def GetOutputSizeOfPooling2dMax(in_size,
kernel_size,
stride,
padding,
ceil_mode=False,
rigor=False,
verbose=False):
"""
Returns satus and the size of output tensor
:param in_size:
:param kernel_size:
:param stride:
:param padding:
:param ceil_mode: use floor when false, otherwize cel when true
:param rigor:
:param verbose:
:return: the size of output tensor
"""
err = 0
if rigor:
if (in_size < 1):
err += 1
if verbose:
dlr_common.DpuError(f"in_size should be positive: {in_size}",
flush=True)
if (kernel_size < 1):
err += 1
if verbose:
dlr_common.DpuError(
f"kernel_size should be positive: {kernel_size}",
flush=True)
if ((kernel_size % 2) == 1):
err += 1
if verbose:
dlr_common.DpuError(
f"kernel_size should be even: {kernel_size}", flush=True)
if (stride < 1):
err += 1
if verbose:
dlr_common.DpuError(
f"stride should be larger than 0: {stride}", flush=True)
if (padding < 0):
err += 1
if verbose:
dlr_common.DpuError(f"padding should be positive: {padding}",
flush=True)
if ceil_mode:
out_size = math.ceil(((in_size - kernel_size + 2 * padding) / stride) +
1)
else:
out_size = math.floor((
(in_size - kernel_size + 2 * padding) / stride) + 1)
if err > 0: return False, out_size
else: return True, out_size
def GetOutputSizeOfDeconvolution2d(in_size,
kernel_size,
stride,
padding,
output_padding=0,
dilation=1,
rigor=False,
verbose=False):
"""
Returns satus and the size of output tensor
:param in_size:
:param kernel_size:
:param stride:
:param padding:
:param rigor:
:param verbose:
:return: the size of output tensor
"""
err = 0
if rigor or dlr_common.rigor:
if (in_size < 1):
err += 1
if verbose:
dlr_common.DpuError(f"in_size should be positive: {in_size}",
flush=True)
if (kernel_size < 1):
err += 1
if verbose:
dlr_common.DpuError(
f"kernel_size should be positive: {kernel_size}",
flush=True)
if ((kernel_size % 2) != 1):
err += 1
if verbose:
dlr_common.DpuError(
f"kernel_size should be odd: {kernel_size}", flush=True)
if (stride < 1):
err += 1
if verbose:
dlr_common.DpuError(
f"stride should be larger than 0: {stride}", flush=True)
if (padding < 0):
err += 1
if verbose:
dlr_common.DpuError(f"padding should be positive: {padding}",
flush=True)
out_size = int((in_size - 1) * stride - 2 * padding + dilation *
(kernel_size - 1) + 1)
if err > 0: return False, out_size
else: return True, out_size
def LinearNd(
out_data # ndim x out_size
,
in_data # ndim x in_size
,
weight # out_size x in_size
,
bias=None # out_size
,
rigor=False,
verbose=False):
"""
Returns True on success, otherwize returns False
Applies a 1D matrix multiplication over an input data data.
Note that all nd-array lists are NumPy (mutable), not PyTorch tensor (immutable).
:param out_data: <mutable> output data, out_data[ndim][out_size]
:param in_data: input data, in_data[ndim][in_size]
:param weight: weight[out_size][in_size]
:param bias: bias for each output, bias[out_size]
:param rigor: check values rigorously when 'True'
:param verbose: output message more when 'True'
:return: 'True' on success, 'False' on failure.
Follwoings are derived from input arguments
. ndim: first dimension of out/in_data
. out_size: array size of out_data
. in_size: array size of in_data
. weight_size: dimension of weight
. bias_size: array size of bias
Following is an example usage for PyTorch.
LinearNd( tensor_out_data.data.numpy() # ndim x out_size
, tenso_in_data.data.numpy() # ndim x in_size
, tensor_weight.data.numpy() # out_size x in_size
, tensor_bias.data.numpy() # out_size
, rigor=True
, verbose=True)
"""
if rigor:
error = 0
if (out_data.ndim != 2):
error += 1
if verbose:
dlr_common.DpuError("out_data is not 1 dim", flush=True)
if (in_data.ndim != 2):
error += 1
if verbose: dlr_common.DpuError("in_data is not 1 dim", flush=True)
if (weight.ndim != 2):
error += 1
if verbose: dlr_common.DpuError("weight is not 2 dim", flush=True)
if (bias is not None) and (bias.ndim != 1):
error += 1
if verbose:
dlr_common.DpuError(f"bias should be 1 dim: {bias.ndim}",
flush=True)
t_out_ndim = out_data.shape[0]
t_out_size = out_data.shape[1] # note ndim (i.e., rank) is 1
t_in_ndim = in_data.shape[0]
t_in_size = in_data.shape[1] # note ndim (i.e., rank) is 1
t_weight_size_row = weight.shape[0] # note ndim (i.e., rank) is 2
t_weight_size_col = weight.shape[1] # note ndim (i.e., rank) is 2
if (t_out_ndim != t_in_ndim):
error += 1
dlr_common.DpuError(f"dimension mis-match", flush=True)
if (t_out_size != t_weight_size_row):
error += 1
dlr_common.DpuError(f"out_size mis-match", flush=True)
if (t_in_size != t_weight_size_col):
error += 1
dlr_common.DpuError(f"out_size mis-match", flush=True)
if verbose:
dlr_common.DpuInfo(f"out_size ={t_out_size} {out_data.shape}")
dlr_common.DpuInfo(f"in_size ={t_in_size} {in_data.shape}")
dlr_common.DpuInfo(
f"weight_size={t_weitht_dize_row} {t_weight_size_col}")
if (error != 0):
dlr_common.DpuError(" parameter mis-match", flush=True)
return False
#_fname=''
#_ctype=''
if out_data.dtype.type == np.int32:
_fname = 'LinearNdInt'
_ctype = ctypes.c_int
elif out_data.dtype.type == np.float32:
_fname = 'LinearNdFloat'
_ctype = ctypes.c_float
elif out_data.dtype.type == np.float64:
_fname = 'LinearNdDouble'
_ctype = ctypes.c_double
else:
dlr_common.DpuError(" not support " + str(out_data.dtype.type),
flush=True)
return False
_LinearNd = dlr_common.WrapFunction(
dlr_common._dlr,
_fname,
None # return type
,
[
ctypes.POINTER(_ctype) # out data
,
ctypes.POINTER(_ctype) # in data
,
ctypes.POINTER(_ctype) # weight
,
ctypes.POINTER(_ctype) # bias
,
ctypes.c_ushort # out_size
,
ctypes.c_ushort # in_size
,
ctypes.c_ushort # bias_size
,
ctypes.c_ubyte # ndim
,
ctypes.c_int # rigor
,
ctypes.c_int
]) # verbose
CP_out_data = out_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_in_data = in_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_weight = weight.ctypes.data_as(ctypes.POINTER(_ctype))
CP_out_size = ctypes.c_ushort(
out_data.shape[1]) # note ndim (i.e., rank) is 2
CP_in_size = ctypes.c_ushort(
in_data.shape[1]) # note ndim (i.e., rank) is 2
CP_ndim = ctypes.c_ubyte(in_data.shape[0]) # note ndim (i.e., rank) is 2
CP_rigor = 1 if rigor else 0
CP_verbose = 1 if verbose else 0
if (bias is None) or (bias.size == 0):
CP_bias = ctypes.POINTER(_ctype)()
CP_bias_size = ctypes.c_ushort(0)
else:
CP_bias = bias.ctypes.data_as(ctypes.POINTER(_ctype))
CP_bias_size = ctypes.c_ushort(bias.shape[0])
_LinearNd(CP_out_data, CP_in_data, CP_weight, CP_bias, CP_out_size,
CP_in_size, CP_bias_size, CP_ndim, CP_rigor, CP_verbose)
return True
def Norm1dBatch(
out_data # in_channel x in_size (contiguous)
,
in_data # in_channel x in_size (contiguous)
,
running_mean # in_channel
,
running_var # in_channel
,
scale=None # None or in_channel (default 1)
,
bias=None # None or in_channel (default 0)
,
epsilon=1E-5,
rigor=False,
verbose=False):
"""
Returns True on success, otherwize returns False
Applies a 1D matrix multiplication over an input data data.
Note that all nd-array lists are NumPy (mutable), not PyTorch tensor (immutable).
:param out_data: <mutable> output data, out_data[channel][size][size]
:param in_data: input data, in_data[channel][size][size]
:param running_mean:
:param running_var:
:param scale:
:param bias:
:param epsilon:
:param rigor: check values rigorously when 'True'
:param verbose: output message more when 'True'
:return: 'True' on success, 'False' on failure.
Following is an example usage for PyTorch.
Norm1dBatch( tensor_out_data.data.numpy() # ndim x out_size
, tenso_in_data.data.numpy() # ndim x in_size
, tensor_running_mean.data.numpy() # out_size x in_size
, tensor_running_var.data.numpy() # out_size
, tensor_scale.data.numpy() # out_size
, tensor_bias.data.numpy() # out_size
, epsilon
, rigor=True
, verbose=True)
"""
if rigor:
error = 0
if (out_data.ndim != in_data.ndim):
error += 1
if verbose:
dlr_common.DpuError("out_data in_data dimension mis-match",
flush=True)
if (running_mean.ndim != 1):
error += 1
if verbose:
dlr_common.DpuError(
f"running_mean dimension mis-match {running_mean.ndim}",
flush=True)
if (running_mean.size != in_data.shape[0]): # num of channels
error += 1
if verbose:
dlr_common.DpuError(
f"running_mean size mis-match {running_mean.size} {in_data.shape}",
flush=True)
if (running_var.ndim != 1):
error += 1
if verbose:
dlr_common.DpuError("running_var dimension mis-match",
flush=True)
if (running_var.size != in_data.shape[0]): # num of channels
error += 1
if verbose:
dlr_common.DpuError("running_var size mis-match", flush=True)
if (scale is not None) and (scale.ndim != 1):
error += 1
if verbose:
dlr_common.DpuError(f"scale should be 1 dim: {scale.ndim}",
flush=True)
if (bias is not None) and (bias.ndim != 1):
error += 1
if verbose:
dlr_common.DpuError(f"bias should be 1 dim: {bias.ndim}",
flush=True)
t_out_channel = out_data.shape[0]
t_out_size = out_data.shape[1]
t_in_channel = in_data.shape[0]
t_in_size = in_data.shape[1]
if (t_out_channel != t_in_channel):
error += 1
dlr_common.DpuError(f"channel mis-match", flush=True)
if (t_out_size != t_in_size):
error += 1
dlr_common.DpuError(f"channel mis-match", flush=True)
if (error != 0):
dlr_common.DpuError("parameter mis-match", flush=True)
return False
#_fname=''
#_ctype=''
if out_data.dtype.type == np.int32:
_fname = 'Norm1dBatchInt'
_ctype = ctypes.c_int
elif out_data.dtype.type == np.float32:
_fname = 'Norm1dBatchFloat'
_ctype = ctypes.c_float
elif out_data.dtype.type == np.float64:
_fname = 'Norm1dBatchDouble'
_ctype = ctypes.c_double
else:
dlr_common.DpuError(" not support " + str(out_data.dtype.type),
flush=True)
return False
_Norm1dBatch = dlr_common.WrapFunction(
dlr_common._dlr,
_fname,
None # return type
,
[
ctypes.POINTER(_ctype) # out data
,
ctypes.POINTER(_ctype) # in data
,
ctypes.POINTER(_ctype) # running_mean
,
ctypes.POINTER(_ctype) # running_var
,
ctypes.POINTER(_ctype) # scale
,
ctypes.POINTER(_ctype) # bias
,
ctypes.c_uint # in_size
,
ctypes.c_ushort # scale_size
,
ctypes.c_ushort # bias_size
,
ctypes.c_ushort # in_channel
,
ctypes.c_float # epsilon
,
ctypes.c_int # rigor
,
ctypes.c_int
]) # verbose
in_channel = in_data.shape[0]
in_size = int(in_data.size / in_channel) # num of elements per channel
CP_out_data = out_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_in_data = in_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_running_mean = running_mean.ctypes.data_as(ctypes.POINTER(_ctype))
CP_running_var = running_var.ctypes.data_as(ctypes.POINTER(_ctype))
CP_in_size = ctypes.c_uint(in_size)
CP_in_channel = ctypes.c_ushort(in_channel)
CP_epsilon = ctypes.c_float(epsilon)
CP_rigor = 1 if rigor else 0
CP_verbose = 1 if verbose else 0
if (scale is None) or (scale.size == 0):
CP_scale = ctypes.POINTER(_ctype)()
CP_scale_size = ctypes.c_ushort(0)
else:
CP_scale = scale.ctypes.data_as(ctypes.POINTER(_ctype))
CP_scale_size = ctypes.c_ushort(scale.shape[1])
if (bias is None) or (bias.size == 0):
CP_bias = ctypes.POINTER(_ctype)()
CP_bias_size = ctypes.c_ushort(0)
else:
CP_bias = bias.ctypes.data_as(ctypes.POINTER(_ctype))
CP_bias_size = ctypes.c_ushort(bias.shape[1])
_Norm1dBatch(CP_out_data, CP_in_data, CP_running_mean, CP_running_var,
CP_scale, CP_bias, CP_in_size, CP_scale_size, CP_bias_size,
CP_in_channel, CP_epsilon, CP_rigor, CP_verbose)
return True
def Pooling2dMax(
out_data # out_channel x out_size x out_size
,
in_data # in_channel x in_size x in_size
,
kernel_size # kernel_size x kernel_size
,
stride=1,
padding=0,
ceil_mode=False,
rigor=False,
verbose=False):
"""
Returns True on success, otherwize returns False
Applies a 2D mAXpolling over an input data composed of several input channels.
Note that all nd-array lists are NumPy (mutable), not PyTorch tensor (immutable).
:param out_data: <mutable> output data, out_data[out_channel][out_size][out_size]
:param in_data: input data, in_data[in_channel][in_size][in_size]
:param kernel_size:
:param stride: num of skips to apply next filter
:param padding: num of pixes at the boundary
:param ceil_mode: use floor() when false, otherwize ceil()
:param rigor: check values rigorously when 'True'
:param verbose: output message more when 'True'
:return: 'True' on success, 'False' on failure.
Follwoings are derived from input arguments
. out_size: array size of out_data
. in_size: array size of in_data
. in_chnannels: num of input channels
. out_channels: num of output channels (it should be the same as in_channel)
Following is an example usage for PyTorch.
Pooling2dMax( tensor_out_data.data.numpy() # out_channel x out_size x out_size
, tenso_in_data.data.numpy() # in_channel x in_size x in_size
, kernel_size
, stride
, padding
, rigor=True
, verbose=True)
"""
if rigor:
error = 0
if (out_data.ndim != 3):
error += 1
if verbose: dlr_common.DpuError("out_data is not 3 dim")
if (in_data.ndim != 3):
error += 1
if verbose: dlr_common.DpuError("in_data is not 3 dim")
if (kernel_size < 2):
error += 1
if verbose: dlr_common.DpuError("kernel_size should be >=2")
if (stride < 1):
error += 1
if verbose: dlr_common.DpuError("stride should be >=1")
if (padding < 0):
error += 1
if verbose: dlr_common.DpuError("stride should be >=0")
t_out_size = out_data.shape[2] # note ndim (i.e., rank) is 3
t_in_size = in_data.shape[2] # note ndim (i.e., rank) is 3
t_kernel_size = kernel_size
t_in_channel = in_data.shape[0]
t_out_channel = out_data.shape[0]
t_stride = stride
t_padding = padding
if (t_in_channel != t_out_channel):
error += 1
if verbose:
dlr_common.DpuError("in/out channel should be the same")
status, t_out_size_expect = GetOutputSizeOfPooling2dMax(
t_in_size, t_kernel_size, t_stride, t_padding)
if not status: return False # something wrong with arguments
if (t_out_size != t_out_size_expect):
error += 1
if verbose:
dlr_common.DpuError(
f"out_size mis-match {t_out_size} {t_out_size_expect}")
if ((t_kernel_size % 2) == 1):
error += 1
if verbose: dlr_common.DpuError(f"kernel_size should be even")
if verbose:
dlr_common.DpuInfo(f"out_channel={t_out_channel} {out_data.shape}")
dlr_common.DpuInfo(f"in_channel ={t_in_channel} {in_data.shape}")
dlr_common.DpuInfo(f"out_size ={t_out_size} {out_data.shape}")
dlr_common.DpuInfo(f"in_size ={t_in_size} {in_data.shape}")
dlr_common.DpuInfo(f"kernel_size={t_kernel_size}")
dlr_common.DpuInfo(f"stride ={t_stride} {stride}")
dlr_common.DpuInfo(f"padding ={t_padding} {padding}")
if (error != 0):
dlr_common.DpuError("parameter mis-match")
return False
#_fname=''
#_ctype=''
if out_data.dtype.type == np.int32:
_fname = 'Pooling2dMaxInt'
_ctype = ctypes.c_int
elif out_data.dtype.type == np.float32:
_fname = 'Pooling2dMaxFloat'
_ctype = ctypes.c_float
elif out_data.dtype.type == np.float64:
_fname = 'Pooling2dMaxDouble'
_ctype = ctypes.c_double
else:
dlr_common.DpuError("not support " + str(out_data.dtype.type))
return False
_Pooling2dMax = dlr_common.WrapFunction(
dlr_common._dlr,
_fname,
None # return type
,
[
ctypes.POINTER(_ctype) # output features
,
ctypes.POINTER(_ctype) # input image
,
ctypes.c_ushort # out_size
,
ctypes.c_ushort # in_size
,
ctypes.c_ubyte # kernel_size (only for square filter)
,
ctypes.c_ushort # channel
,
ctypes.c_ubyte # stride
,
ctypes.c_ubyte # padding
,
ctypes.c_int # ceil_mode
,
ctypes.c_int # rigor
,
ctypes.c_int # verbose
])
CP_out_data = out_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_in_data = in_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_out_size = ctypes.c_ushort(
out_data.shape[2]) # note ndim (i.e., rank) is 3
CP_in_size = ctypes.c_ushort(
in_data.shape[2]) # note ndim (i.e., rank) is 3
CP_kernel_size = ctypes.c_ubyte(kernel_size)
CP_channel = ctypes.c_ushort(in_data.shape[0])
CP_stride = ctypes.c_ubyte(stride)
CP_padding = ctypes.c_ubyte(padding)
CP_ceil_mode = 1 if ceil_mode else 0
CP_rigor = 1 if rigor else 0
CP_verbose = 1 if verbose else 0
_Pooling2dMax(CP_out_data, CP_in_data, CP_out_size, CP_in_size,
CP_kernel_size, CP_channel, CP_stride, CP_padding,
CP_ceil_mode, CP_rigor, CP_verbose)
return True
def Activations( func_name
, out_data # any dimension
, in_data # any dimension
, negative_slope=0.01 # for LeakyReLu
, rigor=False
, verbose=False):
"""
Returns True on success, otherwize returns False
Applies a non-linear activation function over an input data composed of several input channels.
Note that all nd-array lists are NumPy (mutable), not PyTorch tensor (immutable).
:param out_data: <mutable> output data, out_data[...]
:param in_data: input data, in_data[...]
:param rigor: check values rigorously when 'True'
:param verbose: output message more when 'True'
:return: 'True' on success, 'False' on failure.
Follwoings are derived from input arguments
. out_size: array size of out_data
. in_size: array size of in_data
Following is an example usage for PyTorch.
Activation'FUNC_NAME'( tensor_out_data.data.numpy() # contiguous array
, tenso_in_data.data.numpy() # contiguous array
, rigor=True
, verbose=True)
"""
if rigor:
error =0
if (out_data.ndim!=in_data.ndim):
error += 1
if verbose: dlr_common.DpuError("data dimension mis-match")
if (out_data.size!=in_data.size):
error += 1
if verbose: dlr_common.DpuError(f"data size mis-match {in_data.size} {out_data.size}")
for dim in range(in_data.ndim):
if (in_data.shape[dim]!=out_data.shape[dim]):
error += 1
if verbose: dlr_common.DpuError("data dimension size mis-match")
if (error!=0):
dlr_common.DpuError("parameter mis-match");
return False
if (out_data.ndim==0) or (out_data.ndim==1):
channel = 1
size = out_data.size
elif (out_data.ndim==2):
channel = out_data.shape[0]
size = out_data.shape[1]
else:
channel = out_data.shape[0]
size = np.prod(out_data.shape[1:])
#_fname=''
#_ctype=''
if out_data.dtype.type == np.int32:
_fname = 'Activation'+func_name+'Int'
_ctype = ctypes.c_int
elif out_data.dtype.type == np.float32:
_fname = 'Activation'+func_name+'Float'
_ctype = ctypes.c_float
elif out_data.dtype.type == np.float64:
_fname = 'Activation'+func_name+'Double'
_ctype = ctypes.c_double
else:
dlr_common.DpuError("not support "+str(out_data.dtype.type))
return False
if func_name == 'LeakyReLu':
_Activation=dlr_common.WrapFunction(dlr_common._dlr
,_fname
, None # return type
,[ctypes.POINTER(_ctype) # output
,ctypes.POINTER(_ctype) # input
,ctypes.c_uint # number of elements
,ctypes.c_ushort # number of channels
,ctypes.c_uint # negative slope
,ctypes.c_int # rigor
,ctypes.c_int # verbose
])
CP_out_data = out_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_in_data = in_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_size = ctypes.c_uint(size)
CP_channel = ctypes.c_ushort(channel)
CP_negative_slope = ctypes.c_uint.from_buffer(ctypes.c_float(negative_slope)).value
CP_rigor = 1 if rigor else 0
CP_verbose = 1 if verbose else 0
_Activation(CP_out_data
,CP_in_data
,CP_size
,CP_channel
,CP_negative_slope
,CP_rigor
,CP_verbose
)
else :
_Activation=dlr_common.WrapFunction(dlr_common._dlr
,_fname
, None # return type
,[ctypes.POINTER(_ctype) # output
,ctypes.POINTER(_ctype) # input
,ctypes.c_uint # number of elements
,ctypes.c_ushort # number of channels
,ctypes.c_int # rigor
,ctypes.c_int # verbose
])
CP_out_data = out_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_in_data = in_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_size = ctypes.c_uint(size)
CP_channel = ctypes.c_ushort(channel)
CP_rigor = 1 if rigor else 0
CP_verbose = 1 if verbose else 0
_Activation(CP_out_data
,CP_in_data
,CP_size
,CP_channel
,CP_rigor
,CP_verbose
)
return True
def Deconvolution2d(
out_data # out_channel x out_size x out_size
,
in_data # in_channel x in_size x in_size
,
kernel # in_channel x out_channel x kernel_size x kernel_size
,
bias=None # out_channel
,
stride=1,
padding=0,
rigor=False,
verbose=False):
"""
Returns True on success, otherwize returns False
Applies a 2D deconvolution (transpose convolution) over an input data composed of several input channels.
Note that all nd-array lists are NumPy (mutable), not PyTorch tensor (immutable).
:param out_data: <mutable> output data, out_data[out_channel][out_size][out_size]
:param in_data: input data, in_data[in_channel][in_size][in_size]
:param kernel: kernel (or filter), kernel[in_channel][out_channel][kernel_size][kernel_size]
:param bias: bias for each filter (kernel), bias[out_channel]
:param stride: num of skips to apply next filter
:param padding: num of pixes at the boundary
:param rigor: check values rigorously when 'True'
:param verbose: output message more when 'True'
:return: 'True' on success, 'False' on failure.
Follwoings are derived from input arguments
. out_size: array size of out_data
. in_size: array size of in_data
. kernel_size: dimension of filter, e.g., 3 means 3x3 kernel
. in_chnannels: num of input channels, e.g., 3 for RGB, 1 for gray
. out_channels: num of filters
. bias_size: array size of bias
Following is an example usage for PyTorch.
deconvolution2d( tensor_out_data.data.numpy()
, tenso_in_data.data.numpy()
, tensor_kernel.data.numpy()
, tensor_bias.data.numpy()
, stride
, padding
, rigor=True
, verbose=True)
"""
if rigor or dlr_common.rigor:
error = 0
if (out_data.ndim != 3):
error += 1
if verbose:
dlr_common.DpuError("out_data is not 3 dim", flush=True)
if (in_data.ndim != 3):
error += 1
if verbose: dlr_common.DpuError("in_data is not 3 dim", flush=True)
if (kernel.ndim != 4):
error += 1
if verbose: dlr_common.DpuError("kernel is not 4 dim", flush=True)
if (bias is not None) and (bias.ndim != 1):
error += 1
if verbose:
dlr_common.DpuError(f"bias should be 1 dim: {bias.ndim}",
flush=True)
if (stride < 1):
error += 1
if verbose:
dlr_common.DpuError(f"stride should be >=1: {stride}",
flush=True)
if (padding < 0):
error += 1
if verbose:
dlr_common.DpuError(f"padding should be >=0: {padding}",
flush=True)
t_out_size = out_data.shape[2] # note ndim (i.e., rank) is 3
t_in_size = in_data.shape[2] # note ndim (i.e., rank) is 3
t_kernel_size = kernel.shape[3] # note ndim (i.e., rank) is 4
t_in_channel = in_data.shape[0]
t_out_channel = out_data.shape[0]
t_stride = stride
t_padding = padding
status, t_out_size_expect = GetOutputSizeOfDeconvolution2d(
in_size=t_in_size,
kernel_size=t_kernel_size,
stride=t_stride,
padding=t_padding,
output_padding=0,
dilation=1,
rigor=rigor,
verbose=verbose)
if not status: return False # something wrong with arguments
if (t_out_size != t_out_size_expect):
error += 1
dlr_common.DpuError(
f"out_size mis-match: {t_out_size, t_out_size_expect}",
flush=True)
if ((t_kernel_size % 2) != 1):
error += 1
dlr_common.DpuError(f"kernel_size should be odd: {t_kernel_size}",
flush=True)
if verbose:
dlr_common.DpuInfo(f"out_channel={t_out_channel} {out_data.shape}")
dlr_common.DpuInfo(f"in_channel ={t_in_channel} {in_data.shape}")
dlr_common.DpuInfo(f"out_size ={t_out_size} {out_data.shape}")
dlr_common.DpuInfo(f"in_size ={t_in_size} {in_data.shape}")
dlr_common.DpuInfo(f"kernel_size={t_kernel_size} {kernel.shape}")
dlr_common.DpuInfo(f"stride ={t_stride} {stride}")
dlr_common.DpuInfo(f"padding ={t_padding} {padding}")
if (error != 0):
dlr_common.DpuError(" parameter mis-match", flush=True)
return False
#_fname=''
#_ctype=''
if out_data.dtype.type == np.int32:
_fname = 'Deconvolution2dInt'
_ctype = ctypes.c_int
elif out_data.dtype.type == np.float32:
_fname = 'Deconvolution2dFloat'
_ctype = ctypes.c_float
elif out_data.dtype.type == np.float64:
_fname = 'Deconvolution2dDouble'
_ctype = ctypes.c_double
else:
dlr_common.DpuError(" not support " + str(out_data.dtype.type),
flush=True)
return False
_Deconv2d = dlr_common.WrapFunction(
dlr_common._dlr,
_fname,
None # return type
,
[
ctypes.POINTER(_ctype) # output features
,
ctypes.POINTER(_ctype) # input image
,
ctypes.POINTER(_ctype) # kernels
,
ctypes.POINTER(_ctype) # bias
,
ctypes.c_ushort # out_size
,
ctypes.c_ushort # in_size
,
ctypes.c_ubyte # kernel_size (only for square filter)
,
ctypes.c_ushort # bias_size
,
ctypes.c_ushort # in_channel
,
ctypes.c_ushort # out_channel
,
ctypes.c_ubyte # stride
,
ctypes.c_ubyte # padding
,
ctypes.c_int # rigor
,
ctypes.c_int
]) # verbose
CP_out_data = out_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_in_data = in_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_kernel = kernel.ctypes.data_as(ctypes.POINTER(_ctype))
CP_out_size = ctypes.c_ushort(
out_data.shape[2]) # note ndim (i.e., rank) is 3
CP_in_size = ctypes.c_ushort(
in_data.shape[2]) # note ndim (i.e., rank) is 3
CP_kernel_size = ctypes.c_ubyte(
kernel.shape[3]) # note ndim (i.e., rank) is 4
CP_in_channel = ctypes.c_ushort(in_data.shape[0]) # kernel.shape[0]
CP_out_channel = ctypes.c_ushort(kernel.shape[1])
CP_stride = ctypes.c_ubyte(stride)
CP_padding = ctypes.c_ubyte(padding)
CP_rigor = 1 if rigor else 0
CP_verbose = 1 if verbose else 0
if (bias is None) or (bias.size == 0):
CP_bias = ctypes.POINTER(_ctype)()
CP_bias_size = ctypes.c_ushort(0)
else:
CP_bias = bias.ctypes.data_as(ctypes.POINTER(_ctype))
CP_bias_size = ctypes.c_ushort(bias.shape[0])
_Deconv2d(CP_out_data, CP_in_data, CP_kernel, CP_bias, CP_out_size,
CP_in_size, CP_kernel_size, CP_bias_size, CP_in_channel,
CP_out_channel, CP_stride, CP_padding, CP_rigor, CP_verbose)
return True
def Concat2d(
out_data #
,
in_dataA # in_rowsA x in_colsA
,
in_dataB # in_rowsB x in_colsB
,
dim=0,
rigor=False,
verbose=False):
"""
Returns True on success, otherwize returns False
Applies a 2D Concatenation over two 2-dimensional input data
Note that all nd-array lists are NumPy (mutable), not PyTorch tensor (immutable).
:param out_data: <mutable> output data, out_data[][]
:param in_dataA: input data, in_dataA[in_rowsA][in_colsA]
:param in_dataB: input data, in_dataB[in_rowsB][in_colsB]
:param dim: dimension to concatenate, 0 or 1
:param rigor: check values rigorously when 'True'
:param verbose: output message more when 'True'
:return: 'True' on success, 'False' on failure.
Follwoings are derived from input arguments
. out_rows:
. out_cols:
. in_rowsA:
. in_colsA:
. in_rowsB:
. in_colsB:
. dim:
Following is an example usage for PyTorch.
Concat2d( tensor_out_data.data.numpy()
, tenso_in_dataA.data.numpy()
, tenso_in_dataB.data.numpy()
, dim
, rigor=True
, verbose=True)
"""
if rigor:
error = 0
if (out_data.ndim != 2):
error += 1
if verbose: dlr_common.DpuError("out_data is not 2 dim")
if (in_dataA.ndim != 2):
error += 1
if verbose: dlr_common.DpuError("in_data is not 2 dim")
if (in_dataB.ndim != 2):
error += 1
if verbose: dlr_common.DpuError("in_data is not 2 dim")
if (dim != 0) and (dim != 1):
error += 1
if verbose: dlr_common.DpuError("dim should be 0 or 1")
t_in_rowsA = in_dataA.shape[0]
t_in_colsA = in_dataA.shape[1]
t_in_rowsB = in_dataB.shape[0]
t_in_colsB = in_dataB.shape[1]
if dim == 0:
t_out_rows = in_dataA.shape[0] + in_dataB.shape[0]
t_out_cols = in_dataA.shape[1]
else:
t_out_rows = in_dataA.shape[0]
t_out_cols = in_dataA.shape[1] + in_dataB.shape[1]
if (t_out_rows != out_data.shape[0]):
error += 1
if verbose: dlr_common.DpuError("out data row count error")
if (t_out_cols != out_data.shape[1]):
error += 1
if verbose: dlr_common.DpuError("out data column count error")
if dim == 0:
if (t_in_colsA != t_in_colsB):
error += 1
if verbose: dlr_common.DpuError("in dimension eror")
else:
if (t_in_rowsA != t_in_rowsB):
error += 1
if verbose: dlr_common.DpuError("in dimension eror")
if verbose:
dlr_common.DpuInfo(f"out_data={out_data.shape}")
dlr_common.DpuInfo(f"in_dataA={in_dataA.shape}")
dlr_common.DpuInfo(f"in_dataB={in_dataB.shape}")
dlr_common.DpuInfo(f"dim ={dim}")
if (error != 0):
dlr_common.DpuError("parameter mis-match")
return False
#_fname=''
#_ctype=''
if out_data.dtype.type == np.int32:
_fname = 'Concat2dInt'
_ctype = ctypes.c_int
elif out_data.dtype.type == np.float32:
_fname = 'Concat2dFloat'
_ctype = ctypes.c_float
elif out_data.dtype.type == np.float64:
_fname = 'Concat2dDouble'
_ctype = ctypes.c_double
else:
dlr_common.DpuError("not support " + str(out_data.dtype.type))
return False
_Concat2d = dlr_common.WrapFunction(
dlr_common._dlr,
_fname,
None # return type
,
[
ctypes.POINTER(_ctype) # output
,
ctypes.POINTER(_ctype) # input
,
ctypes.POINTER(_ctype) # input
,
ctypes.c_ushort # in_rowsA
,
ctypes.c_ushort # in_colsA
,
ctypes.c_ushort # in_rowsB
,
ctypes.c_ushort # in_colsB
,
ctypes.c_ubyte # dim
,
ctypes.c_int # rigor
,
ctypes.c_int # verbose
])
CP_out_data = out_data.ctypes.data_as(ctypes.POINTER(_ctype))
CP_in_dataA = in_dataA.ctypes.data_as(ctypes.POINTER(_ctype))
CP_in_dataB = in_dataB.ctypes.data_as(ctypes.POINTER(_ctype))
CP_in_rowsA = ctypes.c_ushort(in_dataA.shape[0])
CP_in_colsA = ctypes.c_ushort(in_dataA.shape[1])
CP_in_rowsB = ctypes.c_ushort(in_dataB.shape[0])
CP_in_colsB = ctypes.c_ushort(in_dataB.shape[1])
CP_dim = ctypes.c_ubyte(dim)
CP_rigor = 1 if rigor else 0
CP_verbose = 1 if verbose else 0
_Concat2d(CP_out_data, CP_in_dataA, CP_in_dataB, CP_in_rowsA, CP_in_colsA,
CP_in_rowsB, CP_in_colsB, CP_dim, CP_rigor, CP_verbose)
return True
