python tf_eager.py

Posted 2021-05-09

X = tfe.Variable(X_train)
Y = tfe.Variable(y_train)

X_t = tfe.Variable(X_test)
Y_t = tfe.Variable(y_test)

def compute_accuracy(preds, label):
    p = tf.argmax(preds, axis=1)
    y = tf.cast(label, tf.int64)
    size = int(p.shape[0])
    return (tf.reduce_sum(tf.cast(tf.equal(p, y), tf.int64)) / size).numpy()

class Model(tf.keras.Model):
    def __init__(self):
        super(Model, self).__init__()
        self.dense1 = tf.layers.Dense(30, activation=tf.nn.relu)
        self.dense2 = tf.layers.Dense(20, activation=tf.nn.relu)
        self.dense3 = tf.layers.Dense(len(np.unique(y_train)))
        self.optimizer = tf.train.AdamOptimizer()
        self.cross_entropy = tf.losses.sparse_softmax_cross_entropy

    def predict(self, input_data):
        out = self.dense1(input_data)
        out = self.dense2(out)
        logits = self.dense3(out)
        return logits

    def forward_pass(self, input_data):
        preds = self.predict(input_data)
        return preds

    def backward_pass(self, loss, tape):
        grad = tape.gradient(loss, self.variables)
        self.optimizer.apply_gradients(zip(grad, self.variables))

    def fit(self, X, Y, val_X=None, val_Y=None, epoch=1000, print_every=100):
        for i in range(epoch):
            with tfe.GradientTape() as tape:
                preds = self.forward_pass(X)
                loss = self.cross_entropy(labels=Y, logits=preds)
            self.backward_pass(loss, tape)
            if (i+1) % print_every == 0:
                loss = tf.cast(loss, tf.float64).numpy()
                acc = compute_accuracy(preds, Y)
                if val_X is not None and val_Y is not None:
                    p = self.predict(val_X)
                    val_acc = compute_accuracy(p, val_Y)
                    print("epoch", i+1, "loss", round(loss, 5), "acc", acc, "val_acc", round(val_acc, 2))
                else:
                    print("epoch", i+1, "loss", round(loss, 5), "acc", round(acc,2))


model = Model()
model.fit(X, Y, X_t, Y_t, 500, 50)