entry_point
stringlengths 1
65
| original_triton_python_code
stringlengths 208
619k
| optimised_triton_code
stringlengths 1.15k
275k
| repo_name
stringlengths 7
115
| module_name
stringlengths 1
65
| synthetic
bool 1
class | uuid
int64 0
18.5k
| licenses
listlengths 1
6
| stars
int64 0
19.8k
| sha
stringlengths 40
40
| repo_link
stringlengths 72
180
|
|---|---|---|---|---|---|---|---|---|---|---|
BinaryReg
|
import torch
import torch.nn as nn
import torch.utils.data
class BinaryReg(nn.Module):
"""Regularization for encouraging the outputs to be binary.
"""
def __init__(self, alpha=1.0):
super().__init__()
self.alpha = alpha
def forward(self, input):
diff = input - 0.5
diff = torch.clamp(torch.abs(diff), min=0.01)
loss = 1.0 / diff.sum()
return self.alpha * loss
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
import torch.utils.data
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_abs_clamp_mul_reciprocal_sub_sum_0(in_out_ptr0,
in_ptr0, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = 0.5
tmp2 = tmp0 - tmp1
tmp3 = tl_math.abs(tmp2)
tmp4 = 0.01
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp6 = tl.broadcast_to(tmp5, [RBLOCK])
tmp8 = triton_helpers.promote_to_tensor(tl.sum(tmp6, 0))
tmp9 = tl.full([1], 1, tl.int32)
tmp10 = tmp9 / tmp8
tmp11 = 1.0
tmp12 = tmp10 * tmp11
tmp13 = tmp12 * tmp11
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp13, None)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_abs_clamp_mul_reciprocal_sub_sum_0[grid(1)](buf1,
arg0_1, 1, 256, num_warps=2, num_stages=1)
del arg0_1
return buf1,
class BinaryRegNew(nn.Module):
"""Regularization for encouraging the outputs to be binary.
"""
def __init__(self, alpha=1.0):
super().__init__()
self.alpha = alpha
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
pragyasingh7/pytorch_connectomics
|
BinaryReg
| false
| 4,134
|
[
"MIT"
] | 0
|
fdc8e1900b0a38d19ea50f78f8c81da2a4f015a9
|
https://github.com/pragyasingh7/pytorch_connectomics/tree/fdc8e1900b0a38d19ea50f78f8c81da2a4f015a9
|
DepthWiseSeparableConvBlock
|
import torch
import torch.nn as nn
class DepthWiseSeparableConvBlock(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, stride=1,
padding=0, dilation=1, bias=True, padding_mode='zeros',
inner_kernel_size=1, inner_stride=1, inner_padding=0):
"""Depthwise separable 2D Convolution.
:param in_channels: Input channels.
:type in_channels: int
:param out_channels: Output channels.
:type out_channels: int
:param kernel_size: Kernel shape/size.
:type kernel_size: int|tuple|list
:param stride: Stride.
:type stride: int|tuple|list
:param padding: Padding.
:type padding: int|tuple|list
:param dilation: Dilation.
:type dilation: int
:param bias: Bias.
:type bias: bool
:param padding_mode: Padding mode.
:type padding_mode: str
:param inner_kernel_size: Kernel shape/size of the second convolution.
:type inner_kernel_size: int|tuple|list
:param inner_stride: Inner stride.
:type inner_stride: int|tuple|list
:param inner_padding: Inner padding.
:type inner_padding: int|tuple|list
"""
super(DepthWiseSeparableConvBlock, self).__init__()
self.depth_wise_conv: 'nn.Module' = nn.Conv2d(in_channels=
in_channels, out_channels=out_channels, kernel_size=kernel_size,
stride=stride, padding=padding, dilation=dilation, groups=
in_channels if out_channels >= in_channels else out_channels,
bias=bias, padding_mode=padding_mode)
self.non_linearity: 'nn.Module' = nn.LeakyReLU()
self.point_wise: 'nn.Module' = nn.Conv2d(in_channels=out_channels,
out_channels=out_channels, kernel_size=inner_kernel_size,
stride=inner_stride, padding=inner_padding, dilation=1, groups=
1, bias=bias, padding_mode=padding_mode)
if inner_kernel_size != 1:
None
None
raise ValueError
self.layers = nn.Sequential(self.depth_wise_conv, self.
non_linearity, self.point_wise)
def forward(self, x):
"""Forward pass of the module.
:param x: Input tensor.
:type x: torch.Tensor
:return: Output tensor.
:rtype: torch.Tensor
"""
return self.layers(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4, 'kernel_size': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_convolution_leaky_relu_0(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp5 = 0.01
tmp6 = tmp2 * tmp5
tmp7 = tl.where(tmp4, tmp2, tmp6)
tl.store(out_ptr0 + x2, tmp4, xmask)
tl.store(out_ptr1 + x2, tmp7, xmask)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 1, 4, 4), (16, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=4, bias=None)
assert_size_stride(buf0, (4, 4, 1, 1), (4, 1, 1, 1))
buf1 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.bool)
buf2 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_convolution_leaky_relu_0[grid(16)](buf0, primals_2,
buf1, buf2, 16, XBLOCK=16, num_warps=1, num_stages=1)
del buf0
del primals_2
buf3 = extern_kernels.convolution(buf2, primals_4, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf3, (4, 4, 1, 1), (4, 1, 1, 1))
buf4 = buf3
del buf3
triton_poi_fused_convolution_1[grid(16)](buf4, primals_5, 16,
XBLOCK=16, num_warps=1, num_stages=1)
del primals_5
return buf4, primals_1, primals_3, primals_4, buf1, buf2
class DepthWiseSeparableConvBlockNew(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, stride=1,
padding=0, dilation=1, bias=True, padding_mode='zeros',
inner_kernel_size=1, inner_stride=1, inner_padding=0):
"""Depthwise separable 2D Convolution.
:param in_channels: Input channels.
:type in_channels: int
:param out_channels: Output channels.
:type out_channels: int
:param kernel_size: Kernel shape/size.
:type kernel_size: int|tuple|list
:param stride: Stride.
:type stride: int|tuple|list
:param padding: Padding.
:type padding: int|tuple|list
:param dilation: Dilation.
:type dilation: int
:param bias: Bias.
:type bias: bool
:param padding_mode: Padding mode.
:type padding_mode: str
:param inner_kernel_size: Kernel shape/size of the second convolution.
:type inner_kernel_size: int|tuple|list
:param inner_stride: Inner stride.
:type inner_stride: int|tuple|list
:param inner_padding: Inner padding.
:type inner_padding: int|tuple|list
"""
super(DepthWiseSeparableConvBlockNew, self).__init__()
self.depth_wise_conv: 'nn.Module' = nn.Conv2d(in_channels=
in_channels, out_channels=out_channels, kernel_size=kernel_size,
stride=stride, padding=padding, dilation=dilation, groups=
in_channels if out_channels >= in_channels else out_channels,
bias=bias, padding_mode=padding_mode)
self.non_linearity: 'nn.Module' = nn.LeakyReLU()
self.point_wise: 'nn.Module' = nn.Conv2d(in_channels=out_channels,
out_channels=out_channels, kernel_size=inner_kernel_size,
stride=inner_stride, padding=inner_padding, dilation=1, groups=
1, bias=bias, padding_mode=padding_mode)
if inner_kernel_size != 1:
None
None
raise ValueError
self.layers = nn.Sequential(self.depth_wise_conv, self.
non_linearity, self.point_wise)
def forward(self, input_0):
primals_1 = self.depth_wise_conv.weight
primals_2 = self.depth_wise_conv.bias
primals_4 = self.point_wise.weight
primals_5 = self.point_wise.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
pppyykknen/LFDisplay-PyTorch
|
DepthWiseSeparableConvBlock
| false
| 4,135
|
[
"MIT"
] | 0
|
d19261dac1717a799bb5ba5f96563be1d2383340
|
https://github.com/pppyykknen/LFDisplay-PyTorch/tree/d19261dac1717a799bb5ba5f96563be1d2383340
|
MDN
|
from torch.nn import Module
import torch
from torch.nn.modules import Module
from torch.nn.modules import Linear
class MDN(Module):
def __init__(self, input_size, num_mixtures):
super(MDN, self).__init__()
self.input_size = input_size
self.num_mixtures = num_mixtures
self.parameter_layer = Linear(in_features=input_size, out_features=
1 + 6 * num_mixtures)
def forward(self, input_, bias=None):
mixture_parameters = self.parameter_layer(input_)
eos_hat = mixture_parameters[:, :, 0:1]
pi_hat, mu1_hat, mu2_hat, sigma1_hat, sigma2_hat, rho_hat = (torch.
chunk(mixture_parameters[:, :, 1:], 6, dim=2))
eos = torch.sigmoid(-eos_hat)
mu1 = mu1_hat
mu2 = mu2_hat
rho = torch.tanh(rho_hat)
if bias is None:
bias = torch.zeros_like(rho)
pi = torch.softmax(pi_hat * (1 + bias), dim=2)
sigma1 = torch.exp(sigma1_hat - bias)
sigma2 = torch.exp(sigma2_hat - bias)
return eos, pi, mu1, mu2, sigma1, sigma2, rho
def __repr__(self):
s = '{name}(input_size={input_size}, num_mixtures={num_mixtures})'
return s.format(name=self.__class__.__name__, **self.__dict__)
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'input_size': 4, 'num_mixtures': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch.nn import Module
from torch.nn.modules import Module
from torch.nn.modules import Linear
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_neg_sigmoid_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 25 * x0, xmask, eviction_policy='evict_last')
tmp1 = -tmp0
tmp2 = tl.sigmoid(tmp1)
tl.store(out_ptr0 + x0, tmp2, xmask)
@triton.jit
def triton_poi_fused_tanh_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (21 + x0 + 25 * x1), xmask)
tmp1 = libdevice.tanh(tmp0)
tl.store(out_ptr0 + x2, tmp1, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (1 + x0 + 25 * x1), xmask)
tmp1 = tl.load(in_ptr0 + (1 + 25 * x1), xmask, eviction_policy='evict_last'
)
tmp2 = tl.load(in_ptr0 + (2 + 25 * x1), xmask, eviction_policy='evict_last'
)
tmp4 = tl.load(in_ptr0 + (3 + 25 * x1), xmask, eviction_policy='evict_last'
)
tmp6 = tl.load(in_ptr0 + (4 + 25 * x1), xmask, eviction_policy='evict_last'
)
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_exp_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (13 + x0 + 25 * x1), xmask)
tmp1 = tl_math.exp(tmp0)
tl.store(out_ptr0 + x2, tmp1, xmask)
@triton.jit
def triton_poi_fused_exp_5(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (17 + x0 + 25 * x1), xmask)
tmp1 = tl_math.exp(tmp0)
tl.store(out_ptr0 + x2, tmp1, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (25, 4), (4, 1))
assert_size_stride(primals_2, (25,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 25), (25, 1), torch.float32)
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (16,
4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 25), (1, 4),
0), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((4, 4, 1), (4, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_neg_sigmoid_0[grid(16)](buf0, buf1, 16, XBLOCK=16,
num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_tanh_1[grid(64)](buf0, buf2, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf3 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused__softmax_2[grid(64)](buf0, buf3, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf4 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused__softmax_3[grid(64)](buf3, buf4, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf5 = buf3
del buf3
triton_poi_fused_exp_4[grid(64)](buf0, buf5, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf6 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_exp_5[grid(64)](buf0, buf6, 64, XBLOCK=64,
num_warps=1, num_stages=1)
return buf1, buf4, reinterpret_tensor(buf0, (4, 4, 4), (100, 25, 1), 5
), reinterpret_tensor(buf0, (4, 4, 4), (100, 25, 1), 9
), buf5, buf6, buf2, reinterpret_tensor(primals_3, (16, 4), (4, 1), 0
), buf1, buf2, buf4, buf5, buf6
class MDNNew(Module):
def __init__(self, input_size, num_mixtures):
super(MDNNew, self).__init__()
self.input_size = input_size
self.num_mixtures = num_mixtures
self.parameter_layer = Linear(in_features=input_size, out_features=
1 + 6 * num_mixtures)
def __repr__(self):
s = '{name}(input_size={input_size}, num_mixtures={num_mixtures})'
return s.format(name=self.__class__.__name__, **self.__dict__)
def forward(self, input_0):
primals_1 = self.parameter_layer.weight
primals_2 = self.parameter_layer.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0], output[1], output[2], output[3], output[4], output[5
], output[6]
|
poctaviano/Handwriting-Model
|
MDN
| false
| 4,136
|
[
"MIT"
] | 0
|
30311ea0f4cb6e7bc0114cf0b2a96dc915dd9795
|
https://github.com/poctaviano/Handwriting-Model/tree/30311ea0f4cb6e7bc0114cf0b2a96dc915dd9795
|
KARAttention
|
from _paritybench_helpers import _mock_config
import math
import torch
from torch import nn
class KARMultiHeadAttention(nn.Module):
def __init__(self, config, hidden_size):
super(KARMultiHeadAttention, self).__init__()
if hidden_size % config.num_attention_heads != 0:
raise ValueError(
'The hidden size (%d) is not a multiple of the number of attention heads (%d)'
% (hidden_size, config.num_attention_heads))
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(hidden_size / config.num_attention_heads
)
self.all_head_size = (self.num_attention_heads * self.
attention_head_size)
self.query = nn.Linear(hidden_size, self.all_head_size)
self.key = nn.Linear(hidden_size, self.all_head_size)
self.value = nn.Linear(hidden_size, self.all_head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
self.config = config
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, hidden_states, attention_mask, s_i_tag_in):
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(s_i_tag_in)
mixed_value_layer = self.value(s_i_tag_in)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.
all_head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
return context_layer
class KARAttention(nn.Module):
def __init__(self, config, kar_size):
super(KARAttention, self).__init__()
self.multihead = KARMultiHeadAttention(config, kar_size)
def forward(self, input_tensor, attention_mask, s_m_tag):
attention_output = self.multihead(input_tensor, attention_mask, s_m_tag
)
return attention_output
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4, 4])
]
def get_init_inputs():
return [[], {'config': _mock_config(num_attention_heads=4,
attention_probs_dropout_prob=0.5), 'kar_size': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import math
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(out_ptr0 + (x2 + 4 * y3), tmp4, xmask & ymask)
@triton.jit
def triton_poi_fused_1(in_ptr0, in_ptr1, out_ptr0, out_ptr1, out_ptr2,
xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 16
tmp0 = tl.load(in_ptr0 + 4 * x2, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x2), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x2), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = triton_helpers.maximum(tmp2, tmp5)
tmp9 = tmp7 + tmp8
tmp10 = triton_helpers.maximum(tmp6, tmp9)
tmp13 = tmp11 + tmp12
tmp14 = triton_helpers.maximum(tmp10, tmp13)
tmp15 = tmp2 - tmp14
tmp16 = tl_math.exp(tmp15)
tmp17 = tmp5 - tmp14
tmp18 = tl_math.exp(tmp17)
tmp19 = tmp16 + tmp18
tmp20 = tmp9 - tmp14
tmp21 = tl_math.exp(tmp20)
tmp22 = tmp19 + tmp21
tmp23 = tmp13 - tmp14
tmp24 = tl_math.exp(tmp23)
tmp25 = tmp22 + tmp24
tmp26 = float('-inf')
tmp27 = tmp2 == tmp26
tmp28 = tmp27 == 0
tmp29 = tmp28.to(tl.int64)
tmp30 = tmp29 != 0
tmp31 = tmp5 == tmp26
tmp32 = tmp31 == 0
tmp33 = tmp32.to(tl.int64)
tmp34 = tmp33 != 0
tmp35 = tmp30 | tmp34
tmp36 = tmp9 == tmp26
tmp37 = tmp36 == 0
tmp38 = tmp37.to(tl.int64)
tmp39 = tmp38 != 0
tmp40 = tmp35 | tmp39
tmp41 = tmp13 == tmp26
tmp42 = tmp41 == 0
tmp43 = tmp42.to(tl.int64)
tmp44 = tmp43 != 0
tmp45 = tmp40 | tmp44
tl.store(out_ptr0 + x2, tmp14, xmask)
tl.store(out_ptr1 + x2, tmp25, xmask)
tl.store(out_ptr2 + x2, tmp45, xmask)
@triton.jit
def triton_poi_fused_2(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, in_ptr3,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex // 4
x4 = xindex
x5 = xindex % 64
tmp0 = tl.load(in_ptr0 + x3, xmask, eviction_policy='evict_last').to(tl
.int1)
tmp2 = tl.load(in_out_ptr0 + x4, xmask)
tmp3 = tl.load(in_ptr1 + x5, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr2 + x3, xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr3 + x3, xmask, eviction_policy='evict_last')
tmp1 = tmp0 == 0
tmp4 = tmp2 + tmp3
tmp6 = tmp4 - tmp5
tmp7 = tl_math.exp(tmp6)
tmp9 = tmp7 / tmp8
tmp10 = 0.0
tmp11 = tl.where(tmp1, tmp10, tmp9)
tl.store(in_out_ptr0 + x4, tmp11, xmask)
@triton.jit
def triton_poi_fused_3(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_7, (4, 4), (4, 1))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_6, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 4), (1, 4), 0), out=buf1)
del primals_4
buf2 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_6, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 4), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(16, 4)](buf0, primals_2, buf3, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_2
buf4 = reinterpret_tensor(buf0, (4, 4, 1, 4), (16, 4, 4, 1), 0)
del buf0
triton_poi_fused_0[grid(16, 4)](buf1, primals_5, buf4, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_5
buf5 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf4, (16, 1, 4), (4, 0, 1), 0), out=buf5)
buf6 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 64), 0)
del buf1
buf7 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf8 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.bool)
triton_poi_fused_1[grid(64)](buf5, primals_9, buf6, buf7, buf8, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf9 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused_2[grid(256)](buf9, buf8, primals_9, buf6, buf7,
256, XBLOCK=128, num_warps=4, num_stages=1)
del buf8
del primals_9
buf10 = reinterpret_tensor(buf7, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf7
triton_poi_fused_3[grid(16, 4)](buf2, primals_8, buf10, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_8
buf11 = reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf9, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf10, (16, 4, 1), (4, 1, 0), 0), out=buf11)
buf12 = reinterpret_tensor(buf6, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf6
triton_poi_fused_clone_4[grid(16, 4)](buf11, buf12, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
del buf11
return reinterpret_tensor(buf12, (4, 4, 4), (16, 4, 1), 0
), reinterpret_tensor(primals_3, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_6, (16, 4), (4, 1), 0
), buf9, reinterpret_tensor(buf10, (16, 1, 4), (4, 1, 1), 0
), reinterpret_tensor(buf3, (16, 1, 4), (4, 1, 1), 0
), reinterpret_tensor(buf4, (16, 4, 1), (4, 1, 4), 0)
class KARMultiHeadAttention(nn.Module):
def __init__(self, config, hidden_size):
super(KARMultiHeadAttention, self).__init__()
if hidden_size % config.num_attention_heads != 0:
raise ValueError(
'The hidden size (%d) is not a multiple of the number of attention heads (%d)'
% (hidden_size, config.num_attention_heads))
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(hidden_size / config.num_attention_heads
)
self.all_head_size = (self.num_attention_heads * self.
attention_head_size)
self.query = nn.Linear(hidden_size, self.all_head_size)
self.key = nn.Linear(hidden_size, self.all_head_size)
self.value = nn.Linear(hidden_size, self.all_head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
self.config = config
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, hidden_states, attention_mask, s_i_tag_in):
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(s_i_tag_in)
mixed_value_layer = self.value(s_i_tag_in)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.
all_head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
return context_layer
class KARAttentionNew(nn.Module):
def __init__(self, config, kar_size):
super(KARAttentionNew, self).__init__()
self.multihead = KARMultiHeadAttention(config, kar_size)
def forward(self, input_0, input_1, input_2):
primals_1 = self.multihead.query.weight
primals_2 = self.multihead.query.bias
primals_4 = self.multihead.key.weight
primals_5 = self.multihead.key.bias
primals_7 = self.multihead.value.weight
primals_8 = self.multihead.value.bias
primals_3 = input_0
primals_6 = input_1
primals_9 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0]
|
ohadrozen/inferbert
|
KARAttention
| false
| 4,137
|
[
"Apache-2.0"
] | 0
|
2e450aba894937e5769dcf028e4a8a597991fe43
|
https://github.com/ohadrozen/inferbert/tree/2e450aba894937e5769dcf028e4a8a597991fe43
|
AttentiveTrans2d
|
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.data
class AttentiveTrans2d(nn.Module):
def __init__(self, num_features, hidden_channels=32):
super(AttentiveTrans2d, self).__init__()
self.avgpool = nn.AdaptiveAvgPool2d(1)
self.smooth_gamma = 1
self.smooth_beta = 0
self.matrix1 = nn.Parameter(torch.ones(num_features, hidden_channels))
self.matrix2 = nn.Parameter(torch.ones(hidden_channels, num_features))
self.sigmoid = nn.Sigmoid()
self.conv1 = nn.Conv2d(2, 1, 3, padding=1, bias=False)
self.conv2 = nn.Conv2d(1, 1, 3, padding=1, bias=False)
self.conv3 = nn.Conv2d(num_features, num_features, 1, bias=False)
self.conv4 = nn.Conv2d(num_features, num_features, 1, bias=False)
self.IN_norm = nn.InstanceNorm2d(num_features, affine=False,
track_running_stats=False)
def forward(self, feature):
output = self.IN_norm(feature)
feature_nc = self.avgpool(feature).view(feature.size()[0], feature.
size()[1])
channel_wise_response = self.sigmoid(feature_nc @ self.matrix1
) @ self.matrix2
channel_wise_response = channel_wise_response.unsqueeze(-1).unsqueeze(
-1).expand(output.size())
avg_out = F.adaptive_avg_pool3d(feature, (1, feature.size()[2],
feature.size()[3]))
max_out = F.adaptive_max_pool3d(feature, (1, feature.size()[2],
feature.size()[3]))
avg_max_concat = torch.cat([avg_out, max_out], dim=1)
spatial_wise_response = self.conv2(self.sigmoid(self.conv1(
avg_max_concat))).expand(output.size())
pixel_wise_response = channel_wise_response * spatial_wise_response
importance_gamma = self.conv3(pixel_wise_response) + self.smooth_gamma
importance_beta = self.conv4(pixel_wise_response) + self.smooth_beta
out_in = output * importance_gamma + importance_beta
return out_in
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'num_features': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
import torch.utils.data
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused__native_batch_norm_legit_mean_view_0(in_out_ptr0,
in_out_ptr1, in_ptr0, out_ptr0, xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 16
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 16 * x0), xmask, other=0.0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp3 = tl.where(xmask, tmp1, 0)
tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp6 = tl.where(xmask, tmp4, 0)
tmp7 = tl.sum(tmp6, 1)[:, None]
tmp8 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp9 = tmp8.to(tl.float32)
tmp10 = tmp7 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tl.broadcast_to(tmp12, [XBLOCK, RBLOCK])
tmp15 = tl.where(xmask, tmp13, 0)
tmp16 = tl.sum(tmp15, 1)[:, None]
tmp18 = tl.sum(tmp3, 1)[:, None]
tmp19 = 16.0
tmp20 = tmp16 / tmp19
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tmp24 = tmp18 / tmp19
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp23, xmask)
tl.debug_barrier()
tl.store(in_out_ptr1 + x0, tmp24, xmask)
tl.store(out_ptr0 + x0, tmp10, xmask)
@triton.jit
def triton_poi_fused_sigmoid_1(in_out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.sigmoid(tmp0)
tl.store(in_out_ptr0 + x0, tmp1, xmask)
@triton.jit
def triton_poi_fused_cat_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 16 % 2
x0 = xindex % 16
x2 = xindex // 32
x3 = xindex
tmp0 = x1
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + 16 * x2), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 2, tl.int64)
tmp9 = tl.load(in_ptr1 + (x0 + 16 * x2), tmp6 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x3, tmp10, xmask)
@triton.jit
def triton_poi_fused_sigmoid_3(in_out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.sigmoid(tmp0)
tl.store(in_out_ptr0 + x0, tmp1, xmask)
@triton.jit
def triton_poi_fused_mul_4(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex // 16
x0 = xindex % 16
x2 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + x3, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x4, tmp2, xmask)
@triton.jit
def triton_poi_fused_add_mul_5(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2,
in_ptr3, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 16
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_out_ptr0 + x2, xmask)
tmp9 = tl.load(in_ptr3 + x2, xmask)
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp4 * tmp7
tmp10 = 0.0
tmp11 = tmp9 + tmp10
tmp12 = tmp8 + tmp11
tl.store(in_out_ptr0 + x2, tmp12, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 32), (32, 1))
assert_size_stride(primals_3, (32, 4), (4, 1))
assert_size_stride(primals_4, (1, 2, 3, 3), (18, 9, 3, 1))
assert_size_stride(primals_5, (1, 1, 3, 3), (9, 9, 3, 1))
assert_size_stride(primals_6, (4, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_7, (4, 4, 1, 1), (4, 1, 1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 1, 1), torch.float32)
buf1 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch.float32
)
buf4 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 16, 16), torch.float32)
buf3 = reinterpret_tensor(buf1, (1, 16, 1, 1), (16, 1, 1, 1), 0)
del buf1
buf5 = reinterpret_tensor(buf4, (4, 4), (4, 1), 0)
del buf4
get_raw_stream(0)
triton_per_fused__native_batch_norm_legit_mean_view_0[grid(16)](buf3,
buf5, primals_1, buf0, 16, 16, XBLOCK=8, num_warps=2, num_stages=1)
buf6 = empty_strided_cuda((4, 32), (32, 1), torch.float32)
extern_kernels.mm(buf5, primals_2, out=buf6)
del primals_2
buf7 = buf6
del buf6
triton_poi_fused_sigmoid_1[grid(128)](buf7, 128, XBLOCK=128,
num_warps=4, num_stages=1)
buf8 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.mm(buf7, primals_3, out=buf8)
buf9 = torch.ops.aten._adaptive_avg_pool3d.default(primals_1, [1, 4, 4]
)
buf10 = buf9
del buf9
buf11 = torch.ops.aten.adaptive_max_pool3d.default(primals_1, [1, 4, 4]
)
buf12 = buf11[0]
del buf11
buf14 = empty_strided_cuda((4, 2, 4, 4), (32, 16, 4, 1), torch.float32)
triton_poi_fused_cat_2[grid(128)](buf10, buf12, buf14, 128, XBLOCK=
128, num_warps=4, num_stages=1)
del buf10
del buf12
buf15 = extern_kernels.convolution(buf14, primals_4, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf15, (4, 1, 4, 4), (16, 16, 4, 1))
buf16 = buf15
del buf15
triton_poi_fused_sigmoid_3[grid(64)](buf16, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf17 = extern_kernels.convolution(buf16, primals_5, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf17, (4, 1, 4, 4), (16, 16, 4, 1))
buf18 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_mul_4[grid(256)](buf8, buf17, buf18, 256, XBLOCK=
128, num_warps=4, num_stages=1)
buf19 = extern_kernels.convolution(buf18, primals_6, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf19, (4, 4, 4, 4), (64, 16, 4, 1))
buf20 = extern_kernels.convolution(buf18, primals_7, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf20, (4, 4, 4, 4), (64, 16, 4, 1))
buf21 = buf19
del buf19
triton_poi_fused_add_mul_5[grid(256)](buf21, primals_1, buf0, buf3,
buf20, 256, XBLOCK=128, num_warps=4, num_stages=1)
del buf20
return (buf21, primals_1, primals_4, primals_5, primals_6, primals_7,
buf0, buf3, buf7, buf8, buf14, buf16, buf17, buf18,
reinterpret_tensor(primals_3, (4, 32), (1, 4), 0),
reinterpret_tensor(buf5, (4, 4), (1, 4), 0))
class AttentiveTrans2dNew(nn.Module):
def __init__(self, num_features, hidden_channels=32):
super(AttentiveTrans2dNew, self).__init__()
self.avgpool = nn.AdaptiveAvgPool2d(1)
self.smooth_gamma = 1
self.smooth_beta = 0
self.matrix1 = nn.Parameter(torch.ones(num_features, hidden_channels))
self.matrix2 = nn.Parameter(torch.ones(hidden_channels, num_features))
self.sigmoid = nn.Sigmoid()
self.conv1 = nn.Conv2d(2, 1, 3, padding=1, bias=False)
self.conv2 = nn.Conv2d(1, 1, 3, padding=1, bias=False)
self.conv3 = nn.Conv2d(num_features, num_features, 1, bias=False)
self.conv4 = nn.Conv2d(num_features, num_features, 1, bias=False)
self.IN_norm = nn.InstanceNorm2d(num_features, affine=False,
track_running_stats=False)
def forward(self, input_0):
primals_2 = self.matrix1
primals_3 = self.matrix2
primals_4 = self.conv1.weight
primals_5 = self.conv2.weight
primals_6 = self.conv3.weight
primals_7 = self.conv4.weight
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
|
ppomelo/Attentive-Transformation-Based-Normalization
|
AttentiveTrans2d
| false
| 4,138
|
[
"Apache-2.0"
] | 0
|
62ad02eb025613e90f4fe0e0a9f0f85839e53092
|
https://github.com/ppomelo/Attentive-Transformation-Based-Normalization/tree/62ad02eb025613e90f4fe0e0a9f0f85839e53092
|
DepthLogLoss
|
import torch
import torch.nn as nn
class DepthLogLoss(nn.Module):
def __init__(self, balance_factor):
super(DepthLogLoss, self).__init__()
self.balance_factor = balance_factor
def forward(self, inputs, targets):
n, _, h, w = inputs.shape
n_pixel = n * h * w
inputs = torch.log(inputs + 1e-08)
targets = torch.log(targets)
d = inputs - targets
loss = torch.sum(d ** 2) / n_pixel - self.balance_factor * torch.sum(d
) ** 2 / n_pixel ** 2
return loss
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'balance_factor': 4}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_add_div_log_mul_pow_sub_sum_0(in_out_ptr0, in_ptr0,
in_ptr1, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp4 = tl.load(in_ptr1 + r0, None)
tmp1 = 1e-08
tmp2 = tmp0 + tmp1
tmp3 = tl_math.log(tmp2)
tmp5 = tl_math.log(tmp4)
tmp6 = tmp3 - tmp5
tmp7 = tmp6 * tmp6
tmp8 = tl.broadcast_to(tmp7, [RBLOCK])
tmp10 = triton_helpers.promote_to_tensor(tl.sum(tmp8, 0))
tmp11 = tl.broadcast_to(tmp6, [RBLOCK])
tmp13 = triton_helpers.promote_to_tensor(tl.sum(tmp11, 0))
tmp14 = 0.015625
tmp15 = tmp10 * tmp14
tmp16 = tmp13 * tmp13
tmp17 = 4.0
tmp18 = tmp16 * tmp17
tmp19 = 0.000244140625
tmp20 = tmp18 * tmp19
tmp21 = tmp15 - tmp20
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp21, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf2 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_add_div_log_mul_pow_sub_sum_0[grid(1)](buf2,
arg0_1, arg1_1, 1, 256, num_warps=2, num_stages=1)
del arg0_1
del arg1_1
return buf2,
class DepthLogLossNew(nn.Module):
def __init__(self, balance_factor):
super(DepthLogLossNew, self).__init__()
self.balance_factor = balance_factor
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
|
pystokes/depth_estimation
|
DepthLogLoss
| false
| 4,140
|
[
"MIT"
] | 0
|
b5b1955bcb5b3f1a1f1c8ddde45431cf38514f90
|
https://github.com/pystokes/depth_estimation/tree/b5b1955bcb5b3f1a1f1c8ddde45431cf38514f90
|
ConditionalBottleNeck
|
from _paritybench_helpers import _mock_config
import torch
import torch.nn as nn
class FiLM(nn.Module):
""" Feature-wise Linear Modulation (FiLM) layer"""
def __init__(self, input_size, output_size, num_film_layers=1,
layer_norm=False):
"""
:param input_size: feature size of x_cond
:param output_size: feature size of x_to_film
:param layer_norm: true or false
"""
super(FiLM, self).__init__()
self.input_size = input_size
self.output_size = output_size
self.num_film_layers = num_film_layers
self.layer_norm = nn.LayerNorm(output_size) if layer_norm else None
film_output_size = self.output_size * num_film_layers * 2
self.gb_weights = nn.Linear(self.input_size, film_output_size)
self.gb_weights.bias.data.fill_(0)
def forward(self, x_cond, x_to_film):
gb = self.gb_weights(x_cond).unsqueeze(1)
gamma, beta = torch.chunk(gb, 2, dim=-1)
out = (1 + gamma) * x_to_film + beta
if self.layer_norm is not None:
out = self.layer_norm(out)
return out
class ConditionalBottleNeck(nn.Module):
"""Down projection and up projection with FiLM layers within Transformer layer."""
def __init__(self, config):
super(ConditionalBottleNeck, self).__init__()
self.emb_transf = nn.Linear(config.hidden_size, config.hidden_size)
self.hidden_modulation = FiLM(config.hidden_size, config.hidden_size)
self.down_proj_layer = nn.Linear(config.hidden_size, config.
hidden_size // 3)
self.up_proj_layer = nn.Linear(config.hidden_size // 3, config.
hidden_size)
def forward(self, x_cond, hidden_states):
x_cond = self.emb_transf(x_cond)
hidden_states = self.hidden_modulation(x_cond=x_cond, x_to_film=
hidden_states)
hidden_states = self.down_proj_layer(hidden_states)
hidden_states = self.up_proj_layer(hidden_states)
return hidden_states
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'config': _mock_config(hidden_size=4)}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_add_mul_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 16
x3 = xindex // 256
x4 = xindex % 256
x6 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 8 * x1 + 128 * x3), xmask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr2 + x4, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (4 + x0 + 8 * x1 + 128 * x3), xmask,
eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (4 + x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 + tmp3
tmp6 = tmp4 * tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tl.store(out_ptr0 + x6, tmp10, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (8, 4), (4, 1))
assert_size_stride(primals_5, (8,), (1,))
assert_size_stride(primals_6, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_7, (1, 4), (4, 1))
assert_size_stride(primals_8, (1,), (1,))
assert_size_stride(primals_9, (4, 1), (1, 1))
assert_size_stride(primals_10, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (64,
4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0
), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((64, 8), (8, 1), torch.float32)
extern_kernels.mm(buf0, reinterpret_tensor(primals_4, (4, 8), (1, 4
), 0), out=buf1)
buf2 = empty_strided_cuda((4, 4, 4, 4, 4), (256, 64, 16, 4, 1),
torch.float32)
get_raw_stream(0)
triton_poi_fused_add_mul_0[grid(1024)](buf1, primals_5, primals_6,
buf2, 1024, XBLOCK=256, num_warps=4, num_stages=1)
del buf1
del primals_5
buf4 = empty_strided_cuda((256, 1), (1, 1), torch.float32)
extern_kernels.addmm(primals_8, reinterpret_tensor(buf2, (256, 4),
(4, 1), 0), reinterpret_tensor(primals_7, (4, 1), (1, 4), 0),
alpha=1, beta=1, out=buf4)
del primals_8
buf5 = empty_strided_cuda((256, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_10, buf4, reinterpret_tensor(primals_9,
(1, 4), (1, 1), 0), alpha=1, beta=1, out=buf5)
del primals_10
return reinterpret_tensor(buf5, (4, 4, 4, 4, 4), (256, 64, 16, 4, 1), 0
), primals_6, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf0, reinterpret_tensor(buf2, (256, 4), (4, 1), 0
), buf4, primals_9, primals_7, primals_4
class FiLM(nn.Module):
""" Feature-wise Linear Modulation (FiLM) layer"""
def __init__(self, input_size, output_size, num_film_layers=1,
layer_norm=False):
"""
:param input_size: feature size of x_cond
:param output_size: feature size of x_to_film
:param layer_norm: true or false
"""
super(FiLM, self).__init__()
self.input_size = input_size
self.output_size = output_size
self.num_film_layers = num_film_layers
self.layer_norm = nn.LayerNorm(output_size) if layer_norm else None
film_output_size = self.output_size * num_film_layers * 2
self.gb_weights = nn.Linear(self.input_size, film_output_size)
self.gb_weights.bias.data.fill_(0)
def forward(self, x_cond, x_to_film):
gb = self.gb_weights(x_cond).unsqueeze(1)
gamma, beta = torch.chunk(gb, 2, dim=-1)
out = (1 + gamma) * x_to_film + beta
if self.layer_norm is not None:
out = self.layer_norm(out)
return out
class ConditionalBottleNeckNew(nn.Module):
"""Down projection and up projection with FiLM layers within Transformer layer."""
def __init__(self, config):
super(ConditionalBottleNeckNew, self).__init__()
self.emb_transf = nn.Linear(config.hidden_size, config.hidden_size)
self.hidden_modulation = FiLM(config.hidden_size, config.hidden_size)
self.down_proj_layer = nn.Linear(config.hidden_size, config.
hidden_size // 3)
self.up_proj_layer = nn.Linear(config.hidden_size // 3, config.
hidden_size)
def forward(self, input_0, input_1):
primals_1 = self.emb_transf.weight
primals_2 = self.emb_transf.bias
primals_4 = self.hidden_modulation.gb_weights.weight
primals_5 = self.hidden_modulation.gb_weights.bias
primals_7 = self.down_proj_layer.weight
primals_8 = self.down_proj_layer.bias
primals_9 = self.up_proj_layer.weight
primals_10 = self.up_proj_layer.bias
primals_3 = input_0
primals_6 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9, primals_10])
return output[0]
|
Daupler/CA-MTL
|
ConditionalBottleNeck
| false
| 4,141
|
[
"MIT"
] | 0
|
d417b039dee973e32f42ba5c1c346738cd29ab3c
|
https://github.com/Daupler/CA-MTL/tree/d417b039dee973e32f42ba5c1c346738cd29ab3c
|
TextureFinder
|
import torch
import torch.nn as nn
import torch.nn.functional as F
class TextureFinder(nn.Module):
def __init__(self):
super(TextureFinder, self).__init__()
self.encoder_conv1 = nn.Conv2d(in_channels=1, out_channels=4,
kernel_size=4, stride=2, padding=1, dilation=1, groups=1, bias=True
)
self.encoder_conv1.bias.data.zero_()
self.encoder_conv1.weight.data[:, :, :, :] = 1 / 0.32 + torch.normal(
mean=torch.tensor(0.0), std=torch.tensor(0.0001))
self.encoder_normalization1 = nn.GroupNorm(1, 4, eps=1e-05, affine=True
)
self.encoder_conv2 = nn.Conv2d(in_channels=4, out_channels=16,
kernel_size=4, stride=2, padding=1, dilation=1, groups=1, bias=True
)
self.encoder_conv2.bias.data.zero_()
self.encoder_conv2.weight.data[:, :, :, :] = 1 / (8 * 16
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.0001))
self.encoder_normalization2 = nn.GroupNorm(1, 16, eps=1e-05, affine
=True)
self.encoder_conv3 = nn.Conv2d(in_channels=16, out_channels=32,
kernel_size=4, stride=2, padding=1, dilation=1, groups=1, bias=True
)
self.encoder_conv3.bias.data.zero_()
self.encoder_conv3.weight.data[:, :, :, :] = 1 / (8 * 16
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.001))
self.encoder_normalization3 = nn.GroupNorm(1, 32, eps=1e-05, affine
=True)
self.encoder_mu = nn.Conv2d(in_channels=32, out_channels=16,
kernel_size=4, stride=2, padding=1, dilation=1, groups=1, bias=True
)
self.encoder_mu.bias.data.zero_()
self.encoder_mu.weight.data[:, :, :, :] = 1 / (8 * 16) + torch.normal(
mean=torch.tensor(0.0), std=torch.tensor(0.0001))
self.encoder_log_var = nn.Conv2d(in_channels=32, out_channels=16,
kernel_size=4, stride=2, padding=1, dilation=1, groups=1, bias=True
)
self.encoder_log_var.bias.data[:] = -2.3
self.encoder_log_var.weight.data.zero_()
self.decoder_conv1 = nn.ConvTranspose2d(16, 32, kernel_size=8,
stride=2, padding=3)
self.decoder_conv1.bias.data.zero_()
self.decoder_conv1.weight.data[:, :, :, :] = 1 / (8 * 8 * 8
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.0001))
self.decoder_normalization1 = nn.GroupNorm(1, 32, eps=1e-05, affine
=True)
self.decoder_conv2 = nn.ConvTranspose2d(32, 64, kernel_size=8,
stride=2, padding=3, output_padding=0, groups=1, bias=True,
dilation=1)
self.decoder_conv2.bias.data.zero_()
self.decoder_conv2.weight.data[:, :, :, :] = 1 / (8 * 8 * 8
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.0001))
self.decoder_normalization2 = nn.GroupNorm(1, 64, eps=1e-05, affine
=True)
self.decoder_conv3 = nn.ConvTranspose2d(64, 32, kernel_size=8,
stride=2, padding=3, output_padding=0, groups=1, bias=True,
dilation=1)
self.decoder_conv3.bias.data.zero_()
self.decoder_conv3.weight.data[:, :, :, :] = 1 / (4 * 8 * 8
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.001))
self.decoder_normalization3 = nn.GroupNorm(1, 32, eps=1e-05, affine
=True)
self.decoder_conv5 = nn.ConvTranspose2d(32, 1, kernel_size=8,
stride=2, padding=3, output_padding=0, groups=1, bias=True,
dilation=1)
self.decoder_conv5.bias.data[:] = -(0.5 / 0.24)
self.decoder_conv5.weight.data[:, :, :, :] = 1 / (32 * 8 * 8 * 0.24
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.001))
def forward(self, input):
embeddings_enc0 = F.relu(self.encoder_conv1(input))
embeddings_enc0 = self.encoder_normalization1(embeddings_enc0)
embeddings_enc1 = F.relu(self.encoder_conv2(embeddings_enc0))
embeddings_enc1 = self.encoder_normalization2(embeddings_enc1)
embeddings_enc2 = F.relu(self.encoder_conv3(embeddings_enc1))
embeddings_enc2 = self.encoder_normalization3(embeddings_enc2)
mu = self.encoder_mu(embeddings_enc2)
log_var = self.encoder_log_var(embeddings_enc2)
sample = self.sample_from_mu_log_var(mu, log_var)
embeddings_dec1 = F.relu(self.decoder_conv1(sample, output_size=
embeddings_enc2.size()))
embeddings_dec1 = self.decoder_normalization1(embeddings_dec1)
embeddings_dec2 = F.relu(self.decoder_conv2(embeddings_dec1,
output_size=embeddings_enc1.size()))
embeddings_dec2 = self.decoder_normalization2(embeddings_dec2)
embeddings_dec3 = F.relu(self.decoder_conv3(embeddings_dec2,
output_size=embeddings_enc0.size()))
embeddings_dec3 = self.decoder_normalization3(embeddings_dec3)
reconstructed = F.sigmoid(self.decoder_conv5(embeddings_dec3,
output_size=input.size()))
return (reconstructed, mu, log_var, sample, embeddings_enc1,
embeddings_dec2)
def sample_from_mu_log_var(self, mu, log_var):
std = torch.exp(0.5 * log_var)
eps = torch.randn_like(std)
sample = mu + eps * std
return sample
def get_inputs():
return [torch.rand([4, 1, 64, 64])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch import device
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_red_fused_convolution_native_group_norm_0(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, out_ptr0, out_ptr2, out_ptr3, xnumel, rnumel, XBLOCK:
tl.constexpr, RBLOCK: tl.constexpr):
xnumel = 4
rnumel = 4096
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rbase = tl.arange(0, RBLOCK)[None, :]
x0 = xindex
tmp6_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp6_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp6_weight = tl.zeros([XBLOCK, RBLOCK], tl.float32)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r3 = rindex
r2 = rindex // 1024
tmp0 = tl.load(in_out_ptr0 + (r3 + 4096 * x0), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp1 = tl.load(in_ptr0 + r2, rmask, eviction_policy='evict_last',
other=0.0)
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1, 1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = tl.broadcast_to(tmp4, [XBLOCK, RBLOCK])
tmp6_mean_next, tmp6_m2_next, tmp6_weight_next = (triton_helpers.
welford_reduce(tmp5, tmp6_mean, tmp6_m2, tmp6_weight, roffset == 0)
)
tmp6_mean = tl.where(rmask & xmask, tmp6_mean_next, tmp6_mean)
tmp6_m2 = tl.where(rmask & xmask, tmp6_m2_next, tmp6_m2)
tmp6_weight = tl.where(rmask & xmask, tmp6_weight_next, tmp6_weight)
tl.store(in_out_ptr0 + (r3 + 4096 * x0), tmp2, rmask & xmask)
tmp6_tmp, tmp7_tmp, tmp8_tmp = triton_helpers.welford(tmp6_mean,
tmp6_m2, tmp6_weight, 1)
tmp6 = tmp6_tmp[:, None]
tmp7 = tmp7_tmp[:, None]
tmp8_tmp[:, None]
tl.store(out_ptr0 + x0, tmp6, xmask)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r3 = rindex
r2 = rindex // 1024
tmp9 = tl.load(in_out_ptr0 + (r3 + 4096 * x0), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp19 = tl.load(in_ptr1 + r2, rmask, eviction_policy='evict_last',
other=0.0)
tmp21 = tl.load(in_ptr2 + r2, rmask, eviction_policy='evict_last',
other=0.0)
tmp10 = tl.full([1, 1], 0, tl.int32)
tmp11 = triton_helpers.maximum(tmp10, tmp9)
tmp12 = tmp11 - tmp6
tmp13 = 4096.0
tmp14 = tmp7 / tmp13
tmp15 = 1e-05
tmp16 = tmp14 + tmp15
tmp17 = libdevice.rsqrt(tmp16)
tmp18 = tmp12 * tmp17
tmp20 = tmp18 * tmp19
tmp22 = tmp20 + tmp21
tl.store(out_ptr2 + (r3 + 4096 * x0), tmp22, rmask & xmask)
tmp23 = 4096.0
tmp24 = tmp7 / tmp23
tmp25 = 1e-05
tmp26 = tmp24 + tmp25
tmp27 = libdevice.rsqrt(tmp26)
tl.store(out_ptr3 + x0, tmp27, xmask)
@triton.jit
def triton_red_fused_convolution_native_group_norm_1(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, out_ptr0, out_ptr2, out_ptr3, xnumel, rnumel, XBLOCK:
tl.constexpr, RBLOCK: tl.constexpr):
xnumel = 4
rnumel = 4096
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rbase = tl.arange(0, RBLOCK)[None, :]
x0 = xindex
tmp6_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp6_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp6_weight = tl.zeros([XBLOCK, RBLOCK], tl.float32)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r3 = rindex
r2 = rindex // 256
tmp0 = tl.load(in_out_ptr0 + (r3 + 4096 * x0), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp1 = tl.load(in_ptr0 + r2, rmask, eviction_policy='evict_last',
other=0.0)
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1, 1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = tl.broadcast_to(tmp4, [XBLOCK, RBLOCK])
tmp6_mean_next, tmp6_m2_next, tmp6_weight_next = (triton_helpers.
welford_reduce(tmp5, tmp6_mean, tmp6_m2, tmp6_weight, roffset == 0)
)
tmp6_mean = tl.where(rmask & xmask, tmp6_mean_next, tmp6_mean)
tmp6_m2 = tl.where(rmask & xmask, tmp6_m2_next, tmp6_m2)
tmp6_weight = tl.where(rmask & xmask, tmp6_weight_next, tmp6_weight)
tl.store(in_out_ptr0 + (r3 + 4096 * x0), tmp2, rmask & xmask)
tmp6_tmp, tmp7_tmp, tmp8_tmp = triton_helpers.welford(tmp6_mean,
tmp6_m2, tmp6_weight, 1)
tmp6 = tmp6_tmp[:, None]
tmp7 = tmp7_tmp[:, None]
tmp8_tmp[:, None]
tl.store(out_ptr0 + x0, tmp6, xmask)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r3 = rindex
r2 = rindex // 256
tmp9 = tl.load(in_out_ptr0 + (r3 + 4096 * x0), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp19 = tl.load(in_ptr1 + r2, rmask, eviction_policy='evict_last',
other=0.0)
tmp21 = tl.load(in_ptr2 + r2, rmask, eviction_policy='evict_last',
other=0.0)
tmp10 = tl.full([1, 1], 0, tl.int32)
tmp11 = triton_helpers.maximum(tmp10, tmp9)
tmp12 = tmp11 - tmp6
tmp13 = 4096.0
tmp14 = tmp7 / tmp13
tmp15 = 1e-05
tmp16 = tmp14 + tmp15
tmp17 = libdevice.rsqrt(tmp16)
tmp18 = tmp12 * tmp17
tmp20 = tmp18 * tmp19
tmp22 = tmp20 + tmp21
tl.store(out_ptr2 + (r3 + 4096 * x0), tmp22, rmask & xmask)
tmp23 = 4096.0
tmp24 = tmp7 / tmp23
tmp25 = 1e-05
tmp26 = tmp24 + tmp25
tmp27 = libdevice.rsqrt(tmp26)
tl.store(out_ptr3 + x0, tmp27, xmask)
@triton.jit
def triton_red_fused_convolution_native_group_norm_2(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, out_ptr0, out_ptr2, out_ptr3, xnumel, rnumel, XBLOCK:
tl.constexpr, RBLOCK: tl.constexpr):
xnumel = 4
rnumel = 2048
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rbase = tl.arange(0, RBLOCK)[None, :]
x0 = xindex
tmp6_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp6_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp6_weight = tl.zeros([XBLOCK, RBLOCK], tl.float32)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r3 = rindex
r2 = rindex // 64
tmp0 = tl.load(in_out_ptr0 + (r3 + 2048 * x0), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp1 = tl.load(in_ptr0 + r2, rmask, eviction_policy='evict_last',
other=0.0)
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1, 1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = tl.broadcast_to(tmp4, [XBLOCK, RBLOCK])
tmp6_mean_next, tmp6_m2_next, tmp6_weight_next = (triton_helpers.
welford_reduce(tmp5, tmp6_mean, tmp6_m2, tmp6_weight, roffset == 0)
)
tmp6_mean = tl.where(rmask & xmask, tmp6_mean_next, tmp6_mean)
tmp6_m2 = tl.where(rmask & xmask, tmp6_m2_next, tmp6_m2)
tmp6_weight = tl.where(rmask & xmask, tmp6_weight_next, tmp6_weight)
tl.store(in_out_ptr0 + (r3 + 2048 * x0), tmp2, rmask & xmask)
tmp6_tmp, tmp7_tmp, tmp8_tmp = triton_helpers.welford(tmp6_mean,
tmp6_m2, tmp6_weight, 1)
tmp6 = tmp6_tmp[:, None]
tmp7 = tmp7_tmp[:, None]
tmp8_tmp[:, None]
tl.store(out_ptr0 + x0, tmp6, xmask)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r3 = rindex
r2 = rindex // 64
tmp9 = tl.load(in_out_ptr0 + (r3 + 2048 * x0), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp19 = tl.load(in_ptr1 + r2, rmask, eviction_policy='evict_last',
other=0.0)
tmp21 = tl.load(in_ptr2 + r2, rmask, eviction_policy='evict_last',
other=0.0)
tmp10 = tl.full([1, 1], 0, tl.int32)
tmp11 = triton_helpers.maximum(tmp10, tmp9)
tmp12 = tmp11 - tmp6
tmp13 = 2048.0
tmp14 = tmp7 / tmp13
tmp15 = 1e-05
tmp16 = tmp14 + tmp15
tmp17 = libdevice.rsqrt(tmp16)
tmp18 = tmp12 * tmp17
tmp20 = tmp18 * tmp19
tmp22 = tmp20 + tmp21
tl.store(out_ptr2 + (r3 + 2048 * x0), tmp22, rmask & xmask)
tmp23 = 2048.0
tmp24 = tmp7 / tmp23
tmp25 = 1e-05
tmp26 = tmp24 + tmp25
tmp27 = libdevice.rsqrt(tmp26)
tl.store(out_ptr3 + x0, tmp27, xmask)
@triton.jit
def triton_poi_fused_add_convolution_exp_mul_3(in_out_ptr0, in_out_ptr1,
in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 16
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_out_ptr1 + x3, xmask)
tmp4 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr2 + x3, xmask)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp7 = 0.5
tmp8 = tmp5 * tmp7
tmp9 = tl_math.exp(tmp8)
tmp10 = tmp6 * tmp9
tmp11 = tmp2 + tmp10
tl.store(in_out_ptr0 + x3, tmp2, xmask)
tl.store(in_out_ptr1 + x3, tmp5, xmask)
tl.store(out_ptr0 + x3, tmp11, xmask)
@triton.jit
def triton_red_fused_convolution_native_group_norm_4(in_out_ptr0, in_ptr0,
out_ptr0, out_ptr1, out_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr,
RBLOCK: tl.constexpr):
xnumel = 8
rnumel = 8192
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rbase = tl.arange(0, RBLOCK)[None, :]
x4 = xindex
x0 = xindex % 2
tmp6_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp6_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp6_weight = tl.zeros([XBLOCK, RBLOCK], tl.float32)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r5 = rindex
r3 = rindex // 256
tmp0 = tl.load(in_out_ptr0 + (r5 + 8192 * x4), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp1 = tl.load(in_ptr0 + (r3 + 32 * x0), rmask & xmask,
eviction_policy='evict_last', other=0.0)
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1, 1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = tl.broadcast_to(tmp4, [XBLOCK, RBLOCK])
tmp6_mean_next, tmp6_m2_next, tmp6_weight_next = (triton_helpers.
welford_reduce(tmp5, tmp6_mean, tmp6_m2, tmp6_weight, roffset == 0)
)
tmp6_mean = tl.where(rmask & xmask, tmp6_mean_next, tmp6_mean)
tmp6_m2 = tl.where(rmask & xmask, tmp6_m2_next, tmp6_m2)
tmp6_weight = tl.where(rmask & xmask, tmp6_weight_next, tmp6_weight)
tl.store(in_out_ptr0 + (r5 + 8192 * x4), tmp2, rmask & xmask)
tmp6_tmp, tmp7_tmp, tmp8_tmp = triton_helpers.welford(tmp6_mean,
tmp6_m2, tmp6_weight, 1)
tmp6 = tmp6_tmp[:, None]
tmp7 = tmp7_tmp[:, None]
tmp8 = tmp8_tmp[:, None]
tl.store(out_ptr0 + x4, tmp6, xmask)
tl.store(out_ptr1 + x4, tmp7, xmask)
tl.store(out_ptr2 + x4, tmp8, xmask)
@triton.jit
def triton_per_fused_native_group_norm_5(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, out_ptr1, out_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 2
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 2 * x0), xmask, other=0.0)
tmp1 = tl.load(in_ptr1 + (r1 + 2 * x0), xmask, other=0.0)
tmp2 = tl.load(in_ptr2 + (r1 + 2 * x0), xmask, other=0.0)
tmp3 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp5 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tmp7 = tl.where(xmask, tmp3, 0)
tmp8 = tl.where(xmask, tmp4, 0)
tmp9 = tl.where(xmask, tmp5, 0)
tmp10, tmp11, tmp12 = triton_helpers.welford(tmp7, tmp8, tmp9, 1)
tmp13 = tmp10[:, None]
tmp14 = tmp11[:, None]
tmp12[:, None]
tmp16 = 16384.0
tmp17 = tmp14 / tmp16
tmp18 = 1e-05
tmp19 = tmp17 + tmp18
tmp20 = libdevice.rsqrt(tmp19)
tl.store(out_ptr2 + x0, tmp20, xmask)
tl.store(out_ptr0 + x0, tmp13, xmask)
tl.store(out_ptr1 + x0, tmp14, xmask)
@triton.jit
def triton_poi_fused_native_group_norm_6(in_ptr0, in_ptr1, in_ptr2, in_ptr3,
in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x2 = xindex // 16384
x1 = xindex // 256 % 64
tmp0 = tl.load(in_ptr0 + x3, None)
tmp3 = tl.load(in_ptr1 + x2, None, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr2 + x2, None, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr3 + x1, None, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr4 + x1, None, eviction_policy='evict_last')
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = tmp2 - tmp3
tmp6 = 16384.0
tmp7 = tmp5 / tmp6
tmp8 = 1e-05
tmp9 = tmp7 + tmp8
tmp10 = libdevice.rsqrt(tmp9)
tmp11 = tmp4 * tmp10
tmp13 = tmp11 * tmp12
tmp15 = tmp13 + tmp14
tl.store(out_ptr0 + x3, tmp15, None)
@triton.jit
def triton_red_fused_convolution_native_group_norm_7(in_out_ptr0, in_ptr0,
out_ptr0, out_ptr1, out_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr,
RBLOCK: tl.constexpr):
xnumel = 16
rnumel = 8192
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rbase = tl.arange(0, RBLOCK)[None, :]
x4 = xindex
x0 = xindex % 4
tmp6_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp6_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
tmp6_weight = tl.zeros([XBLOCK, RBLOCK], tl.float32)
for roffset in range(0, rnumel, RBLOCK):
rindex = roffset + rbase
rmask = rindex < rnumel
r5 = rindex
r3 = rindex // 1024
tmp0 = tl.load(in_out_ptr0 + (r5 + 8192 * x4), rmask & xmask,
eviction_policy='evict_first', other=0.0)
tmp1 = tl.load(in_ptr0 + (r3 + 8 * x0), rmask & xmask,
eviction_policy='evict_last', other=0.0)
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1, 1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = tl.broadcast_to(tmp4, [XBLOCK, RBLOCK])
tmp6_mean_next, tmp6_m2_next, tmp6_weight_next = (triton_helpers.
welford_reduce(tmp5, tmp6_mean, tmp6_m2, tmp6_weight, roffset == 0)
)
tmp6_mean = tl.where(rmask & xmask, tmp6_mean_next, tmp6_mean)
tmp6_m2 = tl.where(rmask & xmask, tmp6_m2_next, tmp6_m2)
tmp6_weight = tl.where(rmask & xmask, tmp6_weight_next, tmp6_weight)
tl.store(in_out_ptr0 + (r5 + 8192 * x4), tmp2, rmask & xmask)
tmp6_tmp, tmp7_tmp, tmp8_tmp = triton_helpers.welford(tmp6_mean,
tmp6_m2, tmp6_weight, 1)
tmp6 = tmp6_tmp[:, None]
tmp7 = tmp7_tmp[:, None]
tmp8 = tmp8_tmp[:, None]
tl.store(out_ptr0 + x4, tmp6, xmask)
tl.store(out_ptr1 + x4, tmp7, xmask)
tl.store(out_ptr2 + x4, tmp8, xmask)
@triton.jit
def triton_per_fused_native_group_norm_8(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, out_ptr1, out_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 4 * x0), xmask, other=0.0)
tmp1 = tl.load(in_ptr1 + (r1 + 4 * x0), xmask, other=0.0)
tmp2 = tl.load(in_ptr2 + (r1 + 4 * x0), xmask, other=0.0)
tmp3 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp5 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tmp7 = tl.where(xmask, tmp3, 0)
tmp8 = tl.where(xmask, tmp4, 0)
tmp9 = tl.where(xmask, tmp5, 0)
tmp10, tmp11, tmp12 = triton_helpers.welford(tmp7, tmp8, tmp9, 1)
tmp13 = tmp10[:, None]
tmp14 = tmp11[:, None]
tmp12[:, None]
tmp16 = 32768.0
tmp17 = tmp14 / tmp16
tmp18 = 1e-05
tmp19 = tmp17 + tmp18
tmp20 = libdevice.rsqrt(tmp19)
tl.store(out_ptr2 + x0, tmp20, xmask)
tl.store(out_ptr0 + x0, tmp13, xmask)
tl.store(out_ptr1 + x0, tmp14, xmask)
@triton.jit
def triton_poi_fused_native_group_norm_9(in_ptr0, in_ptr1, in_ptr2, in_ptr3,
in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x2 = xindex // 32768
x1 = xindex // 1024 % 32
tmp0 = tl.load(in_ptr0 + x3, None)
tmp3 = tl.load(in_ptr1 + x2, None, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr2 + x2, None, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr3 + x1, None, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr4 + x1, None, eviction_policy='evict_last')
tmp1 = tl.full([1], 0, tl.int32)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = tmp2 - tmp3
tmp6 = 32768.0
tmp7 = tmp5 / tmp6
tmp8 = 1e-05
tmp9 = tmp7 + tmp8
tmp10 = libdevice.rsqrt(tmp9)
tmp11 = tmp4 * tmp10
tmp13 = tmp11 * tmp12
tmp15 = tmp13 + tmp14
tl.store(out_ptr0 + x3, tmp15, None)
@triton.jit
def triton_poi_fused_convolution_sigmoid_10(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, None)
tmp1 = tl.load(in_ptr0 + 0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK])
tmp3 = tmp0 + tmp2
tmp4 = tl.sigmoid(tmp3)
tl.store(in_out_ptr0 + x0, tmp4, None)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13, primals_14, primals_15, primals_16, primals_17,
primals_18, primals_19, primals_20, primals_21, primals_22,
primals_23, primals_24, primals_25, primals_26, primals_27,
primals_28, primals_29, primals_30, primals_31) = args
args.clear()
assert_size_stride(primals_1, (4, 1, 4, 4), (16, 16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 1, 64, 64), (4096, 4096, 64, 1))
assert_size_stride(primals_4, (4,), (1,))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (16, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_7, (16,), (1,))
assert_size_stride(primals_8, (16,), (1,))
assert_size_stride(primals_9, (16,), (1,))
assert_size_stride(primals_10, (32, 16, 4, 4), (256, 16, 4, 1))
assert_size_stride(primals_11, (32,), (1,))
assert_size_stride(primals_12, (32,), (1,))
assert_size_stride(primals_13, (32,), (1,))
assert_size_stride(primals_14, (16, 32, 4, 4), (512, 16, 4, 1))
assert_size_stride(primals_15, (16,), (1,))
assert_size_stride(primals_16, (16, 32, 4, 4), (512, 16, 4, 1))
assert_size_stride(primals_17, (16,), (1,))
assert_size_stride(primals_18, (16, 32, 8, 8), (2048, 64, 8, 1))
assert_size_stride(primals_19, (32,), (1,))
assert_size_stride(primals_20, (32,), (1,))
assert_size_stride(primals_21, (32,), (1,))
assert_size_stride(primals_22, (32, 64, 8, 8), (4096, 64, 8, 1))
assert_size_stride(primals_23, (64,), (1,))
assert_size_stride(primals_24, (64,), (1,))
assert_size_stride(primals_25, (64,), (1,))
assert_size_stride(primals_26, (64, 32, 8, 8), (2048, 64, 8, 1))
assert_size_stride(primals_27, (32,), (1,))
assert_size_stride(primals_28, (32,), (1,))
assert_size_stride(primals_29, (32,), (1,))
assert_size_stride(primals_30, (32, 1, 8, 8), (64, 64, 8, 1))
assert_size_stride(primals_31, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(2,
2), padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 32, 32), (4096, 1024, 32, 1))
buf1 = buf0
del buf0
buf2 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf5 = empty_strided_cuda((4, 4, 32, 32), (4096, 1024, 32, 1),
torch.float32)
buf6 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
get_raw_stream(0)
triton_red_fused_convolution_native_group_norm_0[grid(4)](buf1,
primals_2, primals_4, primals_5, buf2, buf5, buf6, 4, 4096,
XBLOCK=1, RBLOCK=2048, num_warps=16, num_stages=1)
del primals_2
del primals_5
buf7 = extern_kernels.convolution(buf5, primals_6, stride=(2, 2),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf7, (4, 16, 16, 16), (4096, 256, 16, 1))
buf8 = buf7
del buf7
buf9 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf12 = empty_strided_cuda((4, 16, 16, 16), (4096, 256, 16, 1),
torch.float32)
buf13 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
triton_red_fused_convolution_native_group_norm_1[grid(4)](buf8,
primals_7, primals_8, primals_9, buf9, buf12, buf13, 4, 4096,
XBLOCK=1, RBLOCK=2048, num_warps=16, num_stages=1)
del primals_7
del primals_9
buf14 = extern_kernels.convolution(buf12, primals_10, stride=(2, 2),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf14, (4, 32, 8, 8), (2048, 64, 8, 1))
buf15 = buf14
del buf14
buf16 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf19 = empty_strided_cuda((4, 32, 8, 8), (2048, 64, 8, 1), torch.
float32)
buf20 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
triton_red_fused_convolution_native_group_norm_2[grid(4)](buf15,
primals_11, primals_12, primals_13, buf16, buf19, buf20, 4,
2048, XBLOCK=1, RBLOCK=2048, num_warps=16, num_stages=1)
del primals_11
del primals_13
buf21 = extern_kernels.convolution(buf19, primals_14, stride=(2, 2),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf21, (4, 16, 4, 4), (256, 16, 4, 1))
buf23 = extern_kernels.convolution(buf19, primals_16, stride=(2, 2),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf23, (4, 16, 4, 4), (256, 16, 4, 1))
buf25 = torch.ops.aten.randn.default([4, 16, 4, 4], dtype=torch.
float32, device=device(type='cuda', index=0), pin_memory=False)
buf26 = buf25
del buf25
buf22 = buf21
del buf21
buf24 = buf23
del buf23
buf27 = empty_strided_cuda((4, 16, 4, 4), (256, 16, 4, 1), torch.
float32)
triton_poi_fused_add_convolution_exp_mul_3[grid(1024)](buf22, buf24,
primals_15, primals_17, buf26, buf27, 1024, XBLOCK=128,
num_warps=4, num_stages=1)
del primals_15
del primals_17
buf28 = extern_kernels.convolution(buf27, primals_18, stride=(2, 2),
padding=(3, 3), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf28, (4, 32, 8, 8), (2048, 64, 8, 1))
buf29 = buf28
del buf28
buf30 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf33 = empty_strided_cuda((4, 32, 8, 8), (2048, 64, 8, 1), torch.
float32)
buf34 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
triton_red_fused_convolution_native_group_norm_2[grid(4)](buf29,
primals_19, primals_20, primals_21, buf30, buf33, buf34, 4,
2048, XBLOCK=1, RBLOCK=2048, num_warps=16, num_stages=1)
del primals_19
del primals_21
buf35 = extern_kernels.convolution(buf33, primals_22, stride=(2, 2),
padding=(3, 3), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf35, (4, 64, 16, 16), (16384, 256, 16, 1))
buf36 = buf35
del buf35
buf37 = empty_strided_cuda((4, 1, 1, 1, 2), (2, 8, 8, 8, 1), torch.
float32)
buf38 = empty_strided_cuda((4, 1, 1, 1, 2), (2, 8, 8, 8, 1), torch.
float32)
buf39 = empty_strided_cuda((4, 1, 1, 1, 2), (2, 8, 8, 8, 1), torch.
float32)
triton_red_fused_convolution_native_group_norm_4[grid(8)](buf36,
primals_23, buf37, buf38, buf39, 8, 8192, XBLOCK=1, RBLOCK=2048,
num_warps=16, num_stages=1)
del primals_23
buf40 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf41 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf44 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
triton_per_fused_native_group_norm_5[grid(4)](buf37, buf38, buf39,
buf40, buf41, buf44, 4, 2, XBLOCK=1, num_warps=2, num_stages=1)
del buf37
del buf38
del buf39
buf43 = empty_strided_cuda((4, 64, 16, 16), (16384, 256, 16, 1),
torch.float32)
triton_poi_fused_native_group_norm_6[grid(65536)](buf36, buf40,
buf41, primals_24, primals_25, buf43, 65536, XBLOCK=512,
num_warps=4, num_stages=1)
del primals_25
buf45 = extern_kernels.convolution(buf43, primals_26, stride=(2, 2),
padding=(3, 3), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf45, (4, 32, 32, 32), (32768, 1024, 32, 1))
buf46 = buf45
del buf45
buf47 = empty_strided_cuda((4, 1, 1, 1, 4), (4, 16, 16, 16, 1),
torch.float32)
buf48 = empty_strided_cuda((4, 1, 1, 1, 4), (4, 16, 16, 16, 1),
torch.float32)
buf49 = empty_strided_cuda((4, 1, 1, 1, 4), (4, 16, 16, 16, 1),
torch.float32)
triton_red_fused_convolution_native_group_norm_7[grid(16)](buf46,
primals_27, buf47, buf48, buf49, 16, 8192, XBLOCK=1, RBLOCK=
2048, num_warps=16, num_stages=1)
del primals_27
buf50 = buf41
del buf41
buf51 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf54 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
triton_per_fused_native_group_norm_8[grid(4)](buf47, buf48, buf49,
buf50, buf51, buf54, 4, 4, XBLOCK=1, num_warps=2, num_stages=1)
del buf47
del buf48
del buf49
buf53 = empty_strided_cuda((4, 32, 32, 32), (32768, 1024, 32, 1),
torch.float32)
triton_poi_fused_native_group_norm_9[grid(131072)](buf46, buf50,
buf51, primals_28, primals_29, buf53, 131072, XBLOCK=512,
num_warps=8, num_stages=1)
del buf51
del primals_29
buf55 = extern_kernels.convolution(buf53, primals_30, stride=(2, 2),
padding=(3, 3), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf55, (4, 1, 64, 64), (4096, 4096, 64, 1))
buf56 = buf55
del buf55
triton_poi_fused_convolution_sigmoid_10[grid(16384)](buf56,
primals_31, 16384, XBLOCK=256, num_warps=4, num_stages=1)
del primals_31
return (buf56, buf22, buf24, buf27, buf12, buf43, primals_1, primals_3,
primals_4, primals_6, primals_8, primals_10, primals_12, primals_14,
primals_16, primals_18, primals_20, primals_22, primals_24,
primals_26, primals_28, primals_30, buf1, buf5, reinterpret_tensor(
buf2, (4, 1), (1, 1), 0), reinterpret_tensor(buf6, (4, 1), (1, 1),
0), buf8, buf12, reinterpret_tensor(buf9, (4, 1), (1, 1), 0),
reinterpret_tensor(buf13, (4, 1), (1, 1), 0), buf15, buf19,
reinterpret_tensor(buf16, (4, 1), (1, 1), 0), reinterpret_tensor(
buf20, (4, 1), (1, 1), 0), buf24, buf26, buf27, buf29, buf33,
reinterpret_tensor(buf30, (4, 1), (1, 1), 0), reinterpret_tensor(
buf34, (4, 1), (1, 1), 0), buf36, buf43, reinterpret_tensor(buf40,
(4, 1), (1, 1), 0), reinterpret_tensor(buf44, (4, 1), (1, 1), 0),
buf46, buf53, reinterpret_tensor(buf50, (4, 1), (1, 1), 0),
reinterpret_tensor(buf54, (4, 1), (1, 1), 0), buf56)
class TextureFinderNew(nn.Module):
def __init__(self):
super(TextureFinderNew, self).__init__()
self.encoder_conv1 = nn.Conv2d(in_channels=1, out_channels=4,
kernel_size=4, stride=2, padding=1, dilation=1, groups=1, bias=True
)
self.encoder_conv1.bias.data.zero_()
self.encoder_conv1.weight.data[:, :, :, :] = 1 / 0.32 + torch.normal(
mean=torch.tensor(0.0), std=torch.tensor(0.0001))
self.encoder_normalization1 = nn.GroupNorm(1, 4, eps=1e-05, affine=True
)
self.encoder_conv2 = nn.Conv2d(in_channels=4, out_channels=16,
kernel_size=4, stride=2, padding=1, dilation=1, groups=1, bias=True
)
self.encoder_conv2.bias.data.zero_()
self.encoder_conv2.weight.data[:, :, :, :] = 1 / (8 * 16
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.0001))
self.encoder_normalization2 = nn.GroupNorm(1, 16, eps=1e-05, affine
=True)
self.encoder_conv3 = nn.Conv2d(in_channels=16, out_channels=32,
kernel_size=4, stride=2, padding=1, dilation=1, groups=1, bias=True
)
self.encoder_conv3.bias.data.zero_()
self.encoder_conv3.weight.data[:, :, :, :] = 1 / (8 * 16
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.001))
self.encoder_normalization3 = nn.GroupNorm(1, 32, eps=1e-05, affine
=True)
self.encoder_mu = nn.Conv2d(in_channels=32, out_channels=16,
kernel_size=4, stride=2, padding=1, dilation=1, groups=1, bias=True
)
self.encoder_mu.bias.data.zero_()
self.encoder_mu.weight.data[:, :, :, :] = 1 / (8 * 16) + torch.normal(
mean=torch.tensor(0.0), std=torch.tensor(0.0001))
self.encoder_log_var = nn.Conv2d(in_channels=32, out_channels=16,
kernel_size=4, stride=2, padding=1, dilation=1, groups=1, bias=True
)
self.encoder_log_var.bias.data[:] = -2.3
self.encoder_log_var.weight.data.zero_()
self.decoder_conv1 = nn.ConvTranspose2d(16, 32, kernel_size=8,
stride=2, padding=3)
self.decoder_conv1.bias.data.zero_()
self.decoder_conv1.weight.data[:, :, :, :] = 1 / (8 * 8 * 8
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.0001))
self.decoder_normalization1 = nn.GroupNorm(1, 32, eps=1e-05, affine
=True)
self.decoder_conv2 = nn.ConvTranspose2d(32, 64, kernel_size=8,
stride=2, padding=3, output_padding=0, groups=1, bias=True,
dilation=1)
self.decoder_conv2.bias.data.zero_()
self.decoder_conv2.weight.data[:, :, :, :] = 1 / (8 * 8 * 8
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.0001))
self.decoder_normalization2 = nn.GroupNorm(1, 64, eps=1e-05, affine
=True)
self.decoder_conv3 = nn.ConvTranspose2d(64, 32, kernel_size=8,
stride=2, padding=3, output_padding=0, groups=1, bias=True,
dilation=1)
self.decoder_conv3.bias.data.zero_()
self.decoder_conv3.weight.data[:, :, :, :] = 1 / (4 * 8 * 8
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.001))
self.decoder_normalization3 = nn.GroupNorm(1, 32, eps=1e-05, affine
=True)
self.decoder_conv5 = nn.ConvTranspose2d(32, 1, kernel_size=8,
stride=2, padding=3, output_padding=0, groups=1, bias=True,
dilation=1)
self.decoder_conv5.bias.data[:] = -(0.5 / 0.24)
self.decoder_conv5.weight.data[:, :, :, :] = 1 / (32 * 8 * 8 * 0.24
) + torch.normal(mean=torch.tensor(0.0), std=torch.tensor(0.001))
def sample_from_mu_log_var(self, mu, log_var):
std = torch.exp(0.5 * log_var)
eps = torch.randn_like(std)
sample = mu + eps * std
return sample
def forward(self, input_0):
primals_1 = self.encoder_conv1.weight
primals_2 = self.encoder_conv1.bias
primals_4 = self.encoder_normalization1.weight
primals_5 = self.encoder_normalization1.bias
primals_6 = self.encoder_conv2.weight
primals_7 = self.encoder_conv2.bias
primals_8 = self.encoder_normalization2.weight
primals_9 = self.encoder_normalization2.bias
primals_10 = self.encoder_conv3.weight
primals_11 = self.encoder_conv3.bias
primals_12 = self.encoder_normalization3.weight
primals_13 = self.encoder_normalization3.bias
primals_14 = self.encoder_mu.weight
primals_15 = self.encoder_mu.bias
primals_16 = self.encoder_log_var.weight
primals_17 = self.encoder_log_var.bias
primals_18 = self.decoder_conv1.weight
primals_19 = self.decoder_conv1.bias
primals_20 = self.decoder_normalization1.weight
primals_21 = self.decoder_normalization1.bias
primals_22 = self.decoder_conv2.weight
primals_23 = self.decoder_conv2.bias
primals_24 = self.decoder_normalization2.weight
primals_25 = self.decoder_normalization2.bias
primals_26 = self.decoder_conv3.weight
primals_27 = self.decoder_conv3.bias
primals_28 = self.decoder_normalization3.weight
primals_29 = self.decoder_normalization3.bias
primals_30 = self.decoder_conv5.weight
primals_31 = self.decoder_conv5.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13, primals_14,
primals_15, primals_16, primals_17, primals_18, primals_19,
primals_20, primals_21, primals_22, primals_23, primals_24,
primals_25, primals_26, primals_27, primals_28, primals_29,
primals_30, primals_31])
return output[0], output[1], output[2], output[3], output[4], output[5]
|
paucarre/staal
|
TextureFinder
| false
| 4,142
|
[
"MIT"
] | 0
|
1635e514f0ed978a08c078afd258980bcb6f0cec
|
https://github.com/paucarre/staal/tree/1635e514f0ed978a08c078afd258980bcb6f0cec
|
C3D
|
import torch
import torch.nn as nn
import torch.nn
class C3D(nn.Module):
"""
The C3D network as described in [1].
"""
def __init__(self):
super(C3D, self).__init__()
self.conv1 = nn.Conv3d(3, 64, kernel_size=(3, 3, 3), padding=(1, 1, 1))
self.pool1 = nn.MaxPool3d(kernel_size=(1, 2, 2), stride=(1, 2, 2))
self.conv2 = nn.Conv3d(64, 128, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.pool2 = nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2))
self.conv3a = nn.Conv3d(128, 256, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.conv3b = nn.Conv3d(256, 256, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.pool3 = nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2))
self.conv4a = nn.Conv3d(256, 512, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.conv4b = nn.Conv3d(512, 512, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.pool4 = nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2))
self.conv5a = nn.Conv3d(512, 512, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.conv5b = nn.Conv3d(512, 512, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.pool5 = nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2),
padding=(0, 1, 1))
self.fc6 = nn.Linear(8192, 4096)
self.fc7 = nn.Linear(4096, 1024)
self.fc8 = nn.Linear(1024, 5)
self.dropout = nn.Dropout(p=0.5)
self.relu = nn.ReLU()
self.softmax = nn.Softmax()
def forward(self, x):
h = self.relu(self.conv1(x))
h = self.pool1(h)
h = self.relu(self.conv2(h))
h = self.pool2(h)
h = self.relu(self.conv3a(h))
h = self.relu(self.conv3b(h))
h = self.pool3(h)
h = self.relu(self.conv4a(h))
h = self.relu(self.conv4b(h))
h = self.pool4(h)
h = self.relu(self.conv5a(h))
h = self.relu(self.conv5b(h))
h = self.pool5(h)
h = h.view(-1, 8192)
h = self.relu(self.fc6(h))
h = self.dropout(h)
h = self.relu(self.fc7(h))
h = self.dropout(h)
logits = self.fc8(h)
probs = self.softmax(logits)
return probs
def get_inputs():
return [torch.rand([4, 3, 64, 64, 64])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
import torch.nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_convolution_relu_0(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = xindex // 262144 % 64
tmp0 = tl.load(in_out_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_relu_1(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = xindex // 65536 % 128
tmp0 = tl.load(in_out_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_relu_2(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = xindex // 8192 % 256
tmp0 = tl.load(in_out_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_relu_3(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = xindex // 1024 % 512
tmp0 = tl.load(in_out_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_relu_4(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x1 = xindex // 128 % 512
tmp0 = tl.load(in_out_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, None)
@triton.jit
def triton_poi_fused_relu_5(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 4096
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_relu_6(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 9216
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 1024
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused__softmax_7(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK:
tl.constexpr):
xnumel = 9
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 5 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 5 * x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 5 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 5 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (4 + 5 * x0), xmask, eviction_policy='evict_last')
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp4 = triton_helpers.maximum(tmp2, tmp3)
tmp6 = triton_helpers.maximum(tmp4, tmp5)
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = tmp0 - tmp8
tmp10 = tl_math.exp(tmp9)
tmp11 = tmp1 - tmp8
tmp12 = tl_math.exp(tmp11)
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tl_math.exp(tmp14)
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tl_math.exp(tmp17)
tmp19 = tmp16 + tmp18
tmp20 = tmp7 - tmp8
tmp21 = tl_math.exp(tmp20)
tmp22 = tmp19 + tmp21
tl.store(out_ptr0 + x0, tmp8, xmask)
tl.store(out_ptr1 + x0, tmp22, xmask)
@triton.jit
def triton_poi_fused__softmax_8(in_out_ptr0, in_ptr0, in_ptr1, xnumel,
XBLOCK: tl.constexpr):
xnumel = 45
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 5
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp3 = tl_math.exp(tmp2)
tmp5 = tmp3 / tmp4
tl.store(in_out_ptr0 + x2, tmp5, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13, primals_14, primals_15, primals_16, primals_17,
primals_18, primals_19, primals_20, primals_21, primals_22, primals_23
) = args
args.clear()
assert_size_stride(primals_1, (64, 3, 3, 3, 3), (81, 27, 9, 3, 1))
assert_size_stride(primals_2, (64,), (1,))
assert_size_stride(primals_3, (4, 3, 64, 64, 64), (786432, 262144, 4096,
64, 1))
assert_size_stride(primals_4, (128, 64, 3, 3, 3), (1728, 27, 9, 3, 1))
assert_size_stride(primals_5, (128,), (1,))
assert_size_stride(primals_6, (256, 128, 3, 3, 3), (3456, 27, 9, 3, 1))
assert_size_stride(primals_7, (256,), (1,))
assert_size_stride(primals_8, (256, 256, 3, 3, 3), (6912, 27, 9, 3, 1))
assert_size_stride(primals_9, (256,), (1,))
assert_size_stride(primals_10, (512, 256, 3, 3, 3), (6912, 27, 9, 3, 1))
assert_size_stride(primals_11, (512,), (1,))
assert_size_stride(primals_12, (512, 512, 3, 3, 3), (13824, 27, 9, 3, 1))
assert_size_stride(primals_13, (512,), (1,))
assert_size_stride(primals_14, (512, 512, 3, 3, 3), (13824, 27, 9, 3, 1))
assert_size_stride(primals_15, (512,), (1,))
assert_size_stride(primals_16, (512, 512, 3, 3, 3), (13824, 27, 9, 3, 1))
assert_size_stride(primals_17, (512,), (1,))
assert_size_stride(primals_18, (4096, 8192), (8192, 1))
assert_size_stride(primals_19, (4096,), (1,))
assert_size_stride(primals_20, (1024, 4096), (4096, 1))
assert_size_stride(primals_21, (1024,), (1,))
assert_size_stride(primals_22, (5, 1024), (1024, 1))
assert_size_stride(primals_23, (5,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,
1, 1), padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 64, 64, 64, 64), (16777216, 262144,
4096, 64, 1))
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_convolution_relu_0[grid(67108864)](buf1, primals_2,
67108864, XBLOCK=512, num_warps=8, num_stages=1)
del primals_2
buf2 = torch.ops.aten.max_pool3d_with_indices.default(buf1, [1, 2,
2], [1, 2, 2])
buf3 = buf2[0]
buf4 = buf2[1]
del buf2
buf5 = extern_kernels.convolution(buf3, primals_4, stride=(1, 1, 1),
padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf5, (4, 128, 64, 32, 32), (8388608, 65536,
1024, 32, 1))
buf6 = buf5
del buf5
triton_poi_fused_convolution_relu_1[grid(33554432)](buf6, primals_5,
33554432, XBLOCK=512, num_warps=8, num_stages=1)
del primals_5
buf7 = torch.ops.aten.max_pool3d_with_indices.default(buf6, [2, 2,
2], [2, 2, 2])
buf8 = buf7[0]
buf9 = buf7[1]
del buf7
buf10 = extern_kernels.convolution(buf8, primals_6, stride=(1, 1, 1
), padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf10, (4, 256, 32, 16, 16), (2097152, 8192, 256,
16, 1))
buf11 = buf10
del buf10
triton_poi_fused_convolution_relu_2[grid(8388608)](buf11, primals_7,
8388608, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_7
buf12 = extern_kernels.convolution(buf11, primals_8, stride=(1, 1,
1), padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf12, (4, 256, 32, 16, 16), (2097152, 8192, 256,
16, 1))
buf13 = buf12
del buf12
triton_poi_fused_convolution_relu_2[grid(8388608)](buf13, primals_9,
8388608, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_9
buf14 = torch.ops.aten.max_pool3d_with_indices.default(buf13, [2, 2,
2], [2, 2, 2])
buf15 = buf14[0]
buf16 = buf14[1]
del buf14
buf17 = extern_kernels.convolution(buf15, primals_10, stride=(1, 1,
1), padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf17, (4, 512, 16, 8, 8), (524288, 1024, 64, 8, 1))
buf18 = buf17
del buf17
triton_poi_fused_convolution_relu_3[grid(2097152)](buf18,
primals_11, 2097152, XBLOCK=512, num_warps=8, num_stages=1)
del primals_11
buf19 = extern_kernels.convolution(buf18, primals_12, stride=(1, 1,
1), padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf19, (4, 512, 16, 8, 8), (524288, 1024, 64, 8, 1))
buf20 = buf19
del buf19
triton_poi_fused_convolution_relu_3[grid(2097152)](buf20,
primals_13, 2097152, XBLOCK=512, num_warps=8, num_stages=1)
del primals_13
buf21 = torch.ops.aten.max_pool3d_with_indices.default(buf20, [2, 2,
2], [2, 2, 2])
buf22 = buf21[0]
buf23 = buf21[1]
del buf21
buf24 = extern_kernels.convolution(buf22, primals_14, stride=(1, 1,
1), padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf24, (4, 512, 8, 4, 4), (65536, 128, 16, 4, 1))
buf25 = buf24
del buf24
triton_poi_fused_convolution_relu_4[grid(262144)](buf25, primals_15,
262144, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_15
buf26 = extern_kernels.convolution(buf25, primals_16, stride=(1, 1,
1), padding=(1, 1, 1), dilation=(1, 1, 1), transposed=False,
output_padding=(0, 0, 0), groups=1, bias=None)
assert_size_stride(buf26, (4, 512, 8, 4, 4), (65536, 128, 16, 4, 1))
buf27 = buf26
del buf26
triton_poi_fused_convolution_relu_4[grid(262144)](buf27, primals_17,
262144, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_17
buf28 = torch.ops.aten.max_pool3d_with_indices.default(buf27, [2, 2,
2], [2, 2, 2], [0, 1, 1])
buf29 = buf28[0]
buf30 = buf28[1]
del buf28
buf31 = empty_strided_cuda((9, 4096), (4096, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf29, (9, 8192), (8192, 1), 0
), reinterpret_tensor(primals_18, (8192, 4096), (1, 8192), 0),
out=buf31)
buf32 = buf31
del buf31
triton_poi_fused_relu_5[grid(36864)](buf32, primals_19, 36864,
XBLOCK=512, num_warps=4, num_stages=1)
del primals_19
buf33 = empty_strided_cuda((9, 1024), (1024, 1), torch.float32)
extern_kernels.mm(buf32, reinterpret_tensor(primals_20, (4096, 1024
), (1, 4096), 0), out=buf33)
buf34 = buf33
del buf33
triton_poi_fused_relu_6[grid(9216)](buf34, primals_21, 9216, XBLOCK
=256, num_warps=4, num_stages=1)
del primals_21
buf35 = empty_strided_cuda((9, 5), (5, 1), torch.float32)
extern_kernels.addmm(primals_23, buf34, reinterpret_tensor(
primals_22, (1024, 5), (1, 1024), 0), alpha=1, beta=1, out=buf35)
del primals_23
buf36 = empty_strided_cuda((9, 1), (1, 9), torch.float32)
buf37 = empty_strided_cuda((9, 1), (1, 9), torch.float32)
triton_poi_fused__softmax_7[grid(9)](buf35, buf36, buf37, 9, XBLOCK
=16, num_warps=1, num_stages=1)
buf38 = buf35
del buf35
triton_poi_fused__softmax_8[grid(45)](buf38, buf36, buf37, 45,
XBLOCK=64, num_warps=1, num_stages=1)
del buf36
del buf37
return (buf38, primals_1, primals_3, primals_4, primals_6, primals_8,
primals_10, primals_12, primals_14, primals_16, buf1, buf3, buf4,
buf6, buf8, buf9, buf11, buf13, buf15, buf16, buf18, buf20, buf22,
buf23, buf25, buf27, buf30, reinterpret_tensor(buf29, (9, 8192), (
8192, 1), 0), buf32, buf34, buf38, primals_22, primals_20, primals_18)
class C3DNew(nn.Module):
"""
The C3D network as described in [1].
