returnn.util.pprint#
Alternative to the original pprint module.
This one has different behavior for indentation, specifically for dicts.
Also the order of dict items are kept as-is
(which is fine for newer Python versions, which will be the insertion order).
Compare (via our pprint):
{
'melgan': {
'class': 'subnetwork',
'from': 'data',
'subnetwork': {
'l0': {'class': 'pad', 'mode': 'reflect', 'axes': 'spatial', 'padding': (3, 3), 'from': 'data'},
'la1': {
'class': 'conv',
'from': 'l0',
'activation': None,
'with_bias': True,
'n_out': 384,
'filter_size': (7,),
'padding': 'valid',
'strides': (1,),
'dilation_rate': (1,)
},
'lay2': {'class': 'eval', 'eval': 'tf.nn.leaky_relu(source(0), alpha=0.2)', 'from': 'la1'},
'layer3_xxx': {
'class': 'transposed_conv',
'from': 'lay2',
'activation': None,
'with_bias': True,
'n_out': 192,
'filter_size': (10,),
'strides': (5,),
'padding': 'valid',
'output_padding': (1,),
'remove_padding': (3,)
},
'output': {'class': 'copy', 'from': 'layer3_xxx'}
}
},
'output': {'class': 'copy', 'from': 'melgan'}
}
- Vs (via original
pprint):: - {‘melgan’: {‘class’: ‘subnetwork’,
‘from’: ‘data’, ‘subnetwork’: {‘l0’: {‘axes’: ‘spatial’,
‘class’: ‘pad’, ‘from’: ‘data’, ‘mode’: ‘reflect’, ‘padding’: (3, 3)},
- ‘la1’: {‘activation’: None,
‘class’: ‘conv’, ‘dilation_rate’: (1,), ‘filter_size’: (7,), ‘from’: ‘l0’, ‘n_out’: 384, ‘padding’: ‘valid’, ‘strides’: (1,), ‘with_bias’: True},
- ‘lay2’: {‘class’: ‘eval’,
- ‘eval’: ‘tf.nn.leaky_relu(source(0), ‘
‘alpha=0.2)’,
‘from’: ‘la1’},
- ‘layer3_xxx’: {‘activation’: None,
‘class’: ‘transposed_conv’, ‘filter_size’: (10,), ‘from’: ‘lay2’, ‘n_out’: 192, ‘output_padding’: (1,), ‘padding’: ‘valid’, ‘remove_padding’: (3,), ‘strides’: (5,), ‘with_bias’: True},
‘output’: {‘class’: ‘copy’, ‘from’: ‘layer3_xxx’}}},
‘output’: {‘class’: ‘copy’, ‘from’: ‘melgan’}}
This is a very simple implementation. There are other similar alternatives: * [Rich](https://github.com/willmcgugan/rich) * [pprint++](https://github.com/wolever/pprintpp)