Negative dimension error while using keras Convolutional1D Layer

Question

I'm trying to create a char cnn using Keras. That type of cnn requires you to use Convolutional1D layer. But all the ways I try to add them to my model, it gives me errors at creation stage. Here is my code:

def char_cnn(n_vocab, max_len, n_classes):
    conv_layers = [[256, 7, 3],
                   [256, 7, 3],
                   [256, 3, None],
                   [256, 3, None],
                   [256, 3, None],
                   [256, 3, 3]]
    fully_layers = [1024, 1024]
    th = 1e-6

    embedding_size = 128

    inputs = Input(shape=(max_len,), name='sent_input', dtype='int64')

    # Embedding layer

    x = Embedding(n_vocab, embedding_size, input_length=max_len)(inputs)

    # Convolution layers
    for cl in conv_layers:
        x = Convolution1D(cl[0], cl[1])(x)
        x = ThresholdedReLU(th)(x)
        if not cl[2] is None:
            x = MaxPooling1D(cl[2])(x)


    x = Flatten()(x)


    #Fully connected layers

    for fl in fully_layers:
        x = Dense(fl)(x)
        x = ThresholdedReLU(th)(x)
        x = Dropout(0.5)(x)


    predictions = Dense(n_classes, activation='softmax')(x)

    model = Model(input=inputs, output=predictions)

    model.compile(optimizer='adam', loss='categorical_crossentropy')

    return model

And here is the error I receive when I try to call char_cnn function

InvalidArgumentError                      Traceback (most recent call last)
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/common_shapes.py in _call_cpp_shape_fn_impl(op, input_tensors_needed, input_tensors_as_shapes_needed, require_shape_fn)
    685           graph_def_version, node_def_str, input_shapes, input_tensors,
--> 686           input_tensors_as_shapes, status)
    687   except errors.InvalidArgumentError as err:

/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/errors_impl.py in __exit__(self, type_arg, value_arg, traceback_arg)
    515             compat.as_text(c_api.TF_Message(self.status.status)),
--> 516             c_api.TF_GetCode(self.status.status))
    517     # Delete the underlying status object from memory otherwise it stays alive

InvalidArgumentError: Negative dimension size caused by subtracting 3 from 1 for 'conv1d_26/convolution/Conv2D' (op: 'Conv2D') with input shapes: [?,1,1,256], [1,3,256,256].

How to fix it?

Answer 1

Your downsampling is too aggressive and the key argument here is max_len : when it's too small, the sequence becomes too short to perform either a convolution or a max-pooling. You set pool_size=3 , hence it shrinks the sequence by a factor of 3 after each pooling (see the example below). I suggest you try pool_size=2 .

The minimal max_len that this network can handle is max_len=123 . In this case x shape is transformed in the following way (according to conv_layers ):

(?, 123, 128)
(?, 39, 256)
(?, 11, 256)
(?, 9, 256)
(?, 7, 256)
(?, 5, 256)

Setting a smaller value, like max_len=120 causes x.shape=(?, 4, 256) before the last layer and this can't be performed.