"""
def __init__(self):
super(C3DNew, self).__init__()
self.conv1 = nn.Conv3d(3, 64, kernel_size=(3, 3, 3), padding=(1, 1, 1))
self.pool1 = nn.MaxPool3d(kernel_size=(1, 2, 2), stride=(1, 2, 2))
self.conv2 = nn.Conv3d(64, 128, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.pool2 = nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2))
self.conv3a = nn.Conv3d(128, 256, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.conv3b = nn.Conv3d(256, 256, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.pool3 = nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2))
self.conv4a = nn.Conv3d(256, 512, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.conv4b = nn.Conv3d(512, 512, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.pool4 = nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2))
self.conv5a = nn.Conv3d(512, 512, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.conv5b = nn.Conv3d(512, 512, kernel_size=(3, 3, 3), padding=(1,
1, 1))
self.pool5 = nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2),
padding=(0, 1, 1))
self.fc6 = nn.Linear(8192, 4096)
self.fc7 = nn.Linear(4096, 1024)
self.fc8 = nn.Linear(1024, 5)
self.dropout = nn.Dropout(p=0.5)
self.relu = nn.ReLU()
self.softmax = nn.Softmax()
def forward(self, input_0):
primals_1 = self.conv1.weight
primals_2 = self.conv1.bias
primals_4 = self.conv2.weight
primals_5 = self.conv2.bias
primals_6 = self.conv3a.weight
primals_7 = self.conv3a.bias
primals_8 = self.conv3b.weight
primals_9 = self.conv3b.bias
primals_10 = self.conv4a.weight
primals_11 = self.conv4a.bias
primals_12 = self.conv4b.weight
primals_13 = self.conv4b.bias
primals_14 = self.conv5a.weight
primals_15 = self.conv5a.bias
primals_16 = self.conv5b.weight
primals_17 = self.conv5b.bias
primals_18 = self.fc6.weight
primals_19 = self.fc6.bias
primals_20 = self.fc7.weight
primals_21 = self.fc7.bias
primals_22 = self.fc8.weight
primals_23 = self.fc8.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13, primals_14,
primals_15, primals_16, primals_17, primals_18, primals_19,
primals_20, primals_21, primals_22, primals_23])
return output[0]
|
kar98kbang/c3d-pytorch
|
C3D
| false
| 4,143
|
[
"MIT"
] | 0
|
22b3564798cb9249ad6fdb6c9d929bff3fdfa567
|
https://github.com/kar98kbang/c3d-pytorch/tree/22b3564798cb9249ad6fdb6c9d929bff3fdfa567
|
Model
|
import torch
import torch.nn as nn
import torch.nn.functional as F
class Model(nn.Module):
"""conv. autoencoder"""
def __init__(self):
"""constructor"""
super().__init__()
self.conv1 = nn.Conv2d(3, 32, 5, padding=2)
self.conv2 = nn.Conv2d(32, 64, 3, padding=1)
self.conv3 = nn.Conv2d(64, 128, 3, padding=1)
self.deconv1 = nn.ConvTranspose2d(128, 64, 2, stride=2)
self.deconv2 = nn.ConvTranspose2d(64, 32, 2, stride=2)
self.deconv3 = nn.ConvTranspose2d(32, 3, 2, stride=2)
self.conv4 = nn.Conv2d(3, 3, 5, padding=2)
def forward(self, x):
"""forward prop."""
x = F.max_pool2d(F.relu(self.conv1(x)), 2)
x = F.max_pool2d(F.relu(self.conv2(x)), 2)
x = F.max_pool2d(F.relu(self.conv3(x)), 2)
x = F.relu(self.deconv1(x))
x = F.relu(self.deconv2(x))
x = F.relu(self.deconv3(x))
x = torch.sigmoid(self.conv4(x))
return x
def get_inputs():
return [torch.rand([4, 3, 64, 64])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 96
xnumel = 25
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 3
y1 = yindex // 3
tmp0 = tl.load(in_ptr0 + (x2 + 25 * y3), xmask & ymask, eviction_policy
='evict_last')
tl.store(out_ptr0 + (y0 + 3 * x2 + 75 * y1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 12
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
x2 = xindex
y3 = yindex
y0 = yindex % 3
y1 = yindex // 3
tmp0 = tl.load(in_ptr0 + (x2 + 4096 * y3), ymask, eviction_policy=
'evict_last')
tl.store(out_ptr0 + (y0 + 3 * x2 + 12288 * y1), tmp0, ymask)
@triton.jit
def triton_poi_fused_2(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 32
y1 = yindex // 32
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 32 * x2 + 288 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_3(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 64
y1 = yindex // 64
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 64 * x2 + 576 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 64
y1 = yindex // 64
tmp0 = tl.load(in_ptr0 + (x2 + 4 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 64 * x2 + 256 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_5(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 32
y1 = yindex // 32
tmp0 = tl.load(in_ptr0 + (x2 + 4 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 32 * x2 + 128 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_6(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 96
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 3
y1 = yindex // 3
tmp0 = tl.load(in_ptr0 + (x2 + 4 * y3), xmask & ymask, eviction_policy=
'evict_last')
tl.store(out_ptr0 + (y0 + 3 * x2 + 12 * y1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_7(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 9
xnumel = 25
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 3
y1 = yindex // 3
tmp0 = tl.load(in_ptr0 + (x2 + 25 * y3), xmask & ymask, eviction_policy
='evict_last')
tl.store(out_ptr0 + (y0 + 3 * x2 + 75 * y1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_convolution_relu_8(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 32
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_9(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex % 32
x1 = xindex // 32 % 32
x2 = xindex // 1024
x3 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1 + 4096 * x2), None)
tmp1 = tl.load(in_ptr0 + (32 + x0 + 64 * x1 + 4096 * x2), None)
tmp3 = tl.load(in_ptr0 + (2048 + x0 + 64 * x1 + 4096 * x2), None)
tmp5 = tl.load(in_ptr0 + (2080 + x0 + 64 * x1 + 4096 * x2), None)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1], 1, tl.int8)
tmp9 = tl.full([1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + x3, tmp6, None)
tl.store(out_ptr1 + x3, tmp16, None)
@triton.jit
def triton_poi_fused_convolution_relu_10(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 64
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_11(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex % 64
x1 = xindex // 64 % 16
x2 = xindex // 1024
x3 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 128 * x1 + 4096 * x2), None)
tmp1 = tl.load(in_ptr0 + (64 + x0 + 128 * x1 + 4096 * x2), None)
tmp3 = tl.load(in_ptr0 + (2048 + x0 + 128 * x1 + 4096 * x2), None)
tmp5 = tl.load(in_ptr0 + (2112 + x0 + 128 * x1 + 4096 * x2), None)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1], 1, tl.int8)
tmp9 = tl.full([1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + x3, tmp6, None)
tl.store(out_ptr1 + x3, tmp16, None)
@triton.jit
def triton_poi_fused_convolution_relu_12(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 128
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_13(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex % 128
x1 = xindex // 128 % 8
x2 = xindex // 1024
x3 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 256 * x1 + 4096 * x2), None)
tmp1 = tl.load(in_ptr0 + (128 + x0 + 256 * x1 + 4096 * x2), None)
tmp3 = tl.load(in_ptr0 + (2048 + x0 + 256 * x1 + 4096 * x2), None)
tmp5 = tl.load(in_ptr0 + (2176 + x0 + 256 * x1 + 4096 * x2), None)
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1], 1, tl.int8)
tmp9 = tl.full([1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + x3, tmp6, None)
tl.store(out_ptr1 + x3, tmp16, None)
@triton.jit
def triton_poi_fused_convolution_relu_14(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 64
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_relu_15(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 32
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_relu_16(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 3
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_sigmoid_17(in_ptr0, in_ptr1, out_ptr0,
ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 12
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
x2 = xindex
y0 = yindex % 3
y1 = yindex // 3
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 3 * x2 + 12288 * y1), ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.sigmoid(tmp2)
tl.store(out_ptr0 + (x2 + 4096 * y3), tmp3, ymask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13, primals_14, primals_15) = args
args.clear()
assert_size_stride(primals_1, (32, 3, 5, 5), (75, 25, 5, 1))
assert_size_stride(primals_2, (32,), (1,))
assert_size_stride(primals_3, (4, 3, 64, 64), (12288, 4096, 64, 1))
assert_size_stride(primals_4, (64, 32, 3, 3), (288, 9, 3, 1))
assert_size_stride(primals_5, (64,), (1,))
assert_size_stride(primals_6, (128, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_7, (128,), (1,))
assert_size_stride(primals_8, (128, 64, 2, 2), (256, 4, 2, 1))
assert_size_stride(primals_9, (64,), (1,))
assert_size_stride(primals_10, (64, 32, 2, 2), (128, 4, 2, 1))
assert_size_stride(primals_11, (32,), (1,))
assert_size_stride(primals_12, (32, 3, 2, 2), (12, 4, 2, 1))
assert_size_stride(primals_13, (3,), (1,))
assert_size_stride(primals_14, (3, 3, 5, 5), (75, 25, 5, 1))
assert_size_stride(primals_15, (3,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((32, 3, 5, 5), (75, 1, 15, 3), torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(96, 25)](primals_1, buf0, 96, 25, XBLOCK=32,
YBLOCK=32, num_warps=4, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((4, 3, 64, 64), (12288, 1, 192, 3), torch
.float32)
triton_poi_fused_1[grid(12, 4096)](primals_3, buf1, 12, 4096,
XBLOCK=64, YBLOCK=16, num_warps=4, num_stages=1)
del primals_3
buf2 = empty_strided_cuda((64, 32, 3, 3), (288, 1, 96, 32), torch.
float32)
triton_poi_fused_2[grid(2048, 9)](primals_4, buf2, 2048, 9, XBLOCK=
16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_4
buf3 = empty_strided_cuda((128, 64, 3, 3), (576, 1, 192, 64), torch
.float32)
triton_poi_fused_3[grid(8192, 9)](primals_6, buf3, 8192, 9, XBLOCK=
16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_6
buf4 = empty_strided_cuda((128, 64, 2, 2), (256, 1, 128, 64), torch
.float32)
triton_poi_fused_4[grid(8192, 4)](primals_8, buf4, 8192, 4, XBLOCK=
4, YBLOCK=256, num_warps=4, num_stages=1)
del primals_8
buf5 = empty_strided_cuda((64, 32, 2, 2), (128, 1, 64, 32), torch.
float32)
triton_poi_fused_5[grid(2048, 4)](primals_10, buf5, 2048, 4, XBLOCK
=4, YBLOCK=256, num_warps=4, num_stages=1)
del primals_10
buf6 = empty_strided_cuda((32, 3, 2, 2), (12, 1, 6, 3), torch.float32)
triton_poi_fused_6[grid(96, 4)](primals_12, buf6, 96, 4, XBLOCK=4,
YBLOCK=32, num_warps=4, num_stages=1)
del primals_12
buf7 = empty_strided_cuda((3, 3, 5, 5), (75, 1, 15, 3), torch.float32)
triton_poi_fused_7[grid(9, 25)](primals_14, buf7, 9, 25, XBLOCK=32,
YBLOCK=16, num_warps=4, num_stages=1)
del primals_14
buf8 = extern_kernels.convolution(buf1, buf0, stride=(1, 1),
padding=(2, 2), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf8, (4, 32, 64, 64), (131072, 1, 2048, 32))
buf9 = buf8
del buf8
triton_poi_fused_convolution_relu_8[grid(524288)](buf9, primals_2,
524288, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_2
buf10 = empty_strided_cuda((4, 32, 32, 32), (32768, 1, 1024, 32),
torch.float32)
buf11 = empty_strided_cuda((4, 32, 32, 32), (32768, 1, 1024, 32),
torch.int8)
triton_poi_fused_max_pool2d_with_indices_9[grid(131072)](buf9,
buf10, buf11, 131072, XBLOCK=1024, num_warps=4, num_stages=1)
buf12 = extern_kernels.convolution(buf10, buf2, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf12, (4, 64, 32, 32), (65536, 1, 2048, 64))
buf13 = buf12
del buf12
triton_poi_fused_convolution_relu_10[grid(262144)](buf13, primals_5,
262144, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_5
buf14 = empty_strided_cuda((4, 64, 16, 16), (16384, 1, 1024, 64),
torch.float32)
buf15 = empty_strided_cuda((4, 64, 16, 16), (16384, 1, 1024, 64),
torch.int8)
triton_poi_fused_max_pool2d_with_indices_11[grid(65536)](buf13,
buf14, buf15, 65536, XBLOCK=256, num_warps=4, num_stages=1)
buf16 = extern_kernels.convolution(buf14, buf3, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf16, (4, 128, 16, 16), (32768, 1, 2048, 128))
buf17 = buf16
del buf16
triton_poi_fused_convolution_relu_12[grid(131072)](buf17, primals_7,
131072, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_7
buf18 = empty_strided_cuda((4, 128, 8, 8), (8192, 1, 1024, 128),
torch.float32)
buf19 = empty_strided_cuda((4, 128, 8, 8), (8192, 1, 1024, 128),
torch.int8)
triton_poi_fused_max_pool2d_with_indices_13[grid(32768)](buf17,
buf18, buf19, 32768, XBLOCK=128, num_warps=4, num_stages=1)
buf20 = extern_kernels.convolution(buf18, buf4, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf20, (4, 64, 16, 16), (16384, 1, 1024, 64))
buf21 = buf20
del buf20
triton_poi_fused_convolution_relu_14[grid(65536)](buf21, primals_9,
65536, XBLOCK=512, num_warps=4, num_stages=1)
del primals_9
buf22 = extern_kernels.convolution(buf21, buf5, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf22, (4, 32, 32, 32), (32768, 1, 1024, 32))
buf23 = buf22
del buf22
triton_poi_fused_convolution_relu_15[grid(131072)](buf23,
primals_11, 131072, XBLOCK=512, num_warps=8, num_stages=1)
del primals_11
buf24 = extern_kernels.convolution(buf23, buf6, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf24, (4, 3, 64, 64), (12288, 1, 192, 3))
buf25 = buf24
del buf24
triton_poi_fused_convolution_relu_16[grid(49152)](buf25, primals_13,
49152, XBLOCK=512, num_warps=4, num_stages=1)
del primals_13
buf26 = extern_kernels.convolution(buf25, buf7, stride=(1, 1),
padding=(2, 2), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf26, (4, 3, 64, 64), (12288, 1, 192, 3))
buf27 = empty_strided_cuda((4, 3, 64, 64), (12288, 4096, 64, 1),
torch.float32)
triton_poi_fused_convolution_sigmoid_17[grid(12, 4096)](buf26,
primals_15, buf27, 12, 4096, XBLOCK=64, YBLOCK=16, num_warps=4,
num_stages=1)
del buf26
del primals_15
return (buf27, buf0, buf1, buf2, buf3, buf4, buf5, buf6, buf7, buf9,
buf10, buf11, buf13, buf14, buf15, buf17, buf18, buf19, buf21,
buf23, buf25, buf27)
class ModelNew(nn.Module):
"""conv. autoencoder"""
def __init__(self):
"""constructor"""
super().__init__()
self.conv1 = nn.Conv2d(3, 32, 5, padding=2)
self.conv2 = nn.Conv2d(32, 64, 3, padding=1)
self.conv3 = nn.Conv2d(64, 128, 3, padding=1)
self.deconv1 = nn.ConvTranspose2d(128, 64, 2, stride=2)
self.deconv2 = nn.ConvTranspose2d(64, 32, 2, stride=2)
self.deconv3 = nn.ConvTranspose2d(32, 3, 2, stride=2)
self.conv4 = nn.Conv2d(3, 3, 5, padding=2)
def forward(self, input_0):
primals_1 = self.conv1.weight
primals_2 = self.conv1.bias
primals_4 = self.conv2.weight
primals_5 = self.conv2.bias
primals_6 = self.conv3.weight
primals_7 = self.conv3.bias
primals_8 = self.deconv1.weight
primals_9 = self.deconv1.bias
primals_10 = self.deconv2.weight
primals_11 = self.deconv2.bias
primals_12 = self.deconv3.weight
primals_13 = self.deconv3.bias
primals_14 = self.conv4.weight
primals_15 = self.conv4.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13, primals_14,
primals_15])
return output[0]
|
positivevaib/semi-supervised-imagenet-classification
|
Model
| false
| 4,144
|
[
"MIT"
] | 0
|
4fb6427f5a72951c1b866a1ddbc2599811bb5770
|
https://github.com/positivevaib/semi-supervised-imagenet-classification/tree/4fb6427f5a72951c1b866a1ddbc2599811bb5770
|
ActorCritic
|
import torch
import torch.nn.functional as F
import torch.nn as nn
def swish(x):
return x * F.sigmoid(x)
class ActorCritic(nn.Module):
"""Actor (Policy) Model."""
def __init__(self, state_size, action_size, seed, fc1_units=64,
fc2_units=64):
"""Initialize parameters and build model.
Params
======
state_size (int): Dimension of each state
action_size (int): Dimension of each action
seed (int): Random seed
"""
super().__init__()
self.seed = torch.manual_seed(seed)
self.fc1 = nn.Linear(state_size, fc1_units)
self.fc2 = nn.Linear(fc1_units, fc2_units)
self.actor_fc = nn.Linear(fc2_units, action_size)
self.critic_fc = nn.Linear(fc2_units, 1)
self.std = nn.Parameter(torch.zeros(action_size))
def forward(self, state, actions=None):
"""Build a network that maps state -> actions mu."""
h = swish(self.fc1(state))
h = swish(self.fc2(h))
mu = F.tanh(self.actor_fc(h))
values = self.critic_fc(h).squeeze(-1)
dist = torch.distributions.Normal(mu, F.softplus(self.std))
if actions is None:
actions = dist.sample()
log_prob = dist.log_prob(actions)
log_prob = torch.sum(log_prob, dim=-1)
entropy = torch.sum(dist.entropy(), dim=-1)
return actions, log_prob, entropy, values
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'state_size': 4, 'action_size': 4, 'seed': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.nn.functional as F
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_sigmoid_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, None)
tmp1 = tl.sigmoid(tmp0)
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x0, tmp2, None)
@triton.jit
def triton_poi_fused_tanh_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = libdevice.tanh(tmp0)
tl.store(out_ptr0 + x0, tmp1, xmask)
@triton.jit
def triton_poi_fused_softplus_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 20.0
tmp2 = tmp0 > tmp1
tmp3 = tl_math.exp(tmp0)
tmp4 = libdevice.log1p(tmp3)
tmp5 = tl.where(tmp2, tmp0, tmp4)
tl.store(out_ptr0 + x0, tmp5, xmask)
@triton.jit
def triton_poi_fused_sub_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tl.load(in_out_ptr0 + x0, xmask)
tmp2 = tmp0 - tmp1
tl.store(in_out_ptr0 + x0, tmp2, xmask)
@triton.jit
def triton_poi_fused_mul_pow_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp1 = tmp0 * tmp0
tmp2 = 2.0
tmp3 = tmp1 * tmp2
tl.store(out_ptr0 + x2, tmp3, xmask)
@triton.jit
def triton_poi_fused_add_div_log_neg_pow_sub_sum_5(in_ptr0, in_ptr1,
in_ptr2, out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr2 + 0)
tmp6 = tl.broadcast_to(tmp5, [XBLOCK])
tmp11 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp14 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp16 = tl.load(in_ptr2 + 1)
tmp17 = tl.broadcast_to(tmp16, [XBLOCK])
tmp22 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp25 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp27 = tl.load(in_ptr2 + 2)
tmp28 = tl.broadcast_to(tmp27, [XBLOCK])
tmp33 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp36 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp38 = tl.load(in_ptr2 + 3)
tmp39 = tl.broadcast_to(tmp38, [XBLOCK])
tmp1 = tmp0 * tmp0
tmp2 = -tmp1
tmp4 = tmp2 / tmp3
tmp7 = tl_math.log(tmp6)
tmp8 = tmp4 - tmp7
tmp9 = 0.9189385332046727
tmp10 = tmp8 - tmp9
tmp12 = tmp11 * tmp11
tmp13 = -tmp12
tmp15 = tmp13 / tmp14
tmp18 = tl_math.log(tmp17)
tmp19 = tmp15 - tmp18
tmp20 = tmp19 - tmp9
tmp21 = tmp10 + tmp20
tmp23 = tmp22 * tmp22
tmp24 = -tmp23
tmp26 = tmp24 / tmp25
tmp29 = tl_math.log(tmp28)
tmp30 = tmp26 - tmp29
tmp31 = tmp30 - tmp9
tmp32 = tmp21 + tmp31
tmp34 = tmp33 * tmp33
tmp35 = -tmp34
tmp37 = tmp35 / tmp36
tmp40 = tl_math.log(tmp39)
tmp41 = tmp37 - tmp40
tmp42 = tmp41 - tmp9
tmp43 = tmp32 + tmp42
tmp44 = 1.4189385332046727
tmp45 = tmp7 + tmp44
tmp46 = tmp18 + tmp44
tmp47 = tmp45 + tmp46
tmp48 = tmp29 + tmp44
tmp49 = tmp47 + tmp48
tmp50 = tmp40 + tmp44
tmp51 = tmp49 + tmp50
tl.store(out_ptr0 + x0, tmp43, xmask)
tl.store(out_ptr1 + x0, tmp51, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10) = args
args.clear()
assert_size_stride(primals_1, (64, 4), (4, 1))
assert_size_stride(primals_2, (64,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (64, 64), (64, 1))
assert_size_stride(primals_5, (64,), (1,))
assert_size_stride(primals_6, (4, 64), (64, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (1, 64), (64, 1))
assert_size_stride(primals_9, (1,), (1,))
assert_size_stride(primals_10, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 64), (64, 1), torch.float32)
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (64,
4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 64), (1, 4),
0), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((4, 4, 4, 64), (1024, 256, 64, 1), torch.
float32)
get_raw_stream(0)
triton_poi_fused_mul_sigmoid_0[grid(4096)](buf0, buf1, 4096, XBLOCK
=128, num_warps=4, num_stages=1)
buf2 = empty_strided_cuda((64, 64), (64, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 64),
(64, 1), 0), reinterpret_tensor(primals_4, (64, 64), (1, 64), 0
), alpha=1, beta=1, out=buf2)
del primals_5
buf3 = empty_strided_cuda((4, 4, 4, 64), (1024, 256, 64, 1), torch.
float32)
triton_poi_fused_mul_sigmoid_0[grid(4096)](buf2, buf3, 4096, XBLOCK
=128, num_warps=4, num_stages=1)
buf4 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_7, reinterpret_tensor(buf3, (64, 64),
(64, 1), 0), reinterpret_tensor(primals_6, (64, 4), (1, 64), 0),
alpha=1, beta=1, out=buf4)
del primals_7
buf6 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.addmm(primals_9, reinterpret_tensor(buf3, (64, 64),
(64, 1), 0), reinterpret_tensor(primals_8, (64, 1), (1, 64), 0),
alpha=1, beta=1, out=buf6)
del primals_9
buf7 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_tanh_1[grid(256)](buf4, buf7, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf8 = empty_strided_cuda((4,), (1,), torch.float32)
triton_poi_fused_softplus_2[grid(4)](primals_10, buf8, 4, XBLOCK=4,
num_warps=1, num_stages=1)
buf9 = torch.ops.aten.normal.Tensor_Tensor(buf7, reinterpret_tensor
(buf8, (4, 4, 4, 4), (0, 0, 0, 1), 0))
buf10 = buf9
del buf9
buf11 = buf7
del buf7
triton_poi_fused_sub_3[grid(256)](buf11, buf10, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf12 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_mul_pow_4[grid(256)](buf8, buf12, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf13 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf14 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_div_log_neg_pow_sub_sum_5[grid(64)](buf11,
buf12, buf8, buf13, buf14, 64, XBLOCK=64, num_warps=1, num_stages=1
)
del buf8
return buf10, buf13, buf14, reinterpret_tensor(buf6, (4, 4, 4), (16, 4,
1), 0), primals_10, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf0, reinterpret_tensor(buf1, (64, 64), (64, 1), 0
), buf2, reinterpret_tensor(buf3, (64, 64), (64, 1), 0
), buf4, buf11, buf12, primals_8, primals_6, primals_4
def swish(x):
return x * F.sigmoid(x)
class ActorCriticNew(nn.Module):
"""Actor (Policy) Model."""
def __init__(self, state_size, action_size, seed, fc1_units=64,
fc2_units=64):
"""Initialize parameters and build model.
Params
======
state_size (int): Dimension of each state
action_size (int): Dimension of each action
seed (int): Random seed
"""
super().__init__()
self.seed = torch.manual_seed(seed)
self.fc1 = nn.Linear(state_size, fc1_units)
self.fc2 = nn.Linear(fc1_units, fc2_units)
self.actor_fc = nn.Linear(fc2_units, action_size)
self.critic_fc = nn.Linear(fc2_units, 1)
self.std = nn.Parameter(torch.zeros(action_size))
def forward(self, input_0):
primals_7 = self.std
primals_1 = self.fc1.weight
primals_2 = self.fc1.bias
primals_4 = self.fc2.weight
primals_5 = self.fc2.bias
primals_6 = self.actor_fc.weight
primals_10 = self.actor_fc.bias
primals_8 = self.critic_fc.weight
primals_9 = self.critic_fc.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9, primals_10])
return output[0], output[1], output[2], output[3]
|
postBG/deep-reinforcement-learning
|
ActorCritic
| false
| 4,145
|
[
"MIT"
] | 0
|
5df5662b091c4c3f00beba1aa6f9ce8a52001c93
|
https://github.com/postBG/deep-reinforcement-learning/tree/5df5662b091c4c3f00beba1aa6f9ce8a52001c93
|
ODEfunc
|
import torch
import torch.nn as nn
def norm(dim):
"""
Group normalization to improve model accuracy and training speed.
"""
return nn.GroupNorm(min(1, dim), dim)
class ConcatConv1d(nn.Module):
"""
1d convolution concatenated with time for usage in ODENet.
"""
def __init__(self, dim_in, dim_out, kernel_size=3, stride=1, padding=0,
bias=True, transpose=False):
super(ConcatConv1d, self).__init__()
module = nn.ConvTranspose1d if transpose else nn.Conv1d
self._layer = module(dim_in + 1, dim_out, kernel_size=kernel_size,
stride=stride, padding=padding, bias=bias)
def forward(self, t, x):
tt = torch.ones_like(x[:, :1, :]) * t
ttx = torch.cat([tt, x], 1)
return self._layer(ttx)
class ODEfunc(nn.Module):
"""
Network architecture for ODENet.
"""
def __init__(self, dim):
super(ODEfunc, self).__init__()
self.norm1 = norm(dim)
self.relu = nn.ReLU(inplace=True)
self.conv1 = ConcatConv1d(dim, dim, 3, 1, 1)
self.norm2 = norm(dim)
self.conv2 = ConcatConv1d(dim, dim, 3, 1, 1)
self.norm3 = norm(dim)
self.nfe = 0
def forward(self, t, x):
self.nfe += 1
out = self.norm1(x)
out = self.relu(out)
out = self.conv1(t, out)
out = self.norm2(out)
out = self.relu(out)
out = self.conv2(t, out)
out = self.norm3(out)
return out
def get_inputs():
return [torch.rand([4, 1, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'dim': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused_native_group_norm_relu_0(in_out_ptr0, in_ptr0, in_ptr1,
in_ptr2, out_ptr0, out_ptr1, xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
r3 = rindex // 4
tmp0 = tl.load(in_ptr0 + (r1 + 16 * x0), xmask, other=0.0)
tmp24 = tl.load(in_ptr1 + r3, None, eviction_policy='evict_last')
tmp26 = tl.load(in_ptr2 + r3, None, eviction_policy='evict_last')
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tl.where(xmask, tmp1, 0)
tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp6 = tl.where(xmask, tmp4, 0)
tmp7 = tl.sum(tmp6, 1)[:, None]
tmp8 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp9 = tmp8.to(tl.float32)
tmp10 = tmp7 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tl.broadcast_to(tmp12, [XBLOCK, RBLOCK])
tmp15 = tl.where(xmask, tmp13, 0)
tmp16 = tl.sum(tmp15, 1)[:, None]
tmp17 = 16.0
tmp18 = tmp16 / tmp17
tmp19 = 1e-05
tmp20 = tmp18 + tmp19
tmp21 = libdevice.rsqrt(tmp20)
tmp22 = tmp0 - tmp10
tmp23 = tmp22 * tmp21
tmp25 = tmp23 * tmp24
tmp27 = tmp25 + tmp26
tmp28 = tl.full([1, 1], 0, tl.int32)
tmp29 = triton_helpers.maximum(tmp28, tmp27)
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp21, xmask)
tl.store(out_ptr1 + (r1 + 20 * x0), tmp29, xmask)
tl.store(out_ptr0 + x0, tmp10, xmask)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK: tl.
constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tl.store(out_ptr0 + (x0 + 20 * x1), tmp0, xmask)
tl.store(out_ptr1 + (x0 + 20 * x1), tmp0, xmask)
@triton.jit
def triton_per_fused_convolution_native_group_norm_relu_2(in_out_ptr0,
in_out_ptr1, in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1, xnumel,
rnumel, XBLOCK: tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r3 = rindex
x0 = xindex
r2 = rindex // 4
tmp0 = tl.load(in_out_ptr0 + (r3 + 16 * x0), xmask, other=0.0)
tmp1 = tl.load(in_ptr0 + r2, None, eviction_policy='evict_last')
tmp26 = tl.load(in_ptr1 + r2, None, eviction_policy='evict_last')
tmp28 = tl.load(in_ptr2 + r2, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tl.where(xmask, tmp3, 0)
tmp6 = tl.broadcast_to(tmp3, [XBLOCK, RBLOCK])
tmp8 = tl.where(xmask, tmp6, 0)
tmp9 = tl.sum(tmp8, 1)[:, None]
tmp10 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp11 = tmp10.to(tl.float32)
tmp12 = tmp9 / tmp11
tmp13 = tmp3 - tmp12
tmp14 = tmp13 * tmp13
tmp15 = tl.broadcast_to(tmp14, [XBLOCK, RBLOCK])
tmp17 = tl.where(xmask, tmp15, 0)
tmp18 = tl.sum(tmp17, 1)[:, None]
tmp19 = 16.0
tmp20 = tmp18 / tmp19
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tmp24 = tmp2 - tmp12
tmp25 = tmp24 * tmp23
tmp27 = tmp25 * tmp26
tmp29 = tmp27 + tmp28
tmp30 = tl.full([1, 1], 0, tl.int32)
tmp31 = triton_helpers.maximum(tmp30, tmp29)
tl.store(in_out_ptr0 + (r3 + 16 * x0), tmp2, xmask)
tl.debug_barrier()
tl.store(in_out_ptr1 + x0, tmp23, xmask)
tl.store(out_ptr1 + (r3 + 20 * x0), tmp31, xmask)
tl.store(out_ptr0 + x0, tmp12, xmask)
@triton.jit
def triton_per_fused_convolution_native_group_norm_3(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, out_ptr0, out_ptr2, out_ptr3, xnumel, rnumel, XBLOCK:
tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r3 = rindex
x0 = xindex
r2 = rindex // 4
tmp0 = tl.load(in_out_ptr0 + (r3 + 16 * x0), xmask, other=0.0)
tmp1 = tl.load(in_ptr0 + r2, None, eviction_policy='evict_last')
tmp26 = tl.load(in_ptr1 + r2, None, eviction_policy='evict_last')
tmp28 = tl.load(in_ptr2 + r2, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tl.where(xmask, tmp3, 0)
tmp6 = tl.broadcast_to(tmp3, [XBLOCK, RBLOCK])
tmp8 = tl.where(xmask, tmp6, 0)
tmp9 = tl.sum(tmp8, 1)[:, None]
tmp10 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp11 = tmp10.to(tl.float32)
tmp12 = tmp9 / tmp11
tmp13 = tmp3 - tmp12
tmp14 = tmp13 * tmp13
tmp15 = tl.broadcast_to(tmp14, [XBLOCK, RBLOCK])
tmp17 = tl.where(xmask, tmp15, 0)
tmp18 = tl.sum(tmp17, 1)[:, None]
tmp19 = tmp2 - tmp12
tmp20 = 16.0
tmp21 = tmp18 / tmp20
tmp22 = 1e-05
tmp23 = tmp21 + tmp22
tmp24 = libdevice.rsqrt(tmp23)
tmp25 = tmp19 * tmp24
tmp27 = tmp25 * tmp26
tmp29 = tmp27 + tmp28
tl.store(in_out_ptr0 + (r3 + 16 * x0), tmp2, xmask)
tl.store(out_ptr2 + (r3 + 16 * x0), tmp29, xmask)
tl.store(out_ptr3 + x0, tmp24, xmask)
tl.store(out_ptr0 + x0, tmp12, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12
) = args
args.clear()
assert_size_stride(primals_1, (4,), (1,))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 1, 4), (4, 4, 1))
assert_size_stride(primals_5, (4, 5, 3), (15, 3, 1))
assert_size_stride(primals_6, (4,), (1,))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (4, 5, 3), (15, 3, 1))
assert_size_stride(primals_10, (4,), (1,))
assert_size_stride(primals_11, (4,), (1,))
assert_size_stride(primals_12, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 1, 1, 1), (1, 1, 1, 1), torch.float32)
buf1 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf3 = reinterpret_tensor(buf1, (4, 1, 1, 1), (1, 1, 1, 1), 0)
del buf1
buf6 = empty_strided_cuda((4, 5, 4), (20, 4, 1), torch.float32)
buf5 = reinterpret_tensor(buf6, (4, 4, 4), (20, 4, 1), 4)
get_raw_stream(0)
triton_per_fused_native_group_norm_relu_0[grid(4)](buf3, primals_3,
primals_1, primals_2, buf0, buf5, 4, 16, XBLOCK=1, num_warps=2,
num_stages=1)
buf4 = reinterpret_tensor(buf6, (4, 1, 4), (20, 4, 1), 0)
buf15 = empty_strided_cuda((4, 5, 4), (20, 4, 1), torch.float32)
buf13 = reinterpret_tensor(buf15, (4, 1, 4), (20, 4, 1), 0)
triton_poi_fused_cat_1[grid(16)](primals_4, buf4, buf13, 16, XBLOCK
=16, num_warps=1, num_stages=1)
del primals_4
buf7 = extern_kernels.convolution(buf6, primals_5, stride=(1,),
padding=(1,), dilation=(1,), transposed=False, output_padding=(
0,), groups=1, bias=None)
assert_size_stride(buf7, (4, 4, 4), (16, 4, 1))
buf8 = buf7
del buf7
buf9 = empty_strided_cuda((4, 1, 1, 1), (1, 1, 1, 1), torch.float32)
buf10 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf12 = reinterpret_tensor(buf10, (4, 1, 1, 1), (1, 1, 1, 1), 0)
del buf10
buf14 = reinterpret_tensor(buf15, (4, 4, 4), (20, 4, 1), 4)
triton_per_fused_convolution_native_group_norm_relu_2[grid(4)](buf8,
buf12, primals_6, primals_7, primals_8, buf9, buf14, 4, 16,
XBLOCK=1, num_warps=2, num_stages=1)
del primals_6
buf16 = extern_kernels.convolution(buf15, primals_9, stride=(1,),
padding=(1,), dilation=(1,), transposed=False, output_padding=(
0,), groups=1, bias=None)
assert_size_stride(buf16, (4, 4, 4), (16, 4, 1))
buf17 = buf16
del buf16
buf18 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
buf21 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf22 = empty_strided_cuda((4, 1, 1, 1), (1, 4, 4, 4), torch.float32)
triton_per_fused_convolution_native_group_norm_3[grid(4)](buf17,
primals_10, primals_11, primals_12, buf18, buf21, buf22, 4, 16,
XBLOCK=1, num_warps=2, num_stages=1)
del primals_10
del primals_12
return (buf21, primals_1, primals_2, primals_3, primals_5, primals_7,
primals_8, primals_9, primals_11, buf0, buf3, buf6, buf8, buf9,
buf12, buf15, buf17, reinterpret_tensor(buf18, (4, 1), (1, 1), 0),
reinterpret_tensor(buf22, (4, 1), (1, 1), 0))
def norm(dim):
"""
Group normalization to improve model accuracy and training speed.
"""
return nn.GroupNorm(min(1, dim), dim)
class ConcatConv1d(nn.Module):
"""
1d convolution concatenated with time for usage in ODENet.
"""
def __init__(self, dim_in, dim_out, kernel_size=3, stride=1, padding=0,
bias=True, transpose=False):
super(ConcatConv1d, self).__init__()
module = nn.ConvTranspose1d if transpose else nn.Conv1d
self._layer = module(dim_in + 1, dim_out, kernel_size=kernel_size,
stride=stride, padding=padding, bias=bias)
def forward(self, t, x):
tt = torch.ones_like(x[:, :1, :]) * t
ttx = torch.cat([tt, x], 1)
return self._layer(ttx)
class ODEfuncNew(nn.Module):
"""
Network architecture for ODENet.
"""
def __init__(self, dim):
super(ODEfuncNew, self).__init__()
self.norm1 = norm(dim)
self.relu = nn.ReLU(inplace=True)
self.conv1 = ConcatConv1d(dim, dim, 3, 1, 1)
self.norm2 = norm(dim)
self.conv2 = ConcatConv1d(dim, dim, 3, 1, 1)
self.norm3 = norm(dim)
self.nfe = 0
def forward(self, input_0, input_1):
primals_1 = self.norm1.weight
primals_2 = self.norm1.bias
primals_5 = self.conv1._layer.weight
primals_6 = self.conv1._layer.bias
primals_7 = self.norm2.weight
primals_8 = self.norm2.bias
primals_9 = self.conv2._layer.weight
primals_10 = self.conv2._layer.bias
primals_11 = self.norm3.weight
primals_12 = self.norm3.bias
primals_4 = input_0
primals_3 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12])
return output[0]
|
puneat/SS-using-NODE
|
ODEfunc
| false
| 4,146
|
[
"MIT"
] | 0
|
29f053769420a2d1cab1ad45f59a912c2ac737da
|
https://github.com/puneat/SS-using-NODE/tree/29f053769420a2d1cab1ad45f59a912c2ac737da
|
ConcatConv1d
|
import torch
import torch.nn as nn
class ConcatConv1d(nn.Module):
"""
1d convolution concatenated with time for usage in ODENet.
"""
def __init__(self, dim_in, dim_out, kernel_size=3, stride=1, padding=0,
bias=True, transpose=False):
super(ConcatConv1d, self).__init__()
module = nn.ConvTranspose1d if transpose else nn.Conv1d
self._layer = module(dim_in + 1, dim_out, kernel_size=kernel_size,
stride=stride, padding=padding, bias=bias)
def forward(self, t, x):
tt = torch.ones_like(x[:, :1, :]) * t
ttx = torch.cat([tt, x], 1)
return self._layer(ttx)
def get_inputs():
return [torch.rand([4, 1, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'dim_in': 4, 'dim_out': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 80
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4 % 5
x0 = xindex % 4
x2 = xindex // 20
x3 = xindex
tmp0 = x1
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + 4 * x2), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 5, tl.int64)
tmp9 = tl.load(in_ptr1 + (x0 + 4 * (-1 + x1) + 16 * x2), tmp6 & xmask,
other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x3, tmp10, xmask)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 2 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 1, 4), (4, 4, 1))
assert_size_stride(primals_3, (4, 5, 3), (15, 3, 1))
assert_size_stride(primals_4, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 5, 4), (20, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(80)](primals_2, primals_1, buf0, 80,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_1
del primals_2
buf1 = extern_kernels.convolution(buf0, primals_3, stride=(1,),
padding=(0,), dilation=(1,), transposed=False, output_padding=(
0,), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 2), (8, 2, 1))
buf2 = buf1
del buf1
triton_poi_fused_convolution_1[grid(32)](buf2, primals_4, 32,
XBLOCK=32, num_warps=1, num_stages=1)
del primals_4
return buf2, primals_3, buf0
class ConcatConv1dNew(nn.Module):
"""
1d convolution concatenated with time for usage in ODENet.
"""
def __init__(self, dim_in, dim_out, kernel_size=3, stride=1, padding=0,
bias=True, transpose=False):
super(ConcatConv1dNew, self).__init__()
module = nn.ConvTranspose1d if transpose else nn.Conv1d
self._layer = module(dim_in + 1, dim_out, kernel_size=kernel_size,
stride=stride, padding=padding, bias=bias)
def forward(self, input_0, input_1):
primals_3 = self._layer.weight
primals_4 = self._layer.bias
primals_2 = input_0
primals_1 = input_1
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0]
|
puneat/SS-using-NODE
|
ConcatConv1d
| false
| 4,147
|
[
"MIT"
] | 0
|
29f053769420a2d1cab1ad45f59a912c2ac737da
|
https://github.com/puneat/SS-using-NODE/tree/29f053769420a2d1cab1ad45f59a912c2ac737da
|
AdversarialNetwork
|
import torch
import torch.nn as nn
class AdversarialNetwork(nn.Module):
def __init__(self, in_feature):
super(AdversarialNetwork, self).__init__()
self.ad_layer1 = nn.Linear(in_feature, 32)
self.ad_layer2 = nn.Linear(32, 32)
self.ad_layer3 = nn.Linear(32, 1)
self.ad_layer1.weight.data.normal_(0, 0.01)
self.ad_layer2.weight.data.normal_(0, 0.01)
self.ad_layer3.weight.data.normal_(0, 0.3)
self.ad_layer1.bias.data.fill_(0.0)
self.ad_layer2.bias.data.fill_(0.0)
self.ad_layer3.bias.data.fill_(0.0)
self.relu1 = nn.LeakyReLU()
self.relu2 = nn.LeakyReLU()
self.dropout1 = nn.Dropout(0.5)
self.dropout2 = nn.Dropout(0.5)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
x = self.ad_layer1(x)
x = self.relu1(x)
x = self.dropout1(x)
x = self.ad_layer2(x)
x = self.relu2(x)
x = self.dropout2(x)
x = self.ad_layer3(x)
x = self.sigmoid(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_feature': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_leaky_relu_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 32
tmp0 = tl.load(in_ptr0 + x2, None)
tmp1 = tl.load(in_ptr1 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp5 = 0.01
tmp6 = tmp2 * tmp5
tmp7 = tl.where(tmp4, tmp2, tmp6)
tl.store(out_ptr0 + x2, tmp4, None)
tl.store(out_ptr1 + x2, tmp7, None)
@triton.jit
def triton_poi_fused_sigmoid_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr0 + 0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK])
tmp3 = tmp0 + tmp2
tmp4 = tl.sigmoid(tmp3)
tl.store(in_out_ptr0 + x0, tmp4, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (32, 4), (4, 1))
assert_size_stride(primals_2, (32,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (32, 32), (32, 1))
assert_size_stride(primals_5, (32,), (1,))
assert_size_stride(primals_6, (1, 32), (32, 1))
assert_size_stride(primals_7, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 32), (32, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 32), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((4, 4, 4, 32), (512, 128, 32, 1), torch.bool)
buf2 = empty_strided_cuda((4, 4, 4, 32), (512, 128, 32, 1), torch.
float32)
get_raw_stream(0)
triton_poi_fused_leaky_relu_0[grid(2048)](buf0, primals_2, buf1,
buf2, 2048, XBLOCK=256, num_warps=4, num_stages=1)
del primals_2
buf3 = buf0
del buf0
extern_kernels.mm(reinterpret_tensor(buf2, (64, 32), (32, 1), 0),
reinterpret_tensor(primals_4, (32, 32), (1, 32), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 32), (512, 128, 32, 1), torch.bool)
buf5 = empty_strided_cuda((4, 4, 4, 32), (512, 128, 32, 1), torch.
float32)
triton_poi_fused_leaky_relu_0[grid(2048)](buf3, primals_5, buf4,
buf5, 2048, XBLOCK=256, num_warps=4, num_stages=1)
del buf3
del primals_5
buf6 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf5, (64, 32), (32, 1), 0),
reinterpret_tensor(primals_6, (32, 1), (1, 32), 0), out=buf6)
buf7 = reinterpret_tensor(buf6, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf6
triton_poi_fused_sigmoid_1[grid(64)](buf7, primals_7, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del primals_7
return buf7, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf1, reinterpret_tensor(buf2, (64, 32), (32, 1), 0
), buf4, reinterpret_tensor(buf5, (64, 32), (32, 1), 0
), buf7, primals_6, primals_4
class AdversarialNetworkNew(nn.Module):
def __init__(self, in_feature):
super(AdversarialNetworkNew, self).__init__()
self.ad_layer1 = nn.Linear(in_feature, 32)
self.ad_layer2 = nn.Linear(32, 32)
self.ad_layer3 = nn.Linear(32, 1)
self.ad_layer1.weight.data.normal_(0, 0.01)
self.ad_layer2.weight.data.normal_(0, 0.01)
self.ad_layer3.weight.data.normal_(0, 0.3)
self.ad_layer1.bias.data.fill_(0.0)
self.ad_layer2.bias.data.fill_(0.0)
self.ad_layer3.bias.data.fill_(0.0)
self.relu1 = nn.LeakyReLU()
self.relu2 = nn.LeakyReLU()
self.dropout1 = nn.Dropout(0.5)
self.dropout2 = nn.Dropout(0.5)
self.sigmoid = nn.Sigmoid()
def forward(self, input_0):
primals_1 = self.ad_layer1.weight
primals_2 = self.ad_layer1.bias
primals_4 = self.ad_layer2.weight
primals_5 = self.ad_layer2.bias
primals_6 = self.ad_layer3.weight
primals_7 = self.ad_layer3.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
|
pwjworks/MS-MDA
|
AdversarialNetwork
| false
| 4,148
|
[
"MIT"
] | 0
|
21f921a933a318820239541adb26b9fc6feba699
|
https://github.com/pwjworks/MS-MDA/tree/21f921a933a318820239541adb26b9fc6feba699
|
CollaborativeAttention
|
import math
import torch
import torch.utils.data
from enum import Enum
import torch.nn as nn
class MixingMatrixInit(Enum):
CONCATENATE = 1
ALL_ONES = 2
UNIFORM = 3
class CollaborativeAttention(nn.Module):
def __init__(self, dim_input: 'int', dim_value_all: 'int',
dim_key_query_all: 'int', num_attention_heads: 'int',
mixing_initialization: 'MixingMatrixInit'=MixingMatrixInit.UNIFORM):
super().__init__()
if dim_value_all % num_attention_heads != 0:
raise ValueError(
'Value dimension ({}) should be divisible by number of heads ({})'
.format(dim_value_all, num_attention_heads))
self.dim_input = dim_input
self.dim_value_all = dim_value_all
self.dim_key_query_all = dim_key_query_all
self.num_attention_heads = num_attention_heads
self.mixing_initialization = mixing_initialization
self.dim_value_per_head = dim_value_all // num_attention_heads
self.attention_head_size = dim_key_query_all / num_attention_heads
self.query = nn.Linear(dim_input, dim_key_query_all, bias=False)
self.key = nn.Linear(dim_input, dim_key_query_all, bias=False)
self.value = nn.Linear(dim_input, dim_value_all)
self.mixing = self.init_mixing_matrix()
def forward(self, hidden_states, attention_mask=None, head_mask=None,
encoder_hidden_states=None, encoder_attention_mask=None):
from_sequence = hidden_states
to_sequence = hidden_states
if encoder_hidden_states is not None:
to_sequence = encoder_hidden_states
attention_mask = encoder_attention_mask
query_layer = self.query(from_sequence)
key_layer = self.key(to_sequence)
mixed_query = query_layer[..., None, :, :] * self.mixing[..., :,
None, :]
mixed_key = key_layer[..., None, :, :]
attention_scores = torch.matmul(mixed_query, mixed_key.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
if head_mask is not None:
attention_probs = attention_probs * head_mask
value_layer = self.value(to_sequence)
value_layer = self.transpose_for_scores(value_layer)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.
dim_value_all,)
context_layer = context_layer.view(*new_context_layer_shape)
return context_layer
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, -1)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def init_mixing_matrix(self, scale=0.2):
mixing = torch.zeros(self.num_attention_heads, self.dim_key_query_all)
if self.mixing_initialization is MixingMatrixInit.CONCATENATE:
dim_head = int(math.ceil(self.dim_key_query_all / self.
num_attention_heads))
for i in range(self.num_attention_heads):
mixing[i, i * dim_head:(i + 1) * dim_head] = 1.0
elif self.mixing_initialization is MixingMatrixInit.ALL_ONES:
mixing.one_()
elif self.mixing_initialization is MixingMatrixInit.UNIFORM:
mixing.normal_(std=scale)
else:
raise ValueError('Unknown mixing matrix initialization: {}'.
format(self.mixing_initialization))
return nn.Parameter(mixing)
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'dim_input': 4, 'dim_value_all': 4, 'dim_key_query_all': 4,
'num_attention_heads': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import math
import torch.utils.data
from enum import Enum
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex // 64
x4 = xindex % 16
x0 = xindex % 4
x2 = xindex // 16 % 4
x5 = xindex
tmp0 = tl.load(in_ptr0 + (x4 + 16 * x3), xmask, eviction_policy=
'evict_last')
tmp1 = tl.load(in_ptr1 + (x0 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x5, tmp2, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x1 + 4 * x0 + 16 * x3), xmask,
eviction_policy='evict_last')
tl.store(out_ptr0 + x4, tmp0, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp3 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = 1.0
tmp2 = tmp0 * tmp1
tmp4 = tmp3 * tmp1
tmp6 = tmp5 * tmp1
tmp7 = triton_helpers.maximum(tmp4, tmp6)
tmp9 = tmp8 * tmp1
tmp10 = triton_helpers.maximum(tmp7, tmp9)
tmp12 = tmp11 * tmp1
tmp13 = triton_helpers.maximum(tmp10, tmp12)
tmp14 = tmp2 - tmp13
tmp15 = tmp14 * tmp1
tmp16 = tl_math.exp(tmp15)
tl.store(out_ptr0 + x2, tmp16, xmask)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused_clone_5(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4, 4), (4, 1))
assert_size_stride(primals_6, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), out=buf0)
del primals_2
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_3, (4, 4), (1, 4), 0), out=buf1)
del primals_3
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(256)](buf0, primals_4, buf2, 256,
XBLOCK=256, num_warps=4, num_stages=1)
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_1[grid(256)](buf1, buf3, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf4 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1), 0), out=buf4)
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__softmax_2[grid(256)](buf4, buf5, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf6 = reinterpret_tensor(buf4, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf4
triton_poi_fused__softmax_3[grid(256)](buf5, buf6, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf5
buf7 = buf1
del buf1
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_5, (4, 4), (1, 4), 0), out=buf7)
del primals_5
buf8 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
triton_poi_fused_clone_4[grid(16, 4)](buf7, primals_6, buf8, 16, 4,
XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1)
del primals_6
buf9 = reinterpret_tensor(buf7, (16, 4, 1), (4, 1, 1), 0)
del buf7
extern_kernels.bmm(reinterpret_tensor(buf6, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf8, (16, 4, 1), (4, 1, 0), 0), out=buf9)
buf10 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
triton_poi_fused_clone_5[grid(16, 4)](buf9, buf10, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
del buf9
return reinterpret_tensor(buf10, (4, 4, 4), (16, 4, 1), 0
), primals_4, reinterpret_tensor(primals_1, (16, 4), (4, 1), 0
), buf0, buf6, reinterpret_tensor(buf8, (16, 1, 4), (4, 1, 1), 0
), reinterpret_tensor(buf2, (16, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf3, (16, 4, 4), (16, 1, 4), 0)
class MixingMatrixInit(Enum):
CONCATENATE = 1
ALL_ONES = 2
UNIFORM = 3
class CollaborativeAttentionNew(nn.Module):
def __init__(self, dim_input: 'int', dim_value_all: 'int',
dim_key_query_all: 'int', num_attention_heads: 'int',
mixing_initialization: 'MixingMatrixInit'=MixingMatrixInit.UNIFORM):
super().__init__()
if dim_value_all % num_attention_heads != 0:
raise ValueError(
'Value dimension ({}) should be divisible by number of heads ({})'
.format(dim_value_all, num_attention_heads))
self.dim_input = dim_input
self.dim_value_all = dim_value_all
self.dim_key_query_all = dim_key_query_all
self.num_attention_heads = num_attention_heads
self.mixing_initialization = mixing_initialization
self.dim_value_per_head = dim_value_all // num_attention_heads
self.attention_head_size = dim_key_query_all / num_attention_heads
self.query = nn.Linear(dim_input, dim_key_query_all, bias=False)
self.key = nn.Linear(dim_input, dim_key_query_all, bias=False)
self.value = nn.Linear(dim_input, dim_value_all)
self.mixing = self.init_mixing_matrix()
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, -1)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def init_mixing_matrix(self, scale=0.2):
mixing = torch.zeros(self.num_attention_heads, self.dim_key_query_all)
if self.mixing_initialization is MixingMatrixInit.CONCATENATE:
dim_head = int(math.ceil(self.dim_key_query_all / self.
num_attention_heads))
for i in range(self.num_attention_heads):
mixing[i, i * dim_head:(i + 1) * dim_head] = 1.0
elif self.mixing_initialization is MixingMatrixInit.ALL_ONES:
mixing.one_()
elif self.mixing_initialization is MixingMatrixInit.UNIFORM:
mixing.normal_(std=scale)
else:
raise ValueError('Unknown mixing matrix initialization: {}'.
format(self.mixing_initialization))
return nn.Parameter(mixing)
def forward(self, input_0):
primals_2 = self.mixing
primals_3 = self.query.weight
primals_4 = self.key.weight
primals_5 = self.value.weight
primals_6 = self.value.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6])
return output[0]
|
prattcmp/NonAttentiveTacotron2
|
CollaborativeAttention
| false
| 4,149
|
[
"BSD-3-Clause"
] | 0
|
c65722133c392fba233b5003b480ee498fc0a44a
|
https://github.com/prattcmp/NonAttentiveTacotron2/tree/c65722133c392fba233b5003b480ee498fc0a44a
|
UpSample
|
import torch
import torch.nn as nn
import torch.nn.functional as F
class UpSample(nn.Sequential):
def __init__(self, skip_input, output_features):
super(UpSample, self).__init__()
self.convA = nn.Conv2d(skip_input, output_features, kernel_size=3,
stride=1, padding=1)
self.leakyreluA = nn.LeakyReLU(0.2)
self.convB = nn.Conv2d(output_features, output_features,
kernel_size=3, stride=1, padding=1)
self.leakyreluB = nn.LeakyReLU(0.2)
def forward(self, x, concat_with):
x = F.interpolate(x, size=[concat_with.size(2), concat_with.size(3)
], mode='bilinear', align_corners=True)
x = self.convA(torch.cat([x, concat_with], dim=1))
x = self.leakyreluA(x)
x = self.convB(x)
x = self.leakyreluB(x)
return x
def get_inputs():
return [torch.rand([4, 1, 4, 4]), torch.rand([4, 3, 4, 4])]
def get_init_inputs():
return [[], {'skip_input': 4, 'output_features': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused__to_copy__unsafe_index_add_arange_clamp_mul_sub_0(in_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4 % 4
x0 = xindex % 4
x2 = xindex // 16
x3 = xindex % 16
tmp0 = x1
tmp1 = tmp0.to(tl.float32)
tmp2 = 1.0
tmp3 = tmp1 * tmp2
tmp4 = 0.0
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp6 = tmp5.to(tl.int32)
tmp7 = tl.full([1], 1, tl.int64)
tmp8 = tmp6 + tmp7
tmp9 = tl.full([1], 3, tl.int64)
tmp10 = triton_helpers.minimum(tmp8, tmp9)
tmp11 = x0
tmp12 = tmp11.to(tl.float32)
tmp13 = tmp12 * tmp2
tmp14 = triton_helpers.maximum(tmp13, tmp4)
tmp15 = tmp14.to(tl.int32)
tmp16 = tl.load(in_ptr0 + (tmp15 + 4 * tmp10 + 16 * x2), xmask,
eviction_policy='evict_last')
tmp17 = tmp15 + tmp7
tmp18 = triton_helpers.minimum(tmp17, tmp9)
tmp19 = tl.load(in_ptr0 + (tmp18 + 4 * tmp10 + 16 * x2), xmask,
eviction_policy='evict_last')
tmp20 = tmp19 - tmp16
tmp21 = tmp15.to(tl.float32)
tmp22 = tmp14 - tmp21
tmp23 = triton_helpers.maximum(tmp22, tmp4)
tmp24 = triton_helpers.minimum(tmp23, tmp2)
tmp25 = tmp20 * tmp24
tmp26 = tmp16 + tmp25
tmp27 = tl.load(in_ptr0 + (tmp15 + 4 * tmp6 + 16 * x2), xmask,
eviction_policy='evict_last')
tmp28 = tl.load(in_ptr0 + (tmp18 + 4 * tmp6 + 16 * x2), xmask,
eviction_policy='evict_last')
tmp29 = tmp28 - tmp27
tmp30 = tmp29 * tmp24
tmp31 = tmp27 + tmp30
tmp32 = tmp26 - tmp31
tmp33 = tmp6.to(tl.float32)
tmp34 = tmp5 - tmp33
tmp35 = triton_helpers.maximum(tmp34, tmp4)
tmp36 = triton_helpers.minimum(tmp35, tmp2)
tmp37 = tmp32 * tmp36
tmp38 = tmp31 + tmp37
tl.store(out_ptr1 + (x3 + 64 * x2), tmp38, xmask)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 192
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 48
x1 = xindex // 48
tmp0 = tl.load(in_ptr0 + x2, xmask)
tl.store(out_ptr0 + (x0 + 64 * x1), tmp0, xmask)
@triton.jit
def triton_poi_fused_convolution_leaky_relu_2(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 4
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp5 = 0.2
tmp6 = tmp2 * tmp5
tmp7 = tl.where(tmp4, tmp2, tmp6)
tl.store(out_ptr0 + x3, tmp4, xmask)
tl.store(out_ptr1 + x3, tmp7, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6 = args
args.clear()
assert_size_stride(primals_1, (4, 3, 4, 4), (48, 16, 4, 1))
assert_size_stride(primals_2, (4, 1, 4, 4), (16, 16, 4, 1))
assert_size_stride(primals_3, (4, 4, 3, 3), (36, 9, 3, 1))
assert_size_stride(primals_4, (4,), (1,))
assert_size_stride(primals_5, (4, 4, 3, 3), (36, 9, 3, 1))
assert_size_stride(primals_6, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf1 = reinterpret_tensor(buf3, (4, 1, 4, 4), (64, 16, 4, 1), 0)
get_raw_stream(0)
triton_poi_fused__to_copy__unsafe_index_add_arange_clamp_mul_sub_0[grid
(64)](primals_2, buf1, 64, XBLOCK=64, num_warps=1, num_stages=1)
del primals_2
buf2 = reinterpret_tensor(buf3, (4, 3, 4, 4), (64, 16, 4, 1), 16)
triton_poi_fused_cat_1[grid(192)](primals_1, buf2, 192, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_1
buf4 = extern_kernels.convolution(buf3, primals_3, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf4, (4, 4, 4, 4), (64, 16, 4, 1))
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_convolution_leaky_relu_2[grid(256)](buf4,
primals_4, buf5, buf6, 256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_4
buf7 = extern_kernels.convolution(buf6, primals_5, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf7, (4, 4, 4, 4), (64, 16, 4, 1))
buf8 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
buf9 = buf4
del buf4
triton_poi_fused_convolution_leaky_relu_2[grid(256)](buf7,
primals_6, buf8, buf9, 256, XBLOCK=256, num_warps=4, num_stages=1)
del buf7
del primals_6
return buf9, primals_3, primals_5, buf3, buf5, buf6, buf8
class UpSampleNew(nn.Sequential):
def __init__(self, skip_input, output_features):
super(UpSampleNew, self).__init__()
self.convA = nn.Conv2d(skip_input, output_features, kernel_size=3,
stride=1, padding=1)
self.leakyreluA = nn.LeakyReLU(0.2)
self.convB = nn.Conv2d(output_features, output_features,
kernel_size=3, stride=1, padding=1)
self.leakyreluB = nn.LeakyReLU(0.2)
def forward(self, input_0, input_1):
primals_3 = self.convA.weight
primals_4 = self.convA.bias
primals_5 = self.convB.weight
primals_6 = self.convB.bias
primals_2 = input_0
primals_1 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6])
return output[0]
|
pystokes/depth_estimation
|
UpSample
| false
| 4,150
|
[
"MIT"
] | 0
|
b5b1955bcb5b3f1a1f1c8ddde45431cf38514f90
|
https://github.com/pystokes/depth_estimation/tree/b5b1955bcb5b3f1a1f1c8ddde45431cf38514f90
|
SelfExpression
|
import torch
import torch.nn as nn
class SelfExpression(nn.Module):
def __init__(self, n):
super(SelfExpression, self).__init__()
self.Coefficient = nn.Parameter(0.0001 * torch.ones(n, n, dtype=
torch.float32), requires_grad=True)
def forward(self, x):
y = torch.matmul(self.Coefficient, x)
return y
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'n': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(64, 4)](primals_2, buf0, 64, 4,
XBLOCK=4, YBLOCK=32, num_warps=4, num_stages=1)
del primals_2
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf1)
del primals_1
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_0[grid(64, 4)](buf1, buf2, 64, 4, XBLOCK=4,
YBLOCK=32, num_warps=4, num_stages=1)
del buf1
return buf2, reinterpret_tensor(buf0, (64, 4), (4, 1), 0)
class SelfExpressionNew(nn.Module):
def __init__(self, n):
super(SelfExpressionNew, self).__init__()
self.Coefficient = nn.Parameter(0.0001 * torch.ones(n, n, dtype=
torch.float32), requires_grad=True)
def forward(self, input_0):
primals_1 = self.Coefficient
primals_2 = input_0
output = call([primals_1, primals_2])
return output[0]
|
qilinli/DSC-Net
|
SelfExpression
| false
| 4,151
|
[
"MIT"
] | 0
|
c0e7a3cae3e07c34b2989234f568c7007cf0fc55
|
https://github.com/qilinli/DSC-Net/tree/c0e7a3cae3e07c34b2989234f568c7007cf0fc55
|
Net
|
import torch
import torch.nn as nn
import torch.nn.functional as F
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(3, 16, 3, stride=3)
self.conv2 = nn.Conv2d(16, 32, 3, stride=3)
self.conv3 = nn.Conv2d(32, 64, 3, stride=3)
self.pool = nn.MaxPool2d(2, 2)
self.fc1 = nn.Linear(64, 500)
self.fc2 = nn.Linear(500, 256)
self.fc3 = nn.Linear(256, 2)
self.drop_out = nn.Dropout(0.25)
def forward(self, x):
x = self.pool(F.relu(self.conv1(x)))
x = self.pool(F.relu(self.conv2(x)))
x = self.pool(F.relu(self.conv3(x)))
x = x.view(-1, 64)
x = self.drop_out(x)
x = F.relu(self.fc1(x))
x = self.drop_out(x)
x = F.relu(self.fc2(x))
x = self.drop_out(x)
x = self.fc3(x)
return x
def get_inputs():
return [torch.rand([4, 3, 256, 256])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_convolution_relu_0(in_ptr0, in_ptr1, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 462400
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 7225 % 16
x0 = xindex % 7225
x4 = xindex // 7225
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(out_ptr0 + (x0 + 7232 * x4), tmp4, xmask)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_1(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 112896
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 42
x1 = xindex // 42 % 42
x2 = xindex // 1764
x3 = xindex % 1764
tmp0 = tl.load(in_ptr0 + (2 * x0 + 170 * x1 + 7232 * x2), xmask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 170 * x1 + 7232 * x2), xmask,
eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (85 + 2 * x0 + 170 * x1 + 7232 * x2), xmask,
eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (86 + 2 * x0 + 170 * x1 + 7232 * x2), xmask,
eviction_policy='evict_last')
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1], 1, tl.int8)
tmp9 = tl.full([1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + (x3 + 1792 * x2), tmp6, xmask)
tl.store(out_ptr1 + (x3 + 1792 * x2), tmp16, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_2(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 25088
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 196 % 32
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, xmask)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_3(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 6272
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 7
x1 = xindex // 7
x4 = xindex
x3 = xindex // 1568
x5 = xindex % 1568
tmp0 = tl.load(in_ptr0 + (2 * x0 + 28 * x1), xmask, eviction_policy=
'evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 28 * x1), xmask, eviction_policy
='evict_last')
tmp3 = tl.load(in_ptr0 + (14 + 2 * x0 + 28 * x1), xmask,
eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (15 + 2 * x0 + 28 * x1), xmask,
eviction_policy='evict_last')
tmp2 = triton_helpers.maximum(tmp1, tmp0)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tmp7 = tmp1 > tmp0
tmp8 = tl.full([1], 1, tl.int8)
tmp9 = tl.full([1], 0, tl.int8)
tmp10 = tl.where(tmp7, tmp8, tmp9)
tmp11 = tmp3 > tmp2
tmp12 = tl.full([1], 2, tl.int8)
tmp13 = tl.where(tmp11, tmp12, tmp10)
tmp14 = tmp5 > tmp4
tmp15 = tl.full([1], 3, tl.int8)
tmp16 = tl.where(tmp14, tmp15, tmp13)
tl.store(out_ptr0 + x4, tmp6, xmask)
tl.store(out_ptr1 + (x5 + 1664 * x3), tmp16, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_4(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 4 % 64
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, xmask)
@triton.jit
def triton_poi_fused_max_pool2d_with_indices_5(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp1 > tmp0
tmp3 = tl.full([1], 1, tl.int8)
tmp4 = tl.full([1], 0, tl.int8)
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = triton_helpers.maximum(tmp1, tmp0)
tmp8 = tmp7 > tmp6
tmp9 = tl.full([1], 2, tl.int8)
tmp10 = tl.where(tmp8, tmp9, tmp5)
tmp11 = triton_helpers.maximum(tmp7, tmp6)
tmp13 = tmp12 > tmp11
tmp14 = tl.full([1], 3, tl.int8)
tmp15 = tl.where(tmp13, tmp14, tmp10)
tmp16 = triton_helpers.maximum(tmp12, tmp11)
tl.store(out_ptr0 + x0, tmp15, xmask)
tl.store(out_ptr1 + x0, tmp16, xmask)
@triton.jit
def triton_poi_fused_relu_6(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 2000
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 500
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_relu_7(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 256
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13) = args
args.clear()
assert_size_stride(primals_1, (16, 3, 3, 3), (27, 9, 3, 1))
assert_size_stride(primals_2, (16,), (1,))
assert_size_stride(primals_3, (4, 3, 256, 256), (196608, 65536, 256, 1))
assert_size_stride(primals_4, (32, 16, 3, 3), (144, 9, 3, 1))
assert_size_stride(primals_5, (32,), (1,))
assert_size_stride(primals_6, (64, 32, 3, 3), (288, 9, 3, 1))
assert_size_stride(primals_7, (64,), (1,))
assert_size_stride(primals_8, (500, 64), (64, 1))
assert_size_stride(primals_9, (500,), (1,))
assert_size_stride(primals_10, (256, 500), (500, 1))
assert_size_stride(primals_11, (256,), (1,))
assert_size_stride(primals_12, (2, 256), (256, 1))
assert_size_stride(primals_13, (2,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(3,
3), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 16, 85, 85), (115600, 7225, 85, 1))
buf1 = empty_strided_cuda((4, 16, 85, 85), (115712, 7232, 85, 1),
torch.float32)
get_raw_stream(0)
triton_poi_fused_convolution_relu_0[grid(462400)](buf0, primals_2,
buf1, 462400, XBLOCK=1024, num_warps=4, num_stages=1)
del buf0
del primals_2
buf2 = empty_strided_cuda((4, 16, 42, 42), (28672, 1792, 42, 1),
torch.float32)
buf3 = empty_strided_cuda((4, 16, 42, 42), (28672, 1792, 42, 1),
torch.int8)
triton_poi_fused_max_pool2d_with_indices_1[grid(112896)](buf1, buf2,
buf3, 112896, XBLOCK=512, num_warps=8, num_stages=1)
buf4 = extern_kernels.convolution(buf2, primals_4, stride=(3, 3),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf4, (4, 32, 14, 14), (6272, 196, 14, 1))
buf5 = buf4
del buf4
triton_poi_fused_convolution_relu_2[grid(25088)](buf5, primals_5,
25088, XBLOCK=256, num_warps=4, num_stages=1)
del primals_5
buf6 = empty_strided_cuda((4, 32, 7, 7), (1568, 49, 7, 1), torch.
float32)
buf7 = empty_strided_cuda((4, 32, 7, 7), (1664, 49, 7, 1), torch.int8)
triton_poi_fused_max_pool2d_with_indices_3[grid(6272)](buf5, buf6,
buf7, 6272, XBLOCK=256, num_warps=4, num_stages=1)
buf8 = extern_kernels.convolution(buf6, primals_6, stride=(3, 3),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf8, (4, 64, 2, 2), (256, 4, 2, 1))
buf9 = buf8
del buf8
triton_poi_fused_convolution_relu_4[grid(1024)](buf9, primals_7,
1024, XBLOCK=256, num_warps=4, num_stages=1)
del primals_7
buf10 = empty_strided_cuda((4, 64, 1, 1), (64, 1, 1, 1), torch.int8)
buf11 = empty_strided_cuda((4, 64, 1, 1), (64, 1, 256, 256), torch.
float32)
triton_poi_fused_max_pool2d_with_indices_5[grid(256)](buf9, buf10,
buf11, 256, XBLOCK=128, num_warps=4, num_stages=1)
buf12 = empty_strided_cuda((4, 500), (500, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf11, (4, 64), (64, 1), 0),
reinterpret_tensor(primals_8, (64, 500), (1, 64), 0), out=buf12)
buf13 = buf12
del buf12
triton_poi_fused_relu_6[grid(2000)](buf13, primals_9, 2000, XBLOCK=
256, num_warps=4, num_stages=1)
del primals_9
buf14 = empty_strided_cuda((4, 256), (256, 1), torch.float32)
extern_kernels.mm(buf13, reinterpret_tensor(primals_10, (500, 256),
(1, 500), 0), out=buf14)
buf15 = buf14
del buf14
triton_poi_fused_relu_7[grid(1024)](buf15, primals_11, 1024, XBLOCK
=256, num_warps=4, num_stages=1)
del primals_11
buf16 = empty_strided_cuda((4, 2), (2, 1), torch.float32)
extern_kernels.addmm(primals_13, buf15, reinterpret_tensor(
primals_12, (256, 2), (1, 256), 0), alpha=1, beta=1, out=buf16)
del primals_13
return (buf16, primals_1, primals_3, primals_4, primals_6, buf1, buf2,
buf3, buf5, buf6, buf7, buf9, buf10, reinterpret_tensor(buf11, (4,
64), (64, 1), 0), buf13, buf15, primals_12, primals_10, primals_8)
class NetNew(nn.Module):
def __init__(self):
super(NetNew, self).__init__()
self.conv1 = nn.Conv2d(3, 16, 3, stride=3)
self.conv2 = nn.Conv2d(16, 32, 3, stride=3)
self.conv3 = nn.Conv2d(32, 64, 3, stride=3)
self.pool = nn.MaxPool2d(2, 2)
self.fc1 = nn.Linear(64, 500)
self.fc2 = nn.Linear(500, 256)
self.fc3 = nn.Linear(256, 2)
self.drop_out = nn.Dropout(0.25)
def forward(self, input_0):
primals_1 = self.conv1.weight
primals_2 = self.conv1.bias
primals_4 = self.conv2.weight
primals_5 = self.conv2.bias
primals_6 = self.conv3.weight
primals_7 = self.conv3.bias
primals_8 = self.fc1.weight
primals_9 = self.fc1.bias
primals_10 = self.fc2.weight
primals_11 = self.fc2.bias
primals_12 = self.fc3.weight
primals_13 = self.fc3.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13])
return output[0]
|
prasad5141/cat_vs_dog_webapp
|
Net
| false
| 4,152
|
[
"MIT"
] | 0
|
29c82addbc62104c3b9250af5f465b269cf68039
|
https://github.com/prasad5141/cat_vs_dog_webapp/tree/29c82addbc62104c3b9250af5f465b269cf68039
|
LearnedPositionalEmbedding
|
import torch
import torch.nn as nn
import torch.nn.functional as F
class LearnedPositionalEmbedding(nn.Embedding):
"""
This module learns positional embeddings up to a fixed maximum size.
Padding ids are ignored by either offsetting based on padding_idx
or by setting padding_idx to None and ensuring that the appropriate
position ids are passed to the forward function.
"""
def __init__(self, num_embeddings: 'int', embedding_dim: 'int',
padding_idx: 'int'):
if padding_idx is not None:
num_embeddings_ = num_embeddings + padding_idx + 1
else:
num_embeddings_ = num_embeddings
super().__init__(num_embeddings_, embedding_dim, padding_idx)
self.max_positions = num_embeddings
def forward(self, input: 'torch.Tensor'):
"""Input is expected to be of size [bsz x seqlen]."""
mask = input.ne(self.padding_idx).int()
positions = (torch.cumsum(mask, dim=1).type_as(mask) * mask).long(
) + self.padding_idx
return F.embedding(positions, self.weight, self.padding_idx, self.
max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'num_embeddings': 4, 'embedding_dim': 4, 'padding_idx': 4}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def _triton_helper_fn_add0(arg0_0, arg1_0):
tmp0 = arg0_0 + arg1_0
return tmp0
@triton.jit
def triton_per_fused__to_copy_cumsum_ne_0(in_ptr0, out_ptr0, xnumel, rnumel,
XBLOCK: tl.constexpr):
xnumel = 64
RBLOCK: tl.constexpr = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r2 = rindex
x0 = xindex % 16
x1 = xindex // 16
tmp0 = tl.load(in_ptr0 + (x0 + 16 * r2 + 64 * x1), xmask, other=0.0)
tmp1 = 4.0
tmp2 = tmp0 != tmp1
tmp3 = tmp2.to(tl.int32)
tmp4 = tmp3.to(tl.int64)
tmp5 = tmp4.to(tl.int64)
tmp6 = tl.broadcast_to(tmp5, [XBLOCK, RBLOCK])
tmp7, = tl.associative_scan((tmp6,), 1, _triton_helper_fn_add0)
tl.store(out_ptr0 + (x0 + 16 * r2 + 64 * x1), tmp7, xmask)
@triton.jit
def triton_poi_fused__to_copy_add_mul_ne_1(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp2 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tmp0.to(tl.int32)
tmp3 = 4.0
tmp4 = tmp2 != tmp3
tmp5 = tmp4.to(tl.int32)
tmp6 = tmp1 * tmp5
tmp7 = tmp6.to(tl.int64)
tmp8 = tl.full([1], 4, tl.int64)
tmp9 = tmp7 + tmp8
tl.store(in_out_ptr0 + x0, tmp9, xmask)
@triton.jit
def triton_poi_fused_embedding_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x0 = xindex % 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp1 = tl.full([XBLOCK], 9, tl.int32)
tmp2 = tmp0 + tmp1
tmp3 = tmp0 < 0
tmp4 = tl.where(tmp3, tmp2, tmp0)
tl.device_assert((0 <= tmp4) & (tmp4 < 9) | ~xmask,
'index out of bounds: 0 <= tmp4 < 9')
tmp6 = tl.load(in_ptr1 + (x0 + 4 * tmp4), xmask)
tl.store(out_ptr0 + x2, tmp6, xmask)
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (9, 4), (4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.int64)
get_raw_stream(0)
triton_per_fused__to_copy_cumsum_ne_0[grid(64)](primals_1, buf0, 64,
4, XBLOCK=1, num_warps=2, num_stages=1)
buf1 = buf0
del buf0
triton_poi_fused__to_copy_add_mul_ne_1[grid(256)](buf1, primals_1,
256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_1
buf2 = empty_strided_cuda((4, 4, 4, 4, 4), (256, 64, 16, 4, 1),
torch.float32)
triton_poi_fused_embedding_2[grid(1024)](buf1, primals_2, buf2,
1024, XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
return buf2, buf1
class LearnedPositionalEmbeddingNew(nn.Embedding):
"""
This module learns positional embeddings up to a fixed maximum size.
Padding ids are ignored by either offsetting based on padding_idx
or by setting padding_idx to None and ensuring that the appropriate
position ids are passed to the forward function.
"""
def __init__(self, num_embeddings: 'int', embedding_dim: 'int',
padding_idx: 'int'):
if padding_idx is not None:
num_embeddings_ = num_embeddings + padding_idx + 1
else:
num_embeddings_ = num_embeddings
super().__init__(num_embeddings_, embedding_dim, padding_idx)
self.max_positions = num_embeddings
def forward(self, input_0):
primals_2 = self.weight
primals_1 = input_0
output = call([primals_1, primals_2])
return output[0]
|
qinwang-ai/Contact-Distil
|
LearnedPositionalEmbedding
| false
| 4,153
|
[
"Apache-2.0"
] | 0
|
5e98389de70e0d9c4d16bd91ca1326689dc220a6
|
https://github.com/qinwang-ai/Contact-Distil/tree/5e98389de70e0d9c4d16bd91ca1326689dc220a6
|
ConvAE
|
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
class Conv2dSamePad(nn.Module):
"""
Implement Tensorflow's 'SAME' padding mode in Conv2d.
When an odd number, say `m`, of pixels are need to pad, Tensorflow will pad one more column at right or one more
row at bottom. But Pytorch will pad `m+1` pixels, i.e., Pytorch always pads in both sides.
So we can pad the tensor in the way of Tensorflow before call the Conv2d module.
"""
def __init__(self, kernel_size, stride):
super(Conv2dSamePad, self).__init__()
self.kernel_size = kernel_size if type(kernel_size) in [list, tuple
] else [kernel_size, kernel_size]
self.stride = stride if type(stride) in [list, tuple] else [stride,
stride]
def forward(self, x):
in_height = x.size(2)
in_width = x.size(3)
out_height = math.ceil(float(in_height) / float(self.stride[0]))
out_width = math.ceil(float(in_width) / float(self.stride[1]))
pad_along_height = (out_height - 1) * self.stride[0
] + self.kernel_size[0] - in_height
pad_along_width = (out_width - 1) * self.stride[1] + self.kernel_size[1
] - in_width
pad_top = math.floor(pad_along_height / 2)
pad_left = math.floor(pad_along_width / 2)
pad_bottom = pad_along_height - pad_top
pad_right = pad_along_width - pad_left
return F.pad(x, [pad_left, pad_right, pad_top, pad_bottom],
'constant', 0)
class ConvTranspose2dSamePad(nn.Module):
"""
This module implements the "SAME" padding mode for ConvTranspose2d as in Tensorflow.
A tensor with width w_in, feed it to ConvTranspose2d(ci, co, kernel, stride), the width of output tensor T_nopad:
w_nopad = (w_in - 1) * stride + kernel
If we use padding, i.e., ConvTranspose2d(ci, co, kernel, stride, padding, output_padding), the width of T_pad:
w_pad = (w_in - 1) * stride + kernel - (2*padding - output_padding) = w_nopad - (2*padding - output_padding)
Yes, in ConvTranspose2d, more padding, the resulting tensor is smaller, i.e., the padding is actually deleting row/col.
If `pad`=(2*padding - output_padding) is odd, Pytorch deletes more columns in the left, i.e., the first ceil(pad/2) and
last `pad - ceil(pad/2)` columns of T_nopad are deleted to get T_pad.
In contrast, Tensorflow deletes more columns in the right, i.e., the first floor(pad/2) and last `pad - floor(pad/2)`
columns are deleted.
For the height, Pytorch deletes more rows at top, while Tensorflow at bottom.
In practice, we usually want `w_pad = w_in * stride` or `w_pad = w_in * stride - 1`, i.e., the "SAME" padding mode
in Tensorflow. To determine the value of `w_pad`, we should pass it to this function.
So the number of columns to delete:
pad = 2*padding - output_padding = w_nopad - w_pad
If pad is even, we can directly set padding=pad/2 and output_padding=0 in ConvTranspose2d.
If pad is odd, we can use ConvTranspose2d to get T_nopad, and then delete `pad` rows/columns by
ourselves.
This module should be called after the ConvTranspose2d module with shared kernel_size and stride values.
"""
def __init__(self, output_size):
super(ConvTranspose2dSamePad, self).__init__()
self.output_size = output_size
def forward(self, x):
in_height = x.size(2)
in_width = x.size(3)
pad_height = in_height - self.output_size[0]
pad_width = in_width - self.output_size[1]
pad_top = pad_height // 2
pad_bottom = pad_height - pad_top
pad_left = pad_width // 2
pad_right = pad_width - pad_left
return x[:, :, pad_top:in_height - pad_bottom, pad_left:in_width -
pad_right]
class ConvAE(nn.Module):
def __init__(self, channels, kernels):
"""
:param channels: a list containing all channels including the input image channel (1 for gray, 3 for RGB)
:param kernels: a list containing all kernel sizes, it should satisfy: len(kernels) = len(channels) - 1.
