Keras 密集层在 kaggle 上显示太多参数

Question

def my_model():
    inputs = keras.Input(shape=(height,width,3))
    x = layers.Conv2D(32,3)(inputs)
    x = layers.BatchNormalization(input_shape=(32,32,3))(x)
    x = keras.activations.tanh(x)
    x = layers.MaxPooling2D(pool_size=(2,2))(x)
    x = layers.Conv2D(filters=64,kernel_size=(5,5),padding="valid")(x)
    x = layers.BatchNormalization()(x)
    x = keras.activations.tanh(x)
    x = layers.Conv2D(filters=128,kernel_size=(3,3),padding="valid")(x)
    x = layers.BatchNormalization()(x)
    x = layers.Flatten()(x)
    x = layers.Dense(64,activation="tanh")(x)
    outputs = layers.Dense(35)(x)
    
    model = keras.Model(inputs=inputs, outputs = x)
    return model

model = my_model()
model.compile(
    loss='categorical_crossentropy',
    optimizer=keras.optimizers.Adam(lr=3e-4),
    metrics=["accuracy"]
)
model.summary()

this results in output:

Model: "functional_11"

---

Layer (type) Output Shape Param #

input_7 (InputLayer) [(None, 200, 200, 3)] 0

---

conv2d_18 (Conv2D) (None, 198, 198, 32) 896

---

batch_normalization_18 (Batc (None, 198, 198, 32) 128

---

tf_op_layer_Tanh_12 (TensorF [(None, 198, 198, 32)] 0

---

max_pooling2d_6 (MaxPooling2 (None, 99, 99, 32) 0

---

conv2d_19 (Conv2D) (None, 95, 95, 64) 51264

---

batch_normalization_19 (Batc (None, 95, 95, 64) 256

---

tf_op_layer_Tanh_13 (TensorF [(None, 95, 95, 64)] 0

---

conv2d_20 (Conv2D) (None, 93, 93, 128) 73856

---

batch_normalization_20 (Batc (None, 93, 93, 128) 512

---

flatten_6 (Flatten) (None, 1107072) 0

---

dense_6 (Dense) (None, 64) 70852672

---

dense_7 (Dense) (None, 35) 2275

Total params: 70,981,859 Trainable params: 70,981,411 Non-trainable params: 448

---

Here, dense_6 layer with 64 units is showing a huge number of parameters. I'm using kaggle, can somebody help me out by pointing out my mistake?

Answer 1

It is normal, you have a 1,107,072 neurons fully connected to 64 neurons. So the number of parameters is equal to:

• input_length * output_length + output_length = 1,107,072 * 64 + 64 = 70,852,672.

If it's too much for your problem, you should reduce the size before the flatten() layer. For instance, use more pooling layers or calculate less filters.