使用 Transformer 模型进行时间序列分类
作者： Theodoros Ntakouris
创建时间 2021/06/25
最后修改时间 2021/08/05
描述：此笔记本演示了如何使用 Transformer 模型进行时间序列分类。
简介
这是来自 Attention Is All You Need 的 Transformer 架构，应用于时间序列而不是自然语言。
此示例需要 TensorFlow 2.4 或更高版本。
加载数据集
我们将使用与 从零开始的时间序列分类 示例相同的数据集和预处理。
import numpy as np
import keras
from keras import layers
def readucr(filename):
data = np.loadtxt(filename, delimiter="\t")
y = data[:, 0]
x = data[:, 1:]
return x, y.astype(int)
root_url = "https://raw.githubusercontent.com/hfawaz/cd-diagram/master/FordA/"
x_train, y_train = readucr(root_url + "FordA_TRAIN.tsv")
x_test, y_test = readucr(root_url + "FordA_TEST.tsv")
x_train = x_train.reshape((x_train.shape[0], x_train.shape[1], 1))
x_test = x_test.reshape((x_test.shape[0], x_test.shape[1], 1))
n_classes = len(np.unique(y_train))
idx = np.random.permutation(len(x_train))
x_train = x_train[idx]
y_train = y_train[idx]
y_train[y_train == -1] = 0
y_test[y_test == -1] = 0
构建模型
我们的模型处理形状为 (批次大小，序列长度，特征) 的张量，其中 序列长度 是时间步长数，特征 是每个输入时间序列。
您可以用此模型替换您的分类 RNN 层：输入是完全兼容的！
我们包含残差连接、层归一化和 dropout。生成的层可以堆叠多次。
投影层是通过 keras.layers.Conv1D 实现的。
def transformer_encoder(inputs, head_size, num_heads, ff_dim, dropout=0):
# Attention and Normalization
x = layers.MultiHeadAttention(
key_dim=head_size, num_heads=num_heads, dropout=dropout
)(inputs, inputs)
x = layers.Dropout(dropout)(x)
x = layers.LayerNormalization(epsilon=1e-6)(x)
res = x + inputs
# Feed Forward Part
x = layers.Conv1D(filters=ff_dim, kernel_size=1, activation="relu")(res)
x = layers.Dropout(dropout)(x)
x = layers.Conv1D(filters=inputs.shape[-1], kernel_size=1)(x)
x = layers.LayerNormalization(epsilon=1e-6)(x)
return x + res
我们模型的主要部分现在已经完成。我们可以堆叠多个 transformer_encoder 块，我们还可以继续添加最终的多层感知器分类头。除了 Dense 层的堆叠之外，我们还需要将 TransformerEncoder 模型部分的输出张量缩减为当前批次中每个数据点的特征向量。实现此目标的一种常见方法是使用池化层。对于此示例，GlobalAveragePooling1D 层就足够了。
def build_model(
input_shape,
head_size,
num_heads,
ff_dim,
num_transformer_blocks,
mlp_units,
dropout=0,
mlp_dropout=0,
):
inputs = keras.Input(shape=input_shape)
x = inputs
for _ in range(num_transformer_blocks):
x = transformer_encoder(x, head_size, num_heads, ff_dim, dropout)
x = layers.GlobalAveragePooling1D(data_format="channels_last")(x)
for dim in mlp_units:
x = layers.Dense(dim, activation="relu")(x)
x = layers.Dropout(mlp_dropout)(x)
outputs = layers.Dense(n_classes, activation="softmax")(x)
return keras.Model(inputs, outputs)
训练和评估
input_shape = x_train.shape[1:]
model = build_model(
input_shape,
head_size=256,
num_heads=4,
ff_dim=4,
num_transformer_blocks=4,
mlp_units=[128],
mlp_dropout=0.4,
dropout=0.25,
)
model.compile(
loss="sparse_categorical_crossentropy",
optimizer=keras.optimizers.Adam(learning_rate=1e-4),
metrics=["sparse_categorical_accuracy"],
)
model.summary()
callbacks = [keras.callbacks.EarlyStopping(patience=10, restore_best_weights=True)]
model.fit(
x_train,
y_train,
validation_split=0.2,
epochs=150,
batch_size=64,
callbacks=callbacks,
)
model.evaluate(x_test, y_test, verbose=1)
Model: "functional_1"
┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┓ ┃ Layer (type) ┃ Output Shape ┃ Param # ┃ Connected to ┃ ┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━┩ │ input_layer │ (None, 500, 1) │ 0 │ - │ │ (InputLayer) │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ multi_head_attenti… │ (None, 500, 1) │ 7,169 │ input_layer[0][0], │ │ (MultiHeadAttentio… │ │ │ input_layer[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dropout_1 (Dropout) │ (None, 500, 1) │ 0 │ multi_head_attentio… │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ layer_normalization │ (None, 500, 1) │ 2 │ dropout_1[0][0] │ │ (LayerNormalizatio… │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ add (Add) │ (None, 500, 1) │ 0 │ layer_normalization… │ │ │ │ │ input_layer[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ conv1d (Conv1D) │ (None, 500, 4) │ 8 │ add[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dropout_2 (Dropout) │ (None, 500, 4) │ 0 │ conv1d[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ conv1d_1 (Conv1D) │ (None, 500, 1) │ 5 │ dropout_2[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ layer_normalizatio… │ (None, 500, 1) │ 2 │ conv1d_1[0][0] │ │ (LayerNormalizatio… │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ add_1 (Add) │ (None, 500, 1) │ 0 │ layer_normalization… │ │ │ │ │ add[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ multi_head_attenti… │ (None, 500, 1) │ 7,169 │ add_1[0][0], │ │ (MultiHeadAttentio… │ │ │ add_1[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dropout_4 (Dropout) │ (None, 500, 1) │ 0 │ multi_head_attentio… │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ layer_normalizatio… │ (None, 500, 1) │ 2 │ dropout_4[0][0] │ │ (LayerNormalizatio… │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ add_2 (Add) │ (None, 500, 1) │ 0 │ layer_normalization… │ │ │ │ │ add_1[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ conv1d_2 (Conv1D) │ (None, 500, 4) │ 8 │ add_2[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dropout_5 (Dropout) │ (None, 500, 4) │ 0 │ conv1d_2[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ conv1d_3 (Conv1D) │ (None, 500, 1) │ 5 │ dropout_5[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ layer_normalizatio… │ (None, 500, 1) │ 2 │ conv1d_3[0][0] │ │ (LayerNormalizatio… │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ add_3 (Add) │ (None, 500, 1) │ 0 │ layer_normalization… │ │ │ │ │ add_2[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ multi_head_attenti… │ (None, 500, 1) │ 7,169 │ add_3[0][0], │ │ (MultiHeadAttentio… │ │ │ add_3[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dropout_7 (Dropout) │ (None, 500, 1) │ 0 │ multi_head_attentio… │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ layer_normalizatio… │ (None, 500, 1) │ 2 │ dropout_7[0][0] │ │ (LayerNormalizatio… │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ add_4 (Add) │ (None, 500, 1) │ 0 │ layer_normalization… │ │ │ │ │ add_3[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ conv1d_4 (Conv1D) │ (None, 500, 4) │ 8 │ add_4[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dropout_8 (Dropout) │ (None, 500, 4) │ 0 │ conv1d_4[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ conv1d_5 (Conv1D) │ (None, 500, 1) │ 5 │ dropout_8[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ layer_normalizatio… │ (None, 500, 1) │ 2 │ conv1d_5[0][0] │ │ (LayerNormalizatio… │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ add_5 (Add) │ (None, 500, 1) │ 0 │ layer_normalization… │ │ │ │ │ add_4[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ multi_head_attenti… │ (None, 500, 1) │ 7,169 │ add_5[0][0], │ │ (MultiHeadAttentio… │ │ │ add_5[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dropout_10 │ (None, 500, 1) │ 0 │ multi_head_attentio… │ │ (Dropout) │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ layer_normalizatio… │ (None, 500, 1) │ 2 │ dropout_10[0][0] │ │ (LayerNormalizatio… │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ add_6 (Add) │ (None, 500, 1) │ 0 │ layer_normalization… │ │ │ │ │ add_5[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ conv1d_6 (Conv1D) │ (None, 500, 4) │ 8 │ add_6[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dropout_11 │ (None, 500, 4) │ 0 │ conv1d_6[0][0] │ │ (Dropout) │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ conv1d_7 (Conv1D) │ (None, 500, 1) │ 5 │ dropout_11[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ layer_normalizatio… │ (None, 500, 1) │ 2 │ conv1d_7[0][0] │ │ (LayerNormalizatio… │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ add_7 (Add) │ (None, 500, 1) │ 0 │ layer_normalization… │ │ │ │ │ add_6[0][0] │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ global_average_poo… │ (None, 500) │ 0 │ add_7[0][0] │ │ (GlobalAveragePool… │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dense (Dense) │ (None, 128) │ 64,128 │ global_average_pool… │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dropout_12 │ (None, 128) │ 0 │ dense[0][0] │ │ (Dropout) │ │ │ │ ├─────────────────────┼───────────────────┼─────────┼──────────────────────┤ │ dense_1 (Dense) │ (None, 2) │ 258 │ dropout_12[0][0] │ └─────────────────────┴───────────────────┴─────────┴──────────────────────┘
Total params: 93,130 (363.79 KB)
Trainable params: 93,130 (363.79 KB)
Non-trainable params: 0 (0.00 B)