"""
super(ConvAE, self).__init__()
assert isinstance(channels, list) and isinstance(kernels, list)
self.encoder = nn.Sequential()
for i in range(1, len(channels)):
self.encoder.add_module('pad%d' % i, Conv2dSamePad(kernels[i -
1], 2))
self.encoder.add_module('conv%d' % i, nn.Conv2d(channels[i - 1],
channels[i], kernel_size=kernels[i - 1], stride=2))
self.encoder.add_module('relu%d' % i, nn.ReLU(True))
self.decoder = nn.Sequential()
channels = list(reversed(channels))
kernels = list(reversed(kernels))
sizes = [[12, 11], [24, 21], [48, 42]]
for i in range(len(channels) - 1):
self.decoder.add_module('deconv%d' % (i + 1), nn.
ConvTranspose2d(channels[i], channels[i + 1], kernel_size=
kernels[i], stride=2))
self.decoder.add_module('padd%d' % i, ConvTranspose2dSamePad(
sizes[i]))
self.decoder.add_module('relud%d' % i, nn.ReLU(True))
def forward(self, x):
h = self.encoder(x)
y = self.decoder(h)
return y
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'channels': [4, 4], 'kernels': [4, 4]}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import math
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_constant_pad_nd_0(in_ptr0, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 576
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 6 % 6
x0 = xindex % 6
x2 = xindex // 36
x4 = xindex
tmp0 = -1 + x1
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = -1 + x0
tmp6 = tmp5 >= tmp1
tmp7 = tmp5 < tmp3
tmp8 = tmp2 & tmp4
tmp9 = tmp8 & tmp6
tmp10 = tmp9 & tmp7
tmp11 = tl.load(in_ptr0 + (-5 + x0 + 4 * x1 + 16 * x2), tmp10 & xmask,
other=0.0)
tl.store(out_ptr0 + x4, tmp11, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_1(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 4 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x3, tmp4, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_2(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 576
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 6 % 6
x0 = xindex % 6
x4 = xindex
x2 = xindex // 36 % 4
tmp19 = tl.load(in_out_ptr0 + x4, xmask)
tmp20 = tl.load(in_ptr0 + x2, xmask, eviction_policy='evict_last')
tmp0 = x1
tmp1 = tl.full([1], 3, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = x0
tmp4 = tmp3 >= tmp1
tmp5 = tmp4 & tmp2
tmp6 = tl.load(in_out_ptr0 + x4, tmp5 & xmask, other=0.0)
tmp7 = tl.load(in_ptr0 + x2, tmp5 & xmask, eviction_policy='evict_last',
other=0.0)
tmp8 = tmp6 + tmp7
tmp9 = tl.full([1], 0, tl.int32)
tmp10 = triton_helpers.maximum(tmp9, tmp8)
tmp11 = tl.full(tmp10.shape, 0.0, tmp10.dtype)
tmp12 = tl.where(tmp5, tmp10, tmp11)
tmp13 = tl.load(in_out_ptr0 + x4, tmp2 & xmask, other=0.0)
tmp14 = tl.load(in_ptr0 + x2, tmp2 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp15 = tmp13 + tmp14
tmp16 = tl.where(tmp4, tmp12, tmp15)
tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)
tmp18 = tl.where(tmp2, tmp16, tmp17)
tmp21 = tmp19 + tmp20
tmp22 = tl.where(tmp2, tmp18, tmp21)
tl.store(in_out_ptr0 + x4, tmp22, xmask)
@triton.jit
def triton_poi_fused_threshold_backward_3(in_ptr0, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 144
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 3
x1 = xindex // 3 % 3
x2 = xindex // 9
x3 = xindex
tmp0 = tl.load(in_ptr0 + (21 + x0 + 6 * x1 + 36 * x2), xmask)
tmp1 = 0.0
tmp2 = tmp0 <= tmp1
tl.store(out_ptr0 + x3, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 6, 6), (144, 36, 6, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_constant_pad_nd_0[grid(576)](primals_1, buf0, 576,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_1
buf1 = extern_kernels.convolution(buf0, primals_2, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 2, 2), (16, 4, 2, 1))
buf2 = buf1
del buf1
triton_poi_fused_convolution_relu_1[grid(64)](buf2, primals_3, 64,
XBLOCK=64, num_warps=1, num_stages=1)
del primals_3
buf3 = extern_kernels.convolution(buf2, primals_4, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=True,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf3, (4, 4, 6, 6), (144, 36, 6, 1))
buf4 = buf3
del buf3
triton_poi_fused_convolution_relu_2[grid(576)](buf4, primals_5, 576,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_5
buf5 = empty_strided_cuda((4, 4, 3, 3), (36, 9, 3, 1), torch.bool)
triton_poi_fused_threshold_backward_3[grid(144)](buf4, buf5, 144,
XBLOCK=128, num_warps=4, num_stages=1)
return reinterpret_tensor(buf4, (4, 4, 3, 3), (144, 36, 6, 1), 21
), primals_2, primals_4, buf0, buf2, buf5
class Conv2dSamePad(nn.Module):
"""
Implement Tensorflow's 'SAME' padding mode in Conv2d.
When an odd number, say `m`, of pixels are need to pad, Tensorflow will pad one more column at right or one more
row at bottom. But Pytorch will pad `m+1` pixels, i.e., Pytorch always pads in both sides.
So we can pad the tensor in the way of Tensorflow before call the Conv2d module.
"""
def __init__(self, kernel_size, stride):
super(Conv2dSamePad, self).__init__()
self.kernel_size = kernel_size if type(kernel_size) in [list, tuple
] else [kernel_size, kernel_size]
self.stride = stride if type(stride) in [list, tuple] else [stride,
stride]
def forward(self, x):
in_height = x.size(2)
in_width = x.size(3)
out_height = math.ceil(float(in_height) / float(self.stride[0]))
out_width = math.ceil(float(in_width) / float(self.stride[1]))
pad_along_height = (out_height - 1) * self.stride[0
] + self.kernel_size[0] - in_height
pad_along_width = (out_width - 1) * self.stride[1] + self.kernel_size[1
] - in_width
pad_top = math.floor(pad_along_height / 2)
pad_left = math.floor(pad_along_width / 2)
pad_bottom = pad_along_height - pad_top
pad_right = pad_along_width - pad_left
return F.pad(x, [pad_left, pad_right, pad_top, pad_bottom],
'constant', 0)
class ConvTranspose2dSamePad(nn.Module):
"""
This module implements the "SAME" padding mode for ConvTranspose2d as in Tensorflow.
A tensor with width w_in, feed it to ConvTranspose2d(ci, co, kernel, stride), the width of output tensor T_nopad:
w_nopad = (w_in - 1) * stride + kernel
If we use padding, i.e., ConvTranspose2d(ci, co, kernel, stride, padding, output_padding), the width of T_pad:
w_pad = (w_in - 1) * stride + kernel - (2*padding - output_padding) = w_nopad - (2*padding - output_padding)
Yes, in ConvTranspose2d, more padding, the resulting tensor is smaller, i.e., the padding is actually deleting row/col.
If `pad`=(2*padding - output_padding) is odd, Pytorch deletes more columns in the left, i.e., the first ceil(pad/2) and
last `pad - ceil(pad/2)` columns of T_nopad are deleted to get T_pad.
In contrast, Tensorflow deletes more columns in the right, i.e., the first floor(pad/2) and last `pad - floor(pad/2)`
columns are deleted.
For the height, Pytorch deletes more rows at top, while Tensorflow at bottom.
In practice, we usually want `w_pad = w_in * stride` or `w_pad = w_in * stride - 1`, i.e., the "SAME" padding mode
in Tensorflow. To determine the value of `w_pad`, we should pass it to this function.
So the number of columns to delete:
pad = 2*padding - output_padding = w_nopad - w_pad
If pad is even, we can directly set padding=pad/2 and output_padding=0 in ConvTranspose2d.
If pad is odd, we can use ConvTranspose2d to get T_nopad, and then delete `pad` rows/columns by
ourselves.
This module should be called after the ConvTranspose2d module with shared kernel_size and stride values.
"""
def __init__(self, output_size):
super(ConvTranspose2dSamePad, self).__init__()
self.output_size = output_size
def forward(self, x):
in_height = x.size(2)
in_width = x.size(3)
pad_height = in_height - self.output_size[0]
pad_width = in_width - self.output_size[1]
pad_top = pad_height // 2
pad_bottom = pad_height - pad_top
pad_left = pad_width // 2
pad_right = pad_width - pad_left
return x[:, :, pad_top:in_height - pad_bottom, pad_left:in_width -
pad_right]
class ConvAENew(nn.Module):
def __init__(self, channels, kernels):
"""
:param channels: a list containing all channels including the input image channel (1 for gray, 3 for RGB)
:param kernels: a list containing all kernel sizes, it should satisfy: len(kernels) = len(channels) - 1.
"""
super(ConvAENew, self).__init__()
assert isinstance(channels, list) and isinstance(kernels, list)
self.encoder = nn.Sequential()
for i in range(1, len(channels)):
self.encoder.add_module('pad%d' % i, Conv2dSamePad(kernels[i -
1], 2))
self.encoder.add_module('conv%d' % i, nn.Conv2d(channels[i - 1],
channels[i], kernel_size=kernels[i - 1], stride=2))
self.encoder.add_module('relu%d' % i, nn.ReLU(True))
self.decoder = nn.Sequential()
channels = list(reversed(channels))
kernels = list(reversed(kernels))
sizes = [[12, 11], [24, 21], [48, 42]]
for i in range(len(channels) - 1):
self.decoder.add_module('deconv%d' % (i + 1), nn.
ConvTranspose2d(channels[i], channels[i + 1], kernel_size=
kernels[i], stride=2))
self.decoder.add_module('padd%d' % i, ConvTranspose2dSamePad(
sizes[i]))
self.decoder.add_module('relud%d' % i, nn.ReLU(True))
def forward(self, input_0):
primals_1 = self.encoder.conv1.weight
primals_3 = self.encoder.conv1.bias
primals_2 = self.decoder.deconv1.weight
primals_5 = self.decoder.deconv1.bias
primals_4 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
qilinli/DSC-Net
|
ConvAE
| false
| 4,154
|
[
"MIT"
] | 0
|
c0e7a3cae3e07c34b2989234f568c7007cf0fc55
|
https://github.com/qilinli/DSC-Net/tree/c0e7a3cae3e07c34b2989234f568c7007cf0fc55
|
MultiHeadedAttention
|
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
class MultiHeadedAttention(nn.Module):
def __init__(self, num_head, d_model, dropout=0.1):
super(MultiHeadedAttention, self).__init__()
assert d_model % num_head == 0
self.d_k = d_model // num_head
self.h = num_head
self.linear_key = nn.Linear(d_model, d_model)
self.linear_value = nn.Linear(d_model, d_model)
self.linear_query = nn.Linear(d_model, d_model)
self.linear_out = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(p=dropout)
def attention(self, query, key, value, mask, dropout=None):
d_k = query.size(-1)
scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
scores = scores.masked_fill(mask == 0, -1000000000.0)
p_attn = F.softmax(scores, dim=-1)
if dropout is not None:
p_attn = dropout(p_attn)
return torch.matmul(p_attn, value), p_attn
def forward(self, query, key, value, mask):
nbatches = query.size(0)
query = self.linear_query(query).view(nbatches, -1, self.h, self.d_k
).transpose(1, 2)
key = self.linear_key(key).view(nbatches, -1, self.h, self.d_k
).transpose(1, 2)
value = self.linear_value(value).view(nbatches, -1, self.h, self.d_k
).transpose(1, 2)
mask = mask.unsqueeze(1)
x, _attn = self.attention(query, key, value, mask, dropout=self.dropout
)
x = x.transpose(1, 2).contiguous().view(nbatches, -1, self.h * self.d_k
)
return self.linear_out(x)
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4,
4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'num_head': 4, 'd_model': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import math
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused_eq_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.0
tmp2 = tmp0 == tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
@triton.jit
def triton_poi_fused__softmax_div_masked_fill_2(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex // 16
x3 = xindex
tmp0 = tl.load(in_ptr0 + (4 * x0 + 16 * x2), xmask, eviction_policy=
'evict_last').to(tl.int1)
tmp1 = tl.load(in_ptr1 + 4 * x3, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (1 + 4 * x0 + 16 * x2), xmask, eviction_policy
='evict_last').to(tl.int1)
tmp7 = tl.load(in_ptr1 + (1 + 4 * x3), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (2 + 4 * x0 + 16 * x2), xmask,
eviction_policy='evict_last').to(tl.int1)
tmp12 = tl.load(in_ptr1 + (2 + 4 * x3), xmask, eviction_policy='evict_last'
)
tmp16 = tl.load(in_ptr0 + (3 + 4 * x0 + 16 * x2), xmask,
eviction_policy='evict_last').to(tl.int1)
tmp17 = tl.load(in_ptr1 + (3 + 4 * x3), xmask, eviction_policy='evict_last'
)
tmp2 = 1.0
tmp3 = tmp1 * tmp2
tmp4 = -1000000000.0
tmp5 = tl.where(tmp0, tmp4, tmp3)
tmp8 = tmp7 * tmp2
tmp9 = tl.where(tmp6, tmp4, tmp8)
tmp10 = triton_helpers.maximum(tmp5, tmp9)
tmp13 = tmp12 * tmp2
tmp14 = tl.where(tmp11, tmp4, tmp13)
tmp15 = triton_helpers.maximum(tmp10, tmp14)
tmp18 = tmp17 * tmp2
tmp19 = tl.where(tmp16, tmp4, tmp18)
tmp20 = triton_helpers.maximum(tmp15, tmp19)
tmp21 = tmp5 - tmp20
tmp22 = tl_math.exp(tmp21)
tmp23 = tmp9 - tmp20
tmp24 = tl_math.exp(tmp23)
tmp25 = tmp22 + tmp24
tmp26 = tmp14 - tmp20
tmp27 = tl_math.exp(tmp26)
tmp28 = tmp25 + tmp27
tmp29 = tmp19 - tmp20
tmp30 = tl_math.exp(tmp29)
tmp31 = tmp28 + tmp30
tl.store(out_ptr0 + x3, tmp20, xmask)
tl.store(out_ptr1 + x3, tmp31, xmask)
@triton.jit
def triton_poi_fused__softmax_div_masked_fill_3(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex // 64
x4 = xindex % 16
x5 = xindex
x6 = xindex // 4
tmp0 = tl.load(in_ptr0 + (x4 + 16 * x3), xmask, eviction_policy=
'evict_last').to(tl.int1)
tmp1 = tl.load(in_out_ptr0 + x5, xmask)
tmp6 = tl.load(in_ptr1 + x6, xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + x6, xmask, eviction_policy='evict_last')
tmp2 = 1.0
tmp3 = tmp1 * tmp2
tmp4 = -1000000000.0
tmp5 = tl.where(tmp0, tmp4, tmp3)
tmp7 = tmp5 - tmp6
tmp8 = tl_math.exp(tmp7)
tmp10 = tmp8 / tmp9
tl.store(in_out_ptr0 + x5, tmp10, xmask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12
) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_7, (4, 4), (4, 1))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_10, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_11, (4, 4), (4, 1))
assert_size_stride(primals_12, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), out=buf0)
del primals_2
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_6, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 4), (1, 4), 0), out=buf1)
del primals_4
buf2 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_9, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 4), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(16, 4)](buf0, primals_3, buf3, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_3
buf4 = reinterpret_tensor(buf0, (4, 4, 1, 4), (16, 4, 4, 1), 0)
del buf0
triton_poi_fused_clone_0[grid(16, 4)](buf1, primals_5, buf4, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_5
buf5 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf4, (16, 1, 4), (4, 0, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 1, 4, 4), (16, 16, 4, 1), torch.bool)
triton_poi_fused_eq_1[grid(64)](primals_10, buf6, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del primals_10
buf7 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 64), 0)
del buf1
buf8 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
triton_poi_fused__softmax_div_masked_fill_2[grid(64)](buf6, buf5,
buf7, buf8, 64, XBLOCK=64, num_warps=1, num_stages=1)
buf9 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused__softmax_div_masked_fill_3[grid(256)](buf9, buf6,
buf7, buf8, 256, XBLOCK=128, num_warps=4, num_stages=1)
buf10 = reinterpret_tensor(buf8, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf8
triton_poi_fused_clone_0[grid(16, 4)](buf2, primals_8, buf10, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_8
buf11 = reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf9, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf10, (16, 4, 1), (4, 1, 0), 0), out=buf11)
buf12 = reinterpret_tensor(buf7, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf7
triton_poi_fused_clone_4[grid(16, 4)](buf11, buf12, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf13 = reinterpret_tensor(buf11, (16, 4), (4, 1), 0)
del buf11
extern_kernels.addmm(primals_12, reinterpret_tensor(buf12, (16, 4),
(4, 1), 0), reinterpret_tensor(primals_11, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf13)
del primals_12
return reinterpret_tensor(buf13, (4, 4, 4), (16, 4, 1), 0
), reinterpret_tensor(primals_1, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_6, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0
), buf6, buf9, reinterpret_tensor(buf12, (16, 4), (4, 1), 0
), primals_11, reinterpret_tensor(buf10, (16, 1, 4), (4, 1, 1), 0
), reinterpret_tensor(buf3, (16, 1, 4), (4, 1, 1), 0
), reinterpret_tensor(buf4, (16, 4, 1), (4, 1, 4), 0)
class MultiHeadedAttentionNew(nn.Module):
def __init__(self, num_head, d_model, dropout=0.1):
super(MultiHeadedAttentionNew, self).__init__()
assert d_model % num_head == 0
self.d_k = d_model // num_head
self.h = num_head
self.linear_key = nn.Linear(d_model, d_model)
self.linear_value = nn.Linear(d_model, d_model)
self.linear_query = nn.Linear(d_model, d_model)
self.linear_out = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(p=dropout)
def attention(self, query, key, value, mask, dropout=None):
d_k = query.size(-1)
scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
scores = scores.masked_fill(mask == 0, -1000000000.0)
p_attn = F.softmax(scores, dim=-1)
if dropout is not None:
p_attn = dropout(p_attn)
return torch.matmul(p_attn, value), p_attn
def forward(self, input_0, input_1, input_2, input_3):
primals_2 = self.linear_key.weight
primals_3 = self.linear_key.bias
primals_4 = self.linear_value.weight
primals_5 = self.linear_value.bias
primals_7 = self.linear_query.weight
primals_8 = self.linear_query.bias
primals_11 = self.linear_out.weight
primals_12 = self.linear_out.bias
primals_1 = input_0
primals_6 = input_1
primals_9 = input_2
primals_10 = input_3
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12])
return output[0]
|
qi700/my_point_summarize
|
MultiHeadedAttention
| false
| 4,155
|
[
"Apache-2.0"
] | 0
|
e269c2d0411fc61ea34055c3080472bc9111bcaa
|
https://github.com/qi700/my_point_summarize/tree/e269c2d0411fc61ea34055c3080472bc9111bcaa
|
Attention
|
import torch
import torch.utils.data
from torch import nn
import torch.nn.functional as F
import torch.hub
class Attention(nn.Module):
def forward(self, query, key, value, mask=None, dropout=None):
scale = query.size(-1) ** -0.5
scores = query.matmul(key.transpose(-2, -1)) / scale
if mask is not None:
scores = scores.masked_fill(mask == 0, -1000000000.0)
p_attn = F.softmax(scores, dim=-1)
if dropout is not None:
p_attn = dropout(p_attn)
return p_attn.matmul(value), p_attn
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.utils.data
from torch import nn
import torch.hub
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused__softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp3 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = 1.0
tmp2 = tmp0 * tmp1
tmp4 = tmp3 * tmp1
tmp6 = tmp5 * tmp1
tmp7 = triton_helpers.maximum(tmp4, tmp6)
tmp9 = tmp8 * tmp1
tmp10 = triton_helpers.maximum(tmp7, tmp9)
tmp12 = tmp11 * tmp1
tmp13 = triton_helpers.maximum(tmp10, tmp12)
tmp14 = tmp2 - tmp13
tmp15 = 2.0
tmp16 = tmp14 * tmp15
tmp17 = tl_math.exp(tmp16)
tl.store(out_ptr0 + x2, tmp17, xmask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
def call(args):
arg0_1, arg1_1, arg2_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg2_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(arg0_1, (16, 4, 4), (16, 4, 1
), 0), reinterpret_tensor(arg1_1, (16, 4, 4), (16, 1, 4), 0),
out=buf0)
del arg0_1
del arg1_1
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__softmax_0[grid(256)](buf0, buf1, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf2 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
triton_poi_fused__softmax_1[grid(256)](buf1, buf2, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf3 = reinterpret_tensor(buf1, (16, 4, 4), (16, 4, 1), 0)
del buf1
extern_kernels.bmm(reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(arg2_1, (16, 4, 4), (16, 4, 1), 0), out=buf3
)
del arg2_1
return reinterpret_tensor(buf3, (4, 4, 4, 4), (64, 16, 4, 1), 0), buf2
class AttentionNew(nn.Module):
def forward(self, input_0, input_1, input_2):
arg0_1 = input_0
arg1_1 = input_1
arg2_1 = input_2
output = call([arg0_1, arg1_1, arg2_1])
return output[0], output[1]
|
opqi/VMZ
|
Attention
| false
| 4,156
|
[
"Apache-2.0"
] | 0
|
bc9c3bf5f7d9e7d0ef433f9d9b4a3155ac5ed969
|
https://github.com/opqi/VMZ/tree/bc9c3bf5f7d9e7d0ef433f9d9b4a3155ac5ed969
|
MultiHeadAttention
|
import torch
import torch.nn as nn
class MultiHeadAttention(nn.Module):
def __init__(self, hidden_state, num_heads=1):
super().__init__()
self.q_linear = nn.Linear(hidden_state, hidden_state)
self.v_linear = nn.Linear(hidden_state, hidden_state)
self.k_linear = nn.Linear(hidden_state, hidden_state)
self.attention = nn.MultiheadAttention(hidden_state, num_heads)
def forward(self, query_input, input, mask=None):
query = self.q_linear(query_input)
key = self.k_linear(input)
value = self.v_linear(input)
attn_output, _attn_output_weights = self.attention(query, key,
value, mask)
return attn_output
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'hidden_state': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.5
tmp4 = tmp2 * tmp3
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12
) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4, 4), (4, 1))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (12, 4), (4, 1))
assert_size_stride(primals_10, (12,), (1,))
assert_size_stride(primals_11, (4, 4), (4, 1))
assert_size_stride(primals_12, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_2, primals_3, reinterpret_tensor(
primals_1, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_5, primals_6, reinterpret_tensor(
primals_4, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf1)
del primals_4
del primals_5
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_8, primals_6, reinterpret_tensor(
primals_7, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf2)
del primals_7
del primals_8
buf3 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.mm(buf0, reinterpret_tensor(primals_9, (4, 4), (1, 4
), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(reinterpret_tensor(primals_10, (4,), (1,), 4),
buf1, reinterpret_tensor(primals_9, (4, 4), (1, 4), 16), alpha=
1, beta=1, out=buf4)
buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(reinterpret_tensor(primals_10, (4,), (1,), 8),
buf2, reinterpret_tensor(primals_9, (4, 4), (1, 4), 32), alpha=
1, beta=1, out=buf5)
buf6 = reinterpret_tensor(buf3, (1, 4, 4), (16, 4, 1), 0)
del buf3
get_raw_stream(0)
triton_poi_fused_mul_0[grid(16)](buf6, primals_10, 16, XBLOCK=16,
num_warps=1, num_stages=1)
del primals_10
buf7 = empty_strided_cuda((1, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(buf6, reinterpret_tensor(buf4, (1, 4, 4), (4, 1,
4), 0), out=buf7)
buf8 = empty_strided_cuda((1, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused__softmax_1[grid(16)](buf7, buf8, 16, XBLOCK=16,
num_warps=1, num_stages=1)
buf9 = buf7
del buf7
triton_poi_fused__softmax_2[grid(16)](buf8, buf9, 16, XBLOCK=16,
num_warps=1, num_stages=1)
buf10 = buf8
del buf8
extern_kernels.bmm(buf9, reinterpret_tensor(buf5, (1, 4, 4), (4, 4,
1), 0), out=buf10)
buf11 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_12, reinterpret_tensor(buf10, (4, 4),
(4, 1), 0), reinterpret_tensor(primals_11, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf11)
del primals_12
return (buf11, primals_3, primals_6, buf0, buf1, buf2, buf9,
reinterpret_tensor(buf10, (4, 4), (4, 1), 0), primals_11,
reinterpret_tensor(buf5, (1, 4, 4), (4, 1, 4), 0),
reinterpret_tensor(buf6, (1, 4, 4), (16, 1, 4), 0),
reinterpret_tensor(buf4, (1, 4, 4), (4, 4, 1), 0),
reinterpret_tensor(primals_9, (4, 4), (4, 1), 32),
reinterpret_tensor(primals_9, (4, 4), (4, 1), 16),
reinterpret_tensor(primals_9, (4, 4), (4, 1), 0))
class MultiHeadAttentionNew(nn.Module):
def __init__(self, hidden_state, num_heads=1):
super().__init__()
self.q_linear = nn.Linear(hidden_state, hidden_state)
self.v_linear = nn.Linear(hidden_state, hidden_state)
self.k_linear = nn.Linear(hidden_state, hidden_state)
self.attention = nn.MultiheadAttention(hidden_state, num_heads)
def forward(self, input_0, input_1):
primals_1 = self.q_linear.weight
primals_2 = self.q_linear.bias
primals_3 = self.v_linear.weight
primals_5 = self.v_linear.bias
primals_4 = self.k_linear.weight
primals_8 = self.k_linear.bias
primals_9 = self.attention.in_proj_weight
primals_10 = self.attention.in_proj_bias
primals_6 = self.attention.out_proj.weight
primals_12 = self.attention.out_proj.bias
primals_7 = input_0
primals_11 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12])
return output[0]
|
qinyiwei/MuTual
|
MultiHeadAttention
| false
| 4,157
|
[
"MIT"
] | 0
|
3bdd13c1388d6136b8944666dfd434870760cc93
|
https://github.com/qinyiwei/MuTual/tree/3bdd13c1388d6136b8944666dfd434870760cc93
|
_SubPixelBlock
|
import torch
import torch.nn as nn
class _SubPixelBlock(nn.Module):
def __init__(self, in_channels: 'int'=64, out_channels: 'int'=64,
scale_factor: 'int'=2):
super(_SubPixelBlock, self).__init__()
n_out = out_channels * scale_factor ** 2
self.conv = nn.Conv2d(in_channels, n_out, kernel_size=3, stride=1,
padding=1)
self.shuffle = nn.PixelShuffle(scale_factor)
self.prelu = nn.PReLU(out_channels)
def forward(self, x):
hid = self.conv(x)
hid = self.shuffle(hid)
out = self.prelu(hid)
return out
def get_inputs():
return [torch.rand([4, 64, 64, 64])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 64
y1 = yindex // 64
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 64 * x2 + 576 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 256
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
x2 = xindex
y3 = yindex
y0 = yindex % 64
y1 = yindex // 64
tmp0 = tl.load(in_ptr0 + (x2 + 4096 * y3), ymask, eviction_policy=
'evict_last')
tl.store(out_ptr0 + (y0 + 64 * x2 + 262144 * y1), tmp0, ymask)
@triton.jit
def triton_poi_fused_convolution_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 256
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, None)
@triton.jit
def triton_poi_fused__prelu_kernel_3(in_ptr0, in_ptr1, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex % 128
x1 = xindex // 128 % 128
x2 = xindex // 16384 % 64
x3 = xindex // 1048576
x4 = xindex
tmp0 = tl.load(in_ptr0 + (2 * (x1 % 2) + 4 * x2 + 256 * (x0 // 2) +
16384 * (x1 // 2) + 1048576 * x3 + x0 % 2), None)
tmp3 = tl.load(in_ptr1 + x2, None, eviction_policy='evict_last')
tmp1 = 0.0
tmp2 = tmp0 > tmp1
tmp4 = tmp3 * tmp0
tmp5 = tl.where(tmp2, tmp0, tmp4)
tl.store(out_ptr0 + x4, tmp5, None)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (256, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_2, (256,), (1,))
assert_size_stride(primals_3, (4, 64, 64, 64), (262144, 4096, 64, 1))
assert_size_stride(primals_4, (64,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((256, 64, 3, 3), (576, 1, 192, 64), torch
.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(16384, 9)](primals_1, buf0, 16384, 9,
XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((4, 64, 64, 64), (262144, 1, 4096, 64),
torch.float32)
triton_poi_fused_1[grid(256, 4096)](primals_3, buf1, 256, 4096,
XBLOCK=32, YBLOCK=32, num_warps=4, num_stages=1)
del primals_3
buf2 = extern_kernels.convolution(buf1, buf0, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf2, (4, 256, 64, 64), (1048576, 1, 16384, 256))
buf3 = buf2
del buf2
triton_poi_fused_convolution_2[grid(4194304)](buf3, primals_2,
4194304, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_2
buf4 = empty_strided_cuda((4, 64, 128, 128), (1048576, 16384, 128,
1), torch.float32)
triton_poi_fused__prelu_kernel_3[grid(4194304)](buf3, primals_4,
buf4, 4194304, XBLOCK=1024, num_warps=4, num_stages=1)
return buf4, buf0, buf1, primals_4, buf3
class _SubPixelBlockNew(nn.Module):
def __init__(self, in_channels: 'int'=64, out_channels: 'int'=64,
scale_factor: 'int'=2):
super(_SubPixelBlockNew, self).__init__()
n_out = out_channels * scale_factor ** 2
self.conv = nn.Conv2d(in_channels, n_out, kernel_size=3, stride=1,
padding=1)
self.shuffle = nn.PixelShuffle(scale_factor)
self.prelu = nn.PReLU(out_channels)
def forward(self, input_0):
primals_1 = self.conv.weight
primals_2 = self.conv.bias
primals_4 = self.prelu.weight
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0]
|
pvrancx/torch_isr
|
_SubPixelBlock
| false
| 4,158
|
[
"MIT"
] | 0
|
831278ae5c3b939b4147bae1a99bc3f3d4fc415d
|
https://github.com/pvrancx/torch_isr/tree/831278ae5c3b939b4147bae1a99bc3f3d4fc415d
|
LocalContextNorm
|
import math
import torch
import torch.utils.data
from torchvision.transforms import functional as F
from torch import nn
from torch.nn import functional as F
class LocalContextNorm(nn.Module):
def __init__(self, num_features, channels_per_group=2, window_size=(227,
227), eps=1e-05):
super(LocalContextNorm, self).__init__()
self.weight = nn.Parameter(torch.ones(1, num_features, 1, 1))
self.bias = nn.Parameter(torch.zeros(1, num_features, 1, 1))
self.channels_per_group = channels_per_group
self.eps = eps
self.window_size = window_size
def forward(self, x):
N, C, H, W = x.size()
G = C // self.channels_per_group
assert C % self.channels_per_group == 0
if self.window_size[0] < H and self.window_size[1] < W:
torch.device(torch.cuda.current_device() if torch.cuda.
is_available() else 'cpu')
x_squared = x * x
integral_img = x.cumsum(dim=2).cumsum(dim=3)
integral_img_sq = x_squared.cumsum(dim=2).cumsum(dim=3)
d = 1, self.window_size[0], self.window_size[1]
integral_img = torch.unsqueeze(integral_img, dim=1)
integral_img_sq = torch.unsqueeze(integral_img_sq, dim=1)
kernel = torch.tensor([[[[[1.0, -1.0], [-1.0, 1.0]]]]])
c_kernel = torch.ones((1, 1, self.channels_per_group, 1, 1))
with torch.no_grad():
sums = F.conv3d(integral_img, kernel, stride=[1, 1, 1],
dilation=d)
sums = F.conv3d(sums, c_kernel, stride=[self.
channels_per_group, 1, 1])
squares = F.conv3d(integral_img_sq, kernel, stride=[1, 1, 1
], dilation=d)
squares = F.conv3d(squares, c_kernel, stride=[self.
channels_per_group, 1, 1])
n = self.window_size[0] * self.window_size[1
] * self.channels_per_group
means = torch.squeeze(sums / n, dim=1)
var = torch.squeeze(1.0 / n * (squares - sums * sums / n), dim=1)
_, _, h, w = means.size()
pad2d = int(math.floor((W - w) / 2)), int(math.ceil((W - w) / 2)
), int(math.floor((H - h) / 2)), int(math.ceil((H - h) / 2))
padded_means = F.pad(means, pad2d, 'replicate')
padded_vars = F.pad(var, pad2d, 'replicate') + self.eps
for i in range(G):
x[:, i * self.channels_per_group:i * self.
channels_per_group + self.channels_per_group, :, :] = (x
[:, i * self.channels_per_group:i * self.
channels_per_group + self.channels_per_group, :, :] -
torch.unsqueeze(padded_means[:, i, :, :], dim=1)
) / torch.unsqueeze(padded_vars[:, i, :, :], dim=1).sqrt()
del integral_img
del integral_img_sq
else:
x = x.view(N, G, -1)
mean = x.mean(-1, keepdim=True)
var = x.var(-1, keepdim=True)
x = (x - mean) / (var + self.eps).sqrt()
x = x.view(N, C, H, W)
return x * self.weight + self.bias
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'num_features': 4}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.utils.data
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused_add_mean_sqrt_var_0(in_out_ptr0, in_out_ptr1, in_ptr0,
xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 8
RBLOCK: tl.constexpr = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 32 * x0), xmask, other=0.0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp3 = tl.where(xmask, tmp1, 0)
tmp4 = tl.sum(tmp3, 1)[:, None]
tmp6 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp8 = tl.where(xmask, tmp6, 0)
tmp9 = tl.sum(tmp8, 1)[:, None]
tmp10 = tl.full([XBLOCK, 1], 32, tl.int32)
tmp11 = tmp10.to(tl.float32)
tmp12 = tmp9 / tmp11
tmp13 = tmp1 - tmp12
tmp14 = tmp13 * tmp13
tmp15 = tl.broadcast_to(tmp14, [XBLOCK, RBLOCK])
tmp17 = tl.where(xmask, tmp15, 0)
tmp18 = tl.sum(tmp17, 1)[:, None]
tmp19 = 32.0
tmp20 = tmp4 / tmp19
tmp21 = 31.0
tmp22 = tmp18 / tmp21
tmp23 = 1e-05
tmp24 = tmp22 + tmp23
tmp25 = libdevice.sqrt(tmp24)
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp20, xmask)
tl.debug_barrier()
tl.store(in_out_ptr1 + x0, tmp25, xmask)
@triton.jit
def triton_poi_fused_add_mul_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x1 = xindex // 16 % 4
tmp0 = tl.load(in_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr1 + x4 // 32, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x4 // 32, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 / tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + x4, tmp8, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (1, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_3, (1, 4, 1, 1), (4, 1, 1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 2, 1), (2, 1, 8), torch.float32)
buf3 = empty_strided_cuda((4, 2, 1), (2, 1, 8), torch.float32)
buf1 = reinterpret_tensor(buf0, (4, 2, 1), (2, 1, 1), 0)
del buf0
buf5 = reinterpret_tensor(buf3, (4, 2, 1), (2, 1, 1), 0)
del buf3
get_raw_stream(0)
triton_per_fused_add_mean_sqrt_var_0[grid(8)](buf1, buf5, primals_1,
8, 32, XBLOCK=8, num_warps=2, num_stages=1)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_add_mul_1[grid(256)](primals_1, buf1, buf5,
primals_2, primals_3, buf6, 256, XBLOCK=256, num_warps=4,
num_stages=1)
del primals_2
del primals_3
return buf6, primals_1, buf1, buf5
class LocalContextNormNew(nn.Module):
def __init__(self, num_features, channels_per_group=2, window_size=(227,
227), eps=1e-05):
super(LocalContextNormNew, self).__init__()
self.weight = nn.Parameter(torch.ones(1, num_features, 1, 1))
self.bias = nn.Parameter(torch.zeros(1, num_features, 1, 1))
self.channels_per_group = channels_per_group
self.eps = eps
self.window_size = window_size
def forward(self, input_0):
primals_2 = self.weight
primals_3 = self.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
|
pjh4993/FCOS
|
LocalContextNorm
| false
| 4,159
|
[
"BSD-2-Clause"
] | 0
|
27f79e3fd3f5043796450b9a2201b42c744fd3df
|
https://github.com/pjh4993/FCOS/tree/27f79e3fd3f5043796450b9a2201b42c744fd3df
|
NeuralNet
|
import torch
class NeuralNet(torch.nn.Module):
def __init__(self, input_features, hidden_layer_size, output_classes):
super(NeuralNet, self).__init__()
self.l1 = torch.nn.Linear(input_features, hidden_layer_size)
self.l2 = torch.nn.Linear(hidden_layer_size, output_classes)
def forward(self, X):
hidden_layer = torch.sigmoid(self.l1(X))
return self.l2(hidden_layer)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'input_features': 4, 'hidden_layer_size': 1,
'output_classes': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_sigmoid_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr0 + 0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK])
tmp3 = tmp0 + tmp2
tmp4 = tl.sigmoid(tmp3)
tl.store(in_out_ptr0 + x0, tmp4, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (1, 4), (4, 1))
assert_size_stride(primals_2, (1,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 1), (1, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 1), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf0
get_raw_stream(0)
triton_poi_fused_sigmoid_0[grid(64)](buf1, primals_2, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 1), (
1, 1), 0), reinterpret_tensor(primals_4, (1, 4), (1, 1), 0),
alpha=1, beta=1, out=buf2)
del primals_5
return reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0
), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), buf1, primals_4
class NeuralNetNew(torch.nn.Module):
def __init__(self, input_features, hidden_layer_size, output_classes):
super(NeuralNetNew, self).__init__()
self.l1 = torch.nn.Linear(input_features, hidden_layer_size)
self.l2 = torch.nn.Linear(hidden_layer_size, output_classes)
def forward(self, input_0):
primals_1 = self.l1.weight
primals_2 = self.l1.bias
primals_4 = self.l2.weight
primals_5 = self.l2.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
rahimftd/digit_recognizer
|
NeuralNet
| false
| 4,160
|
[
"MIT"
] | 0
|
a134efa915670308ad7a77c8ace2662e5c775913
|
https://github.com/rahimftd/digit_recognizer/tree/a134efa915670308ad7a77c8ace2662e5c775913
|
FCNet
|
import torch
import torch.nn.functional
from torch import nn
from torch.nn.utils import weight_norm
class FCNet(nn.Module):
def __init__(self, in_size, out_size, activate=None, drop=0.0):
super(FCNet, self).__init__()
self.lin = weight_norm(nn.Linear(in_size, out_size), dim=None)
self.drop_value = drop
self.drop = nn.Dropout(drop)
self.activate = activate.lower() if activate is not None else None
if activate == 'relu':
self.ac_fn = nn.ReLU()
elif activate == 'sigmoid':
self.ac_fn = nn.Sigmoid()
elif activate == 'tanh':
self.ac_fn = nn.Tanh()
def forward(self, x):
if self.drop_value > 0:
x = self.drop(x)
x = self.lin(x)
if self.activate is not None:
x = self.ac_fn(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_size': 4, 'out_size': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn.functional
from torch import nn
from torch.nn.utils import weight_norm
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused_div_mul_norm_0(in_out_ptr0, in_ptr0, in_ptr1, out_ptr0,
xnumel, rnumel, XBLOCK: tl.constexpr):
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp6 = tl.load(in_ptr1 + 0)
tmp7 = tl.broadcast_to(tmp6, [XBLOCK, RBLOCK])
tmp1 = tmp0 * tmp0
tmp2 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp4 = tl.sum(tmp2, 1)[:, None]
tmp5 = libdevice.sqrt(tmp4)
tmp8 = tmp7 / tmp5
tmp9 = tmp0 * tmp8
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp5, None)
tl.store(out_ptr0 + tl.broadcast_to(r0, [XBLOCK, RBLOCK]), tmp9, None)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (), ())
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = buf0
del buf0
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
get_raw_stream(0)
triton_per_fused_div_mul_norm_0[grid(1)](buf1, primals_2, primals_1,
buf2, 1, 16, XBLOCK=1, num_warps=2, num_stages=1)
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_3, reinterpret_tensor(primals_4, (64,
4), (4, 1), 0), reinterpret_tensor(buf2, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf3)
del primals_3
return reinterpret_tensor(buf3, (4, 4, 4, 4), (64, 16, 4, 1), 0
), buf2, primals_1, primals_2, buf1, reinterpret_tensor(primals_4,
(64, 4), (4, 1), 0)
class FCNetNew(nn.Module):
def __init__(self, in_size, out_size, activate=None, drop=0.0):
super(FCNetNew, self).__init__()
self.lin = weight_norm(nn.Linear(in_size, out_size), dim=None)
self.drop_value = drop
self.drop = nn.Dropout(drop)
self.activate = activate.lower() if activate is not None else None
if activate == 'relu':
self.ac_fn = nn.ReLU()
elif activate == 'sigmoid':
self.ac_fn = nn.Sigmoid()
elif activate == 'tanh':
self.ac_fn = nn.Tanh()
def forward(self, input_0):
primals_3 = self.lin.bias
primals_1 = self.lin.weight_g
primals_2 = self.lin.weight_v
primals_4 = input_0
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0]
|
rafiberlin/clp-sose21-pm-vision
|
FCNet
| false
| 4,161
|
[
"MIT"
] | 0
|
55c786182ed4568cdeda4bb3676fa02b9580d68d
|
https://github.com/rafiberlin/clp-sose21-pm-vision/tree/55c786182ed4568cdeda4bb3676fa02b9580d68d
|
SharpenedCosineSimilarity
|
import torch
import torch.nn as nn
import torch.nn.functional as F
def unfold2d(x, kernel_size: 'int', stride: 'int', padding: 'int'):
x = F.pad(x, [padding] * 4)
bs, in_c, h, w = x.size()
ks = kernel_size
strided_x = x.as_strided((bs, in_c, (h - ks) // stride + 1, (w - ks) //
stride + 1, ks, ks), (in_c * h * w, h * w, stride * w, stride, w, 1))
return strided_x
class SharpenedCosineSimilarity(nn.Module):
def __init__(self, in_channels=1, out_channels=1, kernel_size=1, stride
=1, padding=0, eps=1e-12):
super(SharpenedCosineSimilarity, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.eps = eps
self.padding = int(padding)
w = torch.empty(out_channels, in_channels, kernel_size, kernel_size)
nn.init.xavier_uniform_(w)
self.w = nn.Parameter(w.view(out_channels, in_channels, -1),
requires_grad=True)
self.p_scale = 10
p_init = 2 ** 0.5 * self.p_scale
self.register_parameter('p', nn.Parameter(torch.empty(out_channels)))
nn.init.constant_(self.p, p_init)
self.q_scale = 100
self.register_parameter('q', nn.Parameter(torch.empty(1)))
nn.init.constant_(self.q, 10)
def forward(self, x):
x = unfold2d(x, kernel_size=self.kernel_size, stride=self.stride,
padding=self.padding)
n, c, h, w, _, _ = x.shape
x = x.reshape(n, c, h, w, -1)
square_sum = torch.sum(torch.square(x), [1, 4], keepdim=True)
x_norm = torch.add(torch.sqrt(square_sum + self.eps), torch.square(
self.q / self.q_scale))
square_sum = torch.sum(torch.square(self.w), [1, 2], keepdim=True)
w_norm = torch.add(torch.sqrt(square_sum + self.eps), torch.square(
self.q / self.q_scale))
x = torch.einsum('nchwl,vcl->nvhw', x / x_norm, self.w / w_norm)
sign = torch.sign(x)
x = torch.abs(x) + self.eps
x = x.pow(torch.square(self.p / self.p_scale).view(1, -1, 1, 1))
return sign * x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_abs_add_div_mul_pow_sign_sqrt_sum_0(in_out_ptr0,
in_ptr0, in_ptr1, in_ptr2, in_ptr3, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = xindex // 16
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask)
tmp2 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask)
tmp5 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask)
tmp8 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask)
tmp14 = tl.load(in_ptr1 + 0)
tmp15 = tl.broadcast_to(tmp14, [XBLOCK])
tmp27 = tl.load(in_ptr2 + 0)
tmp28 = tl.broadcast_to(tmp27, [XBLOCK])
tmp44 = tl.load(in_ptr3 + 0)
tmp45 = tl.broadcast_to(tmp44, [XBLOCK])
tmp1 = tmp0 * tmp0
tmp3 = tmp2 * tmp2
tmp4 = tmp1 + tmp3
tmp6 = tmp5 * tmp5
tmp7 = tmp4 + tmp6
tmp9 = tmp8 * tmp8
tmp10 = tmp7 + tmp9
tmp11 = 1e-12
tmp12 = tmp10 + tmp11
tmp13 = libdevice.sqrt(tmp12)
tmp16 = 0.01
tmp17 = tmp15 * tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp13 + tmp18
tmp20 = tmp0 / tmp19
tmp21 = tmp2 / tmp19
tmp22 = tmp20 + tmp21
tmp23 = tmp5 / tmp19
tmp24 = tmp22 + tmp23
tmp25 = tmp8 / tmp19
tmp26 = tmp24 + tmp25
tmp29 = tmp28 * tmp28
tmp30 = tmp29 + tmp11
tmp31 = libdevice.sqrt(tmp30)
tmp32 = tmp31 + tmp18
tmp33 = tmp28 / tmp32
tmp34 = tmp26 * tmp33
tmp35 = tl.full([1], 0, tl.int32)
tmp36 = tmp35 < tmp34
tmp37 = tmp36.to(tl.int8)
tmp38 = tmp34 < tmp35
tmp39 = tmp38.to(tl.int8)
tmp40 = tmp37 - tmp39
tmp41 = tmp40.to(tmp34.dtype)
tmp42 = tl_math.abs(tmp34)
tmp43 = tmp42 + tmp11
tmp46 = 0.1
tmp47 = tmp45 * tmp46
tmp48 = tmp47 * tmp47
tmp49 = libdevice.pow(tmp43, tmp48)
tmp50 = tmp41 * tmp49
tl.store(in_out_ptr0 + x2, tmp50, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (1,), (1,))
assert_size_stride(primals_3, (1, 1, 1), (1, 1, 1))
assert_size_stride(primals_4, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 1, 4, 4, 1), (16, 64, 4, 1, 64),
torch.float32)
buf1 = reinterpret_tensor(buf0, (4, 1, 4, 4, 1, 1), (16, 64, 4, 1,
64, 64), 0)
del buf0
buf2 = reinterpret_tensor(buf1, (4, 1, 4, 4), (16, 16, 4, 1), 0)
del buf1
get_raw_stream(0)
triton_poi_fused_abs_add_div_mul_pow_sign_sqrt_sum_0[grid(64)](buf2,
primals_1, primals_2, primals_3, primals_4, 64, XBLOCK=64,
num_warps=1, num_stages=1)
return buf2, primals_1, primals_2, primals_3, primals_4
def unfold2d(x, kernel_size: 'int', stride: 'int', padding: 'int'):
x = F.pad(x, [padding] * 4)
bs, in_c, h, w = x.size()
ks = kernel_size
strided_x = x.as_strided((bs, in_c, (h - ks) // stride + 1, (w - ks) //
stride + 1, ks, ks), (in_c * h * w, h * w, stride * w, stride, w, 1))
return strided_x
class SharpenedCosineSimilarityNew(nn.Module):
def __init__(self, in_channels=1, out_channels=1, kernel_size=1, stride
=1, padding=0, eps=1e-12):
super(SharpenedCosineSimilarityNew, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.eps = eps
self.padding = int(padding)
w = torch.empty(out_channels, in_channels, kernel_size, kernel_size)
nn.init.xavier_uniform_(w)
self.w = nn.Parameter(w.view(out_channels, in_channels, -1),
requires_grad=True)
self.p_scale = 10
p_init = 2 ** 0.5 * self.p_scale
self.register_parameter('p', nn.Parameter(torch.empty(out_channels)))
nn.init.constant_(self.p, p_init)
self.q_scale = 100
self.register_parameter('q', nn.Parameter(torch.empty(1)))
nn.init.constant_(self.q, 10)
def forward(self, input_0):
primals_3 = self.w
primals_2 = self.p
primals_4 = self.q
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0]
|
quickgrid/sharpened_cosine_similarity_torch
|
SharpenedCosineSimilarity
| false
| 4,162
|
[
"MIT"
] | 0
|
d652d76a4994a0b3817e248d5899827d35a5ebeb
|
https://github.com/quickgrid/sharpened_cosine_similarity_torch/tree/d652d76a4994a0b3817e248d5899827d35a5ebeb
|
EncoderLayer
|
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
class AffineLayer(nn.Module):
def __init__(self, dropout, d_model, d_ff):
super(AffineLayer, self).__init__()
self.w_1 = nn.Linear(d_model, d_ff)
self.w_2 = nn.Linear(d_ff, d_model)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
return self.w_2(self.dropout(F.relu(self.w_1(x))))
class MultiHeadedAttention(nn.Module):
def __init__(self, num_head, d_model, dropout=0.1):
super(MultiHeadedAttention, self).__init__()
assert d_model % num_head == 0
self.d_k = d_model // num_head
self.h = num_head
self.linear_key = nn.Linear(d_model, d_model)
self.linear_value = nn.Linear(d_model, d_model)
self.linear_query = nn.Linear(d_model, d_model)
self.linear_out = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(p=dropout)
def attention(self, query, key, value, mask, dropout=None):
d_k = query.size(-1)
scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
scores = scores.masked_fill(mask == 0, -1000000000.0)
p_attn = F.softmax(scores, dim=-1)
if dropout is not None:
p_attn = dropout(p_attn)
return torch.matmul(p_attn, value), p_attn
def forward(self, query, key, value, mask):
nbatches = query.size(0)
query = self.linear_query(query).view(nbatches, -1, self.h, self.d_k
).transpose(1, 2)
key = self.linear_key(key).view(nbatches, -1, self.h, self.d_k
).transpose(1, 2)
value = self.linear_value(value).view(nbatches, -1, self.h, self.d_k
).transpose(1, 2)
mask = mask.unsqueeze(1)
x, _attn = self.attention(query, key, value, mask, dropout=self.dropout
)
x = x.transpose(1, 2).contiguous().view(nbatches, -1, self.h * self.d_k
)
return self.linear_out(x)
class EncoderLayer(nn.Module):
def __init__(self, num_head, dropout, d_model, d_ff):
super(EncoderLayer, self).__init__()
self.att_layer = MultiHeadedAttention(num_head, d_model, dropout)
self.norm_att = nn.LayerNorm(d_model)
self.dropout_att = nn.Dropout(dropout)
self.affine_layer = AffineLayer(dropout, d_model, d_ff)
self.norm_affine = nn.LayerNorm(d_model)
self.dropout_affine = nn.Dropout(dropout)
def forward(self, x, mask):
x_att = self.norm_att(x * mask)
x_att = self.att_layer(x_att, x_att, x_att, mask)
x = x + self.dropout_att(x_att)
x_affine = self.norm_affine(x * mask)
x_affine = self.affine_layer(x_affine)
return x + self.dropout_affine(x_affine)
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'num_head': 4, 'dropout': 0.5, 'd_model': 4, 'd_ff': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import math
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_native_layer_norm_0(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 * tmp1
tmp5 = tmp3 * tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 * tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 * tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(out_ptr1 + x0, tmp28, xmask)
@triton.jit
def triton_poi_fused_mul_native_layer_norm_1(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 * tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-05
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
@triton.jit
def triton_poi_fused_clone_2(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused__softmax_div_eq_masked_fill_3(in_ptr0, in_ptr1,
out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex // 16
x3 = xindex
tmp0 = tl.load(in_ptr0 + (4 * x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tl.load(in_ptr1 + 4 * x3, xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (1 + 4 * x0 + 16 * x2), xmask, eviction_policy
='evict_last')
tmp10 = tl.load(in_ptr1 + (1 + 4 * x3), xmask, eviction_policy='evict_last'
)
tmp14 = tl.load(in_ptr0 + (2 + 4 * x0 + 16 * x2), xmask,
eviction_policy='evict_last')
tmp16 = tl.load(in_ptr1 + (2 + 4 * x3), xmask, eviction_policy='evict_last'
)
tmp20 = tl.load(in_ptr0 + (3 + 4 * x0 + 16 * x2), xmask,
eviction_policy='evict_last')
tmp22 = tl.load(in_ptr1 + (3 + 4 * x3), xmask, eviction_policy='evict_last'
)
tmp1 = 0.0
tmp2 = tmp0 == tmp1
tmp4 = 1.0
tmp5 = tmp3 * tmp4
tmp6 = -1000000000.0
tmp7 = tl.where(tmp2, tmp6, tmp5)
tmp9 = tmp8 == tmp1
tmp11 = tmp10 * tmp4
tmp12 = tl.where(tmp9, tmp6, tmp11)
tmp13 = triton_helpers.maximum(tmp7, tmp12)
tmp15 = tmp14 == tmp1
tmp17 = tmp16 * tmp4
tmp18 = tl.where(tmp15, tmp6, tmp17)
tmp19 = triton_helpers.maximum(tmp13, tmp18)
tmp21 = tmp20 == tmp1
tmp23 = tmp22 * tmp4
tmp24 = tl.where(tmp21, tmp6, tmp23)
tmp25 = triton_helpers.maximum(tmp19, tmp24)
tmp26 = tmp7 - tmp25
tmp27 = tl_math.exp(tmp26)
tmp28 = tmp12 - tmp25
tmp29 = tl_math.exp(tmp28)
tmp30 = tmp27 + tmp29
tmp31 = tmp18 - tmp25
tmp32 = tl_math.exp(tmp31)
tmp33 = tmp30 + tmp32
tmp34 = tmp24 - tmp25
tmp35 = tl_math.exp(tmp34)
tmp36 = tmp33 + tmp35
tl.store(out_ptr0 + x3, tmp25, xmask)
tl.store(out_ptr1 + x3, tmp36, xmask)
@triton.jit
def triton_poi_fused__softmax_div_eq_masked_fill_4(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex // 64
x4 = xindex % 16
x5 = xindex
x6 = xindex // 4
tmp0 = tl.load(in_ptr0 + (x4 + 16 * x3), xmask, eviction_policy=
'evict_last')
tmp3 = tl.load(in_out_ptr0 + x5, xmask)
tmp8 = tl.load(in_ptr1 + x6, xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr2 + x6, xmask, eviction_policy='evict_last')
tmp1 = 0.0
tmp2 = tmp0 == tmp1
tmp4 = 1.0
tmp5 = tmp3 * tmp4
tmp6 = -1000000000.0
tmp7 = tl.where(tmp2, tmp6, tmp5)
tmp9 = tmp7 - tmp8
tmp10 = tl_math.exp(tmp9)
tmp12 = tmp10 / tmp11
tl.store(in_out_ptr0 + x5, tmp12, xmask)
@triton.jit
def triton_poi_fused_clone_5(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_mul_native_layer_norm_6(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + 4 * x0, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr2 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp14 = tl.load(in_ptr2 + (2 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp17 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp18 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp20 = tl.load(in_ptr2 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 * tmp3
tmp7 = tmp5 + tmp6
tmp9 = tmp7 * tmp8
tmp10 = tmp4 + tmp9
tmp13 = tmp11 + tmp12
tmp15 = tmp13 * tmp14
tmp16 = tmp10 + tmp15
tmp19 = tmp17 + tmp18
tmp21 = tmp19 * tmp20
tmp22 = tmp16 + tmp21
tmp23 = 4.0
tmp24 = tmp22 / tmp23
tmp25 = tmp4 - tmp24
tmp26 = tmp25 * tmp25
tmp27 = tmp9 - tmp24
tmp28 = tmp27 * tmp27
tmp29 = tmp26 + tmp28
tmp30 = tmp15 - tmp24
tmp31 = tmp30 * tmp30
tmp32 = tmp29 + tmp31
tmp33 = tmp21 - tmp24
tmp34 = tmp33 * tmp33
tmp35 = tmp32 + tmp34
tmp36 = tmp35 / tmp23
tl.store(out_ptr0 + x0, tmp24, xmask)
tl.store(out_ptr1 + x0, tmp36, xmask)
@triton.jit
def triton_poi_fused_add_mul_native_layer_norm_7(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, in_ptr6, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_ptr2 + x2, xmask)
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x1, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr6 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 - tmp5
tmp8 = 1e-05
tmp9 = tmp7 + tmp8
tmp10 = libdevice.rsqrt(tmp9)
tmp11 = tmp6 * tmp10
tmp13 = tmp11 * tmp12
tmp15 = tmp13 + tmp14
tl.store(out_ptr0 + x2, tmp15, xmask)
@triton.jit
def triton_poi_fused_relu_threshold_backward_8(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, xmask)
tl.store(out_ptr0 + x2, tmp6, xmask)
@triton.jit
def triton_poi_fused_add_9(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel,
XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_out_ptr0 + x2, xmask)
tmp4 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tl.store(in_out_ptr0 + x2, tmp6, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13, primals_14, primals_15, primals_16, primals_17, primals_18
) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4,), (1,))
assert_size_stride(primals_5, (4, 4), (4, 1))
assert_size_stride(primals_6, (4,), (1,))
assert_size_stride(primals_7, (4, 4), (4, 1))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (4, 4), (4, 1))
assert_size_stride(primals_10, (4,), (1,))
assert_size_stride(primals_11, (4, 4), (4, 1))
assert_size_stride(primals_12, (4,), (1,))
assert_size_stride(primals_13, (4,), (1,))
assert_size_stride(primals_14, (4,), (1,))
assert_size_stride(primals_15, (4, 4), (4, 1))
assert_size_stride(primals_16, (4,), (1,))
assert_size_stride(primals_17, (4, 4), (4, 1))
assert_size_stride(primals_18, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf1 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_native_layer_norm_0[grid(16)](primals_1,
primals_2, buf0, buf1, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_mul_native_layer_norm_1[grid(64)](primals_1,
primals_2, buf0, buf1, primals_3, primals_4, buf2, 64, XBLOCK=
64, num_warps=1, num_stages=1)
del primals_3
del primals_4
buf3 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_5, (4, 4), (1, 4), 0), out=buf3)
buf4 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 4), (1, 4), 0), out=buf4)
buf5 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_9, (4, 4), (1, 4), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
triton_poi_fused_clone_2[grid(16, 4)](buf3, primals_6, buf6, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_6
buf7 = reinterpret_tensor(buf3, (4, 4, 1, 4), (16, 4, 4, 1), 0)
del buf3
triton_poi_fused_clone_2[grid(16, 4)](buf4, primals_8, buf7, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_8
buf8 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf6, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf7, (16, 1, 4), (4, 0, 1), 0), out=buf8)
buf9 = reinterpret_tensor(buf4, (4, 4, 4, 1), (16, 4, 1, 64), 0)
del buf4
buf10 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
triton_poi_fused__softmax_div_eq_masked_fill_3[grid(64)](primals_2,
buf8, buf9, buf10, 64, XBLOCK=64, num_warps=1, num_stages=1)
buf11 = reinterpret_tensor(buf8, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf8
triton_poi_fused__softmax_div_eq_masked_fill_4[grid(256)](buf11,
primals_2, buf9, buf10, 256, XBLOCK=128, num_warps=4, num_stages=1)
buf12 = reinterpret_tensor(buf9, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf9
triton_poi_fused_clone_2[grid(16, 4)](buf5, primals_10, buf12, 16,
4, XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_10
buf13 = reinterpret_tensor(buf5, (16, 4, 1), (4, 1, 1), 0)
del buf5
extern_kernels.bmm(reinterpret_tensor(buf11, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf12, (16, 4, 1), (4, 1, 0), 0), out=buf13)
buf14 = reinterpret_tensor(buf10, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf10
triton_poi_fused_clone_5[grid(16, 4)](buf13, buf14, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf15 = reinterpret_tensor(buf13, (16, 4), (4, 1), 0)
del buf13
extern_kernels.addmm(primals_12, reinterpret_tensor(buf14, (16, 4),
(4, 1), 0), reinterpret_tensor(primals_11, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf15)
del primals_12
buf16 = buf1
del buf1
buf17 = buf0
del buf0
triton_poi_fused_add_mul_native_layer_norm_6[grid(16)](primals_1,
buf15, primals_2, buf16, buf17, 16, XBLOCK=16, num_warps=1,
num_stages=1)
buf18 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_mul_native_layer_norm_7[grid(64)](primals_1,
buf15, primals_2, buf16, buf17, primals_13, primals_14, buf18,
64, XBLOCK=64, num_warps=1, num_stages=1)
del buf16
del buf17
del primals_14
buf19 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf18, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_15, (4, 4), (1, 4), 0), out=buf19)
buf20 = reinterpret_tensor(buf19, (4, 4, 4), (16, 4, 1), 0)
del buf19
buf23 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.bool)
triton_poi_fused_relu_threshold_backward_8[grid(64)](buf20,
primals_16, buf23, 64, XBLOCK=64, num_warps=1, num_stages=1)
del primals_16
buf21 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf20, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_17, (4, 4), (1, 4), 0), out=buf21)
buf22 = reinterpret_tensor(buf21, (4, 4, 4), (16, 4, 1), 0)
del buf21
triton_poi_fused_add_9[grid(64)](buf22, primals_1, buf15,
primals_18, 64, XBLOCK=64, num_warps=1, num_stages=1)
del primals_18
return buf22, primals_1, primals_2, primals_13, reinterpret_tensor(buf2,
(16, 4), (4, 1), 0), buf11, reinterpret_tensor(buf14, (16, 4), (4,
1), 0), buf15, reinterpret_tensor(buf18, (16, 4), (4, 1), 0
), reinterpret_tensor(buf20, (16, 4), (4, 1), 0
), primals_17, buf23, primals_15, primals_11, reinterpret_tensor(buf12,
(16, 1, 4), (4, 1, 1), 0), reinterpret_tensor(buf6, (16, 1, 4), (4,
1, 1), 0), reinterpret_tensor(buf7, (16, 4, 1), (4, 1, 4), 0
), primals_9, primals_7, primals_5
class AffineLayer(nn.Module):
def __init__(self, dropout, d_model, d_ff):
super(AffineLayer, self).__init__()
self.w_1 = nn.Linear(d_model, d_ff)
self.w_2 = nn.Linear(d_ff, d_model)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
return self.w_2(self.dropout(F.relu(self.w_1(x))))
class MultiHeadedAttention(nn.Module):
def __init__(self, num_head, d_model, dropout=0.1):
super(MultiHeadedAttention, self).__init__()
assert d_model % num_head == 0
self.d_k = d_model // num_head
self.h = num_head
self.linear_key = nn.Linear(d_model, d_model)
self.linear_value = nn.Linear(d_model, d_model)
self.linear_query = nn.Linear(d_model, d_model)
self.linear_out = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(p=dropout)
def attention(self, query, key, value, mask, dropout=None):
d_k = query.size(-1)
scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
scores = scores.masked_fill(mask == 0, -1000000000.0)
p_attn = F.softmax(scores, dim=-1)
if dropout is not None:
p_attn = dropout(p_attn)
return torch.matmul(p_attn, value), p_attn
def forward(self, query, key, value, mask):
nbatches = query.size(0)
query = self.linear_query(query).view(nbatches, -1, self.h, self.d_k
).transpose(1, 2)
key = self.linear_key(key).view(nbatches, -1, self.h, self.d_k
).transpose(1, 2)
value = self.linear_value(value).view(nbatches, -1, self.h, self.d_k
).transpose(1, 2)
mask = mask.unsqueeze(1)
x, _attn = self.attention(query, key, value, mask, dropout=self.dropout
)
x = x.transpose(1, 2).contiguous().view(nbatches, -1, self.h * self.d_k
)
return self.linear_out(x)
class EncoderLayerNew(nn.Module):
def __init__(self, num_head, dropout, d_model, d_ff):
super(EncoderLayerNew, self).__init__()
self.att_layer = MultiHeadedAttention(num_head, d_model, dropout)
self.norm_att = nn.LayerNorm(d_model)
self.dropout_att = nn.Dropout(dropout)
self.affine_layer = AffineLayer(dropout, d_model, d_ff)
self.norm_affine = nn.LayerNorm(d_model)
self.dropout_affine = nn.Dropout(dropout)
def forward(self, input_0, input_1):
primals_5 = self.att_layer.linear_key.weight
primals_3 = self.att_layer.linear_key.bias
primals_7 = self.att_layer.linear_value.weight
primals_4 = self.att_layer.linear_value.bias
primals_9 = self.att_layer.linear_query.weight
primals_6 = self.att_layer.linear_query.bias
primals_11 = self.att_layer.linear_out.weight
primals_8 = self.att_layer.linear_out.bias
primals_10 = self.norm_att.weight
primals_12 = self.norm_att.bias
primals_15 = self.affine_layer.w_1.weight
primals_13 = self.affine_layer.w_1.bias
primals_17 = self.affine_layer.w_2.weight
primals_14 = self.affine_layer.w_2.bias
primals_16 = self.norm_affine.weight
primals_18 = self.norm_affine.bias
primals_1 = input_0
primals_2 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13, primals_14,
primals_15, primals_16, primals_17, primals_18])
return output[0]
|
qi700/my_point_summarize
|
EncoderLayer
| false
| 4,163
|
[
"Apache-2.0"
] | 0
|
e269c2d0411fc61ea34055c3080472bc9111bcaa
|
https://github.com/qi700/my_point_summarize/tree/e269c2d0411fc61ea34055c3080472bc9111bcaa
|
DSCNet
|
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
class Conv2dSamePad(nn.Module):
"""
Implement Tensorflow's 'SAME' padding mode in Conv2d.
When an odd number, say `m`, of pixels are need to pad, Tensorflow will pad one more column at right or one more
row at bottom. But Pytorch will pad `m+1` pixels, i.e., Pytorch always pads in both sides.
So we can pad the tensor in the way of Tensorflow before call the Conv2d module.
"""
def __init__(self, kernel_size, stride):
super(Conv2dSamePad, self).__init__()
self.kernel_size = kernel_size if type(kernel_size) in [list, tuple
] else [kernel_size, kernel_size]
self.stride = stride if type(stride) in [list, tuple] else [stride,
stride]
def forward(self, x):
in_height = x.size(2)
in_width = x.size(3)
out_height = math.ceil(float(in_height) / float(self.stride[0]))
out_width = math.ceil(float(in_width) / float(self.stride[1]))
pad_along_height = (out_height - 1) * self.stride[0
] + self.kernel_size[0] - in_height
pad_along_width = (out_width - 1) * self.stride[1] + self.kernel_size[1
] - in_width
pad_top = math.floor(pad_along_height / 2)
pad_left = math.floor(pad_along_width / 2)
pad_bottom = pad_along_height - pad_top
pad_right = pad_along_width - pad_left
return F.pad(x, [pad_left, pad_right, pad_top, pad_bottom],
'constant', 0)
class ConvTranspose2dSamePad(nn.Module):
"""
This module implements the "SAME" padding mode for ConvTranspose2d as in Tensorflow.
A tensor with width w_in, feed it to ConvTranspose2d(ci, co, kernel, stride), the width of output tensor T_nopad:
w_nopad = (w_in - 1) * stride + kernel
If we use padding, i.e., ConvTranspose2d(ci, co, kernel, stride, padding, output_padding), the width of T_pad:
w_pad = (w_in - 1) * stride + kernel - (2*padding - output_padding) = w_nopad - (2*padding - output_padding)
Yes, in ConvTranspose2d, more padding, the resulting tensor is smaller, i.e., the padding is actually deleting row/col.
If `pad`=(2*padding - output_padding) is odd, Pytorch deletes more columns in the left, i.e., the first ceil(pad/2) and
last `pad - ceil(pad/2)` columns of T_nopad are deleted to get T_pad.
In contrast, Tensorflow deletes more columns in the right, i.e., the first floor(pad/2) and last `pad - floor(pad/2)`
columns are deleted.
For the height, Pytorch deletes more rows at top, while Tensorflow at bottom.
In practice, we usually want `w_pad = w_in * stride` or `w_pad = w_in * stride - 1`, i.e., the "SAME" padding mode
in Tensorflow. To determine the value of `w_pad`, we should pass it to this function.
So the number of columns to delete:
pad = 2*padding - output_padding = w_nopad - w_pad
If pad is even, we can directly set padding=pad/2 and output_padding=0 in ConvTranspose2d.
If pad is odd, we can use ConvTranspose2d to get T_nopad, and then delete `pad` rows/columns by
ourselves.
This module should be called after the ConvTranspose2d module with shared kernel_size and stride values.
"""
def __init__(self, output_size):
super(ConvTranspose2dSamePad, self).__init__()
self.output_size = output_size
def forward(self, x):
in_height = x.size(2)
in_width = x.size(3)
pad_height = in_height - self.output_size[0]
pad_width = in_width - self.output_size[1]
pad_top = pad_height // 2
pad_bottom = pad_height - pad_top
pad_left = pad_width // 2
pad_right = pad_width - pad_left
return x[:, :, pad_top:in_height - pad_bottom, pad_left:in_width -
pad_right]
class ConvAE(nn.Module):
def __init__(self, channels, kernels):
"""
:param channels: a list containing all channels including the input image channel (1 for gray, 3 for RGB)
:param kernels: a list containing all kernel sizes, it should satisfy: len(kernels) = len(channels) - 1.
"""
super(ConvAE, self).__init__()
assert isinstance(channels, list) and isinstance(kernels, list)
self.encoder = nn.Sequential()
for i in range(1, len(channels)):
self.encoder.add_module('pad%d' % i, Conv2dSamePad(kernels[i -
1], 2))
self.encoder.add_module('conv%d' % i, nn.Conv2d(channels[i - 1],
channels[i], kernel_size=kernels[i - 1], stride=2))
self.encoder.add_module('relu%d' % i, nn.ReLU(True))
self.decoder = nn.Sequential()
channels = list(reversed(channels))
kernels = list(reversed(kernels))
sizes = [[12, 11], [24, 21], [48, 42]]
for i in range(len(channels) - 1):
self.decoder.add_module('deconv%d' % (i + 1), nn.
ConvTranspose2d(channels[i], channels[i + 1], kernel_size=
kernels[i], stride=2))
self.decoder.add_module('padd%d' % i, ConvTranspose2dSamePad(
sizes[i]))
self.decoder.add_module('relud%d' % i, nn.ReLU(True))
def forward(self, x):
h = self.encoder(x)
y = self.decoder(h)
return y
class SelfExpression(nn.Module):
def __init__(self, n):
super(SelfExpression, self).__init__()
self.Coefficient = nn.Parameter(0.0001 * torch.ones(n, n, dtype=
torch.float32), requires_grad=True)
def forward(self, x):
y = torch.matmul(self.Coefficient, x)
return y
class DSCNet(nn.Module):
def __init__(self, channels, kernels, num_sample):
super(DSCNet, self).__init__()
self.n = num_sample
self.ae = ConvAE(channels, kernels)
self.self_expression = SelfExpression(self.n)
def forward(self, x):
z = self.ae.encoder(x)
shape = z.shape
z = z.view(self.n, -1)
z_recon = self.self_expression(z)
z_recon_reshape = z_recon.view(shape)
x_recon = self.ae.decoder(z_recon_reshape)
return x_recon, z, z_recon
def loss_fn(self, x, x_recon, z, z_recon, weight_coef, weight_selfExp):
loss_ae = 0.5 * F.mse_loss(x_recon, x, reduction='sum')
loss_coef = torch.sum(torch.pow(self.self_expression.Coefficient, 2))
loss_selfExp = 0.5 * F.mse_loss(z_recon, z, reduction='sum')
loss = (loss_ae + weight_coef * loss_coef + weight_selfExp *
loss_selfExp)
loss /= x.size(0)
return loss
def smoothLoss(self, z):
Z = torch.pow(z.unsqueeze(1) - z.unsqueeze(0), 2).sum(-1)
C = torch.abs(self.self_expression.Coefficient)
C = 0.5 * (C + torch.transpose(C, 0, 1))
C = C.fill_diagonal_(0)
loss_smooth = (Z * C).sum() / z.shape[0]
return loss_smooth
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'channels': [4, 4], 'kernels': [4, 4], 'num_sample': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import math
import torch.nn as nn
import torch.nn.functional as F
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_constant_pad_nd_0(in_ptr0, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 576
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 6 % 6
x0 = xindex % 6
x2 = xindex // 36
x4 = xindex
tmp0 = -1 + x1
tmp1 = tl.full([1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = -1 + x0
tmp6 = tmp5 >= tmp1
tmp7 = tmp5 < tmp3
tmp8 = tmp2 & tmp4
tmp9 = tmp8 & tmp6
tmp10 = tmp9 & tmp7
tmp11 = tl.load(in_ptr0 + (-5 + x0 + 4 * x1 + 16 * x2), tmp10 & xmask,
other=0.0)
tl.store(out_ptr0 + x4, tmp11, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_threshold_backward_1(in_out_ptr0,
in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 4 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x3, tmp4, xmask)
tl.store(out_ptr0 + x3, tmp6, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_2(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 576
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 6 % 6
x0 = xindex % 6
x4 = xindex
x2 = xindex // 36 % 4
tmp19 = tl.load(in_out_ptr0 + x4, xmask)
tmp20 = tl.load(in_ptr0 + x2, xmask, eviction_policy='evict_last')
tmp0 = x1
tmp1 = tl.full([1], 3, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = x0
tmp4 = tmp3 >= tmp1
tmp5 = tmp4 & tmp2
tmp6 = tl.load(in_out_ptr0 + x4, tmp5 & xmask, other=0.0)
tmp7 = tl.load(in_ptr0 + x2, tmp5 & xmask, eviction_policy='evict_last',
other=0.0)
tmp8 = tmp6 + tmp7
tmp9 = tl.full([1], 0, tl.int32)
tmp10 = triton_helpers.maximum(tmp9, tmp8)
tmp11 = tl.full(tmp10.shape, 0.0, tmp10.dtype)
tmp12 = tl.where(tmp5, tmp10, tmp11)
tmp13 = tl.load(in_out_ptr0 + x4, tmp2 & xmask, other=0.0)
tmp14 = tl.load(in_ptr0 + x2, tmp2 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp15 = tmp13 + tmp14
tmp16 = tl.where(tmp4, tmp12, tmp15)
tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)
tmp18 = tl.where(tmp2, tmp16, tmp17)
tmp21 = tmp19 + tmp20
tmp22 = tl.where(tmp2, tmp18, tmp21)
tl.store(in_out_ptr0 + x4, tmp22, xmask)
@triton.jit
def triton_poi_fused_threshold_backward_3(in_ptr0, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 144
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 3
x1 = xindex // 3 % 3
x2 = xindex // 9
x3 = xindex
tmp0 = tl.load(in_ptr0 + (21 + x0 + 6 * x1 + 36 * x2), xmask)
tmp1 = 0.0
tmp2 = tmp0 <= tmp1
tl.store(out_ptr0 + x3, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_6, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 6, 6), (144, 36, 6, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_constant_pad_nd_0[grid(576)](primals_1, buf0, 576,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_1
buf1 = extern_kernels.convolution(buf0, primals_2, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 2, 2), (16, 4, 2, 1))
buf2 = buf1
del buf1
buf7 = empty_strided_cuda((4, 4, 2, 2), (16, 4, 2, 1), torch.bool)
triton_poi_fused_convolution_relu_threshold_backward_1[grid(64)](buf2,
primals_3, buf7, 64, XBLOCK=64, num_warps=1, num_stages=1)
del primals_3
buf3 = empty_strided_cuda((4, 16), (16, 1), torch.float32)
extern_kernels.mm(primals_4, reinterpret_tensor(buf2, (4, 16), (16,
1), 0), out=buf3)
buf4 = extern_kernels.convolution(reinterpret_tensor(buf3, (4, 4, 2,
2), (16, 4, 2, 1), 0), primals_5, stride=(2, 2), padding=(0, 0),
dilation=(1, 1), transposed=True, output_padding=(0, 0), groups
=1, bias=None)
assert_size_stride(buf4, (4, 4, 6, 6), (144, 36, 6, 1))
buf5 = buf4
del buf4
triton_poi_fused_convolution_relu_2[grid(576)](buf5, primals_6, 576,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_6
buf6 = empty_strided_cuda((4, 4, 3, 3), (36, 9, 3, 1), torch.bool)
triton_poi_fused_threshold_backward_3[grid(144)](buf5, buf6, 144,
XBLOCK=128, num_warps=4, num_stages=1)
return reinterpret_tensor(buf5, (4, 4, 3, 3), (144, 36, 6, 1), 21
), reinterpret_tensor(buf2, (4, 16), (16, 1), 0
), buf3, primals_2, primals_5, buf0, reinterpret_tensor(buf3, (4, 4,
2, 2), (16, 4, 2, 1), 0), buf6, reinterpret_tensor(primals_4, (4, 4
), (1, 4), 0), reinterpret_tensor(buf2, (16, 4), (1, 16), 0), buf7
class Conv2dSamePad(nn.Module):
"""
Implement Tensorflow's 'SAME' padding mode in Conv2d.
When an odd number, say `m`, of pixels are need to pad, Tensorflow will pad one more column at right or one more
row at bottom. But Pytorch will pad `m+1` pixels, i.e., Pytorch always pads in both sides.
So we can pad the tensor in the way of Tensorflow before call the Conv2d module.
"""
def __init__(self, kernel_size, stride):
super(Conv2dSamePad, self).__init__()
self.kernel_size = kernel_size if type(kernel_size) in [list, tuple
] else [kernel_size, kernel_size]
self.stride = stride if type(stride) in [list, tuple] else [stride,
stride]
def forward(self, x):
in_height = x.size(2)
in_width = x.size(3)
out_height = math.ceil(float(in_height) / float(self.stride[0]))
out_width = math.ceil(float(in_width) / float(self.stride[1]))
pad_along_height = (out_height - 1) * self.stride[0
] + self.kernel_size[0] - in_height
pad_along_width = (out_width - 1) * self.stride[1] + self.kernel_size[1
] - in_width
pad_top = math.floor(pad_along_height / 2)
pad_left = math.floor(pad_along_width / 2)
pad_bottom = pad_along_height - pad_top
pad_right = pad_along_width - pad_left
return F.pad(x, [pad_left, pad_right, pad_top, pad_bottom],
'constant', 0)
class ConvTranspose2dSamePad(nn.Module):
"""
This module implements the "SAME" padding mode for ConvTranspose2d as in Tensorflow.
A tensor with width w_in, feed it to ConvTranspose2d(ci, co, kernel, stride), the width of output tensor T_nopad:
w_nopad = (w_in - 1) * stride + kernel
If we use padding, i.e., ConvTranspose2d(ci, co, kernel, stride, padding, output_padding), the width of T_pad:
w_pad = (w_in - 1) * stride + kernel - (2*padding - output_padding) = w_nopad - (2*padding - output_padding)
Yes, in ConvTranspose2d, more padding, the resulting tensor is smaller, i.e., the padding is actually deleting row/col.
If `pad`=(2*padding - output_padding) is odd, Pytorch deletes more columns in the left, i.e., the first ceil(pad/2) and
last `pad - ceil(pad/2)` columns of T_nopad are deleted to get T_pad.
In contrast, Tensorflow deletes more columns in the right, i.e., the first floor(pad/2) and last `pad - floor(pad/2)`
columns are deleted.
For the height, Pytorch deletes more rows at top, while Tensorflow at bottom.
In practice, we usually want `w_pad = w_in * stride` or `w_pad = w_in * stride - 1`, i.e., the "SAME" padding mode
in Tensorflow. To determine the value of `w_pad`, we should pass it to this function.
So the number of columns to delete:
pad = 2*padding - output_padding = w_nopad - w_pad
If pad is even, we can directly set padding=pad/2 and output_padding=0 in ConvTranspose2d.
If pad is odd, we can use ConvTranspose2d to get T_nopad, and then delete `pad` rows/columns by
ourselves.
This module should be called after the ConvTranspose2d module with shared kernel_size and stride values.
"""
def __init__(self, output_size):
super(ConvTranspose2dSamePad, self).__init__()
self.output_size = output_size
def forward(self, x):
in_height = x.size(2)
in_width = x.size(3)
pad_height = in_height - self.output_size[0]
pad_width = in_width - self.output_size[1]
pad_top = pad_height // 2
pad_bottom = pad_height - pad_top
pad_left = pad_width // 2
pad_right = pad_width - pad_left
return x[:, :, pad_top:in_height - pad_bottom, pad_left:in_width -
pad_right]
class ConvAE(nn.Module):
def __init__(self, channels, kernels):
"""
:param channels: a list containing all channels including the input image channel (1 for gray, 3 for RGB)
:param kernels: a list containing all kernel sizes, it should satisfy: len(kernels) = len(channels) - 1.
"""
super(ConvAE, self).__init__()
assert isinstance(channels, list) and isinstance(kernels, list)
self.encoder = nn.Sequential()
for i in range(1, len(channels)):
self.encoder.add_module('pad%d' % i, Conv2dSamePad(kernels[i -
1], 2))
self.encoder.add_module('conv%d' % i, nn.Conv2d(channels[i - 1],
channels[i], kernel_size=kernels[i - 1], stride=2))
self.encoder.add_module('relu%d' % i, nn.ReLU(True))
self.decoder = nn.Sequential()
channels = list(reversed(channels))
kernels = list(reversed(kernels))
sizes = [[12, 11], [24, 21], [48, 42]]
for i in range(len(channels) - 1):
self.decoder.add_module('deconv%d' % (i + 1), nn.
