Keras Maxpooling2d layer gives ValueError

Question

I am trying to replicate VGG16 model in keras, the following is my code:

model = Sequential()
model.add(ZeroPadding2D((1,1),input_shape=(3,224,224)))
model.add(Convolution2D(64, 3, 3, activation='relu'))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(64, 3, 3, activation='relu'))
model.add(MaxPooling2D((2,2), strides=(2,2)))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(128, 3, 3, activation='relu'))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(128, 3, 3, activation='relu'))
model.add(MaxPooling2D((2,2), strides=(2,2))) ###This line gives error
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(256, 3, 3, activation='relu'))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(256, 3, 3, activation='relu'))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(256, 3, 3, activation='relu'))
model.add(MaxPooling2D((2,2), strides=(2,2)))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(512, 3, 3, activation='relu'))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(512, 3, 3, activation='relu'))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(512, 3, 3, activation='relu'))
model.add(MaxPooling2D((2,2), strides=(2,2)))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(512, 3, 3, activation='relu'))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(512, 3, 3, activation='relu'))
model.add(ZeroPadding2D((1,1)))
model.add(Convolution2D(512, 3, 3, activation='relu'))
model.add(MaxPooling2D((2,2), strides=(2,2)))
model.add(Flatten())
model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(1000, activation='softmax'))

The maxpooling2d layer gives an error at the line which is commented

The error says:

ValueError: Negative dimension size caused by subtracting 2 from 1 for 'MaxPool_7' (op: 'MaxPool') with input shapes: [?,1,112,128].

What might be the reason behind this? How to solve this?

Edit: A more detailed error log:

---

ValueError Traceback (most recent call last) in () 12 model.add(Convolution2D(128, 3, 3, activation='relu')) 13 ---> 14 model.add(MaxPooling2D((2,2), strides=(2,2))) 15 16 model.add(ZeroPadding2D((1,1)))

/usr/local/lib/python2.7/dist-packages/keras/models.pyc in add(self, layer) 306 output_shapes=[self.outputs[0]._keras_shape]) 307 else: --> 308 output_tensor = layer(self.outputs[0]) 309 if type(output_tensor) is list: 310 raise Exception('All layers in a Sequential model '

/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in call(self, x, mask) 512 if inbound_layers: 513 # this will call layer.build() if necessary --> 514 self.add_inbound_node(inbound_layers, node_indices, tensor_indices) 515 input_added = True 516

/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in add_inbound_node(self, inbound_layers, node_indices, tensor_indices) 570 # creating the node automatically updates self.inbound_nodes 571 # as well as outbound_nodes on inbound layers. --> 572 Node.create_node(self, inbound_layers, node_indices, tensor_indices) 573 574 def get_output_shape_for(self, input_shape):

/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in create_node(cls, outbound_layer, inbound_layers, node_indices, tensor_indices) 147 148 if len(input_tensors) == 1: --> 149 output_tensors = to_list(outbound_layer.call(input_tensors[0], mask=input_masks[0])) 150 output_masks = to_list(outbound_layer.compute_mask(input_tensors[0], input_masks[0])) 151 # TODO: try to auto-infer shape if exception is raised by get_output_shape_for

/usr/local/lib/python2.7/dist-packages/keras/layers/pooling.pyc in call(self, x, mask) 160 strides=self.strides, 161 border_mode=self.border_mode, --> 162 dim_ordering=self.dim_ordering) 163 return output 164

/usr/local/lib/python2.7/dist-packages/keras/layers/pooling.pyc in _pooling_function(self, inputs, pool_size, strides, border_mode, dim_ordering) 210 border_mode, dim_ordering): 211 output = K.pool2d(inputs, pool_size, strides, --> 212 border_mode, dim_ordering, pool_mode='max') 213 return output 214

/usr/local/lib/python2.7/dist-packages/keras/backend/tensorflow_backend.pyc in pool2d(x, pool_size, strides, border_mode, dim_ordering, pool_mode) 1699 1700 if pool_mode == 'max': -> 1701 x = tf.nn.max_pool(x, pool_size, strides, padding=padding) 1702 elif pool_mode == 'avg': 1703
x = tf.nn.avg_pool(x, pool_size, strides, padding=padding)

/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/nn_ops.pyc in max_pool(value, ksize, strides, padding, data_format, name) 1391 padding=padding, 1392
data_format=data_format, -> 1393 name=name) 1394 1395

