
          Digit 7: 375.756
          Digit 8: -2.763
          Digit 9: -2.763
        
        🎯 PREDICTED: 4 (Expected: 7) ❌ WRONG!
        
        ==================================================
        
        🎉 All tests complete! ExecuTorch inference of neural network works on Pico2!

[GregoryComer]

@psiddh Do you have any suggestions?

[psiddh]

@landazur Firstly thanks for trying out the Pico2 tutorials and raising this very detailed task explaining your thought process.

If I understand your concern correctly , the following is your main concern

CHANGE main.cpp allocator size to 280KB. and now it WORKS . BUT Why did the memory needs went up so much from an unquantized model to the quantized model

The likely memory increase you're seeing is due to how ExecuTorch implements quantization using decomposed quantize/dequantize (q/dq) operations, which creates additional intermediate tensors during execution. You could also profile it via this tool if you want to.

[landazur]

@psiddh Thank you for your reply. I will try the profile tool and share the results. And yes that is main concern as if I try other models to put on the pico2 this doubling of memory is going to not let me fit other models

But on the quantization step , for raspberry PICO2 , do you believe the choice of XNNPACKQuantizer() and per channel quantization is correct with is_per_channel = True?. There is new documentation out for the 1.0 Executorch release in https://docs.pytorch.org/executorch/stable/backends/xnnpack/xnnpack-overview.html and explains XNNPACK Backend is for

Target Requirements

• ARM64 on Android, iOS, macOS, Linux, and Windows.
• ARMv7 (with NEON) on Android.
• ARMv6 (with VFPv2) on Linux.
• x86 and x86-64 (up to AVX512) on Windows, Linux, Android.

So I am not really sure I am choosing the right quantizer for the prepara_pt2e() step below

        def quantize(
		model, example_inputs, quant_type: QuantType = QuantType.STATIC_PER_TENSOR
	):
		"""This is the official recommended flow for quantization in pytorch 2.0 export"""
		logging.info(f"Original model: {model}")
		quantizer = XNNPACKQuantizer()
		operator_config = get_symmetric_quantization_config(
			is_per_channel=True,
			is_dynamic=is_dynamic,
		)
		quantizer.set_global(operator_config)
		m = prepare_pt2e(model, quantizer)
		# calibration
		m(*example_inputs)
		m = convert_pt2e(m)
		logging.info(f"Quantized model: {m}")
		# make sure we can export to flat buffer
		return m

[digantdesai]

@landazur - glad you managed to get the quantized network work! @psiddh - we should document and repro this!

static uint8_t method_allocator_pool[120 * 1024];
static uint8_t activation_pool[280 * 1024];

I would guess the quantized model may have more intermediate tensors - i.e. you need to bump up from 200KiB to 280KiB as @psiddh already pointed out. Consuming q/dq ops, and running quantized ops without fp32 intermediate tensors is the way to get reduced memory footprint of activation/intermediate tensors for the quantized network, however our op coverage for Cortex-M is just ramping up. If you have a model in mind, and if you can share, or even an op-list, then it would help us prioritize.

XNNPACKQuantizer() and per channel quantization is correct with is_per_channel = True?

per_channel=True implies weights for Linear, Conv2D like ops quantized with finer granularity. It should work with Cortex-M backend and the op should support this -

executorch/backends/cortex_m/ops/op_quantized_linear.cpp

Line 97 in 72b1fa1

if (is_per_channel) {

. Although I did not see any tests at a quick glance. @psiddh - can we add a test for this just to make sure it works.

[landazur]

@digantdesai My original use case was to replicate what I can do with Edge Impulse and Tensor Flow Lite Micro with deploying a vision transfer learning model with MobilenetV2 on an arduino H7 Portenta device to classify 4 types of pictures (4 classed instead of the 1000 classes mv2 supports)

I tried to follow the steps from @psiddh for Pico2 but could not really get a build compiling for Arduino H7 with the arduino-cli + Executorch

In pytorch my original use case is a torchvision mv2 with a width multiplier of 0.1 or even 0.05 so that it fits.
A width of 0.1 fits fine both ROM and RAM on Arduinio H7 portenta (1 MB of RAM) using both edge impulse or TF Lite Micro

Since I was not able to build for Arduino , then I went to build Executorch for pico2 (raspberry pi) and I am not sure why the mv2 model would not fit in RAM , and my quantization was not working

Something like this

diff --git a/examples/models/mobilenet_v2/model.py b/examples/models/mobilenet_v2/model.py
index 32e82197e4..fce33500e0 100644
--- a/examples/models/mobilenet_v2/model.py
+++ b/examples/models/mobilenet_v2/model.py
@@ -20,13 +20,20 @@ class MV2Model(EagerModelBase):
         pass

     def get_eager_model(self) -> torch.nn.Module:
-        logging.info("Loading mobilenet_v2 model")
-        mv2 = mobilenet_v2(weights=MobileNet_V2_Weights.DEFAULT)
+        logging.info("Loading mobilenet_v2 model multiplier=0.1")
+        #mv2 = mobilenet_v2(weights=MobileNet_V2_Weights.DEFAULT)
+        mv2 = mobilenet_v2(weights=None, width_mult=0.05)
+        num_ftrs = mv2.classifier[-1].in_features
+
+        # Create a new linear layer with 4 output features
+        mv2.classifier[-1] = torch.nn.Linear(num_ftrs, 4)
+
+        #print(f"Classifier modified to output {4} classes.")
         logging.info("Loaded mobilenet_v2 model")
         return mv2

     def get_example_inputs(self):
-        tensor_size = (1, 3, 224, 224)
+        tensor_size = (1, 3, 96, 96)
         input_batch = (torch.randn(tensor_size),)
         if self.use_real_input:
             logging.info("Loaded real input image dog.jpg")
@@ -47,7 +54,7 @@ class MV2Model(EagerModelBase):
             preprocess = transforms.Compose(
                 [
                     transforms.Resize(256),
-                    transforms.CenterCrop(224),
+                    transforms.CenterCrop(96),
                     transforms.ToTensor(),
                     transforms.Normalize(
                         mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
@@ -65,9 +72,9 @@ class MV2UntrainedModel(EagerModelBase):
         pass

     def get_eager_model(self) -> torch.nn.Module:
-        mv2 = mobilenet_v2()
+        mv2 = mobilenet_v2(weights=None, width_mult=0.05)
         return mv2

Another use case to explore is Keyword Spotting using BC-ResNet1 or 2 as explored here by TensorFlow Lite Micro

bc_resnet_1 with bc_resnet topology | 96.4% | ~10K param
bc_resnet_2 with bc_resnet topology | 97.6% | 30K param
https://github.com/google-research/google-research/tree/master/kws_streaming

This is much smaller at 10K to 30K and should perform well on the raspberry pi pico2

BCresnet pytorch implementations here from the original paper
https://github.com/Qualcomm-AI-research/bcresnet/blob/main/main.py

When will the op coverage for Cortex-M be ramping up? And that does mean there will be at some point a M-Cortex Quantizer that replaces what I used to today XNNPACKQuantizer()?

Thanks
Javier

[AdrianLundell]

Hi, we have started with this work, currently ADD and LINEAR are implemented, and MUL is on it's way in. As you said, there is Cortex-M quantizer connected to this work as well, you can check it out in https://github.com/pytorch/executorch/tree/main/backends/cortex_m. We are mainly targeting MV3 at the moment.