Epoch 1/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 17s 183ms/step - loss: 1.0039 - sparse_categorical_accuracy: 0.5180 - val_loss: 0.7024 - val_sparse_categorical_accuracy: 0.5908
Epoch 2/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.8639 - sparse_categorical_accuracy: 0.5625 - val_loss: 0.6370 - val_sparse_categorical_accuracy: 0.6241
Epoch 3/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.7701 - sparse_categorical_accuracy: 0.6118 - val_loss: 0.6042 - val_sparse_categorical_accuracy: 0.6602
Epoch 4/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.7522 - sparse_categorical_accuracy: 0.6167 - val_loss: 0.5794 - val_sparse_categorical_accuracy: 0.6782
Epoch 5/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.6845 - sparse_categorical_accuracy: 0.6606 - val_loss: 0.5609 - val_sparse_categorical_accuracy: 0.6893
Epoch 6/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.6760 - sparse_categorical_accuracy: 0.6653 - val_loss: 0.5520 - val_sparse_categorical_accuracy: 0.7046
Epoch 7/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.6589 - sparse_categorical_accuracy: 0.6558 - val_loss: 0.5390 - val_sparse_categorical_accuracy: 0.7129
Epoch 8/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.6416 - sparse_categorical_accuracy: 0.6675 - val_loss: 0.5299 - val_sparse_categorical_accuracy: 0.7171
Epoch 9/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.6270 - sparse_categorical_accuracy: 0.6861 - val_loss: 0.5202 - val_sparse_categorical_accuracy: 0.7295
Epoch 10/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.5995 - sparse_categorical_accuracy: 0.6969 - val_loss: 0.5135 - val_sparse_categorical_accuracy: 0.7323
Epoch 11/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.5846 - sparse_categorical_accuracy: 0.6927 - val_loss: 0.5084 - val_sparse_categorical_accuracy: 0.7420
Epoch 12/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.5837 - sparse_categorical_accuracy: 0.7163 - val_loss: 0.5042 - val_sparse_categorical_accuracy: 0.7420
Epoch 13/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.5407 - sparse_categorical_accuracy: 0.7323 - val_loss: 0.4984 - val_sparse_categorical_accuracy: 0.7462
Epoch 14/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.5302 - sparse_categorical_accuracy: 0.7446 - val_loss: 0.4958 - val_sparse_categorical_accuracy: 0.7462
Epoch 15/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.5041 - sparse_categorical_accuracy: 0.7459 - val_loss: 0.4905 - val_sparse_categorical_accuracy: 0.7503
Epoch 16/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.5122 - sparse_categorical_accuracy: 0.7506 - val_loss: 0.4842 - val_sparse_categorical_accuracy: 0.7642
Epoch 17/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.5042 - sparse_categorical_accuracy: 0.7565 - val_loss: 0.4824 - val_sparse_categorical_accuracy: 0.7656
Epoch 18/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.4965 - sparse_categorical_accuracy: 0.7709 - val_loss: 0.4794 - val_sparse_categorical_accuracy: 0.7587
Epoch 19/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.4860 - sparse_categorical_accuracy: 0.7649 - val_loss: 0.4733 - val_sparse_categorical_accuracy: 0.7614
Epoch 20/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.4797 - sparse_categorical_accuracy: 0.7716 - val_loss: 0.4700 - val_sparse_categorical_accuracy: 0.7642
Epoch 21/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.4946 - sparse_categorical_accuracy: 0.7638 - val_loss: 0.4668 - val_sparse_categorical_accuracy: 0.7670
Epoch 22/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.4443 - sparse_categorical_accuracy: 0.7949 - val_loss: 0.4640 - val_sparse_categorical_accuracy: 0.7670
Epoch 23/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.4495 - sparse_categorical_accuracy: 0.7897 - val_loss: 0.4597 - val_sparse_categorical_accuracy: 0.7739
Epoch 24/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.4284 - sparse_categorical_accuracy: 0.8085 - val_loss: 0.4572 - val_sparse_categorical_accuracy: 0.7739
Epoch 25/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.4353 - sparse_categorical_accuracy: 0.8060 - val_loss: 0.4548 - val_sparse_categorical_accuracy: 0.7795
Epoch 26/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.4332 - sparse_categorical_accuracy: 0.8024 - val_loss: 0.4531 - val_sparse_categorical_accuracy: 0.7781
Epoch 27/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.4399 - sparse_categorical_accuracy: 0.7992 - val_loss: 0.4462 - val_sparse_categorical_accuracy: 0.7864
Epoch 28/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.4143 - sparse_categorical_accuracy: 0.8098 - val_loss: 0.4433 - val_sparse_categorical_accuracy: 0.7850
Epoch 29/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3950 - sparse_categorical_accuracy: 0.8373 - val_loss: 0.4421 - val_sparse_categorical_accuracy: 0.7850
Epoch 30/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.4050 - sparse_categorical_accuracy: 0.8186 - val_loss: 0.4392 - val_sparse_categorical_accuracy: 0.7878
Epoch 31/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.4152 - sparse_categorical_accuracy: 0.8162 - val_loss: 0.4361 - val_sparse_categorical_accuracy: 0.7947
Epoch 32/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3870 - sparse_categorical_accuracy: 0.8290 - val_loss: 0.4335 - val_sparse_categorical_accuracy: 0.7961
Epoch 33/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3966 - sparse_categorical_accuracy: 0.8239 - val_loss: 0.4295 - val_sparse_categorical_accuracy: 0.7961
Epoch 34/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3797 - sparse_categorical_accuracy: 0.8320 - val_loss: 0.4252 - val_sparse_categorical_accuracy: 0.8031
Epoch 35/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3798 - sparse_categorical_accuracy: 0.8336 - val_loss: 0.4222 - val_sparse_categorical_accuracy: 0.8003
Epoch 36/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3652 - sparse_categorical_accuracy: 0.8437 - val_loss: 0.4217 - val_sparse_categorical_accuracy: 0.8044
Epoch 37/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3590 - sparse_categorical_accuracy: 0.8394 - val_loss: 0.4203 - val_sparse_categorical_accuracy: 0.8072
Epoch 38/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3457 - sparse_categorical_accuracy: 0.8562 - val_loss: 0.4182 - val_sparse_categorical_accuracy: 0.8100
Epoch 39/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3668 - sparse_categorical_accuracy: 0.8379 - val_loss: 0.4147 - val_sparse_categorical_accuracy: 0.8072
Epoch 40/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3382 - sparse_categorical_accuracy: 0.8612 - val_loss: 0.4116 - val_sparse_categorical_accuracy: 0.8128
Epoch 41/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.3454 - sparse_categorical_accuracy: 0.8525 - val_loss: 0.4076 - val_sparse_categorical_accuracy: 0.8155
Epoch 42/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.3359 - sparse_categorical_accuracy: 0.8672 - val_loss: 0.4075 - val_sparse_categorical_accuracy: 0.8100
Epoch 43/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3420 - sparse_categorical_accuracy: 0.8538 - val_loss: 0.4033 - val_sparse_categorical_accuracy: 0.8197
Epoch 44/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3325 - sparse_categorical_accuracy: 0.8642 - val_loss: 0.4010 - val_sparse_categorical_accuracy: 0.8197
Epoch 45/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3201 - sparse_categorical_accuracy: 0.8715 - val_loss: 0.3993 - val_sparse_categorical_accuracy: 0.8211
Epoch 46/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.3342 - sparse_categorical_accuracy: 0.8597 - val_loss: 0.3966 - val_sparse_categorical_accuracy: 0.8294
Epoch 47/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.3171 - sparse_categorical_accuracy: 0.8714 - val_loss: 0.3955 - val_sparse_categorical_accuracy: 0.8280
Epoch 48/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.3213 - sparse_categorical_accuracy: 0.8698 - val_loss: 0.3919 - val_sparse_categorical_accuracy: 0.8294
Epoch 49/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.3063 - sparse_categorical_accuracy: 0.8822 - val_loss: 0.3907 - val_sparse_categorical_accuracy: 0.8322
Epoch 50/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2966 - sparse_categorical_accuracy: 0.8826 - val_loss: 0.3888 - val_sparse_categorical_accuracy: 0.8322
Epoch 51/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2946 - sparse_categorical_accuracy: 0.8844 - val_loss: 0.3885 - val_sparse_categorical_accuracy: 0.8308