ConvTranspose2d(channels[i], channels[i + 1], kernel_size=
kernels[i], stride=2))
self.decoder.add_module('padd%d' % i, ConvTranspose2dSamePad(
sizes[i]))
self.decoder.add_module('relud%d' % i, nn.ReLU(True))
def forward(self, x):
h = self.encoder(x)
y = self.decoder(h)
return y
class SelfExpression(nn.Module):
def __init__(self, n):
super(SelfExpression, self).__init__()
self.Coefficient = nn.Parameter(0.0001 * torch.ones(n, n, dtype=
torch.float32), requires_grad=True)
def forward(self, x):
y = torch.matmul(self.Coefficient, x)
return y
class DSCNetNew(nn.Module):
def __init__(self, channels, kernels, num_sample):
super(DSCNetNew, self).__init__()
self.n = num_sample
self.ae = ConvAE(channels, kernels)
self.self_expression = SelfExpression(self.n)
def loss_fn(self, x, x_recon, z, z_recon, weight_coef, weight_selfExp):
loss_ae = 0.5 * F.mse_loss(x_recon, x, reduction='sum')
loss_coef = torch.sum(torch.pow(self.self_expression.Coefficient, 2))
loss_selfExp = 0.5 * F.mse_loss(z_recon, z, reduction='sum')
loss = (loss_ae + weight_coef * loss_coef + weight_selfExp *
loss_selfExp)
loss /= x.size(0)
return loss
def smoothLoss(self, z):
Z = torch.pow(z.unsqueeze(1) - z.unsqueeze(0), 2).sum(-1)
C = torch.abs(self.self_expression.Coefficient)
C = 0.5 * (C + torch.transpose(C, 0, 1))
C = C.fill_diagonal_(0)
loss_smooth = (Z * C).sum() / z.shape[0]
return loss_smooth
def forward(self, input_0):
primals_1 = self.ae.encoder.conv1.weight
primals_3 = self.ae.encoder.conv1.bias
primals_2 = self.ae.decoder.deconv1.weight
primals_6 = self.ae.decoder.deconv1.bias
primals_4 = self.self_expression.Coefficient
primals_5 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6])
return output[0], output[1], output[2]
|
qilinli/DSC-Net
|
DSCNet
| false
| 4,164
|
[
"MIT"
] | 0
|
c0e7a3cae3e07c34b2989234f568c7007cf0fc55
|
https://github.com/qilinli/DSC-Net/tree/c0e7a3cae3e07c34b2989234f568c7007cf0fc55
|
FuseLayer
|
from _paritybench_helpers import _mock_config
import torch
import torch.nn as nn
class FuseLayer(nn.Module):
def __init__(self, config):
super().__init__()
self.linear1 = nn.Linear(4 * config.hidden_size, config.hidden_size)
self.linear2 = nn.Linear(4 * config.hidden_size, config.hidden_size)
self.linear3 = nn.Linear(2 * config.hidden_size, config.hidden_size)
self.activation = nn.ReLU()
self.gate = nn.Sigmoid()
def forward(self, orig, input1, input2):
out1 = self.activation(self.linear1(torch.cat([orig, input1, orig -
input1, orig * input1], dim=-1)))
out2 = self.activation(self.linear2(torch.cat([orig, input2, orig -
input2, orig * input2], dim=-1)))
fuse_prob = self.gate(self.linear3(torch.cat([out1, out2], dim=-1)))
return fuse_prob * input1 + (1 - fuse_prob) * input2
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand(
[4, 4, 4, 4])]
def get_init_inputs():
return [[], {'config': _mock_config(hidden_size=4)}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = xindex // 16
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tmp12 = tl.full([1], 12, tl.int64)
tmp13 = tmp0 < tmp12
tmp14 = tmp11 & tmp13
tmp15 = tl.load(in_ptr0 + (4 * x1 + (-8 + x0)), tmp14 & xmask,
eviction_policy='evict_last', other=0.0)
tmp16 = tl.load(in_ptr1 + (4 * x1 + (-8 + x0)), tmp14 & xmask,
eviction_policy='evict_last', other=0.0)
tmp17 = tmp15 - tmp16
tmp18 = tl.full(tmp17.shape, 0.0, tmp17.dtype)
tmp19 = tl.where(tmp14, tmp17, tmp18)
tmp20 = tmp0 >= tmp12
tl.full([1], 16, tl.int64)
tmp23 = tl.load(in_ptr0 + (4 * x1 + (-12 + x0)), tmp20 & xmask,
eviction_policy='evict_last', other=0.0)
tmp24 = tl.load(in_ptr1 + (4 * x1 + (-12 + x0)), tmp20 & xmask,
eviction_policy='evict_last', other=0.0)
tmp25 = tmp23 * tmp24
tmp26 = tl.full(tmp25.shape, 0.0, tmp25.dtype)
tmp27 = tl.where(tmp20, tmp25, tmp26)
tmp28 = tl.where(tmp14, tmp19, tmp27)
tmp29 = tl.where(tmp9, tmp10, tmp28)
tmp30 = tl.where(tmp4, tmp5, tmp29)
tmp31 = tl.load(in_ptr2 + (4 * x1 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp32 = tl.load(in_ptr2 + (4 * x1 + (-8 + x0)), tmp14 & xmask,
eviction_policy='evict_last', other=0.0)
tmp33 = tmp15 - tmp32
tmp34 = tl.full(tmp33.shape, 0.0, tmp33.dtype)
tmp35 = tl.where(tmp14, tmp33, tmp34)
tmp36 = tl.load(in_ptr2 + (4 * x1 + (-12 + x0)), tmp20 & xmask,
eviction_policy='evict_last', other=0.0)
tmp37 = tmp23 * tmp36
tmp38 = tl.full(tmp37.shape, 0.0, tmp37.dtype)
tmp39 = tl.where(tmp20, tmp37, tmp38)
tmp40 = tl.where(tmp14, tmp35, tmp39)
tmp41 = tl.where(tmp9, tmp31, tmp40)
tmp42 = tl.where(tmp4, tmp5, tmp41)
tl.store(out_ptr0 + x2, tmp30, xmask)
tl.store(out_ptr1 + x2, tmp42, xmask)
@triton.jit
def triton_poi_fused_cat_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = xindex // 8
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tl.load(in_ptr1 + x0, tmp4 & xmask, eviction_policy='evict_last',
other=0.0)
tmp7 = tmp5 + tmp6
tmp8 = tl.full([1], 0, tl.int32)
tmp9 = triton_helpers.maximum(tmp8, tmp7)
tmp10 = tl.full(tmp9.shape, 0.0, tmp9.dtype)
tmp11 = tl.where(tmp4, tmp9, tmp10)
tmp12 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp15 = tl.load(in_ptr2 + (4 * x1 + (-4 + x0)), tmp12 & xmask,
eviction_policy='evict_last', other=0.0)
tmp16 = tl.load(in_ptr3 + (-4 + x0), tmp12 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp17 = tmp15 + tmp16
tmp18 = triton_helpers.maximum(tmp8, tmp17)
tmp19 = tl.full(tmp18.shape, 0.0, tmp18.dtype)
tmp20 = tl.where(tmp12, tmp18, tmp19)
tmp21 = tl.where(tmp4, tmp11, tmp20)
tl.store(out_ptr0 + x2, tmp21, xmask)
@triton.jit
def triton_poi_fused_add_mul_rsub_sigmoid_2(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp2 = tl.load(in_ptr1 + x0, xmask)
tmp6 = tl.load(in_ptr2 + x0, xmask)
tmp1 = tl.sigmoid(tmp0)
tmp3 = tmp1 * tmp2
tmp4 = 1.0
tmp5 = tmp4 - tmp1
tmp7 = tmp5 * tmp6
tmp8 = tmp3 + tmp7
tl.store(out_ptr0 + x0, tmp8, xmask)
@triton.jit
def triton_poi_fused_relu_threshold_backward_3(in_ptr0, in_ptr1, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(out_ptr0 + x2, tmp6, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, 16), (16, 1))
assert_size_stride(primals_4, (4,), (1,))
assert_size_stride(primals_5, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_6, (4, 16), (16, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 8), (8, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.
float32)
buf2 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.
float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(1024)](primals_1, primals_2, primals_5,
buf0, buf2, 1024, XBLOCK=256, num_warps=4, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (64, 16), (16, 1), 0),
reinterpret_tensor(primals_3, (16, 4), (1, 16), 0), out=buf1)
del primals_3
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (64, 16), (16, 1), 0),
reinterpret_tensor(primals_6, (16, 4), (1, 16), 0), out=buf3)
del primals_6
buf4 = empty_strided_cuda((4, 4, 4, 8), (128, 32, 8, 1), torch.float32)
triton_poi_fused_cat_1[grid(512)](buf1, primals_4, buf3, primals_7,
buf4, 512, XBLOCK=128, num_warps=4, num_stages=1)
buf5 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_9, reinterpret_tensor(buf4, (64, 8), (
8, 1), 0), reinterpret_tensor(primals_8, (8, 4), (1, 8), 0),
alpha=1, beta=1, out=buf5)
del primals_9
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_add_mul_rsub_sigmoid_2[grid(256)](buf5, primals_2,
primals_5, buf6, 256, XBLOCK=256, num_warps=4, num_stages=1)
buf7 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
triton_poi_fused_relu_threshold_backward_3[grid(256)](buf3,
primals_7, buf7, 256, XBLOCK=128, num_warps=4, num_stages=1)
del buf3
del primals_7
buf8 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
triton_poi_fused_relu_threshold_backward_3[grid(256)](buf1,
primals_4, buf8, 256, XBLOCK=128, num_warps=4, num_stages=1)
del buf1
del primals_4
return buf6, primals_2, primals_5, reinterpret_tensor(buf0, (64, 16), (
16, 1), 0), reinterpret_tensor(buf2, (64, 16), (16, 1), 0
), reinterpret_tensor(buf4, (64, 8), (8, 1), 0
), buf5, primals_8, buf7, buf8
class FuseLayerNew(nn.Module):
def __init__(self, config):
super().__init__()
self.linear1 = nn.Linear(4 * config.hidden_size, config.hidden_size)
self.linear2 = nn.Linear(4 * config.hidden_size, config.hidden_size)
self.linear3 = nn.Linear(2 * config.hidden_size, config.hidden_size)
self.activation = nn.ReLU()
self.gate = nn.Sigmoid()
def forward(self, input_0, input_1, input_2):
primals_3 = self.linear1.weight
primals_4 = self.linear1.bias
primals_6 = self.linear2.weight
primals_7 = self.linear2.bias
primals_8 = self.linear3.weight
primals_9 = self.linear3.bias
primals_1 = input_0
primals_2 = input_1
primals_5 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0]
|
qinyiwei/MuTual
|
FuseLayer
| false
| 4,165
|
[
"MIT"
] | 0
|
3bdd13c1388d6136b8944666dfd434870760cc93
|
https://github.com/qinyiwei/MuTual/tree/3bdd13c1388d6136b8944666dfd434870760cc93
|
AbsModule
|
import torch
class AbsModule(torch.nn.Module):
def __init__(self):
super(AbsModule, self).__init__()
def forward(self, x):
return torch.abs(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import math as tl_math
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_abs_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tl_math.abs(tmp0)
tl.store(out_ptr0 + x0, tmp1, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_abs_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del arg0_1
return buf0,
class AbsModuleNew(torch.nn.Module):
def __init__(self):
super(AbsModuleNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
mirecta/nncase
|
AbsModule
| false
| 4,166
|
[
"Apache-2.0"
] | 0
|
d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
https://github.com/mirecta/nncase/tree/d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
Tanh
|
import math
import torch
class Tanh(torch.nn.Tanh):
"""
Class that extends ``torch.nn.Tanh`` additionally computing the log diagonal
blocks of the Jacobian.
"""
def forward(self, inputs, grad: 'torch.Tensor'=None):
"""
Parameters
----------
inputs : ``torch.Tensor``, required.
The input tensor.
grad : ``torch.Tensor``, optional (default = None).
The log diagonal blocks of the partial Jacobian of previous transformations.
Returns
-------
The output tensor and the log diagonal blocks of the partial log-Jacobian of previous
transformations combined with this transformation.
"""
g = -2 * (inputs - math.log(2) + torch.nn.functional.softplus(-2 *
inputs))
return torch.tanh(inputs), g.view(grad.shape
) + grad if grad is not None else g
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_mul_softplus_sub_tanh_0(in_ptr0, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = libdevice.tanh(tmp0)
tmp2 = 0.6931471805599453
tmp3 = tmp0 - tmp2
tmp4 = -2.0
tmp5 = tmp0 * tmp4
tmp6 = 20.0
tmp7 = tmp5 > tmp6
tmp8 = tl_math.exp(tmp5)
tmp9 = libdevice.log1p(tmp8)
tmp10 = tl.where(tmp7, tmp5, tmp9)
tmp11 = tmp3 + tmp10
tmp12 = tmp11 * tmp4
tl.store(out_ptr0 + x0, tmp1, xmask)
tl.store(out_ptr1 + x0, tmp12, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_mul_softplus_sub_tanh_0[grid(256)](arg0_1,
buf0, buf1, 256, XBLOCK=256, num_warps=4, num_stages=1)
del arg0_1
return buf0, buf1
class TanhNew(torch.nn.Tanh):
"""
Class that extends ``torch.nn.Tanh`` additionally computing the log diagonal
blocks of the Jacobian.
"""
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0], output[1]
|
ralphc1212/BNAF
|
Tanh
| false
| 4,167
|
[
"MIT"
] | 0
|
b6e331aa96cdd4496b6eed6c6ce65512a99f4149
|
https://github.com/ralphc1212/BNAF/tree/b6e331aa96cdd4496b6eed6c6ce65512a99f4149
|
MHA
|
from _paritybench_helpers import _mock_config
import math
import torch
import torch.nn as nn
class MHA(nn.Module):
def __init__(self, config):
super().__init__()
self.num_attention_heads = config.num_attention_heads
self.hidden_size = config.hidden_size
self.attention_head_size = (config.hidden_size // config.
num_attention_heads)
self.all_head_size = (self.num_attention_heads * self.
attention_head_size)
self.query = nn.Linear(config.hidden_size, self.all_head_size)
self.key = nn.Linear(config.hidden_size, self.all_head_size)
self.value = nn.Linear(config.hidden_size, self.all_head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.
layer_norm_eps)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, input_ids_a, input_ids_b, attention_mask=None,
head_mask=None, output_attentions=False):
mixed_query_layer = self.query(input_ids_a)
mixed_key_layer = self.key(input_ids_b)
mixed_value_layer = self.value(input_ids_b)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
if head_mask is not None:
attention_probs = attention_probs * head_mask
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
w = self.dense.weight.t().view(self.num_attention_heads, self.
attention_head_size, self.hidden_size)
b = self.dense.bias
projected_context_layer = torch.einsum('bfnd,ndh->bfh',
context_layer, w) + b
projected_context_layer_dropout = self.dropout(projected_context_layer)
layernormed_context_layer = self.LayerNorm(input_ids_a +
projected_context_layer_dropout)
return (layernormed_context_layer, attention_probs
) if output_attentions else (layernormed_context_layer,)
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'config': _mock_config(num_attention_heads=4, hidden_size=
4, attention_probs_dropout_prob=0.5, layer_norm_eps=1)}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(out_ptr0 + (x2 + 4 * y3), tmp4, xmask & ymask)
@triton.jit
def triton_poi_fused_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp18 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp25 = tl.load(in_ptr1 + x2, xmask)
tmp26 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')
tmp27 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp29 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp31 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = float('-inf')
tmp2 = tmp0 == tmp1
tmp3 = tmp2 == 0
tmp4 = tmp3.to(tl.int64)
tmp5 = tmp4 != 0
tmp7 = tmp6 == tmp1
tmp8 = tmp7 == 0
tmp9 = tmp8.to(tl.int64)
tmp10 = tmp9 != 0
tmp11 = tmp5 | tmp10
tmp13 = tmp12 == tmp1
tmp14 = tmp13 == 0
tmp15 = tmp14.to(tl.int64)
tmp16 = tmp15 != 0
tmp17 = tmp11 | tmp16
tmp19 = tmp18 == tmp1
tmp20 = tmp19 == 0
tmp21 = tmp20.to(tl.int64)
tmp22 = tmp21 != 0
tmp23 = tmp17 | tmp22
tmp24 = tmp23 == 0
tmp28 = tmp26 + tmp27
tmp30 = tmp28 + tmp29
tmp32 = tmp30 + tmp31
tmp33 = tmp25 / tmp32
tmp34 = 0.0
tmp35 = tl.where(tmp24, tmp34, tmp33)
tl.store(out_ptr0 + x2, tmp35, xmask)
@triton.jit
def triton_poi_fused_3(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_5(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr2 + 0)
tmp3 = tl.broadcast_to(tmp2, [XBLOCK])
tmp6 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr2 + 1)
tmp9 = tl.broadcast_to(tmp8, [XBLOCK])
tmp13 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp14 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp15 = tl.load(in_ptr2 + 2)
tmp16 = tl.broadcast_to(tmp15, [XBLOCK])
tmp20 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp21 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp22 = tl.load(in_ptr2 + 3)
tmp23 = tl.broadcast_to(tmp22, [XBLOCK])
tmp4 = tmp1 + tmp3
tmp5 = tmp0 + tmp4
tmp10 = tmp7 + tmp9
tmp11 = tmp6 + tmp10
tmp12 = tmp5 + tmp11
tmp17 = tmp14 + tmp16
tmp18 = tmp13 + tmp17
tmp19 = tmp12 + tmp18
tmp24 = tmp21 + tmp23
tmp25 = tmp20 + tmp24
tmp26 = tmp19 + tmp25
tmp27 = 4.0
tmp28 = tmp26 / tmp27
tmp29 = tmp5 - tmp28
tmp30 = tmp29 * tmp29
tmp31 = tmp11 - tmp28
tmp32 = tmp31 * tmp31
tmp33 = tmp30 + tmp32
tmp34 = tmp18 - tmp28
tmp35 = tmp34 * tmp34
tmp36 = tmp33 + tmp35
tmp37 = tmp25 - tmp28
tmp38 = tmp37 * tmp37
tmp39 = tmp36 + tmp38
tmp40 = tmp39 / tmp27
tl.store(out_ptr0 + x0, tmp28, xmask)
tl.store(out_ptr1 + x0, tmp40, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_6(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, in_ptr6, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x1, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr6 + x0, xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp4 = tmp0 + tmp3
tmp6 = tmp4 - tmp5
tmp8 = 1.0
tmp9 = tmp7 + tmp8
tmp10 = libdevice.rsqrt(tmp9)
tmp11 = tmp6 * tmp10
tmp13 = tmp11 * tmp12
tmp15 = tmp13 + tmp14
tl.store(out_ptr0 + x2, tmp15, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12
) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_7, (4, 4), (4, 1))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (4, 4), (4, 1))
assert_size_stride(primals_10, (4,), (1,))
assert_size_stride(primals_11, (4,), (1,))
assert_size_stride(primals_12, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_6, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 4), (1, 4), 0), out=buf1)
del primals_4
buf2 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_6, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 4), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(16, 4)](buf0, primals_2, buf3, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_2
buf4 = reinterpret_tensor(buf0, (4, 4, 1, 4), (16, 4, 4, 1), 0)
del buf0
triton_poi_fused_0[grid(16, 4)](buf1, primals_5, buf4, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_5
buf5 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf4, (16, 1, 4), (4, 0, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_1[grid(256)](buf5, buf6, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf7 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_2[grid(256)](buf5, buf6, buf7, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf5
del buf6
buf8 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf1
triton_poi_fused_3[grid(16, 4)](buf2, primals_8, buf8, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_8
buf9 = reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf8, (16, 4, 1), (4, 1, 0), 0), out=buf9)
buf10 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
triton_poi_fused_clone_4[grid(16, 4)](buf9, buf10, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf11 = reinterpret_tensor(buf9, (1, 16, 4), (64, 4, 1), 0)
del buf9
extern_kernels.bmm(reinterpret_tensor(buf10, (1, 16, 4), (0, 4, 1),
0), reinterpret_tensor(primals_9, (1, 4, 4), (0, 1, 4), 0), out
=buf11)
buf12 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf13 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
triton_poi_fused_add_native_layer_norm_5[grid(16)](primals_3, buf11,
primals_10, buf12, buf13, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf14 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_native_layer_norm_6[grid(64)](primals_3, buf11,
primals_10, buf12, buf13, primals_11, primals_12, buf14, 64,
XBLOCK=64, num_warps=1, num_stages=1)
del buf12
del buf13
del primals_12
return buf14, primals_3, primals_10, primals_11, reinterpret_tensor(
primals_6, (16, 4), (4, 1), 0), buf7, reinterpret_tensor(buf8, (16,
1, 4), (4, 1, 1), 0), reinterpret_tensor(buf3, (16, 1, 4), (4, 1, 1), 0
), reinterpret_tensor(buf4, (16, 4, 1), (4, 1, 4), 0
), buf11, reinterpret_tensor(buf10, (1, 4, 16), (64, 1, 4), 0
), reinterpret_tensor(primals_9, (1, 4, 4), (4, 4, 1), 0)
class MHANew(nn.Module):
def __init__(self, config):
super().__init__()
self.num_attention_heads = config.num_attention_heads
self.hidden_size = config.hidden_size
self.attention_head_size = (config.hidden_size // config.
num_attention_heads)
self.all_head_size = (self.num_attention_heads * self.
attention_head_size)
self.query = nn.Linear(config.hidden_size, self.all_head_size)
self.key = nn.Linear(config.hidden_size, self.all_head_size)
self.value = nn.Linear(config.hidden_size, self.all_head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.
layer_norm_eps)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, input_0, input_1):
primals_1 = self.query.weight
primals_2 = self.query.bias
primals_4 = self.key.weight
primals_5 = self.key.bias
primals_7 = self.value.weight
primals_8 = self.value.bias
primals_9 = self.dense.weight
primals_10 = self.dense.bias
primals_11 = self.LayerNorm.weight
primals_12 = self.LayerNorm.bias
primals_3 = input_0
primals_6 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12])
return output[0]
|
qinyiwei/MuTual
|
MHA
| false
| 4,168
|
[
"MIT"
] | 0
|
3bdd13c1388d6136b8944666dfd434870760cc93
|
https://github.com/qinyiwei/MuTual/tree/3bdd13c1388d6136b8944666dfd434870760cc93
|
CosModule
|
import torch
class CosModule(torch.nn.Module):
def __init__(self):
super(CosModule, self).__init__()
def forward(self, x):
return torch.cos(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import math as tl_math
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_cos_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tl_math.cos(tmp0)
tl.store(out_ptr0 + x0, tmp1, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cos_0[grid(256)](arg0_1, buf0, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del arg0_1
return buf0,
class CosModuleNew(torch.nn.Module):
def __init__(self):
super(CosModuleNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
mirecta/nncase
|
CosModule
| false
| 4,169
|
[
"Apache-2.0"
] | 0
|
d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
https://github.com/mirecta/nncase/tree/d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
PopArt
|
import torch
import numpy as np
import torch.nn as nn
class PopArt(nn.Module):
"""Normalize a vector of observations - across the first norm_axes dimensions"""
def __init__(self, input_shape, norm_axes=1, beta=0.99999,
per_element_update=False, epsilon=1e-05, device=torch.device('cpu')):
super(PopArt, self).__init__()
self.input_shape = input_shape
self.norm_axes = norm_axes
self.epsilon = epsilon
self.beta = beta
self.per_element_update = per_element_update
self.tpdv = dict(dtype=torch.float32, device=device)
self.running_mean = nn.Parameter(torch.zeros(input_shape),
requires_grad=False)
self.running_mean_sq = nn.Parameter(torch.zeros(input_shape),
requires_grad=False)
self.debiasing_term = nn.Parameter(torch.tensor(0.0), requires_grad
=False)
def reset_parameters(self):
self.running_mean.zero_()
self.running_mean_sq.zero_()
self.debiasing_term.zero_()
def running_mean_var(self):
debiased_mean = self.running_mean / self.debiasing_term.clamp(min=
self.epsilon)
debiased_mean_sq = self.running_mean_sq / self.debiasing_term.clamp(min
=self.epsilon)
debiased_var = (debiased_mean_sq - debiased_mean ** 2).clamp(min=0.01)
return debiased_mean, debiased_var
def forward(self, input_vector, train=True):
if type(input_vector) == np.ndarray:
input_vector = torch.from_numpy(input_vector)
input_vector = input_vector
if train:
detached_input = input_vector.detach()
batch_mean = detached_input.mean(dim=tuple(range(self.norm_axes)))
batch_sq_mean = (detached_input ** 2).mean(dim=tuple(range(self
.norm_axes)))
if self.per_element_update:
batch_size = np.prod(detached_input.size()[:self.norm_axes])
weight = self.beta ** batch_size
else:
weight = self.beta
self.running_mean.mul_(weight).add_(batch_mean * (1.0 - weight))
self.running_mean_sq.mul_(weight).add_(batch_sq_mean * (1.0 -
weight))
self.debiasing_term.mul_(weight).add_(1.0 * (1.0 - weight))
mean, var = self.running_mean_var()
out = (input_vector - mean[(None,) * self.norm_axes]) / torch.sqrt(var
)[(None,) * self.norm_axes]
return out
def denormalize(self, input_vector):
"""Transform normalized data back into original distribution"""
if type(input_vector) == np.ndarray:
input_vector = torch.from_numpy(input_vector)
input_vector = input_vector
mean, var = self.running_mean_var()
out = input_vector * torch.sqrt(var)[(None,) * self.norm_axes] + mean[
(None,) * self.norm_axes]
out = out.cpu().numpy()
return out
def get_inputs():
return [torch.rand([4, 4])]
def get_init_inputs():
return [[], {'input_shape': 4}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice
import numpy as np
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_mean_mul_pow_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0,
out_ptr1, out_ptr2, out_ptr3, xnumel, XBLOCK: tl.constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp3 = tl.load(in_ptr1 + x0, xmask)
tmp4 = tl.load(in_ptr1 + (4 + x0), xmask)
tmp6 = tl.load(in_ptr1 + (8 + x0), xmask)
tmp8 = tl.load(in_ptr1 + (12 + x0), xmask)
tmp15 = tl.load(in_ptr2 + x0, xmask)
tmp1 = 0.99999
tmp2 = tmp0 * tmp1
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp9 = tmp7 + tmp8
tmp10 = 4.0
tmp11 = tmp9 / tmp10
tmp12 = 9.99999999995449e-06
tmp13 = tmp11 * tmp12
tmp14 = tmp2 + tmp13
tmp16 = tmp15 * tmp1
tmp17 = tmp3 * tmp3
tmp18 = tmp4 * tmp4
tmp19 = tmp17 + tmp18
tmp20 = tmp6 * tmp6
tmp21 = tmp19 + tmp20
tmp22 = tmp8 * tmp8
tmp23 = tmp21 + tmp22
tmp24 = tmp23 / tmp10
tmp25 = tmp24 * tmp12
tmp26 = tmp16 + tmp25
tl.store(out_ptr0 + x0, tmp14, xmask)
tl.store(out_ptr1 + x0, tmp26, xmask)
tl.store(out_ptr2 + x0, tmp14, xmask)
tl.store(out_ptr3 + x0, tmp26, xmask)
@triton.jit
def triton_poi_fused_div_sub_1(in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr2 + 0)
tmp3 = tl.broadcast_to(tmp2, [XBLOCK])
tmp12 = tl.load(in_ptr3 + x0, xmask, eviction_policy='evict_last')
tmp4 = 0.99999
tmp5 = tmp3 * tmp4
tmp6 = 9.99999999995449e-06
tmp7 = tmp5 + tmp6
tmp8 = 1e-05
tmp9 = triton_helpers.maximum(tmp7, tmp8)
tmp10 = tmp1 / tmp9
tmp11 = tmp0 - tmp10
tmp13 = tmp12 / tmp9
tmp14 = tmp10 * tmp10
tmp15 = tmp13 - tmp14
tmp16 = 0.01
tmp17 = triton_helpers.maximum(tmp15, tmp16)
tmp18 = libdevice.sqrt(tmp17)
tmp19 = tmp11 / tmp18
tl.store(out_ptr0 + x2, tmp19, xmask)
@triton.jit
def triton_poi_fused_add_mul_2(in_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr
):
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
tmp0 = tl.load(in_ptr0 + 0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
tmp2 = 0.99999
tmp3 = tmp1 * tmp2
tmp4 = 9.99999999995449e-06
tmp5 = tmp3 + tmp4
tl.store(out_ptr1 + tl.full([XBLOCK], 0, tl.int32), tmp5, None)
def call(args):
arg0_1, arg1_1, arg2_1, arg3_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4), (4, 1))
assert_size_stride(arg1_1, (4,), (1,))
assert_size_stride(arg2_1, (4,), (1,))
assert_size_stride(arg3_1, (), ())
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4,), (1,), torch.float32)
buf1 = empty_strided_cuda((4,), (1,), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_mean_mul_pow_0[grid(4)](arg1_1, arg0_1, arg2_1,
buf0, buf1, arg1_1, arg2_1, 4, XBLOCK=4, num_warps=1, num_stages=1)
del arg1_1
del arg2_1
buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
triton_poi_fused_div_sub_1[grid(16)](arg0_1, buf0, arg3_1, buf1,
buf2, 16, XBLOCK=16, num_warps=1, num_stages=1)
del arg0_1
del buf0
del buf1
triton_poi_fused_add_mul_2[grid(1)](arg3_1, arg3_1, 1, XBLOCK=1,
num_warps=1, num_stages=1)
del arg3_1
return buf2,
class PopArtNew(nn.Module):
"""Normalize a vector of observations - across the first norm_axes dimensions"""
def __init__(self, input_shape, norm_axes=1, beta=0.99999,
per_element_update=False, epsilon=1e-05, device=torch.device('cpu')):
super(PopArtNew, self).__init__()
self.input_shape = input_shape
self.norm_axes = norm_axes
self.epsilon = epsilon
self.beta = beta
self.per_element_update = per_element_update
self.tpdv = dict(dtype=torch.float32, device=device)
self.running_mean = nn.Parameter(torch.zeros(input_shape),
requires_grad=False)
self.running_mean_sq = nn.Parameter(torch.zeros(input_shape),
requires_grad=False)
self.debiasing_term = nn.Parameter(torch.tensor(0.0), requires_grad
=False)
def reset_parameters(self):
self.running_mean.zero_()
self.running_mean_sq.zero_()
self.debiasing_term.zero_()
def running_mean_var(self):
debiased_mean = self.running_mean / self.debiasing_term.clamp(min=
self.epsilon)
debiased_mean_sq = self.running_mean_sq / self.debiasing_term.clamp(min
=self.epsilon)
debiased_var = (debiased_mean_sq - debiased_mean ** 2).clamp(min=0.01)
return debiased_mean, debiased_var
def denormalize(self, input_vector):
"""Transform normalized data back into original distribution"""
if type(input_vector) == np.ndarray:
input_vector = torch.from_numpy(input_vector)
input_vector = input_vector
mean, var = self.running_mean_var()
out = input_vector * torch.sqrt(var)[(None,) * self.norm_axes] + mean[
(None,) * self.norm_axes]
out = out.cpu().numpy()
return out
def forward(self, input_0):
arg1_1 = self.running_mean
arg2_1 = self.running_mean_sq
arg3_1 = self.debiasing_term
arg0_1 = input_0
output = call([arg0_1, arg1_1, arg2_1, arg3_1])
return output[0]
|
rainwangphy/TRPO-in-MARL
|
PopArt
| false
| 4,170
|
[
"MIT"
] | 0
|
22229abba417708922ecf6455c1c5180dbe80391
|
https://github.com/rainwangphy/TRPO-in-MARL/tree/22229abba417708922ecf6455c1c5180dbe80391
|
RegressionHead
|
import abc
import torch
import torch.nn as nn
from torch.nn.functional import *
import torch.utils.data.dataset
class BaseHead(nn.Module, metaclass=abc.ABCMeta):
"""Absract class for task heads"""
@abc.abstractmethod
def __init__(self):
super().__init__()
class RegressionHead(BaseHead):
def __init__(self, task, hidden_size, hidden_dropout_prob, **kwargs):
"""From RobertaClassificationHead"""
super().__init__()
self.dense = nn.Linear(hidden_size, hidden_size)
self.dropout = nn.Dropout(hidden_dropout_prob)
self.out_proj = nn.Linear(hidden_size, 1)
def forward(self, pooled):
x = self.dropout(pooled)
x = self.dense(x)
x = torch.tanh(x)
x = self.dropout(x)
scores = self.out_proj(x)
return scores
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'task': 4, 'hidden_size': 4, 'hidden_dropout_prob': 0.5}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import abc
import torch.nn as nn
from torch.nn.functional import *
import torch.utils.data.dataset
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_tanh_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = libdevice.tanh(tmp2)
tl.store(in_out_ptr0 + x2, tmp3, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (1, 4), (4, 1))
assert_size_stride(primals_5, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), out=buf0)
del primals_2
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
get_raw_stream(0)
triton_poi_fused_tanh_0[grid(256)](buf1, primals_3, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_3
buf3 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 4), (
4, 1), 0), reinterpret_tensor(primals_4, (4, 1), (1, 4), 0),
alpha=1, beta=1, out=buf3)
del primals_5
return reinterpret_tensor(buf3, (4, 4, 4, 1), (16, 4, 1, 1), 0
), reinterpret_tensor(primals_1, (64, 4), (4, 1), 0), buf1, primals_4
class BaseHead(nn.Module, metaclass=abc.ABCMeta):
"""Absract class for task heads"""
@abc.abstractmethod
def __init__(self):
super().__init__()
class RegressionHeadNew(BaseHead):
def __init__(self, task, hidden_size, hidden_dropout_prob, **kwargs):
"""From RobertaClassificationHead"""
super().__init__()
self.dense = nn.Linear(hidden_size, hidden_size)
self.dropout = nn.Dropout(hidden_dropout_prob)
self.out_proj = nn.Linear(hidden_size, 1)
def forward(self, input_0):
primals_2 = self.dense.weight
primals_3 = self.dense.bias
primals_4 = self.out_proj.weight
primals_5 = self.out_proj.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
mfk3138/jiant
|
RegressionHead
| false
| 4,171
|
[
"MIT"
] | 0
|
6e67ff1ecb1bb98533c1019a86af4ad2c04c6a64
|
https://github.com/mfk3138/jiant/tree/6e67ff1ecb1bb98533c1019a86af4ad2c04c6a64
|
CeilModule
|
import torch
class CeilModule(torch.nn.Module):
def __init__(self):
super(CeilModule, self).__init__()
def forward(self, x):
return torch.ceil(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_ceil_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = libdevice.ceil(tmp0)
tl.store(out_ptr0 + x0, tmp1, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_ceil_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del arg0_1
return buf0,
class CeilModuleNew(torch.nn.Module):
def __init__(self):
super(CeilModuleNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
mirecta/nncase
|
CeilModule
| false
| 4,172
|
[
"Apache-2.0"
] | 0
|
d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
https://github.com/mirecta/nncase/tree/d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
AttFlowLayer
|
import torch
import torch.nn as nn
import torch.nn.functional as F
class AttFlowLayer(nn.Module):
def __init__(self, embed_length):
super(AttFlowLayer, self).__init__()
self.embed_length = embed_length
self.alpha = nn.Linear(3 * embed_length, 1, bias=False)
def forward(self, context, query):
batch_size = context.shape[0]
query = query.unsqueeze(0).expand((batch_size, query.shape[0], self
.embed_length))
shape = batch_size, context.shape[1], query.shape[1], self.embed_length
context_extended = context.unsqueeze(2).expand(shape)
query_extended = query.unsqueeze(1).expand(shape)
multiplied = torch.mul(context_extended, query_extended)
cated = torch.cat((context_extended, query_extended, multiplied), 3)
S = self.alpha(cated).view(batch_size, context.shape[1], query.shape[1]
)
S_softmax_row = F.softmax(S, dim=1).permute(0, 2, 1)
F.softmax(S, dim=2)
query_masks = torch.sign(torch.abs(torch.sum(query, dim=-1)))
query_masks = torch.unsqueeze(query_masks, 2).repeat(1, 1, context.
size()[1])
S_softmax_row = S_softmax_row * query_masks
S_softmax_row_1 = S_softmax_row.unsqueeze(3).expand(S_softmax_row.
shape[0], S_softmax_row.shape[1], S_softmax_row.shape[2], self.
embed_length)
context_1 = context_extended.permute(0, 2, 1, 3)
attd = torch.mul(S_softmax_row_1, context_1)
G = torch.sum(attd, 1)
H = torch.sum(attd, 2)
G = torch.cat((context, G), 2)
return G, H
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'embed_length': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 768
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 12
x2 = xindex // 48
x1 = xindex // 12 % 4
x3 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x2 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp9 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tl.full([1], 12, tl.int64)
tmp14 = tl.load(in_ptr0 + (4 * x2 + (-8 + x0)), tmp11 & xmask,
eviction_policy='evict_last', other=0.0)
tmp15 = tl.load(in_ptr1 + (4 * x1 + (-8 + x0)), tmp11 & xmask,
eviction_policy='evict_last', other=0.0)
tmp16 = tmp14 * tmp15
tmp17 = tl.full(tmp16.shape, 0.0, tmp16.dtype)
tmp18 = tl.where(tmp11, tmp16, tmp17)
tmp19 = tl.where(tmp9, tmp10, tmp18)
tmp20 = tl.where(tmp4, tmp5, tmp19)
tl.store(out_ptr0 + x3, tmp20, xmask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 4
x2 = xindex // 16
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (4 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (8 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (12 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x3, tmp9, xmask)
@triton.jit
def triton_poi_fused_mul_repeat_2(in_ptr0, in_ptr1, out_ptr0, out_ptr1,
ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
y0 = yindex % 4
x2 = xindex
y3 = yindex
y1 = yindex // 4
tmp0 = tl.load(in_ptr0 + 4 * y0, ymask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 4 * y0), ymask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 4 * y0), ymask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 4 * y0), ymask, eviction_policy='evict_last')
tmp15 = tl.load(in_ptr1 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp16 = tl.load(in_ptr1 + (y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp17 = tl.load(in_ptr1 + (4 + y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp19 = tl.load(in_ptr1 + (8 + y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp21 = tl.load(in_ptr1 + (12 + y0 + 16 * y1), ymask, eviction_policy=
'evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = tl_math.abs(tmp6)
tmp8 = tl.full([1, 1], 0, tl.int32)
tmp9 = tmp8 < tmp7
tmp10 = tmp9.to(tl.int8)
tmp11 = tmp7 < tmp8
tmp12 = tmp11.to(tl.int8)
tmp13 = tmp10 - tmp12
tmp14 = tmp13.to(tmp7.dtype)
tmp18 = tmp16 + tmp17
tmp20 = tmp18 + tmp19
tmp22 = tmp20 + tmp21
tmp23 = tmp15 / tmp22
tmp24 = tmp23 * tmp14
tl.store(out_ptr0 + (x2 + 4 * y3), tmp14, xmask & ymask)
tl.store(out_ptr1 + (x2 + 4 * y3), tmp24, xmask & ymask)
@triton.jit
def triton_poi_fused_mul_sum_3(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex // 4
x0 = xindex % 4
x2 = xindex // 16
x4 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x3, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x3), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (4 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x3), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (8 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x3), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (12 + x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tmp0 * tmp1
tmp5 = tmp3 * tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 * tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 * tmp12
tmp14 = tmp10 + tmp13
tl.store(out_ptr0 + x4, tmp14, xmask)
@triton.jit
def triton_poi_fused_cat_4(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x3 = xindex // 8
x1 = xindex // 8 % 4
x2 = xindex // 32
x4 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x3 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (x1 + 16 * x2), tmp6 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp10 = tl.load(in_ptr0 + (4 * x3 + (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = tmp9 * tmp10
tmp12 = tl.load(in_ptr1 + (4 + x1 + 16 * x2), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp13 = tmp12 * tmp10
tmp14 = tmp11 + tmp13
tmp15 = tl.load(in_ptr1 + (8 + x1 + 16 * x2), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp16 = tmp15 * tmp10
tmp17 = tmp14 + tmp16
tmp18 = tl.load(in_ptr1 + (12 + x1 + 16 * x2), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp19 = tmp18 * tmp10
tmp20 = tmp17 + tmp19
tmp21 = tl.full(tmp20.shape, 0.0, tmp20.dtype)
tmp22 = tl.where(tmp6, tmp20, tmp21)
tmp23 = tl.where(tmp4, tmp5, tmp22)
tl.store(out_ptr0 + x4, tmp23, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (1, 12), (12, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 12), (192, 48, 12, 1), torch.
float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(768)](primals_1, primals_2, buf0, 768,
XBLOCK=128, num_warps=4, num_stages=1)
buf1 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (64, 12), (12, 1), 0),
reinterpret_tensor(primals_3, (12, 1), (1, 12), 0), out=buf1)
del primals_3
buf2 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused__softmax_1[grid(64)](buf1, buf2, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf3 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
buf4 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_mul_repeat_2[grid(16, 4)](primals_2, buf2, buf3,
buf4, 16, 4, XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_2
buf5 = buf2
del buf2
triton_poi_fused_mul_sum_3[grid(64)](buf4, primals_1, buf5, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf6 = empty_strided_cuda((4, 4, 8), (32, 8, 1), torch.float32)
triton_poi_fused_cat_4[grid(128)](primals_1, buf4, buf6, 128,
XBLOCK=128, num_warps=4, num_stages=1)
del buf4
return buf6, buf5, primals_1, reinterpret_tensor(buf0, (64, 12), (12, 1), 0
), buf1, buf3
class AttFlowLayerNew(nn.Module):
def __init__(self, embed_length):
super(AttFlowLayerNew, self).__init__()
self.embed_length = embed_length
self.alpha = nn.Linear(3 * embed_length, 1, bias=False)
def forward(self, input_0, input_1):
primals_3 = self.alpha.weight
primals_1 = input_0
primals_2 = input_1
output = call([primals_1, primals_2, primals_3])
return output[0], output[1]
|
qtxcm/Joint_NER_with_NTP
|
AttFlowLayer
| false
| 4,173
|
[
"Apache-2.0"
] | 0
|
02f26f2cc891d36808b2e28f337cc4846524e5df
|
https://github.com/qtxcm/Joint_NER_with_NTP/tree/02f26f2cc891d36808b2e28f337cc4846524e5df
|
SqrtModule
|
import torch
class SqrtModule(torch.nn.Module):
def __init__(self):
super(SqrtModule, self).__init__()
def forward(self, x):
return torch.sqrt(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_sqrt_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = libdevice.sqrt(tmp0)
tl.store(out_ptr0 + x0, tmp1, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_sqrt_0[grid(256)](arg0_1, buf0, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del arg0_1
return buf0,
class SqrtModuleNew(torch.nn.Module):
def __init__(self):
super(SqrtModuleNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
mirecta/nncase
|
SqrtModule
| false
| 4,174
|
[
"Apache-2.0"
] | 0
|
d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
https://github.com/mirecta/nncase/tree/d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
ReduceMaxModule
|
import torch
class ReduceMaxModule(torch.nn.Module):
def __init__(self):
super(ReduceMaxModule, self).__init__()
def forward(self, x):
return torch.max(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_max_0(in_ptr0, out_ptr0, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.broadcast_to(tmp0, [RBLOCK])
tmp3 = triton_helpers.promote_to_tensor(triton_helpers.max2(tmp1, 0))
tl.store(out_ptr0 + tl.full([1], 0, tl.int32), tmp3, None)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
get_raw_stream(0)
triton_per_fused_max_0[grid(1)](arg0_1, buf0, 1, 256, num_warps=2,
num_stages=1)
del arg0_1
return buf0,
class ReduceMaxModuleNew(torch.nn.Module):
def __init__(self):
super(ReduceMaxModuleNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
mirecta/nncase
|
ReduceMaxModule
| false
| 4,175
|
[
"Apache-2.0"
] | 0
|
d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
https://github.com/mirecta/nncase/tree/d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
NegModule
|
import torch
class NegModule(torch.nn.Module):
def __init__(self):
super(NegModule, self).__init__()
def forward(self, x):
return -x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_neg_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = -tmp0
tl.store(out_ptr0 + x0, tmp1, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_neg_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del arg0_1
return buf0,
class NegModuleNew(torch.nn.Module):
def __init__(self):
super(NegModuleNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
mirecta/nncase
|
NegModule
| false
| 4,176
|
[
"Apache-2.0"
] | 0
|
d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
https://github.com/mirecta/nncase/tree/d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
FloorModule
|
import torch
class FloorModule(torch.nn.Module):
def __init__(self):
super(FloorModule, self).__init__()
def forward(self, x):
return torch.floor(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_floor_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = libdevice.floor(tmp0)
tl.store(out_ptr0 + x0, tmp1, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_floor_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del arg0_1
return buf0,
class FloorModuleNew(torch.nn.Module):
def __init__(self):
super(FloorModuleNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
mirecta/nncase
|
FloorModule
| false
| 4,177
|
[
"Apache-2.0"
] | 0
|
d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
https://github.com/mirecta/nncase/tree/d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
ReduceMeanModule
|
import torch
class ReduceMeanModule(torch.nn.Module):
def __init__(self):
super(ReduceMeanModule, self).__init__()
def forward(self, x):
return torch.mean(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_mean_0(in_out_ptr0, in_ptr0, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.broadcast_to(tmp0, [RBLOCK])
tmp3 = triton_helpers.promote_to_tensor(tl.sum(tmp1, 0))
tmp4 = 256.0
tmp5 = tmp3 / tmp4
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp5, None)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_mean_0[grid(1)](buf1, arg0_1, 1, 256, num_warps=2,
num_stages=1)
del arg0_1
return buf1,
class ReduceMeanModuleNew(torch.nn.Module):
def __init__(self):
super(ReduceMeanModuleNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
mirecta/nncase
|
ReduceMeanModule
| false
| 4,178
|
[
"Apache-2.0"
] | 0
|
d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
https://github.com/mirecta/nncase/tree/d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
ReduceMinModule
|
import torch
class ReduceMinModule(torch.nn.Module):
def __init__(self):
super(ReduceMinModule, self).__init__()
def forward(self, x):
return torch.min(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_min_0(in_ptr0, out_ptr0, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.broadcast_to(tmp0, [RBLOCK])
tmp3 = triton_helpers.promote_to_tensor(triton_helpers.min2(tmp1, 0))
tl.store(out_ptr0 + tl.full([1], 0, tl.int32), tmp3, None)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
get_raw_stream(0)
triton_per_fused_min_0[grid(1)](arg0_1, buf0, 1, 256, num_warps=2,
num_stages=1)
del arg0_1
return buf0,
class ReduceMinModuleNew(torch.nn.Module):
def __init__(self):
super(ReduceMinModuleNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
mirecta/nncase
|
ReduceMinModule
| false
| 4,179
|
[
"Apache-2.0"
] | 0
|
d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
https://github.com/mirecta/nncase/tree/d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
DenseSAGEConv
|
import math
import torch
import torch.nn.functional as F
from torch.nn import Parameter
import torch.utils.data
def uniform(size, tensor):
bound = 1.0 / math.sqrt(size)
if tensor is not None:
tensor.data.uniform_(-bound, bound)
class DenseSAGEConv(torch.nn.Module):
"""See :class:`torch_geometric.nn.conv.SAGEConv`.
"""
def __init__(self, in_channels, out_channels, normalize=False, bias=True):
super(DenseSAGEConv, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.normalize = normalize
self.weight = Parameter(torch.Tensor(self.in_channels, out_channels))
if bias:
self.bias = Parameter(torch.Tensor(out_channels))
else:
self.register_parameter('bias', None)
self.reset_parameters()
def reset_parameters(self):
uniform(self.in_channels, self.weight)
uniform(self.in_channels, self.bias)
def forward(self, x, adj, mask=None, add_loop=True):
"""
Args:
x (Tensor): Node feature tensor :math:`\\mathbf{X} \\in \\mathbb{R}^{B
\\times N \\times F}`, with batch-size :math:`B`, (maximum)
number of nodes :math:`N` for each graph, and feature
dimension :math:`F`.
adj (Tensor): Adjacency tensor :math:`\\mathbf{A} \\in \\mathbb{R}^{B
\\times N \\times N}`. The adjacency tensor is broadcastable in
the batch dimension, resulting in a shared adjacency matrix for
the complete batch.
mask (BoolTensor, optional): Mask matrix
:math:`\\mathbf{M} \\in {\\{ 0, 1 \\}}^{B \\times N}` indicating
the valid nodes for each graph. (default: :obj:`None`)
add_loop (bool, optional): If set to :obj:`False`, the layer will
not automatically add self-loops to the adjacency matrices.
(default: :obj:`True`)
"""
x = x.unsqueeze(0) if x.dim() == 2 else x
adj = adj.unsqueeze(0) if adj.dim() == 2 else adj
B, N, _ = adj.size()
if add_loop:
adj = adj.clone()
idx = torch.arange(N, dtype=torch.long, device=adj.device)
adj[:, idx, idx] = 1
out = torch.matmul(adj, x)
out = out / adj.sum(dim=-1, keepdim=True).clamp(min=1)
out = torch.matmul(out, self.weight)
if self.bias is not None:
out = out + self.bias
if self.normalize:
out = F.normalize(out, p=2, dim=-1)
if mask is not None:
out = out * mask.view(B, N, 1)
return out
def __repr__(self):
return '{}({}, {})'.format(self.__class__.__name__, self.
in_channels, self.out_channels)
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import math
from torch.nn import Parameter
import torch.utils.data
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_index_put_lift_fresh_0(in_ptr0, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tl.store(out_ptr0 + x0, tmp0, xmask)
@triton.jit
def triton_poi_fused_index_put_lift_fresh_1(out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
tmp0 = 1.0
tl.store(out_ptr0 + (5 * x0 + 16 * x1), tmp0, xmask)
@triton.jit
def triton_poi_fused_clone_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 64
x2 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tl.store(out_ptr0 + x2, tmp0, xmask)
@triton.jit
def triton_poi_fused_clamp_div_sum_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 4 % 16
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = 1.0
tmp9 = triton_helpers.maximum(tmp7, tmp8)
tmp10 = tmp0 / tmp9
tl.store(in_out_ptr0 + x3, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_4(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_index_put_lift_fresh_0[grid(64)](primals_2, buf0,
64, XBLOCK=64, num_warps=1, num_stages=1)
del primals_2
triton_poi_fused_index_put_lift_fresh_1[grid(16)](buf0, 16, XBLOCK=
16, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_2[grid(256)](buf0, buf2, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf3 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(primals_1, (16, 4, 4), (16, 4, 1), 0),
out=buf3)
del primals_1
buf4 = reinterpret_tensor(buf3, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf3
triton_poi_fused_clamp_div_sum_3[grid(256)](buf4, buf0, 256, XBLOCK
=128, num_warps=4, num_stages=1)
del buf0
buf5 = reinterpret_tensor(buf2, (64, 4), (4, 1), 0)
del buf2
extern_kernels.mm(reinterpret_tensor(buf4, (64, 4), (4, 1), 0),
primals_3, out=buf5)
del primals_3
buf6 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused_add_4[grid(256)](buf6, primals_4, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_4
return buf6, reinterpret_tensor(buf4, (4, 64), (1, 4), 0)
def uniform(size, tensor):
bound = 1.0 / math.sqrt(size)
if tensor is not None:
tensor.data.uniform_(-bound, bound)
class DenseSAGEConvNew(torch.nn.Module):
"""See :class:`torch_geometric.nn.conv.SAGEConv`.
"""
def __init__(self, in_channels, out_channels, normalize=False, bias=True):
super(DenseSAGEConvNew, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.normalize = normalize
self.weight = Parameter(torch.Tensor(self.in_channels, out_channels))
if bias:
self.bias = Parameter(torch.Tensor(out_channels))
else:
self.register_parameter('bias', None)
self.reset_parameters()
def reset_parameters(self):
uniform(self.in_channels, self.weight)
uniform(self.in_channels, self.bias)
def __repr__(self):
return '{}({}, {})'.format(self.__class__.__name__, self.
in_channels, self.out_channels)
def forward(self, input_0, input_1):
primals_3 = self.weight
primals_4 = self.bias
primals_1 = input_0
primals_2 = input_1
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0]
|
rbshi/pytorch_geometric
|
DenseSAGEConv
| false
| 4,180
|
[
"MIT"
] | 0
|
fcfbad49219974689eb5c6e32365939ae09ace84
|
https://github.com/rbshi/pytorch_geometric/tree/fcfbad49219974689eb5c6e32365939ae09ace84
|
ResizeModule
|
import torch
class ResizeModule(torch.nn.Module):
def __init__(self):
super(ResizeModule, self).__init__()
def forward(self, x):
return torch.nn.functional.interpolate(x, size=(3, 4))
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused__unsafe_index_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 192
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4 % 3
x0 = xindex % 4
x2 = xindex // 12
x4 = xindex
tmp0 = x1
tmp1 = tmp0.to(tl.float32)
tmp2 = 1.3333333333333333
tmp3 = tmp1 * tmp2
tmp4 = tmp3.to(tl.int32)
tmp5 = x0
tmp6 = tmp5.to(tl.float32)
tmp7 = 1.0
tmp8 = tmp6 * tmp7
tmp9 = tmp8.to(tl.int32)
tmp10 = tl.load(in_ptr0 + (tmp9 + 4 * tmp4 + 16 * x2), xmask,
eviction_policy='evict_last')
tl.store(out_ptr0 + x4, tmp10, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 3, 4), (48, 12, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__unsafe_index_0[grid(192)](arg0_1, buf0, 192,
XBLOCK=256, num_warps=4, num_stages=1)
del arg0_1
return buf0,
class ResizeModuleNew(torch.nn.Module):
def __init__(self):
super(ResizeModuleNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
mirecta/nncase
|
ResizeModule
| false
| 4,181
|
[
"Apache-2.0"
] | 0
|
d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
https://github.com/mirecta/nncase/tree/d2efa59677a26f4259b3b6a5b6ec05ea16d4e40c
|
RMSELoss
|
import torch
from torch import nn
import torch.cuda
class RMSELoss(nn.Module):
def __init__(self, eps=1e-06):
super().__init__()
self.mse = nn.MSELoss()
self.eps = eps
def forward(self, yhat, y):
loss = torch.sqrt(self.mse(yhat, y) + self.eps)
return loss
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice
from torch import nn
import torch.cuda
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_add_mse_loss_sqrt_0(in_out_ptr0, in_ptr0, in_ptr1,
xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.load(in_ptr1 + r0, None)
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp4 = tl.broadcast_to(tmp3, [RBLOCK])
tmp6 = triton_helpers.promote_to_tensor(tl.sum(tmp4, 0))
tmp7 = 256.0
tmp8 = tmp6 / tmp7
tmp9 = 1e-06
tmp10 = tmp8 + tmp9
tmp11 = libdevice.sqrt(tmp10)
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp11, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_add_mse_loss_sqrt_0[grid(1)](buf1, arg1_1, arg0_1,
1, 256, num_warps=2, num_stages=1)
del arg0_1
del arg1_1
return buf1,
class RMSELossNew(nn.Module):
def __init__(self, eps=1e-06):
super().__init__()
self.mse = nn.MSELoss()
self.eps = eps
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
|
rgbayrak/multi-task-physio
|
RMSELoss
| false
| 4,182
|
[
"MIT"
] | 0
|
01ea98f26cc9b96ec94105d5213cb1ef93673c2c
|
https://github.com/rgbayrak/multi-task-physio/tree/01ea98f26cc9b96ec94105d5213cb1ef93673c2c
|
_ASPPModule
|
import torch
import torch.nn as nn
class _ASPPModule(nn.Module):
"""Atrous Spatial Pyramid Pooling"""
def __init__(self, in_channels, out_channels, pyramids):
super(_ASPPModule, self).__init__()
self.stages = nn.Module()
for i, (dilation, padding) in enumerate(zip(pyramids, pyramids)):
self.stages.add_module('c{}'.format(i), nn.Conv2d(in_channels=
in_channels, out_channels=out_channels, kernel_size=3,
stride=1, padding=padding, dilation=dilation, bias=True))
for m in self.stages.children():
nn.init.normal(m.weight, mean=0, std=0.01)
nn.init.constant(m.bias, 0)
def forward(self, x):
h = 0
for stage in self.stages.children():
h += stage(x)
return h
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'out_channels': 4, 'pyramids': [4, 4]}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
@triton.jit
def triton_poi_fused_add_convolution_0(in_out_ptr0, in_ptr0, in_ptr1,
in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + x3, xmask)
tmp6 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 + tmp3
tmp7 = tmp5 + tmp6
tmp8 = tmp4 + tmp7
tl.store(in_out_ptr0 + x3, tmp8, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 3, 3), (36, 9, 3, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 4, 3, 3), (36, 9, 3, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,
1), padding=(4, 4), dilation=(4, 4), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 4, 4), (64, 16, 4, 1))
buf1 = extern_kernels.convolution(primals_3, primals_4, stride=(1,
1), padding=(4, 4), dilation=(4, 4), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 4, 4), (64, 16, 4, 1))
buf2 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_add_convolution_0[grid(256)](buf2, primals_2, buf1,
primals_5, 256, XBLOCK=256, num_warps=4, num_stages=1)
del buf1
del primals_2
del primals_5
return buf2, primals_1, primals_3, primals_4
class _ASPPModuleNew(nn.Module):
"""Atrous Spatial Pyramid Pooling"""
def __init__(self, in_channels, out_channels, pyramids):
super(_ASPPModuleNew, self).__init__()
self.stages = nn.Module()
for i, (dilation, padding) in enumerate(zip(pyramids, pyramids)):
self.stages.add_module('c{}'.format(i), nn.Conv2d(in_channels=
in_channels, out_channels=out_channels, kernel_size=3,
stride=1, padding=padding, dilation=dilation, bias=True))
for m in self.stages.children():
nn.init.normal(m.weight, mean=0, std=0.01)
nn.init.constant(m.bias, 0)
def forward(self, input_0):
primals_1 = self.stages.c0.weight
primals_2 = self.stages.c0.bias
primals_4 = self.stages.c1.weight
primals_5 = self.stages.c1.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
reyuwei/deeplab-pytorch
|
_ASPPModule
| false
| 4,183
|
[
"MIT"
] | 0
|
f4e241c83be5f85f0f2e1be5d76160b8c2d7ec9a
|
https://github.com/reyuwei/deeplab-pytorch/tree/f4e241c83be5f85f0f2e1be5d76160b8c2d7ec9a
|
Net
|
import torch
import torch.nn as nn
class Net(nn.Module):
def __init__(self, input_size):
super(Net, self).__init__()
hlayer1 = int(input_size * 10)
hlayer2 = int(input_size * 10 / 2)
self.fc1 = nn.Linear(input_size, hlayer1)
self.relu1 = nn.ReLU()
self.fc2 = nn.Linear(hlayer1, hlayer2)
self.prelu = nn.PReLU(1)
self.out = nn.Linear(hlayer2, 1)
self.out_act = nn.Sigmoid()
def forward(self, input_):
a1 = self.fc1(input_)
h1 = self.relu1(a1)
a2 = self.fc2(h1)
h2 = self.prelu(a2)
a3 = self.out(h2)
y = self.out_act(a3)
return y
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'input_size': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_relu_threshold_backward_0(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 2560
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 40
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, xmask)
tl.store(out_ptr0 + x2, tmp6, xmask)
@triton.jit
def triton_poi_fused__prelu_kernel_1(in_ptr0, in_ptr1, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 1280
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp3 = tl.load(in_ptr1 + 0)
tmp4 = tl.broadcast_to(tmp3, [XBLOCK])
tmp1 = 0.0
tmp2 = tmp0 > tmp1
tmp5 = tmp4 * tmp0
tmp6 = tl.where(tmp2, tmp0, tmp5)
tl.store(out_ptr0 + x0, tmp6, xmask)
@triton.jit
def triton_poi_fused_sigmoid_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr0 + 0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK])
tmp3 = tmp0 + tmp2
tmp4 = tl.sigmoid(tmp3)
tl.store(in_out_ptr0 + x0, tmp4, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8) = args
args.clear()
assert_size_stride(primals_1, (40, 4), (4, 1))
assert_size_stride(primals_2, (40,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (20, 40), (40, 1))
assert_size_stride(primals_5, (20,), (1,))
assert_size_stride(primals_6, (1,), (1,))
assert_size_stride(primals_7, (1, 20), (20, 1))
assert_size_stride(primals_8, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 40), (40, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 40), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 40), (640, 160, 40, 1), 0)
del buf0
buf6 = empty_strided_cuda((4, 4, 4, 40), (640, 160, 40, 1), torch.bool)
get_raw_stream(0)
triton_poi_fused_relu_threshold_backward_0[grid(2560)](buf1,
primals_2, buf6, 2560, XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((64, 20), (20, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 40),
(40, 1), 0), reinterpret_tensor(primals_4, (40, 20), (1, 40), 0
), alpha=1, beta=1, out=buf2)
del primals_5
buf3 = empty_strided_cuda((4, 4, 4, 20), (320, 80, 20, 1), torch.
float32)
triton_poi_fused__prelu_kernel_1[grid(1280)](buf2, primals_6, buf3,
1280, XBLOCK=128, num_warps=4, num_stages=1)
buf4 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf3, (64, 20), (20, 1), 0),
reinterpret_tensor(primals_7, (20, 1), (1, 20), 0), out=buf4)
buf5 = reinterpret_tensor(buf4, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf4
triton_poi_fused_sigmoid_2[grid(64)](buf5, primals_8, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del primals_8
return buf5, primals_6, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), reinterpret_tensor(buf1, (64, 40), (40, 1), 0
), buf2, reinterpret_tensor(buf3, (64, 20), (20, 1), 0
), buf5, primals_7, primals_4, buf6
class NetNew(nn.Module):
def __init__(self, input_size):
super(NetNew, self).__init__()
hlayer1 = int(input_size * 10)
hlayer2 = int(input_size * 10 / 2)
self.fc1 = nn.Linear(input_size, hlayer1)
self.relu1 = nn.ReLU()
self.fc2 = nn.Linear(hlayer1, hlayer2)
self.prelu = nn.PReLU(1)
self.out = nn.Linear(hlayer2, 1)
self.out_act = nn.Sigmoid()
def forward(self, input_0):
primals_1 = self.fc1.weight
primals_2 = self.fc1.bias
primals_4 = self.fc2.weight
primals_5 = self.fc2.bias
primals_6 = self.prelu.weight
primals_7 = self.out.weight
primals_8 = self.out.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8])
return output[0]
|
rcaborges/music-cold-start
|
Net
| false
| 4,184
|
[
"Apache-2.0"
] | 0
|
a2b321e8b5ef7b894b5e0659c5da2f9ae3df25d8
|
https://github.com/rcaborges/music-cold-start/tree/a2b321e8b5ef7b894b5e0659c5da2f9ae3df25d8
|
L2Loss
|
import torch
import torch.nn as nn
import torch.utils.data
class L2Loss(nn.Module):
"""
Compute the l2 distance
"""
def __init__(self):
super(L2Loss, self).__init__()
def forward(self, h_pred, h_target):
return torch.norm(h_target - h_pred, p=2)
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
import torch.utils.data
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_per_fused_linalg_vector_norm_sub_0(in_out_ptr0, in_ptr0, in_ptr1,
xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 256
xoffset = tl.program_id(0) * XBLOCK
tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.load(in_ptr1 + r0, None)
tmp2 = tmp0 - tmp1
tmp3 = tmp2 * tmp2
tmp4 = tl.broadcast_to(tmp3, [RBLOCK])
tmp6 = triton_helpers.promote_to_tensor(tl.sum(tmp4, 0))
tmp7 = libdevice.sqrt(tmp6)
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp7, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((), (), torch.float32)
buf1 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_linalg_vector_norm_sub_0[grid(1)](buf1, arg0_1,
arg1_1, 1, 256, num_warps=2, num_stages=1)
del arg0_1
del arg1_1
return buf1,
class L2LossNew(nn.Module):
"""
Compute the l2 distance
"""
def __init__(self):
super(L2LossNew, self).__init__()
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
|
riokt/video-paragraph
|
L2Loss
| false
| 4,185
|
[
"MIT"
] | 0
|
2da3298819e73809af495457db2cf1dfffad712f
|
https://github.com/riokt/video-paragraph/tree/2da3298819e73809af495457db2cf1dfffad712f
|
SNNBlock
|
from torch.nn import Module
import math
import torch
from torch.nn import SELU
from torch.nn import AlphaDropout
from torch.nn import Identity
from torch.nn import Parameter
from torch.nn.functional import conv2d
class SNNBlock(Module):
"""Block for a self-normalizing fully-connected layer.
This block consists of:
* AlphaDropout
* Linear
* SELU
"""
def __init__(self, in_features: 'int', out_features: 'int', dropout:
'float'=0.0, activation: 'bool'=True):
"""Initialize the layers.
Args:
in_features: The no. of input features
out_features: The no. of output features
dropout: The probability of dropping out the inputs
activation: Whether to add the activation function
"""
super().__init__()
self.dropout = AlphaDropout(dropout)
self.activation = SELU() if activation else Identity()
stddev = math.sqrt(1 / in_features)
weight = torch.randn(out_features, in_features, 1, 1) * stddev
bias = torch.zeros(out_features)
self.weight = Parameter(weight)
self.bias = Parameter(bias)
def forward(self, inputs: 'torch.Tensor') ->torch.Tensor:
"""Get the block's outputs."""
outputs = self.dropout(inputs)
outputs = conv2d(outputs, self.weight, self.bias)
return self.activation(outputs)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_features': 4, 'out_features': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
from torch.nn import Module
import math
from torch.nn import SELU
from torch.nn import AlphaDropout
from torch.nn import Identity
from torch.nn import Parameter
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_convolution_elu_0(in_out_ptr0, in_ptr0, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp5 = 1.0507009873554805
tmp6 = tmp2 * tmp5
tmp7 = 1.0
tmp8 = tmp2 * tmp7
tmp9 = libdevice.expm1(tmp8)
tmp10 = 1.7580993408473766
tmp11 = tmp9 * tmp10
tmp12 = tl.where(tmp4, tmp6, tmp11)
tl.store(in_out_ptr0 + x3, tmp2, xmask)
tl.store(out_ptr0 + x3, tmp12, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_1, primals_2, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 4, 4), (64, 16, 4, 1))
buf1 = buf0
del buf0
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_convolution_elu_0[grid(256)](buf1, primals_3, buf2,
256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_3
return buf2, primals_1, primals_2, buf1
class SNNBlockNew(Module):
"""Block for a self-normalizing fully-connected layer.
This block consists of:
* AlphaDropout
* Linear
* SELU
"""
def __init__(self, in_features: 'int', out_features: 'int', dropout:
'float'=0.0, activation: 'bool'=True):
"""Initialize the layers.
Args:
in_features: The no. of input features
out_features: The no. of output features
dropout: The probability of dropping out the inputs
activation: Whether to add the activation function
"""
super().__init__()
self.dropout = AlphaDropout(dropout)
self.activation = SELU() if activation else Identity()
stddev = math.sqrt(1 / in_features)
weight = torch.randn(out_features, in_features, 1, 1) * stddev
bias = torch.zeros(out_features)
self.weight = Parameter(weight)
self.bias = Parameter(bias)
def forward(self, input_0):
primals_2 = self.weight
primals_3 = self.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
|
rharish101/CIL-Project
|
SNNBlock
| false
| 4,186
|
[
"MIT"
] | 0
|
fed1be8b22bb4228329b719a301f74459a7bf13b
|
https://github.com/rharish101/CIL-Project/tree/fed1be8b22bb4228329b719a301f74459a7bf13b
|
FinalPool
|
import torch
import torch.utils.data
class FinalPool(torch.nn.Module):
def __init__(self):
super(FinalPool, self).__init__()
def forward(self, input):
"""
input : Tensor of shape (batch size, T, Cin)
Outputs a Tensor of shape (batch size, Cin).
"""
return input.max(dim=1)[0]
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.utils.data
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_max_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x1 = xindex // 16
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask)
tmp1 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask)
tmp3 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask)
tmp5 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask)
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp4 = triton_helpers.maximum(tmp2, tmp3)
tmp6 = triton_helpers.maximum(tmp4, tmp5)
tl.store(out_ptr0 + x2, tmp6, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_max_0[grid(64)](arg0_1, buf0, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del arg0_1
return buf0,
class FinalPoolNew(torch.nn.Module):
def __init__(self):
super(FinalPoolNew, self).__init__()
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
praesc/end-to-end-SLU
|
FinalPool
| false
| 4,187
|
[
"Apache-2.0"
] | 0
|
c4e8a5be0ea6a8d93ea7cfd3a5bdab0560c50848
|
https://github.com/praesc/end-to-end-SLU/tree/c4e8a5be0ea6a8d93ea7cfd3a5bdab0560c50848
|
CAE_ENC
|
import torch
import torch.nn as nn
import torch.nn.functional as F
class CAE_ENC(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(3, 32, kernel_size=5, padding=2, stride=2)
self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1, stride=2)
self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1, stride=2)
self.conv4 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2)
self.fc1 = nn.Linear(256 * 6 * 6, 1000)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = F.relu(self.conv3(x))
x = F.relu(self.conv4(x))
x = x.view(-1, 256 * 6 * 6)
x = self.fc1(x)
return x
def get_inputs():
return [torch.rand([4, 3, 96, 96])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 96
xnumel = 25
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 3
y1 = yindex // 3
tmp0 = tl.load(in_ptr0 + (x2 + 25 * y3), xmask & ymask, eviction_policy
='evict_last')
tl.store(out_ptr0 + (y0 + 3 * x2 + 75 * y1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 12
xnumel = 9216
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 3
y1 = yindex // 3
tmp0 = tl.load(in_ptr0 + (x2 + 9216 * y3), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + 3 * x2 + 27648 * y1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_2(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 32
y1 = yindex // 32
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 32 * x2 + 288 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_3(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 64
y1 = yindex // 64
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 64 * x2 + 576 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 128
y1 = yindex // 128
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 128 * x2 + 1152 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_5(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 32
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_relu_6(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 64
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_relu_7(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 128
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, None)
@triton.jit
def triton_poi_fused_convolution_relu_threshold_backward_8(in_ptr0, in_ptr1,
out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.
constexpr):
xnumel = 36
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 256
y1 = yindex // 256
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 256 * x2 + 9216 * y1), xmask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1, 1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(out_ptr0 + (x2 + 36 * y3), tmp4, xmask)
tl.store(out_ptr1 + (y0 + 256 * x2 + 9216 * y1), tmp6, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11) = args
args.clear()
assert_size_stride(primals_1, (32, 3, 5, 5), (75, 25, 5, 1))
assert_size_stride(primals_2, (32,), (1,))
assert_size_stride(primals_3, (4, 3, 96, 96), (27648, 9216, 96, 1))
assert_size_stride(primals_4, (64, 32, 3, 3), (288, 9, 3, 1))
assert_size_stride(primals_5, (64,), (1,))
assert_size_stride(primals_6, (128, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_7, (128,), (1,))
assert_size_stride(primals_8, (256, 128, 3, 3), (1152, 9, 3, 1))
assert_size_stride(primals_9, (256,), (1,))
assert_size_stride(primals_10, (1000, 9216), (9216, 1))
assert_size_stride(primals_11, (1000,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((32, 3, 5, 5), (75, 1, 15, 3), torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(96, 25)](primals_1, buf0, 96, 25, XBLOCK=32,
YBLOCK=32, num_warps=4, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((4, 3, 96, 96), (27648, 1, 288, 3), torch
.float32)
triton_poi_fused_1[grid(12, 9216)](primals_3, buf1, 12, 9216,
XBLOCK=64, YBLOCK=16, num_warps=4, num_stages=1)
del primals_3
buf2 = empty_strided_cuda((64, 32, 3, 3), (288, 1, 96, 32), torch.
float32)
triton_poi_fused_2[grid(2048, 9)](primals_4, buf2, 2048, 9, XBLOCK=
16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_4
buf3 = empty_strided_cuda((128, 64, 3, 3), (576, 1, 192, 64), torch
.float32)
triton_poi_fused_3[grid(8192, 9)](primals_6, buf3, 8192, 9, XBLOCK=
16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_6
buf4 = empty_strided_cuda((256, 128, 3, 3), (1152, 1, 384, 128),
torch.float32)
triton_poi_fused_4[grid(32768, 9)](primals_8, buf4, 32768, 9,
XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_8
buf5 = extern_kernels.convolution(buf1, buf0, stride=(2, 2),
padding=(2, 2), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf5, (4, 32, 48, 48), (73728, 1, 1536, 32))
buf6 = buf5
del buf5
triton_poi_fused_convolution_relu_5[grid(294912)](buf6, primals_2,
294912, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_2
buf7 = extern_kernels.convolution(buf6, buf2, stride=(2, 2),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf7, (4, 64, 24, 24), (36864, 1, 1536, 64))
buf8 = buf7
del buf7
triton_poi_fused_convolution_relu_6[grid(147456)](buf8, primals_5,
147456, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_5
buf9 = extern_kernels.convolution(buf8, buf3, stride=(2, 2),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf9, (4, 128, 12, 12), (18432, 1, 1536, 128))
buf10 = buf9
del buf9
triton_poi_fused_convolution_relu_7[grid(73728)](buf10, primals_7,
73728, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_7
buf11 = extern_kernels.convolution(buf10, buf4, stride=(2, 2),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf11, (4, 256, 6, 6), (9216, 1, 1536, 256))
buf12 = empty_strided_cuda((4, 256, 6, 6), (9216, 36, 6, 1), torch.
float32)
buf14 = empty_strided_cuda((4, 256, 6, 6), (9216, 1, 1536, 256),
torch.bool)
triton_poi_fused_convolution_relu_threshold_backward_8[grid(1024, 36)](
buf11, primals_9, buf12, buf14, 1024, 36, XBLOCK=64, YBLOCK=8,
num_warps=4, num_stages=1)
del buf11
del primals_9
buf13 = empty_strided_cuda((4, 1000), (1000, 1), torch.float32)
extern_kernels.addmm(primals_11, reinterpret_tensor(buf12, (4, 9216
), (9216, 1), 0), reinterpret_tensor(primals_10, (9216, 1000),
(1, 9216), 0), alpha=1, beta=1, out=buf13)
del primals_11
return (buf13, buf0, buf1, buf2, buf3, buf4, buf6, buf8, buf10,
reinterpret_tensor(buf12, (4, 9216), (9216, 1), 0), primals_10, buf14)
class CAE_ENCNew(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(3, 32, kernel_size=5, padding=2, stride=2)
self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1, stride=2)
self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1, stride=2)
self.conv4 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2)
self.fc1 = nn.Linear(256 * 6 * 6, 1000)
def forward(self, input_0):
primals_1 = self.conv1.weight
primals_2 = self.conv1.bias
primals_4 = self.conv2.weight
primals_5 = self.conv2.bias
primals_6 = self.conv3.weight
primals_7 = self.conv3.bias
primals_8 = self.conv4.weight
primals_9 = self.conv4.bias
primals_10 = self.fc1.weight
primals_11 = self.fc1.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11])
return output[0]
|
positivevaib/semi-supervised-imagenet-classification
|
CAE_ENC
| false
| 4,188
|
[
"MIT"
] | 0
|
4fb6427f5a72951c1b866a1ddbc2599811bb5770
|
https://github.com/positivevaib/semi-supervised-imagenet-classification/tree/4fb6427f5a72951c1b866a1ddbc2599811bb5770
|
PSA_p
|
import torch
import torch.nn as nn
import torch._utils
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
def kaiming_init(module, a=0, mode='fan_out', nonlinearity='relu', bias=0,
distribution='normal'):
assert distribution in ['uniform', 'normal']
if distribution == 'uniform':
nn.init.kaiming_uniform_(module.weight, a=a, mode=mode,
nonlinearity=nonlinearity)
else:
nn.init.kaiming_normal_(module.weight, a=a, mode=mode, nonlinearity
=nonlinearity)
if hasattr(module, 'bias') and module.bias is not None:
nn.init.constant_(module.bias, bias)
class PSA_p(nn.Module):
def __init__(self, inplanes, planes, kernel_size=1, stride=1):
super(PSA_p, self).__init__()
self.inplanes = inplanes
self.inter_planes = planes // 2
self.planes = planes
self.kernel_size = kernel_size
self.stride = stride
self.padding = (kernel_size - 1) // 2
self.conv_q_right = nn.Conv2d(self.inplanes, 1, kernel_size=1,
stride=stride, padding=0, bias=False)
self.conv_v_right = nn.Conv2d(self.inplanes, self.inter_planes,
kernel_size=1, stride=stride, padding=0, bias=False)
self.conv_up = nn.Conv2d(self.inter_planes, self.planes,
kernel_size=1, stride=1, padding=0, bias=False)
self.softmax_right = nn.Softmax(dim=2)
self.sigmoid = nn.Sigmoid()
self.conv_q_left = nn.Conv2d(self.inplanes, self.inter_planes,
kernel_size=1, stride=stride, padding=0, bias=False)
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.conv_v_left = nn.Conv2d(self.inplanes, self.inter_planes,
kernel_size=1, stride=stride, padding=0, bias=False)
self.softmax_left = nn.Softmax(dim=2)
self.reset_parameters()
def reset_parameters(self):
kaiming_init(self.conv_q_right, mode='fan_in')
kaiming_init(self.conv_v_right, mode='fan_in')
kaiming_init(self.conv_q_left, mode='fan_in')
kaiming_init(self.conv_v_left, mode='fan_in')
self.conv_q_right.inited = True
self.conv_v_right.inited = True
self.conv_q_left.inited = True
self.conv_v_left.inited = True
def spatial_pool(self, x):
input_x = self.conv_v_right(x)
batch, channel, height, width = input_x.size()
input_x = input_x.view(batch, channel, height * width)
context_mask = self.conv_q_right(x)
context_mask = context_mask.view(batch, 1, height * width)
context_mask = self.softmax_right(context_mask)
context = torch.matmul(input_x, context_mask.transpose(1, 2))
context = context.unsqueeze(-1)
context = self.conv_up(context)
mask_ch = self.sigmoid(context)
out = x * mask_ch
return out
def channel_pool(self, x):
g_x = self.conv_q_left(x)
batch, channel, height, width = g_x.size()
avg_x = self.avg_pool(g_x)
batch, channel, avg_x_h, avg_x_w = avg_x.size()
avg_x = avg_x.view(batch, channel, avg_x_h * avg_x_w).permute(0, 2, 1)
theta_x = self.conv_v_left(x).view(batch, self.inter_planes, height *
width)
context = torch.matmul(avg_x, theta_x)
context = self.softmax_left(context)
context = context.view(batch, 1, height, width)
mask_sp = self.sigmoid(context)
out = x * mask_sp
return out
def forward(self, x):
context_channel = self.spatial_pool(x)
context_spatial = self.channel_pool(x)
out = context_spatial + context_channel
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'inplanes': 4, 'planes': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
import torch._utils
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused__softmax_0(in_ptr0, out_ptr2, xnumel, rnumel, XBLOCK:
tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 16 * x0), xmask, other=0.0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp3 = tl.where(xmask, tmp1, float('-inf'))
tmp4 = triton_helpers.max2(tmp3, 1)[:, None]
tmp5 = tmp0 - tmp4
tmp6 = tl_math.exp(tmp5)
tmp7 = tl.broadcast_to(tmp6, [XBLOCK, RBLOCK])
tmp9 = tl.where(xmask, tmp7, 0)
tmp10 = tl.sum(tmp9, 1)[:, None]
tmp11 = tmp6 / tmp10
tl.store(out_ptr2 + (r1 + 16 * x0), tmp11, xmask)
@triton.jit
def triton_per_fused_mean_1(in_out_ptr0, in_ptr0, xnumel, rnumel, XBLOCK:
tl.constexpr):
xnumel = 8
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 16 * x0), xmask, other=0.0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp3 = tl.where(xmask, tmp1, 0)
tmp4 = tl.sum(tmp3, 1)[:, None]
tmp5 = 16.0
tmp6 = tmp4 / tmp5
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp6, xmask)
@triton.jit
def triton_per_fused__softmax_2(in_ptr0, out_ptr0, out_ptr1, xnumel, rnumel,
XBLOCK: tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 16 * x0), xmask, other=0.0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp3 = tl.where(xmask, tmp1, float('-inf'))
tmp4 = triton_helpers.max2(tmp3, 1)[:, None]
tmp5 = tmp0 - tmp4
tmp6 = tl_math.exp(tmp5)
tmp7 = tl.broadcast_to(tmp6, [XBLOCK, RBLOCK])
tmp9 = tl.where(xmask, tmp7, 0)
tmp10 = tl.sum(tmp9, 1)[:, None]
tl.store(out_ptr0 + x0, tmp4, xmask)
tl.store(out_ptr1 + x0, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_mul_sigmoid_3(in_ptr0, in_ptr1, in_ptr2, in_ptr3,
in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
x4 = xindex // 16
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr1 + (x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr2 + x2, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x2, xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr4 + x4, xmask, eviction_policy='evict_last')
tmp3 = tmp1 - tmp2
tmp4 = tl_math.exp(tmp3)
tmp6 = tmp4 / tmp5
tmp7 = tl.sigmoid(tmp6)
tmp8 = tmp0 * tmp7
tmp10 = tl.sigmoid(tmp9)
tmp11 = tmp0 * tmp10
tmp12 = tmp8 + tmp11
tl.store(out_ptr0 + x3, tmp12, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6 = args
args.clear()
assert_size_stride(primals_1, (2, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (1, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_4, (4, 2, 1, 1), (2, 1, 1, 1))
assert_size_stride(primals_5, (2, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_6, (2, 4, 1, 1), (4, 1, 1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_2, primals_1, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 2, 4, 4), (32, 16, 4, 1))
buf1 = extern_kernels.convolution(primals_2, primals_3, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 1, 4, 4), (16, 16, 4, 1))
buf4 = empty_strided_cuda((4, 1, 16), (16, 16, 1), torch.float32)
get_raw_stream(0)
triton_per_fused__softmax_0[grid(4)](buf1, buf4, 4, 16, XBLOCK=1,
num_warps=2, num_stages=1)
buf5 = empty_strided_cuda((4, 2, 1), (2, 1, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf0, (4, 2, 16), (32, 16, 1),
0), reinterpret_tensor(buf4, (4, 16, 1), (16, 1, 16), 0), out=buf5)
buf6 = extern_kernels.convolution(reinterpret_tensor(buf5, (4, 2, 1,
1), (2, 1, 1, 1), 0), primals_4, stride=(1, 1), padding=(0, 0),
dilation=(1, 1), transposed=False, output_padding=(0, 0),
groups=1, bias=None)
assert_size_stride(buf6, (4, 4, 1, 1), (4, 1, 1, 1))
buf7 = extern_kernels.convolution(primals_2, primals_5, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf7, (4, 2, 4, 4), (32, 16, 4, 1))
buf8 = empty_strided_cuda((4, 2, 1, 1), (2, 1, 8, 8), torch.float32)
buf10 = buf8
del buf8
triton_per_fused_mean_1[grid(8)](buf10, buf7, 8, 16, XBLOCK=8,
num_warps=2, num_stages=1)
del buf7
buf9 = extern_kernels.convolution(primals_2, primals_6, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf9, (4, 2, 4, 4), (32, 16, 4, 1))
buf11 = reinterpret_tensor(buf1, (4, 1, 16), (16, 16, 1), 0)
del buf1
extern_kernels.bmm(reinterpret_tensor(buf10, (4, 1, 2), (2, 0, 1),
0), reinterpret_tensor(buf9, (4, 2, 16), (32, 16, 1), 0), out=buf11
)
buf12 = empty_strided_cuda((4, 1, 1), (1, 1, 1), torch.float32)
buf13 = empty_strided_cuda((4, 1, 1), (1, 1, 1), torch.float32)
triton_per_fused__softmax_2[grid(4)](buf11, buf12, buf13, 4, 16,
XBLOCK=1, num_warps=2, num_stages=1)
buf14 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_add_mul_sigmoid_3[grid(256)](primals_2, buf11,
buf12, buf13, buf6, buf14, 256, XBLOCK=128, num_warps=4,
num_stages=1)
return (buf14, primals_1, primals_2, primals_3, primals_4, primals_5,
primals_6, buf4, reinterpret_tensor(buf5, (4, 2, 1, 1), (2, 1, 1, 1
), 0), buf6, buf11, buf12, buf13, reinterpret_tensor(buf10, (4, 2,
1), (2, 1, 1), 0), reinterpret_tensor(buf9, (4, 16, 2), (32, 1, 16),
0), reinterpret_tensor(buf0, (4, 16, 2), (32, 1, 16), 0))
def kaiming_init(module, a=0, mode='fan_out', nonlinearity='relu', bias=0,
distribution='normal'):
assert distribution in ['uniform', 'normal']
if distribution == 'uniform':
nn.init.kaiming_uniform_(module.weight, a=a, mode=mode,
nonlinearity=nonlinearity)
else:
nn.init.kaiming_normal_(module.weight, a=a, mode=mode, nonlinearity
=nonlinearity)
if hasattr(module, 'bias') and module.bias is not None:
nn.init.constant_(module.bias, bias)
class PSA_pNew(nn.Module):
def __init__(self, inplanes, planes, kernel_size=1, stride=1):
super(PSA_pNew, self).__init__()
self.inplanes = inplanes
self.inter_planes = planes // 2
self.planes = planes
self.kernel_size = kernel_size
self.stride = stride
self.padding = (kernel_size - 1) // 2
self.conv_q_right = nn.Conv2d(self.inplanes, 1, kernel_size=1,
stride=stride, padding=0, bias=False)
self.conv_v_right = nn.Conv2d(self.inplanes, self.inter_planes,
kernel_size=1, stride=stride, padding=0, bias=False)
self.conv_up = nn.Conv2d(self.inter_planes, self.planes,
kernel_size=1, stride=1, padding=0, bias=False)
self.softmax_right = nn.Softmax(dim=2)
self.sigmoid = nn.Sigmoid()
self.conv_q_left = nn.Conv2d(self.inplanes, self.inter_planes,
kernel_size=1, stride=stride, padding=0, bias=False)
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.conv_v_left = nn.Conv2d(self.inplanes, self.inter_planes,
kernel_size=1, stride=stride, padding=0, bias=False)
self.softmax_left = nn.Softmax(dim=2)
self.reset_parameters()
def reset_parameters(self):
kaiming_init(self.conv_q_right, mode='fan_in')
kaiming_init(self.conv_v_right, mode='fan_in')
kaiming_init(self.conv_q_left, mode='fan_in')
kaiming_init(self.conv_v_left, mode='fan_in')
self.conv_q_right.inited = True
self.conv_v_right.inited = True
self.conv_q_left.inited = True
self.conv_v_left.inited = True
def spatial_pool(self, x):
input_x = self.conv_v_right(x)
batch, channel, height, width = input_x.size()
input_x = input_x.view(batch, channel, height * width)
context_mask = self.conv_q_right(x)
context_mask = context_mask.view(batch, 1, height * width)
context_mask = self.softmax_right(context_mask)
context = torch.matmul(input_x, context_mask.transpose(1, 2))
context = context.unsqueeze(-1)
context = self.conv_up(context)
mask_ch = self.sigmoid(context)
out = x * mask_ch
return out
def channel_pool(self, x):
g_x = self.conv_q_left(x)
batch, channel, height, width = g_x.size()
avg_x = self.avg_pool(g_x)
batch, channel, avg_x_h, avg_x_w = avg_x.size()
avg_x = avg_x.view(batch, channel, avg_x_h * avg_x_w).permute(0, 2, 1)
theta_x = self.conv_v_left(x).view(batch, self.inter_planes, height *
width)
context = torch.matmul(avg_x, theta_x)
context = self.softmax_left(context)
context = context.view(batch, 1, height, width)
mask_sp = self.sigmoid(context)
out = x * mask_sp
return out
def forward(self, input_0):
primals_3 = self.conv_q_right.weight
primals_1 = self.conv_v_right.weight
primals_4 = self.conv_up.weight
primals_5 = self.conv_q_left.weight
primals_6 = self.conv_v_left.weight
primals_2 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6])
return output[0]
|
realphongha/human-pose-estimation.pytorch
|
PSA_p
| false
| 4,189
|
[
"MIT"
] | 0
|
29b106d3e6c6e12325a7d4bca4abc56ecbc12b1f
|
https://github.com/realphongha/human-pose-estimation.pytorch/tree/29b106d3e6c6e12325a7d4bca4abc56ecbc12b1f
|
ContrastiveLoss
|
from torch.nn import Module
import torch
from torch.nn import LogSoftmax
from torch.nn.functional import cosine_similarity
class ContrastiveLoss(Module):
"""A contrastive loss adapted from SimCLR.
Link to SimCLR: https://arxiv.org/abs/2002.05709v3.
"""
def __init__(self, temperature: 'float'=1.0):
"""Save hyper-params."""
super().__init__()
self.temperature = temperature
self._log_softmax_fn = LogSoftmax(dim=-1)
def forward(self, inputs: 'torch.Tensor', targets: 'torch.Tensor'
) ->torch.Tensor:
"""Get the loss."""
inputs = inputs.permute(2, 3, 0, 1)
targets = targets.permute(2, 3, 0, 1)
batch_size = inputs.shape[-2]
left = torch.cat([inputs, targets], -2).unsqueeze(-1)
right = left.permute(0, 1, 4, 3, 2)
similarity = cosine_similarity(left, right, dim=-2, eps=torch.finfo
(left.dtype).eps)
mask = torch.eye(2 * batch_size, device=similarity.device).bool()
mask_nd = mask.unsqueeze(0).unsqueeze(0).tile(*similarity.shape[:2],
1, 1)
neg_inf = float('-inf') * torch.ones_like(similarity)
similarity = torch.where(mask_nd, neg_inf, similarity)
log_softmax = self._log_softmax_fn(similarity / self.temperature)
positive_pairs = torch.cat([torch.diagonal(log_softmax, offset=
batch_size, dim1=-2, dim2=-1), torch.diagonal(log_softmax,
offset=-batch_size, dim1=-2, dim2=-1)], -1)
return -positive_pairs.mean()
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch.nn import Module
from torch.nn import LogSoftmax
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_linalg_vector_norm_0(in_ptr0, in_ptr1, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 8 % 8
x2 = xindex // 64
x3 = xindex
tmp0 = x1
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x2 + 64 * x1), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (x2 + 64 * (-4 + x1)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tmp11 = tmp10 * tmp10
tmp12 = tl.load(in_ptr0 + (16 + x2 + 64 * x1), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp13 = tl.load(in_ptr1 + (16 + x2 + 64 * (-4 + x1)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp14 = tl.where(tmp4, tmp12, tmp13)
tmp15 = tmp14 * tmp14
tmp16 = tmp11 + tmp15
tmp17 = tl.load(in_ptr0 + (32 + x2 + 64 * x1), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp18 = tl.load(in_ptr1 + (32 + x2 + 64 * (-4 + x1)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp19 = tl.where(tmp4, tmp17, tmp18)
tmp20 = tmp19 * tmp19
tmp21 = tmp16 + tmp20
tmp22 = tl.load(in_ptr0 + (48 + x2 + 64 * x1), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp23 = tl.load(in_ptr1 + (48 + x2 + 64 * (-4 + x1)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp24 = tl.where(tmp4, tmp22, tmp23)
tmp25 = tmp24 * tmp24
tmp26 = tmp21 + tmp25
tl.store(out_ptr0 + x3, tmp26, xmask)
@triton.jit
def triton_poi_fused_linalg_vector_norm_1(in_ptr0, in_ptr1, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x2 = xindex // 64
x3 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x2 + 64 * x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (x2 + 64 * (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tmp11 = tmp10 * tmp10
tmp12 = tl.load(in_ptr0 + (16 + x2 + 64 * x0), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp13 = tl.load(in_ptr1 + (16 + x2 + 64 * (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp14 = tl.where(tmp4, tmp12, tmp13)
tmp15 = tmp14 * tmp14
tmp16 = tmp11 + tmp15
tmp17 = tl.load(in_ptr0 + (32 + x2 + 64 * x0), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp18 = tl.load(in_ptr1 + (32 + x2 + 64 * (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp19 = tl.where(tmp4, tmp17, tmp18)
tmp20 = tmp19 * tmp19
tmp21 = tmp16 + tmp20
tmp22 = tl.load(in_ptr0 + (48 + x2 + 64 * x0), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp23 = tl.load(in_ptr1 + (48 + x2 + 64 * (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp24 = tl.where(tmp4, tmp22, tmp23)
tmp25 = tmp24 * tmp24
tmp26 = tmp21 + tmp25
tl.store(out_ptr0 + x3, tmp26, xmask)
@triton.jit
def triton_poi_fused_clamp_min_div_linalg_vector_norm_mul_2(in_ptr0,
in_ptr1, in_ptr2, in_ptr3, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex // 32 % 8
x1 = xindex // 8 % 4
x3 = xindex // 256
x0 = xindex % 8
x4 = xindex // 32
x5 = xindex
tmp11 = tl.load(in_ptr2 + (x0 + 8 * x4), None, eviction_policy='evict_last'
)
tmp24 = tl.load(in_ptr3 + (x0 + 8 * x4), None, eviction_policy='evict_last'
)
tmp0 = x2
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x3 + 16 * x1 + 64 * x2), tmp4,
eviction_policy='evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (x3 + 16 * x1 + 64 * (-4 + x2)), tmp6,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tmp12 = libdevice.sqrt(tmp11)
tmp13 = 1.1920928955078125e-07
tmp14 = triton_helpers.maximum(tmp12, tmp13)
tmp15 = tmp10 / tmp14
tmp16 = x0
tmp18 = tmp16 < tmp3
tmp19 = tl.load(in_ptr0 + (x3 + 16 * x1 + 64 * x0), tmp18,
eviction_policy='evict_last', other=0.0)
tmp20 = tmp16 >= tmp3
tmp22 = tl.load(in_ptr1 + (x3 + 16 * x1 + 64 * (-4 + x0)), tmp20,
eviction_policy='evict_last', other=0.0)
tmp23 = tl.where(tmp18, tmp19, tmp22)
tmp25 = libdevice.sqrt(tmp24)
tmp26 = triton_helpers.maximum(tmp25, tmp13)
tmp27 = tmp23 / tmp26
tmp28 = tmp15 * tmp27
tl.store(out_ptr0 + x5, tmp28, None)
@triton.jit
def triton_per_fused__log_softmax_mul_repeat_sum_where_3(in_ptr0, out_ptr1,
out_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 128
RBLOCK: tl.constexpr = 8
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
x0 = xindex % 8
r2 = rindex
x3 = xindex
tmp7 = tl.load(in_ptr0 + (r2 + 32 * x3), xmask, other=0.0)
tmp8 = tl.load(in_ptr0 + (8 + r2 + 32 * x3), xmask, other=0.0)
tmp10 = tl.load(in_ptr0 + (16 + r2 + 32 * x3), xmask, other=0.0)
tmp12 = tl.load(in_ptr0 + (24 + r2 + 32 * x3), xmask, other=0.0)
tmp0 = x0
tmp1 = r2
tmp2 = tmp0 == tmp1
tmp3 = 1.0
tmp4 = 0.0
tmp5 = tl.where(tmp2, tmp3, tmp4)
tmp6 = tmp5 != 0
tmp9 = tmp7 + tmp8
tmp11 = tmp9 + tmp10
tmp13 = tmp11 + tmp12
tmp14 = float('-inf')
tmp15 = tl.where(tmp6, tmp14, tmp13)
tmp16 = tmp15 * tmp3
tmp17 = tl.broadcast_to(tmp16, [XBLOCK, RBLOCK])
tmp19 = tl.where(xmask, tmp17, float('-inf'))
tmp20 = triton_helpers.max2(tmp19, 1)[:, None]
tmp21 = tmp16 - tmp20
tmp22 = tmp21 * tmp3
tmp23 = tl_math.exp(tmp22)
tmp24 = tl.broadcast_to(tmp23, [XBLOCK, RBLOCK])
tmp26 = tl.where(xmask, tmp24, 0)
tmp27 = tl.sum(tmp26, 1)[:, None]
tl.store(out_ptr1 + (r2 + 8 * x3), tmp22, xmask)
tl.store(out_ptr2 + x3, tmp27, xmask)
@triton.jit
def triton_poi_fused_cat_4(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + (4 + 9 * x0 + 64 * x1), xmask, eviction_policy
='evict_last')
tmp1 = tl.load(in_ptr1 + (x0 + 8 * x1), xmask)
tmp2 = tl_math.log(tmp1)
tmp3 = tmp0 - tmp2
tl.store(out_ptr0 + (x0 + 8 * x1), tmp3, xmask)
@triton.jit
def triton_poi_fused_cat_5(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + (32 + 9 * x0 + 64 * x1), xmask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (4 + x0 + 8 * x1), xmask)
tmp2 = tl_math.log(tmp1)
tmp3 = tmp0 - tmp2
tl.store(out_ptr0 + (x0 + 8 * x1), tmp3, xmask)
@triton.jit
def triton_per_fused_mean_neg_6(in_out_ptr0, in_ptr0, xnumel, rnumel,
XBLOCK: tl.constexpr):
RBLOCK: tl.constexpr = 128
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp3 = tl.sum(tmp1, 1)[:, None]
tmp4 = 128.0
tmp5 = tmp3 / tmp4
tmp6 = -tmp5
tl.debug_barrier()
tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp6, None)
def call(args):
arg0_1, arg1_1 = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(arg1_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 8, 1, 8), (256, 64, 8, 1024, 1),
torch.float32)
get_raw_stream(0)
triton_poi_fused_linalg_vector_norm_0[grid(1024)](arg0_1, arg1_1,
buf0, 1024, XBLOCK=256, num_warps=4, num_stages=1)
buf1 = empty_strided_cuda((4, 4, 8, 1, 8), (256, 64, 8, 1024, 1),
torch.float32)
triton_poi_fused_linalg_vector_norm_1[grid(1024)](arg0_1, arg1_1,
buf1, 1024, XBLOCK=128, num_warps=4, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 8, 4, 8), (1024, 256, 32, 8, 1),
torch.float32)
triton_poi_fused_clamp_min_div_linalg_vector_norm_mul_2[grid(4096)](
arg0_1, arg1_1, buf0, buf1, buf2, 4096, XBLOCK=128, num_warps=4,
num_stages=1)
del arg0_1
del arg1_1
del buf0
buf4 = reinterpret_tensor(buf1, (4, 4, 8, 8), (256, 64, 8, 1), 0)
del buf1
buf5 = empty_strided_cuda((4, 4, 8, 1), (32, 8, 1, 128), torch.float32)
triton_per_fused__log_softmax_mul_repeat_sum_where_3[grid(128)](buf2,
buf4, buf5, 128, 8, XBLOCK=8, num_warps=2, num_stages=1)
del buf2
buf8 = empty_strided_cuda((4, 4, 8), (32, 8, 1), torch.float32)
buf6 = reinterpret_tensor(buf8, (4, 4, 4), (32, 8, 1), 0)
triton_poi_fused_cat_4[grid(64)](buf4, buf5, buf6, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf7 = reinterpret_tensor(buf8, (4, 4, 4), (32, 8, 1), 4)
triton_poi_fused_cat_5[grid(64)](buf4, buf5, buf7, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf4
del buf5
buf9 = empty_strided_cuda((), (), torch.float32)
buf10 = buf9
del buf9
triton_per_fused_mean_neg_6[grid(1)](buf10, buf8, 1, 128, XBLOCK=1,
num_warps=2, num_stages=1)
del buf6
del buf7
del buf8
return buf10,
class ContrastiveLossNew(Module):
"""A contrastive loss adapted from SimCLR.
Link to SimCLR: https://arxiv.org/abs/2002.05709v3.
"""
def __init__(self, temperature: 'float'=1.0):
"""Save hyper-params."""
super().__init__()
self.temperature = temperature
self._log_softmax_fn = LogSoftmax(dim=-1)
def forward(self, input_0, input_1):
arg0_1 = input_0
arg1_1 = input_1
output = call([arg0_1, arg1_1])
return output[0]
|
rharish101/CIL-Project
|
ContrastiveLoss
| false
| 4,190
|
[
"MIT"
] | 0
|
fed1be8b22bb4228329b719a301f74459a7bf13b
|
https://github.com/rharish101/CIL-Project/tree/fed1be8b22bb4228329b719a301f74459a7bf13b
|
FilterNorm
|
import torch
import torch.nn as nn
from torch.nn.init import calculate_gain
import torch.nn.parallel
class FilterNorm(nn.Module):
def __init__(self, in_channels, kernel_size, filter_type, nonlinearity=
'linear', running_std=False, running_mean=False):
assert filter_type in ('spatial', 'channel')
assert in_channels >= 1
super(FilterNorm, self).__init__()
self.in_channels = in_channels
self.filter_type = filter_type
self.runing_std = running_std
self.runing_mean = running_mean
std = calculate_gain(nonlinearity) / kernel_size
if running_std:
self.std = nn.Parameter(torch.randn(in_channels * kernel_size **
2) * std, requires_grad=True)
else:
self.std = std
if running_mean:
self.mean = nn.Parameter(torch.randn(in_channels * kernel_size **
2), requires_grad=True)
def forward(self, x):
if self.filter_type == 'spatial':
b, _, h, w = x.size()
x = x.reshape(b, self.in_channels, -1, h, w)
x = x - x.mean(dim=2).reshape(b, self.in_channels, 1, h, w)
x = x / (x.std(dim=2).reshape(b, self.in_channels, 1, h, w) + 1e-10
)
x = x.reshape(b, _, h, w)
if self.runing_std:
x = x * self.std[None, :, None, None]
else:
x = x * self.std
if self.runing_mean:
x = x + self.mean[None, :, None, None]
elif self.filter_type == 'channel':
b = x.size(0)
c = self.in_channels
x = x.reshape(b, c, -1)
x = x - x.mean(dim=2).reshape(b, c, 1)
x = x / (x.std(dim=2).reshape(b, c, 1) + 1e-10)
x = x.reshape(b, -1)
if self.runing_std:
x = x * self.std[None, :]
else:
x = x * self.std
if self.runing_mean:
x = x + self.mean[None, :]
else:
raise RuntimeError('Unsupported filter type {}'.format(self.
filter_type))
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'kernel_size': 4, 'filter_type': 'spatial'}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
from torch.nn.init import calculate_gain
import torch.nn.parallel
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 1.0
tmp2 = tmp0 / tmp1
tmp3 = tmp0 - tmp2
tmp4 = tmp3 / tmp1
tmp5 = tmp3 - tmp4
tmp6 = tmp5 * tmp5
tmp7 = 0.0
tmp8 = tmp6 / tmp7
tmp9 = libdevice.sqrt(tmp8)
tmp10 = 1e-10
tmp11 = tmp9 + tmp10
tmp12 = tmp3 / tmp11
tmp13 = 0.25
tmp14 = tmp12 * tmp13
tl.store(out_ptr0 + x0, tmp14, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del arg0_1
return buf0,
class FilterNormNew(nn.Module):
def __init__(self, in_channels, kernel_size, filter_type, nonlinearity=
'linear', running_std=False, running_mean=False):
assert filter_type in ('spatial', 'channel')
assert in_channels >= 1
super(FilterNormNew, self).__init__()
self.in_channels = in_channels
self.filter_type = filter_type
self.runing_std = running_std
self.runing_mean = running_mean
std = calculate_gain(nonlinearity) / kernel_size
if running_std:
self.std = nn.Parameter(torch.randn(in_channels * kernel_size **
2) * std, requires_grad=True)
else:
self.std = std
if running_mean:
self.mean = nn.Parameter(torch.randn(in_channels * kernel_size **
2), requires_grad=True)
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
rightchose/ddfnet
|
FilterNorm
| false
| 4,191
|
[
"MIT"
] | 0
|
44a2f63933c1784a53f26a10c1157a164d044485
|
https://github.com/rightchose/ddfnet/tree/44a2f63933c1784a53f26a10c1157a164d044485
|
Actor
|
import torch
import numpy as np
import torch.nn.functional as F
import torch.nn as nn
def hidden_init(layer):
fan_in = layer.weight.data.size()[0]
lim = 1.0 / np.sqrt(fan_in)
return -lim, lim
class Actor(nn.Module):
"""Actor (Policy) Model."""
def __init__(self, state_size, action_size, seed, fc1_units=350,
fc2_units=300):
"""Initialize parameters and build model.
Params
======
state_size (int): Dimension of each state
action_size (int): Dimension of each action
seed (int): Random seed
fc1_units (int): Number of nodes in first hidden layer
fc2_units (int): Number of nodes in second hidden layer
"""
super(Actor, self).__init__()
self.seed = torch.manual_seed(seed)
self.fc1 = nn.Linear(state_size, fc1_units)
self.fc2 = nn.Linear(fc1_units, fc2_units)
self.fc3 = nn.Linear(fc2_units, action_size)
self.reset_parameters()
def reset_parameters(self):
self.fc1.weight.data.uniform_(*hidden_init(self.fc1))
self.fc2.weight.data.uniform_(*hidden_init(self.fc2))
self.fc3.weight.data.uniform_(-0.003, 0.003)
def forward(self, state):
"""Build an actor (policy) network that maps states -> actions."""
x = F.relu(self.fc1(state))
x = F.relu(self.fc2(x))
return torch.tanh(self.fc3(x))
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'state_size': 4, 'action_size': 4, 'seed': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice
import numpy as np
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_relu_threshold_backward_0(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 22400
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 350
x2 = xindex // 1400
x3 = xindex % 1400
tmp0 = tl.load(in_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(out_ptr0 + (x3 + 1408 * x2), tmp4, xmask)
tl.store(out_ptr1 + (x3 + 1408 * x2), tmp6, xmask)
@triton.jit
def triton_poi_fused_relu_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 22400
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 350
x1 = xindex // 350
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 350 * (x1 % 4) + 1408 * (x1 // 4)), xmask)
tl.store(out_ptr0 + x2, tmp0, xmask)
@triton.jit
def triton_poi_fused_relu_threshold_backward_2(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 19200
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 300
x2 = xindex // 1200
x3 = xindex % 1200
tmp0 = tl.load(in_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(out_ptr0 + (x3 + 1216 * x2), tmp4, xmask)
tl.store(out_ptr1 + (x3 + 1280 * x2), tmp6, xmask)
@triton.jit
def triton_poi_fused_relu_view_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 19200
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 300
x1 = xindex // 300
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 300 * (x1 % 4) + 1216 * (x1 // 4)), xmask)
tl.store(out_ptr0 + x2, tmp0, xmask)
@triton.jit
def triton_poi_fused_tanh_4(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = libdevice.tanh(tmp2)
tl.store(in_out_ptr0 + x2, tmp3, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (350, 4), (4, 1))
assert_size_stride(primals_2, (350,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (300, 350), (350, 1))
assert_size_stride(primals_5, (300,), (1,))
assert_size_stride(primals_6, (4, 300), (300, 1))
assert_size_stride(primals_7, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 350), (350, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 350), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((4, 4, 4, 350), (5632, 1408, 350, 1),
torch.float32)
buf9 = empty_strided_cuda((4, 4, 4, 350), (5632, 1408, 350, 1),
torch.bool)
get_raw_stream(0)
triton_poi_fused_relu_threshold_backward_0[grid(22400)](buf0,
primals_2, buf1, buf9, 22400, XBLOCK=256, num_warps=4, num_stages=1
)
del primals_2
buf2 = buf0
del buf0
triton_poi_fused_relu_view_1[grid(22400)](buf1, buf2, 22400, XBLOCK
=256, num_warps=4, num_stages=1)
del buf1
buf3 = empty_strided_cuda((64, 300), (300, 1), torch.float32)
extern_kernels.mm(buf2, reinterpret_tensor(primals_4, (350, 300), (
1, 350), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 300), (4864, 1216, 300, 1),
torch.float32)
buf8 = empty_strided_cuda((4, 4, 4, 300), (5120, 1280, 300, 1),
torch.bool)
triton_poi_fused_relu_threshold_backward_2[grid(19200)](buf3,
primals_5, buf4, buf8, 19200, XBLOCK=256, num_warps=4, num_stages=1
)
del primals_5
buf5 = buf3
del buf3
triton_poi_fused_relu_view_3[grid(19200)](buf4, buf5, 19200, XBLOCK
=256, num_warps=4, num_stages=1)
del buf4
buf6 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(buf5, reinterpret_tensor(primals_6, (300, 4), (1,
300), 0), out=buf6)
buf7 = reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf6
triton_poi_fused_tanh_4[grid(256)](buf7, primals_7, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_7
return buf7, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf2, buf5, buf7, primals_6, buf8, primals_4, buf9
def hidden_init(layer):
fan_in = layer.weight.data.size()[0]
lim = 1.0 / np.sqrt(fan_in)
return -lim, lim
class ActorNew(nn.Module):
"""Actor (Policy) Model."""
def __init__(self, state_size, action_size, seed, fc1_units=350,
fc2_units=300):
"""Initialize parameters and build model.
Params
======
state_size (int): Dimension of each state
action_size (int): Dimension of each action
seed (int): Random seed
fc1_units (int): Number of nodes in first hidden layer
fc2_units (int): Number of nodes in second hidden layer
"""
super(ActorNew, self).__init__()
self.seed = torch.manual_seed(seed)
self.fc1 = nn.Linear(state_size, fc1_units)
self.fc2 = nn.Linear(fc1_units, fc2_units)
self.fc3 = nn.Linear(fc2_units, action_size)
self.reset_parameters()
def reset_parameters(self):
self.fc1.weight.data.uniform_(*hidden_init(self.fc1))
self.fc2.weight.data.uniform_(*hidden_init(self.fc2))
self.fc3.weight.data.uniform_(-0.003, 0.003)
def forward(self, input_0):
primals_1 = self.fc1.weight
primals_2 = self.fc1.bias
primals_4 = self.fc2.weight
primals_5 = self.fc2.bias
primals_6 = self.fc3.weight
primals_7 = self.fc3.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
|
ricklentz/deep-reinforcement-learning
|
Actor
| false
| 4,192
|
[
"MIT"
] | 0
|
4a034a955c64a630e0fd72f4380d81e2c25a4c68
|
https://github.com/ricklentz/deep-reinforcement-learning/tree/4a034a955c64a630e0fd72f4380d81e2c25a4c68
|
TransformerLayer
|
import math
import torch
import uuid
from torch import Tensor
from typing import Tuple
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional
from typing import Dict
from torch.nn import Parameter
def gelu(x):
"""Implementation of the gelu activation function.
For information: OpenAI GPT's gelu is slightly different
(and gives slightly different results):
0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
"""
return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
def utils_softmax(x, dim: 'int', onnx_trace: 'bool'=False):
if onnx_trace:
return F.softmax(x.float(), dim=dim)
else:
return F.softmax(x, dim=dim, dtype=torch.float32)
def with_incremental_state(cls):
cls.__bases__ = (FairseqIncrementalState,) + tuple(b for b in cls.
__bases__ if b != FairseqIncrementalState)
return cls
class ESM1LayerNorm(nn.Module):
def __init__(self, hidden_size, eps=1e-12, affine=True):
"""Construct a layernorm layer in the TF style (eps inside the sqrt)."""
super().__init__()
self.hidden_size = (hidden_size,) if isinstance(hidden_size, int
) else tuple(hidden_size)
self.eps = eps
self.affine = bool(affine)
if self.affine:
self.weight = nn.Parameter(torch.ones(hidden_size))
self.bias = nn.Parameter(torch.zeros(hidden_size))
else:
self.weight, self.bias = None, None
def forward(self, x):
dims = tuple(-(i + 1) for i in range(len(self.hidden_size)))
means = x.mean(dims, keepdim=True)
x_zeromean = x - means
variances = x_zeromean.pow(2).mean(dims, keepdim=True)
x = x_zeromean / torch.sqrt(variances + self.eps)
if self.affine:
x = self.weight * x + self.bias
return x
class FairseqIncrementalState(object):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.init_incremental_state()
def init_incremental_state(self):
self._incremental_state_id = str(uuid.uuid4())
def _get_full_incremental_state_key(self, key: 'str') ->str:
return '{}.{}'.format(self._incremental_state_id, key)
def get_incremental_state(self, incremental_state:
'Optional[Dict[str, Dict[str, Optional[Tensor]]]]', key: 'str'
) ->Optional[Dict[str, Optional[Tensor]]]:
"""Helper for getting incremental state for an nn.Module."""
full_key = self._get_full_incremental_state_key(key)
if incremental_state is None or full_key not in incremental_state:
return None
return incremental_state[full_key]
def set_incremental_state(self, incremental_state:
'Optional[Dict[str, Dict[str, Optional[Tensor]]]]', key: 'str',
value: 'Dict[str, Optional[Tensor]]') ->Optional[Dict[str, Dict[str,
Optional[Tensor]]]]:
"""Helper for setting incremental state for an nn.Module."""
if incremental_state is not None:
full_key = self._get_full_incremental_state_key(key)
incremental_state[full_key] = value
return incremental_state
@with_incremental_state
class MultiheadAttention(nn.Module):
"""Multi-headed attention.
See "Attention Is All You Need" for more details.
"""
def __init__(self, embed_dim, num_heads, kdim=None, vdim=None, dropout=
0.0, bias=True, add_bias_kv=False, add_zero_attn=False,
self_attention=False, encoder_decoder_attention=False):
super().__init__()
self.embed_dim = embed_dim
self.kdim = kdim if kdim is not None else embed_dim
self.vdim = vdim if vdim is not None else embed_dim
self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
assert self.head_dim * num_heads == self.embed_dim, 'embed_dim must be divisible by num_heads'
self.scaling = self.head_dim ** -0.5
self.self_attention = self_attention
self.encoder_decoder_attention = encoder_decoder_attention
assert not self.self_attention or self.qkv_same_dim, 'Self-attention requires query, key and value to be of the same size'
self.k_proj = nn.Linear(self.kdim, embed_dim, bias=bias)
self.v_proj = nn.Linear(self.vdim, embed_dim, bias=bias)
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
if add_bias_kv:
self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))
self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))
else:
self.bias_k = self.bias_v = None
self.add_zero_attn = add_zero_attn
self.reset_parameters()
self.onnx_trace = False
self.enable_torch_version = False
if hasattr(F, 'multi_head_attention_forward'):
self.enable_torch_version = True
else:
self.enable_torch_version = False
def prepare_for_onnx_export_(self):
self.onnx_trace = True
def reset_parameters(self):
if self.qkv_same_dim:
nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))
nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))
nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))
else:
nn.init.xavier_uniform_(self.k_proj.weight)
nn.init.xavier_uniform_(self.v_proj.weight)
nn.init.xavier_uniform_(self.q_proj.weight)
nn.init.xavier_uniform_(self.out_proj.weight)
if self.out_proj.bias is not None:
nn.init.constant_(self.out_proj.bias, 0.0)
if self.bias_k is not None:
nn.init.xavier_normal_(self.bias_k)
if self.bias_v is not None:
nn.init.xavier_normal_(self.bias_v)
def forward(self, query, key: 'Optional[Tensor]', value:
'Optional[Tensor]', key_padding_mask: 'Optional[Tensor]'=None,
incremental_state:
'Optional[Dict[str, Dict[str, Optional[Tensor]]]]'=None,
need_weights: 'bool'=True, static_kv: 'bool'=False, attn_mask:
'Optional[Tensor]'=None, before_softmax: 'bool'=False,
need_head_weights: 'bool'=False) ->Tuple[Tensor, Optional[Tensor]]:
"""Input shape: Time x Batch x Channel
Args:
key_padding_mask (ByteTensor, optional): mask to exclude
keys that are pads, of shape `(batch, src_len)`, where
padding elements are indicated by 1s.
need_weights (bool, optional): return the attention weights,
averaged over heads (default: False).
attn_mask (ByteTensor, optional): typically used to
implement causal attention, where the mask prevents the
attention from looking forward in time (default: None).
before_softmax (bool, optional): return the raw attention
weights and values before the attention softmax.
need_head_weights (bool, optional): return the attention
weights for each head. Implies *need_weights*. Default:
return the average attention weights over all heads.
"""
if need_head_weights:
need_weights = True
tgt_len, bsz, embed_dim = query.size()
assert embed_dim == self.embed_dim
assert list(query.size()) == [tgt_len, bsz, embed_dim]
if (self.enable_torch_version and not self.onnx_trace and
incremental_state is None and not static_kv and not torch.jit.
is_scripting() and not need_head_weights):
assert key is not None and value is not None
return F.multi_head_attention_forward(query, key, value, self.
embed_dim, self.num_heads, torch.empty([0]), torch.cat((
self.q_proj.bias, self.k_proj.bias, self.v_proj.bias)),
self.bias_k, self.bias_v, self.add_zero_attn, self.dropout,
self.out_proj.weight, self.out_proj.bias, self.training,
key_padding_mask, need_weights, attn_mask,
use_separate_proj_weight=True, q_proj_weight=self.q_proj.
weight, k_proj_weight=self.k_proj.weight, v_proj_weight=
self.v_proj.weight)
if incremental_state is not None:
saved_state = self._get_input_buffer(incremental_state)
if saved_state is not None and 'prev_key' in saved_state:
if static_kv:
assert self.encoder_decoder_attention and not self.self_attention
key = value = None
else:
saved_state = None
if self.self_attention:
q = self.q_proj(query)
k = self.k_proj(query)
v = self.v_proj(query)
elif self.encoder_decoder_attention:
q = self.q_proj(query)
if key is None:
assert value is None
k = v = None
else:
k = self.k_proj(key)
v = self.v_proj(key)
else:
assert key is not None and value is not None
q = self.q_proj(query)
k = self.k_proj(key)
v = self.v_proj(value)
q *= self.scaling
if self.bias_k is not None:
assert self.bias_v is not None
k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])
v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])
if attn_mask is not None:
attn_mask = torch.cat([attn_mask, attn_mask.new_zeros(
attn_mask.size(0), 1)], dim=1)
if key_padding_mask is not None:
key_padding_mask = torch.cat([key_padding_mask,
key_padding_mask.new_zeros(key_padding_mask.size(0), 1)
], dim=1)
q = q.contiguous().view(tgt_len, bsz * self.num_heads, self.head_dim
).transpose(0, 1)
if k is not None:
k = k.contiguous().view(-1, bsz * self.num_heads, self.head_dim
).transpose(0, 1)
if v is not None:
v = v.contiguous().view(-1, bsz * self.num_heads, self.head_dim
).transpose(0, 1)
if saved_state is not None:
if 'prev_key' in saved_state:
_prev_key = saved_state['prev_key']
assert _prev_key is not None
prev_key = _prev_key.view(bsz * self.num_heads, -1, self.
head_dim)
if static_kv:
k = prev_key
else:
assert k is not None
k = torch.cat([prev_key, k], dim=1)
if 'prev_value' in saved_state:
_prev_value = saved_state['prev_value']
assert _prev_value is not None
prev_value = _prev_value.view(bsz * self.num_heads, -1,
self.head_dim)
if static_kv:
v = prev_value
else:
assert v is not None
v = torch.cat([prev_value, v], dim=1)
prev_key_padding_mask: 'Optional[Tensor]' = None
if 'prev_key_padding_mask' in saved_state:
prev_key_padding_mask = saved_state['prev_key_padding_mask']
assert k is not None and v is not None
key_padding_mask = (MultiheadAttention.
_append_prev_key_padding_mask(key_padding_mask=
key_padding_mask, prev_key_padding_mask=
prev_key_padding_mask, batch_size=bsz, src_len=k.size(1),
static_kv=static_kv))
saved_state['prev_key'] = k.view(bsz, self.num_heads, -1, self.
head_dim)
saved_state['prev_value'] = v.view(bsz, self.num_heads, -1,
self.head_dim)
saved_state['prev_key_padding_mask'] = key_padding_mask
assert incremental_state is not None
incremental_state = self._set_input_buffer(incremental_state,
saved_state)
assert k is not None
src_len = k.size(1)
if key_padding_mask is not None and key_padding_mask.dim() == 0:
key_padding_mask = None
if key_padding_mask is not None:
assert key_padding_mask.size(0) == bsz
assert key_padding_mask.size(1) == src_len
if self.add_zero_attn:
assert v is not None
src_len += 1
k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])],
dim=1)
v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])],
dim=1)
if attn_mask is not None:
attn_mask = torch.cat([attn_mask, attn_mask.new_zeros(
attn_mask.size(0), 1)], dim=1)
if key_padding_mask is not None:
key_padding_mask = torch.cat([key_padding_mask, torch.zeros
(key_padding_mask.size(0), 1).type_as(key_padding_mask)
], dim=1)
attn_weights = torch.bmm(q, k.transpose(1, 2))
attn_weights = MultiheadAttention.apply_sparse_mask(attn_weights,
tgt_len, src_len, bsz)
assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len,
src_len]
if attn_mask is not None:
attn_mask = attn_mask.unsqueeze(0)
if self.onnx_trace:
attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1)
attn_weights += attn_mask
if key_padding_mask is not None:
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len,
src_len)
attn_weights = attn_weights.masked_fill(key_padding_mask.
unsqueeze(1).unsqueeze(2), float('-inf'))
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len,
src_len)
if before_softmax:
return attn_weights, v
attn_weights_float = utils_softmax(attn_weights, dim=-1, onnx_trace
=self.onnx_trace)
attn_weights = attn_weights_float.type_as(attn_weights)
attn_probs = F.dropout(attn_weights_float.type_as(attn_weights), p=
self.dropout, training=self.training)
assert v is not None
attn = torch.bmm(attn_probs, v)
assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.
head_dim]
if self.onnx_trace and attn.size(1) == 1:
attn = attn.contiguous().view(tgt_len, bsz, embed_dim)
else:
attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz,
embed_dim)
attn = self.out_proj(attn)
attn_weights: 'Optional[Tensor]' = None
if need_weights:
attn_weights = attn_weights_float.view(bsz, self.num_heads,
tgt_len, src_len).transpose(1, 0)
if not need_head_weights:
attn_weights = attn_weights.mean(dim=0)
return attn, attn_weights
@staticmethod
def _append_prev_key_padding_mask(key_padding_mask: 'Optional[Tensor]',
prev_key_padding_mask: 'Optional[Tensor]', batch_size: 'int',
src_len: 'int', static_kv: 'bool') ->Optional[Tensor]:
if prev_key_padding_mask is not None and static_kv:
new_key_padding_mask = prev_key_padding_mask
elif prev_key_padding_mask is not None and key_padding_mask is not None:
new_key_padding_mask = torch.cat([prev_key_padding_mask.float(),
key_padding_mask.float()], dim=1)
elif prev_key_padding_mask is not None:
filler = torch.zeros((batch_size, src_len -
prev_key_padding_mask.size(1)), device=
prev_key_padding_mask.device)
new_key_padding_mask = torch.cat([prev_key_padding_mask.float(),
filler.float()], dim=1)
elif key_padding_mask is not None:
filler = torch.zeros((batch_size, src_len - key_padding_mask.
size(1)), device=key_padding_mask.device)
new_key_padding_mask = torch.cat([filler.float(),
key_padding_mask.float()], dim=1)
else:
new_key_padding_mask = prev_key_padding_mask
return new_key_padding_mask
@torch.jit.export
def reorder_incremental_state(self, incremental_state:
'Dict[str, Dict[str, Optional[Tensor]]]', new_order: 'Tensor'):
"""Reorder buffered internal state (for incremental generation)."""
input_buffer = self._get_input_buffer(incremental_state)
if input_buffer is not None:
for k in input_buffer.keys():
input_buffer_k = input_buffer[k]
if input_buffer_k is not None:
if self.encoder_decoder_attention and input_buffer_k.size(0
) == new_order.size(0):
break
input_buffer[k] = input_buffer_k.index_select(0, new_order)
incremental_state = self._set_input_buffer(incremental_state,
input_buffer)
return incremental_state
def _get_input_buffer(self, incremental_state:
'Optional[Dict[str, Dict[str, Optional[Tensor]]]]') ->Dict[str,
Optional[Tensor]]:
result = self.get_incremental_state(incremental_state, 'attn_state')
if result is not None:
return result
else:
empty_result: 'Dict[str, Optional[Tensor]]' = {}
return empty_result
def _set_input_buffer(self, incremental_state:
'Dict[str, Dict[str, Optional[Tensor]]]', buffer:
'Dict[str, Optional[Tensor]]'):
return self.set_incremental_state(incremental_state, 'attn_state',
buffer)
def apply_sparse_mask(attn_weights, tgt_len: 'int', src_len: 'int', bsz:
'int'):
return attn_weights
def upgrade_state_dict_named(self, state_dict, name):
prefix = name + '.' if name != '' else ''
items_to_add = {}
keys_to_remove = []
for k in state_dict.keys():
if k.endswith(prefix + 'in_proj_weight'):
dim = int(state_dict[k].shape[0] / 3)
items_to_add[prefix + 'q_proj.weight'] = state_dict[k][:dim]
items_to_add[prefix + 'k_proj.weight'] = state_dict[k][dim:
2 * dim]
items_to_add[prefix + 'v_proj.weight'] = state_dict[k][2 * dim:
]
keys_to_remove.append(k)
k_bias = prefix + 'in_proj_bias'
if k_bias in state_dict.keys():
dim = int(state_dict[k].shape[0] / 3)
items_to_add[prefix + 'q_proj.bias'] = state_dict[k_bias][:
dim]
items_to_add[prefix + 'k_proj.bias'] = state_dict[k_bias][
dim:2 * dim]
items_to_add[prefix + 'v_proj.bias'] = state_dict[k_bias][
2 * dim:]
keys_to_remove.append(prefix + 'in_proj_bias')
for k in keys_to_remove:
del state_dict[k]
for key, value in items_to_add.items():
state_dict[key] = value
class TransformerLayer(nn.Module):
"""Transformer layer block."""
def __init__(self, embed_dim, ffn_embed_dim, attention_heads,
add_bias_kv=True, use_esm1b_layer_norm=False):
super().__init__()
self.embed_dim = embed_dim
self.ffn_embed_dim = ffn_embed_dim
self.attention_heads = attention_heads
self._init_submodules(add_bias_kv, use_esm1b_layer_norm)
def _init_submodules(self, add_bias_kv, use_esm1b_layer_norm):
BertLayerNorm = (ESM1bLayerNorm if use_esm1b_layer_norm else
ESM1LayerNorm)
self.self_attn = MultiheadAttention(self.embed_dim, self.
attention_heads, add_bias_kv=add_bias_kv, add_zero_attn=False)
self.self_attn_layer_norm = BertLayerNorm(self.embed_dim)
self.fc1 = nn.Linear(self.embed_dim, self.ffn_embed_dim)
self.fc2 = nn.Linear(self.ffn_embed_dim, self.embed_dim)
self.final_layer_norm = BertLayerNorm(self.embed_dim)
def forward(self, x, self_attn_mask=None, self_attn_padding_mask=None,
need_head_weights=False):
residual = x
x = self.self_attn_layer_norm(x)
x, attn = self.self_attn(query=x, key=x, value=x, key_padding_mask=
self_attn_padding_mask, need_weights=True, need_head_weights=
need_head_weights, attn_mask=self_attn_mask)
x = residual + x
residual = x
x = self.final_layer_norm(x)
x = gelu(self.fc1(x))
x = self.fc2(x)
x = residual + x
return x, attn
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'embed_dim': 4, 'ffn_embed_dim': 4, 'attention_heads': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import math
import uuid
from torch import Tensor
from typing import Tuple
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional
from typing import Dict
from torch.nn import Parameter
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mean_sub_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = 4.0
tmp9 = tmp7 / tmp8
tmp10 = tmp0 - tmp9
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_add_div_mean_mul_pow_sqrt_1(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp20 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp3 = tmp2 * tmp2
tmp5 = tmp4 * tmp4
tmp6 = tmp3 + tmp5
tmp8 = tmp7 * tmp7
tmp9 = tmp6 + tmp8
tmp11 = tmp10 * tmp10
tmp12 = tmp9 + tmp11
tmp13 = 4.0
tmp14 = tmp12 / tmp13
tmp15 = 1e-12
tmp16 = tmp14 + tmp15
tmp17 = libdevice.sqrt(tmp16)
tmp18 = tmp1 / tmp17
tmp19 = tmp0 * tmp18
tmp21 = tmp19 + tmp20
tl.store(out_ptr0 + x2, tmp21, xmask)
@triton.jit
def triton_poi_fused_cat_2(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 12
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + x0, tmp4 & xmask, eviction_policy='evict_last',
other=0.0)
tmp6 = tmp0 >= tmp3
tmp7 = tl.full([1], 8, tl.int64)
tmp8 = tmp0 < tmp7
tmp9 = tmp6 & tmp8
tmp10 = tl.load(in_ptr1 + (-4 + x0), tmp9 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp11 = tmp0 >= tmp7
tl.full([1], 12, tl.int64)
tmp14 = tl.load(in_ptr2 + (-8 + x0), tmp11 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp15 = tl.where(tmp9, tmp10, tmp14)
tmp16 = tl.where(tmp4, tmp5, tmp15)
tl.store(out_ptr0 + x0, tmp16, xmask)
@triton.jit
def triton_poi_fused_cat_3(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 80
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex // 16
x3 = xindex % 16
x0 = xindex % 4
x4 = xindex
tmp0 = x2
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x3 + 16 * x2), tmp4 & xmask, other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 5, tl.int64)
tmp9 = tl.load(in_ptr1 + x0, tmp6 & xmask, eviction_policy='evict_last',
other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x4, tmp10, xmask)
@triton.jit
def triton_poi_fused_mul_4(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel,
XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 16
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = x2 % 4
tl.full([1], 0, tl.int64)
tmp4 = tl.full([1], 4, tl.int64)
tmp5 = tmp1 < tmp4
tmp6 = tl.load(in_ptr0 + x0 % 4, tmp5 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp7 = tmp1 >= tmp4
tmp8 = tl.full([1], 8, tl.int64)
tmp9 = tmp1 < tmp8
tmp10 = tmp7 & tmp9
tmp11 = tl.load(in_ptr1 + (-4 + x0 % 4), tmp10 & xmask, eviction_policy
='evict_last', other=0.0)
tmp12 = tmp1 >= tmp8
tl.full([1], 12, tl.int64)
tmp15 = tl.load(in_ptr2 + (-8 + x0 % 4), tmp12 & xmask, eviction_policy
='evict_last', other=0.0)
tmp16 = tl.where(tmp10, tmp11, tmp15)
tmp17 = tl.where(tmp5, tmp6, tmp16)
tmp18 = tmp0 + tmp17
tmp19 = 1.0
tmp20 = tmp18 * tmp19
tl.store(in_out_ptr0 + x2, tmp20, xmask)
@triton.jit
def triton_poi_fused__softmax_5(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK:
tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 5 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 5 * x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 5 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 5 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (4 + 5 * x0), xmask, eviction_policy='evict_last')
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp4 = triton_helpers.maximum(tmp2, tmp3)
tmp6 = triton_helpers.maximum(tmp4, tmp5)
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = tmp0 - tmp8
tmp10 = tl_math.exp(tmp9)
tmp11 = tmp1 - tmp8
tmp12 = tl_math.exp(tmp11)
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tl_math.exp(tmp14)
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tl_math.exp(tmp17)
tmp19 = tmp16 + tmp18
tmp20 = tmp7 - tmp8
tmp21 = tl_math.exp(tmp20)
tmp22 = tmp19 + tmp21
tl.store(out_ptr0 + x0, tmp8, xmask)
tl.store(out_ptr1 + x0, tmp22, xmask)
@triton.jit
def triton_poi_fused__softmax_6(in_out_ptr0, in_ptr0, in_ptr1, xnumel,
XBLOCK: tl.constexpr):
xnumel = 320
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 5
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp3 = tl_math.exp(tmp2)
tmp5 = tmp3 / tmp4
tl.store(in_out_ptr0 + x2, tmp5, xmask)
@triton.jit
def triton_poi_fused_clone_7(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask, eviction_policy=
'evict_last')
tl.store(out_ptr0 + (x1 + 16 * y0), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_mean_8(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 80
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 20
x1 = xindex // 20
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 80 * x1), xmask)
tmp1 = tl.load(in_ptr0 + (20 + x0 + 80 * x1), xmask)
tmp3 = tl.load(in_ptr0 + (40 + x0 + 80 * x1), xmask)
tmp5 = tl.load(in_ptr0 + (60 + x0 + 80 * x1), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_add_mean_pow_sub_9(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(out_ptr1 + x0, tmp28, xmask)
@triton.jit
def triton_poi_fused_add_div_mean_mul_sqrt_sub_10(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tl.load(in_ptr2 + x2, xmask)
tmp4 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr4 + x1, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 - tmp4
tmp7 = 1e-12
tmp8 = tmp6 + tmp7
tmp9 = libdevice.sqrt(tmp8)
tmp10 = tmp5 / tmp9
tmp11 = tmp0 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
@triton.jit
def triton_poi_fused_add_div_erf_mul_11(in_ptr0, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865475
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + x0, tmp8, xmask)
@triton.jit
def triton_poi_fused_add_12(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel,
XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_out_ptr0 + x2, xmask)
tmp4 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tl.store(in_out_ptr0 + x2, tmp6, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13, primals_14, primals_15, primals_16, primals_17,
primals_18, primals_19) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4,), (1,))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4,), (1,))
assert_size_stride(primals_7, (1, 1, 4), (4, 4, 1))
assert_size_stride(primals_8, (1, 1, 4), (4, 4, 1))
assert_size_stride(primals_9, (4, 4), (4, 1))
assert_size_stride(primals_10, (4,), (1,))
assert_size_stride(primals_11, (4, 4), (4, 1))
assert_size_stride(primals_12, (4, 4), (4, 1))
assert_size_stride(primals_13, (4, 4), (4, 1))
assert_size_stride(primals_14, (4,), (1,))
assert_size_stride(primals_15, (4,), (1,))
assert_size_stride(primals_16, (4, 4), (4, 1))
assert_size_stride(primals_17, (4,), (1,))
assert_size_stride(primals_18, (4, 4), (4, 1))
assert_size_stride(primals_19, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mean_sub_0[grid(64)](primals_1, buf0, 64, XBLOCK=
64, num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_div_mean_mul_pow_sqrt_1[grid(64)](primals_2,
buf0, primals_3, buf1, 64, XBLOCK=64, num_warps=1, num_stages=1)
del primals_2
del primals_3
buf2 = reinterpret_tensor(buf0, (16, 4), (4, 1), 0)
del buf0
extern_kernels.mm(reinterpret_tensor(buf1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_11, (4, 4), (1, 4), 0), out=buf2)
buf3 = empty_strided_cuda((12,), (1,), torch.float32)
triton_poi_fused_cat_2[grid(12)](primals_4, primals_5, primals_6,
buf3, 12, XBLOCK=16, num_warps=1, num_stages=1)
buf4 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(reinterpret_tensor(buf3, (4,), (1,), 4),
reinterpret_tensor(buf1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_12, (4, 4), (1, 4), 0), alpha=1,
beta=1, out=buf4)
buf5 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(reinterpret_tensor(buf3, (4,), (1,), 8),
reinterpret_tensor(buf1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_13, (4, 4), (1, 4), 0), alpha=1,
beta=1, out=buf5)
del buf3
buf6 = empty_strided_cuda((5, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_cat_3[grid(80)](buf5, primals_8, buf6, 80, XBLOCK=
128, num_warps=4, num_stages=1)
del primals_8
buf7 = reinterpret_tensor(buf2, (16, 4, 1), (1, 16, 64), 0)
del buf2
triton_poi_fused_mul_4[grid(64)](buf7, primals_4, primals_5,
primals_6, 64, XBLOCK=64, num_warps=1, num_stages=1)
del primals_4
del primals_5
del primals_6
buf8 = empty_strided_cuda((5, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_cat_3[grid(80)](buf4, primals_7, buf8, 80, XBLOCK=
128, num_warps=4, num_stages=1)
del primals_7
buf9 = empty_strided_cuda((16, 4, 5), (20, 5, 1), torch.float32)
extern_kernels.bmm(buf7, reinterpret_tensor(buf8, (16, 1, 5), (1, 0,
16), 0), out=buf9)
buf10 = reinterpret_tensor(buf4, (16, 4, 1), (4, 1, 64), 0)
del buf4
buf11 = reinterpret_tensor(buf5, (16, 4, 1), (4, 1, 64), 0)
del buf5
triton_poi_fused__softmax_5[grid(64)](buf9, buf10, buf11, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf12 = buf9
del buf9
triton_poi_fused__softmax_6[grid(320)](buf12, buf10, buf11, 320,
XBLOCK=256, num_warps=4, num_stages=1)
buf13 = reinterpret_tensor(buf11, (16, 4, 1), (4, 1, 1), 0)
del buf11
extern_kernels.bmm(buf12, reinterpret_tensor(buf6, (16, 5, 1), (1,
16, 0), 0), out=buf13)
buf14 = reinterpret_tensor(buf10, (4, 16, 1), (16, 1, 1), 0)
del buf10
triton_poi_fused_clone_7[grid(4, 16)](buf13, buf14, 4, 16, XBLOCK=
16, YBLOCK=4, num_warps=1, num_stages=1)
buf15 = reinterpret_tensor(buf13, (16, 4), (4, 1), 0)
del buf13
extern_kernels.addmm(primals_10, reinterpret_tensor(buf14, (16, 4),
(4, 1), 0), reinterpret_tensor(primals_9, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf15)
del primals_10
buf16 = empty_strided_cuda((4, 4, 5), (20, 5, 1), torch.float32)
triton_poi_fused_mean_8[grid(80)](buf12, buf16, 80, XBLOCK=128,
num_warps=4, num_stages=1)
buf17 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf18 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
triton_poi_fused_add_mean_pow_sub_9[grid(16)](primals_1, buf15,
buf17, buf18, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf19 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_div_mean_mul_sqrt_sub_10[grid(64)](primals_14,
primals_1, buf15, buf17, buf18, primals_15, buf19, 64, XBLOCK=
64, num_warps=1, num_stages=1)
del buf17
del buf18
del primals_15
buf20 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_17, reinterpret_tensor(buf19, (16, 4),
(4, 1), 0), reinterpret_tensor(primals_16, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf20)
del primals_17
buf21 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_div_erf_mul_11[grid(64)](buf20, buf21, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf22 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf21, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_18, (4, 4), (1, 4), 0), out=buf22)
buf23 = reinterpret_tensor(buf22, (4, 4, 4), (16, 4, 1), 0)
del buf22
triton_poi_fused_add_12[grid(64)](buf23, primals_1, buf15,
primals_19, 64, XBLOCK=64, num_warps=1, num_stages=1)
del primals_19
return buf23, buf16, primals_1, primals_14, reinterpret_tensor(buf1, (
16, 4), (4, 1), 0), buf12, reinterpret_tensor(buf14, (16, 4), (4, 1), 0
), buf15, reinterpret_tensor(buf19, (16, 4), (4, 1), 0
), buf20, reinterpret_tensor(buf21, (16, 4), (4, 1), 0
), primals_18, primals_16, primals_9, reinterpret_tensor(buf6, (16,
1, 5), (1, 1, 16), 0), reinterpret_tensor(buf7, (16, 1, 4), (1, 1,
16), 0), reinterpret_tensor(buf8, (16, 5, 1), (1, 16, 1), 0
), primals_13, primals_12, primals_11
def gelu(x):
"""Implementation of the gelu activation function.
For information: OpenAI GPT's gelu is slightly different
(and gives slightly different results):
0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
"""
return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
def utils_softmax(x, dim: 'int', onnx_trace: 'bool'=False):
if onnx_trace:
return F.softmax(x.float(), dim=dim)
else:
return F.softmax(x, dim=dim, dtype=torch.float32)
def with_incremental_state(cls):
cls.__bases__ = (FairseqIncrementalState,) + tuple(b for b in cls.
__bases__ if b != FairseqIncrementalState)
return cls
class ESM1LayerNorm(nn.Module):
def __init__(self, hidden_size, eps=1e-12, affine=True):
"""Construct a layernorm layer in the TF style (eps inside the sqrt)."""
super().__init__()
self.hidden_size = (hidden_size,) if isinstance(hidden_size, int
) else tuple(hidden_size)
self.eps = eps
self.affine = bool(affine)
if self.affine:
self.weight = nn.Parameter(torch.ones(hidden_size))
self.bias = nn.Parameter(torch.zeros(hidden_size))
else:
self.weight, self.bias = None, None
def forward(self, x):
dims = tuple(-(i + 1) for i in range(len(self.hidden_size)))
means = x.mean(dims, keepdim=True)
x_zeromean = x - means
variances = x_zeromean.pow(2).mean(dims, keepdim=True)
x = x_zeromean / torch.sqrt(variances + self.eps)
if self.affine:
x = self.weight * x + self.bias
return x
class FairseqIncrementalState(object):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.init_incremental_state()
def init_incremental_state(self):
self._incremental_state_id = str(uuid.uuid4())
def _get_full_incremental_state_key(self, key: 'str') ->str:
return '{}.{}'.format(self._incremental_state_id, key)
def get_incremental_state(self, incremental_state:
'Optional[Dict[str, Dict[str, Optional[Tensor]]]]', key: 'str'
) ->Optional[Dict[str, Optional[Tensor]]]:
"""Helper for getting incremental state for an nn.Module."""
full_key = self._get_full_incremental_state_key(key)
if incremental_state is None or full_key not in incremental_state:
return None
return incremental_state[full_key]
def set_incremental_state(self, incremental_state:
'Optional[Dict[str, Dict[str, Optional[Tensor]]]]', key: 'str',
value: 'Dict[str, Optional[Tensor]]') ->Optional[Dict[str, Dict[str,
Optional[Tensor]]]]:
"""Helper for setting incremental state for an nn.Module."""
if incremental_state is not None:
full_key = self._get_full_incremental_state_key(key)
incremental_state[full_key] = value
return incremental_state
@with_incremental_state
class MultiheadAttention(nn.Module):
"""Multi-headed attention.
See "Attention Is All You Need" for more details.
"""
def __init__(self, embed_dim, num_heads, kdim=None, vdim=None, dropout=
0.0, bias=True, add_bias_kv=False, add_zero_attn=False,
self_attention=False, encoder_decoder_attention=False):
super().__init__()
self.embed_dim = embed_dim
self.kdim = kdim if kdim is not None else embed_dim
self.vdim = vdim if vdim is not None else embed_dim
self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
assert self.head_dim * num_heads == self.embed_dim, 'embed_dim must be divisible by num_heads'
self.scaling = self.head_dim ** -0.5
self.self_attention = self_attention
self.encoder_decoder_attention = encoder_decoder_attention
assert not self.self_attention or self.qkv_same_dim, 'Self-attention requires query, key and value to be of the same size'
self.k_proj = nn.Linear(self.kdim, embed_dim, bias=bias)
self.v_proj = nn.Linear(self.vdim, embed_dim, bias=bias)
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
if add_bias_kv:
self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))
self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))
else:
self.bias_k = self.bias_v = None
self.add_zero_attn = add_zero_attn
self.reset_parameters()
self.onnx_trace = False
self.enable_torch_version = False
if hasattr(F, 'multi_head_attention_forward'):
self.enable_torch_version = True
else:
self.enable_torch_version = False
def prepare_for_onnx_export_(self):
self.onnx_trace = True
def reset_parameters(self):
if self.qkv_same_dim:
nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))
nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))
nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))
else:
nn.init.xavier_uniform_(self.k_proj.weight)
nn.init.xavier_uniform_(self.v_proj.weight)
nn.init.xavier_uniform_(self.q_proj.weight)
nn.init.xavier_uniform_(self.out_proj.weight)
if self.out_proj.bias is not None:
nn.init.constant_(self.out_proj.bias, 0.0)
if self.bias_k is not None:
nn.init.xavier_normal_(self.bias_k)
if self.bias_v is not None:
nn.init.xavier_normal_(self.bias_v)
def forward(self, query, key: 'Optional[Tensor]', value:
'Optional[Tensor]', key_padding_mask: 'Optional[Tensor]'=None,
incremental_state:
'Optional[Dict[str, Dict[str, Optional[Tensor]]]]'=None,
need_weights: 'bool'=True, static_kv: 'bool'=False, attn_mask:
'Optional[Tensor]'=None, before_softmax: 'bool'=False,
need_head_weights: 'bool'=False) ->Tuple[Tensor, Optional[Tensor]]:
"""Input shape: Time x Batch x Channel
Args:
key_padding_mask (ByteTensor, optional): mask to exclude
keys that are pads, of shape `(batch, src_len)`, where
padding elements are indicated by 1s.
need_weights (bool, optional): return the attention weights,
averaged over heads (default: False).
attn_mask (ByteTensor, optional): typically used to
implement causal attention, where the mask prevents the
attention from looking forward in time (default: None).
before_softmax (bool, optional): return the raw attention
weights and values before the attention softmax.
need_head_weights (bool, optional): return the attention
weights for each head. Implies *need_weights*. Default:
return the average attention weights over all heads.
"""
if need_head_weights:
need_weights = True
tgt_len, bsz, embed_dim = query.size()
assert embed_dim == self.embed_dim
assert list(query.size()) == [tgt_len, bsz, embed_dim]
if (self.enable_torch_version and not self.onnx_trace and
incremental_state is None and not static_kv and not torch.jit.
is_scripting() and not need_head_weights):
assert key is not None and value is not None
return F.multi_head_attention_forward(query, key, value, self.
embed_dim, self.num_heads, torch.empty([0]), torch.cat((
self.q_proj.bias, self.k_proj.bias, self.v_proj.bias)),
self.bias_k, self.bias_v, self.add_zero_attn, self.dropout,
self.out_proj.weight, self.out_proj.bias, self.training,
key_padding_mask, need_weights, attn_mask,
use_separate_proj_weight=True, q_proj_weight=self.q_proj.
weight, k_proj_weight=self.k_proj.weight, v_proj_weight=
self.v_proj.weight)
if incremental_state is not None:
saved_state = self._get_input_buffer(incremental_state)
if saved_state is not None and 'prev_key' in saved_state:
if static_kv:
assert self.encoder_decoder_attention and not self.self_attention
key = value = None
else:
saved_state = None
if self.self_attention:
q = self.q_proj(query)
k = self.k_proj(query)
v = self.v_proj(query)
elif self.encoder_decoder_attention:
q = self.q_proj(query)
if key is None:
assert value is None
k = v = None
else:
k = self.k_proj(key)
v = self.v_proj(key)
else:
assert key is not None and value is not None
q = self.q_proj(query)
k = self.k_proj(key)
v = self.v_proj(value)
q *= self.scaling
if self.bias_k is not None:
assert self.bias_v is not None
k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])
v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])
if attn_mask is not None:
attn_mask = torch.cat([attn_mask, attn_mask.new_zeros(
attn_mask.size(0), 1)], dim=1)
if key_padding_mask is not None:
key_padding_mask = torch.cat([key_padding_mask,
key_padding_mask.new_zeros(key_padding_mask.size(0), 1)
], dim=1)
q = q.contiguous().view(tgt_len, bsz * self.num_heads, self.head_dim
).transpose(0, 1)
if k is not None:
k = k.contiguous().view(-1, bsz * self.num_heads, self.head_dim
).transpose(0, 1)
if v is not None:
v = v.contiguous().view(-1, bsz * self.num_heads, self.head_dim
).transpose(0, 1)
if saved_state is not None:
if 'prev_key' in saved_state:
_prev_key = saved_state['prev_key']
assert _prev_key is not None
prev_key = _prev_key.view(bsz * self.num_heads, -1, self.
head_dim)
if static_kv:
k = prev_key
else:
assert k is not None
k = torch.cat([prev_key, k], dim=1)
if 'prev_value' in saved_state:
_prev_value = saved_state['prev_value']
assert _prev_value is not None
prev_value = _prev_value.view(bsz * self.num_heads, -1,
self.head_dim)
if static_kv:
v = prev_value
else:
assert v is not None
v = torch.cat([prev_value, v], dim=1)
prev_key_padding_mask: 'Optional[Tensor]' = None
if 'prev_key_padding_mask' in saved_state:
prev_key_padding_mask = saved_state['prev_key_padding_mask']
assert k is not None and v is not None
key_padding_mask = (MultiheadAttention.
_append_prev_key_padding_mask(key_padding_mask=
key_padding_mask, prev_key_padding_mask=
prev_key_padding_mask, batch_size=bsz, src_len=k.size(1),
static_kv=static_kv))
saved_state['prev_key'] = k.view(bsz, self.num_heads, -1, self.
head_dim)
saved_state['prev_value'] = v.view(bsz, self.num_heads, -1,
self.head_dim)
saved_state['prev_key_padding_mask'] = key_padding_mask
assert incremental_state is not None
incremental_state = self._set_input_buffer(incremental_state,
saved_state)
assert k is not None
src_len = k.size(1)
if key_padding_mask is not None and key_padding_mask.dim() == 0:
key_padding_mask = None
if key_padding_mask is not None:
assert key_padding_mask.size(0) == bsz
assert key_padding_mask.size(1) == src_len
if self.add_zero_attn:
assert v is not None
src_len += 1
k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])],
dim=1)
v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])],
dim=1)
if attn_mask is not None:
attn_mask = torch.cat([attn_mask, attn_mask.new_zeros(
attn_mask.size(0), 1)], dim=1)
if key_padding_mask is not None:
key_padding_mask = torch.cat([key_padding_mask, torch.zeros
(key_padding_mask.size(0), 1).type_as(key_padding_mask)
], dim=1)
attn_weights = torch.bmm(q, k.transpose(1, 2))
attn_weights = MultiheadAttention.apply_sparse_mask(attn_weights,
tgt_len, src_len, bsz)
assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len,
src_len]
if attn_mask is not None:
attn_mask = attn_mask.unsqueeze(0)
if self.onnx_trace:
attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1)
attn_weights += attn_mask
if key_padding_mask is not None:
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len,
src_len)
attn_weights = attn_weights.masked_fill(key_padding_mask.
unsqueeze(1).unsqueeze(2), float('-inf'))
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len,
src_len)
if before_softmax:
return attn_weights, v
attn_weights_float = utils_softmax(attn_weights, dim=-1, onnx_trace
=self.onnx_trace)
attn_weights = attn_weights_float.type_as(attn_weights)
attn_probs = F.dropout(attn_weights_float.type_as(attn_weights), p=
self.dropout, training=self.training)
assert v is not None
attn = torch.bmm(attn_probs, v)
assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.
head_dim]
if self.onnx_trace and attn.size(1) == 1:
attn = attn.contiguous().view(tgt_len, bsz, embed_dim)
else:
attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz,
embed_dim)
attn = self.out_proj(attn)
attn_weights: 'Optional[Tensor]' = None
if need_weights:
attn_weights = attn_weights_float.view(bsz, self.num_heads,
tgt_len, src_len).transpose(1, 0)
if not need_head_weights:
attn_weights = attn_weights.mean(dim=0)
return attn, attn_weights
@staticmethod
def _append_prev_key_padding_mask(key_padding_mask: 'Optional[Tensor]',
prev_key_padding_mask: 'Optional[Tensor]', batch_size: 'int',
src_len: 'int', static_kv: 'bool') ->Optional[Tensor]:
if prev_key_padding_mask is not None and static_kv:
new_key_padding_mask = prev_key_padding_mask
elif prev_key_padding_mask is not None and key_padding_mask is not None:
new_key_padding_mask = torch.cat([prev_key_padding_mask.float(),
key_padding_mask.float()], dim=1)
elif prev_key_padding_mask is not None:
filler = torch.zeros((batch_size, src_len -
prev_key_padding_mask.size(1)), device=
prev_key_padding_mask.device)
new_key_padding_mask = torch.cat([prev_key_padding_mask.float(),
filler.float()], dim=1)
elif key_padding_mask is not None:
filler = torch.zeros((batch_size, src_len - key_padding_mask.
size(1)), device=key_padding_mask.device)
new_key_padding_mask = torch.cat([filler.float(),
key_padding_mask.float()], dim=1)
else:
new_key_padding_mask = prev_key_padding_mask
return new_key_padding_mask
@torch.jit.export
def reorder_incremental_state(self, incremental_state:
'Dict[str, Dict[str, Optional[Tensor]]]', new_order: 'Tensor'):
"""Reorder buffered internal state (for incremental generation)."""
input_buffer = self._get_input_buffer(incremental_state)
if input_buffer is not None:
for k in input_buffer.keys():
input_buffer_k = input_buffer[k]
if input_buffer_k is not None:
if self.encoder_decoder_attention and input_buffer_k.size(0
) == new_order.size(0):
break
input_buffer[k] = input_buffer_k.index_select(0, new_order)
incremental_state = self._set_input_buffer(incremental_state,
input_buffer)
return incremental_state
def _get_input_buffer(self, incremental_state:
'Optional[Dict[str, Dict[str, Optional[Tensor]]]]') ->Dict[str,
Optional[Tensor]]:
result = self.get_incremental_state(incremental_state, 'attn_state')
if result is not None:
return result
else:
empty_result: 'Dict[str, Optional[Tensor]]' = {}
return empty_result
def _set_input_buffer(self, incremental_state:
'Dict[str, Dict[str, Optional[Tensor]]]', buffer:
'Dict[str, Optional[Tensor]]'):
return self.set_incremental_state(incremental_state, 'attn_state',
buffer)
def apply_sparse_mask(attn_weights, tgt_len: 'int', src_len: 'int', bsz:
'int'):
return attn_weights
def upgrade_state_dict_named(self, state_dict, name):
prefix = name + '.' if name != '' else ''
items_to_add = {}
keys_to_remove = []
for k in state_dict.keys():
if k.endswith(prefix + 'in_proj_weight'):
dim = int(state_dict[k].shape[0] / 3)
items_to_add[prefix + 'q_proj.weight'] = state_dict[k][:dim]
items_to_add[prefix + 'k_proj.weight'] = state_dict[k][dim:
2 * dim]
items_to_add[prefix + 'v_proj.weight'] = state_dict[k][2 * dim:
]
keys_to_remove.append(k)
k_bias = prefix + 'in_proj_bias'
if k_bias in state_dict.keys():
dim = int(state_dict[k].shape[0] / 3)
items_to_add[prefix + 'q_proj.bias'] = state_dict[k_bias][:
dim]
items_to_add[prefix + 'k_proj.bias'] = state_dict[k_bias][
dim:2 * dim]
items_to_add[prefix + 'v_proj.bias'] = state_dict[k_bias][
2 * dim:]
keys_to_remove.append(prefix + 'in_proj_bias')
for k in keys_to_remove:
del state_dict[k]
for key, value in items_to_add.items():
state_dict[key] = value
class TransformerLayerNew(nn.Module):
"""Transformer layer block."""
def __init__(self, embed_dim, ffn_embed_dim, attention_heads,
add_bias_kv=True, use_esm1b_layer_norm=False):
super().__init__()
self.embed_dim = embed_dim
self.ffn_embed_dim = ffn_embed_dim
self.attention_heads = attention_heads
self._init_submodules(add_bias_kv, use_esm1b_layer_norm)
def _init_submodules(self, add_bias_kv, use_esm1b_layer_norm):
BertLayerNorm = (ESM1bLayerNorm if use_esm1b_layer_norm else
ESM1LayerNorm)
self.self_attn = MultiheadAttention(self.embed_dim, self.
attention_heads, add_bias_kv=add_bias_kv, add_zero_attn=False)
self.self_attn_layer_norm = BertLayerNorm(self.embed_dim)
self.fc1 = nn.Linear(self.embed_dim, self.ffn_embed_dim)
self.fc2 = nn.Linear(self.ffn_embed_dim, self.embed_dim)
self.final_layer_norm = BertLayerNorm(self.embed_dim)
def forward(self, input_0):
primals_7 = self.self_attn.bias_k
primals_8 = self.self_attn.bias_v
primals_9 = self.self_attn.k_proj.weight
primals_2 = self.self_attn.k_proj.bias
primals_11 = self.self_attn.v_proj.weight
primals_3 = self.self_attn.v_proj.bias
primals_12 = self.self_attn.q_proj.weight
primals_4 = self.self_attn.q_proj.bias
primals_13 = self.self_attn.out_proj.weight
primals_5 = self.self_attn.out_proj.bias
primals_6 = self.self_attn_layer_norm.weight
primals_10 = self.self_attn_layer_norm.bias
primals_16 = self.fc1.weight
primals_14 = self.fc1.bias
primals_18 = self.fc2.weight
primals_15 = self.fc2.bias
primals_17 = self.final_layer_norm.weight
primals_19 = self.final_layer_norm.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13, primals_14,
primals_15, primals_16, primals_17, primals_18, primals_19])
return output[0], output[1]
|
qinwang-ai/Contact-Distil
|
TransformerLayer
| false
| 4,193
|
[
"Apache-2.0"
] | 0
|
5e98389de70e0d9c4d16bd91ca1326689dc220a6
|
https://github.com/qinwang-ai/Contact-Distil/tree/5e98389de70e0d9c4d16bd91ca1326689dc220a6
|
MLP
|
import torch
import torch as th
import torch.nn as nn
class MLP(nn.Module):
def __init__(self, input_size, output_size, hidden=128):
super(MLP, self).__init__()
self.linear1 = nn.Linear(input_size, hidden, bias=False)
self.linear2 = nn.Linear(hidden, output_size, bias=False)
def forward(self, x):
x = self.linear1(x)
x = th.tanh(x)
x = self.linear2(x)
x = th.tanh(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'input_size': 4, 'output_size': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_tanh_0(in_out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, None)
tmp1 = libdevice.tanh(tmp0)
tl.store(in_out_ptr0 + x0, tmp1, None)
@triton.jit
def triton_poi_fused_tanh_1(in_out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = libdevice.tanh(tmp0)
tl.store(in_out_ptr0 + x0, tmp1, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (128, 4), (4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, 128), (128, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 128), (128, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_2, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 128), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 128), (2048, 512, 128, 1), 0)
del buf0
get_raw_stream(0)
triton_poi_fused_tanh_0[grid(8192)](buf1, 8192, XBLOCK=256,
num_warps=4, num_stages=1)
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf1, (64, 128), (128, 1), 0),
reinterpret_tensor(primals_3, (128, 4), (1, 128), 0), out=buf2)
buf3 = reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf2
triton_poi_fused_tanh_1[grid(256)](buf3, 256, XBLOCK=128, num_warps
=4, num_stages=1)
return buf3, reinterpret_tensor(primals_2, (64, 4), (4, 1), 0
), buf1, buf3, primals_3
class MLPNew(nn.Module):
def __init__(self, input_size, output_size, hidden=128):
super(MLPNew, self).__init__()
self.linear1 = nn.Linear(input_size, hidden, bias=False)
self.linear2 = nn.Linear(hidden, output_size, bias=False)
def forward(self, input_0):
primals_1 = self.linear1.weight
primals_3 = self.linear2.weight
primals_2 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
|
ngoby/cherry
|
MLP
| false
| 4,194
|
[
"Apache-2.0"
] | 0
|
ec88bac03bf3ac3fae1010c5db8329db595dc5d6
|
https://github.com/ngoby/cherry/tree/ec88bac03bf3ac3fae1010c5db8329db595dc5d6
|
EncoderLayer
|
import math
import torch
import torch.nn.functional as F
import torch.nn as nn
def attention(q, k, v, d_k, mask=None, dropout=None):
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(d_k)
if mask is not None:
mask = mask.unsqueeze(1)
scores = scores.masked_fill(mask == 0, -1000000000.0)
scores = F.softmax(scores, dim=-1)
if dropout is not None:
scores = dropout(scores)
output = torch.matmul(scores, v)
return output
class FeedForward(nn.Module):
def __init__(self, d_model, d_ff=2048, dropout=0.1):
super().__init__()
self.linear_1 = nn.Linear(d_model, d_ff)
self.dropout = nn.Dropout(dropout)
self.linear_2 = nn.Linear(d_ff, d_model)
def forward(self, x):
x = self.dropout(F.relu(self.linear_1(x)))
x = self.linear_2(x)
return x
class MultiHeadAttention(nn.Module):
def __init__(self, heads, d_model, dropout=0.1):
super().__init__()
self.d_model = d_model
self.d_k = d_model // heads
self.h = heads
self.q_linear = nn.Linear(d_model, d_model)
self.v_linear = nn.Linear(d_model, d_model)
self.k_linear = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(dropout)
self.out = nn.Linear(d_model, d_model)
def forward(self, q, k, v, mask=None):
bs = q.size(0)
k = self.k_linear(k).view(bs, -1, self.h, self.d_k)
q = self.q_linear(q).view(bs, -1, self.h, self.d_k)
v = self.v_linear(v).view(bs, -1, self.h, self.d_k)
k = k.transpose(1, 2)
q = q.transpose(1, 2)
v = v.transpose(1, 2)
scores = attention(q, k, v, self.d_k, mask, self.dropout)
concat = scores.transpose(1, 2).contiguous().view(bs, -1, self.d_model)
output = self.out(concat)
return output
class Norm(nn.Module):
def __init__(self, d_model, eps=1e-06):
super().__init__()
self.size = d_model
self.alpha = nn.Parameter(torch.ones(self.size))
self.bias = nn.Parameter(torch.zeros(self.size))
self.eps = eps
def forward(self, x):
norm = self.alpha * (x - x.mean(dim=-1, keepdim=True)) / (x.std(dim
=-1, keepdim=True) + self.eps) + self.bias
return norm
class EncoderLayer(nn.Module):
def __init__(self, d_model, heads, dropout=0.1):
super().__init__()
self.norm_1 = Norm(d_model)
self.norm_2 = Norm(d_model)
self.attn = MultiHeadAttention(heads, d_model, dropout=dropout)
self.ff = FeedForward(d_model, dropout=dropout)
self.dropout_1 = nn.Dropout(dropout)
self.dropout_2 = nn.Dropout(dropout)
def forward(self, x, mask):
x2 = self.norm_1(x)
x = x + self.dropout_1(self.attn(x2, x2, x2, mask))
x2 = self.norm_2(x)
x = x + self.dropout_2(self.ff(x2))
return x
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'d_model': 4, 'heads': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import math
import torch.nn.functional as F
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_add_div_mean_mul_std_sub_0(in_ptr0, in_ptr1, in_ptr2,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp30 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp8 = tmp6 + tmp7
tmp9 = 4.0
tmp10 = tmp8 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp0 * tmp11
tmp13 = tmp2 - tmp10
tmp14 = tmp13 * tmp13
tmp15 = tmp3 - tmp10
tmp16 = tmp15 * tmp15
tmp17 = tmp14 + tmp16
tmp18 = tmp5 - tmp10
tmp19 = tmp18 * tmp18
tmp20 = tmp17 + tmp19
tmp21 = tmp7 - tmp10
tmp22 = tmp21 * tmp21
tmp23 = tmp20 + tmp22
tmp24 = 3.0
tmp25 = tmp23 / tmp24
tmp26 = libdevice.sqrt(tmp25)
tmp27 = 1e-06
tmp28 = tmp26 + tmp27
tmp29 = tmp12 / tmp28
tmp31 = tmp29 + tmp30
tl.store(out_ptr0 + x2, tmp31, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused_eq_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.0
tmp2 = tmp0 == tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
@triton.jit
def triton_poi_fused__softmax_div_masked_fill_3(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex // 16
x3 = xindex
tmp0 = tl.load(in_ptr0 + (4 * x0 + 16 * x2), xmask, eviction_policy=
'evict_last').to(tl.int1)
tmp1 = tl.load(in_ptr1 + 4 * x3, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (1 + 4 * x0 + 16 * x2), xmask, eviction_policy
='evict_last').to(tl.int1)
tmp7 = tl.load(in_ptr1 + (1 + 4 * x3), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (2 + 4 * x0 + 16 * x2), xmask,
eviction_policy='evict_last').to(tl.int1)
tmp12 = tl.load(in_ptr1 + (2 + 4 * x3), xmask, eviction_policy='evict_last'
)
tmp16 = tl.load(in_ptr0 + (3 + 4 * x0 + 16 * x2), xmask,
eviction_policy='evict_last').to(tl.int1)
tmp17 = tl.load(in_ptr1 + (3 + 4 * x3), xmask, eviction_policy='evict_last'
)
tmp2 = 1.0
tmp3 = tmp1 * tmp2
tmp4 = -1000000000.0
tmp5 = tl.where(tmp0, tmp4, tmp3)
tmp8 = tmp7 * tmp2
tmp9 = tl.where(tmp6, tmp4, tmp8)
tmp10 = triton_helpers.maximum(tmp5, tmp9)
tmp13 = tmp12 * tmp2
tmp14 = tl.where(tmp11, tmp4, tmp13)
tmp15 = triton_helpers.maximum(tmp10, tmp14)
tmp18 = tmp17 * tmp2
tmp19 = tl.where(tmp16, tmp4, tmp18)
tmp20 = triton_helpers.maximum(tmp15, tmp19)
tmp21 = tmp5 - tmp20
tmp22 = tl_math.exp(tmp21)
tmp23 = tmp9 - tmp20
tmp24 = tl_math.exp(tmp23)
tmp25 = tmp22 + tmp24
tmp26 = tmp14 - tmp20
tmp27 = tl_math.exp(tmp26)
tmp28 = tmp25 + tmp27
tmp29 = tmp19 - tmp20
tmp30 = tl_math.exp(tmp29)
tmp31 = tmp28 + tmp30
tl.store(out_ptr0 + x3, tmp20, xmask)
tl.store(out_ptr1 + x3, tmp31, xmask)
@triton.jit
def triton_poi_fused__softmax_div_masked_fill_4(in_out_ptr0, in_ptr0,
in_ptr1, in_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex // 64
x4 = xindex % 16
x5 = xindex
x6 = xindex // 4
tmp0 = tl.load(in_ptr0 + (x4 + 16 * x3), xmask, eviction_policy=
'evict_last').to(tl.int1)
tmp1 = tl.load(in_out_ptr0 + x5, xmask)
tmp6 = tl.load(in_ptr1 + x6, xmask, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr2 + x6, xmask, eviction_policy='evict_last')
tmp2 = 1.0
tmp3 = tmp1 * tmp2
tmp4 = -1000000000.0
tmp5 = tl.where(tmp0, tmp4, tmp3)
tmp7 = tmp5 - tmp6
tmp8 = tl_math.exp(tmp7)
tmp10 = tmp8 / tmp9
tl.store(in_out_ptr0 + x5, tmp10, xmask)
@triton.jit
def triton_poi_fused_clone_5(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_mean_std_6(in_out_ptr0, in_ptr0, in_ptr1, out_ptr0,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = 3.0
tmp29 = tmp27 / tmp28
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(in_out_ptr0 + x0, tmp29, xmask)
@triton.jit
def triton_poi_fused_add_div_mean_mul_std_sub_7(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tl.load(in_ptr2 + x2, xmask)
tmp4 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x1, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 - tmp4
tmp6 = tmp0 * tmp5
tmp8 = libdevice.sqrt(tmp7)
tmp9 = 1e-06
tmp10 = tmp8 + tmp9
tmp11 = tmp6 / tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
@triton.jit
def triton_poi_fused_relu_threshold_backward_8(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 2048
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, None)
tl.store(out_ptr0 + x2, tmp6, None)
@triton.jit
def triton_poi_fused_add_9(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel,
XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_out_ptr0 + x2, xmask)
tmp4 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tl.store(in_out_ptr0 + x2, tmp6, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13, primals_14, primals_15, primals_16, primals_17, primals_18
) = args
args.clear()
assert_size_stride(primals_1, (4,), (1,))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 4), (4, 1))
assert_size_stride(primals_9, (4,), (1,))
assert_size_stride(primals_10, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_11, (4, 4), (4, 1))
assert_size_stride(primals_12, (4,), (1,))
assert_size_stride(primals_13, (4,), (1,))
assert_size_stride(primals_14, (4,), (1,))
assert_size_stride(primals_15, (2048, 4), (4, 1))
assert_size_stride(primals_16, (2048,), (1,))
assert_size_stride(primals_17, (4, 2048), (2048, 1))
assert_size_stride(primals_18, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_div_mean_mul_std_sub_0[grid(64)](primals_1,
primals_2, primals_3, buf0, 64, XBLOCK=64, num_warps=1,
num_stages=1)
del primals_1
del primals_3
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 4), (1, 4), 0), out=buf1)
buf2 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), out=buf2)
buf3 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_8, (4, 4), (1, 4), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
triton_poi_fused_clone_1[grid(16, 4)](buf2, primals_7, buf4, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_7
buf5 = reinterpret_tensor(buf2, (4, 4, 1, 4), (16, 4, 4, 1), 0)
del buf2
triton_poi_fused_clone_1[grid(16, 4)](buf1, primals_5, buf5, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_5
buf6 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf4, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf5, (16, 1, 4), (4, 0, 1), 0), out=buf6)
buf7 = empty_strided_cuda((4, 1, 4, 4), (16, 16, 4, 1), torch.bool)
triton_poi_fused_eq_2[grid(64)](primals_10, buf7, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del primals_10
buf8 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 64), 0)
del buf1
buf9 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
triton_poi_fused__softmax_div_masked_fill_3[grid(64)](buf7, buf6,
buf8, buf9, 64, XBLOCK=64, num_warps=1, num_stages=1)
buf10 = reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf6
triton_poi_fused__softmax_div_masked_fill_4[grid(256)](buf10, buf7,
buf8, buf9, 256, XBLOCK=128, num_warps=4, num_stages=1)
buf11 = reinterpret_tensor(buf9, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf9
triton_poi_fused_clone_1[grid(16, 4)](buf3, primals_9, buf11, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_9
buf12 = reinterpret_tensor(buf3, (16, 4, 1), (4, 1, 1), 0)
del buf3
extern_kernels.bmm(reinterpret_tensor(buf10, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf11, (16, 4, 1), (4, 1, 0), 0), out=buf12)
buf13 = reinterpret_tensor(buf8, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf8
triton_poi_fused_clone_5[grid(16, 4)](buf12, buf13, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf14 = reinterpret_tensor(buf12, (16, 4), (4, 1), 0)
del buf12
extern_kernels.addmm(primals_12, reinterpret_tensor(buf13, (16, 4),
(4, 1), 0), reinterpret_tensor(primals_11, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf14)
del primals_12
buf15 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf16 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf17 = buf16
del buf16
triton_poi_fused_add_mean_std_6[grid(16)](buf17, primals_2, buf14,
buf15, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf18 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_div_mean_mul_std_sub_7[grid(64)](primals_13,
primals_2, buf14, buf15, buf17, primals_14, buf18, 64, XBLOCK=
64, num_warps=1, num_stages=1)
del buf15
del buf17
del primals_14
buf19 = empty_strided_cuda((16, 2048), (2048, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf18, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_15, (4, 2048), (1, 4), 0), out=buf19)
buf20 = reinterpret_tensor(buf19, (4, 4, 2048), (8192, 2048, 1), 0)
del buf19
buf23 = empty_strided_cuda((4, 4, 2048), (8192, 2048, 1), torch.bool)
triton_poi_fused_relu_threshold_backward_8[grid(32768)](buf20,
primals_16, buf23, 32768, XBLOCK=128, num_warps=4, num_stages=1)
del primals_16
buf21 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf20, (16, 2048), (2048, 1),
0), reinterpret_tensor(primals_17, (2048, 4), (1, 2048), 0),
out=buf21)
buf22 = reinterpret_tensor(buf21, (4, 4, 4), (16, 4, 1), 0)
del buf21
triton_poi_fused_add_9[grid(64)](buf22, primals_2, buf14,
primals_18, 64, XBLOCK=64, num_warps=1, num_stages=1)
del primals_18
return buf22, primals_2, primals_13, reinterpret_tensor(buf0, (16, 4),
(4, 1), 0), buf7, buf10, reinterpret_tensor(buf13, (16, 4), (4, 1), 0
), buf14, reinterpret_tensor(buf18, (16, 4), (4, 1), 0
), reinterpret_tensor(buf20, (16, 2048), (2048, 1), 0
), primals_17, buf23, primals_15, primals_11, reinterpret_tensor(buf11,
(16, 1, 4), (4, 1, 1), 0), reinterpret_tensor(buf4, (16, 1, 4), (4,
1, 1), 0), reinterpret_tensor(buf5, (16, 4, 1), (4, 1, 4), 0
), primals_8, primals_6, primals_4
def attention(q, k, v, d_k, mask=None, dropout=None):
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(d_k)
if mask is not None:
mask = mask.unsqueeze(1)
scores = scores.masked_fill(mask == 0, -1000000000.0)
scores = F.softmax(scores, dim=-1)
if dropout is not None:
scores = dropout(scores)
output = torch.matmul(scores, v)
return output
class FeedForward(nn.Module):
def __init__(self, d_model, d_ff=2048, dropout=0.1):
super().__init__()
self.linear_1 = nn.Linear(d_model, d_ff)
self.dropout = nn.Dropout(dropout)
self.linear_2 = nn.Linear(d_ff, d_model)
def forward(self, x):
x = self.dropout(F.relu(self.linear_1(x)))
x = self.linear_2(x)
return x
class MultiHeadAttention(nn.Module):
def __init__(self, heads, d_model, dropout=0.1):
super().__init__()
self.d_model = d_model
self.d_k = d_model // heads
self.h = heads
self.q_linear = nn.Linear(d_model, d_model)
self.v_linear = nn.Linear(d_model, d_model)
self.k_linear = nn.Linear(d_model, d_model)
self.dropout = nn.Dropout(dropout)
self.out = nn.Linear(d_model, d_model)
def forward(self, q, k, v, mask=None):
bs = q.size(0)
k = self.k_linear(k).view(bs, -1, self.h, self.d_k)
q = self.q_linear(q).view(bs, -1, self.h, self.d_k)
v = self.v_linear(v).view(bs, -1, self.h, self.d_k)
k = k.transpose(1, 2)
q = q.transpose(1, 2)
v = v.transpose(1, 2)
scores = attention(q, k, v, self.d_k, mask, self.dropout)
concat = scores.transpose(1, 2).contiguous().view(bs, -1, self.d_model)
output = self.out(concat)
return output
class Norm(nn.Module):
def __init__(self, d_model, eps=1e-06):
super().__init__()
self.size = d_model
self.alpha = nn.Parameter(torch.ones(self.size))
self.bias = nn.Parameter(torch.zeros(self.size))
self.eps = eps
def forward(self, x):
norm = self.alpha * (x - x.mean(dim=-1, keepdim=True)) / (x.std(dim
=-1, keepdim=True) + self.eps) + self.bias
return norm
class EncoderLayerNew(nn.Module):
def __init__(self, d_model, heads, dropout=0.1):
super().__init__()
self.norm_1 = Norm(d_model)
self.norm_2 = Norm(d_model)
self.attn = MultiHeadAttention(heads, d_model, dropout=dropout)
self.ff = FeedForward(d_model, dropout=dropout)
self.dropout_1 = nn.Dropout(dropout)
self.dropout_2 = nn.Dropout(dropout)
def forward(self, input_0, input_1):
primals_1 = self.norm_1.alpha
primals_3 = self.norm_1.bias
primals_5 = self.norm_2.alpha
primals_7 = self.norm_2.bias
primals_4 = self.attn.q_linear.weight
primals_9 = self.attn.q_linear.bias
primals_6 = self.attn.v_linear.weight
primals_12 = self.attn.v_linear.bias
primals_8 = self.attn.k_linear.weight
primals_13 = self.attn.k_linear.bias
primals_11 = self.attn.out.weight
primals_14 = self.attn.out.bias
primals_15 = self.ff.linear_1.weight
primals_16 = self.ff.linear_1.bias
primals_17 = self.ff.linear_2.weight
primals_18 = self.ff.linear_2.bias
primals_2 = input_0
primals_10 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13, primals_14,
primals_15, primals_16, primals_17, primals_18])
return output[0]
|
rcasero/Transformer
|
EncoderLayer
| false
| 4,195
|
[
"Apache-2.0"
] | 0
|
82f51e04f80634d56b134e0ac87f67d6ba8c736a
|
https://github.com/rcasero/Transformer/tree/82f51e04f80634d56b134e0ac87f67d6ba8c736a
|
ResidualBlock
|
import torch
import torch.nn as nn
import torch.nn.functional as F
class ResidualBlock(nn.Module):
"""
Vanilla convolutional residual block from seminal paper by He et al.
