Training settings
Please provide a valid training processor option
Audio training options
Neural network architecture
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, InputLayer, Dropout, Conv1D, Conv2D, Flatten, Reshape, MaxPooling1D, MaxPooling2D, AveragePooling2D, BatchNormalization, Permute, ReLU, Softmax
from tensorflow.keras.optimizers.legacy import Adam
EPOCHS = args.epochs or 5
LEARNING_RATE = args.learning_rate or 0.0001
# If True, non-deterministic functions (e.g. shuffling batches) are not used.
# This is False by default.
ENSURE_DETERMINISM = args.ensure_determinism
# this controls the batch size, or you can manipulate the tf.data.Dataset objects yourself
BATCH_SIZE = args.batch_size or 32
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)
# model architecture
model = Sequential()
model.add(Conv2D(16, kernel_size=3, strides=2, kernel_constraint=tf.keras.constraints.MaxNorm(1), padding='same'))
model.add(ReLU())
model.add(Conv2D(16, kernel_size=3, strides=2, kernel_constraint=tf.keras.constraints.MaxNorm(1), padding='same'))
model.add(ReLU())
model.add(Conv2D(16, kernel_size=3, strides=2, kernel_constraint=tf.keras.constraints.MaxNorm(1), padding='same'))
model.add(ReLU())
model.add(Dropout(0.4))
model.add(Conv2D(8, kernel_size=3, strides=2, kernel_constraint=tf.keras.constraints.MaxNorm(1), padding='same'))
model.add(ReLU())
model.add(Dropout(0.2))
model.add(Conv2D(16, kernel_size=3, strides=2, kernel_constraint=tf.keras.constraints.MaxNorm(1), padding='same'))
model.add(ReLU())
model.add(Dropout(0.1))
model.add(Flatten())
model.add(Dense(classes, name='y_pred'))
model.add(Softmax())
# this controls the learning rate
opt = Adam(learning_rate=LEARNING_RATE, beta_1=0.9, beta_2=0.999)
callbacks.append(BatchLoggerCallback(BATCH_SIZE, train_sample_count, epochs=EPOCHS, ensure_determinism=ENSURE_DETERMINISM))
# train the neural network
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
model.fit(train_dataset, epochs=EPOCHS, validation_data=validation_dataset, verbose=2, callbacks=callbacks)
import tensorflow as tf
def akida_quantize_model(
keras_model,
weight_quantization: int = 4,
activ_quantization: int = 4,
input_weight_quantization: int = 8,
):
import cnn2snn
print("Performing post-training quantization...")
akida_model = cnn2snn.quantize(
keras_model,
weight_quantization=weight_quantization,
activ_quantization=activ_quantization,
input_weight_quantization=input_weight_quantization,
)
print("Performing post-training quantization OK")
print("")
return akida_model
def akida_perform_qat(
akida_model,
train_dataset: tf.data.Dataset,
validation_dataset: tf.data.Dataset,
optimizer: str,
fine_tune_loss: str,
fine_tune_metrics: "list[str]",
callbacks,
stopping_metric: str = "val_accuracy",
fit_verbose: int = 2,
qat_epochs: int = 30,
):
early_stopping = tf.keras.callbacks.EarlyStopping(
monitor=stopping_metric,
mode="max",
verbose=1,
min_delta=0,
patience=10,
restore_best_weights=True,
)
callbacks.append(early_stopping)
print("Running quantization-aware training...")
akida_model.compile(
optimizer=optimizer, loss=fine_tune_loss, metrics=fine_tune_metrics
)
akida_model.fit(
train_dataset,
epochs=qat_epochs,
verbose=fit_verbose,
validation_data=validation_dataset,
callbacks=callbacks,
)
print("Running quantization-aware training OK")
print("")
return akida_model
akida_model = akida_quantize_model(model)
akida_model = akida_perform_qat(
akida_model,
train_dataset=train_dataset,
validation_dataset=validation_dataset,
optimizer=opt,
fine_tune_loss='categorical_crossentropy',
fine_tune_metrics=['accuracy'],
callbacks=callbacks)
Input layer (12,935 features)
2D conv / pool layer (16 filters, 3 kernel size, 3 layers)
Dropout (rate 0.4)
2D conv / pool layer (8 filters, 3 kernel size, 1 layer)
Dropout (rate 0.2)
2D conv / pool layer (16 filters, 3 kernel size, 1 layer)
Dropout (rate 0.1)
Flatten layer
Output layer (2 classes)
Model
Model version: