def dense_ae():
"""
Test mlp on mnist data.
"""
import nets.dense as dense
# load mnist data, flatten, and normalize to 0-1
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train = x_train.reshape((x_train.shape[0], 784))
x_train = x_train / 255.0
# create a batch feed from the train tensors
train_ds = tf.data.Dataset.from_tensor_slices((x_train, y_train)) \
.shuffle(10000) \
.batch(32)
config = {
"encoding_dims": [64, 32],
"latent_dim": 10,
"activation": "relu",
"optimizer": {
"Adam": {
"learning_rate": 0.001
}
},
"loss": {
"MeanAbsoluteError": {}
},
"activity_regularizer": {
"l1_l2": {
"l1": 0.0075,
"l2": 0.0025
}
},
"epochs": 2,
"sparse_flag": False
}
compiled_model = train.model_init(dense.DenseAE(config), config["loss"],
config["optimizer"], (None, 784))
for _ in range(config["epochs"]):
loss = 0.0
step = 0
for x, y in train_ds:
loss_, grads = train.grad(compiled_model, x, x)
updates = zip(grads, compiled_model.trainable_variables)
compiled_model.optimizer.apply_gradients(updates)
loss += loss_
step += 1
print("Epoch loss: {loss}".format(loss=loss / float(step)))
def resnet():
"""
Test mlp on mnist data.
"""
import nets.image as img
# load mnist data, reshape to have channel dim, and normalize to 0-1
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train = x_train.reshape((x_train.shape[0], 28, 28, 1))
x_train = x_train / 255.0
# create a batch feed from the train tensors
train_ds = tf.data.Dataset.from_tensor_slices((x_train, y_train)) \
.shuffle(10000) \
.batch(32)
config = {
"conv_filters": [24],
"res_filters": [8, 24],
"res_depth": 1,
"dense_dims": [24],
"output_dim": 10,
"output_activation": "softmax",
"optimizer": {
"Adam": {
"learning_rate": 0.001
}
},
"loss": {
"SparseCategoricalCrossentropy": {}
},
"epochs": 2
}
compiled_model = train.model_init(img.ResNet(config), config["loss"],
config["optimizer"], (None, 28, 28, 1))
for _ in range(config["epochs"]):
loss = 0.0
step = 0
for x, y in train_ds:
loss_, grads = train.grad(compiled_model, x, y)
updates = zip(grads, compiled_model.trainable_variables)
compiled_model.optimizer.apply_gradients(updates)
loss += loss_
step += 1
print("Epoch loss: {loss}".format(loss=loss / float(step)))