/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_nn_ops.pyc in _max_pool(input, ksize, strides, padding, data_format, name)
1593 result = _op_def_lib.apply_op("MaxPool", input=input, ksize=ksize, 1594 strides=strides, padding=padding, -> 1595 data_format=data_format, name=name) 1596 return result 1597

/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/op_def_library.pyc in apply_op(self, op_type_name, name, **keywords) 747 op = g.create_op(op_type_name, inputs, output_types, name=scope, 748 input_types=input_types, attrs=attr_protos, --> 749 op_def=op_def) 750 outputs = op.outputs 751 return _Restructure(ops.convert_n_to_tensor(outputs),

/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.pyc in create_op(self, op_type, inputs, dtypes, input_types, name, attrs, op_def, compute_shapes, compute_device) 2388
original_op=self._default_original_op, op_def=op_def) 2389 if compute_shapes: -> 2390 set_shapes_for_outputs(ret) 2391 self._add_op(ret) 2392
self._record_op_seen_by_control_dependencies(ret)

/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.pyc in set_shapes_for_outputs(op) 1783 raise RuntimeError("No shape function registered for standard op: %s" 1784
% op.type) -> 1785 shapes = shape_func(op) 1786 if shapes is None: 1787 raise RuntimeError(

/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/common_shapes.pyc in call_cpp_shape_fn(op, input_tensors_needed, debug_python_shape_fn) 594 status) 595 except errors.InvalidArgumentError as err: --> 596 raise ValueError(err.message) 597 598 # Convert TensorShapeProto values in output_shapes.

ValueError: Negative dimension size caused by subtracting 2 from 1 for 'MaxPool_7' (op: 'MaxPool') with input shapes: [?,1,112,128].

Solution 1

Quoting an answer mentioned in github, you need to specify the dimension ordering:

Keras is a wrapper over Theano or Tensorflow libraries. Keras uses the setting variable image_dim_ordering to decide if the input layer is Theano or Tensorflow format. This setting can be specified in 2 ways -

1. specify 'tf' or 'th' in ~/.keras/keras.json like so - image_dim_ordering: 'th'. Note: this is a json file.
2. or specify the image_dim_ordering in your model like so: model.add(MaxPooling2D(pool_size=(2, 2), dim_ordering="th"))

Update: Apr 2020 Keras 2.2.5 link seems to have an updated API where dim_ordering is changed to data_format so:

keras.layers.MaxPooling2D(pool_size=(2, 2), strides=None, padding='valid', data_format='channels_first') to get NCHW or use channels_last to get NHWC

Appendix: image_dim_ordering in 'th' mode the channels dimension (the depth) is at index 1 (e.g. 3, 256, 256). In 'tf' mode is it at index 3 (e.g. 256, 256, 3). Quoting @naoko from comments.

Solution 2

I faced the same issue, I solved it by changing my padding: 'valid' to padding:'SAME': I guess its enough to add a parameter padding:' same'

model.add(MaxPooling2D((2,2), strides=(2,2), padding='same'))

Solution 3

For keras with TensorFlow try following:

model.add(ZeroPadding2D((1, 1), input_shape=(img_rows, img_cols, channel)))

Solution 4

You are using the input shape as (3,x,y) should change it to input_shape=x,y,3

Solution 5

The accepted answer works. But you can also do the following:

    model.add(MaxPooling2D((2, 2), name='block1_pool', data_format='channels_last')

Keras assumes input to be (width, height, channels) for TensorFlow backend and (channel, width, height) for Theano backend. Since your input_shape=(3,224,224), specifying data_format='channels_last' should do the trick.

View more solutions

Comments

[Srikar Appalaraju]

@PranayMathur did not notice that!

[naoko]

Note: dim_ordering, In 'th' mode, the channels dimension (the depth) is at index 1, in 'tf' mode is it at index 3. For tensorflow 0.10 would use "tf" (default) and 0.11 would use "th". You can set this behavior in ~/.keras/keras.json

[kRazzy R]

please expand your answer, i'm getting error like : ValueError: Negative dimension size caused by subtracting 3 from 1 for 'conv2d_2/convolution' (op: 'Conv2D') with input shapes: [?,20,1,32], [3,3,32,32].

[Alexander Rossa]

@SrikarAppalaraju In you appendix, wouldn't that be Index 0 and Index 2 instead of 1 and 3?

[xdola]

TF2: Keyword argument not understood: 'dim_ordering'

[Farid Alijani]

is this answer valid for keras 2.2.5 ? cuz MaxPooling2D does not seem to have dim_ordering argument! keras.layers.MaxPooling2D(pool_size=(2, 2), strides=None, padding='valid', data_format=None)