/fake_tf

模仿Tensorflow的API，自己实现自动求导、训练等功能

Primary LanguagePython

fake ft

This is a self practice to imitate Google's tensorflow. The fucntion of this model is mainly to realize automatic derivation, optimization, etc.

Of course, all the calculations are done on CPU without multiprocess.

TODO list

  • cache for repeated calculation
  • operations such as abs, pow
  • cache for repeated constants and operations
  • random initialization
  • None in placeholder shape
  • activation functions
  • SGD
  • higher end apis

Usage

from faketf import reduce_mean, Placeholder
from faketf.highend.layers import fully_connected
from faketf.train import SGD
import numpy as np

x = Placeholder([100, 5], name="x")
y_real = Placeholder([100, 1], name="real")
y = fully_connected(x, 1)
loss = reduce_mean((y - y_real) ** 2)
trainer = SGD(loss, learning_rate=0.01)

x_values = np.random.randn(100, 5)
true_weights = np.array([[1.0], [0.5], [-0.5], [2.0], [-0.3]])
y_values = x_values @ true_weights
feed_dict = {x: x_values, y_real: y_values}

for i in range(1000):
    trainer.train(feed_dict)
    if i % 100 == 0:
        trainer.learning_rate *= 0.8
        print("Loss: %0.4f" % np.asscalar(loss.eval(feed_dict)))
print(y.W.eval(), y.b.eval())