espdlx

Train models in PyTorch and export them to .espdl — the native model format for ESP32 boards.

espdlx is the Python training side: build models from espdlx.layers inside an espdlx.Model, train with your own PyTorch loop, then export via the standard path: torch.onnx.export -> esp-dl quantization (espdl_quantize_onnx -> .espdl).

Deployment is paired with the espdlx Arduino library (esp-dl port for ESP32-S3, installable from the Arduino Library Manager) for maximum ease of use: embed the .espdl file as a C array, load it with dl::Model, call run().

Installation

pip install espdlx
# with esp-dl quantization support:
pip install "espdlx[convert]"

Example

Simple CNN trained on MNIST:

import torch
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torchvision import datasets, transforms

import espdlx
import espdlx.layers as L

device = torch.device("mps")

model = espdlx.Model(
    [
        L.Conv2d(1, 16, 3, padding=1), L.ReLU(),
        L.MaxPool2d(2),
        L.Conv2d(16, 32, 3, padding=1), L.ReLU(),
        L.MaxPool2d(2),
        L.Conv2d(32, 10, 1),
        L.Mean(),
        L.Flatten(),
    ],
    name="mnist_cnn",
).to(device)

train_ds = datasets.MNIST(root="./data", train=True, download=True, transform=transforms.ToTensor())
test_ds = datasets.MNIST(root="./data", train=False, download=True, transform=transforms.ToTensor())
train_ld = DataLoader(train_ds, batch_size=128, shuffle=True)
test_ld = DataLoader(test_ds, batch_size=512)

opt = torch.optim.Adam(model.parameters(), lr=1e-3)
model.train()
for epoch in range(3):
    for x, y in train_ld:
        x, y = x.to(device), y.to(device)
        opt.zero_grad()
        loss = F.cross_entropy(model(x), y)
        loss.backward()
        opt.step()

model.eval()
correct = total = 0
with torch.no_grad():
    for x, y in test_ld:
        correct += (model(x.to(device)).argmax(-1) == y.to(device)).sum().item()
        total += len(y)
print(f"test acc: {correct / total:.3f}")

Export to .espdl

Quantize a trained model to .espdl (needs the convert extra):

from torch.utils.data import DataLoader, TensorDataset
from espdlx.convert import convert

calib_loader = DataLoader(TensorDataset(calib_images), batch_size=8)

report = convert(
    model.cpu(),              # any nn.Module; espdlx.Model or plain torch
    example_input,            # dummy input for torch.onnx.export
    calib_loader,             # calibration batches
    "model.espdl",            # also writes model.onnx + espdl_report.json
    calib_steps=16,
)
print(report["espdl_bytes"], report["input_scale"], report["input_zero_point"])

Deploy with the espdlx Arduino library

Install espdlx from the Arduino Library Manager (ESP32-S3), then embed the exported .espdl file as a C header next to your sketch:

# model_data.h — generated from model.espdl (xxd-style, one line per 16 bytes)
xxd -i model.espdl > model_data.h

Run inference with dl::Model (pattern from the espdlx Arduino library examples):

#include <Arduino.h>
#include <espdlx.h>
#include "model_data.h";  // const unsigned char model_espdl[] = { 0x45, 0x44, ... };

dl::Model *model = nullptr;

void setup() {
  Serial.begin(115200);

  // Load the .espdl model straight from flash.
  model = new dl::Model((const char *)model_espdl,
                        fbs::MODEL_LOCATION_IN_FLASH_RODATA);

  auto inputs = model->get_inputs();
  auto outputs = model->get_outputs();
  dl::TensorBase *in = inputs.begin()->second;
  dl::TensorBase *out = outputs.begin()->second;
  Serial.printf("in_elems=%d out_elems=%d\n", in->size, out->size);

  memcpy(in->data, my_input, in->size);  // int8 NHWC, quantized with input_scale
  model->run();                          // or model->run(dl::RUNTIME_MODE_MULTI_CORE)

  float *logits = (float *)out->data;
  int pred = 0;
  for (int i = 1; i < out->size; i++)
    if (logits[i] > logits[pred]) pred = i;
  Serial.printf("pred=%d\n", pred);
}

void loop() { delay(1000); }

The library also ships a SelfTest example (examples/SelfTest/SelfTest.ino) that exercises core + vision + audio through the single espdlx.h header:

#include <Arduino.h>
#include <espdlx.h>

void setup() {
  Serial.begin(115200);
  float data[4] = {1.0f, 2.0f, 3.0f, 4.0f};
  dl::TensorBase t({4}, data, 0, dl::DATA_TYPE_FLOAT, true);
  dl::math::softmax((float *)t.data, 4);  // core check via espdlx.h
}

void loop() { delay(5000); }

Model zoo

Prebuilt architectures (espdlx/zoo.py, all residual-free variants — no runtime tensor Add on-device, so skip connections are omitted):

from espdlx.zoo import MobileNetV2, MobileNetV1, VGG, DSCNN

model = MobileNetV2.Slim(num_classes=5)  # flowers, 128px RGB
model = MobileNetV2.Base(num_classes=32) # CIFAR-100 pretrain backbone, 32px
model = MobileNetV2.Fomo()               # EXPERIMENTAL centroid detector, 96px

• All device numbers measured 2026-09-12 on ESP32-S3 @ 240 MHz via esp-dl (.espdl w8a8, OPI PSRAM, huge_app, N=20 runs, micros() around model->run()).
• .espdl from espdlx.convert after 1 epoch of synthetic-noise training per model; sizes and latency are weights-independent (accuracy on noise weights is meaningless, so no accuracy column).
• PSRAM ctx = model context + working set (free_psram drop after load). Sketch flash is ~1.1 MB framework floor + model + test vector.
• MobileNetV2.Fomo / espdlx.fomo are experimental: synthetic unit tests only, no on-device accuracy verification yet (latency above just proves it runs). Best real-data run so far: P=0.03/R=0.28 — not a working detector yet.

License

MIT — see LICENSE.