Epoch 52/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.2930 - sparse_categorical_accuracy: 0.8948 - val_loss: 0.3865 - val_sparse_categorical_accuracy: 0.8322
Epoch 53/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2715 - sparse_categorical_accuracy: 0.9141 - val_loss: 0.3835 - val_sparse_categorical_accuracy: 0.8280
Epoch 54/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2960 - sparse_categorical_accuracy: 0.8848 - val_loss: 0.3806 - val_sparse_categorical_accuracy: 0.8252
Epoch 55/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2813 - sparse_categorical_accuracy: 0.8989 - val_loss: 0.3808 - val_sparse_categorical_accuracy: 0.8239
Epoch 56/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2708 - sparse_categorical_accuracy: 0.9076 - val_loss: 0.3784 - val_sparse_categorical_accuracy: 0.8363
Epoch 57/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2895 - sparse_categorical_accuracy: 0.8882 - val_loss: 0.3786 - val_sparse_categorical_accuracy: 0.8336
Epoch 58/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2905 - sparse_categorical_accuracy: 0.8810 - val_loss: 0.3780 - val_sparse_categorical_accuracy: 0.8363
Epoch 59/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2732 - sparse_categorical_accuracy: 0.9023 - val_loss: 0.3738 - val_sparse_categorical_accuracy: 0.8419
Epoch 60/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2698 - sparse_categorical_accuracy: 0.8962 - val_loss: 0.3733 - val_sparse_categorical_accuracy: 0.8308
Epoch 61/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.2741 - sparse_categorical_accuracy: 0.9025 - val_loss: 0.3724 - val_sparse_categorical_accuracy: 0.8391
Epoch 62/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 128ms/step - loss: 0.2713 - sparse_categorical_accuracy: 0.8973 - val_loss: 0.3698 - val_sparse_categorical_accuracy: 0.8308
Epoch 63/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.2682 - sparse_categorical_accuracy: 0.9004 - val_loss: 0.3681 - val_sparse_categorical_accuracy: 0.8363
Epoch 64/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2673 - sparse_categorical_accuracy: 0.9006 - val_loss: 0.3692 - val_sparse_categorical_accuracy: 0.8377
Epoch 65/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2585 - sparse_categorical_accuracy: 0.9056 - val_loss: 0.3684 - val_sparse_categorical_accuracy: 0.8322
Epoch 66/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2696 - sparse_categorical_accuracy: 0.8958 - val_loss: 0.3654 - val_sparse_categorical_accuracy: 0.8336
Epoch 67/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2489 - sparse_categorical_accuracy: 0.9182 - val_loss: 0.3630 - val_sparse_categorical_accuracy: 0.8405
Epoch 68/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2475 - sparse_categorical_accuracy: 0.9121 - val_loss: 0.3626 - val_sparse_categorical_accuracy: 0.8433
Epoch 69/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2398 - sparse_categorical_accuracy: 0.9195 - val_loss: 0.3607 - val_sparse_categorical_accuracy: 0.8433
Epoch 70/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2379 - sparse_categorical_accuracy: 0.9138 - val_loss: 0.3598 - val_sparse_categorical_accuracy: 0.8474
Epoch 71/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2343 - sparse_categorical_accuracy: 0.9162 - val_loss: 0.3568 - val_sparse_categorical_accuracy: 0.8447
Epoch 72/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2497 - sparse_categorical_accuracy: 0.9104 - val_loss: 0.3554 - val_sparse_categorical_accuracy: 0.8419
Epoch 73/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.2399 - sparse_categorical_accuracy: 0.9070 - val_loss: 0.3552 - val_sparse_categorical_accuracy: 0.8433
Epoch 74/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2300 - sparse_categorical_accuracy: 0.9190 - val_loss: 0.3572 - val_sparse_categorical_accuracy: 0.8419
Epoch 75/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2370 - sparse_categorical_accuracy: 0.9109 - val_loss: 0.3523 - val_sparse_categorical_accuracy: 0.8419
Epoch 76/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2324 - sparse_categorical_accuracy: 0.9172 - val_loss: 0.3512 - val_sparse_categorical_accuracy: 0.8391
