PSPNet, Tensorflow.kerasに関してのエラー(Negative dimension size)

以下のコードを参考に, tensorflow.kerasで動くように書き直しています.
https://github.com/ykamikawa/tf-keras-PSPNet

以下のエラーがでて, 手も足もでない状態です.
解決方法は何かございますでしょうか.

Negative dimension size caused by subtracting 30 from 16 for '{{node pool_level_6_16/AvgPool}} = AvgPoolT=DT_FLOAT, data_format="NHWC", ksize=[1, 30, 30, 1], padding="VALID", strides=[1, 30, 30, 1]' with input shapes: [?,16,16,2048].

書き直し中のmodel.pyです.

Python
1ef interp_block(
2    x, num_filters=64, level=1, input_shape=(256,256, 3), output_stride=16
3):
4    """ interpolation block """
5    feature_map_shape = (input_shape[0] / output_stride, input_shape[1] / output_stride)
6 
7    # compute dataformat
8    if K.image_data_format() == "channels_last":
9        bn_axis = 3
10    else:
11        bn_axis = 1
12 
13    if output_stride == 16:
14        scale = 5
15    elif output_stride == 8:
16        scale = 10
17 
18    kernel = (level * scale, level * scale)
19    strides = (level * scale, level * scale)
20    global_feat = AveragePooling2D(
21        kernel, strides=strides, name="pool_level_%s_%s" % (level, output_stride)
22)(x)
23    global_feat = _conv(
24        filters=num_filters,
25        kernel_size=(1, 1),
26        padding="same",
27        name="conv_level_%s_%s" % (level, output_stride),
28    )(global_feat)
29    global_feat = BatchNormalization(
30        axis=bn_axis, name="bn_level_%s_%s" % (level, output_stride)
31    )(global_feat)
32    global_feat = Lambda(Interp, arguments={"shape": feature_map_shape})(global_feat)
33 
34    return global_feat
35
36
37
38
39
40コード