def conv_transpose(x, kernel, output_shape, strides, padding, data_format, dilation_rate):
# Keras gives every dim on the output_shape, but PlaidML expects to infer the channel dims; so restrict to spatial dims
data_format = _normalize_data_format(data_format)
if data_format == 'nxc':
output_shape = output_shape[1:-1]
elif data_format == 'ncx':
output_shape = output_shape[2:]
else:
raise ValueError('Could not parse data_format "{}"'.format(data_format))
rank = x.tensor.shape.ndims - 2
if strides is None:
strides = tuple(1 for _ in range(rank))
if dilation_rate is None:
dilation_rate = tuple(1 for _ in range(rank))
return _KerasNode(
'conv',
tensor=plaidml_op.convolution(
x.tensor,
kernel.tensor,
strides,
dilation_rate,
[1] * len(strides),
[],
1,
_normalize_padding(padding),
[],
data_format,
'xck',
'none',
False, # winograd_allowed
cur_name(),
'ungrouped',
'data',
output_shape))
def conv_transpose(x, kernel, output_shape, strides, padding, data_format, dilation_rate):
logger.debug(
'conv_transpose(x: {}, kernel: {}, output_shape: {}, strides: {}, padding: {}, data_format: {}, dilation_rate: {}'
.format(x, kernel, output_shape, strides, padding, data_format, dilation_rate))
# Keras gives every dim on the output_shape, but PlaidML expects to infer the channel dims; so restrict to spatial dims
data_format = _normalize_data_format(data_format)
if data_format == 'nxc':
output_shape = output_shape[1:-1]
elif data_format == 'ncx':
output_shape = output_shape[2:]
else:
raise ValueError("Could not parse data_format '{}'".format(data_format))
return _KerasNode(
'conv',
tensor=plaidml_op.convolution(
x.tensor,
kernel.tensor,
strides,
dilation_rate,
[1] * len(strides),
[],
1,
_normalize_padding(padding),
[],
data_format,
'xck',
'none',
False, # winograd_allowed
cur_name(),
'ungrouped',
'data',
output_shape))
def conv(x,
kernel,
strides=None,
padding='valid',
data_format=None,
dilation_rate=None,
channelwise=False):
if channelwise:
group_layout = 'in_C'
autogroup_mode = 'max'
else:
group_layout = 'none'
autogroup_mode = 'ungrouped'
return _KerasNode(
'conv',
tensor=plaidml_op.convolution(
x.tensor,
kernel.tensor,
strides,
dilation_rate,
[1] * len(strides),
[],
1,
_normalize_padding(padding),
[],
_normalize_data_format(data_format),
'xck',
group_layout,
False, # winograd_allowed
cur_name(),
autogroup_mode,
'none',
[]))