Epoch 77/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2262 - sparse_categorical_accuracy: 0.9210 - val_loss: 0.3488 - val_sparse_categorical_accuracy: 0.8391
Epoch 78/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2262 - sparse_categorical_accuracy: 0.9175 - val_loss: 0.3495 - val_sparse_categorical_accuracy: 0.8419
Epoch 79/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.2226 - sparse_categorical_accuracy: 0.9270 - val_loss: 0.3487 - val_sparse_categorical_accuracy: 0.8433
Epoch 80/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2181 - sparse_categorical_accuracy: 0.9247 - val_loss: 0.3501 - val_sparse_categorical_accuracy: 0.8474
Epoch 81/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2220 - sparse_categorical_accuracy: 0.9181 - val_loss: 0.3479 - val_sparse_categorical_accuracy: 0.8460
Epoch 82/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2114 - sparse_categorical_accuracy: 0.9254 - val_loss: 0.3464 - val_sparse_categorical_accuracy: 0.8460
Epoch 83/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2148 - sparse_categorical_accuracy: 0.9196 - val_loss: 0.3467 - val_sparse_categorical_accuracy: 0.8460
Epoch 84/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.2262 - sparse_categorical_accuracy: 0.9181 - val_loss: 0.3446 - val_sparse_categorical_accuracy: 0.8474
Epoch 85/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2121 - sparse_categorical_accuracy: 0.9205 - val_loss: 0.3452 - val_sparse_categorical_accuracy: 0.8460
Epoch 86/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.2057 - sparse_categorical_accuracy: 0.9238 - val_loss: 0.3460 - val_sparse_categorical_accuracy: 0.8350
Epoch 87/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2081 - sparse_categorical_accuracy: 0.9342 - val_loss: 0.3455 - val_sparse_categorical_accuracy: 0.8488
Epoch 88/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2153 - sparse_categorical_accuracy: 0.9211 - val_loss: 0.3421 - val_sparse_categorical_accuracy: 0.8488
Epoch 89/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1977 - sparse_categorical_accuracy: 0.9366 - val_loss: 0.3413 - val_sparse_categorical_accuracy: 0.8474
Epoch 90/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1928 - sparse_categorical_accuracy: 0.9410 - val_loss: 0.3428 - val_sparse_categorical_accuracy: 0.8405
Epoch 91/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1968 - sparse_categorical_accuracy: 0.9327 - val_loss: 0.3411 - val_sparse_categorical_accuracy: 0.8474
Epoch 92/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1909 - sparse_categorical_accuracy: 0.9308 - val_loss: 0.3404 - val_sparse_categorical_accuracy: 0.8488
Epoch 93/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2067 - sparse_categorical_accuracy: 0.9285 - val_loss: 0.3371 - val_sparse_categorical_accuracy: 0.8488
Epoch 94/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1990 - sparse_categorical_accuracy: 0.9329 - val_loss: 0.3385 - val_sparse_categorical_accuracy: 0.8502
Epoch 95/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1986 - sparse_categorical_accuracy: 0.9267 - val_loss: 0.3368 - val_sparse_categorical_accuracy: 0.8433
Epoch 96/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.2069 - sparse_categorical_accuracy: 0.9235 - val_loss: 0.3346 - val_sparse_categorical_accuracy: 0.8502
Epoch 97/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1971 - sparse_categorical_accuracy: 0.9296 - val_loss: 0.3340 - val_sparse_categorical_accuracy: 0.8544
Epoch 98/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.2042 - sparse_categorical_accuracy: 0.9250 - val_loss: 0.3352 - val_sparse_categorical_accuracy: 0.8419
Epoch 99/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1998 - sparse_categorical_accuracy: 0.9271 - val_loss: 0.3334 - val_sparse_categorical_accuracy: 0.8474
Epoch 100/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1832 - sparse_categorical_accuracy: 0.9406 - val_loss: 0.3317 - val_sparse_categorical_accuracy: 0.8474
Epoch 101/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1917 - sparse_categorical_accuracy: 0.9340 - val_loss: 0.3343 - val_sparse_categorical_accuracy: 0.8433
Epoch 102/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1811 - sparse_categorical_accuracy: 0.9286 - val_loss: 0.3317 - val_sparse_categorical_accuracy: 0.8530
