03 LR solution

gistfile1.txt

# recall the training loop pattern

# 1. Batch of training data
# Batch of training data
x_values = tf.range(0,1,0.01)
# 1B: desired outputs or 'labels'
y_values = 3*x_values + 1 + tf.random_uniform(x_values.shape)

# 2. Model
# 2A: Variables that will be trained
# 2B: Prediction function combines training inputs, ops, and variables, generates prediction

A = tfe.Variable(0.0)
B = tfe.Variable(0.0)

def predict_y_values(x_values):
  return A*x_values+B


# the training loop
for i in range(200):
  with tfe.GradientTape() as tape:
      # 3. Loss computes error between prediction and desired output
      predicted_y_values = predict_y_values(x_values)
      loss_value = loss_function(predicted_y_values, y_values)
      # 5. Metrics tell you how well your model is training
      if i % 20 == 0:
        print("Loss at step {:03d}: {:.3f}".format(i, loss_value)) 
  # 4. Update step 
  # 4A: Gradient tells us which direction to change the variables to reduce loss
  gradient_A, gradient_B = tape.gradient(loss_value, [A, B])
  # 4B: Nudge the variables by a small step in the right direction
  A.assign_sub(gradient_A*0.1)
  B.assign_sub(gradient_B*0.1)