Training settings
Please provide a valid training processor option
Neural network architecture
# Modelo LEVE para ESP32 - Classificação de ECG
# Use este se o modelo principal for muito grande para o ESP32
# Cole no Edge Impulse: Impulse Design -> Learning Block -> Expert Mode
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import (
Dense, InputLayer, Dropout, Conv1D, Flatten,
MaxPooling1D, BatchNormalization, ReLU, Reshape
)
from tensorflow.keras.optimizers.legacy import Adam
# ============================================================================
# HIPERPARÂMETROS
# ============================================================================
EPOCHS = args.epochs or 40
LEARNING_RATE = args.learning_rate or 0.001
BATCH_SIZE = args.batch_size or 64
ENSURE_DETERMINISM = args.ensure_determinism
# ============================================================================
# PREPARAÇÃO DOS DADOS
# ============================================================================
if not ENSURE_DETERMINISM:
train_dataset = train_dataset.shuffle(buffer_size=BATCH_SIZE * 4)
train_dataset = train_dataset.batch(BATCH_SIZE, drop_remainder=False)
validation_dataset = validation_dataset.batch(BATCH_SIZE, drop_remainder=False)
# ============================================================================
# ARQUITETURA LITE - Otimizado para ESP32
# ============================================================================
# Menos parâmetros, menor uso de RAM
# Estimativa: ~15-25KB de Flash, ~5-10KB de RAM
model = Sequential()
# Input
model.add(InputLayer(input_shape=(input_length,)))
model.add(Reshape((input_length, 1)))
# Bloco 1: Conv pequena
model.add(Conv1D(16, kernel_size=7, padding='same'))
model.add(BatchNormalization())
model.add(ReLU())
model.add(MaxPooling1D(pool_size=4))
# Bloco 2
model.add(Conv1D(32, kernel_size=5, padding='same'))
model.add(BatchNormalization())
model.add(ReLU())
model.add(MaxPooling1D(pool_size=4))
# Bloco 3
model.add(Conv1D(32, kernel_size=3, padding='same'))
model.add(ReLU())
model.add(MaxPooling1D(pool_size=2))
# Classificador compacto
model.add(Flatten())
model.add(Dense(32, activation='relu'))
model.add(Dropout(0.3))
model.add(Dense(classes, name='y_pred', activation='softmax'))
# ============================================================================
# COMPILAÇÃO E TREINAMENTO
# ============================================================================
opt = Adam(learning_rate=LEARNING_RATE)
callbacks.append(BatchLoggerCallback(
BATCH_SIZE,
train_sample_count,
epochs=EPOCHS,
ensure_determinism=ENSURE_DETERMINISM
))
callbacks.append(tf.keras.callbacks.EarlyStopping(
monitor='val_accuracy',
patience=8,
restore_best_weights=True
))
model.compile(
loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy']
)
model.fit(
train_dataset,
epochs=EPOCHS,
validation_data=validation_dataset,
verbose=2,
callbacks=callbacks
)
disable_per_channel_quantization = False
Input layer (134 features)
Select a scoring function
Dense layer (20 neurons)
Dense layer (10 neurons)
Output layer (5 classes)
Model
Model version:
