def __init__(self,
lib,
dtype,
N,
C,
K,
D=1,
H=1,
W=1,
T=1,
R=1,
S=1,
pad_d=0,
pad_h=0,
pad_w=0,
str_d=1,
str_h=1,
str_w=1):
super(ConvLayer, self).__init__(lib, dtype, N, np.float32)
# Compute the output spatial dimensions
M = lib.output_dim(D, T, pad_d, str_d)
P = lib.output_dim(H, R, pad_h, str_h)
Q = lib.output_dim(W, S, pad_w, str_w)
self.C = C
self.K = K
self.M = M
self.P = P
self.Q = Q
self.NCK = (N, C, K)
self.TRS = (T, R, S)
self.DHW = (D, H, W)
self.MPQ = (M, P, Q)
self.padding = (pad_d, pad_h, pad_w)
self.strides = (str_d, str_h, str_w)
self.all_params = (N, C, K, D, H, W, T, R, S, pad_d, pad_h, pad_w,
str_d, str_h, str_w)
self.dimI = (C, D, H, W, N)
self.dimF = (C, T, R, S, K)
self.dimFb = (K, T, R, S, C)
self.dimO = (K, M, P, Q, N)
self.dimI2 = (C * D * H * W, N)
self.dimF2 = (C * T * R * S, K)
self.dimF2t = (K, C * T * R * S)
self.dimO2 = (K * M * P * Q, N)
self.dimS = (K, 1)
self.sizeI = reduce(mul, self.dimI, 1)
self.sizeF = reduce(mul, self.dimF, 1)
self.sizeO = reduce(mul, self.dimO, 1)
self.nOut = reduce(mul, self.MPQ, 1) * K
# flop count for benchmarking
self.flops = P * Q * M * K * N * C * R * S * T * 2.0
args = (lib, self.dtype, N, C, K, D, H, W, T, R, S, M, P, Q, pad_d,
pad_h, pad_w, str_d, str_h, str_w)
#lib.enable_winograd = 0
####### Cuda C ###########
if lib.use_cudac_kernels:
#3D conv not supported yet
if T > 1 or D > 1:
raise ValueError(
"3D Convolution not supported by CUDA C kernels and pre-Maxwell GPUs"
)
# TODO small C bprop?
self.fprop_kernels = convolution.FpropCuda(*args)
self.bprop_kernels = convolution.BpropCuda(*args)
self.updat_kernels = convolution.UpdateCuda(*args)
####### Winograd ###########
elif lib.enable_winograd and R == 3 and S == 3 and all(
x == 1 for x in (D, M, T, str_w, str_h, str_d)):
from .winograd_conv import (FpropWinograd_2x2_3x3,
BpropWinograd_2x2_3x3,
UpdateWinograd_3x3_2x2,
FpropWinograd_4x4_3x3,
BpropWinograd_4x4_3x3,
UpdateWinograd_3x3_4x4)
# Temp for now till we can autotune
# 2 is safer for fp16 without batchnorm
if dtype == np.float32 and lib.enable_winograd == 4:
winograd = 4
else:
winograd = 2
if C < 8:
self.fprop_kernels = convolution.FpropDirect(*args)
elif winograd == 4 and H * W < 112 * 112:
self.fprop_kernels = FpropWinograd_4x4_3x3(*args)
else:
self.fprop_kernels = FpropWinograd_2x2_3x3(*args)
if winograd == 4 and H * W < 112 * 112:
self.bprop_kernels = BpropWinograd_4x4_3x3(*args)
else:
self.bprop_kernels = BpropWinograd_2x2_3x3(*args)
if N >= 4 and (C < 8 or H * W > 112 * 112):
self.updat_kernels = convolution.UpdateDirect(*args)
elif winograd == 4:
self.updat_kernels = UpdateWinograd_3x3_4x4(*args)
else:
self.updat_kernels = UpdateWinograd_3x3_2x2(*args)
# elif lib.enable_winograd and not lib.deterministic and N > 1 and \
# R == 5 and S == 5 and all(x == 1 for x in (D,M,T,str_w,str_h,str_d)):
#
# from .winograd_conv import (FpropWinograd_2x2_5x5, BpropWinograd_2x2_5x5)
#
# self.fprop_kernels = FpropWinograd_2x2_5x5(*args)
# self.bprop_kernels = BpropWinograd_2x2_5x5(*args)
# if N >= 4:
# self.updat_kernels = convolution.UpdateDirect(*args)
####### Direct ###########
else:
self.fprop_kernels = convolution.FpropDirect(*args)
self.bprop_kernels = convolution.BpropDirect(*args)
if N >= 4:
self.updat_kernels = convolution.UpdateDirect(*args)
def __init__(self, lib, dtype,
N, C, K,
D=1, H=1, W=1,
T=1, R=1, S=1,
pad_d=0, pad_h=0, pad_w=0,
str_d=1, str_h=1, str_w=1,
relu=False, bsum=False):
super(ConvLayer, self).__init__(lib, dtype, N, np.float32)
# Compute the output spatial dimensions
M = lib.output_dim(D, T, pad_d, str_d)
P = lib.output_dim(H, R, pad_h, str_h)
Q = lib.output_dim(W, S, pad_w, str_w)
self.C = C
self.K = K
self.M = M
self.P = P
self.Q = Q
self.NCK = (N, C, K)
self.TRS = (T, R, S)
self.DHW = (D, H, W)
self.MPQ = (M, P, Q)
self.padding = (pad_d, pad_h, pad_w)
self.strides = (str_d, str_h, str_w)
self.relu = relu
self.bsum = bsum
self.all_params = (N, C, K, D, H, W, T, R, S, pad_d, pad_h, pad_w, str_d, str_h, str_w)
self.dimI = (C, D, H, W, N)
self.dimF = (C, T, R, S, K)
self.dimFb = (K, T, R, S, C)
self.dimO = (K, M, P, Q, N)
self.dimI2 = (C*D*H*W, N)
self.dimF2 = (C*T*R*S, K)
self.dimF2t = (K, C*T*R*S)
self.dimO2 = (K*M*P*Q, N)
self.dimS = (K, 1)
self.sizeI = reduce(mul, self.dimI, 1)
self.sizeF = reduce(mul, self.dimF, 1)
self.sizeO = reduce(mul, self.dimO, 1)
self.nOut = reduce(mul, self.MPQ, 1) * K
# flop count for benchmarking
self.flops = P*Q*M*K*N*C*R*S*T * 2.0
####### Cuda C ###########
if lib.use_cudac_kernels:
#3D conv not supported yet
if T > 1 or D > 1:
raise ValueError("3D Convolution not supported by CUDA C kernels.")
self.fprop_kernels = convolution.FpropCuda(lib, self.dtype, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w, bsum=bsum)
# TODO small C bprop?
self.bprop_kernels = convolution.BpropCuda(lib, self.dtype, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w, bsum=bsum)
self.updat_kernels = convolution.UpdateCuda(lib, self.dtype, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w)
####### Winograd ###########
elif lib.have_winograd and R == 3 and S == 3 and all(x == 1 for x in (D,M,T,str_w,str_h,str_d)):
from winograd.convolution import FpropWinograd, BpropWinograd, UpdateWinograd
if N >= 64 and C < 8 or not lib.have_winograd:
self.fprop_kernels = convolution.FpropDirect(
lib, self.dtype, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w, relu, bsum)
else:
self.fprop_kernels = FpropWinograd(
lib, self.dtype, N, C, K, H, W, P, Q, pad_h, pad_w, relu, bsum)
self.bprop_kernels = BpropWinograd(
lib, self.dtype, N, C, K, H, W, P, Q, pad_h, pad_w, relu, bsum)
if N >=32 and C < 8 or not lib.have_winograd:
self.updat_kernels = convolution.UpdateDirect(
lib, self.dtype, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w)
else:
self.updat_kernels = UpdateWinograd(
lib, self.dtype, N, C, K, H, W, P, Q, pad_h, pad_w)
####### Direct ###########
else:
vec_size = 4 if self.dtype.itemsize == 4 else 8
self.fprop_kernels = convolution.FpropDirect(
lib, self.dtype, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w, relu, bsum)
if C % vec_size == 0:
self.bprop_kernels = convolution.BpropDirect(
lib, self.dtype, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w, relu, bsum)
else:
# special kernel for deconv into first layer
self.bprop_kernels = convolution.BpropDirectSmallC(
lib, self.dtype, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w)
self.updat_kernels = convolution.UpdateDirect(
lib, self.dtype, N, C, K, D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d, str_h, str_w)
logger.debug("%s: %s, %s, %s", str(self), str(self.fprop_kernels), str(self.bprop_kernels), str(self.updat_kernels))
def __init__(self,
lib,
dtype,
N,
C,
K,
D=1,
H=1,
W=1,
T=1,
R=1,
S=1,
pad_d=0,
pad_h=0,
pad_w=0,
str_d=1,
str_h=1,
str_w=1,
bsum=False):
super(ConvLayer, self).__init__(lib, dtype, N, np.float32)
# Compute the output spatial dimensions
M = lib.output_dim(D, T, pad_d, str_d)
P = lib.output_dim(H, R, pad_h, str_h)
Q = lib.output_dim(W, S, pad_w, str_w)
self.C = C
self.K = K
self.M = M
self.P = P
self.Q = Q
self.NCK = (N, C, K)
self.TRS = (T, R, S)
self.DHW = (D, H, W)
self.MPQ = (M, P, Q)
self.padding = (pad_d, pad_h, pad_w)
self.strides = (str_d, str_h, str_w)
self.bsum = bsum
self.all_params = (N, C, K, D, H, W, T, R, S, pad_d, pad_h, pad_w,
str_d, str_h, str_w)
self.dimI = (C, D, H, W, N)
self.dimF = (C, T, R, S, K)
self.dimFb = (K, T, R, S, C)
self.dimO = (K, M, P, Q, N)
self.dimI2 = (C * D * H * W, N)
self.dimF2 = (C * T * R * S, K)
self.dimF2t = (K, C * T * R * S)
self.dimO2 = (K * M * P * Q, N)
self.dimS = (K, 1)
self.sizeI = reduce(mul, self.dimI, 1)
self.sizeF = reduce(mul, self.dimF, 1)
self.sizeO = reduce(mul, self.dimO, 1)
self.nOut = reduce(mul, self.MPQ, 1) * K
# flop count for benchmarking
self.flops = P * Q * M * K * N * C * R * S * T * 2.0
if T > 1 or D > 1:
raise ValueError("3D Convolution not supported by CUDA C kernels.")
self.fprop_kernels = convolution.FpropCuda(lib,
self.dtype,
N,
C,
K,
D,
H,
W,
T,
R,
S,
M,
P,
Q,
pad_d,
pad_h,
pad_w,
str_d,
str_h,
str_w,
bsum=bsum)
# TODO small C bprop?
self.bprop_kernels = convolution.BpropCuda(lib,
self.dtype,
N,
C,
K,
D,
H,
W,
T,
R,
S,
M,
P,
Q,
pad_d,
pad_h,
pad_w,
str_d,
str_h,
str_w,
bsum=bsum)
self.updat_kernels = convolution.UpdateCuda(lib, self.dtype, N, C, K,
D, H, W, T, R, S, M, P, Q,
pad_d, pad_h, pad_w, str_d,
str_h, str_w)