tensorflow 2.0 基础操作 之 不同 dim 对应的具体含义

Scalar (dim=0)

• []
  • 0, 1, 3.2, …
• $loss=mse(out,y)$
• $accuracy$

out = tf.random.uniform([4,10])

y = tf.range(4)
y = tf.one_hot(y, depth=10)

loss = tf.keras.losses.mse(y, out)
loss = tf.reduce_mean(loss)
loss   # tf.Tensor(0.2689442, shape=(), dtype=float32)

Vector (dim=1)

• Bias
  • [out_put]
    
    

net = layers.Dense(10)
net.build([4, 8])
net.kernel.shape  # matrix
# (8, 10)
net.bias   # vector
# <tf.Variable 'bias:0' shape=(10,) dtype=float32, numpy=array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)>

Matrix (dim=2)

• input x:[b, vec_dim]
• weight: [input_dim, output_dim]
  

x = tf.random.normal([4, 784])
net = layers.Dense(10)   # 784 --> 10
net.build([4, 784])
net(x).shape
# (4, 10)

Tensor (dim=3)

• x: [b, seq_len, word_dim] [句子数目， 句子单词数目， 单词编码长度]
  • [b, 5, 5]
    
    

(X_train, y_train), (X_test, y_test) = keras.datasets.imdb.load_data(num_words=10000)
x_train = keras.preprocessing.sequence.pad_sequences(X_train, maxlen=80)
x_train.shape   # (25000, 80)

emb = embedding(x_train)
emb.shape   # (25000, 80, 100)

out = rnn(emb[:4])
out.shape   # (4, 256)

Tensor (dim=4)

• images: [b, h, w, 3]
• feature maps: [b, h, w, c]

x = tf.random.normal([4,28,28,1])
net = layers.Conv2D(16, kernel_size=3)
net(x).shape   # (4, 26, 26, 16)

Tensor (dim=5)

• Single task: [b, h, w, 3]
• meta-learning:
  • [task_b, b, h, w, 3]