keras Conv1D 上的负尺寸

Question

我正在使用Keras 的模型 api 将一维卷积应用于大小为 20 的输入一维向量。我想要五个大小为 3 的内核。输入的形状为(None, 1,20)（可变数量的大小为 20 的一维向量）。

input = Input(shape=(1, 20))
conv = Conv1D(filters=5, kernel_size=3, activation=keras.activations.relu, input_shape=(None,20, 1))(input)
dense =dense(1)(conv)
model = Model(inputs=input, outputs=dense)

model.compile(loss=nn.customLoss, optimizer='adam')

history = model.fit(train_X, train_labels, batch_size=50,
                    epochs=15, validation_split=0.2)

模型的总结是——

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, None, 20)          0         
_________________________________________________________________
conv1d_1 (Conv1D)            (None, None, 5)           305       
_________________________________________________________________
dense_1 (Dense)              (None, None, 1)           6         
=================================================================
Total params: 311
Trainable params: 311
Non-trainable params: 0

train_x 的形状为 (None, 1, 20)，train_labels 的形状为 (None, 1)。

错误来自卷积层-

    Caused by op 'conv1d_1/convolution/Conv2D', defined at:
  File "/home/user/Desktop/hack/imlhack2018/conv_nn.py", line 72, in <module>
    main()
  File "/home/user/Desktop/hack/imlhack2018/conv_nn.py", line 42, in main
    conv = Conv1D(filters=5, kernel_size=3, activation=keras.activations.relu, input_shape=(None,20, 1))(input)
  File "/home/user/anaconda3/lib/python3.6/site-packages/keras/engine/topology.py", line 596, in __call__
    output = self.call(inputs, **kwargs)
  File "/home/user/anaconda3/lib/python3.6/site-packages/keras/layers/convolutional.py", line 156, in call
    dilation_rate=self.dilation_rate[0])
  File "/home/user/anaconda3/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py", line 3116, in conv1d
    data_format=tf_data_format)
  File "/home/user/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py", line 670, in convolution
    op=op)
  File "/home/user/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py", line 338, in with_space_to_batch
    return op(input, num_spatial_dims, padding)
  File "/home/user/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py", line 662, in op
    name=name)
  File "/home/user/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py", line 116, in _non_atrous_convolution
    name=scope)
  File "/home/user/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py", line 2010, in conv1d
    data_format=data_format)
  File "/home/user/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/gen_nn_ops.py", line 399, in conv2d
    data_format=data_format, name=name)
  File "/home/user/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 767, in apply_op
    op_def=op_def)
  File "/home/user/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 2506, in create_op
    original_op=self._default_original_op, op_def=op_def)
  File "/home/user/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1269, in __init__
    self._traceback = _extract_stack()

InvalidArgumentError (see above for traceback): computed output size would be negative
     [[Node: conv1d_1/convolution/Conv2D = Conv2D[T=DT_FLOAT, data_format="NHWC", padding="VALID", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true, _device="/job:localhost/replica:0/task:0/cpu:0"](conv1d_1/convolution/ExpandDims, conv1d_1/convolution/ExpandDims_1)]]

当我将padding="same" 添加到卷积层时，一切似乎都运行良好。这种行为的原因是什么？

Answer 1

您的输入形状是 (1, 20)，它被解释为一个宽度为 1、通道数为 20 的数组。您可能想要相反的，即宽度 20 和 1 通道。由于您的数组只有一个元素，因此在没有填充 SAME 的情况下执行卷积将导致负维度，从而产生错误。

请注意，卷积总是在空间维度上执行，对于 Conv1D，空间维度是形状数组中倒数第二个维度。最后一个维度代表渠道。

Answer 2

在官方文档中，写着“当使用该层作为模型的第一层时，提供一个input_shape参数（整数元组或None，不包括batch轴）”。我很困惑为什么它在输入层之后声明了conv1D的输入形状