Use of instance normalization suggested by Ulyanov et al. in
https://arxiv.org/pdf/1607.08022.pdf%C2%A0%C2%A0%C2%A0%C2%A0.
"""
def __init__(self, filters=128):
super(ResidualBlock, self).__init__()
self.conv1 = nn.Conv2d(filters, filters, (3, 3), padding=(1, 1),
padding_mode='reflect')
self.in_norm1 = nn.InstanceNorm2d(filters, affine=True)
self.conv2 = nn.Conv2d(filters, filters, (3, 3), padding=(1, 1),
padding_mode='reflect')
self.in_norm2 = nn.InstanceNorm2d(filters, affine=True)
def forward(self, x):
a = self.conv1(x)
b = self.in_norm1(a)
c = F.relu(b)
d = self.conv2(c)
e = self.in_norm2(d)
return F.relu(e + x)
def get_inputs():
return [torch.rand([4, 128, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 128
y1 = yindex // 128
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 128 * x2 + 1152 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_reflection_pad2d_1(in_ptr0, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 512
xnumel = 36
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex % 6
x3 = xindex // 6
y4 = yindex
x5 = xindex
y0 = yindex % 128
y1 = yindex // 128
tmp0 = tl.load(in_ptr0 + (15 + -1 * tl_math.abs(-3 + tl_math.abs(-1 +
x2)) + -4 * tl_math.abs(-3 + tl_math.abs(-1 + x3)) + 16 * y4),
xmask & ymask, eviction_policy='evict_last')
tl.store(out_ptr0 + (y0 + 128 * x5 + 4608 * y1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_convolution_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 128
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, None)
@triton.jit
def triton_poi_fused_repeat_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0 % 128, xmask)
tl.store(out_ptr0 + x0, tmp0, xmask)
@triton.jit
def triton_per_fused__native_batch_norm_legit_4(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 512
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (128 * r1 + 2048 * (x0 // 128) + x0 % 128),
xmask, other=0.0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tl.where(xmask, tmp1, 0)
tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp6 = tl.where(xmask, tmp4, 0)
tmp7 = tl.sum(tmp6, 1)[:, None]
tmp8 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp9 = tmp8.to(tl.float32)
tmp10 = tmp7 / tmp9
tmp11 = tmp1 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tl.broadcast_to(tmp12, [XBLOCK, RBLOCK])
tmp15 = tl.where(xmask, tmp13, 0)
tmp16 = tl.sum(tmp15, 1)[:, None]
tmp17 = 16.0
tmp18 = tmp16 / tmp17
tmp19 = 1e-05
tmp20 = tmp18 + tmp19
tmp21 = libdevice.rsqrt(tmp20)
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp21, xmask)
tl.store(out_ptr0 + x0, tmp10, xmask)
@triton.jit
def triton_poi_fused_reflection_pad2d_relu_5(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex % 128
x1 = xindex // 128 % 6
x2 = xindex // 768 % 6
x3 = xindex // 4608
x5 = xindex
tmp0 = tl.load(in_ptr0 + (1920 + x0 + -512 * tl_math.abs(-3 + tl_math.
abs(-1 + x2)) + -128 * tl_math.abs(-3 + tl_math.abs(-1 + x1)) +
2048 * x3), None)
tmp1 = tl.load(in_ptr1 + (x0 + 128 * x3), None, eviction_policy=
'evict_last')
tmp3 = tl.load(in_ptr2 + (x0 + 128 * x3), None, eviction_policy=
'evict_last')
tmp5 = tl.load(in_ptr3 + (x0 + 128 * x3), None, eviction_policy=
'evict_last')
tmp7 = tl.load(in_ptr4 + (x0 + 128 * x3), None, eviction_policy=
'evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tmp9 = tl.full([1], 0, tl.int32)
tmp10 = triton_helpers.maximum(tmp9, tmp8)
tl.store(out_ptr0 + x5, tmp10, None)
@triton.jit
def triton_per_fused__native_batch_norm_legit_add_relu_repeat_threshold_backward_6(
in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0, out_ptr1, out_ptr3,
out_ptr4, out_ptr5, xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 512
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
x0 = xindex
r1 = rindex
x2 = xindex % 128
x3 = xindex // 128
tmp0 = tl.load(in_ptr0 + x0 % 128, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + (128 * r1 + 2048 * (x0 // 128) + x0 % 128),
xmask, other=0.0)
tmp23 = tl.load(in_ptr1 + (x2 + 128 * r1 + 2048 * x3), xmask, other=0.0)
tmp27 = tl.load(in_ptr2 + x2, xmask, eviction_policy='evict_last')
tmp29 = tl.load(in_ptr3 + (r1 + 16 * x0), xmask, other=0.0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tl.where(xmask, tmp2, 0)
tmp5 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
tmp7 = tl.where(xmask, tmp5, 0)
tmp8 = tl.sum(tmp7, 1)[:, None]
tmp9 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp10 = tmp9.to(tl.float32)
tmp11 = tmp8 / tmp10
tmp12 = tmp2 - tmp11
tmp13 = tmp12 * tmp12
tmp14 = tl.broadcast_to(tmp13, [XBLOCK, RBLOCK])
tmp16 = tl.where(xmask, tmp14, 0)
tmp17 = tl.sum(tmp16, 1)[:, None]
tmp18 = 16.0
tmp19 = tmp17 / tmp18
tmp20 = 1e-05
tmp21 = tmp19 + tmp20
tmp22 = libdevice.rsqrt(tmp21)
tmp24 = tmp23 - tmp11
tmp25 = tmp24 * tmp22
tmp26 = tmp25 * tmp0
tmp28 = tmp26 + tmp27
tmp30 = tmp28 + tmp29
tmp31 = tl.full([1, 1], 0, tl.int32)
tmp32 = triton_helpers.maximum(tmp31, tmp30)
tmp33 = 0.0
tmp34 = tmp32 <= tmp33
tl.store(out_ptr0 + x0, tmp0, xmask)
tl.store(out_ptr3 + x0, tmp22, xmask)
tl.store(out_ptr4 + (r1 + 16 * x0), tmp32, xmask)
tl.store(out_ptr5 + (x2 + 128 * r1 + 2048 * x3), tmp34, xmask)
tl.store(out_ptr1 + x0, tmp11, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (128, 128, 3, 3), (1152, 9, 3, 1))
assert_size_stride(primals_2, (128,), (1,))
assert_size_stride(primals_3, (4, 128, 4, 4), (2048, 16, 4, 1))
assert_size_stride(primals_4, (128,), (1,))
assert_size_stride(primals_5, (128,), (1,))
assert_size_stride(primals_6, (128, 128, 3, 3), (1152, 9, 3, 1))
assert_size_stride(primals_7, (128,), (1,))
assert_size_stride(primals_8, (128,), (1,))
assert_size_stride(primals_9, (128,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((128, 128, 3, 3), (1152, 1, 384, 128),
torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(16384, 9)](primals_1, buf0, 16384, 9,
XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((128, 128, 3, 3), (1152, 1, 384, 128),
torch.float32)
triton_poi_fused_0[grid(16384, 9)](primals_6, buf1, 16384, 9,
XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_6
buf2 = empty_strided_cuda((4, 128, 6, 6), (4608, 1, 768, 128),
torch.float32)
triton_poi_fused_reflection_pad2d_1[grid(512, 36)](primals_3, buf2,
512, 36, XBLOCK=32, YBLOCK=32, num_warps=4, num_stages=1)
buf3 = extern_kernels.convolution(buf2, buf0, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf3, (4, 128, 4, 4), (2048, 1, 512, 128))
buf4 = buf3
del buf3
triton_poi_fused_convolution_2[grid(8192)](buf4, primals_2, 8192,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
buf5 = empty_strided_cuda((512,), (1,), torch.float32)
triton_poi_fused_repeat_3[grid(512)](primals_4, buf5, 512, XBLOCK=
256, num_warps=4, num_stages=1)
del primals_4
buf6 = empty_strided_cuda((512,), (1,), torch.float32)
triton_poi_fused_repeat_3[grid(512)](primals_5, buf6, 512, XBLOCK=
256, num_warps=4, num_stages=1)
del primals_5
buf7 = empty_strided_cuda((1, 512, 1, 1), (512, 1, 512, 512), torch
.float32)
buf8 = empty_strided_cuda((1, 512, 1, 1), (512, 1, 512, 512), torch
.float32)
buf10 = buf8
del buf8
triton_per_fused__native_batch_norm_legit_4[grid(512)](buf10, buf4,
buf7, 512, 16, XBLOCK=8, num_warps=2, num_stages=1)
buf11 = empty_strided_cuda((4, 128, 6, 6), (4608, 1, 768, 128),
torch.float32)
triton_poi_fused_reflection_pad2d_relu_5[grid(18432)](buf4, buf7,
buf10, buf5, buf6, buf11, 18432, XBLOCK=256, num_warps=4,
num_stages=1)
buf12 = extern_kernels.convolution(buf11, buf1, stride=(1, 1),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf12, (4, 128, 4, 4), (2048, 1, 512, 128))
buf13 = buf12
del buf12
triton_poi_fused_convolution_2[grid(8192)](buf13, primals_7, 8192,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_7
buf14 = empty_strided_cuda((512,), (1,), torch.float32)
buf15 = empty_strided_cuda((1, 512, 1, 1), (512, 1, 512, 512),
torch.float32)
buf18 = empty_strided_cuda((1, 512, 1, 1), (512, 1, 512, 512),
torch.float32)
buf19 = empty_strided_cuda((4, 128, 4, 4), (2048, 16, 4, 1), torch.
float32)
buf20 = empty_strided_cuda((4, 128, 4, 4), (2048, 1, 512, 128),
torch.bool)
triton_per_fused__native_batch_norm_legit_add_relu_repeat_threshold_backward_6[
grid(512)](primals_8, buf13, primals_9, primals_3, buf14, buf15,
buf18, buf19, buf20, 512, 16, XBLOCK=8, num_warps=2, num_stages=1)
del primals_3
del primals_8
del primals_9
return (buf19, buf0, buf1, buf2, buf4, buf5, buf6, buf7, buf10, buf11,
buf13, buf14, reinterpret_tensor(buf18, (512,), (1,), 0), buf20,
reinterpret_tensor(buf15, (1, 512, 1, 1), (512, 1, 1, 1), 0))
class ResidualBlockNew(nn.Module):
"""
Vanilla convolutional residual block from seminal paper by He et al.
Use of instance normalization suggested by Ulyanov et al. in
https://arxiv.org/pdf/1607.08022.pdf%C2%A0%C2%A0%C2%A0%C2%A0.
"""
def __init__(self, filters=128):
super(ResidualBlockNew, self).__init__()
self.conv1 = nn.Conv2d(filters, filters, (3, 3), padding=(1, 1),
padding_mode='reflect')
self.in_norm1 = nn.InstanceNorm2d(filters, affine=True)
self.conv2 = nn.Conv2d(filters, filters, (3, 3), padding=(1, 1),
padding_mode='reflect')
self.in_norm2 = nn.InstanceNorm2d(filters, affine=True)
def forward(self, input_0):
primals_1 = self.conv1.weight
primals_2 = self.conv1.bias
primals_4 = self.in_norm1.weight
primals_5 = self.in_norm1.bias
primals_6 = self.conv2.weight
primals_7 = self.conv2.bias
primals_8 = self.in_norm2.weight
primals_9 = self.in_norm2.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0]
|
rileypsmith/Fast-Style-Transfer
|
ResidualBlock
| false
| 4,196
|
[
"MIT"
] | 0
|
8b2164f8bc6d63530f914610b6c5c5c1b0f4ffd5
|
https://github.com/rileypsmith/Fast-Style-Transfer/tree/8b2164f8bc6d63530f914610b6c5c5c1b0f4ffd5
|
RegL1
|
import torch
import torch.nn as nn
class RegL1(nn.Module):
"""
Run Regression with L1
"""
def __init__(self, n_input, n_output):
super(RegL1, self).__init__()
self.linear = nn.Linear(n_input, n_output, bias=True)
def forward(self, x, training=True):
self.training = training
x = self.linear(x)
z1 = torch.sum(torch.abs(self.linear.weight))
self.training = True
return x, z1
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'n_input': 4, 'n_output': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused_abs_sum_0(in_ptr0, out_ptr0, xnumel, rnumel, XBLOCK:
tl.constexpr):
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r0 = rindex
tmp0 = tl.load(in_ptr0 + r0, None)
tmp1 = tl_math.abs(tmp0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
tmp4 = tl.sum(tmp2, 1)[:, None]
tl.store(out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp4, None)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (64,
4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0
), alpha=1, beta=1, out=buf0)
del primals_2
buf1 = empty_strided_cuda((), (), torch.float32)
get_raw_stream(0)
triton_per_fused_abs_sum_0[grid(1)](primals_1, buf1, 1, 16, XBLOCK=
1, num_warps=2, num_stages=1)
return reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0
), buf1, primals_1, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0)
class RegL1New(nn.Module):
"""
Run Regression with L1
"""
def __init__(self, n_input, n_output):
super(RegL1New, self).__init__()
self.linear = nn.Linear(n_input, n_output, bias=True)
def forward(self, input_0):
primals_1 = self.linear.weight
primals_2 = self.linear.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0], output[1]
|
rmporsch/ML_genetic_risk
|
RegL1
| false
| 4,197
|
[
"MIT"
] | 0
|
4e1a0510c94260e69f93639ff4104c5f85080d9f
|
https://github.com/rmporsch/ML_genetic_risk/tree/4e1a0510c94260e69f93639ff4104c5f85080d9f
|
DecoderRNN
|
import torch
from torch import nn
import torch.nn.functional as F
class DecoderRNN(nn.Module):
def __init__(self, T, d):
super().__init__()
self.T = T
self.d = d
self.W = nn.Linear(d, d)
self.U = nn.Linear(d, d)
self.V = nn.Linear(d, d)
self.b = nn.Parameter(torch.rand(1, d))
def forward(self, x, h_t):
y_t = None
for t in range(self.T):
a_t = self.b + self.W(h_t) + self.U(x[t])
h_t = torch.tanh(a_t)
o_t = self.V(h_t)
y_t = F.softmax(o_t, 1)
return y_t, h_t
def init_hidden(self):
return torch.zeros(1, self.d)
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'T': 4, 'd': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_add_tanh_0(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2,
in_ptr3, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x3 = xindex
x4 = xindex % 64
tmp0 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_out_ptr0 + x3, xmask)
tmp2 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr2 + x4, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr3 + x0, xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp4 = tmp0 + tmp3
tmp7 = tmp5 + tmp6
tmp8 = tmp4 + tmp7
tmp9 = libdevice.tanh(tmp8)
tl.store(in_out_ptr0 + x3, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x3, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x3, tmp8, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (1, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_5, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 4), (4, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_4, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), out=buf0)
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_5, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), out=buf1)
buf2 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
get_raw_stream(0)
triton_poi_fused_add_tanh_0[grid(256)](buf2, primals_1, primals_3,
buf1, primals_7, 256, XBLOCK=128, num_warps=4, num_stages=1)
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), out=buf3)
buf4 = buf1
del buf1
extern_kernels.mm(reinterpret_tensor(primals_5, (16, 4), (4, 1), 64
), reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), out=buf4)
buf5 = reinterpret_tensor(buf3, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf3
triton_poi_fused_add_tanh_0[grid(256)](buf5, primals_1, primals_3,
buf4, primals_7, 256, XBLOCK=128, num_warps=4, num_stages=1)
buf6 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf5, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), out=buf6)
buf7 = buf4
del buf4
extern_kernels.mm(reinterpret_tensor(primals_5, (16, 4), (4, 1),
128), reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), out=buf7)
buf8 = reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf6
triton_poi_fused_add_tanh_0[grid(256)](buf8, primals_1, primals_3,
buf7, primals_7, 256, XBLOCK=128, num_warps=4, num_stages=1)
buf9 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf8, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), out=buf9)
buf10 = buf7
del buf7
extern_kernels.mm(reinterpret_tensor(primals_5, (16, 4), (4, 1),
192), reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), out=buf10)
del primals_6
buf11 = reinterpret_tensor(buf9, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf9
triton_poi_fused_add_tanh_0[grid(256)](buf11, primals_1, primals_3,
buf10, primals_7, 256, XBLOCK=128, num_warps=4, num_stages=1)
del buf10
del primals_1
del primals_3
del primals_7
buf12 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_9, reinterpret_tensor(buf11, (64, 4),
(4, 1), 0), reinterpret_tensor(primals_8, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf12)
del primals_9
buf13 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__softmax_1[grid(256)](buf12, buf13, 256, XBLOCK=
256, num_warps=4, num_stages=1)
buf14 = reinterpret_tensor(buf12, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf12
triton_poi_fused__softmax_2[grid(256)](buf13, buf14, 256, XBLOCK=
256, num_warps=4, num_stages=1)
del buf13
return buf14, buf11, reinterpret_tensor(primals_4, (64, 4), (4, 1), 0
), reinterpret_tensor(primals_5, (16, 4), (4, 1), 0
), buf2, reinterpret_tensor(primals_5, (16, 4), (4, 1), 64
), buf5, reinterpret_tensor(primals_5, (16, 4), (4, 1), 128
), buf8, reinterpret_tensor(primals_5, (16, 4), (4, 1), 192
), buf11, buf14, primals_8, primals_2
class DecoderRNNNew(nn.Module):
def __init__(self, T, d):
super().__init__()
self.T = T
self.d = d
self.W = nn.Linear(d, d)
self.U = nn.Linear(d, d)
self.V = nn.Linear(d, d)
self.b = nn.Parameter(torch.rand(1, d))
def init_hidden(self):
return torch.zeros(1, self.d)
def forward(self, input_0, input_1):
primals_1 = self.b
primals_2 = self.W.weight
primals_3 = self.W.bias
primals_6 = self.U.weight
primals_7 = self.U.bias
primals_8 = self.V.weight
primals_9 = self.V.bias
primals_4 = input_0
primals_5 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0], output[1]
|
rish-16/SHA-RNN
|
DecoderRNN
| false
| 4,198
|
[
"MIT"
] | 0
|
08c701396217f0b645de043963ff8ec4bf27e835
|
https://github.com/rish-16/SHA-RNN/tree/08c701396217f0b645de043963ff8ec4bf27e835
|
SpatialAttention
|
import torch
import torch.utils.data
import torch
import torch.nn as nn
class SpatialAttention(nn.Module):
def __init__(self):
super(SpatialAttention, self).__init__()
self.conv1 = nn.Conv2d(in_channels=2, out_channels=1, kernel_size=3,
padding=1, bias=False)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
avg_out = torch.mean(x, dim=1, keepdim=True)
max_out, _ = torch.max(x, dim=1, keepdim=True)
temp_x = torch.cat([avg_out, max_out], dim=1)
temp_x = self.conv1(temp_x)
attention = self.sigmoid(temp_x)
x = attention * x
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.utils.data
import torch
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 16 % 2
x0 = xindex % 16
x2 = xindex // 32
x3 = xindex
tmp0 = x1
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + 64 * x2), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp7 = tmp5 + tmp6
tmp8 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp9 = tmp7 + tmp8
tmp10 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = tmp9 + tmp10
tmp12 = 4.0
tmp13 = tmp11 / tmp12
tmp14 = tl.full(tmp13.shape, 0.0, tmp13.dtype)
tmp15 = tl.where(tmp4, tmp13, tmp14)
tmp16 = tmp0 >= tmp3
tl.full([1], 2, tl.int64)
tmp19 = tl.load(in_ptr0 + (x0 + 64 * x2), tmp16 & xmask,
eviction_policy='evict_last', other=0.0)
tmp20 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), tmp16 & xmask,
eviction_policy='evict_last', other=0.0)
tmp21 = triton_helpers.maximum(tmp19, tmp20)
tmp22 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), tmp16 & xmask,
eviction_policy='evict_last', other=0.0)
tmp23 = triton_helpers.maximum(tmp21, tmp22)
tmp24 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), tmp16 & xmask,
eviction_policy='evict_last', other=0.0)
tmp25 = triton_helpers.maximum(tmp23, tmp24)
tmp26 = tl.full(tmp25.shape, 0.0, tmp25.dtype)
tmp27 = tl.where(tmp16, tmp25, tmp26)
tmp28 = tl.where(tmp4, tmp15, tmp27)
tl.store(out_ptr0 + x3, tmp28, xmask)
@triton.jit
def triton_poi_fused_mul_sigmoid_1(in_ptr0, in_ptr1, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 16
x2 = xindex // 64
x3 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr1 + x3, xmask)
tmp1 = tl.sigmoid(tmp0)
tmp3 = tmp1 * tmp2
tl.store(out_ptr0 + x3, tmp3, xmask)
def call(args):
primals_1, primals_2 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (1, 2, 3, 3), (18, 9, 3, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 2, 4, 4), (32, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(128)](primals_1, buf0, 128, XBLOCK=128,
num_warps=4, num_stages=1)
buf1 = extern_kernels.convolution(buf0, primals_2, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 1, 4, 4), (16, 16, 4, 1))
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_mul_sigmoid_1[grid(256)](buf1, primals_1, buf2,
256, XBLOCK=256, num_warps=4, num_stages=1)
return buf2, primals_1, primals_2, buf0, buf1
class SpatialAttentionNew(nn.Module):
def __init__(self):
super(SpatialAttentionNew, self).__init__()
self.conv1 = nn.Conv2d(in_channels=2, out_channels=1, kernel_size=3,
padding=1, bias=False)
self.sigmoid = nn.Sigmoid()
def forward(self, input_0):
primals_2 = self.conv1.weight
primals_1 = input_0
output = call([primals_1, primals_2])
return output[0]
|
robvincen/robot_gradet
|
SpatialAttention
| false
| 4,199
|
[
"BSD-3-Clause"
] | 0
|
a39e3c772c72806dfc99e4d24d8787e0d1bdeef5
|
https://github.com/robvincen/robot_gradet/tree/a39e3c772c72806dfc99e4d24d8787e0d1bdeef5
|
QNet
|
import torch
import torch.nn as nn
class QNet(nn.Module):
def __init__(self, in_size: 'int', out_size: 'int'):
super(QNet, self).__init__()
self.fc1 = nn.Linear(in_size, 16)
self.fc_out = nn.Linear(16, out_size)
self.act = nn.LeakyReLU()
def forward(self, x):
o1 = self.act(self.fc1(x))
o2 = self.act(self.fc_out(o1))
return o2
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_size': 4, 'out_size': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_leaky_relu_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 16
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp5 = 0.01
tmp6 = tmp2 * tmp5
tmp7 = tl.where(tmp4, tmp2, tmp6)
tl.store(out_ptr0 + x2, tmp4, xmask)
tl.store(out_ptr1 + x2, tmp7, xmask)
@triton.jit
def triton_poi_fused_leaky_relu_1(in_ptr0, in_ptr1, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 0.0
tmp4 = tmp2 > tmp3
tmp5 = 0.01
tmp6 = tmp2 * tmp5
tmp7 = tl.where(tmp4, tmp2, tmp6)
tl.store(out_ptr0 + x2, tmp4, xmask)
tl.store(out_ptr1 + x2, tmp7, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (16, 4), (4, 1))
assert_size_stride(primals_2, (16,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 16), (16, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 16), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.bool)
buf2 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.
float32)
get_raw_stream(0)
triton_poi_fused_leaky_relu_0[grid(1024)](buf0, primals_2, buf1,
buf2, 1024, XBLOCK=256, num_warps=4, num_stages=1)
del buf0
del primals_2
buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (64, 16), (16, 1), 0),
reinterpret_tensor(primals_4, (16, 4), (1, 16), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_leaky_relu_1[grid(256)](buf3, primals_5, buf4,
buf5, 256, XBLOCK=128, num_warps=4, num_stages=1)
del buf3
del primals_5
return buf5, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf1, reinterpret_tensor(buf2, (64, 16), (16, 1), 0
), buf4, primals_4
class QNetNew(nn.Module):
def __init__(self, in_size: 'int', out_size: 'int'):
super(QNetNew, self).__init__()
self.fc1 = nn.Linear(in_size, 16)
self.fc_out = nn.Linear(16, out_size)
self.act = nn.LeakyReLU()
def forward(self, input_0):
primals_1 = self.fc1.weight
primals_2 = self.fc1.bias
primals_4 = self.fc_out.weight
primals_5 = self.fc_out.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
rosebin/gymlab
|
QNet
| false
| 4,200
|
[
"BSD-3-Clause"
] | 0
|
de97fc24e0ddf5e328a2aa732cc339b2371d92d1
|
https://github.com/rosebin/gymlab/tree/de97fc24e0ddf5e328a2aa732cc339b2371d92d1
|
L0Linear
|
import torch
import numpy as np
import torch.nn as nn
from torch.nn import functional as F
from torch.autograd import Variable
import logging as lg
def hard_sigmoid(x):
"""Hard Sigmoid function."""
return torch.min(torch.max(x, torch.zeros_like(x)), torch.ones_like(x))
class _L0Norm(nn.Module):
"""L0 norm."""
def __init__(self, origin, loc_mean: 'float'=0.0, loc_sdev: 'float'=
0.01, beta: 'float'=2 / 3, gamma: 'float'=-0.1, zeta: 'float'=1.1,
fix_temp: 'bool'=True, l02: 'bool'=False, l02_alpha=0.01):
"""Class of layers using L0 Norm.
:param origin: original layer such as nn.Linear(..), nn.Conv2d(..)
:param loc_mean: mean of the normal of initial location parameters
:param loc_sdev: standard deviation of initial location parameters
:param beta: initial temperature parameter
:param gamma: lower bound of "stretched" s
:param zeta: upper bound of "stretched" s
:param fix_temp: True if temperature is fixed
"""
super(_L0Norm, self).__init__()
self._origin = origin
self._size = self._origin.weight.size()
self.loc = nn.Parameter(torch.zeros(self._size).normal_(loc_mean,
loc_sdev))
self.temp = beta if fix_temp else nn.Parameter(torch.zeros(1).fill_
(beta))
self.register_buffer('uniform', torch.zeros(self._size))
self.gamma = gamma
self.zeta = zeta
self.gamma_zeta_ratio = np.log(-gamma / zeta)
self.l02_alpha = l02_alpha
self.l02 = l02
if self.l02:
assert self.l02_alpha > 0
lg.info('Using L_0,2 norm')
def _get_mask(self):
if self.training:
self.uniform.uniform_()
u = Variable(self.uniform)
s = F.sigmoid((torch.log(u) - torch.log(1 - u) + self.loc) /
self.temp)
s = s * (self.zeta - self.gamma) + self.gamma
penalty = F.sigmoid(self.loc - self.temp * self.gamma_zeta_ratio
).sum()
if self.l02:
l02Norm = (F.sigmoid(self.loc - self.temp * self.
gamma_zeta_ratio) * self._origin.weight ** 2).sum()
penalty = penalty + self.l02_alpha * l02Norm
else:
s = F.sigmoid(self.loc) * (self.zeta - self.gamma) + self.gamma
penalty = 0
return hard_sigmoid(s), penalty
class L0Linear(_L0Norm):
"""Linear model with L0 norm."""
def __init__(self, in_features: 'int', out_features: 'int', bias:
'bool'=True, **kwargs):
"""Linear model with L0 norm."""
super(L0Linear, self).__init__(nn.Linear(in_features, out_features,
bias=bias), **kwargs)
def forward(self, input):
"""Forward function with mask and penalty."""
mask, penalty = self._get_mask()
out = F.linear(input, self._origin.weight * mask, self._origin.bias)
return out, penalty
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_features': 4, 'out_features': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import numpy as np
import torch.nn as nn
from torch.nn import functional as F
from torch.autograd import Variable
import logging as lg
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_add_maximum_minimum_mul_ones_like_sigmoid_zeros_like_0(
in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask)
tmp2 = tl.sigmoid(tmp1)
tmp3 = 1.2000000000000002
tmp4 = tmp2 * tmp3
tmp5 = -0.1
tmp6 = tmp4 + tmp5
tmp7 = 0.0
tmp8 = triton_helpers.maximum(tmp6, tmp7)
tmp9 = 1.0
tmp10 = triton_helpers.minimum(tmp8, tmp9)
tmp11 = tmp0 * tmp10
tl.store(out_ptr0 + x0, tmp11, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_maximum_minimum_mul_ones_like_sigmoid_zeros_like_0[
grid(16)](primals_2, primals_1, buf0, 16, XBLOCK=16, num_warps=
1, num_stages=1)
buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_3, reinterpret_tensor(primals_4, (64,
4), (4, 1), 0), reinterpret_tensor(buf0, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf1)
del buf0
del primals_3
return reinterpret_tensor(buf1, (4, 4, 4, 4), (64, 16, 4, 1), 0
), primals_1, primals_2, reinterpret_tensor(primals_4, (64, 4), (4,
1), 0)
def hard_sigmoid(x):
"""Hard Sigmoid function."""
return torch.min(torch.max(x, torch.zeros_like(x)), torch.ones_like(x))
class _L0Norm(nn.Module):
"""L0 norm."""
def __init__(self, origin, loc_mean: 'float'=0.0, loc_sdev: 'float'=
0.01, beta: 'float'=2 / 3, gamma: 'float'=-0.1, zeta: 'float'=1.1,
fix_temp: 'bool'=True, l02: 'bool'=False, l02_alpha=0.01):
"""Class of layers using L0 Norm.
:param origin: original layer such as nn.Linear(..), nn.Conv2d(..)
:param loc_mean: mean of the normal of initial location parameters
:param loc_sdev: standard deviation of initial location parameters
:param beta: initial temperature parameter
:param gamma: lower bound of "stretched" s
:param zeta: upper bound of "stretched" s
:param fix_temp: True if temperature is fixed
"""
super(_L0Norm, self).__init__()
self._origin = origin
self._size = self._origin.weight.size()
self.loc = nn.Parameter(torch.zeros(self._size).normal_(loc_mean,
loc_sdev))
self.temp = beta if fix_temp else nn.Parameter(torch.zeros(1).fill_
(beta))
self.register_buffer('uniform', torch.zeros(self._size))
self.gamma = gamma
self.zeta = zeta
self.gamma_zeta_ratio = np.log(-gamma / zeta)
self.l02_alpha = l02_alpha
self.l02 = l02
if self.l02:
assert self.l02_alpha > 0
lg.info('Using L_0,2 norm')
def _get_mask(self):
if self.training:
self.uniform.uniform_()
u = Variable(self.uniform)
s = F.sigmoid((torch.log(u) - torch.log(1 - u) + self.loc) /
self.temp)
s = s * (self.zeta - self.gamma) + self.gamma
penalty = F.sigmoid(self.loc - self.temp * self.gamma_zeta_ratio
).sum()
if self.l02:
l02Norm = (F.sigmoid(self.loc - self.temp * self.
gamma_zeta_ratio) * self._origin.weight ** 2).sum()
penalty = penalty + self.l02_alpha * l02Norm
else:
s = F.sigmoid(self.loc) * (self.zeta - self.gamma) + self.gamma
penalty = 0
return hard_sigmoid(s), penalty
class L0LinearNew(_L0Norm):
"""Linear model with L0 norm."""
def __init__(self, in_features: 'int', out_features: 'int', bias:
'bool'=True, **kwargs):
"""Linear model with L0 norm."""
super(L0LinearNew, self).__init__(nn.Linear(in_features,
out_features, bias=bias), **kwargs)
def forward(self, input_0):
primals_1 = self.loc
primals_2 = self._origin.weight
primals_3 = self._origin.bias
primals_4 = input_0
output = call([primals_1, primals_2, primals_3, primals_4])
return output[0], output[1]
|
rmporsch/ML_genetic_risk
|
L0Linear
| false
| 4,201
|
[
"MIT"
] | 0
|
4e1a0510c94260e69f93639ff4104c5f85080d9f
|
https://github.com/rmporsch/ML_genetic_risk/tree/4e1a0510c94260e69f93639ff4104c5f85080d9f
|
QNetwork
|
import torch
import torch.nn.functional as F
import torch.nn as nn
class QNetwork(nn.Module):
def __init__(self, state_size, action_size, seed):
super(QNetwork, self).__init__()
self.seed = torch.manual_seed(seed)
self.fc1 = nn.Linear(state_size, 128)
self.fc2 = nn.Linear(128, 64)
self.fc3 = nn.Linear(64, action_size)
def forward(self, state):
x = F.relu(self.fc1(state))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'state_size': 4, 'action_size': 4, 'seed': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_relu_threshold_backward_0(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 128
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, None)
tl.store(out_ptr0 + x2, tmp6, None)
@triton.jit
def triton_poi_fused_relu_threshold_backward_1(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 64
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, None)
tl.store(out_ptr0 + x2, tmp6, None)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (128, 4), (4, 1))
assert_size_stride(primals_2, (128,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (64, 128), (128, 1))
assert_size_stride(primals_5, (64,), (1,))
assert_size_stride(primals_6, (4, 64), (64, 1))
assert_size_stride(primals_7, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 128), (128, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 128), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 128), (2048, 512, 128, 1), 0)
del buf0
buf6 = empty_strided_cuda((4, 4, 4, 128), (2048, 512, 128, 1),
torch.bool)
get_raw_stream(0)
triton_poi_fused_relu_threshold_backward_0[grid(8192)](buf1,
primals_2, buf6, 8192, XBLOCK=256, num_warps=4, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((64, 64), (64, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf1, (64, 128), (128, 1), 0),
reinterpret_tensor(primals_4, (128, 64), (1, 128), 0), out=buf2)
buf3 = reinterpret_tensor(buf2, (4, 4, 4, 64), (1024, 256, 64, 1), 0)
del buf2
buf5 = empty_strided_cuda((4, 4, 4, 64), (1024, 256, 64, 1), torch.bool
)
triton_poi_fused_relu_threshold_backward_1[grid(4096)](buf3,
primals_5, buf5, 4096, XBLOCK=256, num_warps=4, num_stages=1)
del primals_5
buf4 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_7, reinterpret_tensor(buf3, (64, 64),
(64, 1), 0), reinterpret_tensor(primals_6, (64, 4), (1, 64), 0),
alpha=1, beta=1, out=buf4)
del primals_7
return reinterpret_tensor(buf4, (4, 4, 4, 4), (64, 16, 4, 1), 0
), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), reinterpret_tensor(buf1, (64, 128), (128, 1), 0
), reinterpret_tensor(buf3, (64, 64), (64, 1), 0
), primals_6, buf5, primals_4, buf6
class QNetworkNew(nn.Module):
def __init__(self, state_size, action_size, seed):
super(QNetworkNew, self).__init__()
self.seed = torch.manual_seed(seed)
self.fc1 = nn.Linear(state_size, 128)
self.fc2 = nn.Linear(128, 64)
self.fc3 = nn.Linear(64, action_size)
def forward(self, input_0):
primals_1 = self.fc1.weight
primals_2 = self.fc1.bias
primals_4 = self.fc2.weight
primals_5 = self.fc2.bias
primals_6 = self.fc3.weight
primals_7 = self.fc3.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
|
royveshovda/deep-reinforcement-learning
|
QNetwork
| false
| 4,202
|
[
"MIT"
] | 0
|
64ba7ef5ab44f095b7e8b29f6c4ff1585025981a
|
https://github.com/royveshovda/deep-reinforcement-learning/tree/64ba7ef5ab44f095b7e8b29f6c4ff1585025981a
|
Discriminator
|
import torch
import torch.nn as nn
class Discriminator(nn.Module):
def __init__(self, state_dim, action_dim):
super(Discriminator, self).__init__()
self.l1 = nn.Linear(state_dim + action_dim, 500)
self.l2 = nn.Linear(500, 300)
self.l3 = nn.Linear(300, 300)
self.l4 = nn.Linear(300, 1)
def forward(self, state, action):
state_action = torch.cat([state, action], 1)
x = torch.tanh(self.l1(state_action))
x = torch.tanh(self.l2(x))
x = torch.tanh(self.l3(x))
x = torch.sigmoid(self.l4(x))
return x
def get_inputs():
return [torch.rand([4, 4]), torch.rand([4, 4])]
def get_init_inputs():
return [[], {'state_dim': 4, 'action_dim': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 32
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 8
x1 = xindex // 8
x2 = xindex
tmp0 = x0
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 4, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (4 * x1 + x0), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tmp0 >= tmp3
tl.full([1], 8, tl.int64)
tmp9 = tl.load(in_ptr1 + (4 * x1 + (-4 + x0)), tmp6 & xmask,
eviction_policy='evict_last', other=0.0)
tmp10 = tl.where(tmp4, tmp5, tmp9)
tl.store(out_ptr0 + x2, tmp10, xmask)
@triton.jit
def triton_poi_fused_tanh_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 2000
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 500
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = libdevice.tanh(tmp2)
tl.store(in_out_ptr0 + x2, tmp3, xmask)
@triton.jit
def triton_poi_fused_tanh_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 1200
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 300
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = libdevice.tanh(tmp2)
tl.store(in_out_ptr0 + x2, tmp3, xmask)
@triton.jit
def triton_poi_fused_sigmoid_3(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 4
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr0 + 0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK])
tmp3 = tmp0 + tmp2
tmp4 = tl.sigmoid(tmp3)
tl.store(in_out_ptr0 + x0, tmp4, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (500, 8), (8, 1))
assert_size_stride(primals_4, (500,), (1,))
assert_size_stride(primals_5, (300, 500), (500, 1))
assert_size_stride(primals_6, (300,), (1,))
assert_size_stride(primals_7, (300, 300), (300, 1))
assert_size_stride(primals_8, (300,), (1,))
assert_size_stride(primals_9, (1, 300), (300, 1))
assert_size_stride(primals_10, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(32)](primals_1, primals_2, buf0, 32,
XBLOCK=32, num_warps=1, num_stages=1)
del primals_1
del primals_2
buf1 = empty_strided_cuda((4, 500), (500, 1), torch.float32)
extern_kernels.mm(buf0, reinterpret_tensor(primals_3, (8, 500), (1,
8), 0), out=buf1)
del primals_3
buf2 = buf1
del buf1
triton_poi_fused_tanh_1[grid(2000)](buf2, primals_4, 2000, XBLOCK=
128, num_warps=4, num_stages=1)
del primals_4
buf3 = empty_strided_cuda((4, 300), (300, 1), torch.float32)
extern_kernels.mm(buf2, reinterpret_tensor(primals_5, (500, 300), (
1, 500), 0), out=buf3)
buf4 = buf3
del buf3
triton_poi_fused_tanh_2[grid(1200)](buf4, primals_6, 1200, XBLOCK=
128, num_warps=4, num_stages=1)
del primals_6
buf5 = empty_strided_cuda((4, 300), (300, 1), torch.float32)
extern_kernels.mm(buf4, reinterpret_tensor(primals_7, (300, 300), (
1, 300), 0), out=buf5)
buf6 = buf5
del buf5
triton_poi_fused_tanh_2[grid(1200)](buf6, primals_8, 1200, XBLOCK=
128, num_warps=4, num_stages=1)
del primals_8
buf7 = empty_strided_cuda((4, 1), (1, 1), torch.float32)
extern_kernels.mm(buf6, reinterpret_tensor(primals_9, (300, 1), (1,
300), 0), out=buf7)
buf8 = buf7
del buf7
triton_poi_fused_sigmoid_3[grid(4)](buf8, primals_10, 4, XBLOCK=4,
num_warps=1, num_stages=1)
del primals_10
return buf8, buf0, buf2, buf4, buf6, buf8, primals_9, primals_7, primals_5
class DiscriminatorNew(nn.Module):
def __init__(self, state_dim, action_dim):
super(DiscriminatorNew, self).__init__()
self.l1 = nn.Linear(state_dim + action_dim, 500)
self.l2 = nn.Linear(500, 300)
self.l3 = nn.Linear(300, 300)
self.l4 = nn.Linear(300, 1)
def forward(self, input_0, input_1):
primals_3 = self.l1.weight
primals_4 = self.l1.bias
primals_5 = self.l2.weight
primals_6 = self.l2.bias
primals_7 = self.l3.weight
primals_8 = self.l3.bias
primals_9 = self.l4.weight
primals_10 = self.l4.bias
primals_1 = input_0
primals_2 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9, primals_10])
return output[0]
|
rortiz9/meleeml
|
Discriminator
| false
| 4,203
|
[
"MIT"
] | 0
|
9be4bf53a377dfb46dbb3b51f102f1bffc0124d2
|
https://github.com/rortiz9/meleeml/tree/9be4bf53a377dfb46dbb3b51f102f1bffc0124d2
|
RelationalTransformerEncoderLayer
|
import torch
import warnings
from torch import Tensor
from torch.nn import TransformerEncoderLayer
from torch.nn.functional import *
from torch.nn.modules.activation import MultiheadAttention
from torch.nn.modules.activation import xavier_uniform_
from torch.nn.modules.activation import xavier_normal_
from torch.nn.modules.activation import constant_
from torch.nn.modules.activation import Parameter
from typing import Optional
import torch.utils.data.dataset
from typing import Tuple
def relational_multi_head_attention_forward(query: 'Tensor', key: 'Tensor',
value: 'Tensor', relation: 'Tensor', embed_dim_to_check: 'int',
num_heads: 'int', in_proj_weight: 'Tensor', in_proj_bias: 'Tensor',
bias_k: 'Optional[Tensor]', bias_v: 'Optional[Tensor]', add_zero_attn:
'bool', dropout_p: 'float', out_proj_weight: 'Tensor', out_proj_bias:
'Tensor', training: 'bool'=True, key_padding_mask: 'Optional[Tensor]'=
None, need_weights: 'bool'=True, attn_mask: 'Optional[Tensor]'=None,
use_separate_proj_weight: 'bool'=False, q_proj_weight:
'Optional[Tensor]'=None, k_proj_weight: 'Optional[Tensor]'=None,
v_proj_weight: 'Optional[Tensor]'=None, static_k: 'Optional[Tensor]'=
None, static_v: 'Optional[Tensor]'=None, relation_type: 'str'=None
) ->Tuple[Tensor, Optional[Tensor]]:
"""
Args:
query, key, value: map a query and a set of key-value pairs to an output.
See "Attention Is All You Need" for more details.
relation: relation between queries and keys.
embed_dim_to_check: total dimension of the model.
num_heads: parallel attention heads.
in_proj_weight, in_proj_bias: input projection weight and bias.
bias_k, bias_v: bias of the key and value sequences to be added at dim=0.
add_zero_attn: add a new batch of zeros to the key and
value sequences at dim=1.
dropout_p: probability of an element to be zeroed.
out_proj_weight, out_proj_bias: the output projection weight and bias.
training: apply dropout if is ``True``.
key_padding_mask: if provided, specified padding elements in the key will
be ignored by the attention. This is an binary mask. When the value is True,
the corresponding value on the attention layer will be filled with -inf.
need_weights: output attn_output_weights.
attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
the batches while a 3D mask allows to specify a different mask for the entries of each batch.
use_separate_proj_weight: the function accept the proj. weights for query, key,
and value in different forms. If false, in_proj_weight will be used, which is
a combination of q_proj_weight, k_proj_weight, v_proj_weight.
q_proj_weight, k_proj_weight, v_proj_weight, in_proj_bias: input projection weight and bias.
static_k, static_v: static key and value used for attention operators.
Shape:
Inputs:
- query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
the embedding dimension.
- key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- relation: :math:`(L, S, N, E)` where L is the target sequence length, where S is the source sequence length,
N is the batch size, E is the embedding dimension.
- key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
will be unchanged. If a BoolTensor is provided, the positions with the
value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
- attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
is provided, it will be added to the attention weight.
- static_k: :math:`(N*num_heads, S, E/num_heads)`, where S is the source sequence length,
N is the batch size, E is the embedding dimension. E/num_heads is the head dimension.
- static_v: :math:`(N*num_heads, S, E/num_heads)`, where S is the source sequence length,
N is the batch size, E is the embedding dimension. E/num_heads is the head dimension.
Outputs:
- attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
E is the embedding dimension.
- attn_output_weights: :math:`(N, L, S)` where N is the batch size,
L is the target sequence length, S is the source sequence length.
"""
tens_ops = (query, key, value, in_proj_weight, in_proj_bias, bias_k,
bias_v, out_proj_weight, out_proj_bias)
if has_torch_function(tens_ops):
return handle_torch_function(multi_head_attention_forward, tens_ops,
query, key, value, embed_dim_to_check, num_heads,
in_proj_weight, in_proj_bias, bias_k, bias_v, add_zero_attn,
dropout_p, out_proj_weight, out_proj_bias, training=training,
key_padding_mask=key_padding_mask, need_weights=need_weights,
attn_mask=attn_mask, use_separate_proj_weight=
use_separate_proj_weight, q_proj_weight=q_proj_weight,
k_proj_weight=k_proj_weight, v_proj_weight=v_proj_weight,
static_k=static_k, static_v=static_v)
tgt_len, bsz, embed_dim = query.size()
assert embed_dim == embed_dim_to_check
assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
head_dim = embed_dim // num_heads
assert head_dim * num_heads == embed_dim, 'embed_dim must be divisible by num_heads'
scaling = float(head_dim) ** -0.5
if not use_separate_proj_weight:
if (query is key or torch.equal(query, key)) and (key is value or
torch.equal(key, value)):
q, k, v = linear(query, in_proj_weight, in_proj_bias).chunk(3,
dim=-1)
elif key is value or torch.equal(key, value):
_b = in_proj_bias
_start = 0
_end = embed_dim
_w = in_proj_weight[_start:_end, :]
if _b is not None:
_b = _b[_start:_end]
q = linear(query, _w, _b)
if key is None:
assert value is None
k = None
v = None
else:
_b = in_proj_bias
_start = embed_dim
_end = None
_w = in_proj_weight[_start:, :]
if _b is not None:
_b = _b[_start:]
k, v = linear(key, _w, _b).chunk(2, dim=-1)
else:
_b = in_proj_bias
_start = 0
_end = embed_dim
_w = in_proj_weight[_start:_end, :]
if _b is not None:
_b = _b[_start:_end]
q = linear(query, _w, _b)
_b = in_proj_bias
_start = embed_dim
_end = embed_dim * 2
_w = in_proj_weight[_start:_end, :]
if _b is not None:
_b = _b[_start:_end]
k = linear(key, _w, _b)
_b = in_proj_bias
_start = embed_dim * 2
_end = None
_w = in_proj_weight[_start:, :]
if _b is not None:
_b = _b[_start:]
v = linear(value, _w, _b)
else:
q_proj_weight_non_opt = torch.jit._unwrap_optional(q_proj_weight)
len1, len2 = q_proj_weight_non_opt.size()
assert len1 == embed_dim and len2 == query.size(-1)
k_proj_weight_non_opt = torch.jit._unwrap_optional(k_proj_weight)
len1, len2 = k_proj_weight_non_opt.size()
assert len1 == embed_dim and len2 == key.size(-1)
v_proj_weight_non_opt = torch.jit._unwrap_optional(v_proj_weight)
len1, len2 = v_proj_weight_non_opt.size()
assert len1 == embed_dim and len2 == value.size(-1)
if in_proj_bias is not None:
q = linear(query, q_proj_weight_non_opt, in_proj_bias[0:embed_dim])
k = linear(key, k_proj_weight_non_opt, in_proj_bias[embed_dim:
embed_dim * 2])
v = linear(value, v_proj_weight_non_opt, in_proj_bias[embed_dim *
2:])
else:
q = linear(query, q_proj_weight_non_opt, in_proj_bias)
k = linear(key, k_proj_weight_non_opt, in_proj_bias)
v = linear(value, v_proj_weight_non_opt, in_proj_bias)
q = q * scaling
if attn_mask is not None:
assert attn_mask.dtype == torch.float32 or attn_mask.dtype == torch.float64 or attn_mask.dtype == torch.float16 or attn_mask.dtype == torch.uint8 or attn_mask.dtype == torch.bool, 'Only float, byte, and bool types are supported for attn_mask, not {}'.format(
attn_mask.dtype)
if attn_mask.dtype == torch.uint8:
warnings.warn(
'Byte tensor for attn_mask in nn.MultiheadAttention is deprecated. Use bool tensor instead.'
)
attn_mask = attn_mask
if attn_mask.dim() == 2:
attn_mask = attn_mask.unsqueeze(0)
if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
raise RuntimeError(
'The size of the 2D attn_mask is not correct.')
elif attn_mask.dim() == 3:
if list(attn_mask.size()) != [bsz * num_heads, query.size(0),
key.size(0)]:
raise RuntimeError(
'The size of the 3D attn_mask is not correct.')
else:
raise RuntimeError("attn_mask's dimension {} is not supported".
format(attn_mask.dim()))
if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
warnings.warn(
'Byte tensor for key_padding_mask in nn.MultiheadAttention is deprecated. Use bool tensor instead.'
)
key_padding_mask = key_padding_mask
if bias_k is not None and bias_v is not None:
if static_k is None and static_v is None:
k = torch.cat([k, bias_k.repeat(1, bsz, 1)])
v = torch.cat([v, bias_v.repeat(1, bsz, 1)])
if attn_mask is not None:
attn_mask = pad(attn_mask, (0, 1))
if key_padding_mask is not None:
key_padding_mask = pad(key_padding_mask, (0, 1))
else:
assert static_k is None, 'bias cannot be added to static key.'
assert static_v is None, 'bias cannot be added to static value.'
else:
assert bias_k is None
assert bias_v is None
r = None
q = q.contiguous().view(tgt_len, bsz * num_heads, head_dim).transpose(0, 1)
if k is not None:
k = k.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
if v is not None:
v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
if static_k is not None:
assert static_k.size(0) == bsz * num_heads
assert static_k.size(2) == head_dim
k = static_k
if static_v is not None:
assert static_v.size(0) == bsz * num_heads
assert static_v.size(2) == head_dim
v = static_v
src_len = k.size(1)
if relation is not None:
if relation_type == 'qk+r':
r = relation.contiguous().view(tgt_len, -1, bsz * num_heads, 1
).squeeze(3).permute(2, 0, 1)
elif relation_type == 'q(k+r)':
r = relation.contiguous().view(tgt_len, src_len, bsz *
num_heads, head_dim).permute(2, 0, 1)
r = r.view(bsz * num_heads, tgt_len * src_len, head_dim)
r = torch.bmm(r, q.unsqueeze(2).repeat(1, 1, src_len, 1).view(
bsz * num_heads, tgt_len * src_len, head_dim).transpose(1, 2)
).view(bsz * num_heads, tgt_len, src_len)
if key_padding_mask is not None:
assert key_padding_mask.size(0) == bsz
assert key_padding_mask.size(1) == src_len
if add_zero_attn:
src_len += 1
k = torch.cat([k, torch.zeros((k.size(0), 1) + k.size()[2:], dtype=
k.dtype, device=k.device)], dim=1)
v = torch.cat([v, torch.zeros((v.size(0), 1) + v.size()[2:], dtype=
v.dtype, device=v.device)], dim=1)
if attn_mask is not None:
attn_mask = pad(attn_mask, (0, 1))
if key_padding_mask is not None:
key_padding_mask = pad(key_padding_mask, (0, 1))
attn_output_weights = torch.bmm(q, k.transpose(1, 2))
if r is not None:
attn_output_weights = attn_output_weights + r
assert list(attn_output_weights.size()) == [bsz * num_heads, tgt_len,
src_len]
if attn_mask is not None:
if attn_mask.dtype == torch.bool:
attn_output_weights.masked_fill_(attn_mask, float('-inf'))
else:
attn_output_weights += attn_mask
if key_padding_mask is not None:
attn_output_weights = attn_output_weights.view(bsz, num_heads,
tgt_len, src_len)
attn_output_weights = attn_output_weights.masked_fill(key_padding_mask
.unsqueeze(1).unsqueeze(2), float('-inf'))
attn_output_weights = attn_output_weights.view(bsz * num_heads,
tgt_len, src_len)
attn_output_weights = softmax(attn_output_weights, dim=-1)
attn_output_weights = dropout(attn_output_weights, p=dropout_p,
training=training)
attn_output = torch.bmm(attn_output_weights, v)
assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len,
bsz, embed_dim)
attn_output = linear(attn_output, out_proj_weight, out_proj_bias)
if need_weights:
attn_output_weights = attn_output_weights.view(bsz, num_heads,
tgt_len, src_len)
return attn_output, attn_output_weights.sum(dim=1) / num_heads
else:
return attn_output, None
class RelationalMultiheadAttention(MultiheadAttention):
"""Allows the model to jointly attend to information
from different representation subspaces.
See `Attention Is All You Need <https://arxiv.org/abs/1706.03762>`_
.. math::
\\text{MultiHead}(Q, K, V) = \\text{Concat}(head_1,\\dots,head_h)W^O
where :math:`head_i = \\text{Attention}(QW_i^Q, KW_i^K, VW_i^V)`.
Args:
embed_dim: total dimension of the model.
num_heads: parallel attention heads.
dropout: a Dropout layer on attn_output_weights. Default: 0.0.
bias: add bias as module parameter. Default: True.
add_bias_kv: add bias to the key and value sequences at dim=0.
add_zero_attn: add a new batch of zeros to the key and
value sequences at dim=1.
kdim: total number of features in key. Default: None.
vdim: total number of features in value. Default: None.
Note that if :attr:`kdim` and :attr:`vdim` are None, they will be set
to :attr:`embed_dim` such that query, key, and value have the same
number of features.
Examples::
>>> realational_multihead_attn = RelationalMultiheadAttention(embed_dim, num_heads)
>>> attn_output, attn_output_weights = realational_multihead_attn(query, key, value)
"""
bias_k: 'Optional[torch.Tensor]'
bias_v: 'Optional[torch.Tensor]'
def __init__(self, embed_dim, num_heads, dropout=0.0, bias=True,
add_bias_kv=False, add_zero_attn=False, kdim=None, vdim=None,
add_relation=False, rdim=None, relation_type=None):
super(RelationalMultiheadAttention, self).__init__(embed_dim=
embed_dim, num_heads=num_heads, dropout=dropout, bias=bias,
add_bias_kv=add_bias_kv, add_zero_attn=add_zero_attn, kdim=kdim,
vdim=vdim)
self.add_relation = add_relation
self.rdim = rdim if rdim is not None else embed_dim
self.relation_type = relation_type if relation_type else 'qk+r'
if self.add_relation:
if relation_type == 'qk+r':
self.r_proj_weight = Parameter(torch.Tensor(num_heads, self
.rdim))
self.r_proj_bias = Parameter(torch.empty(num_heads))
elif relation_type == 'q(k+r)':
self.r_proj_weight = Parameter(torch.Tensor(embed_dim, self
.rdim))
self.r_proj_bias = Parameter(torch.empty(embed_dim))
self._reset_parameters()
def _reset_parameters(self):
if self._qkv_same_embed_dim:
xavier_uniform_(self.in_proj_weight)
else:
xavier_uniform_(self.q_proj_weight)
xavier_uniform_(self.k_proj_weight)
xavier_uniform_(self.v_proj_weight)
if hasattr(self, 'add_relation') and self.add_relation:
xavier_uniform_(self.r_proj_weight)
constant_(self.r_proj_bias, 0.0)
if self.in_proj_bias is not None:
constant_(self.in_proj_bias, 0.0)
constant_(self.out_proj.bias, 0.0)
if self.bias_k is not None:
xavier_normal_(self.bias_k)
if self.bias_v is not None:
xavier_normal_(self.bias_v)
def __setstate__(self, state):
if '_qkv_same_embed_dim' not in state:
state['_qkv_same_embed_dim'] = True
super(MultiheadAttention, self).__setstate__(state)
def forward(self, query: 'Tensor', key: 'Tensor', value: 'Tensor',
relation_dict=None, key_padding_mask: 'Optional[Tensor]'=None,
need_weights: 'bool'=True, attn_mask: 'Optional[Tensor]'=None) ->Tuple[
Tensor, Optional[Tensor]]:
"""
Args:
query, key, value: map a query and a set of key-value pairs to an output.
See "Attention Is All You Need" for more details.
key_padding_mask: if provided, specified padding elements in the key will
be ignored by the attention. When given a binary mask and a value is True,
the corresponding value on the attention layer will be ignored. When given
a byte mask and a value is non-zero, the corresponding value on the attention
layer will be ignored
need_weights: output attn_output_weights.
attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
the batches while a 3D mask allows to specify a different mask for the entries of each batch.
Shapes for inputs:
- query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
the embedding dimension.
- key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
If a ByteTensor is provided, the non-zero positions will be ignored while the position
with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
- attn_mask: if a 2D mask: :math:`(L, S)` where L is the target sequence length, S is the
source sequence length.
If a 3D mask: :math:`(N\\cdot\\text{num\\_heads}, L, S)` where N is the batch size, L is the target sequence
length, S is the source sequence length. ``attn_mask`` ensure that position i is allowed to attend
the unmasked positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
is provided, it will be added to the attention weight.
Shapes for outputs:
- attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
E is the embedding dimension.
- attn_output_weights: :math:`(N, L, S)` where N is the batch size,
L is the target sequence length, S is the source sequence length.
"""
if relation_dict is not None:
relation_labels = relation_dict['relation_labels']
relation_ids = relation_dict['relation_ids']
batch_index = relation_dict['batch_index']
pad_embedding = relation_dict['pad_embedding']
relation_labels = linear(relation_labels, self.r_proj_weight,
self.r_proj_bias)
pad_embedding = linear(pad_embedding.unsqueeze(0), self.
r_proj_weight, self.r_proj_bias).squeeze()
tgt_length, bsz, _ = query.size()
src_length, _, _ = key.size()
relation = pad_embedding.view(1, 1, 1, -1).repeat(bsz,
tgt_length, src_length, 1)
relation[batch_index, relation_ids[:, :, 0], relation_ids[:, :, 1]
] = relation_labels
relation = relation.permute(1, 2, 0, 3)
if not self._qkv_same_embed_dim:
return relational_multi_head_attention_forward(query, key,
value, relation, self.embed_dim, self.num_heads, self.
in_proj_weight, self.in_proj_bias, self.bias_k, self.
bias_v, self.add_zero_attn, self.dropout, self.out_proj
.weight, self.out_proj.bias, training=self.training,
key_padding_mask=key_padding_mask, need_weights=
need_weights, attn_mask=attn_mask,
use_separate_proj_weight=True, q_proj_weight=self.
q_proj_weight, k_proj_weight=self.k_proj_weight,
v_proj_weight=self.v_proj_weight)
else:
return relational_multi_head_attention_forward(query, key,
value, relation, self.embed_dim, self.num_heads, self.
in_proj_weight, self.in_proj_bias, self.bias_k, self.
bias_v, self.add_zero_attn, self.dropout, self.out_proj
.weight, self.out_proj.bias, training=self.training,
key_padding_mask=key_padding_mask, need_weights=
need_weights, attn_mask=attn_mask)
elif not self._qkv_same_embed_dim:
return relational_multi_head_attention_forward(query, key,
value, None, self.embed_dim, self.num_heads, self.
in_proj_weight, self.in_proj_bias, self.bias_k, self.bias_v,
self.add_zero_attn, self.dropout, self.out_proj.weight,
self.out_proj.bias, training=self.training,
key_padding_mask=key_padding_mask, need_weights=
need_weights, attn_mask=attn_mask, use_separate_proj_weight
=True, q_proj_weight=self.q_proj_weight, k_proj_weight=self
.k_proj_weight, v_proj_weight=self.v_proj_weight)
else:
return relational_multi_head_attention_forward(query, key,
value, None, self.embed_dim, self.num_heads, self.
in_proj_weight, self.in_proj_bias, self.bias_k, self.bias_v,
self.add_zero_attn, self.dropout, self.out_proj.weight,
self.out_proj.bias, training=self.training,
key_padding_mask=key_padding_mask, need_weights=
need_weights, attn_mask=attn_mask)
class RelationalTransformerEncoderLayer(TransformerEncoderLayer):
"""TransformerEncoderLayer is made up of self-attn and feedforward network.
This standard encoder layer is based on the paper "Attention Is All You Need".
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in
Neural Information Processing Systems, pages 6000-6010. Users may modify or implement
in a different way during application.
Args:
d_model: the number of expected features in the input (required).
nhead: the number of heads in the multiheadattention models (required).
dim_feedforward: the dimension of the feedforward network model (default=2048).
dropout: the dropout value (default=0.1).
activation: the activation function of intermediate layer, relu or gelu (default=relu).
Examples::
>>> encoder_layer = RelationalTransformerEncoderLayer(d_model=512, nhead=8)
>>> src = torch.rand(10, 32, 512)
>>> rel = torch.rand(10, 10, 32, 512)
>>> out = encoder_layer(src, rel)
"""
def __init__(self, d_model, nhead, add_relation=False, dim_feedforward=
2048, dropout=0.1, activation='relu', relation_type=None):
super(RelationalTransformerEncoderLayer, self).__init__(d_model,
nhead, dim_feedforward=dim_feedforward, dropout=dropout,
activation=activation)
self.self_attn = RelationalMultiheadAttention(d_model, nhead,
add_relation=add_relation, dropout=dropout, relation_type=
relation_type)
def forward(self, src: 'Tensor', relation=None, src_mask:
'Optional[Tensor]'=None, src_key_padding_mask: 'Optional[Tensor]'=None
) ->Tensor:
"""Pass the input through the encoder layer.
Args:
src: the sequence to the encoder layer (required).
src_mask: the mask for the src sequence (optional).
src_key_padding_mask: the mask for the src keys per batch (optional).
Shape:
see the docs in Transformer class.
"""
src2 = self.self_attn(src, src, src, relation_dict=relation,
attn_mask=src_mask, key_padding_mask=src_key_padding_mask)[0]
src = src + self.dropout1(src2)
src = self.norm1(src)
src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
src = src + self.dropout2(src2)
src = self.norm2(src)
return src
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'d_model': 4, 'nhead': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import warnings
from torch import Tensor
from torch.nn import TransformerEncoderLayer
from torch.nn.functional import *
from torch.nn.modules.activation import MultiheadAttention
from torch.nn.modules.activation import xavier_uniform_
from torch.nn.modules.activation import xavier_normal_
from torch.nn.modules.activation import constant_
from torch.nn.modules.activation import Parameter
from typing import Optional
import torch.utils.data.dataset
from typing import Tuple
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mul_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 12 * x1), xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (8 + x0 + 12 * x1), xmask)
tmp1 = tl.load(in_ptr1 + (8 + x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + x2, tmp2, xmask)
@triton.jit
def triton_poi_fused_clone_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (4 + x0 + 12 * x1), xmask)
tmp1 = tl.load(in_ptr1 + (4 + x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + x2, tmp2, xmask)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_clone_5(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 4
xnumel = 16
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x1 = xindex
y0 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask, eviction_policy=
'evict_last')
tl.store(out_ptr0 + (x1 + 16 * y0), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_6(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(out_ptr1 + x0, tmp28, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_7(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-05
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
@triton.jit
def triton_poi_fused_relu_threshold_backward_8(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 2048
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, None)
tl.store(out_ptr0 + x2, tmp6, None)
@triton.jit
def triton_poi_fused_add_9(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK:
tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_out_ptr0 + x2, xmask)
tmp2 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp4 = tmp0 + tmp3
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_native_layer_norm_10(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-05
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + x0, tmp8, xmask)
tl.store(out_ptr1 + x0, tmp23, xmask)
@triton.jit
def triton_poi_fused_native_layer_norm_11(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x0, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13) = args
args.clear()
assert_size_stride(primals_1, (12, 4), (4, 1))
assert_size_stride(primals_2, (12,), (1,))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4,), (1,))
assert_size_stride(primals_5, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_6, (4,), (1,))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (2048, 4), (4, 1))
assert_size_stride(primals_9, (2048,), (1,))
assert_size_stride(primals_10, (4, 2048), (2048, 1))
assert_size_stride(primals_11, (4,), (1,))
assert_size_stride(primals_12, (4,), (1,))
assert_size_stride(primals_13, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 12), (12, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_5, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 12), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mul_0[grid(64)](buf0, primals_2, buf1, 64, XBLOCK=
64, num_warps=1, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_clone_1[grid(64)](buf0, primals_2, buf2, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf3 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_clone_2[grid(64)](buf0, primals_2, buf3, 64,
XBLOCK=64, num_warps=1, num_stages=1)
del buf0
del primals_2
buf4 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf1, (16, 4, 1), (1, 16, 0),
0), reinterpret_tensor(buf3, (16, 1, 4), (1, 0, 16), 0), out=buf4)
buf5 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused__softmax_3[grid(256)](buf4, buf5, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf6 = buf4
del buf4
triton_poi_fused__softmax_4[grid(256)](buf5, buf6, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf5
buf7 = empty_strided_cuda((16, 4, 1), (4, 1, 1), torch.float32)
extern_kernels.bmm(buf6, reinterpret_tensor(buf2, (16, 4, 1), (1,
16, 0), 0), out=buf7)
buf8 = empty_strided_cuda((4, 16, 1), (16, 1, 1), torch.float32)
triton_poi_fused_clone_5[grid(4, 16)](buf7, buf8, 4, 16, XBLOCK=16,
YBLOCK=4, num_warps=1, num_stages=1)
buf9 = reinterpret_tensor(buf7, (16, 4), (4, 1), 0)
del buf7
extern_kernels.addmm(primals_4, reinterpret_tensor(buf8, (16, 4), (
4, 1), 0), reinterpret_tensor(primals_3, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf9)
del primals_4
buf10 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf11 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
triton_poi_fused_add_native_layer_norm_6[grid(16)](primals_5, buf9,
buf10, buf11, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf12 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_native_layer_norm_7[grid(64)](primals_5, buf9,
buf10, buf11, primals_6, primals_7, buf12, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del primals_7
buf13 = empty_strided_cuda((16, 2048), (2048, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf12, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_8, (4, 2048), (1, 4), 0), out=buf13)
buf14 = reinterpret_tensor(buf13, (4, 4, 2048), (8192, 2048, 1), 0)
del buf13
buf20 = empty_strided_cuda((4, 4, 2048), (8192, 2048, 1), torch.bool)
triton_poi_fused_relu_threshold_backward_8[grid(32768)](buf14,
primals_9, buf20, 32768, XBLOCK=128, num_warps=4, num_stages=1)
del primals_9
buf15 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf14, (16, 2048), (2048, 1),
0), reinterpret_tensor(primals_10, (2048, 4), (1, 2048), 0),
out=buf15)
buf16 = reinterpret_tensor(buf15, (4, 4, 4), (16, 4, 1), 0)
del buf15
triton_poi_fused_add_9[grid(64)](buf16, buf12, primals_11, 64,
XBLOCK=64, num_warps=1, num_stages=1)
del primals_11
buf17 = buf11
del buf11
buf18 = buf10
del buf10
triton_poi_fused_native_layer_norm_10[grid(16)](buf16, buf17, buf18,
16, XBLOCK=16, num_warps=1, num_stages=1)
buf19 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_native_layer_norm_11[grid(64)](buf16, buf17, buf18,
primals_12, primals_13, buf19, 64, XBLOCK=64, num_warps=1,
num_stages=1)
del buf17
del buf18
del primals_13
return buf19, primals_5, primals_6, primals_12, buf6, reinterpret_tensor(
buf8, (16, 4), (4, 1), 0), buf9, reinterpret_tensor(buf12, (16, 4),
(4, 1), 0), reinterpret_tensor(buf14, (16, 2048), (2048, 1), 0
), buf16, primals_10, buf20, primals_8, primals_3, reinterpret_tensor(
buf2, (16, 1, 4), (1, 1, 16), 0), reinterpret_tensor(buf1, (16, 1,
4), (1, 1, 16), 0), reinterpret_tensor(buf3, (16, 4, 1), (1, 16, 1), 0)
def relational_multi_head_attention_forward(query: 'Tensor', key: 'Tensor',
value: 'Tensor', relation: 'Tensor', embed_dim_to_check: 'int',
num_heads: 'int', in_proj_weight: 'Tensor', in_proj_bias: 'Tensor',
bias_k: 'Optional[Tensor]', bias_v: 'Optional[Tensor]', add_zero_attn:
'bool', dropout_p: 'float', out_proj_weight: 'Tensor', out_proj_bias:
'Tensor', training: 'bool'=True, key_padding_mask: 'Optional[Tensor]'=
None, need_weights: 'bool'=True, attn_mask: 'Optional[Tensor]'=None,
use_separate_proj_weight: 'bool'=False, q_proj_weight:
'Optional[Tensor]'=None, k_proj_weight: 'Optional[Tensor]'=None,
v_proj_weight: 'Optional[Tensor]'=None, static_k: 'Optional[Tensor]'=
None, static_v: 'Optional[Tensor]'=None, relation_type: 'str'=None
) ->Tuple[Tensor, Optional[Tensor]]:
"""
Args:
query, key, value: map a query and a set of key-value pairs to an output.
See "Attention Is All You Need" for more details.
relation: relation between queries and keys.
embed_dim_to_check: total dimension of the model.
num_heads: parallel attention heads.
in_proj_weight, in_proj_bias: input projection weight and bias.
bias_k, bias_v: bias of the key and value sequences to be added at dim=0.
add_zero_attn: add a new batch of zeros to the key and
value sequences at dim=1.
dropout_p: probability of an element to be zeroed.
out_proj_weight, out_proj_bias: the output projection weight and bias.
training: apply dropout if is ``True``.
key_padding_mask: if provided, specified padding elements in the key will
be ignored by the attention. This is an binary mask. When the value is True,
the corresponding value on the attention layer will be filled with -inf.
need_weights: output attn_output_weights.
attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
the batches while a 3D mask allows to specify a different mask for the entries of each batch.
use_separate_proj_weight: the function accept the proj. weights for query, key,
and value in different forms. If false, in_proj_weight will be used, which is
a combination of q_proj_weight, k_proj_weight, v_proj_weight.
q_proj_weight, k_proj_weight, v_proj_weight, in_proj_bias: input projection weight and bias.
static_k, static_v: static key and value used for attention operators.
Shape:
Inputs:
- query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
the embedding dimension.
- key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- relation: :math:`(L, S, N, E)` where L is the target sequence length, where S is the source sequence length,
N is the batch size, E is the embedding dimension.
- key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions
will be unchanged. If a BoolTensor is provided, the positions with the
value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
- attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length,
S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked
positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
is provided, it will be added to the attention weight.
- static_k: :math:`(N*num_heads, S, E/num_heads)`, where S is the source sequence length,
N is the batch size, E is the embedding dimension. E/num_heads is the head dimension.
- static_v: :math:`(N*num_heads, S, E/num_heads)`, where S is the source sequence length,
N is the batch size, E is the embedding dimension. E/num_heads is the head dimension.
Outputs:
- attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
E is the embedding dimension.
- attn_output_weights: :math:`(N, L, S)` where N is the batch size,
L is the target sequence length, S is the source sequence length.
"""
tens_ops = (query, key, value, in_proj_weight, in_proj_bias, bias_k,
bias_v, out_proj_weight, out_proj_bias)
if has_torch_function(tens_ops):
return handle_torch_function(multi_head_attention_forward, tens_ops,
query, key, value, embed_dim_to_check, num_heads,
in_proj_weight, in_proj_bias, bias_k, bias_v, add_zero_attn,
dropout_p, out_proj_weight, out_proj_bias, training=training,
key_padding_mask=key_padding_mask, need_weights=need_weights,
attn_mask=attn_mask, use_separate_proj_weight=
use_separate_proj_weight, q_proj_weight=q_proj_weight,
k_proj_weight=k_proj_weight, v_proj_weight=v_proj_weight,
static_k=static_k, static_v=static_v)
tgt_len, bsz, embed_dim = query.size()
assert embed_dim == embed_dim_to_check
assert key.size(0) == value.size(0) and key.size(1) == value.size(1)
head_dim = embed_dim // num_heads
assert head_dim * num_heads == embed_dim, 'embed_dim must be divisible by num_heads'
scaling = float(head_dim) ** -0.5
if not use_separate_proj_weight:
if (query is key or torch.equal(query, key)) and (key is value or
torch.equal(key, value)):
q, k, v = linear(query, in_proj_weight, in_proj_bias).chunk(3,
dim=-1)
elif key is value or torch.equal(key, value):
_b = in_proj_bias
_start = 0
_end = embed_dim
_w = in_proj_weight[_start:_end, :]
if _b is not None:
_b = _b[_start:_end]
q = linear(query, _w, _b)
if key is None:
assert value is None
k = None
v = None
else:
_b = in_proj_bias
_start = embed_dim
_end = None
_w = in_proj_weight[_start:, :]
if _b is not None:
_b = _b[_start:]
k, v = linear(key, _w, _b).chunk(2, dim=-1)
else:
_b = in_proj_bias
_start = 0
_end = embed_dim
_w = in_proj_weight[_start:_end, :]
if _b is not None:
_b = _b[_start:_end]
q = linear(query, _w, _b)
_b = in_proj_bias
_start = embed_dim
_end = embed_dim * 2
_w = in_proj_weight[_start:_end, :]
if _b is not None:
_b = _b[_start:_end]
k = linear(key, _w, _b)
_b = in_proj_bias
_start = embed_dim * 2
_end = None
_w = in_proj_weight[_start:, :]
if _b is not None:
_b = _b[_start:]
v = linear(value, _w, _b)
else:
q_proj_weight_non_opt = torch.jit._unwrap_optional(q_proj_weight)
len1, len2 = q_proj_weight_non_opt.size()
assert len1 == embed_dim and len2 == query.size(-1)
k_proj_weight_non_opt = torch.jit._unwrap_optional(k_proj_weight)
len1, len2 = k_proj_weight_non_opt.size()
assert len1 == embed_dim and len2 == key.size(-1)
v_proj_weight_non_opt = torch.jit._unwrap_optional(v_proj_weight)
len1, len2 = v_proj_weight_non_opt.size()
assert len1 == embed_dim and len2 == value.size(-1)
if in_proj_bias is not None:
q = linear(query, q_proj_weight_non_opt, in_proj_bias[0:embed_dim])
k = linear(key, k_proj_weight_non_opt, in_proj_bias[embed_dim:
embed_dim * 2])
v = linear(value, v_proj_weight_non_opt, in_proj_bias[embed_dim *
2:])
else:
q = linear(query, q_proj_weight_non_opt, in_proj_bias)
k = linear(key, k_proj_weight_non_opt, in_proj_bias)
v = linear(value, v_proj_weight_non_opt, in_proj_bias)
q = q * scaling
if attn_mask is not None:
assert attn_mask.dtype == torch.float32 or attn_mask.dtype == torch.float64 or attn_mask.dtype == torch.float16 or attn_mask.dtype == torch.uint8 or attn_mask.dtype == torch.bool, 'Only float, byte, and bool types are supported for attn_mask, not {}'.format(
attn_mask.dtype)
if attn_mask.dtype == torch.uint8:
warnings.warn(
'Byte tensor for attn_mask in nn.MultiheadAttention is deprecated. Use bool tensor instead.'
)
attn_mask = attn_mask
if attn_mask.dim() == 2:
attn_mask = attn_mask.unsqueeze(0)
if list(attn_mask.size()) != [1, query.size(0), key.size(0)]:
raise RuntimeError(
'The size of the 2D attn_mask is not correct.')
elif attn_mask.dim() == 3:
if list(attn_mask.size()) != [bsz * num_heads, query.size(0),
key.size(0)]:
raise RuntimeError(
'The size of the 3D attn_mask is not correct.')
else:
raise RuntimeError("attn_mask's dimension {} is not supported".
format(attn_mask.dim()))
if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8:
warnings.warn(
'Byte tensor for key_padding_mask in nn.MultiheadAttention is deprecated. Use bool tensor instead.'
)
key_padding_mask = key_padding_mask
if bias_k is not None and bias_v is not None:
if static_k is None and static_v is None:
k = torch.cat([k, bias_k.repeat(1, bsz, 1)])
v = torch.cat([v, bias_v.repeat(1, bsz, 1)])
if attn_mask is not None:
attn_mask = pad(attn_mask, (0, 1))
if key_padding_mask is not None:
key_padding_mask = pad(key_padding_mask, (0, 1))
else:
assert static_k is None, 'bias cannot be added to static key.'
assert static_v is None, 'bias cannot be added to static value.'
else:
assert bias_k is None
assert bias_v is None
r = None
q = q.contiguous().view(tgt_len, bsz * num_heads, head_dim).transpose(0, 1)
if k is not None:
k = k.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
if v is not None:
v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
if static_k is not None:
assert static_k.size(0) == bsz * num_heads
assert static_k.size(2) == head_dim
k = static_k
if static_v is not None:
assert static_v.size(0) == bsz * num_heads
assert static_v.size(2) == head_dim
v = static_v
src_len = k.size(1)
if relation is not None:
if relation_type == 'qk+r':
r = relation.contiguous().view(tgt_len, -1, bsz * num_heads, 1
).squeeze(3).permute(2, 0, 1)
elif relation_type == 'q(k+r)':
r = relation.contiguous().view(tgt_len, src_len, bsz *
num_heads, head_dim).permute(2, 0, 1)
r = r.view(bsz * num_heads, tgt_len * src_len, head_dim)
r = torch.bmm(r, q.unsqueeze(2).repeat(1, 1, src_len, 1).view(
bsz * num_heads, tgt_len * src_len, head_dim).transpose(1, 2)
).view(bsz * num_heads, tgt_len, src_len)
if key_padding_mask is not None:
assert key_padding_mask.size(0) == bsz
assert key_padding_mask.size(1) == src_len
if add_zero_attn:
src_len += 1
k = torch.cat([k, torch.zeros((k.size(0), 1) + k.size()[2:], dtype=
k.dtype, device=k.device)], dim=1)
v = torch.cat([v, torch.zeros((v.size(0), 1) + v.size()[2:], dtype=
v.dtype, device=v.device)], dim=1)
if attn_mask is not None:
attn_mask = pad(attn_mask, (0, 1))
if key_padding_mask is not None:
key_padding_mask = pad(key_padding_mask, (0, 1))
attn_output_weights = torch.bmm(q, k.transpose(1, 2))
if r is not None:
attn_output_weights = attn_output_weights + r
assert list(attn_output_weights.size()) == [bsz * num_heads, tgt_len,
src_len]
if attn_mask is not None:
if attn_mask.dtype == torch.bool:
attn_output_weights.masked_fill_(attn_mask, float('-inf'))
else:
attn_output_weights += attn_mask
if key_padding_mask is not None:
attn_output_weights = attn_output_weights.view(bsz, num_heads,
tgt_len, src_len)
attn_output_weights = attn_output_weights.masked_fill(key_padding_mask
.unsqueeze(1).unsqueeze(2), float('-inf'))
attn_output_weights = attn_output_weights.view(bsz * num_heads,
tgt_len, src_len)
attn_output_weights = softmax(attn_output_weights, dim=-1)
attn_output_weights = dropout(attn_output_weights, p=dropout_p,
training=training)
attn_output = torch.bmm(attn_output_weights, v)
assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len,
bsz, embed_dim)
attn_output = linear(attn_output, out_proj_weight, out_proj_bias)
if need_weights:
attn_output_weights = attn_output_weights.view(bsz, num_heads,
tgt_len, src_len)
return attn_output, attn_output_weights.sum(dim=1) / num_heads
else:
return attn_output, None
class RelationalMultiheadAttention(MultiheadAttention):
"""Allows the model to jointly attend to information
from different representation subspaces.
See `Attention Is All You Need <https://arxiv.org/abs/1706.03762>`_
.. math::
\\text{MultiHead}(Q, K, V) = \\text{Concat}(head_1,\\dots,head_h)W^O
where :math:`head_i = \\text{Attention}(QW_i^Q, KW_i^K, VW_i^V)`.
Args:
embed_dim: total dimension of the model.
num_heads: parallel attention heads.
dropout: a Dropout layer on attn_output_weights. Default: 0.0.
bias: add bias as module parameter. Default: True.
add_bias_kv: add bias to the key and value sequences at dim=0.
add_zero_attn: add a new batch of zeros to the key and
value sequences at dim=1.
kdim: total number of features in key. Default: None.
vdim: total number of features in value. Default: None.
Note that if :attr:`kdim` and :attr:`vdim` are None, they will be set
to :attr:`embed_dim` such that query, key, and value have the same
number of features.
