def inplace_sub(x, i, v):
"""Applies an inplace sub on input x at index i with value v.
Note that this function is not actually inplace - it allocates
a copy of x. The utility is not avoiding memory copies but rather
specifying a sparse update.
If i is None, x and v must be the same shape. Computes
y = x; y -= v;
If i is a scalar, x has a rank 1 higher than v's. Computes
y = x; y[i, :] -= v;
Otherwise, x and v must have the same rank. Computes
y = x; y[i, :] -= v;
Args:
x: A Tensor.
i: None, a scalar or a vector.
v: A Tensor.
Returns:
Returns y, which is guaranteed not to be an alias of x.
"""
return alias_inplace_sub(gen_array_ops.deep_copy(x), i, v)
def inplace_sub(x, i, v):
"""Applies an inplace sub on input x at index i with value v.
Note that this function is not actually inplace - it allocates
a copy of x. The utility is not avoiding memory copies but rather
specifying a sparse update.
If i is None, x and v must be the same shape. Computes
y = x; y -= v;
If i is a scalar, x has a rank 1 higher than v's. Computes
y = x; y[i, :] -= v;
Otherwise, x and v must have the same rank. Computes
y = x; y[i, :] -= v;
Args:
x: A Tensor.
i: None, a scalar or a vector.
v: A Tensor.
Returns:
Returns y, which is guaranteed not to be an alias of x.
"""
return alias_inplace_sub(gen_array_ops.deep_copy(x), i, v)
model(np.ones([1, 3]))
# model = keras.models.load_model("cartpole_model_with_TD(lambda)_60000steps.h5")
optimizer = keras.optimizers.RMSprop(learning_rate=0.01,
momentum=0.95,
rho=0.95)
model.summary()
programing_times = 10
replay_memory_length = 1000
replay_memory = []
gamma = 0.99
old_theta = []
for each in model.trainable_variables:
old_theta.append(deep_copy(each))
max_episode_num = 10
for episode in range(max_episode_num):
print(f'Episode: {episode+1}')
t = 0
state = env.reset()
step = 0
rewards = 0
while True:
action = greedy_action(model, state)
state_next, reward, done, _ = env.step(action)
if done and step < 300: