为什么此Keras Conv2D层与输入不兼容？

Question

I am having trouble understanding what input shapes my first convolutional neural network expects.

My training set is 500 grayscale images of 50x50 pixels.

The network starts with a Conv2D layer. Documentation for the argument input_shape says:

Input shape:
    4D tensor with shape:
    `(samples, channels, rows, cols)` if data_format='channels_first'
    or 4D tensor with shape:
    `(samples, rows, cols, channels)` if data_format='channels_last'.

So I expected that I need to supply my images (which are so far stored in a column of a pandas.DataFrame ) as a numpy.array of the shape (500, 1, 50, 50) , since I only have one "color" channel in the images. I reshaped it as follows:

X = np.array([img for img in imgs["img_res"]])
X = X.reshape(-1, 1, img_size, img_size)

X.shape is now: (500, 1, 50, 50). I supplied that as an arguent to Conv2D .

model = tf.keras.models.Sequential([
    tf.keras.layers.Conv2D(filters=64, 
                           kernel_size=(3,3), 
                           input_shape=X.shape[1:], 
                           activation="relu"),
])

This produces the following error. Can you point out what is wrong here?

---------------------------------------------------------------------------
InvalidArgumentError                      Traceback (most recent call last)
/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in _create_c_op(graph, node_def, inputs, control_inputs)
   1566   try:
-> 1567     c_op = c_api.TF_FinishOperation(op_desc)
   1568   except errors.InvalidArgumentError as e:

InvalidArgumentError: Negative dimension size caused by subtracting 3 from 1 for 'conv2d/Conv2D' (op: 'Conv2D') with input shapes: [?,1,50,50], [3,3,50,64].

During handling of the above exception, another exception occurred:

ValueError                                Traceback (most recent call last)
<ipython-input-24-0b665136e60b> in <module>()
      3                            kernel_size=(3,3),
      4                            input_shape=X.shape[1:],
----> 5                            activation="relu"),
      6     #tf.keras.layers.MaxPool2D(pool_size=(2,2)),
      7     #tf.keras.layers.Conv2D(filters=64,

/usr/local/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/engine/sequential.py in __init__(self, layers, name)
     99     if layers:
    100       for layer in layers:
--> 101         self.add(layer)
    102 
    103   @property

/usr/local/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/engine/sequential.py in add(self, layer)
    162           # and create the node connecting the current layer
    163           # to the input layer we just created.
--> 164           layer(x)
    165           set_inputs = True
    166         else:

/usr/local/lib/python3.6/site-packages/tensorflow/python/keras/_impl/keras/engine/base_layer.py in __call__(self, inputs, *args, **kwargs)
    312     """
    313     # Actually call the layer (optionally building it).
--> 314     output = super(Layer, self).__call__(inputs, *args, **kwargs)
    315 
    316     if args and getattr(self, '_uses_inputs_arg', True):

/usr/local/lib/python3.6/site-packages/tensorflow/python/layers/base.py in __call__(self, inputs, *args, **kwargs)
    715 
    716         if not in_deferred_mode:
--> 717           outputs = self.call(inputs, *args, **kwargs)
    718           if outputs is None:
    719             raise ValueError('A layer\'s `call` method should return a Tensor '

/usr/local/lib/python3.6/site-packages/tensorflow/python/layers/convolutional.py in call(self, inputs)
    166 
    167   def call(self, inputs):
--> 168     outputs = self._convolution_op(inputs, self.kernel)
    169 
    170     if self.use_bias:

/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py in __call__(self, inp, filter)
    866 
    867   def __call__(self, inp, filter):  # pylint: disable=redefined-builtin
--> 868     return self.conv_op(inp, filter)
    869 
    870 

/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py in __call__(self, inp, filter)
    518 
    519   def __call__(self, inp, filter):  # pylint: disable=redefined-builtin
--> 520     return self.call(inp, filter)
    521 
    522 

/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/nn_ops.py in __call__(self, inp, filter)
    202         padding=self.padding,
    203         data_format=self.data_format,
--> 204         name=self.name)
    205 
    206 

/usr/local/lib/python3.6/site-packages/tensorflow/python/ops/gen_nn_ops.py in conv2d(input, filter, strides, padding, use_cudnn_on_gpu, data_format, dilations, name)
    954         "Conv2D", input=input, filter=filter, strides=strides,
    955         padding=padding, use_cudnn_on_gpu=use_cudnn_on_gpu,
--> 956         data_format=data_format, dilations=dilations, name=name)
    957     _result = _op.outputs[:]
    958     _inputs_flat = _op.inputs

/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)
    785         op = g.create_op(op_type_name, inputs, output_types, name=scope,
    786                          input_types=input_types, attrs=attr_protos,
--> 787                          op_def=op_def)
    788       return output_structure, op_def.is_stateful, op
    789 

/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in create_op(self, op_type, inputs, dtypes, input_types, name, attrs, op_def, compute_shapes, compute_device)
   3390           input_types=input_types,
   3391           original_op=self._default_original_op,
-> 3392           op_def=op_def)
   3393 
   3394       # Note: shapes are lazily computed with the C API enabled.

/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in __init__(self, node_def, g, inputs, output_types, control_inputs, input_types, original_op, op_def)
   1732           op_def, inputs, node_def.attr)
   1733       self._c_op = _create_c_op(self._graph, node_def, grouped_inputs,
-> 1734                                 control_input_ops)
   1735     else:
   1736       self._c_op = None

/usr/local/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in _create_c_op(graph, node_def, inputs, control_inputs)
   1568   except errors.InvalidArgumentError as e:
   1569     # Convert to ValueError for backwards compatibility.
-> 1570     raise ValueError(str(e))
   1571 
   1572   return c_op

ValueError: Negative dimension size caused by subtracting 3 from 1 for 'conv2d/Conv2D' (op: 'Conv2D') with input shapes: [?,1,50,50], [3,3,50,64].

Answer 1

Specify that you are not using the default data format by passing data_format='channels_first' to Conv2D.

model = tf.keras.models.Sequential([
     tf.keras.layers.Conv2D(filters=64, 
                       kernel_size=(3,3), 
                       input_shape=X.shape[1:], 
                       activation="relu", 
                       data_format='channels_first'),
     ])

Answer 2

You are using TensorFlow, which by default uses the "channels last" input format, meaning that the channels dimension should go at the end:

model = tf.keras.models.Sequential([
tf.keras.layers.Conv2D(filters=64, 
                       kernel_size=(3,3), 
                       input_shape=(50, 50, 1), 
                       activation="relu"),
])

The error happens because the 1 in your input shape was being interpreted as one of the spatial dimensions, producing a negative dimension after performing convolution.