Epoch 103/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1733 - sparse_categorical_accuracy: 0.9396 - val_loss: 0.3340 - val_sparse_categorical_accuracy: 0.8460
Epoch 104/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1661 - sparse_categorical_accuracy: 0.9464 - val_loss: 0.3288 - val_sparse_categorical_accuracy: 0.8488
Epoch 105/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1806 - sparse_categorical_accuracy: 0.9390 - val_loss: 0.3296 - val_sparse_categorical_accuracy: 0.8516
Epoch 106/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1774 - sparse_categorical_accuracy: 0.9401 - val_loss: 0.3291 - val_sparse_categorical_accuracy: 0.8530
Epoch 107/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1689 - sparse_categorical_accuracy: 0.9463 - val_loss: 0.3290 - val_sparse_categorical_accuracy: 0.8488
Epoch 108/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1830 - sparse_categorical_accuracy: 0.9319 - val_loss: 0.3299 - val_sparse_categorical_accuracy: 0.8447
Epoch 109/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1757 - sparse_categorical_accuracy: 0.9304 - val_loss: 0.3315 - val_sparse_categorical_accuracy: 0.8488
Epoch 110/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1810 - sparse_categorical_accuracy: 0.9378 - val_loss: 0.3280 - val_sparse_categorical_accuracy: 0.8502
Epoch 111/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1628 - sparse_categorical_accuracy: 0.9522 - val_loss: 0.3276 - val_sparse_categorical_accuracy: 0.8474
Epoch 112/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1659 - sparse_categorical_accuracy: 0.9484 - val_loss: 0.3285 - val_sparse_categorical_accuracy: 0.8530
Epoch 113/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1814 - sparse_categorical_accuracy: 0.9364 - val_loss: 0.3281 - val_sparse_categorical_accuracy: 0.8474
Epoch 114/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1721 - sparse_categorical_accuracy: 0.9391 - val_loss: 0.3287 - val_sparse_categorical_accuracy: 0.8433
Epoch 115/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 127ms/step - loss: 0.1743 - sparse_categorical_accuracy: 0.9321 - val_loss: 0.3275 - val_sparse_categorical_accuracy: 0.8474
Epoch 116/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1677 - sparse_categorical_accuracy: 0.9415 - val_loss: 0.3297 - val_sparse_categorical_accuracy: 0.8391
Epoch 117/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1657 - sparse_categorical_accuracy: 0.9449 - val_loss: 0.3228 - val_sparse_categorical_accuracy: 0.8419
Epoch 118/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1787 - sparse_categorical_accuracy: 0.9316 - val_loss: 0.3230 - val_sparse_categorical_accuracy: 0.8447
Epoch 119/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1659 - sparse_categorical_accuracy: 0.9408 - val_loss: 0.3233 - val_sparse_categorical_accuracy: 0.8460
Epoch 120/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1615 - sparse_categorical_accuracy: 0.9385 - val_loss: 0.3235 - val_sparse_categorical_accuracy: 0.8460
Epoch 121/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1582 - sparse_categorical_accuracy: 0.9526 - val_loss: 0.3247 - val_sparse_categorical_accuracy: 0.8474
Epoch 122/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1577 - sparse_categorical_accuracy: 0.9497 - val_loss: 0.3263 - val_sparse_categorical_accuracy: 0.8474
Epoch 123/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1593 - sparse_categorical_accuracy: 0.9483 - val_loss: 0.3261 - val_sparse_categorical_accuracy: 0.8433
Epoch 124/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1570 - sparse_categorical_accuracy: 0.9442 - val_loss: 0.3277 - val_sparse_categorical_accuracy: 0.8419
Epoch 125/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1434 - sparse_categorical_accuracy: 0.9460 - val_loss: 0.3257 - val_sparse_categorical_accuracy: 0.8447
Epoch 126/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1589 - sparse_categorical_accuracy: 0.9414 - val_loss: 0.3237 - val_sparse_categorical_accuracy: 0.8447
Epoch 127/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1591 - sparse_categorical_accuracy: 0.9460 - val_loss: 0.3217 - val_sparse_categorical_accuracy: 0.8447
Epoch 128/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1530 - sparse_categorical_accuracy: 0.9450 - val_loss: 0.3203 - val_sparse_categorical_accuracy: 0.8474
Epoch 129/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1464 - sparse_categorical_accuracy: 0.9514 - val_loss: 0.3206 - val_sparse_categorical_accuracy: 0.8474
Epoch 130/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1437 - sparse_categorical_accuracy: 0.9526 - val_loss: 0.3231 - val_sparse_categorical_accuracy: 0.8447
Epoch 131/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1415 - sparse_categorical_accuracy: 0.9510 - val_loss: 0.3226 - val_sparse_categorical_accuracy: 0.8433
Epoch 132/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1539 - sparse_categorical_accuracy: 0.9505 - val_loss: 0.3261 - val_sparse_categorical_accuracy: 0.8405
Epoch 133/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1432 - sparse_categorical_accuracy: 0.9544 - val_loss: 0.3239 - val_sparse_categorical_accuracy: 0.8377
Epoch 134/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1368 - sparse_categorical_accuracy: 0.9567 - val_loss: 0.3200 - val_sparse_categorical_accuracy: 0.8474
Epoch 135/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1319 - sparse_categorical_accuracy: 0.9619 - val_loss: 0.3200 - val_sparse_categorical_accuracy: 0.8433
Epoch 136/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1479 - sparse_categorical_accuracy: 0.9494 - val_loss: 0.3201 - val_sparse_categorical_accuracy: 0.8502
Epoch 137/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1353 - sparse_categorical_accuracy: 0.9573 - val_loss: 0.3208 - val_sparse_categorical_accuracy: 0.8488
Epoch 138/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1349 - sparse_categorical_accuracy: 0.9584 - val_loss: 0.3213 - val_sparse_categorical_accuracy: 0.8474
Epoch 139/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1418 - sparse_categorical_accuracy: 0.9532 - val_loss: 0.3197 - val_sparse_categorical_accuracy: 0.8447
Epoch 140/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1402 - sparse_categorical_accuracy: 0.9534 - val_loss: 0.3204 - val_sparse_categorical_accuracy: 0.8488
Epoch 141/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1344 - sparse_categorical_accuracy: 0.9525 - val_loss: 0.3207 - val_sparse_categorical_accuracy: 0.8474
Epoch 142/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1448 - sparse_categorical_accuracy: 0.9494 - val_loss: 0.3192 - val_sparse_categorical_accuracy: 0.8488
Epoch 143/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1363 - sparse_categorical_accuracy: 0.9552 - val_loss: 0.3219 - val_sparse_categorical_accuracy: 0.8460
Epoch 144/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1380 - sparse_categorical_accuracy: 0.9540 - val_loss: 0.3219 - val_sparse_categorical_accuracy: 0.8474
Epoch 145/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1472 - sparse_categorical_accuracy: 0.9468 - val_loss: 0.3219 - val_sparse_categorical_accuracy: 0.8474
Epoch 146/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1402 - sparse_categorical_accuracy: 0.9622 - val_loss: 0.3217 - val_sparse_categorical_accuracy: 0.8502
Epoch 147/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1236 - sparse_categorical_accuracy: 0.9617 - val_loss: 0.3194 - val_sparse_categorical_accuracy: 0.8460
Epoch 148/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1183 - sparse_categorical_accuracy: 0.9683 - val_loss: 0.3193 - val_sparse_categorical_accuracy: 0.8488
Epoch 149/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 126ms/step - loss: 0.1189 - sparse_categorical_accuracy: 0.9618 - val_loss: 0.3237 - val_sparse_categorical_accuracy: 0.8488
Epoch 150/150
45/45 ━━━━━━━━━━━━━━━━━━━━ 6s 125ms/step - loss: 0.1495 - sparse_categorical_accuracy: 0.9459 - val_loss: 0.3181 - val_sparse_categorical_accuracy: 0.8460
42/42 ━━━━━━━━━━━━━━━━━━━━ 3s 44ms/step - loss: 0.3182 - sparse_categorical_accuracy: 0.8617
[0.3543623089790344, 0.843181848526001]
结论
在大约 110-120 个 epoch（在 Colab 上每个 epoch 25 秒），模型达到约 0.95 的训练准确率、约 84 的验证准确率和约 85 的测试准确率，无需超参数调整。并且这是对于一个参数少于 10 万的参数模型。当然，可以通过超参数搜索和更复杂的学习率计划或不同的优化器来提高参数数量和准确率。