Keras 输入形状和尺寸问题

Question

I'm doing some RL using Keras (I'm a Torch guy and this is my 2nd or 3rd time using Keras), here is the simplified code

model=keras.models.Sequential([
    keras.layers.Dense(10,activation='relu',input_shape=[4],name='layer1'),
    keras.layers.Dense(4,activation='softmax',name='layer2'),
    ])

then I call it on some data

obs=tf.convert_to_tensor([x1,y1,x2,y2],dtype=tf.float32)
pred=model(obs)

where x1 etc are integers and I get the error

WARNING:tensorflow:Model was constructed with shape Tensor("layer1_input:0", shape=(None, 4), dtype=float32) for input (None, 4), but it was re-called on a Tensor with incompatible shape (4,).
Traceback (most recent call last):
  File "C:\Users\milok\ev_rl.py", line 131, in <module>
    all_rewards,all_grads = play_multiple(env,n_episodes_per_update,n_max_steps,model,loss_fn)
  File "C:\Users\milok\ev_rl.py", line 101, in play_multiple
    obs,reward,grad = take_step(env,obs,model,loss_fn)
  File "C:\Users\milok\ev_rl.py", line 81, in take_step
    pred=model(obs.as_tensor())
  File "C:\Users\milok\Anaconda3\lib\site-packages\tensorflow_core\python\keras\engine\base_layer.py", line 822, in __call__
    outputs = self.call(cast_inputs, *args, **kwargs)
  File "C:\Users\milok\Anaconda3\lib\site-packages\tensorflow_core\python\keras\engine\sequential.py", line 267, in call
    return super(Sequential, self).call(inputs, training=training, mask=mask)
  File "C:\Users\milok\Anaconda3\lib\site-packages\tensorflow_core\python\keras\engine\network.py", line 717, in call
    convert_kwargs_to_constants=base_layer_utils.call_context().saving)
  File "C:\Users\milok\Anaconda3\lib\site-packages\tensorflow_core\python\keras\engine\network.py", line 891, in _run_internal_graph
    output_tensors = layer(computed_tensors, **kwargs)
  File "C:\Users\milok\Anaconda3\lib\site-packages\tensorflow_core\python\keras\engine\base_layer.py", line 822, in __call__
    outputs = self.call(cast_inputs, *args, **kwargs)
  File "C:\Users\milok\Anaconda3\lib\site-packages\tensorflow_core\python\keras\layers\core.py", line 1142, in call
    outputs = gen_math_ops.mat_mul(inputs, self.kernel)
  File "C:\Users\milok\Anaconda3\lib\site-packages\tensorflow_core\python\ops\gen_math_ops.py", line 5615, in mat_mul
    _ops.raise_from_not_ok_status(e, name)
  File "C:\Users\milok\Anaconda3\lib\site-packages\tensorflow_core\python\framework\ops.py", line 6606, in raise_from_not_ok_status
    six.raise_from(core._status_to_exception(e.code, message), None)
  File "<string>", line 3, in raise_from
tensorflow.python.framework.errors_impl.InvalidArgumentError: In[0] is not a matrix. Instead it has shape [4] [Op:MatMul]```

Answer 1

take care to manage the batch dimension when you compute the predictions... you have to pass to your model an object of dim (batch_size, n_feat)

model=tf.keras.models.Sequential([
    tf.keras.layers.Dense(10,activation='relu',input_shape=[4],name='layer1'),
    tf.keras.layers.Dense(4,activation='softmax',name='layer2'),
    ])

### Error ###
obs=tf.constant([1,2,3,4],dtype=tf.float32) 
pred=model(obs)

### OK ###
obs=tf.constant([[1,2,3,4]],dtype=tf.float32)
pred=model(obs)

Answer 2

The error message is telling you that you are trying to call your model on a tensor with incompatible shape.

The tensor [x1,y1,x2,y2] has shape [4] , but when you set up the model you used a Dense node that expects objects of shape [batch, 4] .

Answer 3

obs should be a numpy array/tensor with shape (None, 4) . https://keras.io/guides/sequential_model/