我有以下模型：
class VDNMixer(kr.Model):
def __init__(self, hidden_units, observation_shape, action_dim):
super().__init__('VDNMixer')
self.action_dim = action_dim
input_layer = kl.Input(shape=observation_shape)
dense = MLP(hidden_units, 0, observation_shape)(input_layer)
stream_adv, stream_val = tf.split(dense, 2, axis=1)
advantage = kl.Dense(action_dim, activation=None, use_bias=None)(stream_adv)
advantage = tf.subtract(advantage, tf.reduce_mean(advantage, axis=1, keepdims=True))
value = kl.Dense(1, activation=None, use_bias=None)(stream_val)
Qout = value + advantage
self.q = kr.Model(inputs=input_layer, outputs=[Qout])
def __call__(self, list_of_obs_act, training=True):
"""
:param list_of_obs_act: list of n arrays with dimensions [None, obs_dim+1]
:param training: (bool)
:return: n
"""
result = []
for obs in list_of_obs_act:
# n_agents, obs_shape
act = obs[:, -1]
obs = obs[:, :-1]
qout = self.q(obs)
result.append(tf.reduce_sum(tf.one_hot(tf.cast(act, tf.int32), self.action_dim) * qout))
return tf.stack(result)
我可以用它来精确地计算损失函数：
def td_loss(targets, qtot):
return tf.keras.losses.mean_squared_error(targets, qtot)
# model params
observation_shape = (100, )
action_dim = 5
q_kwargs = {'hidden_units': [512, 256, 128],
'observation_shape': observation_shape,
'action_dim': action_dim}
net = VDNMixer(**q_kwargs)
# create dummy data
batch_size = 10
agents = np.random.randint(0, 8, size=batch_size)
list_of_obs_act = [np.hstack((np.random.rand(a, observation_shape[0]), np.random.randint(0, action_dim, size=(a, 1))))
for a in agents]
t_targets = np.random.rand(batch_size)
# compute loss
print(td_loss(net(list_of_obs_act), t_targets))
>> tf.Tensor(1.0777278029536537, shape=(), dtype=float64)
但是，如果我使用keras高级API来训练模型：
net.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), loss=td_loss,
metrics=[])
net.train_on_batch(list_of_obs_act, t_targets)
我收到以下错误：
Traceback (most recent call last):
File "/home/joao/anaconda3/envs/tf2/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3296, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-10-ecf35d2000b8>", line 62, in <module>
net.train_on_batch(list_of_obs_act, t_targets)
File "/home/joao/anaconda3/envs/tf2/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py", line 896, in train_on_batch
extract_tensors_from_dataset=True)
File "/home/joao/anaconda3/envs/tf2/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py", line 2335, in _standardize_user_data
self._set_inputs(cast_inputs)
File "/home/joao/anaconda3/envs/tf2/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py", line 2553, in _set_inputs
outputs = self(inputs, **kwargs)
File "<ipython-input-10-ecf35d2000b8>", line 37, in __call__
qout = self.q(obs)
File "/home/joao/anaconda3/envs/tf2/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py", line 662, in __call__
outputs = call_fn(inputs, *args, **kwargs)
File "/home/joao/anaconda3/envs/tf2/lib/python3.6/site-packages/tensorflow/python/keras/engine/network.py", line 753, in call
return self._run_internal_graph(inputs, training=training, mask=mask)
File "/home/joao/anaconda3/envs/tf2/lib/python3.6/site-packages/tensorflow/python/keras/engine/network.py", line 895, in _run_internal_graph
output_tensors = layer(computed_tensors, **kwargs)
File "/home/joao/anaconda3/envs/tf2/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py", line 662, in __call__
outputs = call_fn(inputs, *args, **kwargs)
File "/home/joao/anaconda3/envs/tf2/lib/python3.6/site-packages/tensorflow/python/autograph/impl/api.py", line 169, in wrapper
raise e.ag_error_metadata.to_exception(type(e))
ValueError: in converted code:
relative to /home/joao:
github/artificial_life/models/base.py:25 call *
inputs = layer(inputs)
anaconda3/envs/tf2/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py:611 __call__
self.name)
anaconda3/envs/tf2/lib/python3.6/site-packages/tensorflow/python/keras/engine/input_spec.py:163 assert_input_compatibility
layer_name + ' is incompatible with the layer: '
ValueError: Input 0 of layer dense_15 is incompatible with the layer: its rank is undefined, but the layer requires a defined rank.
有什么想法吗？
我正在使用Tensorflow：v2.0.0-beta0-16-g1d91213fe7 2.0.0-beta1
Python 3.6.8
Ubuntu 18.04