Examples::
>>> realational_multihead_attn = RelationalMultiheadAttention(embed_dim, num_heads)
>>> attn_output, attn_output_weights = realational_multihead_attn(query, key, value)
"""
bias_k: 'Optional[torch.Tensor]'
bias_v: 'Optional[torch.Tensor]'
def __init__(self, embed_dim, num_heads, dropout=0.0, bias=True,
add_bias_kv=False, add_zero_attn=False, kdim=None, vdim=None,
add_relation=False, rdim=None, relation_type=None):
super(RelationalMultiheadAttention, self).__init__(embed_dim=
embed_dim, num_heads=num_heads, dropout=dropout, bias=bias,
add_bias_kv=add_bias_kv, add_zero_attn=add_zero_attn, kdim=kdim,
vdim=vdim)
self.add_relation = add_relation
self.rdim = rdim if rdim is not None else embed_dim
self.relation_type = relation_type if relation_type else 'qk+r'
if self.add_relation:
if relation_type == 'qk+r':
self.r_proj_weight = Parameter(torch.Tensor(num_heads, self
.rdim))
self.r_proj_bias = Parameter(torch.empty(num_heads))
elif relation_type == 'q(k+r)':
self.r_proj_weight = Parameter(torch.Tensor(embed_dim, self
.rdim))
self.r_proj_bias = Parameter(torch.empty(embed_dim))
self._reset_parameters()
def _reset_parameters(self):
if self._qkv_same_embed_dim:
xavier_uniform_(self.in_proj_weight)
else:
xavier_uniform_(self.q_proj_weight)
xavier_uniform_(self.k_proj_weight)
xavier_uniform_(self.v_proj_weight)
if hasattr(self, 'add_relation') and self.add_relation:
xavier_uniform_(self.r_proj_weight)
constant_(self.r_proj_bias, 0.0)
if self.in_proj_bias is not None:
constant_(self.in_proj_bias, 0.0)
constant_(self.out_proj.bias, 0.0)
if self.bias_k is not None:
xavier_normal_(self.bias_k)
if self.bias_v is not None:
xavier_normal_(self.bias_v)
def __setstate__(self, state):
if '_qkv_same_embed_dim' not in state:
state['_qkv_same_embed_dim'] = True
super(MultiheadAttention, self).__setstate__(state)
def forward(self, query: 'Tensor', key: 'Tensor', value: 'Tensor',
relation_dict=None, key_padding_mask: 'Optional[Tensor]'=None,
need_weights: 'bool'=True, attn_mask: 'Optional[Tensor]'=None) ->Tuple[
Tensor, Optional[Tensor]]:
"""
Args:
query, key, value: map a query and a set of key-value pairs to an output.
See "Attention Is All You Need" for more details.
key_padding_mask: if provided, specified padding elements in the key will
be ignored by the attention. When given a binary mask and a value is True,
the corresponding value on the attention layer will be ignored. When given
a byte mask and a value is non-zero, the corresponding value on the attention
layer will be ignored
need_weights: output attn_output_weights.
attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all
the batches while a 3D mask allows to specify a different mask for the entries of each batch.
Shapes for inputs:
- query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
the embedding dimension.
- key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length.
If a ByteTensor is provided, the non-zero positions will be ignored while the position
with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the
value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged.
- attn_mask: if a 2D mask: :math:`(L, S)` where L is the target sequence length, S is the
source sequence length.
If a 3D mask: :math:`(N\\cdot\\text{num\\_heads}, L, S)` where N is the batch size, L is the target sequence
length, S is the source sequence length. ``attn_mask`` ensure that position i is allowed to attend
the unmasked positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend
while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True``
is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor
is provided, it will be added to the attention weight.
Shapes for outputs:
- attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
E is the embedding dimension.
- attn_output_weights: :math:`(N, L, S)` where N is the batch size,
L is the target sequence length, S is the source sequence length.
"""
if relation_dict is not None:
relation_labels = relation_dict['relation_labels']
relation_ids = relation_dict['relation_ids']
batch_index = relation_dict['batch_index']
pad_embedding = relation_dict['pad_embedding']
relation_labels = linear(relation_labels, self.r_proj_weight,
self.r_proj_bias)
pad_embedding = linear(pad_embedding.unsqueeze(0), self.
r_proj_weight, self.r_proj_bias).squeeze()
tgt_length, bsz, _ = query.size()
src_length, _, _ = key.size()
relation = pad_embedding.view(1, 1, 1, -1).repeat(bsz,
tgt_length, src_length, 1)
relation[batch_index, relation_ids[:, :, 0], relation_ids[:, :, 1]
] = relation_labels
relation = relation.permute(1, 2, 0, 3)
if not self._qkv_same_embed_dim:
return relational_multi_head_attention_forward(query, key,
value, relation, self.embed_dim, self.num_heads, self.
in_proj_weight, self.in_proj_bias, self.bias_k, self.
bias_v, self.add_zero_attn, self.dropout, self.out_proj
.weight, self.out_proj.bias, training=self.training,
key_padding_mask=key_padding_mask, need_weights=
need_weights, attn_mask=attn_mask,
use_separate_proj_weight=True, q_proj_weight=self.
q_proj_weight, k_proj_weight=self.k_proj_weight,
v_proj_weight=self.v_proj_weight)
else:
return relational_multi_head_attention_forward(query, key,
value, relation, self.embed_dim, self.num_heads, self.
in_proj_weight, self.in_proj_bias, self.bias_k, self.
bias_v, self.add_zero_attn, self.dropout, self.out_proj
.weight, self.out_proj.bias, training=self.training,
key_padding_mask=key_padding_mask, need_weights=
need_weights, attn_mask=attn_mask)
elif not self._qkv_same_embed_dim:
return relational_multi_head_attention_forward(query, key,
value, None, self.embed_dim, self.num_heads, self.
in_proj_weight, self.in_proj_bias, self.bias_k, self.bias_v,
self.add_zero_attn, self.dropout, self.out_proj.weight,
self.out_proj.bias, training=self.training,
key_padding_mask=key_padding_mask, need_weights=
need_weights, attn_mask=attn_mask, use_separate_proj_weight
=True, q_proj_weight=self.q_proj_weight, k_proj_weight=self
.k_proj_weight, v_proj_weight=self.v_proj_weight)
else:
return relational_multi_head_attention_forward(query, key,
value, None, self.embed_dim, self.num_heads, self.
in_proj_weight, self.in_proj_bias, self.bias_k, self.bias_v,
self.add_zero_attn, self.dropout, self.out_proj.weight,
self.out_proj.bias, training=self.training,
key_padding_mask=key_padding_mask, need_weights=
need_weights, attn_mask=attn_mask)
class RelationalTransformerEncoderLayerNew(TransformerEncoderLayer):
"""TransformerEncoderLayer is made up of self-attn and feedforward network.
This standard encoder layer is based on the paper "Attention Is All You Need".
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in
Neural Information Processing Systems, pages 6000-6010. Users may modify or implement
in a different way during application.
Args:
d_model: the number of expected features in the input (required).
nhead: the number of heads in the multiheadattention models (required).
dim_feedforward: the dimension of the feedforward network model (default=2048).
dropout: the dropout value (default=0.1).
activation: the activation function of intermediate layer, relu or gelu (default=relu).
Examples::
>>> encoder_layer = RelationalTransformerEncoderLayer(d_model=512, nhead=8)
>>> src = torch.rand(10, 32, 512)
>>> rel = torch.rand(10, 10, 32, 512)
>>> out = encoder_layer(src, rel)
"""
def __init__(self, d_model, nhead, add_relation=False, dim_feedforward=
2048, dropout=0.1, activation='relu', relation_type=None):
super(RelationalTransformerEncoderLayerNew, self).__init__(d_model,
nhead, dim_feedforward=dim_feedforward, dropout=dropout,
activation=activation)
self.self_attn = RelationalMultiheadAttention(d_model, nhead,
add_relation=add_relation, dropout=dropout, relation_type=
relation_type)
def forward(self, input_0):
primals_1 = self.self_attn.in_proj_weight
primals_2 = self.self_attn.in_proj_bias
primals_3 = self.self_attn.out_proj.weight
primals_4 = self.self_attn.out_proj.bias
primals_8 = self.linear1.weight
primals_9 = self.linear1.bias
primals_10 = self.linear2.weight
primals_6 = self.linear2.bias
primals_7 = self.norm1.weight
primals_11 = self.norm1.bias
primals_12 = self.norm2.weight
primals_13 = self.norm2.bias
primals_5 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13])
return output[0]
|
mfk3138/jiant
|
RelationalTransformerEncoderLayer
| false
| 4,204
|
[
"MIT"
] | 0
|
6e67ff1ecb1bb98533c1019a86af4ad2c04c6a64
|
https://github.com/mfk3138/jiant/tree/6e67ff1ecb1bb98533c1019a86af4ad2c04c6a64
|
LxmertAttentionOutput
|
from _paritybench_helpers import _mock_config
import torch
from torch import nn
class LxmertAttentionOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
return hidden_states
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'config': _mock_config(hidden_size=4, hidden_dropout_prob=
0.5)}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_add_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK:
tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + x2, xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_native_layer_norm_1(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-12
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + x0, tmp8, xmask)
tl.store(out_ptr1 + x0, tmp23, xmask)
@triton.jit
def triton_poi_fused_native_layer_norm_2(in_ptr0, in_ptr1, in_ptr2, in_ptr3,
in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x0, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
get_raw_stream(0)
triton_poi_fused_add_0[grid(256)](buf1, primals_2, primals_4, 256,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
del primals_4
buf2 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf3 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
triton_poi_fused_native_layer_norm_1[grid(64)](buf1, buf2, buf3, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf4 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_native_layer_norm_2[grid(256)](buf1, buf2, buf3,
primals_5, primals_6, buf4, 256, XBLOCK=256, num_warps=4,
num_stages=1)
del buf2
del buf3
del primals_6
return buf4, primals_5, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf1
class LxmertAttentionOutputNew(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, input_0, input_1):
primals_1 = self.dense.weight
primals_2 = self.dense.bias
primals_5 = self.LayerNorm.weight
primals_6 = self.LayerNorm.bias
primals_3 = input_0
primals_4 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6])
return output[0]
|
rsgit95/med_kg_txt_multimodal
|
LxmertAttentionOutput
| false
| 4,205
|
[
"Apache-2.0"
] | 0
|
80355b0cf58e0571531ad6f9728c533110ca996d
|
https://github.com/rsgit95/med_kg_txt_multimodal/tree/80355b0cf58e0571531ad6f9728c533110ca996d
|
Block
|
import torch
import torch as th
from torch import nn
def drop_path(x, drop_prob: 'float'=0.0, training: 'bool'=False):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
'survival rate' as the argument.
"""
if drop_prob == 0.0 or not training:
return x
keep_prob = 1 - drop_prob
shape = (x.shape[0],) + (1,) * (x.ndim - 1)
random_tensor = keep_prob + th.rand(shape, dtype=x.dtype, device=x.device)
random_tensor.floor_()
output = x.div(keep_prob) * random_tensor
return output
class Attention(nn.Module):
def __init__(self, dim=2, num_heads=8, qkv_bias=False, qk_scale=None,
attn_drop=0.0, proj_drop=0.0):
super().__init__()
self.proj = nn.Linear(dim, dim)
def forward(self, x):
x = self.proj(x)
return x
class DropPath(th.nn.Module):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
"""
def __init__(self, drop_prob=None):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
def forward(self, x):
return drop_path(x, self.drop_prob, self.training)
class Mlp(th.nn.Module):
""" MLP as used in Vision Transformer, MLP-Mixer and related networks
"""
def __init__(self, in_features, hidden_features=None, out_features=None,
act_layer=th.nn.GELU, drop=0.0):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = th.nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = th.nn.Linear(hidden_features, out_features)
self.drop = th.nn.Dropout(drop)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
class Block(th.nn.Module):
def __init__(self, dim=512, num_heads=8, mlp_ratio=4.0, qkv_bias=False,
qk_scale=None, drop=0.0, attn_drop=0.0, drop_path=0.0, act_layer=th
.nn.GELU, norm_layer=th.nn.LayerNorm):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias,
qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
self.drop_path = DropPath(drop_path
) if drop_path > 0.0 else th.nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim,
act_layer=act_layer, drop=drop)
def forward(self, x):
x = x + self.drop_path(self.attn(self.norm1(x)))
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
def get_inputs():
return [torch.rand([4, 4, 512, 512])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice
import torch as th
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused_native_layer_norm_0(in_out_ptr0, in_ptr0, in_ptr1,
in_ptr2, out_ptr0, out_ptr1, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 512 * x0), None)
tmp21 = tl.load(in_ptr1 + r1, None, eviction_policy='evict_last')
tmp23 = tl.load(in_ptr2 + r1, None, eviction_policy='evict_last')
tmp1 = tl.broadcast_to(tmp0, [RBLOCK])
tmp3 = tl.broadcast_to(tmp1, [RBLOCK])
tmp5 = triton_helpers.promote_to_tensor(tl.sum(tmp3, 0))
tmp6 = tl.full([1], 512, tl.int32)
tmp7 = tmp6.to(tl.float32)
tmp8 = tmp5 / tmp7
tmp9 = tmp1 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tl.broadcast_to(tmp10, [RBLOCK])
tmp13 = triton_helpers.promote_to_tensor(tl.sum(tmp11, 0))
tmp14 = 512.0
tmp15 = tmp13 / tmp14
tmp16 = 1e-05
tmp17 = tmp15 + tmp16
tmp18 = libdevice.rsqrt(tmp17)
tmp19 = tmp0 - tmp8
tmp20 = tmp19 * tmp18
tmp22 = tmp20 * tmp21
tmp24 = tmp22 + tmp23
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp18, None)
tl.store(out_ptr1 + (r1 + 512 * x0), tmp24, None)
tl.store(out_ptr0 + x0, tmp8, None)
@triton.jit
def triton_per_fused_add_native_layer_norm_1(in_out_ptr0, in_ptr0, in_ptr1,
in_ptr2, in_ptr3, out_ptr0, out_ptr1, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 512 * x0), None)
tmp1 = tl.load(in_ptr1 + (r1 + 512 * x0), None)
tmp23 = tl.load(in_ptr2 + r1, None, eviction_policy='evict_last')
tmp25 = tl.load(in_ptr3 + r1, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.broadcast_to(tmp2, [RBLOCK])
tmp5 = tl.broadcast_to(tmp3, [RBLOCK])
tmp7 = triton_helpers.promote_to_tensor(tl.sum(tmp5, 0))
tmp8 = tl.full([1], 512, tl.int32)
tmp9 = tmp8.to(tl.float32)
tmp10 = tmp7 / tmp9
tmp11 = tmp3 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tl.broadcast_to(tmp12, [RBLOCK])
tmp15 = triton_helpers.promote_to_tensor(tl.sum(tmp13, 0))
tmp16 = 512.0
tmp17 = tmp15 / tmp16
tmp18 = 1e-05
tmp19 = tmp17 + tmp18
tmp20 = libdevice.rsqrt(tmp19)
tmp21 = tmp2 - tmp10
tmp22 = tmp21 * tmp20
tmp24 = tmp22 * tmp23
tmp26 = tmp24 + tmp25
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp20, None)
tl.store(out_ptr1 + (r1 + 512 * x0), tmp26, None)
tl.store(out_ptr0 + x0, tmp10, None)
@triton.jit
def triton_poi_fused_gelu_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, None)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865476
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + x0, tmp8, None)
@triton.jit
def triton_poi_fused_add_3(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel,
XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 512
tmp0 = tl.load(in_ptr0 + x2, None)
tmp1 = tl.load(in_ptr1 + x2, None)
tmp3 = tl.load(in_out_ptr0 + x2, None)
tmp4 = tl.load(in_ptr2 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tl.store(in_out_ptr0 + x2, tmp6, None)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11) = args
args.clear()
assert_size_stride(primals_1, (512,), (1,))
assert_size_stride(primals_2, (512,), (1,))
assert_size_stride(primals_3, (4, 4, 512, 512), (1048576, 262144, 512, 1))
assert_size_stride(primals_4, (512, 512), (512, 1))
assert_size_stride(primals_5, (512,), (1,))
assert_size_stride(primals_6, (512,), (1,))
assert_size_stride(primals_7, (512,), (1,))
assert_size_stride(primals_8, (2048, 512), (512, 1))
assert_size_stride(primals_9, (2048,), (1,))
assert_size_stride(primals_10, (512, 2048), (2048, 1))
assert_size_stride(primals_11, (512,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 512, 1), (2048, 512, 1, 1), torch.
float32)
buf1 = empty_strided_cuda((4, 4, 512, 1), (2048, 512, 1, 8192),
torch.float32)
buf3 = reinterpret_tensor(buf1, (4, 4, 512, 1), (2048, 512, 1, 1), 0)
del buf1
buf4 = empty_strided_cuda((4, 4, 512, 512), (1048576, 262144, 512,
1), torch.float32)
get_raw_stream(0)
triton_per_fused_native_layer_norm_0[grid(8192)](buf3, primals_3,
primals_1, primals_2, buf0, buf4, 8192, 512, num_warps=4,
num_stages=1)
del primals_1
del primals_2
buf5 = empty_strided_cuda((8192, 512), (512, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf4, (8192, 512
), (512, 1), 0), reinterpret_tensor(primals_4, (512, 512), (1,
512), 0), alpha=1, beta=1, out=buf5)
del primals_5
buf6 = empty_strided_cuda((4, 4, 512, 1), (2048, 512, 1, 1), torch.
float32)
buf7 = empty_strided_cuda((4, 4, 512, 1), (2048, 512, 1, 8192),
torch.float32)
buf9 = reinterpret_tensor(buf7, (4, 4, 512, 1), (2048, 512, 1, 1), 0)
del buf7
buf10 = empty_strided_cuda((4, 4, 512, 512), (1048576, 262144, 512,
1), torch.float32)
triton_per_fused_add_native_layer_norm_1[grid(8192)](buf9,
primals_3, buf5, primals_6, primals_7, buf6, buf10, 8192, 512,
num_warps=4, num_stages=1)
del primals_7
buf11 = empty_strided_cuda((8192, 2048), (2048, 1), torch.float32)
extern_kernels.addmm(primals_9, reinterpret_tensor(buf10, (8192,
512), (512, 1), 0), reinterpret_tensor(primals_8, (512, 2048),
(1, 512), 0), alpha=1, beta=1, out=buf11)
del primals_9
buf12 = empty_strided_cuda((4, 4, 512, 2048), (4194304, 1048576,
2048, 1), torch.float32)
triton_poi_fused_gelu_2[grid(16777216)](buf11, buf12, 16777216,
XBLOCK=1024, num_warps=4, num_stages=1)
buf13 = empty_strided_cuda((8192, 512), (512, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf12, (8192, 2048), (2048, 1),
0), reinterpret_tensor(primals_10, (2048, 512), (1, 2048), 0),
out=buf13)
buf14 = reinterpret_tensor(buf13, (4, 4, 512, 512), (1048576,
262144, 512, 1), 0)
del buf13
triton_poi_fused_add_3[grid(4194304)](buf14, primals_3, buf5,
primals_11, 4194304, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_11
return buf14, primals_3, primals_6, buf0, buf3, reinterpret_tensor(buf4,
(8192, 512), (512, 1), 0), buf5, buf6, buf9, reinterpret_tensor(buf10,
(8192, 512), (512, 1), 0), buf11, reinterpret_tensor(buf12, (8192,
2048), (2048, 1), 0), primals_10, primals_8, primals_4
def drop_path(x, drop_prob: 'float'=0.0, training: 'bool'=False):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
'survival rate' as the argument.
"""
if drop_prob == 0.0 or not training:
return x
keep_prob = 1 - drop_prob
shape = (x.shape[0],) + (1,) * (x.ndim - 1)
random_tensor = keep_prob + th.rand(shape, dtype=x.dtype, device=x.device)
random_tensor.floor_()
output = x.div(keep_prob) * random_tensor
return output
class Attention(nn.Module):
def __init__(self, dim=2, num_heads=8, qkv_bias=False, qk_scale=None,
attn_drop=0.0, proj_drop=0.0):
super().__init__()
self.proj = nn.Linear(dim, dim)
def forward(self, x):
x = self.proj(x)
return x
class DropPath(th.nn.Module):
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
"""
def __init__(self, drop_prob=None):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
def forward(self, x):
return drop_path(x, self.drop_prob, self.training)
class Mlp(th.nn.Module):
""" MLP as used in Vision Transformer, MLP-Mixer and related networks
"""
def __init__(self, in_features, hidden_features=None, out_features=None,
act_layer=th.nn.GELU, drop=0.0):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = th.nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = th.nn.Linear(hidden_features, out_features)
self.drop = th.nn.Dropout(drop)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
class BlockNew(th.nn.Module):
def __init__(self, dim=512, num_heads=8, mlp_ratio=4.0, qkv_bias=False,
qk_scale=None, drop=0.0, attn_drop=0.0, drop_path=0.0, act_layer=th
.nn.GELU, norm_layer=th.nn.LayerNorm):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias,
qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
self.drop_path = DropPath(drop_path
) if drop_path > 0.0 else th.nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim,
act_layer=act_layer, drop=drop)
def forward(self, input_0):
primals_1 = self.norm1.weight
primals_2 = self.norm1.bias
primals_4 = self.attn.proj.weight
primals_5 = self.attn.proj.bias
primals_6 = self.norm2.weight
primals_7 = self.norm2.bias
primals_8 = self.mlp.fc1.weight
primals_9 = self.mlp.fc1.bias
primals_10 = self.mlp.fc2.weight
primals_11 = self.mlp.fc2.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11])
return output[0]
|
q5628077/Transformer-in-RL
|
Block
| false
| 4,206
|
[
"MIT"
] | 0
|
14679656779a372d91d9fbd89bd802b5ff34c200
|
https://github.com/q5628077/Transformer-in-RL/tree/14679656779a372d91d9fbd89bd802b5ff34c200
|
Net
|
import torch
import torch.nn as nn
import torch.nn.functional as F
class Net(nn.Module):
def __init__(self, state_dim, action_dim):
super(Net, self).__init__()
fc1_dim = 32
fc2_dim = 64
fc3_dim = 128
self.fc1 = nn.Linear(state_dim, fc1_dim)
self.fc2 = nn.Linear(fc1_dim, fc2_dim)
self.fc3 = nn.Linear(fc2_dim, fc3_dim)
self.fc_out = nn.Linear(fc3_dim, action_dim)
def forward(self, x):
x = self.fc1(x)
x = F.relu(x)
x = self.fc2(x)
x = F.relu(x)
x = self.fc3(x)
x = F.relu(x)
return self.fc_out(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'state_dim': 4, 'action_dim': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_relu_threshold_backward_0(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 32
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, None)
tl.store(out_ptr0 + x2, tmp6, None)
@triton.jit
def triton_poi_fused_relu_threshold_backward_1(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 64
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, None)
tl.store(out_ptr0 + x2, tmp6, None)
@triton.jit
def triton_poi_fused_relu_threshold_backward_2(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 128
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, None)
tl.store(out_ptr0 + x2, tmp6, None)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (32, 4), (4, 1))
assert_size_stride(primals_2, (32,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (64, 32), (32, 1))
assert_size_stride(primals_5, (64,), (1,))
assert_size_stride(primals_6, (128, 64), (64, 1))
assert_size_stride(primals_7, (128,), (1,))
assert_size_stride(primals_8, (4, 128), (128, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 32), (32, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 32), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 32), (512, 128, 32, 1), 0)
del buf0
buf9 = empty_strided_cuda((4, 4, 4, 32), (512, 128, 32, 1), torch.bool)
get_raw_stream(0)
triton_poi_fused_relu_threshold_backward_0[grid(2048)](buf1,
primals_2, buf9, 2048, XBLOCK=256, num_warps=4, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((64, 64), (64, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf1, (64, 32), (32, 1), 0),
reinterpret_tensor(primals_4, (32, 64), (1, 32), 0), out=buf2)
buf3 = reinterpret_tensor(buf2, (4, 4, 4, 64), (1024, 256, 64, 1), 0)
del buf2
buf8 = empty_strided_cuda((4, 4, 4, 64), (1024, 256, 64, 1), torch.bool
)
triton_poi_fused_relu_threshold_backward_1[grid(4096)](buf3,
primals_5, buf8, 4096, XBLOCK=256, num_warps=4, num_stages=1)
del primals_5
buf4 = empty_strided_cuda((64, 128), (128, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf3, (64, 64), (64, 1), 0),
reinterpret_tensor(primals_6, (64, 128), (1, 64), 0), out=buf4)
buf5 = reinterpret_tensor(buf4, (4, 4, 4, 128), (2048, 512, 128, 1), 0)
del buf4
buf7 = empty_strided_cuda((4, 4, 4, 128), (2048, 512, 128, 1),
torch.bool)
triton_poi_fused_relu_threshold_backward_2[grid(8192)](buf5,
primals_7, buf7, 8192, XBLOCK=256, num_warps=4, num_stages=1)
del primals_7
buf6 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_9, reinterpret_tensor(buf5, (64, 128),
(128, 1), 0), reinterpret_tensor(primals_8, (128, 4), (1, 128),
0), alpha=1, beta=1, out=buf6)
del primals_9
return reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0
), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), reinterpret_tensor(buf1, (64, 32), (32, 1), 0), reinterpret_tensor(
buf3, (64, 64), (64, 1), 0), reinterpret_tensor(buf5, (64, 128), (
128, 1), 0), primals_8, buf7, primals_6, buf8, primals_4, buf9
class NetNew(nn.Module):
def __init__(self, state_dim, action_dim):
super(NetNew, self).__init__()
fc1_dim = 32
fc2_dim = 64
fc3_dim = 128
self.fc1 = nn.Linear(state_dim, fc1_dim)
self.fc2 = nn.Linear(fc1_dim, fc2_dim)
self.fc3 = nn.Linear(fc2_dim, fc3_dim)
self.fc_out = nn.Linear(fc3_dim, action_dim)
def forward(self, input_0):
primals_1 = self.fc1.weight
primals_2 = self.fc1.bias
primals_4 = self.fc2.weight
primals_5 = self.fc2.bias
primals_6 = self.fc3.weight
primals_7 = self.fc3.bias
primals_8 = self.fc_out.weight
primals_9 = self.fc_out.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0]
|
ronekko/study_reinforcement_learning
|
Net
| false
| 4,207
|
[
"MIT"
] | 0
|
ef5201e3eae69c20f29b7f176b5a6de7ecdb856a
|
https://github.com/ronekko/study_reinforcement_learning/tree/ef5201e3eae69c20f29b7f176b5a6de7ecdb856a
|
IReLU
|
import math
import torch
class IReLU(torch.nn.Module):
__constants__ = ['negative_slope', 'positive_slope']
negative_slope: 'float'
positive_slope: 'float'
def __init__(self, negative_slope=math.tan(math.pi / 8), positive_slope
=math.tan(3 * math.pi / 8)):
super(IReLU, self).__init__()
self.negative_slope = negative_slope
self.positive_slope = positive_slope
def forward(self, x):
return torch.clamp(x, min=0) * self.positive_slope + torch.clamp(x,
max=0) * self.negative_slope
def inv(self, y):
return torch.clamp(y, min=0) / self.positive_slope + torch.clamp(y,
max=0) / self.negative_slope
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import math
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_clamp_mul_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 0.0
tmp2 = triton_helpers.maximum(tmp0, tmp1)
tmp3 = 2.414213562373095
tmp4 = tmp2 * tmp3
tmp5 = triton_helpers.minimum(tmp0, tmp1)
tmp6 = 0.41421356237309503
tmp7 = tmp5 * tmp6
tmp8 = tmp4 + tmp7
tl.store(out_ptr0 + x0, tmp8, xmask)
def call(args):
arg0_1, = args
args.clear()
assert_size_stride(arg0_1, (4, 4, 4, 4), (64, 16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_clamp_mul_0[grid(256)](arg0_1, buf0, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del arg0_1
return buf0,
class IReLUNew(torch.nn.Module):
__constants__ = ['negative_slope', 'positive_slope']
negative_slope: 'float'
positive_slope: 'float'
def __init__(self, negative_slope=math.tan(math.pi / 8), positive_slope
=math.tan(3 * math.pi / 8)):
super(IReLUNew, self).__init__()
self.negative_slope = negative_slope
self.positive_slope = positive_slope
def inv(self, y):
return torch.clamp(y, min=0) / self.positive_slope + torch.clamp(y,
max=0) / self.negative_slope
def forward(self, input_0):
arg0_1 = input_0
output = call([arg0_1])
return output[0]
|
rupumped/DFL
|
IReLU
| false
| 4,208
|
[
"BSD-3-Clause"
] | 0
|
a4e4d96b7ce7522cf7fee3c2cfdbb54eb7a473f2
|
https://github.com/rupumped/DFL/tree/a4e4d96b7ce7522cf7fee3c2cfdbb54eb7a473f2
|
Affine
|
import torch
from torch import nn
class Affine(nn.Module):
def __init__(self, channel):
super().__init__()
self.g = nn.Parameter(torch.ones(1, 1, channel))
self.b = nn.Parameter(torch.zeros(1, 1, channel))
def forward(self, x):
return x * self.g + self.b
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'channel': 4}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_add_mul_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 * tmp1
tmp4 = tmp2 + tmp3
tl.store(out_ptr0 + x2, tmp4, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (1, 1, 4), (4, 4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (1, 1, 4), (4, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_add_mul_0[grid(256)](primals_2, primals_1,
primals_3, buf0, 256, XBLOCK=128, num_warps=4, num_stages=1)
del primals_1
del primals_3
return buf0, primals_2
class AffineNew(nn.Module):
def __init__(self, channel):
super().__init__()
self.g = nn.Parameter(torch.ones(1, 1, channel))
self.b = nn.Parameter(torch.zeros(1, 1, channel))
def forward(self, input_0):
primals_1 = self.g
primals_3 = self.b
primals_2 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
Affine
| false
| 4,209
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
ECAAttention
|
import torch
from torch import nn
from torch.nn import init
class ECAAttention(nn.Module):
def __init__(self, kernel_size=3):
super().__init__()
self.gap = nn.AdaptiveAvgPool2d(1)
self.conv = nn.Conv1d(1, 1, kernel_size=kernel_size, padding=(
kernel_size - 1) // 2)
self.sigmoid = nn.Sigmoid()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, x):
y = self.gap(x)
y = y.squeeze(-1).permute(0, 2, 1)
y = self.conv(y)
y = self.sigmoid(y)
y = y.permute(0, 2, 1).unsqueeze(-1)
return x * y.expand_as(x)
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch import nn
from torch.nn import init
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused_mean_0(in_out_ptr0, in_ptr0, xnumel, rnumel, XBLOCK:
tl.constexpr):
xnumel = 16
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 16 * x0), xmask, other=0.0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp3 = tl.where(xmask, tmp1, 0)
tmp4 = tl.sum(tmp3, 1)[:, None]
tmp5 = 16.0
tmp6 = tmp4 / tmp5
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp6, xmask)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr0 + 0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK])
tmp3 = tmp0 + tmp2
tl.store(in_out_ptr0 + x0, tmp3, xmask)
@triton.jit
def triton_poi_fused_mul_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 16
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp2 = tl.sigmoid(tmp1)
tmp3 = tmp0 * tmp2
tl.store(out_ptr0 + x2, tmp3, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (1, 1, 3), (3, 3, 1))
assert_size_stride(primals_3, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 16, 16), torch.float32)
buf1 = buf0
del buf0
get_raw_stream(0)
triton_per_fused_mean_0[grid(16)](buf1, primals_1, 16, 16, XBLOCK=1,
num_warps=2, num_stages=1)
buf2 = extern_kernels.convolution(reinterpret_tensor(buf1, (4, 1, 4
), (4, 0, 1), 0), primals_2, stride=(1,), padding=(1,),
dilation=(1,), transposed=False, output_padding=(0,), groups=1,
bias=None)
assert_size_stride(buf2, (4, 1, 4), (4, 4, 1))
buf3 = buf2
del buf2
triton_poi_fused_convolution_1[grid(16)](buf3, primals_3, 16,
XBLOCK=16, num_warps=1, num_stages=1)
del primals_3
buf4 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_mul_2[grid(256)](primals_1, buf3, buf4, 256,
XBLOCK=128, num_warps=4, num_stages=1)
return buf4, primals_1, primals_2, reinterpret_tensor(buf1, (4, 1, 4),
(4, 1, 1), 0), buf3
class ECAAttentionNew(nn.Module):
def __init__(self, kernel_size=3):
super().__init__()
self.gap = nn.AdaptiveAvgPool2d(1)
self.conv = nn.Conv1d(1, 1, kernel_size=kernel_size, padding=(
kernel_size - 1) // 2)
self.sigmoid = nn.Sigmoid()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, input_0):
primals_2 = self.conv.weight
primals_3 = self.conv.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
ECAAttention
| false
| 4,210
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
GTXAttentionOutput
|
from _paritybench_helpers import _mock_config
import torch
from torch import nn
class GTXAttentionOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor, KnowMix_indices=None):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
if KnowMix_indices is None:
hidden_states = input_tensor + hidden_states
else:
if isinstance(KnowMix_indices, int):
input_tensor[:, KnowMix_indices] = input_tensor[:,
KnowMix_indices] + hidden_states.squeeze(1)
else:
input_tensor[KnowMix_indices, :] = input_tensor[
KnowMix_indices, :] + hidden_states.squeeze(1)
hidden_states = input_tensor
hidden_states = self.LayerNorm(hidden_states)
return hidden_states
def get_inputs():
return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'config': _mock_config(hidden_size=4, hidden_dropout_prob=
0.5)}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_add_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK:
tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_out_ptr0 + x2, xmask)
tmp2 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp4 = tmp0 + tmp3
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_native_layer_norm_1(in_ptr0, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tmp9 = tmp0 - tmp8
tmp10 = tmp9 * tmp9
tmp11 = tmp1 - tmp8
tmp12 = tmp11 * tmp11
tmp13 = tmp10 + tmp12
tmp14 = tmp3 - tmp8
tmp15 = tmp14 * tmp14
tmp16 = tmp13 + tmp15
tmp17 = tmp5 - tmp8
tmp18 = tmp17 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp19 / tmp7
tmp21 = 1e-12
tmp22 = tmp20 + tmp21
tmp23 = libdevice.rsqrt(tmp22)
tl.store(out_ptr0 + x0, tmp8, xmask)
tl.store(out_ptr1 + x0, tmp23, xmask)
@triton.jit
def triton_poi_fused_native_layer_norm_2(in_ptr0, in_ptr1, in_ptr2, in_ptr3,
in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x0, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 - tmp1
tmp4 = tmp2 * tmp3
tmp6 = tmp4 * tmp5
tmp8 = tmp6 + tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6 = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
get_raw_stream(0)
triton_poi_fused_add_0[grid(256)](buf1, primals_4, primals_2, 256,
XBLOCK=128, num_warps=4, num_stages=1)
del primals_2
del primals_4
buf2 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf3 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
triton_poi_fused_native_layer_norm_1[grid(64)](buf1, buf2, buf3, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf4 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_native_layer_norm_2[grid(256)](buf1, buf2, buf3,
primals_5, primals_6, buf4, 256, XBLOCK=256, num_warps=4,
num_stages=1)
del buf2
del buf3
del primals_6
return buf4, primals_5, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf1
class GTXAttentionOutputNew(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, input_0, input_1):
primals_1 = self.dense.weight
primals_2 = self.dense.bias
primals_5 = self.LayerNorm.weight
primals_6 = self.LayerNorm.bias
primals_3 = input_0
primals_4 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6])
return output[0]
|
rsgit95/med_kg_txt_multimodal
|
GTXAttentionOutput
| false
| 4,211
|
[
"Apache-2.0"
] | 0
|
80355b0cf58e0571531ad6f9728c533110ca996d
|
https://github.com/rsgit95/med_kg_txt_multimodal/tree/80355b0cf58e0571531ad6f9728c533110ca996d
|
Actor
|
import torch
import torch.nn as nn
import torch.nn.functional as F
class Actor(nn.Module):
def __init__(self, state_dim, action_dim):
super(Actor, self).__init__()
self.l1 = nn.Linear(state_dim, 400)
self.l2 = nn.Linear(400, 200)
self.l3 = nn.Linear(200, action_dim)
def forward(self, x):
x = F.relu(self.l1(x))
x = F.relu(self.l2(x))
x = nn.Softmax()(self.l3(x))
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'state_dim': 4, 'action_dim': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_relu_threshold_backward_0(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 25600
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x4 = xindex
x0 = xindex % 400
x2 = xindex % 1600
x3 = xindex // 1600
tmp0 = tl.load(in_out_ptr0 + x4, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x4, tmp4, xmask)
tl.store(out_ptr0 + (x2 + 1664 * x3), tmp6, xmask)
@triton.jit
def triton_poi_fused_relu_threshold_backward_1(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 12800
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 200
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, xmask)
tl.store(out_ptr0 + x2, tmp6, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x3, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy=
'evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x3, tmp8, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (400, 4), (4, 1))
assert_size_stride(primals_2, (400,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (200, 400), (400, 1))
assert_size_stride(primals_5, (200,), (1,))
assert_size_stride(primals_6, (4, 200), (200, 1))
assert_size_stride(primals_7, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 400), (400, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 400), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 400), (6400, 1600, 400, 1), 0
)
del buf0
buf8 = empty_strided_cuda((4, 4, 4, 400), (6656, 1664, 400, 1),
torch.bool)
get_raw_stream(0)
triton_poi_fused_relu_threshold_backward_0[grid(25600)](buf1,
primals_2, buf8, 25600, XBLOCK=256, num_warps=4, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((64, 200), (200, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf1, (64, 400), (400, 1), 0),
reinterpret_tensor(primals_4, (400, 200), (1, 400), 0), out=buf2)
buf3 = reinterpret_tensor(buf2, (4, 4, 4, 200), (3200, 800, 200, 1), 0)
del buf2
buf7 = empty_strided_cuda((4, 4, 4, 200), (3200, 800, 200, 1),
torch.bool)
triton_poi_fused_relu_threshold_backward_1[grid(12800)](buf3,
primals_5, buf7, 12800, XBLOCK=256, num_warps=4, num_stages=1)
del primals_5
buf4 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_7, reinterpret_tensor(buf3, (64, 200),
(200, 1), 0), reinterpret_tensor(primals_6, (200, 4), (1, 200),
0), alpha=1, beta=1, out=buf4)
del primals_7
buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__softmax_2[grid(256)](buf4, buf5, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf6 = reinterpret_tensor(buf4, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf4
triton_poi_fused__softmax_3[grid(256)](buf5, buf6, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf5
return buf6, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), reinterpret_tensor(buf1, (64, 400), (400, 1), 0
), reinterpret_tensor(buf3, (64, 200), (200, 1), 0
), buf6, primals_6, buf7, primals_4, buf8
class ActorNew(nn.Module):
def __init__(self, state_dim, action_dim):
super(ActorNew, self).__init__()
self.l1 = nn.Linear(state_dim, 400)
self.l2 = nn.Linear(400, 200)
self.l3 = nn.Linear(200, action_dim)
def forward(self, input_0):
primals_1 = self.l1.weight
primals_2 = self.l1.bias
primals_4 = self.l2.weight
primals_5 = self.l2.bias
primals_6 = self.l3.weight
primals_7 = self.l3.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
|
rortiz9/meleeml
|
Actor
| false
| 4,212
|
[
"MIT"
] | 0
|
9be4bf53a377dfb46dbb3b51f102f1bffc0124d2
|
https://github.com/rortiz9/meleeml/tree/9be4bf53a377dfb46dbb3b51f102f1bffc0124d2
|
PolicyNetwork
|
import torch
import torch.nn as nn
from torch.nn import functional as F
from torch.distributions import Normal
class PolicyNetwork(nn.Module):
def __init__(self, state_dim, action_dim, hidden_dim, init_w=0.003,
log_std_min=-20, log_std_max=2):
super(PolicyNetwork, self).__init__()
self.log_std_min = log_std_min
self.log_std_max = log_std_max
self.linear1 = nn.Linear(state_dim, hidden_dim)
self.linear2 = nn.Linear(hidden_dim, hidden_dim)
self.mean_linear = nn.Linear(hidden_dim, action_dim)
self.mean_linear.weight.data.uniform_(-init_w, init_w)
self.mean_linear.bias.data.uniform_(-init_w, init_w)
self.log_std_linear = nn.Linear(hidden_dim, action_dim)
self.log_std_linear.weight.data.uniform_(-init_w, init_w)
self.log_std_linear.bias.data.uniform_(-init_w, init_w)
def forward(self, state):
x = F.relu(self.linear1(state))
x = F.relu(self.linear2(x))
mean = self.mean_linear(x)
log_std = self.log_std_linear(x)
log_std = torch.clamp(log_std, self.log_std_min, self.log_std_max)
return mean, log_std
def evaluate(self, state, epsilon=1e-06):
mean, log_std = self.forward(state)
std = log_std.exp()
normal = Normal(0, 1)
z = normal.sample()
action = torch.tanh(mean + std * z)
log_prob = Normal(mean, std).log_prob(mean + std * z) - torch.log(1 -
action.pow(2) + epsilon)
return action, log_prob, z, mean, log_std
def get_action(self, state):
state = torch.FloatTensor(state).unsqueeze(0)
mean, log_std = self.forward(state)
std = log_std.exp()
normal = Normal(mean, std)
z = normal.sample()
action = torch.tanh(z)
return action[0]
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'state_dim': 4, 'action_dim': 4, 'hidden_dim': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn as nn
from torch.distributions import Normal
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_relu_threshold_backward_0(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tmp5 = 0.0
tmp6 = tmp4 <= tmp5
tl.store(in_out_ptr0 + x2, tmp4, xmask)
tl.store(out_ptr0 + x2, tmp6, xmask)
@triton.jit
def triton_poi_fused_clamp_ge_le_logical_and_1(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = -20.0
tmp4 = triton_helpers.maximum(tmp2, tmp3)
tmp5 = 2.0
tmp6 = triton_helpers.minimum(tmp4, tmp5)
tmp7 = tmp2 >= tmp3
tmp8 = tmp2 <= tmp5
tmp9 = tmp7 & tmp8
tl.store(out_ptr0 + x2, tmp6, xmask)
tl.store(out_ptr1 + x2, tmp9, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 4), (4, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
buf9 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
get_raw_stream(0)
triton_poi_fused_relu_threshold_backward_0[grid(256)](buf1,
primals_2, buf9, 256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf1, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 4), (1, 4), 0), out=buf2)
buf3 = reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf2
buf8 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
triton_poi_fused_relu_threshold_backward_0[grid(256)](buf3,
primals_5, buf8, 256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_5
buf4 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_7, reinterpret_tensor(buf3, (64, 4), (
4, 1), 0), reinterpret_tensor(primals_6, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf4)
del primals_7
buf5 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_8, (4, 4), (1, 4), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
buf7 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool)
triton_poi_fused_clamp_ge_le_logical_and_1[grid(256)](buf5,
primals_9, buf6, buf7, 256, XBLOCK=128, num_warps=4, num_stages=1)
del buf5
del primals_9
return reinterpret_tensor(buf4, (4, 4, 4, 4), (64, 16, 4, 1), 0
), buf6, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), reinterpret_tensor(buf1, (64, 4), (4, 1), 0), reinterpret_tensor(
buf3, (64, 4), (4, 1), 0
), buf7, primals_8, primals_6, buf8, primals_4, buf9
class PolicyNetworkNew(nn.Module):
def __init__(self, state_dim, action_dim, hidden_dim, init_w=0.003,
log_std_min=-20, log_std_max=2):
super(PolicyNetworkNew, self).__init__()
self.log_std_min = log_std_min
self.log_std_max = log_std_max
self.linear1 = nn.Linear(state_dim, hidden_dim)
self.linear2 = nn.Linear(hidden_dim, hidden_dim)
self.mean_linear = nn.Linear(hidden_dim, action_dim)
self.mean_linear.weight.data.uniform_(-init_w, init_w)
self.mean_linear.bias.data.uniform_(-init_w, init_w)
self.log_std_linear = nn.Linear(hidden_dim, action_dim)
self.log_std_linear.weight.data.uniform_(-init_w, init_w)
self.log_std_linear.bias.data.uniform_(-init_w, init_w)
def evaluate(self, state, epsilon=1e-06):
mean, log_std = self.forward(state)
std = log_std.exp()
normal = Normal(0, 1)
z = normal.sample()
action = torch.tanh(mean + std * z)
log_prob = Normal(mean, std).log_prob(mean + std * z) - torch.log(1 -
action.pow(2) + epsilon)
return action, log_prob, z, mean, log_std
def get_action(self, state):
state = torch.FloatTensor(state).unsqueeze(0)
mean, log_std = self.forward(state)
std = log_std.exp()
normal = Normal(mean, std)
z = normal.sample()
action = torch.tanh(z)
return action[0]
def forward(self, input_0):
primals_1 = self.linear1.weight
primals_2 = self.linear1.bias
primals_4 = self.linear2.weight
primals_5 = self.linear2.bias
primals_6 = self.mean_linear.weight
primals_7 = self.mean_linear.bias
primals_8 = self.log_std_linear.weight
primals_9 = self.log_std_linear.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0], output[1]
|
rtharungowda/Soft-Actor-Critic-Pytorch
|
PolicyNetwork
| false
| 4,213
|
[
"MIT"
] | 0
|
0d2c20c6cfd4e578e0b7cff4525ddf0bc956812f
|
https://github.com/rtharungowda/Soft-Actor-Critic-Pytorch/tree/0d2c20c6cfd4e578e0b7cff4525ddf0bc956812f
|
Depth_Pointwise_Conv1d
|
import torch
from torch import nn
class Depth_Pointwise_Conv1d(nn.Module):
def __init__(self, in_ch, out_ch, k):
super().__init__()
if k == 1:
self.depth_conv = nn.Identity()
else:
self.depth_conv = nn.Conv1d(in_channels=in_ch, out_channels=
in_ch, kernel_size=k, groups=in_ch, padding=k // 2)
self.pointwise_conv = nn.Conv1d(in_channels=in_ch, out_channels=
out_ch, kernel_size=1, groups=1)
def forward(self, x):
out = self.pointwise_conv(self.depth_conv(x))
return out
def get_inputs():
return [torch.rand([4, 4])]
def get_init_inputs():
return [[], {'in_ch': 4, 'out_ch': 4, 'k': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_convolution_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 20
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 5
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 1, 4), (4, 4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4), (4, 1))
assert_size_stride(primals_4, (4, 4, 1), (4, 1, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(reinterpret_tensor(primals_3, (1,
4, 4), (16, 4, 1), 0), primals_1, stride=(1,), padding=(2,),
dilation=(1,), transposed=False, output_padding=(0,), groups=4,
bias=None)
assert_size_stride(buf0, (1, 4, 5), (20, 5, 1))
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_convolution_0[grid(20)](buf1, primals_2, 20,
XBLOCK=32, num_warps=1, num_stages=1)
del primals_2
buf2 = extern_kernels.convolution(reinterpret_tensor(buf1, (1, 4, 5
), (0, 5, 1), 0), primals_4, stride=(1,), padding=(0,),
dilation=(1,), transposed=False, output_padding=(0,), groups=1,
bias=None)
assert_size_stride(buf2, (1, 4, 5), (20, 5, 1))
buf3 = buf2
del buf2
triton_poi_fused_convolution_0[grid(20)](buf3, primals_5, 20,
XBLOCK=32, num_warps=1, num_stages=1)
del primals_5
return reinterpret_tensor(buf3, (4, 5), (5, 1), 0
), primals_1, primals_4, reinterpret_tensor(primals_3, (1, 4, 4), (
16, 4, 1), 0), buf1
class Depth_Pointwise_Conv1dNew(nn.Module):
def __init__(self, in_ch, out_ch, k):
super().__init__()
if k == 1:
self.depth_conv = nn.Identity()
else:
self.depth_conv = nn.Conv1d(in_channels=in_ch, out_channels=
in_ch, kernel_size=k, groups=in_ch, padding=k // 2)
self.pointwise_conv = nn.Conv1d(in_channels=in_ch, out_channels=
out_ch, kernel_size=1, groups=1)
def forward(self, input_0):
primals_1 = self.depth_conv.weight
primals_2 = self.depth_conv.bias
primals_4 = self.pointwise_conv.weight
primals_5 = self.pointwise_conv.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
Depth_Pointwise_Conv1d
| false
| 4,214
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
ExternalAttention
|
import torch
from torch import nn
from torch.nn import init
class ExternalAttention(nn.Module):
def __init__(self, d_model, S=64):
super().__init__()
self.mk = nn.Linear(d_model, S, bias=False)
self.mv = nn.Linear(S, d_model, bias=False)
self.softmax = nn.Softmax(dim=1)
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, queries):
attn = self.mk(queries)
attn = self.softmax(attn)
attn = attn / torch.sum(attn, dim=2, keepdim=True)
out = self.mv(attn)
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'d_model': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from torch import nn
from torch.nn import init
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused__softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x0 = xindex % 256
x2 = xindex // 1024
tmp0 = tl.load(in_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + (x0 + 1024 * x2), None, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (256 + x0 + 1024 * x2), None, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (512 + x0 + 1024 * x2), None, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (768 + x0 + 1024 * x2), None, eviction_policy=
'evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x3, tmp9, None)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x0 = xindex % 256
x2 = xindex // 1024
tmp0 = tl.load(in_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + (x0 + 1024 * x2), None, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (256 + x0 + 1024 * x2), None, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (512 + x0 + 1024 * x2), None, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (768 + x0 + 1024 * x2), None, eviction_policy=
'evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x3, tmp8, None)
@triton.jit
def triton_poi_fused_div_sum_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x3 = xindex
x0 = xindex % 64
x2 = xindex // 256
tmp0 = tl.load(in_ptr0 + x3, None)
tmp1 = tl.load(in_ptr0 + (x0 + 256 * x2), None, eviction_policy=
'evict_last')
tmp2 = tl.load(in_ptr0 + (64 + x0 + 256 * x2), None, eviction_policy=
'evict_last')
tmp4 = tl.load(in_ptr0 + (128 + x0 + 256 * x2), None, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr0 + (192 + x0 + 256 * x2), None, eviction_policy=
'evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x3, tmp8, None)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (64, 4), (4, 1))
assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_3, (4, 64), (64, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 64), (64, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_2, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 64), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((4, 4, 4, 64), (1024, 256, 64, 1), torch.
float32)
get_raw_stream(0)
triton_poi_fused__softmax_0[grid(4096)](buf0, buf1, 4096, XBLOCK=
256, num_warps=4, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4, 64), (1024, 256, 64, 1), torch.
float32)
triton_poi_fused__softmax_1[grid(4096)](buf1, buf2, 4096, XBLOCK=
128, num_warps=4, num_stages=1)
buf3 = buf1
del buf1
triton_poi_fused_div_sum_2[grid(4096)](buf2, buf3, 4096, XBLOCK=256,
num_warps=4, num_stages=1)
del buf2
buf4 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf3, (64, 64), (64, 1), 0),
reinterpret_tensor(primals_3, (64, 4), (1, 64), 0), out=buf4)
return reinterpret_tensor(buf4, (4, 4, 4, 4), (64, 16, 4, 1), 0
), reinterpret_tensor(primals_2, (64, 4), (4, 1), 0
), buf0, buf3, primals_3
class ExternalAttentionNew(nn.Module):
def __init__(self, d_model, S=64):
super().__init__()
self.mk = nn.Linear(d_model, S, bias=False)
self.mv = nn.Linear(S, d_model, bias=False)
self.softmax = nn.Softmax(dim=1)
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, input_0):
primals_1 = self.mk.weight
primals_3 = self.mv.weight
primals_2 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
ExternalAttention
| false
| 4,215
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
SpatialAttention
|
import torch
from torch import nn
class SpatialAttention(nn.Module):
def __init__(self, kernel_size=7):
super().__init__()
self.conv = nn.Conv2d(2, 1, kernel_size=kernel_size, padding=
kernel_size // 2)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
max_result, _ = torch.max(x, dim=1, keepdim=True)
avg_result = torch.mean(x, dim=1, keepdim=True)
result = torch.cat([max_result, avg_result], 1)
output = self.conv(result)
output = self.sigmoid(output)
return output
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
@triton.jit
def triton_poi_fused_cat_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 128
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 16 % 2
x0 = xindex % 16
x2 = xindex // 32
x3 = xindex
tmp0 = x1
tl.full([1], 0, tl.int64)
tmp3 = tl.full([1], 1, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = tl.load(in_ptr0 + (x0 + 64 * x2), tmp4 & xmask, eviction_policy=
'evict_last', other=0.0)
tmp6 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp9 = triton_helpers.maximum(tmp7, tmp8)
tmp10 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), tmp4 & xmask,
eviction_policy='evict_last', other=0.0)
tmp11 = triton_helpers.maximum(tmp9, tmp10)
tmp12 = tl.full(tmp11.shape, 0.0, tmp11.dtype)
tmp13 = tl.where(tmp4, tmp11, tmp12)
tmp14 = tmp0 >= tmp3
tl.full([1], 2, tl.int64)
tmp17 = tl.load(in_ptr0 + (x0 + 64 * x2), tmp14 & xmask,
eviction_policy='evict_last', other=0.0)
tmp18 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), tmp14 & xmask,
eviction_policy='evict_last', other=0.0)
tmp19 = tmp17 + tmp18
tmp20 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), tmp14 & xmask,
eviction_policy='evict_last', other=0.0)
tmp21 = tmp19 + tmp20
tmp22 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), tmp14 & xmask,
eviction_policy='evict_last', other=0.0)
tmp23 = tmp21 + tmp22
tmp24 = 4.0
tmp25 = tmp23 / tmp24
tmp26 = tl.full(tmp25.shape, 0.0, tmp25.dtype)
tmp27 = tl.where(tmp14, tmp25, tmp26)
tmp28 = tl.where(tmp4, tmp13, tmp27)
tl.store(out_ptr0 + x3, tmp28, xmask)
@triton.jit
def triton_poi_fused_convolution_sigmoid_1(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr0 + 0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK])
tmp3 = tmp0 + tmp2
tmp4 = tl.sigmoid(tmp3)
tl.store(in_out_ptr0 + x0, tmp4, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (1, 2, 7, 7), (98, 49, 7, 1))
assert_size_stride(primals_3, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 2, 4, 4), (32, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_cat_0[grid(128)](primals_1, buf0, 128, XBLOCK=128,
num_warps=4, num_stages=1)
del primals_1
buf1 = extern_kernels.convolution(buf0, primals_2, stride=(1, 1),
padding=(3, 3), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 1, 4, 4), (16, 16, 4, 1))
buf2 = buf1
del buf1
triton_poi_fused_convolution_sigmoid_1[grid(64)](buf2, primals_3,
64, XBLOCK=64, num_warps=1, num_stages=1)
del primals_3
return buf2, primals_2, buf0, buf2
class SpatialAttentionNew(nn.Module):
def __init__(self, kernel_size=7):
super().__init__()
self.conv = nn.Conv2d(2, 1, kernel_size=kernel_size, padding=
kernel_size // 2)
self.sigmoid = nn.Sigmoid()
def forward(self, input_0):
primals_2 = self.conv.weight
primals_3 = self.conv.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
SpatialAttention
| false
| 4,216
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
GTXSelfAttentionLayer
|
from _paritybench_helpers import _mock_config
import math
import torch
from torch import nn
class GTXAttention(nn.Module):
def __init__(self, config, ctx_dim=None):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
'The hidden size (%d) is not a multiple of the number of attention heads (%d)'
% (config.hidden_size, config.num_attention_heads))
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.
num_attention_heads)
self.head_size = self.num_attention_heads * self.attention_head_size
if ctx_dim is None:
ctx_dim = config.hidden_size
self.query = nn.Linear(config.hidden_size, self.head_size)
self.key = nn.Linear(ctx_dim, self.head_size)
self.value = nn.Linear(ctx_dim, self.head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, hidden_states, context, attention_mask=None,
output_attentions=False):
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(context)
mixed_value_layer = self.value(context)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
outputs = (context_layer, attention_probs) if output_attentions else (
context_layer,)
return outputs
class GTXAttentionOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor, KnowMix_indices=None):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
if KnowMix_indices is None:
hidden_states = input_tensor + hidden_states
else:
if isinstance(KnowMix_indices, int):
input_tensor[:, KnowMix_indices] = input_tensor[:,
KnowMix_indices] + hidden_states.squeeze(1)
else:
input_tensor[KnowMix_indices, :] = input_tensor[
KnowMix_indices, :] + hidden_states.squeeze(1)
hidden_states = input_tensor
hidden_states = self.LayerNorm(hidden_states)
return hidden_states
class GTXSelfAttentionLayer(nn.Module):
def __init__(self, config):
super().__init__()
self.self = GTXAttention(config)
self.output = GTXAttentionOutput(config)
def forward(self, input_tensor, attention_mask, output_attentions=False):
output = self.self(input_tensor, input_tensor, attention_mask,
output_attentions=output_attentions)
if output_attentions:
attention_probs = output[1]
attention_output = self.output(output[0], input_tensor)
outputs = (attention_output, attention_probs
) if output_attentions else (attention_output,)
return outputs
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'config': _mock_config(hidden_size=4, num_attention_heads=
4, attention_probs_dropout_prob=0.5, hidden_dropout_prob=0.5)}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import math
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(out_ptr0 + (x2 + 4 * y3), tmp4, xmask & ymask)
@triton.jit
def triton_poi_fused_1(in_ptr0, in_ptr1, out_ptr0, out_ptr1, out_ptr2,
xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 16
tmp0 = tl.load(in_ptr0 + 4 * x2, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x2), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x2), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = triton_helpers.maximum(tmp2, tmp5)
tmp9 = tmp7 + tmp8
tmp10 = triton_helpers.maximum(tmp6, tmp9)
tmp13 = tmp11 + tmp12
tmp14 = triton_helpers.maximum(tmp10, tmp13)
tmp15 = tmp2 - tmp14
tmp16 = tl_math.exp(tmp15)
tmp17 = tmp5 - tmp14
tmp18 = tl_math.exp(tmp17)
tmp19 = tmp16 + tmp18
tmp20 = tmp9 - tmp14
tmp21 = tl_math.exp(tmp20)
tmp22 = tmp19 + tmp21
tmp23 = tmp13 - tmp14
tmp24 = tl_math.exp(tmp23)
tmp25 = tmp22 + tmp24
tmp26 = float('-inf')
tmp27 = tmp2 == tmp26
tmp28 = tmp27 == 0
tmp29 = tmp28.to(tl.int64)
tmp30 = tmp29 != 0
tmp31 = tmp5 == tmp26
tmp32 = tmp31 == 0
tmp33 = tmp32.to(tl.int64)
tmp34 = tmp33 != 0
tmp35 = tmp30 | tmp34
tmp36 = tmp9 == tmp26
tmp37 = tmp36 == 0
tmp38 = tmp37.to(tl.int64)
tmp39 = tmp38 != 0
tmp40 = tmp35 | tmp39
tmp41 = tmp13 == tmp26
tmp42 = tmp41 == 0
tmp43 = tmp42.to(tl.int64)
tmp44 = tmp43 != 0
tmp45 = tmp40 | tmp44
tl.store(out_ptr0 + x2, tmp14, xmask)
tl.store(out_ptr1 + x2, tmp25, xmask)
tl.store(out_ptr2 + x2, tmp45, xmask)
@triton.jit
def triton_poi_fused_2(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, in_ptr3,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex // 4
x4 = xindex
x5 = xindex % 64
tmp0 = tl.load(in_ptr0 + x3, xmask, eviction_policy='evict_last').to(tl
.int1)
tmp2 = tl.load(in_out_ptr0 + x4, xmask)
tmp3 = tl.load(in_ptr1 + x5, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr2 + x3, xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr3 + x3, xmask, eviction_policy='evict_last')
tmp1 = tmp0 == 0
tmp4 = tmp2 + tmp3
tmp6 = tmp4 - tmp5
tmp7 = tl_math.exp(tmp6)
tmp9 = tmp7 / tmp8
tmp10 = 0.0
tmp11 = tl.where(tmp1, tmp10, tmp9)
tl.store(in_out_ptr0 + x4, tmp11, xmask)
@triton.jit
def triton_poi_fused_3(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_5(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(out_ptr1 + x0, tmp28, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_6(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-12
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12
) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_9, (4, 4), (4, 1))
assert_size_stride(primals_10, (4,), (1,))
assert_size_stride(primals_11, (4,), (1,))
assert_size_stride(primals_12, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 4), (1, 4), 0), out=buf1)
del primals_4
buf2 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), out=buf2)
del primals_6
buf3 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(16, 4)](buf0, primals_2, buf3, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_2
buf4 = reinterpret_tensor(buf0, (4, 4, 1, 4), (16, 4, 4, 1), 0)
del buf0
triton_poi_fused_0[grid(16, 4)](buf1, primals_5, buf4, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_5
buf5 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf4, (16, 1, 4), (4, 0, 1), 0), out=buf5)
buf6 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 64), 0)
del buf1
buf7 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf8 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.bool)
triton_poi_fused_1[grid(64)](buf5, primals_8, buf6, buf7, buf8, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf9 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused_2[grid(256)](buf9, buf8, primals_8, buf6, buf7,
256, XBLOCK=128, num_warps=4, num_stages=1)
del buf8
del primals_8
buf10 = reinterpret_tensor(buf7, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf7
triton_poi_fused_3[grid(16, 4)](buf2, primals_7, buf10, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_7
buf11 = reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf9, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf10, (16, 4, 1), (4, 1, 0), 0), out=buf11)
buf12 = reinterpret_tensor(buf6, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf6
triton_poi_fused_clone_4[grid(16, 4)](buf11, buf12, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf13 = reinterpret_tensor(buf11, (16, 4), (4, 1), 0)
del buf11
extern_kernels.addmm(primals_10, reinterpret_tensor(buf12, (16, 4),
(4, 1), 0), reinterpret_tensor(primals_9, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf13)
del primals_10
buf14 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf15 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
triton_poi_fused_add_native_layer_norm_5[grid(16)](primals_3, buf13,
buf14, buf15, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf16 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_native_layer_norm_6[grid(64)](primals_3, buf13,
buf14, buf15, primals_11, primals_12, buf16, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf14
del buf15
del primals_12
return buf16, primals_3, primals_11, buf9, reinterpret_tensor(buf10, (
16, 1, 4), (4, 1, 1), 0), reinterpret_tensor(buf3, (16, 1, 4), (4,
1, 1), 0), reinterpret_tensor(buf4, (16, 4, 1), (4, 1, 4), 0
), reinterpret_tensor(buf12, (16, 4), (4, 1), 0), buf13, primals_9
class GTXAttention(nn.Module):
def __init__(self, config, ctx_dim=None):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
'The hidden size (%d) is not a multiple of the number of attention heads (%d)'
% (config.hidden_size, config.num_attention_heads))
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.
num_attention_heads)
self.head_size = self.num_attention_heads * self.attention_head_size
if ctx_dim is None:
ctx_dim = config.hidden_size
self.query = nn.Linear(config.hidden_size, self.head_size)
self.key = nn.Linear(ctx_dim, self.head_size)
self.value = nn.Linear(ctx_dim, self.head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, hidden_states, context, attention_mask=None,
output_attentions=False):
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(context)
mixed_value_layer = self.value(context)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
outputs = (context_layer, attention_probs) if output_attentions else (
context_layer,)
return outputs
class GTXAttentionOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor, KnowMix_indices=None):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
if KnowMix_indices is None:
hidden_states = input_tensor + hidden_states
else:
if isinstance(KnowMix_indices, int):
input_tensor[:, KnowMix_indices] = input_tensor[:,
KnowMix_indices] + hidden_states.squeeze(1)
else:
input_tensor[KnowMix_indices, :] = input_tensor[
KnowMix_indices, :] + hidden_states.squeeze(1)
hidden_states = input_tensor
hidden_states = self.LayerNorm(hidden_states)
return hidden_states
class GTXSelfAttentionLayerNew(nn.Module):
def __init__(self, config):
super().__init__()
self.self = GTXAttention(config)
self.output = GTXAttentionOutput(config)
def forward(self, input_0, input_1):
primals_1 = self.self.query.weight
primals_2 = self.self.query.bias
primals_4 = self.self.key.weight
primals_5 = self.self.key.bias
primals_6 = self.self.value.weight
primals_7 = self.self.value.bias
primals_9 = self.output.dense.weight
primals_10 = self.output.dense.bias
primals_11 = self.output.LayerNorm.weight
primals_12 = self.output.LayerNorm.bias
primals_3 = input_0
primals_8 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12])
return output[0]
|
rsgit95/med_kg_txt_multimodal
|
GTXSelfAttentionLayer
| false
| 4,217
|
[
"Apache-2.0"
] | 0
|
80355b0cf58e0571531ad6f9728c533110ca996d
|
https://github.com/rsgit95/med_kg_txt_multimodal/tree/80355b0cf58e0571531ad6f9728c533110ca996d
|
MlpBlock
|
import torch
from torch import nn
class MlpBlock(nn.Module):
def __init__(self, input_dim, mlp_dim=512):
super().__init__()
self.fc1 = nn.Linear(input_dim, mlp_dim)
self.gelu = nn.GELU()
self.fc2 = nn.Linear(mlp_dim, input_dim)
def forward(self, x):
return self.fc2(self.gelu(self.fc1(x)))
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'input_dim': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_gelu_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, None)
tmp1 = 0.5
tmp2 = tmp0 * tmp1
tmp3 = 0.7071067811865476
tmp4 = tmp0 * tmp3
tmp5 = libdevice.erf(tmp4)
tmp6 = 1.0
tmp7 = tmp5 + tmp6
tmp8 = tmp2 * tmp7
tl.store(out_ptr0 + x0, tmp8, None)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (512, 4), (4, 1))
assert_size_stride(primals_2, (512,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (4, 512), (512, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 512), (512, 1), torch.float32)
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (64,
4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 512), (1, 4),
0), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((4, 4, 4, 512), (8192, 2048, 512, 1),
torch.float32)
get_raw_stream(0)
triton_poi_fused_gelu_0[grid(32768)](buf0, buf1, 32768, XBLOCK=256,
num_warps=4, num_stages=1)
buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 512),
(512, 1), 0), reinterpret_tensor(primals_4, (512, 4), (1, 512),
0), alpha=1, beta=1, out=buf2)
del primals_5
return reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0
), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf0, reinterpret_tensor(buf1, (64, 512), (512, 1), 0), primals_4
class MlpBlockNew(nn.Module):
def __init__(self, input_dim, mlp_dim=512):
super().__init__()
self.fc1 = nn.Linear(input_dim, mlp_dim)
self.gelu = nn.GELU()
self.fc2 = nn.Linear(mlp_dim, input_dim)
def forward(self, input_0):
primals_1 = self.fc1.weight
primals_2 = self.fc1.bias
primals_4 = self.fc2.weight
primals_5 = self.fc2.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
MlpBlock
| false
| 4,218
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
LxmertCrossAttentionLayer
|
from _paritybench_helpers import _mock_config
import math
import torch
from torch import nn
class LxmertAttention(nn.Module):
def __init__(self, config, ctx_dim=None):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
'The hidden size (%d) is not a multiple of the number of attention heads (%d)'
% (config.hidden_size, config.num_attention_heads))
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.
num_attention_heads)
self.head_size = self.num_attention_heads * self.attention_head_size
if ctx_dim is None:
ctx_dim = config.hidden_size
self.query = nn.Linear(config.hidden_size, self.head_size)
self.key = nn.Linear(ctx_dim, self.head_size)
self.value = nn.Linear(ctx_dim, self.head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, hidden_states, context, attention_mask=None,
output_attentions=False):
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(context)
mixed_value_layer = self.value(context)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
outputs = (context_layer, attention_probs) if output_attentions else (
context_layer,)
return outputs
class LxmertAttentionOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
return hidden_states
class LxmertCrossAttentionLayer(nn.Module):
def __init__(self, config):
super().__init__()
self.att = LxmertAttention(config)
self.output = LxmertAttentionOutput(config)
def forward(self, input_tensor, ctx_tensor, ctx_att_mask=None,
output_attentions=False):
output = self.att(input_tensor, ctx_tensor, ctx_att_mask,
output_attentions=output_attentions)
if output_attentions:
attention_probs = output[1]
attention_output = self.output(output[0], input_tensor)
outputs = (attention_output, attention_probs
) if output_attentions else (attention_output,)
return outputs
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'config': _mock_config(hidden_size=4, num_attention_heads=
4, attention_probs_dropout_prob=0.5, hidden_dropout_prob=0.5)}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import math
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(out_ptr0 + (x2 + 4 * y3), tmp4, xmask & ymask)
@triton.jit
def triton_poi_fused_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp18 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp25 = tl.load(in_ptr1 + x2, xmask)
tmp26 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')
tmp27 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp29 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp31 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = float('-inf')
tmp2 = tmp0 == tmp1
tmp3 = tmp2 == 0
tmp4 = tmp3.to(tl.int64)
tmp5 = tmp4 != 0
tmp7 = tmp6 == tmp1
tmp8 = tmp7 == 0
tmp9 = tmp8.to(tl.int64)
tmp10 = tmp9 != 0
tmp11 = tmp5 | tmp10
tmp13 = tmp12 == tmp1
tmp14 = tmp13 == 0
tmp15 = tmp14.to(tl.int64)
tmp16 = tmp15 != 0
tmp17 = tmp11 | tmp16
tmp19 = tmp18 == tmp1
tmp20 = tmp19 == 0
tmp21 = tmp20.to(tl.int64)
tmp22 = tmp21 != 0
tmp23 = tmp17 | tmp22
tmp24 = tmp23 == 0
tmp28 = tmp26 + tmp27
tmp30 = tmp28 + tmp29
tmp32 = tmp30 + tmp31
tmp33 = tmp25 / tmp32
tmp34 = 0.0
tmp35 = tl.where(tmp24, tmp34, tmp33)
tl.store(out_ptr0 + x2, tmp35, xmask)
@triton.jit
def triton_poi_fused_3(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_5(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(out_ptr1 + x0, tmp28, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_6(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-12
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12
) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_7, (4, 4), (4, 1))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (4, 4), (4, 1))
assert_size_stride(primals_10, (4,), (1,))
assert_size_stride(primals_11, (4,), (1,))
assert_size_stride(primals_12, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_6, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 4), (1, 4), 0), out=buf1)
del primals_4
buf2 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_6, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 4), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(16, 4)](buf0, primals_2, buf3, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_2
buf4 = reinterpret_tensor(buf0, (4, 4, 1, 4), (16, 4, 4, 1), 0)
del buf0
triton_poi_fused_0[grid(16, 4)](buf1, primals_5, buf4, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_5
buf5 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf4, (16, 1, 4), (4, 0, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_1[grid(256)](buf5, buf6, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf7 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_2[grid(256)](buf5, buf6, buf7, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf5
del buf6
buf8 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf1
triton_poi_fused_3[grid(16, 4)](buf2, primals_8, buf8, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_8
buf9 = reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf8, (16, 4, 1), (4, 1, 0), 0), out=buf9)
buf10 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
triton_poi_fused_clone_4[grid(16, 4)](buf9, buf10, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf11 = reinterpret_tensor(buf9, (16, 4), (4, 1), 0)
del buf9
extern_kernels.addmm(primals_10, reinterpret_tensor(buf10, (16, 4),
(4, 1), 0), reinterpret_tensor(primals_9, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf11)
del primals_10
buf12 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf13 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
triton_poi_fused_add_native_layer_norm_5[grid(16)](buf11, primals_3,
buf12, buf13, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf14 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_native_layer_norm_6[grid(64)](buf11, primals_3,
buf12, buf13, primals_11, primals_12, buf14, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf12
del buf13
del primals_12
return buf14, primals_3, primals_11, reinterpret_tensor(primals_6, (16,
4), (4, 1), 0), buf7, reinterpret_tensor(buf8, (16, 1, 4), (4, 1, 1), 0
), reinterpret_tensor(buf3, (16, 1, 4), (4, 1, 1), 0
), reinterpret_tensor(buf4, (16, 4, 1), (4, 1, 4), 0
), reinterpret_tensor(buf10, (16, 4), (4, 1), 0), buf11, primals_9
class LxmertAttention(nn.Module):
def __init__(self, config, ctx_dim=None):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
'The hidden size (%d) is not a multiple of the number of attention heads (%d)'
% (config.hidden_size, config.num_attention_heads))
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.
num_attention_heads)
self.head_size = self.num_attention_heads * self.attention_head_size
if ctx_dim is None:
ctx_dim = config.hidden_size
self.query = nn.Linear(config.hidden_size, self.head_size)
self.key = nn.Linear(ctx_dim, self.head_size)
self.value = nn.Linear(ctx_dim, self.head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, hidden_states, context, attention_mask=None,
output_attentions=False):
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(context)
mixed_value_layer = self.value(context)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
outputs = (context_layer, attention_probs) if output_attentions else (
context_layer,)
return outputs
class LxmertAttentionOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
return hidden_states
class LxmertCrossAttentionLayerNew(nn.Module):
def __init__(self, config):
super().__init__()
self.att = LxmertAttention(config)
self.output = LxmertAttentionOutput(config)
def forward(self, input_0, input_1):
primals_1 = self.att.query.weight
primals_2 = self.att.query.bias
primals_4 = self.att.key.weight
primals_5 = self.att.key.bias
primals_7 = self.att.value.weight
primals_8 = self.att.value.bias
primals_9 = self.output.dense.weight
primals_10 = self.output.dense.bias
primals_11 = self.output.LayerNorm.weight
primals_12 = self.output.LayerNorm.bias
primals_3 = input_0
primals_6 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12])
return output[0]
|
rsgit95/med_kg_txt_multimodal
|
LxmertCrossAttentionLayer
| false
| 4,219
|
[
"Apache-2.0"
] | 0
|
80355b0cf58e0571531ad6f9728c533110ca996d
|
https://github.com/rsgit95/med_kg_txt_multimodal/tree/80355b0cf58e0571531ad6f9728c533110ca996d
|
GTXCrossAttentionLayer
|
from _paritybench_helpers import _mock_config
import math
import torch
from torch import nn
class GTXAttention(nn.Module):
def __init__(self, config, ctx_dim=None):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
'The hidden size (%d) is not a multiple of the number of attention heads (%d)'
% (config.hidden_size, config.num_attention_heads))
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.
num_attention_heads)
self.head_size = self.num_attention_heads * self.attention_head_size
if ctx_dim is None:
ctx_dim = config.hidden_size
self.query = nn.Linear(config.hidden_size, self.head_size)
self.key = nn.Linear(ctx_dim, self.head_size)
self.value = nn.Linear(ctx_dim, self.head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, hidden_states, context, attention_mask=None,
output_attentions=False):
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(context)
mixed_value_layer = self.value(context)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
outputs = (context_layer, attention_probs) if output_attentions else (
context_layer,)
return outputs
class GTXAttentionOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor, KnowMix_indices=None):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
if KnowMix_indices is None:
hidden_states = input_tensor + hidden_states
else:
if isinstance(KnowMix_indices, int):
input_tensor[:, KnowMix_indices] = input_tensor[:,
KnowMix_indices] + hidden_states.squeeze(1)
else:
input_tensor[KnowMix_indices, :] = input_tensor[
KnowMix_indices, :] + hidden_states.squeeze(1)
hidden_states = input_tensor
hidden_states = self.LayerNorm(hidden_states)
return hidden_states
class GTXCrossAttentionLayer(nn.Module):
def __init__(self, config):
super().__init__()
self.att = GTXAttention(config)
self.output = GTXAttentionOutput(config)
def forward(self, input_tensor, ctx_tensor, ctx_att_mask=None,
KnowMix_indices=None, output_attentions=False):
if KnowMix_indices is None:
output = self.att(input_tensor, ctx_tensor, ctx_att_mask,
output_attentions=output_attentions)
elif isinstance(KnowMix_indices, int):
output = self.att(input_tensor[:, KnowMix_indices].unsqueeze(1),
ctx_tensor, ctx_att_mask, output_attentions=output_attentions)
else:
output = self.att(input_tensor[KnowMix_indices, :].unsqueeze(1),
ctx_tensor[KnowMix_indices, :], ctx_att_mask[
KnowMix_indices.unsqueeze(1), :].unsqueeze(1).unsqueeze(2),
output_attentions=output_attentions)
if output_attentions:
attention_probs = output[1]
attention_output = self.output(output[0], input_tensor, KnowMix_indices
)
outputs = (attention_output, attention_probs
) if output_attentions else (attention_output,)
return outputs
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'config': _mock_config(hidden_size=4, num_attention_heads=
4, attention_probs_dropout_prob=0.5, hidden_dropout_prob=0.5)}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import math
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(out_ptr0 + (x2 + 4 * y3), tmp4, xmask & ymask)
@triton.jit
def triton_poi_fused_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp18 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp25 = tl.load(in_ptr1 + x2, xmask)
tmp26 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last')
tmp27 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp29 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp31 = tl.load(in_ptr1 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = float('-inf')
tmp2 = tmp0 == tmp1
tmp3 = tmp2 == 0
tmp4 = tmp3.to(tl.int64)
tmp5 = tmp4 != 0
tmp7 = tmp6 == tmp1
tmp8 = tmp7 == 0
tmp9 = tmp8.to(tl.int64)
tmp10 = tmp9 != 0
tmp11 = tmp5 | tmp10
tmp13 = tmp12 == tmp1
tmp14 = tmp13 == 0
tmp15 = tmp14.to(tl.int64)
tmp16 = tmp15 != 0
tmp17 = tmp11 | tmp16
tmp19 = tmp18 == tmp1
tmp20 = tmp19 == 0
tmp21 = tmp20.to(tl.int64)
tmp22 = tmp21 != 0
tmp23 = tmp17 | tmp22
tmp24 = tmp23 == 0
tmp28 = tmp26 + tmp27
tmp30 = tmp28 + tmp29
tmp32 = tmp30 + tmp31
tmp33 = tmp25 / tmp32
tmp34 = 0.0
tmp35 = tl.where(tmp24, tmp34, tmp33)
tl.store(out_ptr0 + x2, tmp35, xmask)
@triton.jit
def triton_poi_fused_3(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_5(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(out_ptr1 + x0, tmp28, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_6(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-12
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12
) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_7, (4, 4), (4, 1))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (4, 4), (4, 1))
assert_size_stride(primals_10, (4,), (1,))
assert_size_stride(primals_11, (4,), (1,))
assert_size_stride(primals_12, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_6, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 4), (1, 4), 0), out=buf1)
del primals_4
buf2 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_6, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 4), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(16, 4)](buf0, primals_2, buf3, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_2
buf4 = reinterpret_tensor(buf0, (4, 4, 1, 4), (16, 4, 4, 1), 0)
del buf0
triton_poi_fused_0[grid(16, 4)](buf1, primals_5, buf4, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_5
buf5 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf4, (16, 1, 4), (4, 0, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_1[grid(256)](buf5, buf6, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf7 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_2[grid(256)](buf5, buf6, buf7, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf5
del buf6
buf8 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf1
triton_poi_fused_3[grid(16, 4)](buf2, primals_8, buf8, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_8
buf9 = reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf8, (16, 4, 1), (4, 1, 0), 0), out=buf9)
buf10 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
triton_poi_fused_clone_4[grid(16, 4)](buf9, buf10, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf11 = reinterpret_tensor(buf9, (16, 4), (4, 1), 0)
del buf9
extern_kernels.addmm(primals_10, reinterpret_tensor(buf10, (16, 4),
(4, 1), 0), reinterpret_tensor(primals_9, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf11)
del primals_10
buf12 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf13 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
triton_poi_fused_add_native_layer_norm_5[grid(16)](primals_3, buf11,
buf12, buf13, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf14 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_native_layer_norm_6[grid(64)](primals_3, buf11,
buf12, buf13, primals_11, primals_12, buf14, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf12
del buf13
del primals_12
return buf14, primals_3, primals_11, reinterpret_tensor(primals_6, (16,
4), (4, 1), 0), buf7, reinterpret_tensor(buf8, (16, 1, 4), (4, 1, 1), 0
), reinterpret_tensor(buf3, (16, 1, 4), (4, 1, 1), 0
), reinterpret_tensor(buf4, (16, 4, 1), (4, 1, 4), 0
), reinterpret_tensor(buf10, (16, 4), (4, 1), 0), buf11, primals_9
class GTXAttention(nn.Module):
def __init__(self, config, ctx_dim=None):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
'The hidden size (%d) is not a multiple of the number of attention heads (%d)'
% (config.hidden_size, config.num_attention_heads))
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.
num_attention_heads)
self.head_size = self.num_attention_heads * self.attention_head_size
if ctx_dim is None:
ctx_dim = config.hidden_size
self.query = nn.Linear(config.hidden_size, self.head_size)
self.key = nn.Linear(ctx_dim, self.head_size)
self.value = nn.Linear(ctx_dim, self.head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, hidden_states, context, attention_mask=None,
output_attentions=False):
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(context)
mixed_value_layer = self.value(context)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
outputs = (context_layer, attention_probs) if output_attentions else (
context_layer,)
return outputs
class GTXAttentionOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor, KnowMix_indices=None):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
if KnowMix_indices is None:
hidden_states = input_tensor + hidden_states
else:
if isinstance(KnowMix_indices, int):
input_tensor[:, KnowMix_indices] = input_tensor[:,
KnowMix_indices] + hidden_states.squeeze(1)
else:
input_tensor[KnowMix_indices, :] = input_tensor[
KnowMix_indices, :] + hidden_states.squeeze(1)
hidden_states = input_tensor
hidden_states = self.LayerNorm(hidden_states)
return hidden_states
class GTXCrossAttentionLayerNew(nn.Module):
def __init__(self, config):
super().__init__()
self.att = GTXAttention(config)
self.output = GTXAttentionOutput(config)
def forward(self, input_0, input_1):
primals_1 = self.att.query.weight
primals_2 = self.att.query.bias
primals_4 = self.att.key.weight
primals_5 = self.att.key.bias
primals_7 = self.att.value.weight
primals_8 = self.att.value.bias
primals_9 = self.output.dense.weight
primals_10 = self.output.dense.bias
primals_11 = self.output.LayerNorm.weight
primals_12 = self.output.LayerNorm.bias
primals_3 = input_0
primals_6 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12])
return output[0]
|
rsgit95/med_kg_txt_multimodal
|
GTXCrossAttentionLayer
| false
| 4,220
|
[
"Apache-2.0"
] | 0
|
80355b0cf58e0571531ad6f9728c533110ca996d
|
https://github.com/rsgit95/med_kg_txt_multimodal/tree/80355b0cf58e0571531ad6f9728c533110ca996d
|
ConvEncoder
|
import torch
from torch import nn
class ConvEncoder(nn.Module):
""" Simple convolutional encoder network.
It consists of 5 convolutional layers, each downsampling the input by a
factor of 2, and a final fully-connected layer projecting the output to
c_dim dimensions.
Args:
c_dim (int): output dimension of latent embedding
"""
def __init__(self, c_dim=128):
super().__init__()
self.conv0 = nn.Conv2d(3, 32, 3, stride=2)
self.conv1 = nn.Conv2d(32, 64, 3, stride=2)
self.conv2 = nn.Conv2d(64, 128, 3, stride=2)
self.conv3 = nn.Conv2d(128, 256, 3, stride=2)
self.conv4 = nn.Conv2d(256, 512, 3, stride=2)
self.fc_out = nn.Linear(512, c_dim)
self.actvn = nn.ReLU()
def forward(self, x):
batch_size = x.size(0)
net = self.conv0(x)
net = self.conv1(self.actvn(net))
net = self.conv2(self.actvn(net))
net = self.conv3(self.actvn(net))
net = self.conv4(self.actvn(net))
net = net.view(batch_size, 512, -1).mean(2)
out = self.fc_out(self.actvn(net))
return out
def get_inputs():
return [torch.rand([4, 3, 64, 64])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 12
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
x2 = xindex
y3 = yindex
y0 = yindex % 3
y1 = yindex // 3
tmp0 = tl.load(in_ptr0 + (x2 + 4096 * y3), ymask, eviction_policy=
'evict_last')
tl.store(out_ptr0 + (y0 + 3 * x2 + 12288 * y1), tmp0, ymask)
@triton.jit
def triton_poi_fused_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 96
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 3
y1 = yindex // 3
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask & ymask, eviction_policy=
'evict_last')
tl.store(out_ptr0 + (y0 + 3 * x2 + 27 * y1), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_2(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 32
y1 = yindex // 32
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 32 * x2 + 288 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_3(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 64
y1 = yindex // 64
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 64 * x2 + 576 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 128
y1 = yindex // 128
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 128 * x2 + 1152 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_5(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = (tl.program_id(1) + tl.program_id(2) * tl.num_programs(1)
) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 256
y1 = yindex // 256
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 256 * x2 + 2304 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_6(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 123008
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 32
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_7(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 57600
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 64
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_8(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 25088
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 128
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_convolution_relu_9(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 9216
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 256
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = tl.full([1], 0, tl.int32)
tmp4 = triton_helpers.maximum(tmp3, tmp2)
tl.store(in_out_ptr0 + x2, tmp4, xmask)
@triton.jit
def triton_poi_fused_mean_relu_10(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.
constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 512
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 / tmp3
tmp5 = tl.full([1], 0, tl.int32)
tmp6 = triton_helpers.maximum(tmp5, tmp4)
tl.store(in_out_ptr0 + x2, tmp6, None)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12,
primals_13) = args
args.clear()
assert_size_stride(primals_1, (4, 3, 64, 64), (12288, 4096, 64, 1))
assert_size_stride(primals_2, (32, 3, 3, 3), (27, 9, 3, 1))
assert_size_stride(primals_3, (32,), (1,))
assert_size_stride(primals_4, (64, 32, 3, 3), (288, 9, 3, 1))
assert_size_stride(primals_5, (64,), (1,))
assert_size_stride(primals_6, (128, 64, 3, 3), (576, 9, 3, 1))
assert_size_stride(primals_7, (128,), (1,))
assert_size_stride(primals_8, (256, 128, 3, 3), (1152, 9, 3, 1))
assert_size_stride(primals_9, (256,), (1,))
assert_size_stride(primals_10, (512, 256, 3, 3), (2304, 9, 3, 1))
assert_size_stride(primals_11, (512,), (1,))
assert_size_stride(primals_12, (128, 512), (512, 1))
assert_size_stride(primals_13, (128,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 3, 64, 64), (12288, 1, 192, 3), torch
.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(12, 4096)](primals_1, buf0, 12, 4096,
XBLOCK=64, YBLOCK=16, num_warps=4, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((32, 3, 3, 3), (27, 1, 9, 3), torch.float32)
triton_poi_fused_1[grid(96, 9)](primals_2, buf1, 96, 9, XBLOCK=16,
YBLOCK=64, num_warps=4, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((64, 32, 3, 3), (288, 1, 96, 32), torch.
float32)
triton_poi_fused_2[grid(2048, 9)](primals_4, buf2, 2048, 9, XBLOCK=
16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_4
buf3 = empty_strided_cuda((128, 64, 3, 3), (576, 1, 192, 64), torch
.float32)
triton_poi_fused_3[grid(8192, 9)](primals_6, buf3, 8192, 9, XBLOCK=
16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_6
buf4 = empty_strided_cuda((256, 128, 3, 3), (1152, 1, 384, 128),
torch.float32)
triton_poi_fused_4[grid(32768, 9)](primals_8, buf4, 32768, 9,
XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_8
buf5 = empty_strided_cuda((512, 256, 3, 3), (2304, 1, 768, 256),
torch.float32)
triton_poi_fused_5[grid(131072, 9)](primals_10, buf5, 131072, 9,
XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_10
buf6 = extern_kernels.convolution(buf0, buf1, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf6, (4, 32, 31, 31), (30752, 1, 992, 32))
buf7 = buf6
del buf6
triton_poi_fused_convolution_relu_6[grid(123008)](buf7, primals_3,
123008, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_3
buf8 = extern_kernels.convolution(buf7, buf2, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf8, (4, 64, 15, 15), (14400, 1, 960, 64))
buf9 = buf8
del buf8
triton_poi_fused_convolution_relu_7[grid(57600)](buf9, primals_5,
57600, XBLOCK=512, num_warps=4, num_stages=1)
del primals_5
buf10 = extern_kernels.convolution(buf9, buf3, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf10, (4, 128, 7, 7), (6272, 1, 896, 128))
buf11 = buf10
del buf10
triton_poi_fused_convolution_relu_8[grid(25088)](buf11, primals_7,
25088, XBLOCK=256, num_warps=4, num_stages=1)
del primals_7
buf12 = extern_kernels.convolution(buf11, buf4, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf12, (4, 256, 3, 3), (2304, 1, 768, 256))
buf13 = buf12
del buf12
triton_poi_fused_convolution_relu_9[grid(9216)](buf13, primals_9,
9216, XBLOCK=256, num_warps=4, num_stages=1)
del primals_9
buf14 = extern_kernels.convolution(buf13, buf5, stride=(2, 2),
padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf14, (4, 512, 1, 1), (512, 1, 512, 512))
buf15 = reinterpret_tensor(buf14, (4, 512), (512, 1), 0)
del buf14
triton_poi_fused_mean_relu_10[grid(2048)](buf15, primals_11, 2048,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_11
buf16 = empty_strided_cuda((4, 128), (128, 1), torch.float32)
extern_kernels.addmm(primals_13, buf15, reinterpret_tensor(
primals_12, (512, 128), (1, 512), 0), alpha=1, beta=1, out=buf16)
del primals_13
return (buf16, buf0, buf1, buf2, buf3, buf4, buf5, buf7, buf9, buf11,
buf13, buf15, primals_12)
class ConvEncoderNew(nn.Module):
""" Simple convolutional encoder network.
It consists of 5 convolutional layers, each downsampling the input by a
factor of 2, and a final fully-connected layer projecting the output to
c_dim dimensions.
Args:
c_dim (int): output dimension of latent embedding
"""
def __init__(self, c_dim=128):
super().__init__()
self.conv0 = nn.Conv2d(3, 32, 3, stride=2)
self.conv1 = nn.Conv2d(32, 64, 3, stride=2)
self.conv2 = nn.Conv2d(64, 128, 3, stride=2)
self.conv3 = nn.Conv2d(128, 256, 3, stride=2)
self.conv4 = nn.Conv2d(256, 512, 3, stride=2)
self.fc_out = nn.Linear(512, c_dim)
self.actvn = nn.ReLU()
def forward(self, input_0):
primals_2 = self.conv0.weight
primals_3 = self.conv0.bias
primals_4 = self.conv1.weight
primals_5 = self.conv1.bias
primals_6 = self.conv2.weight
primals_7 = self.conv2.bias
primals_8 = self.conv3.weight
primals_9 = self.conv3.bias
primals_10 = self.conv4.weight
primals_11 = self.conv4.bias
primals_12 = self.fc_out.weight
primals_13 = self.fc_out.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12, primals_13])
return output[0]
|
planetceres/differentiable_volumetric_rendering
|
ConvEncoder
| false
| 4,221
|
[
"MIT"
] | 0
|
f2fe46d139244c7642439ced23656db1e7f5c128
|
https://github.com/planetceres/differentiable_volumetric_rendering/tree/f2fe46d139244c7642439ced23656db1e7f5c128
|
DoubleAttention
|
import torch
from torch import nn
from torch.nn import functional as F
from torch.nn import init
class DoubleAttention(nn.Module):
def __init__(self, in_channels, c_m, c_n, reconstruct=True):
super().__init__()
self.in_channels = in_channels
self.reconstruct = reconstruct
self.c_m = c_m
self.c_n = c_n
self.convA = nn.Conv2d(in_channels, c_m, 1)
self.convB = nn.Conv2d(in_channels, c_n, 1)
self.convV = nn.Conv2d(in_channels, c_n, 1)
if self.reconstruct:
self.conv_reconstruct = nn.Conv2d(c_m, in_channels, kernel_size=1)
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, x):
b, c, h, w = x.shape
assert c == self.in_channels
A = self.convA(x)
B = self.convB(x)
V = self.convV(x)
tmpA = A.view(b, self.c_m, -1)
attention_maps = F.softmax(B.view(b, self.c_n, -1))
attention_vectors = F.softmax(V.view(b, self.c_n, -1))
global_descriptors = torch.bmm(tmpA, attention_maps.permute(0, 2, 1))
tmpZ = global_descriptors.matmul(attention_vectors)
tmpZ = tmpZ.view(b, self.c_m, h, w)
if self.reconstruct:
tmpZ = self.conv_reconstruct(tmpZ)
return tmpZ
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'in_channels': 4, 'c_m': 4, 'c_n': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from torch import nn
from torch.nn import init
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_convolution_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, xmask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, in_ptr1, out_ptr0, out_ptr1,
xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 16
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (64 + x2), xmask)
tmp6 = tl.load(in_ptr0 + (128 + x2), xmask)
tmp9 = tl.load(in_ptr0 + (192 + x2), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp3 + tmp1
tmp5 = triton_helpers.maximum(tmp2, tmp4)
tmp7 = tmp6 + tmp1
tmp8 = triton_helpers.maximum(tmp5, tmp7)
tmp10 = tmp9 + tmp1
tmp11 = triton_helpers.maximum(tmp8, tmp10)
tmp12 = tmp2 - tmp11
tmp13 = tl_math.exp(tmp12)
tmp14 = tmp4 - tmp11
tmp15 = tl_math.exp(tmp14)
tmp16 = tmp13 + tmp15
tmp17 = tmp7 - tmp11
tmp18 = tl_math.exp(tmp17)
tmp19 = tmp16 + tmp18
tmp20 = tmp10 - tmp11
tmp21 = tl_math.exp(tmp20)
tmp22 = tmp19 + tmp21
tl.store(out_ptr0 + x2, tmp11, xmask)
tl.store(out_ptr1 + x2, tmp22, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 16 % 4
x4 = xindex % 64
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr1 + x4, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr2 + x4, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp5 = tl_math.exp(tmp4)
tmp7 = tmp5 / tmp6
tl.store(in_out_ptr0 + x3, tmp7, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_3, (4,), (1,))
assert_size_stride(primals_4, (4, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 4, 1, 1), (4, 1, 1, 1))
assert_size_stride(primals_9, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_1, primals_2, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 4, 4), (64, 16, 4, 1))
buf1 = extern_kernels.convolution(primals_1, primals_4, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf1, (4, 4, 4, 4), (64, 16, 4, 1))
buf2 = extern_kernels.convolution(primals_1, primals_6, stride=(1,
1), padding=(0, 0), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf2, (4, 4, 4, 4), (64, 16, 4, 1))
buf3 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_convolution_0[grid(256)](buf3, primals_3, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_3
buf4 = empty_strided_cuda((1, 4, 16), (64, 16, 1), torch.float32)
buf5 = empty_strided_cuda((1, 4, 16), (64, 16, 1), torch.float32)
triton_poi_fused__softmax_1[grid(64)](buf1, primals_5, buf4, buf5,
64, XBLOCK=64, num_warps=1, num_stages=1)
buf6 = reinterpret_tensor(buf1, (4, 4, 16), (64, 16, 1), 0)
del buf1
triton_poi_fused__softmax_2[grid(256)](buf6, primals_5, buf4, buf5,
256, XBLOCK=256, num_warps=4, num_stages=1)
del primals_5
buf7 = buf5
del buf5
buf8 = buf4
del buf4
triton_poi_fused__softmax_1[grid(64)](buf2, primals_7, buf7, buf8,
64, XBLOCK=64, num_warps=1, num_stages=1)
buf9 = reinterpret_tensor(buf2, (4, 4, 16), (64, 16, 1), 0)
del buf2
triton_poi_fused__softmax_2[grid(256)](buf9, primals_7, buf7, buf8,
256, XBLOCK=256, num_warps=4, num_stages=1)
del buf7
del primals_7
buf10 = reinterpret_tensor(buf8, (4, 4, 4), (16, 4, 1), 0)
del buf8
extern_kernels.bmm(reinterpret_tensor(buf3, (4, 4, 16), (64, 16, 1),
0), reinterpret_tensor(buf6, (4, 16, 4), (64, 1, 16), 0), out=buf10
)
buf11 = empty_strided_cuda((4, 4, 16), (64, 16, 1), torch.float32)
extern_kernels.bmm(buf10, buf9, out=buf11)
buf12 = extern_kernels.convolution(reinterpret_tensor(buf11, (4, 4,
4, 4), (64, 16, 4, 1), 0), primals_8, stride=(1, 1), padding=(0,
0), dilation=(1, 1), transposed=False, output_padding=(0, 0),
groups=1, bias=None)
assert_size_stride(buf12, (4, 4, 4, 4), (64, 16, 4, 1))
buf13 = buf12
del buf12
triton_poi_fused_convolution_0[grid(256)](buf13, primals_9, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_9
return (buf13, primals_1, primals_2, primals_4, primals_6, primals_8,
buf6, buf9, reinterpret_tensor(buf11, (4, 4, 4, 4), (64, 16, 4, 1),
0), reinterpret_tensor(buf10, (4, 4, 4), (16, 1, 4), 0),
reinterpret_tensor(buf3, (4, 16, 4), (64, 1, 16), 0))
class DoubleAttentionNew(nn.Module):
def __init__(self, in_channels, c_m, c_n, reconstruct=True):
super().__init__()
self.in_channels = in_channels
self.reconstruct = reconstruct
self.c_m = c_m
self.c_n = c_n
self.convA = nn.Conv2d(in_channels, c_m, 1)
self.convB = nn.Conv2d(in_channels, c_n, 1)
self.convV = nn.Conv2d(in_channels, c_n, 1)
if self.reconstruct:
self.conv_reconstruct = nn.Conv2d(c_m, in_channels, kernel_size=1)
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, input_0):
primals_2 = self.convA.weight
primals_3 = self.convA.bias
primals_4 = self.convB.weight
primals_5 = self.convB.bias
primals_6 = self.convV.weight
primals_7 = self.convV.bias
primals_8 = self.conv_reconstruct.weight
primals_9 = self.conv_reconstruct.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
DoubleAttention
| false
| 4,222
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
LxmertSelfAttentionLayer
|
from _paritybench_helpers import _mock_config
import math
import torch
from torch import nn
class LxmertAttention(nn.Module):
def __init__(self, config, ctx_dim=None):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
'The hidden size (%d) is not a multiple of the number of attention heads (%d)'
% (config.hidden_size, config.num_attention_heads))
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.
num_attention_heads)
self.head_size = self.num_attention_heads * self.attention_head_size
if ctx_dim is None:
ctx_dim = config.hidden_size
self.query = nn.Linear(config.hidden_size, self.head_size)
self.key = nn.Linear(ctx_dim, self.head_size)
self.value = nn.Linear(ctx_dim, self.head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, hidden_states, context, attention_mask=None,
output_attentions=False):
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(context)
mixed_value_layer = self.value(context)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
outputs = (context_layer, attention_probs) if output_attentions else (
context_layer,)
return outputs
class LxmertAttentionOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
return hidden_states
class LxmertSelfAttentionLayer(nn.Module):
def __init__(self, config):
super().__init__()
self.self = LxmertAttention(config)
self.output = LxmertAttentionOutput(config)
def forward(self, input_tensor, attention_mask, output_attentions=False):
output = self.self(input_tensor, input_tensor, attention_mask,
output_attentions=output_attentions)
if output_attentions:
attention_probs = output[1]
attention_output = self.output(output[0], input_tensor)
outputs = (attention_output, attention_probs
) if output_attentions else (attention_output,)
return outputs
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'config': _mock_config(hidden_size=4, num_attention_heads=
4, attention_probs_dropout_prob=0.5, hidden_dropout_prob=0.5)}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import math
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tl.store(out_ptr0 + (x2 + 4 * y3), tmp4, xmask & ymask)
@triton.jit
def triton_poi_fused_1(in_ptr0, in_ptr1, out_ptr0, out_ptr1, out_ptr2,
xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 16
tmp0 = tl.load(in_ptr0 + 4 * x2, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x2), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x2), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = triton_helpers.maximum(tmp2, tmp5)
tmp9 = tmp7 + tmp8
tmp10 = triton_helpers.maximum(tmp6, tmp9)
tmp13 = tmp11 + tmp12
tmp14 = triton_helpers.maximum(tmp10, tmp13)
tmp15 = tmp2 - tmp14
tmp16 = tl_math.exp(tmp15)
tmp17 = tmp5 - tmp14
tmp18 = tl_math.exp(tmp17)
tmp19 = tmp16 + tmp18
tmp20 = tmp9 - tmp14
tmp21 = tl_math.exp(tmp20)
tmp22 = tmp19 + tmp21
tmp23 = tmp13 - tmp14
tmp24 = tl_math.exp(tmp23)
tmp25 = tmp22 + tmp24
tmp26 = float('-inf')
tmp27 = tmp2 == tmp26
tmp28 = tmp27 == 0
tmp29 = tmp28.to(tl.int64)
tmp30 = tmp29 != 0
tmp31 = tmp5 == tmp26
tmp32 = tmp31 == 0
tmp33 = tmp32.to(tl.int64)
tmp34 = tmp33 != 0
tmp35 = tmp30 | tmp34
tmp36 = tmp9 == tmp26
tmp37 = tmp36 == 0
tmp38 = tmp37.to(tl.int64)
tmp39 = tmp38 != 0
tmp40 = tmp35 | tmp39
tmp41 = tmp13 == tmp26
tmp42 = tmp41 == 0
tmp43 = tmp42.to(tl.int64)
tmp44 = tmp43 != 0
tmp45 = tmp40 | tmp44
tl.store(out_ptr0 + x2, tmp14, xmask)
tl.store(out_ptr1 + x2, tmp25, xmask)
tl.store(out_ptr2 + x2, tmp45, xmask)
@triton.jit
def triton_poi_fused_2(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, in_ptr3,
xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex // 4
x4 = xindex
x5 = xindex % 64
tmp0 = tl.load(in_ptr0 + x3, xmask, eviction_policy='evict_last').to(tl
.int1)
tmp2 = tl.load(in_out_ptr0 + x4, xmask)
tmp3 = tl.load(in_ptr1 + x5, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr2 + x3, xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr3 + x3, xmask, eviction_policy='evict_last')
tmp1 = tmp0 == 0
tmp4 = tmp2 + tmp3
tmp6 = tmp4 - tmp5
tmp7 = tl_math.exp(tmp6)
tmp9 = tmp7 / tmp8
tmp10 = 0.0
tmp11 = tl.where(tmp1, tmp10, tmp9)
tl.store(in_out_ptr0 + x4, tmp11, xmask)
@triton.jit
def triton_poi_fused_3(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK:
tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_5(in_ptr0, in_ptr1, out_ptr0,
out_ptr1, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tmp5 = tmp3 + tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 + tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 + tmp12
tmp14 = tmp10 + tmp13
tmp15 = 4.0
tmp16 = tmp14 / tmp15
tmp17 = tmp2 - tmp16
tmp18 = tmp17 * tmp17
tmp19 = tmp5 - tmp16
tmp20 = tmp19 * tmp19
tmp21 = tmp18 + tmp20
tmp22 = tmp9 - tmp16
tmp23 = tmp22 * tmp22
tmp24 = tmp21 + tmp23
tmp25 = tmp13 - tmp16
tmp26 = tmp25 * tmp25
tmp27 = tmp24 + tmp26
tmp28 = tmp27 / tmp15
tl.store(out_ptr0 + x0, tmp16, xmask)
tl.store(out_ptr1 + x0, tmp28, xmask)
@triton.jit
def triton_poi_fused_add_native_layer_norm_6(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, in_ptr5, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
x0 = xindex % 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr1 + x2, xmask)
tmp3 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr4 + x0, xmask, eviction_policy='evict_last')
tmp12 = tl.load(in_ptr5 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp4 = tmp2 - tmp3
tmp6 = 1e-12
tmp7 = tmp5 + tmp6
tmp8 = libdevice.rsqrt(tmp7)
tmp9 = tmp4 * tmp8
tmp11 = tmp9 * tmp10
tmp13 = tmp11 + tmp12
tl.store(out_ptr0 + x2, tmp13, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11, primals_12
) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4), (4, 1))
assert_size_stride(primals_7, (4,), (1,))
assert_size_stride(primals_8, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_9, (4, 4), (4, 1))
assert_size_stride(primals_10, (4,), (1,))
assert_size_stride(primals_11, (4,), (1,))
assert_size_stride(primals_12, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0)
del primals_1
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_4, (4, 4), (1, 4), 0), out=buf1)
del primals_4
buf2 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), out=buf2)
del primals_6
buf3 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(16, 4)](buf0, primals_2, buf3, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_2
buf4 = reinterpret_tensor(buf0, (4, 4, 1, 4), (16, 4, 4, 1), 0)
del buf0
triton_poi_fused_0[grid(16, 4)](buf1, primals_5, buf4, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_5
buf5 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf4, (16, 1, 4), (4, 0, 1), 0), out=buf5)
buf6 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 64), 0)
del buf1
buf7 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32)
buf8 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.bool)
triton_poi_fused_1[grid(64)](buf5, primals_8, buf6, buf7, buf8, 64,
XBLOCK=64, num_warps=1, num_stages=1)
buf9 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused_2[grid(256)](buf9, buf8, primals_8, buf6, buf7,
256, XBLOCK=128, num_warps=4, num_stages=1)
del buf8
del primals_8
buf10 = reinterpret_tensor(buf7, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf7
triton_poi_fused_3[grid(16, 4)](buf2, primals_7, buf10, 16, 4,
XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1)
del primals_7
buf11 = reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf9, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf10, (16, 4, 1), (4, 1, 0), 0), out=buf11)
buf12 = reinterpret_tensor(buf6, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf6
triton_poi_fused_clone_4[grid(16, 4)](buf11, buf12, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf13 = reinterpret_tensor(buf11, (16, 4), (4, 1), 0)
del buf11
extern_kernels.addmm(primals_10, reinterpret_tensor(buf12, (16, 4),
(4, 1), 0), reinterpret_tensor(primals_9, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf13)
del primals_10
buf14 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
buf15 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32)
triton_poi_fused_add_native_layer_norm_5[grid(16)](buf13, primals_3,
buf14, buf15, 16, XBLOCK=16, num_warps=1, num_stages=1)
buf16 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused_add_native_layer_norm_6[grid(64)](buf13, primals_3,
buf14, buf15, primals_11, primals_12, buf16, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf14
del buf15
del primals_12
return buf16, primals_3, primals_11, buf9, reinterpret_tensor(buf10, (
16, 1, 4), (4, 1, 1), 0), reinterpret_tensor(buf3, (16, 1, 4), (4,
1, 1), 0), reinterpret_tensor(buf4, (16, 4, 1), (4, 1, 4), 0
), reinterpret_tensor(buf12, (16, 4), (4, 1), 0), buf13, primals_9
class LxmertAttention(nn.Module):
def __init__(self, config, ctx_dim=None):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
'The hidden size (%d) is not a multiple of the number of attention heads (%d)'
% (config.hidden_size, config.num_attention_heads))
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.
num_attention_heads)
self.head_size = self.num_attention_heads * self.attention_head_size
if ctx_dim is None:
ctx_dim = config.hidden_size
self.query = nn.Linear(config.hidden_size, self.head_size)
self.key = nn.Linear(ctx_dim, self.head_size)
self.value = nn.Linear(ctx_dim, self.head_size)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.
attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(self, hidden_states, context, attention_mask=None,
output_attentions=False):
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(context)
mixed_value_layer = self.value(context)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1,
-2))
attention_scores = attention_scores / math.sqrt(self.
attention_head_size)
if attention_mask is not None:
attention_scores = attention_scores + attention_mask
attention_probs = nn.Softmax(dim=-1)(attention_scores)
attention_probs = self.dropout(attention_probs)
context_layer = torch.matmul(attention_probs, value_layer)
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
outputs = (context_layer, attention_probs) if output_attentions else (
context_layer,)
return outputs
class LxmertAttentionOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=1e-12)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, input_tensor):
hidden_states = self.dense(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.LayerNorm(hidden_states + input_tensor)
return hidden_states
class LxmertSelfAttentionLayerNew(nn.Module):
def __init__(self, config):
super().__init__()
self.self = LxmertAttention(config)
self.output = LxmertAttentionOutput(config)
def forward(self, input_0, input_1):
primals_1 = self.self.query.weight
primals_2 = self.self.query.bias
primals_4 = self.self.key.weight
primals_5 = self.self.key.bias
primals_6 = self.self.value.weight
primals_7 = self.self.value.bias
primals_9 = self.output.dense.weight
primals_10 = self.output.dense.bias
primals_11 = self.output.LayerNorm.weight
primals_12 = self.output.LayerNorm.bias
primals_3 = input_0
primals_8 = input_1
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11, primals_12])
return output[0]
|
rsgit95/med_kg_txt_multimodal
|
LxmertSelfAttentionLayer
| false
| 4,223
|
[
"Apache-2.0"
] | 0
|
80355b0cf58e0571531ad6f9728c533110ca996d
|
https://github.com/rsgit95/med_kg_txt_multimodal/tree/80355b0cf58e0571531ad6f9728c533110ca996d
|
SimplifiedScaledDotProductAttention
|
import torch
import numpy as np
from torch import nn
from torch.nn import init
class SimplifiedScaledDotProductAttention(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, h, dropout=0.1):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(SimplifiedScaledDotProductAttention, self).__init__()
self.d_model = d_model
self.d_k = d_model // h
self.d_v = d_model // h
self.h = h
self.fc_o = nn.Linear(h * self.d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nk, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, nq, nk).
:return:
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = queries.view(b_s, nq, self.h, self.d_k).permute(0, 2, 1, 3)
k = keys.view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = values.view(b_s, nk, self.h, self.d_v).permute(0, 2, 1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
att = torch.softmax(att, -1)
att = self.dropout(att)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
def get_inputs():
return [torch.rand([4, 4, 4, 1]), torch.rand([4, 4, 4, 1]), torch.rand(
[4, 4, 4, 1])]
def get_init_inputs():
return [[], {'d_model': 4, 'h': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from torch import nn
from torch.nn import init
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 16
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 4
y1 = yindex // 4
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask,
eviction_policy='evict_last')
tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 1), (16, 4, 1, 1))
assert_size_stride(primals_2, (4, 4, 4, 1), (16, 4, 1, 1))
assert_size_stride(primals_3, (4, 4, 4, 1), (16, 4, 1, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(16, 4)](primals_1, buf0, 16, 4,
XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((4, 4, 1, 4), (16, 4, 4, 1), torch.float32)
triton_poi_fused_clone_0[grid(16, 4)](primals_2, buf1, 16, 4,
XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((16, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf0, (16, 4, 1), (4, 1, 0),
0), reinterpret_tensor(buf1, (16, 1, 4), (4, 0, 1), 0), out=buf2)
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__softmax_1[grid(256)](buf2, buf3, 256, XBLOCK=128,
num_warps=4, num_stages=1)
buf4 = reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf2
triton_poi_fused__softmax_2[grid(256)](buf3, buf4, 256, XBLOCK=128,
num_warps=4, num_stages=1)
del buf3
buf5 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 1), 0)
del buf1
triton_poi_fused_clone_0[grid(16, 4)](primals_3, buf5, 16, 4,
XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1)
del primals_3
buf6 = reinterpret_tensor(buf0, (16, 4, 1), (4, 1, 1), 0)
del buf0
extern_kernels.bmm(reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf5, (16, 4, 1), (4, 1, 0), 0), out=buf6)
del buf4
buf7 = buf5
del buf5
triton_poi_fused_clone_0[grid(16, 4)](buf6, buf7, 16, 4, XBLOCK=4,
YBLOCK=16, num_warps=1, num_stages=1)
buf8 = reinterpret_tensor(buf6, (16, 4), (4, 1), 0)
del buf6
extern_kernels.addmm(primals_5, reinterpret_tensor(buf7, (16, 4), (
4, 1), 0), reinterpret_tensor(primals_4, (4, 4), (1, 4), 0),
alpha=1, beta=1, out=buf8)
del primals_4
del primals_5
return reinterpret_tensor(buf8, (4, 4, 4), (16, 4, 1), 0
), reinterpret_tensor(buf7, (16, 4), (4, 1), 0)
class SimplifiedScaledDotProductAttentionNew(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, h, dropout=0.1):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(SimplifiedScaledDotProductAttentionNew, self).__init__()
self.d_model = d_model
self.d_k = d_model // h
self.d_v = d_model // h
self.h = h
self.fc_o = nn.Linear(h * self.d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, input_0, input_1, input_2):
primals_4 = self.fc_o.weight
primals_5 = self.fc_o.bias
primals_1 = input_0
primals_2 = input_1
primals_3 = input_2
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
SimplifiedScaledDotProductAttention
| false
| 4,224
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
SpatialGroupEnhance
|
import torch
from torch import nn
from torch.nn import init
class SpatialGroupEnhance(nn.Module):
def __init__(self, groups):
super().__init__()
self.groups = groups
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.weight = nn.Parameter(torch.zeros(1, groups, 1, 1))
self.bias = nn.Parameter(torch.zeros(1, groups, 1, 1))
self.sig = nn.Sigmoid()
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, x):
b, c, h, w = x.shape
x = x.view(b * self.groups, -1, h, w)
xn = x * self.avg_pool(x)
xn = xn.sum(dim=1, keepdim=True)
t = xn.view(b * self.groups, -1)
t = t - t.mean(dim=1, keepdim=True)
std = t.std(dim=1, keepdim=True) + 1e-05
t = t / std
t = t.view(b, self.groups, h, w)
t = t * self.weight + self.bias
t = t.view(b * self.groups, 1, h, w)
x = x * self.sig(t)
x = x.view(b, c, h, w)
return x
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'groups': 1}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
from torch import nn
from torch.nn import init
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused_mean_0(in_out_ptr0, in_ptr0, xnumel, rnumel, XBLOCK:
tl.constexpr):
xnumel = 16
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 16 * x0), xmask, other=0.0)
tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
tmp3 = tl.where(xmask, tmp1, 0)
tmp4 = tl.sum(tmp3, 1)[:, None]
tmp5 = 16.0
tmp6 = tmp4 / tmp5
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp6, xmask)
@triton.jit
def triton_per_fused_add_mean_mul_sigmoid_std_sub_sum_1(in_out_ptr0,
in_out_ptr1, in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0, out_ptr1,
xnumel, rnumel, XBLOCK: tl.constexpr):
xnumel = 4
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 64 * x0), xmask, other=0.0)
tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (16 + r1 + 64 * x0), xmask, other=0.0)
tmp4 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (32 + r1 + 64 * x0), xmask, other=0.0)
tmp8 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (48 + r1 + 64 * x0), xmask, other=0.0)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last'
)
tmp44 = tl.load(in_ptr2 + 0)
tmp45 = tl.broadcast_to(tmp44, [XBLOCK, RBLOCK])
tmp47 = tl.load(in_ptr3 + 0)
tmp48 = tl.broadcast_to(tmp47, [XBLOCK, RBLOCK])
tmp2 = tmp0 * tmp1
tmp5 = tmp3 * tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 * tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 * tmp12
tmp14 = tmp10 + tmp13
tmp15 = tl.broadcast_to(tmp14, [XBLOCK, RBLOCK])
tmp17 = tl.where(xmask, tmp15, 0)
tmp18 = tl.sum(tmp17, 1)[:, None]
tmp19 = 16.0
tmp20 = tmp18 / tmp19
tmp21 = tmp14 - tmp20
tmp22 = tl.broadcast_to(tmp21, [XBLOCK, RBLOCK])
tl.where(xmask, tmp22, 0)
tmp25 = tl.broadcast_to(tmp22, [XBLOCK, RBLOCK])
tmp27 = tl.where(xmask, tmp25, 0)
tmp28 = tl.sum(tmp27, 1)[:, None]
tmp29 = tl.full([XBLOCK, 1], 16, tl.int32)
tmp30 = tmp29.to(tl.float32)
tmp31 = tmp28 / tmp30
tmp32 = tmp22 - tmp31
tmp33 = tmp32 * tmp32
tmp34 = tl.broadcast_to(tmp33, [XBLOCK, RBLOCK])
tmp36 = tl.where(xmask, tmp34, 0)
tmp37 = tl.sum(tmp36, 1)[:, None]
tmp38 = 15.0
tmp39 = tmp37 / tmp38
tmp40 = libdevice.sqrt(tmp39)
tmp41 = 1e-05
tmp42 = tmp40 + tmp41
tmp43 = tmp21 / tmp42
tmp46 = tmp43 * tmp45
tmp49 = tmp46 + tmp48
tmp50 = tl.sigmoid(tmp49)
tl.store(out_ptr0 + (r1 + 16 * x0), tmp14, xmask)
tl.debug_barrier()
tl.store(in_out_ptr0 + x0, tmp20, xmask)
tl.debug_barrier()
tl.store(in_out_ptr1 + x0, tmp42, xmask)
tl.store(out_ptr1 + (r1 + 16 * x0), tmp50, xmask)
@triton.jit
def triton_poi_fused_add_mul_sigmoid_2(in_ptr0, in_ptr1, out_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 16
x2 = xindex // 64
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr1 + (x0 + 16 * x2), xmask, eviction_policy=
'evict_last')
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x3, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (1, 1, 1, 1), (1, 1, 1, 1))
assert_size_stride(primals_3, (1, 1, 1, 1), (1, 1, 1, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 16, 16), torch.float32)
buf1 = reinterpret_tensor(buf0, (4, 4, 1, 1), (4, 1, 1, 1), 0)
del buf0
get_raw_stream(0)
triton_per_fused_mean_0[grid(16)](buf1, primals_1, 16, 16, XBLOCK=1,
num_warps=2, num_stages=1)
buf2 = empty_strided_cuda((4, 1, 4, 4), (16, 64, 4, 1), torch.float32)
buf3 = empty_strided_cuda((4, 1), (1, 4), torch.float32)
buf4 = reinterpret_tensor(buf3, (4, 1), (1, 1), 0)
del buf3
buf6 = empty_strided_cuda((4, 1), (1, 4), torch.float32)
buf8 = reinterpret_tensor(buf6, (4, 1), (1, 1), 0)
del buf6
buf9 = empty_strided_cuda((4, 1, 4, 4), (16, 64, 4, 1), torch.float32)
triton_per_fused_add_mean_mul_sigmoid_std_sub_sum_1[grid(4)](buf4,
buf8, primals_1, buf1, primals_2, primals_3, buf2, buf9, 4, 16,
XBLOCK=1, num_warps=2, num_stages=1)
del buf2
buf10 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_add_mul_sigmoid_2[grid(256)](primals_1, buf9,
buf10, 256, XBLOCK=256, num_warps=4, num_stages=1)
del buf9
return buf10, primals_1, primals_2, primals_3, buf1, buf4, buf8
class SpatialGroupEnhanceNew(nn.Module):
def __init__(self, groups):
super().__init__()
self.groups = groups
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.weight = nn.Parameter(torch.zeros(1, groups, 1, 1))
self.bias = nn.Parameter(torch.zeros(1, groups, 1, 1))
self.sig = nn.Sigmoid()
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, input_0):
primals_2 = self.weight
primals_3 = self.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
SpatialGroupEnhance
| false
| 4,225
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
ScaledDotProductAttention
|
import torch
import numpy as np
from torch import nn
from torch.nn import init
class ScaledDotProductAttention(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(ScaledDotProductAttention, self).__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nk, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, nq, nk).
:return:
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = self.fc_q(queries).view(b_s, nq, self.h, self.d_k).permute(0, 2,
1, 3)
k = self.fc_k(keys).view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = self.fc_v(values).view(b_s, nk, self.h, self.d_v).permute(0, 2,
1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
att = torch.softmax(att, -1)
att = self.dropout(att)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4, 4])
]
def get_init_inputs():
return [[], {'d_model': 4, 'd_k': 4, 'd_v': 4, 'h': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from torch import nn
from torch.nn import init
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tmp1 = tl.load(in_ptr1 + (x0 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + x4, tmp2, xmask)
@triton.jit
def triton_poi_fused_clone_1(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel,
YBLOCK: tl.constexpr, XBLOCK: tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y0 = yindex % 16
y1 = yindex // 16
y3 = yindex
tmp0 = tl.load(in_ptr0 + (y0 + 16 * x2 + 64 * y1), xmask & ymask,
eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + y0, ymask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(out_ptr0 + (x2 + 4 * y3), tmp2, xmask & ymask)
@triton.jit
def triton_poi_fused__softmax_sqrt_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp8 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp13 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp16 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = tl.full([1], 2.0, tl.float64)
tmp2 = tl.full([1], 0.0, tl.float64)
tmp3 = tmp1 >= tmp2
tmp4 = 1.0
tmp5 = -1.0
tmp6 = tl.where(tmp3, tmp4, tmp5)
tmp7 = tmp0 * tmp6
tmp9 = tmp8 * tmp6
tmp11 = tmp10 * tmp6
tmp12 = triton_helpers.maximum(tmp9, tmp11)
tmp14 = tmp13 * tmp6
tmp15 = triton_helpers.maximum(tmp12, tmp14)
tmp17 = tmp16 * tmp6
tmp18 = triton_helpers.maximum(tmp15, tmp17)
tmp19 = tmp7 - tmp18
tmp20 = tmp6.to(tl.float64)
tmp21 = tmp20 * tmp1
tmp22 = tmp21.to(tl.float32)
tmp23 = tmp19 / tmp22
tmp24 = tl_math.exp(tmp23)
tl.store(out_ptr0 + x2, tmp24, xmask)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_clone_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 4
x3 = xindex // 64
x4 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask)
tl.store(out_ptr0 + x4, tmp0, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9, primals_10, primals_11) = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_3, (16, 4), (4, 1))
assert_size_stride(primals_4, (16,), (1,))
assert_size_stride(primals_5, (16, 4), (4, 1))
assert_size_stride(primals_6, (16,), (1,))
assert_size_stride(primals_7, (16, 4), (4, 1))
assert_size_stride(primals_8, (16,), (1,))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_10, (4, 16), (16, 1))
assert_size_stride(primals_11, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_3, (4, 16), (1, 4), 0), out=buf0)
del primals_3
buf1 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_2, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_5, (4, 16), (1, 4), 0), out=buf1)
del primals_5
buf2 = empty_strided_cuda((16, 16), (16, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_9, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_7, (4, 16), (1, 4), 0), out=buf2)
del primals_7
buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_clone_0[grid(256)](buf0, primals_4, buf3, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_4
buf4 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
triton_poi_fused_clone_1[grid(64, 4)](buf1, primals_6, buf4, 64, 4,
XBLOCK=4, YBLOCK=32, num_warps=4, num_stages=1)
del primals_6
buf5 = reinterpret_tensor(buf1, (16, 4, 4), (16, 4, 1), 0)
del buf1
extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf4, (16, 4, 4), (16, 4, 1), 0), out=buf5)
buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused__softmax_sqrt_2[grid(256)](buf5, buf6, 256, XBLOCK
=128, num_warps=4, num_stages=1)
buf7 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf5
triton_poi_fused__softmax_3[grid(256)](buf6, buf7, 256, XBLOCK=256,
num_warps=4, num_stages=1)
buf8 = buf6
del buf6
triton_poi_fused_clone_0[grid(256)](buf2, primals_8, buf8, 256,
XBLOCK=256, num_warps=4, num_stages=1)
del primals_8
buf9 = reinterpret_tensor(buf2, (16, 4, 4), (16, 4, 1), 0)
del buf2
extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf8, (16, 4, 4), (16, 4, 1), 0), out=buf9)
buf10 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_clone_4[grid(256)](buf9, buf10, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del buf9
buf11 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_11, reinterpret_tensor(buf10, (16, 16),
(16, 1), 0), reinterpret_tensor(primals_10, (16, 4), (1, 16), 0
), alpha=1, beta=1, out=buf11)
del primals_11
return reinterpret_tensor(buf11, (4, 4, 4), (16, 4, 1), 0
), reinterpret_tensor(primals_1, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_2, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0
), buf7, reinterpret_tensor(buf10, (16, 16), (16, 1), 0
), primals_10, reinterpret_tensor(buf8, (16, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf3, (16, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf4, (16, 4, 4), (16, 1, 4), 0)
class ScaledDotProductAttentionNew(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, d_k, d_v, h, dropout=0.1):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(ScaledDotProductAttentionNew, self).__init__()
self.fc_q = nn.Linear(d_model, h * d_k)
self.fc_k = nn.Linear(d_model, h * d_k)
self.fc_v = nn.Linear(d_model, h * d_v)
self.fc_o = nn.Linear(h * d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.d_model = d_model
self.d_k = d_k
self.d_v = d_v
self.h = h
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, input_0, input_1, input_2):
primals_3 = self.fc_q.weight
primals_4 = self.fc_q.bias
primals_5 = self.fc_k.weight
primals_6 = self.fc_k.bias
primals_7 = self.fc_v.weight
primals_8 = self.fc_v.bias
primals_10 = self.fc_o.weight
primals_11 = self.fc_o.bias
primals_1 = input_0
primals_2 = input_1
primals_9 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9,
primals_10, primals_11])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
ScaledDotProductAttention
| false
| 4,226
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
AttentionHead
|
import torch
from torch import Tensor
import torch.nn as nn
import torch.nn.functional as F
from torch.functional import Tensor
def scaled_dot_product_attention(query: 'torch.Tensor', key: 'torch.Tensor',
value: 'torch.Tensor') ->Tensor:
temp = query.bmm(key.transpose(1, 2))
scale = query.size(-1) ** 0.5
softmax = F.softmax(temp / scale, dim=-1)
return softmax.bmm(value)
class AttentionHead(nn.Module):
def __init__(self, dim_in: 'int', dim_k: 'int', dim_v: 'int'):
super().__init__()
self.q = nn.Linear(dim_in, dim_k)
self.k = nn.Linear(dim_in, dim_k)
self.v = nn.Linear(dim_in, dim_v)
def forward(self, query: 'torch.Tensor', key: 'torch.Tensor', value:
'Tensor') ->Tensor:
return scaled_dot_product_attention(self.q(query), self.k(key),
self.v(value))
def get_inputs():
return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4, 4])
]
def get_init_inputs():
return [[], {'dim_in': 4, 'dim_k': 4, 'dim_v': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from torch import Tensor
import torch.nn as nn
import torch.nn.functional as F
from torch.functional import Tensor
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused__softmax_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp3 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last'
)
tmp1 = 1.0
tmp2 = tmp0 * tmp1
tmp4 = tmp3 * tmp1
tmp6 = tmp5 * tmp1
tmp7 = triton_helpers.maximum(tmp4, tmp6)
tmp9 = tmp8 * tmp1
tmp10 = triton_helpers.maximum(tmp7, tmp9)
tmp12 = tmp11 * tmp1
tmp13 = triton_helpers.maximum(tmp10, tmp12)
tmp14 = tmp2 - tmp13
tmp15 = 0.5
tmp16 = tmp14 * tmp15
tmp17 = tl_math.exp(tmp16)
tl.store(out_ptr0 + x2, tmp17, xmask)
@triton.jit
def triton_poi_fused__softmax_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7, primals_8, primals_9) = args
args.clear()
assert_size_stride(primals_1, (4, 4), (4, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 4), (4, 1))
assert_size_stride(primals_5, (4,), (1,))
assert_size_stride(primals_6, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_7, (4, 4), (4, 1))
assert_size_stride(primals_8, (4,), (1,))
assert_size_stride(primals_9, (4, 4, 4), (16, 4, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (16,
4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0
), alpha=1, beta=1, out=buf0)
del primals_1
del primals_2
buf1 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(primals_6, (16,
4), (4, 1), 0), reinterpret_tensor(primals_4, (4, 4), (1, 4), 0
), alpha=1, beta=1, out=buf1)
del primals_4
del primals_5
buf2 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_8, reinterpret_tensor(primals_9, (16,
4), (4, 1), 0), reinterpret_tensor(primals_7, (4, 4), (1, 4), 0
), alpha=1, beta=1, out=buf2)
del primals_7
del primals_8
buf3 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf0, (4, 4, 4), (16, 4, 1),
0), reinterpret_tensor(buf1, (4, 4, 4), (16, 1, 4), 0), out=buf3)
buf4 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused__softmax_0[grid(64)](buf3, buf4, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf5 = buf3
del buf3
triton_poi_fused__softmax_1[grid(64)](buf4, buf5, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf6 = buf4
del buf4
extern_kernels.bmm(buf5, reinterpret_tensor(buf2, (4, 4, 4), (16, 4,
1), 0), out=buf6)
return buf6, reinterpret_tensor(primals_3, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_6, (16, 4), (4, 1), 0
), reinterpret_tensor(primals_9, (16, 4), (4, 1), 0
), buf5, reinterpret_tensor(buf2, (4, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf0, (4, 4, 4), (16, 1, 4), 0
), reinterpret_tensor(buf1, (4, 4, 4), (16, 4, 1), 0)
def scaled_dot_product_attention(query: 'torch.Tensor', key: 'torch.Tensor',
value: 'torch.Tensor') ->Tensor:
temp = query.bmm(key.transpose(1, 2))
scale = query.size(-1) ** 0.5
softmax = F.softmax(temp / scale, dim=-1)
return softmax.bmm(value)
class AttentionHeadNew(nn.Module):
def __init__(self, dim_in: 'int', dim_k: 'int', dim_v: 'int'):
super().__init__()
self.q = nn.Linear(dim_in, dim_k)
self.k = nn.Linear(dim_in, dim_k)
self.v = nn.Linear(dim_in, dim_v)
def forward(self, input_0, input_1, input_2):
primals_1 = self.q.weight
primals_2 = self.q.bias
primals_4 = self.k.weight
primals_5 = self.k.bias
primals_7 = self.v.weight
primals_8 = self.v.bias
primals_3 = input_0
primals_6 = input_1
primals_9 = input_2
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7, primals_8, primals_9])
return output[0]
|
sabernn/vit-pytorch
|
AttentionHead
| false
| 4,227
|
[
"MIT"
] | 0
|
21a2671aa92adb941a56ae629f6089f550949fb2
|
https://github.com/sabernn/vit-pytorch/tree/21a2671aa92adb941a56ae629f6089f550949fb2
|
SE_Connect
|
import torch
import torch.nn.functional as F
import torch.nn
import torch.nn as nn
class SE_Connect(nn.Module):
def __init__(self, channels, s=4):
super().__init__()
assert channels % s == 0, '{} % {} != 0'.format(channesl, s)
self.linear1 = nn.Linear(channels, channels // s)
self.linear2 = nn.Linear(channels // s, channels)
def forward(self, x):
out = x.mean(dim=2)
out = F.relu(self.linear1(out))
out = torch.sigmoid(self.linear2(out))
out = x * out.unsqueeze(2)
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'channels': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
import torch.nn
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_mean_0(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 16 * x1), xmask)
tmp1 = tl.load(in_ptr0 + (4 + x0 + 16 * x1), xmask)
tmp3 = tl.load(in_ptr0 + (8 + x0 + 16 * x1), xmask)
tmp5 = tl.load(in_ptr0 + (12 + x0 + 16 * x1), xmask)
tmp2 = tmp0 + tmp1
tmp4 = tmp2 + tmp3
tmp6 = tmp4 + tmp5
tmp7 = 4.0
tmp8 = tmp6 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_relu_threshold_backward_1(in_out_ptr0, in_ptr0,
out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_out_ptr0 + x0, xmask)
tmp1 = tl.load(in_ptr0 + 0)
tmp2 = tl.broadcast_to(tmp1, [XBLOCK])
tmp3 = tmp0 + tmp2
tmp4 = tl.full([1], 0, tl.int32)
tmp5 = triton_helpers.maximum(tmp4, tmp3)
tmp6 = 0.0
tmp7 = tmp5 <= tmp6
tl.store(in_out_ptr0 + x0, tmp5, xmask)
tl.store(out_ptr0 + x0, tmp7, xmask)
@triton.jit
def triton_poi_fused_mul_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x0 = xindex % 4
x2 = xindex // 16
tmp0 = tl.load(in_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr1 + (x0 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp2 = tl.sigmoid(tmp1)
tmp3 = tmp0 * tmp2
tl.store(out_ptr0 + x3, tmp3, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (1, 4), (4, 1))
assert_size_stride(primals_3, (1,), (1,))
assert_size_stride(primals_4, (4, 1), (1, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
get_raw_stream(0)
triton_poi_fused_mean_0[grid(64)](primals_1, buf0, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf1 = empty_strided_cuda((16, 1), (1, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf0, (16, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 1), (1, 4), 0), out=buf1)
del primals_2
buf2 = reinterpret_tensor(buf1, (4, 4, 1), (4, 1, 1), 0)
del buf1
buf5 = empty_strided_cuda((4, 4, 1), (4, 1, 1), torch.bool)
triton_poi_fused_relu_threshold_backward_1[grid(16)](buf2,
primals_3, buf5, 16, XBLOCK=16, num_warps=1, num_stages=1)
del primals_3
buf3 = empty_strided_cuda((16, 4), (4, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf2, (16, 1), (
1, 0), 0), reinterpret_tensor(primals_4, (1, 4), (1, 1), 0),
alpha=1, beta=1, out=buf3)
del primals_5
buf4 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32)
triton_poi_fused_mul_2[grid(256)](primals_1, buf3, buf4, 256,
XBLOCK=256, num_warps=4, num_stages=1)
return buf4, primals_1, reinterpret_tensor(buf0, (16, 4), (4, 1), 0
), reinterpret_tensor(buf2, (16, 1), (1, 1), 0), buf3, primals_4, buf5
class SE_ConnectNew(nn.Module):
def __init__(self, channels, s=4):
super().__init__()
assert channels % s == 0, '{} % {} != 0'.format(channesl, s)
self.linear1 = nn.Linear(channels, channels // s)
self.linear2 = nn.Linear(channels // s, channels)
def forward(self, input_0):
primals_2 = self.linear1.weight
primals_3 = self.linear1.bias
primals_4 = self.linear2.weight
primals_5 = self.linear2.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
qlindazm/asv-subtools
|
SE_Connect
| false
| 4,228
|
[
"Apache-2.0"
] | 0
|
fe1d31db9f3268622016babe944201f6ff81ed56
|
https://github.com/qlindazm/asv-subtools/tree/fe1d31db9f3268622016babe944201f6ff81ed56
|
AttentiveStatsPool
|
import torch
import torch.nn
import torch.nn as nn
class AttentiveStatsPool(nn.Module):
def __init__(self, in_dim, bottleneck_dim):
super().__init__()
self.linear1 = nn.Conv1d(in_dim, bottleneck_dim, kernel_size=1)
self.linear2 = nn.Conv1d(bottleneck_dim, in_dim, kernel_size=1)
def forward(self, x):
alpha = torch.tanh(self.linear1(x))
alpha = torch.softmax(self.linear2(alpha), dim=2)
mean = torch.sum(alpha * x, dim=2)
residuals = torch.sum(alpha * x ** 2, dim=2) - mean ** 2
std = torch.sqrt(residuals.clamp(min=1e-09))
return torch.cat([mean, std], dim=1)
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'in_dim': 4, 'bottleneck_dim': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
import torch.nn
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_convolution_tanh_0(in_out_ptr0, in_ptr0, xnumel,
XBLOCK: tl.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 4 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = libdevice.tanh(tmp2)
tl.store(in_out_ptr0 + x3, tmp3, xmask)
@triton.jit
def triton_poi_fused_convolution_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 4 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, xmask)
@triton.jit
def triton_poi_fused__softmax_2(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = triton_helpers.maximum(tmp1, tmp2)
tmp5 = triton_helpers.maximum(tmp3, tmp4)
tmp7 = triton_helpers.maximum(tmp5, tmp6)
tmp8 = tmp0 - tmp7
tmp9 = tl_math.exp(tmp8)
tl.store(out_ptr0 + x2, tmp9, xmask)
@triton.jit
def triton_poi_fused__softmax_3(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr
):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last')
tmp2 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last')
tmp3 = tmp1 + tmp2
tmp5 = tmp3 + tmp4
tmp7 = tmp5 + tmp6
tmp8 = tmp0 / tmp7
tl.store(out_ptr0 + x2, tmp8, xmask)
@triton.jit
def triton_poi_fused_clamp_mul_pow_sqrt_sub_sum_4(in_ptr0, in_ptr1,
out_ptr0, out_ptr2, xnumel, XBLOCK: tl.constexpr):
xnumel = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
x1 = xindex // 4
tmp0 = tl.load(in_ptr0 + 4 * x2, xmask, eviction_policy='evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x2, xmask, eviction_policy='evict_last')
tmp3 = tl.load(in_ptr0 + (1 + 4 * x2), xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr1 + (1 + 4 * x2), xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr0 + (2 + 4 * x2), xmask, eviction_policy='evict_last')
tmp8 = tl.load(in_ptr1 + (2 + 4 * x2), xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr0 + (3 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp12 = tl.load(in_ptr1 + (3 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 * tmp1
tmp5 = tmp3 * tmp4
tmp6 = tmp2 + tmp5
tmp9 = tmp7 * tmp8
tmp10 = tmp6 + tmp9
tmp13 = tmp11 * tmp12
tmp14 = tmp10 + tmp13
tmp15 = tmp1 * tmp1
tmp16 = tmp0 * tmp15
tmp17 = tmp4 * tmp4
tmp18 = tmp3 * tmp17
tmp19 = tmp16 + tmp18
tmp20 = tmp8 * tmp8
tmp21 = tmp7 * tmp20
tmp22 = tmp19 + tmp21
tmp23 = tmp12 * tmp12
tmp24 = tmp11 * tmp23
tmp25 = tmp22 + tmp24
tmp26 = tmp14 * tmp14
tmp27 = tmp25 - tmp26
tmp28 = 1e-09
tmp29 = triton_helpers.maximum(tmp27, tmp28)
tmp30 = libdevice.sqrt(tmp29)
tl.store(out_ptr0 + (x0 + 8 * x1), tmp14, xmask)
tl.store(out_ptr2 + (x0 + 8 * x1), tmp30, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 1), (4, 1, 1))
assert_size_stride(primals_2, (4,), (1,))
assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_4, (4, 4, 1), (4, 1, 1))
assert_size_stride(primals_5, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1,),
padding=(0,), dilation=(1,), transposed=False, output_padding=(
0,), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 4), (16, 4, 1))
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_convolution_tanh_0[grid(64)](buf1, primals_2, 64,
XBLOCK=64, num_warps=1, num_stages=1)
del primals_2
buf2 = extern_kernels.convolution(buf1, primals_4, stride=(1,),
padding=(0,), dilation=(1,), transposed=False, output_padding=(
0,), groups=1, bias=None)
assert_size_stride(buf2, (4, 4, 4), (16, 4, 1))
buf3 = buf2
del buf2
triton_poi_fused_convolution_1[grid(64)](buf3, primals_5, 64,
XBLOCK=64, num_warps=1, num_stages=1)
del primals_5
buf4 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused__softmax_2[grid(64)](buf3, buf4, 64, XBLOCK=64,
num_warps=1, num_stages=1)
buf5 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32)
triton_poi_fused__softmax_3[grid(64)](buf4, buf5, 64, XBLOCK=64,
num_warps=1, num_stages=1)
del buf4
buf9 = empty_strided_cuda((4, 8), (8, 1), torch.float32)
buf6 = reinterpret_tensor(buf9, (4, 4), (8, 1), 0)
buf8 = reinterpret_tensor(buf9, (4, 4), (8, 1), 4)
triton_poi_fused_clamp_mul_pow_sqrt_sub_sum_4[grid(16)](buf5,
primals_3, buf6, buf8, 16, XBLOCK=16, num_warps=1, num_stages=1)
del buf5
return buf9, primals_1, primals_3, primals_4, buf1, buf3
class AttentiveStatsPoolNew(nn.Module):
def __init__(self, in_dim, bottleneck_dim):
super().__init__()
self.linear1 = nn.Conv1d(in_dim, bottleneck_dim, kernel_size=1)
self.linear2 = nn.Conv1d(bottleneck_dim, in_dim, kernel_size=1)
def forward(self, input_0):
primals_1 = self.linear1.weight
primals_2 = self.linear1.bias
primals_4 = self.linear2.weight
primals_5 = self.linear2.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
qlindazm/asv-subtools
|
AttentiveStatsPool
| false
| 4,229
|
[
"Apache-2.0"
] | 0
|
fe1d31db9f3268622016babe944201f6ff81ed56
|
https://github.com/qlindazm/asv-subtools/tree/fe1d31db9f3268622016babe944201f6ff81ed56
|
OutlookAttention
|
import math
import torch
from torch import nn
from torch.nn import functional as F
class OutlookAttention(nn.Module):
def __init__(self, dim, num_heads=1, kernel_size=3, padding=1, stride=1,
qkv_bias=False, attn_drop=0.1):
super().__init__()
self.dim = dim
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.kernel_size = kernel_size
self.padding = padding
self.stride = stride
self.scale = self.head_dim ** -0.5
self.v_pj = nn.Linear(dim, dim, bias=qkv_bias)
self.attn = nn.Linear(dim, kernel_size ** 4 * num_heads)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(attn_drop)
self.unflod = nn.Unfold(kernel_size, padding, stride)
self.pool = nn.AvgPool2d(kernel_size=stride, stride=stride,
ceil_mode=True)
def forward(self, x):
B, H, W, C = x.shape
v = self.v_pj(x).permute(0, 3, 1, 2)
h, w = math.ceil(H / self.stride), math.ceil(W / self.stride)
v = self.unflod(v).reshape(B, self.num_heads, self.head_dim, self.
kernel_size * self.kernel_size, h * w).permute(0, 1, 4, 3, 2)
attn = self.pool(x.permute(0, 3, 1, 2)).permute(0, 2, 3, 1)
attn = self.attn(attn).reshape(B, h * w, self.num_heads, self.
kernel_size * self.kernel_size, self.kernel_size * self.kernel_size
).permute(0, 2, 1, 3, 4)
attn = self.scale * attn
attn = attn.softmax(-1)
attn = self.attn_drop(attn)
out = (attn @ v).permute(0, 1, 4, 3, 2).reshape(B, C * self.
kernel_size * self.kernel_size, h * w)
out = F.fold(out, output_size=(H, W), kernel_size=self.kernel_size,
padding=self.padding, stride=self.stride)
out = self.proj(out.permute(0, 2, 3, 1))
out = self.proj_drop(out)
return out
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'dim': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
from torch import nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_im2col_0(out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 12
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 4
x1 = xindex // 4
x2 = xindex
tmp0 = x0 + x1
tl.store(out_ptr0 + x2, tmp0, xmask)
@triton.jit
def triton_poi_fused_avg_pool2d_1(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.
constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = tl.load(in_ptr0 + x0, xmask)
tmp1 = 1.0
tmp2 = tmp0 * tmp1
tl.store(out_ptr0 + x0, tmp2, xmask)
@triton.jit
def triton_per_fused__softmax_2(in_ptr0, in_ptr1, out_ptr2, xnumel, rnumel,
XBLOCK: tl.constexpr):
xnumel = 576
rnumel = 9
RBLOCK: tl.constexpr = 16
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
rmask = rindex < rnumel
r2 = rindex
x5 = xindex
x0 = xindex % 9
x4 = xindex // 144
x6 = xindex % 144
tmp0 = tl.load(in_ptr0 + (r2 + 9 * x5), rmask & xmask, other=0.0)
tmp1 = tl.load(in_ptr1 + (r2 + 9 * x0), rmask & xmask, eviction_policy=
'evict_last', other=0.0)
tmp2 = tmp0 + tmp1
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tmp5 = tl.broadcast_to(tmp4, [XBLOCK, RBLOCK])
tmp7 = tl.where(rmask & xmask, tmp5, float('-inf'))
tmp8 = triton_helpers.max2(tmp7, 1)[:, None]
tmp9 = tmp4 - tmp8
tmp10 = 0.5
tmp11 = tmp9 * tmp10
tmp12 = tl_math.exp(tmp11)
tmp13 = tl.broadcast_to(tmp12, [XBLOCK, RBLOCK])
tmp15 = tl.where(rmask & xmask, tmp13, 0)
tmp16 = tl.sum(tmp15, 1)[:, None]
tmp17 = tmp12 / tmp16
tl.store(out_ptr2 + (r2 + 9 * x6 + 1312 * x4), tmp17, rmask & xmask)
@triton.jit
def triton_poi_fused_clone_3(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 2304
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x1 = xindex // 4 % 9
x2 = xindex // 36 % 16
x0 = xindex % 4
x3 = xindex // 576
x4 = xindex
tmp0 = tl.load(in_ptr0 + (4 * (x1 // 3) + x2 // 4), xmask,
eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + (4 * (x1 % 3) + x2 % 4), xmask,
eviction_policy='evict_last')
tmp1 = tl.full([XBLOCK], 6, tl.int32)
tmp2 = tmp0 + tmp1
tmp3 = tmp0 < 0
tmp4 = tl.where(tmp3, tmp2, tmp0)
tl.device_assert((0 <= tmp4) & (tmp4 < 6) | ~xmask,
'index out of bounds: 0 <= tmp4 < 6')
tmp7 = tmp6 + tmp1
tmp8 = tmp6 < 0
tmp9 = tl.where(tmp8, tmp7, tmp6)
tl.device_assert((0 <= tmp9) & (tmp9 < 6) | ~xmask,
'index out of bounds: 0 <= tmp9 < 6')
tmp11 = -1 + tmp4
tmp12 = tl.full([1], 0, tl.int64)
tmp13 = tmp11 >= tmp12
tmp14 = tl.full([1], 4, tl.int64)
tmp15 = tmp11 < tmp14
tmp16 = -1 + tmp9
tmp17 = tmp16 >= tmp12
tmp18 = tmp16 < tmp14
tmp19 = tmp13 & tmp15
tmp20 = tmp19 & tmp17
tmp21 = tmp20 & tmp18
tmp22 = tl.load(in_ptr1 + (-20 + x0 + 4 * tmp9 + 16 * tmp4 + 64 * x3),
tmp21 & xmask, other=0.0)
tl.store(out_ptr0 + x4, tmp22, xmask)
@triton.jit
def triton_poi_fused_bmm_4(in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 5184
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 81
x1 = xindex // 81
x2 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 81 * (x1 % 16) + 1312 * (x1 // 16)), xmask)
tl.store(out_ptr0 + x2, tmp0, xmask)
@triton.jit
def triton_poi_fused_col2im_5(out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 576
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex
tmp0 = 0.0
tl.store(out_ptr0 + x0, tmp0, xmask)
@triton.jit
def triton_poi_fused_col2im_6(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK:
tl.constexpr):
xnumel = 2304
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x7 = xindex // 48 % 12
x9 = xindex // 4 % 12
x0 = xindex % 4
x1 = xindex // 4 % 4
x2 = xindex // 16 % 3
x3 = xindex // 48 % 4
x4 = xindex // 192 % 3
x5 = xindex // 576
tmp0 = tl.load(in_ptr0 + x7, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr0 + x9, xmask, eviction_policy='evict_last')
tmp11 = tl.load(in_ptr1 + (x0 + 4 * x2 + 12 * x4 + 36 * x1 + 144 * x3 +
576 * x5 + (x2 + 3 * x4) // 9), xmask)
tmp1 = tl.full([XBLOCK], 6, tl.int32)
tmp2 = tmp0 + tmp1
tmp3 = tmp0 < 0
tmp4 = tl.where(tmp3, tmp2, tmp0)
tl.device_assert((0 <= tmp4) & (tmp4 < 6) | ~xmask,
'index out of bounds: 0 <= tmp4 < 6')
tmp7 = tmp6 + tmp1
tmp8 = tmp6 < 0
tmp9 = tl.where(tmp8, tmp7, tmp6)
tl.device_assert((0 <= tmp9) & (tmp9 < 6) | ~xmask,
'index out of bounds: 0 <= tmp9 < 6')
tl.atomic_add(out_ptr0 + (tmp9 + 6 * tmp4 + 36 * x0 + 144 * x5), tmp11,
xmask, sem='relaxed')
@triton.jit
def triton_poi_fused_clone_7(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
ynumel = 64
xnumel = 4
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
ymask = yindex < ynumel
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
y1 = yindex // 4 % 4
y0 = yindex % 4
x3 = xindex
y2 = yindex // 16
y5 = yindex
tmp0 = 1 + y1
tmp1 = tl.full([1, 1], 0, tl.int64)
tmp2 = tmp0 >= tmp1
tmp3 = tl.full([1, 1], 6, tl.int64)
tmp4 = tmp0 < tmp3
tmp5 = 1 + y0
tmp6 = tmp5 >= tmp1
tmp7 = tmp5 < tmp3
tmp8 = tmp2 & tmp4
tmp9 = tmp8 & tmp6
tmp10 = tmp9 & tmp7
tmp11 = tl.load(in_ptr0 + (7 + y0 + 6 * y1 + 36 * x3 + 144 * y2), tmp10 &
xmask & ymask, eviction_policy='evict_last', other=0.0)
tl.store(out_ptr0 + (x3 + 4 * y5), tmp11, xmask & ymask)
@triton.jit
def triton_poi_fused_add_8(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 256
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x2 = xindex
x0 = xindex % 4
tmp0 = tl.load(in_out_ptr0 + x2, xmask)
tmp1 = tl.load(in_ptr0 + x0, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5, primals_6 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_2, (4, 4), (4, 1))
assert_size_stride(primals_3, (81, 4), (4, 1))
assert_size_stride(primals_4, (81,), (1,))
assert_size_stride(primals_5, (4, 4), (4, 1))
assert_size_stride(primals_6, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_1, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), out=buf0)
del primals_2
buf1 = empty_strided_cuda((3, 4), (4, 1), torch.int64)
get_raw_stream(0)
triton_poi_fused_im2col_0[grid(12)](buf1, 12, XBLOCK=16, num_warps=
1, num_stages=1)
buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 1, 16, 4), torch.float32)
triton_poi_fused_avg_pool2d_1[grid(256)](primals_1, buf2, 256,
XBLOCK=256, num_warps=4, num_stages=1)
buf3 = empty_strided_cuda((64, 81), (81, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf2, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_3, (4, 81), (1, 4), 0), out=buf3)
del primals_3
buf6 = empty_strided_cuda((4, 1, 16, 9, 9), (1312, 1312, 81, 9, 1),
torch.float32)
triton_per_fused__softmax_2[grid(576)](buf3, primals_4, buf6, 576,
9, XBLOCK=8, num_warps=2, num_stages=1)
del primals_4
buf7 = empty_strided_cuda((4, 1, 16, 9, 4), (576, 1, 36, 4, 1),
torch.float32)
triton_poi_fused_clone_3[grid(2304)](buf1, buf0, buf7, 2304, XBLOCK
=256, num_warps=4, num_stages=1)
buf8 = reinterpret_tensor(buf3, (64, 9, 9), (81, 9, 1), 0)
del buf3
triton_poi_fused_bmm_4[grid(5184)](buf6, buf8, 5184, XBLOCK=128,
num_warps=4, num_stages=1)
buf9 = empty_strided_cuda((64, 9, 4), (36, 4, 1), torch.float32)
extern_kernels.bmm(buf8, reinterpret_tensor(buf7, (64, 9, 4), (36,
4, 1), 0), out=buf9)
del buf8
buf10 = empty_strided_cuda((4, 4, 6, 6), (144, 36, 6, 1), torch.float32
)
triton_poi_fused_col2im_5[grid(576)](buf10, 576, XBLOCK=256,
num_warps=4, num_stages=1)
buf11 = empty_strided_cuda((4, 4, 6, 6), (144, 36, 6, 1), torch.float32
)
triton_poi_fused_col2im_5[grid(576)](buf11, 576, XBLOCK=256,
num_warps=4, num_stages=1)
triton_poi_fused_col2im_6[grid(2304)](buf1, buf9, buf11, 2304,
XBLOCK=128, num_warps=4, num_stages=1)
del buf9
buf13 = reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf0
triton_poi_fused_clone_7[grid(64, 4)](buf11, buf13, 64, 4, XBLOCK=4,
YBLOCK=32, num_warps=4, num_stages=1)
del buf11
buf14 = empty_strided_cuda((64, 4), (4, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf13, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_5, (4, 4), (1, 4), 0), out=buf14)
buf15 = reinterpret_tensor(buf14, (4, 4, 4, 4), (64, 16, 4, 1), 0)
del buf14
triton_poi_fused_add_8[grid(256)](buf15, primals_6, 256, XBLOCK=256,
num_warps=4, num_stages=1)
del primals_6
return buf15, reinterpret_tensor(primals_1, (64, 4), (4, 1), 0
), buf1, reinterpret_tensor(buf2, (64, 4), (4, 1), 0
), buf6, buf10, reinterpret_tensor(buf13, (64, 4), (4, 1), 0
), primals_5, reinterpret_tensor(buf7, (64, 4, 9), (36, 1, 4), 0)
class OutlookAttentionNew(nn.Module):
def __init__(self, dim, num_heads=1, kernel_size=3, padding=1, stride=1,
qkv_bias=False, attn_drop=0.1):
super().__init__()
self.dim = dim
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.kernel_size = kernel_size
self.padding = padding
self.stride = stride
self.scale = self.head_dim ** -0.5
self.v_pj = nn.Linear(dim, dim, bias=qkv_bias)
self.attn = nn.Linear(dim, kernel_size ** 4 * num_heads)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(attn_drop)
self.unflod = nn.Unfold(kernel_size, padding, stride)
self.pool = nn.AvgPool2d(kernel_size=stride, stride=stride,
ceil_mode=True)
def forward(self, input_0):
primals_2 = self.v_pj.weight
primals_3 = self.attn.weight
primals_4 = self.attn.bias
primals_5 = self.proj.weight
primals_6 = self.proj.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
OutlookAttention
| false
| 4,230
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
Critic
|
import torch
import torch.nn as nn
class Critic(nn.Module):
def __init__(self, obs_dim: 'int'):
super().__init__()
self.fc1 = nn.Linear(obs_dim, 64)
self.fc2 = nn.Linear(64, 64)
self.fc3 = nn.Linear(64, 1)
def forward(self, x):
x = torch.tanh(self.fc1(x))
x = torch.tanh(self.fc2(x))
x = self.fc3(x)
return x.squeeze()
def get_inputs():
return [torch.rand([4, 4, 4, 4])]
def get_init_inputs():
return [[], {'obs_dim': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime.triton_helpers import libdevice
import torch.nn as nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_tanh_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr
):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 64
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tmp3 = libdevice.tanh(tmp2)
tl.store(in_out_ptr0 + x2, tmp3, None)
def call(args):
(primals_1, primals_2, primals_3, primals_4, primals_5, primals_6,
primals_7) = args
args.clear()
assert_size_stride(primals_1, (64, 4), (4, 1))
assert_size_stride(primals_2, (64,), (1,))
assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1))
assert_size_stride(primals_4, (64, 64), (64, 1))
assert_size_stride(primals_5, (64,), (1,))
assert_size_stride(primals_6, (1, 64), (64, 1))
assert_size_stride(primals_7, (1,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((64, 64), (64, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0),
reinterpret_tensor(primals_1, (4, 64), (1, 4), 0), out=buf0)
del primals_1
buf1 = reinterpret_tensor(buf0, (4, 4, 4, 64), (1024, 256, 64, 1), 0)
del buf0
get_raw_stream(0)
triton_poi_fused_tanh_0[grid(4096)](buf1, primals_2, 4096, XBLOCK=
256, num_warps=4, num_stages=1)
del primals_2
buf2 = empty_strided_cuda((64, 64), (64, 1), torch.float32)
extern_kernels.mm(reinterpret_tensor(buf1, (64, 64), (64, 1), 0),
reinterpret_tensor(primals_4, (64, 64), (1, 64), 0), out=buf2)
buf3 = reinterpret_tensor(buf2, (4, 4, 4, 64), (1024, 256, 64, 1), 0)
del buf2
triton_poi_fused_tanh_0[grid(4096)](buf3, primals_5, 4096, XBLOCK=
256, num_warps=4, num_stages=1)
del primals_5
buf5 = empty_strided_cuda((64, 1), (1, 1), torch.float32)
extern_kernels.addmm(primals_7, reinterpret_tensor(buf3, (64, 64),
(64, 1), 0), reinterpret_tensor(primals_6, (64, 1), (1, 64), 0),
alpha=1, beta=1, out=buf5)
del primals_7
return reinterpret_tensor(buf5, (4, 4, 4), (16, 4, 1), 0
), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0
), buf1, buf3, primals_6, primals_4
class CriticNew(nn.Module):
def __init__(self, obs_dim: 'int'):
super().__init__()
self.fc1 = nn.Linear(obs_dim, 64)
self.fc2 = nn.Linear(64, 64)
self.fc3 = nn.Linear(64, 1)
def forward(self, input_0):
primals_1 = self.fc1.weight
primals_2 = self.fc1.bias
primals_4 = self.fc2.weight
primals_5 = self.fc2.bias
primals_6 = self.fc3.weight
primals_7 = self.fc3.bias
primals_3 = input_0
output = call([primals_1, primals_2, primals_3, primals_4,
primals_5, primals_6, primals_7])
return output[0]
|
raznem/rlex
|
Critic
| false
| 4,231
|
[
"MIT"
] | 0
|
d24b964d80067becc81d86f6ce87e5be413b7049
|
https://github.com/raznem/rlex/tree/d24b964d80067becc81d86f6ce87e5be413b7049
|
TdnnAffine
|
import torch
import torch.nn.functional as F
import torch.nn
def to_device(device_object, tensor):
"""
Select device for non-parameters tensor w.r.t model or tensor which has been specified a device.
"""
if isinstance(device_object, torch.nn.Module):
next(device_object.parameters()).device
elif isinstance(device_object, torch.Tensor):
pass
return tensor
class TdnnAffine(torch.nn.Module):
""" An implemented tdnn affine component by conv1d
y = splice(w * x, context) + b
@input_dim: number of dims of frame <=> inputs channels of conv
@output_dim: number of layer nodes <=> outputs channels of conv
@context: a list of context
e.g. [-2,0,2]
If context is [0], then the TdnnAffine is equal to linear layer.
"""
def __init__(self, input_dim, output_dim, context=[0], bias=True, pad=
True, stride=1, groups=1, norm_w=False, norm_f=False):
super(TdnnAffine, self).__init__()
assert input_dim % groups == 0
for index in range(0, len(context) - 1):
if context[index] >= context[index + 1]:
raise ValueError(
'Context tuple {} is invalid, such as the order.'.
format(context))
self.input_dim = input_dim
self.output_dim = output_dim
self.context = context
self.bool_bias = bias
self.pad = pad
self.groups = groups
self.norm_w = norm_w
self.norm_f = norm_f
self.stride = stride
self.left_context = context[0] if context[0] < 0 else 0
self.right_context = context[-1] if context[-1] > 0 else 0
self.tot_context = self.right_context - self.left_context + 1
if self.tot_context > 1 and self.norm_f:
self.norm_f = False
None
kernel_size = self.tot_context,
self.weight = torch.nn.Parameter(torch.randn(output_dim, input_dim //
groups, *kernel_size))
if self.bool_bias:
self.bias = torch.nn.Parameter(torch.randn(output_dim))
else:
self.register_parameter('bias', None)
self.init_weight()
if len(context) != self.tot_context:
self.mask = torch.tensor([[[(1 if index in context else 0) for
index in range(self.left_context, self.right_context + 1)]]])
else:
self.mask = None
self.selected_device = False
def init_weight(self):
torch.nn.init.normal_(self.weight, 0.0, 0.01)
if self.bias is not None:
torch.nn.init.constant_(self.bias, 0.0)
def forward(self, inputs):
"""
@inputs: a 3-dimensional tensor (a batch), including [samples-index, frames-dim-index, frames-index]
"""
assert len(inputs.shape) == 3
assert inputs.shape[1] == self.input_dim
if self.pad:
inputs = F.pad(inputs, (-self.left_context, self.right_context),
mode='constant', value=0)
assert inputs.shape[2] >= self.tot_context
if not self.selected_device and self.mask is not None:
self.mask = to_device(self, self.mask)
self.selected_device = True
filters = (self.weight * self.mask if self.mask is not None else
self.weight)
if self.norm_w:
filters = F.normalize(filters, dim=1)
if self.norm_f:
inputs = F.normalize(inputs, dim=1)
outputs = F.conv1d(inputs, filters, self.bias, stride=self.stride,
padding=0, dilation=1, groups=self.groups)
return outputs
def extra_repr(self):
return (
'{input_dim}, {output_dim}, context={context}, bias={bool_bias}, stride={stride}, pad={pad}, groups={groups}, norm_w={norm_w}, norm_f={norm_f}'
.format(**self.__dict__))
@classmethod
def thop_count(self, m, x, y):
x = x[0]
kernel_ops = torch.zeros(m.weight.size()[2:]).numel()
bias_ops = 1 if m.bias is not None else 0
total_ops = y.nelement() * (m.input_dim * kernel_ops + bias_ops)
m.total_ops += torch.DoubleTensor([int(total_ops)])
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'input_dim': 4, 'output_dim': 4}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
import torch.nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
@triton.jit
def triton_poi_fused_convolution_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xnumel = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x3 = xindex
x1 = xindex // 4 % 4
tmp0 = tl.load(in_out_ptr0 + x3, xmask)
tmp1 = tl.load(in_ptr0 + x1, xmask, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x3, tmp2, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 4, 1), (4, 1, 1))
assert_size_stride(primals_3, (4,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = extern_kernels.convolution(primals_1, primals_2, stride=(1,),
padding=(0,), dilation=(1,), transposed=False, output_padding=(
0,), groups=1, bias=None)
assert_size_stride(buf0, (4, 4, 4), (16, 4, 1))
buf1 = buf0
del buf0
get_raw_stream(0)
triton_poi_fused_convolution_0[grid(64)](buf1, primals_3, 64,
XBLOCK=64, num_warps=1, num_stages=1)
del primals_3
return buf1, primals_1, primals_2
def to_device(device_object, tensor):
"""
Select device for non-parameters tensor w.r.t model or tensor which has been specified a device.
"""
if isinstance(device_object, torch.nn.Module):
next(device_object.parameters()).device
elif isinstance(device_object, torch.Tensor):
pass
return tensor
class TdnnAffineNew(torch.nn.Module):
""" An implemented tdnn affine component by conv1d
y = splice(w * x, context) + b
@input_dim: number of dims of frame <=> inputs channels of conv
@output_dim: number of layer nodes <=> outputs channels of conv
@context: a list of context
e.g. [-2,0,2]
If context is [0], then the TdnnAffine is equal to linear layer.
"""
def __init__(self, input_dim, output_dim, context=[0], bias=True, pad=
True, stride=1, groups=1, norm_w=False, norm_f=False):
super(TdnnAffineNew, self).__init__()
assert input_dim % groups == 0
for index in range(0, len(context) - 1):
if context[index] >= context[index + 1]:
raise ValueError(
'Context tuple {} is invalid, such as the order.'.
format(context))
self.input_dim = input_dim
self.output_dim = output_dim
self.context = context
self.bool_bias = bias
self.pad = pad
self.groups = groups
self.norm_w = norm_w
self.norm_f = norm_f
self.stride = stride
self.left_context = context[0] if context[0] < 0 else 0
self.right_context = context[-1] if context[-1] > 0 else 0
self.tot_context = self.right_context - self.left_context + 1
if self.tot_context > 1 and self.norm_f:
self.norm_f = False
None
kernel_size = self.tot_context,
self.weight = torch.nn.Parameter(torch.randn(output_dim, input_dim //
groups, *kernel_size))
if self.bool_bias:
self.bias = torch.nn.Parameter(torch.randn(output_dim))
else:
self.register_parameter('bias', None)
self.init_weight()
if len(context) != self.tot_context:
self.mask = torch.tensor([[[(1 if index in context else 0) for
index in range(self.left_context, self.right_context + 1)]]])
else:
self.mask = None
self.selected_device = False
def init_weight(self):
torch.nn.init.normal_(self.weight, 0.0, 0.01)
if self.bias is not None:
torch.nn.init.constant_(self.bias, 0.0)
def extra_repr(self):
return (
'{input_dim}, {output_dim}, context={context}, bias={bool_bias}, stride={stride}, pad={pad}, groups={groups}, norm_w={norm_w}, norm_f={norm_f}'
.format(**self.__dict__))
@classmethod
def thop_count(self, m, x, y):
x = x[0]
kernel_ops = torch.zeros(m.weight.size()[2:]).numel()
bias_ops = 1 if m.bias is not None else 0
total_ops = y.nelement() * (m.input_dim * kernel_ops + bias_ops)
m.total_ops += torch.DoubleTensor([int(total_ops)])
def forward(self, input_0):
primals_2 = self.weight
primals_3 = self.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
|
qlindazm/asv-subtools
|
TdnnAffine
| false
| 4,232
|
[
"Apache-2.0"
] | 0
|
fe1d31db9f3268622016babe944201f6ff81ed56
|
https://github.com/qlindazm/asv-subtools/tree/fe1d31db9f3268622016babe944201f6ff81ed56
|
ChannelAttentionModule
|
import torch
import numpy as np
from torch import nn
from torch.nn import init
class SimplifiedScaledDotProductAttention(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, h, dropout=0.1):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(SimplifiedScaledDotProductAttention, self).__init__()
self.d_model = d_model
self.d_k = d_model // h
self.d_v = d_model // h
self.h = h
self.fc_o = nn.Linear(h * self.d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nk, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, nq, nk).
:return:
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = queries.view(b_s, nq, self.h, self.d_k).permute(0, 2, 1, 3)
k = keys.view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = values.view(b_s, nk, self.h, self.d_v).permute(0, 2, 1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
att = torch.softmax(att, -1)
att = self.dropout(att)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
class ChannelAttentionModule(nn.Module):
def __init__(self, d_model=512, kernel_size=3, H=7, W=7):
super().__init__()
self.cnn = nn.Conv2d(d_model, d_model, kernel_size=kernel_size,
padding=(kernel_size - 1) // 2)
self.pa = SimplifiedScaledDotProductAttention(H * W, h=1)
def forward(self, x):
bs, c, _h, _w = x.shape
y = self.cnn(x)
y = y.view(bs, c, -1)
y = self.pa(y, y, y)
return y
def get_inputs():
return [torch.rand([4, 512, 1, 49])]
def get_init_inputs():
return [[], {}]
|
import torch
from torch._inductor.select_algorithm import extern_kernels
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import numpy as np
from torch import nn
from torch.nn import init
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_poi_fused_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 49
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 512
y1 = yindex // 512
tmp0 = tl.load(in_ptr0 + (x2 + 49 * y3), xmask, eviction_policy=
'evict_last')
tl.store(out_ptr0 + (y0 + 512 * x2 + 25088 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.
constexpr, XBLOCK: tl.constexpr):
xnumel = 9
yoffset = (tl.program_id(1) + tl.program_id(2) * tl.num_programs(1)
) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
tl.full([XBLOCK, YBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
x2 = xindex
y3 = yindex
y0 = yindex % 512
y1 = yindex // 512
tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last'
)
tl.store(out_ptr0 + (y0 + 512 * x2 + 4608 * y1), tmp0, xmask)
@triton.jit
def triton_poi_fused_convolution_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl
.constexpr):
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
tl.full([XBLOCK], True, tl.int1)
x2 = xindex
x0 = xindex % 512
tmp0 = tl.load(in_out_ptr0 + x2, None)
tmp1 = tl.load(in_ptr0 + x0, None, eviction_policy='evict_last')
tmp2 = tmp0 + tmp1
tl.store(in_out_ptr0 + x2, tmp2, None)
@triton.jit
def triton_per_fused__softmax_sqrt_3(in_ptr0, out_ptr2, xnumel, rnumel):
XBLOCK: tl.constexpr = 1
RBLOCK: tl.constexpr = 512
xoffset = tl.program_id(0) * XBLOCK
xindex = tl.full([1], xoffset, tl.int32)
tl.full([RBLOCK], True, tl.int1)
rindex = tl.arange(0, RBLOCK)[:]
tl.full([RBLOCK], True, tl.int1)
r1 = rindex
x0 = xindex
tmp0 = tl.load(in_ptr0 + (r1 + 512 * x0), None)
tmp1 = tl.full([1], 7.0, tl.float64)
tmp2 = tl.full([1], 0.0, tl.float64)
tmp3 = tmp1 >= tmp2
tmp4 = 1.0
tmp5 = -1.0
tmp6 = tl.where(tmp3, tmp4, tmp5)
tmp7 = tmp0 * tmp6
tmp8 = tl.broadcast_to(tmp7, [RBLOCK])
tmp10 = triton_helpers.promote_to_tensor(triton_helpers.max2(tmp8, 0))
tmp11 = tmp7 - tmp10
tmp12 = tmp6.to(tl.float64)
tmp13 = tmp12 * tmp1
tmp14 = tmp13.to(tl.float32)
tmp15 = tmp11 / tmp14
tmp16 = tl_math.exp(tmp15)
tmp17 = tl.broadcast_to(tmp16, [RBLOCK])
tmp19 = triton_helpers.promote_to_tensor(tl.sum(tmp17, 0))
tmp20 = tmp16 / tmp19
tl.store(out_ptr2 + (r1 + 512 * x0), tmp20, None)
def call(args):
primals_1, primals_2, primals_3, primals_4, primals_5 = args
args.clear()
assert_size_stride(primals_1, (4, 512, 1, 49), (25088, 49, 49, 1))
assert_size_stride(primals_2, (512, 512, 3, 3), (4608, 9, 3, 1))
assert_size_stride(primals_3, (512,), (1,))
assert_size_stride(primals_4, (49, 49), (49, 1))
assert_size_stride(primals_5, (49,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 512, 1, 49), (25088, 1, 25088, 512),
torch.float32)
get_raw_stream(0)
triton_poi_fused_0[grid(2048, 49)](primals_1, buf0, 2048, 49,
XBLOCK=32, YBLOCK=32, num_warps=4, num_stages=1)
del primals_1
buf1 = empty_strided_cuda((512, 512, 3, 3), (4608, 1, 1536, 512),
torch.float32)
triton_poi_fused_1[grid(262144, 9)](primals_2, buf1, 262144, 9,
XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1)
del primals_2
buf2 = extern_kernels.convolution(buf0, buf1, stride=(1, 1),
padding=(1, 1), dilation=(1, 1), transposed=False,
output_padding=(0, 0), groups=1, bias=None)
assert_size_stride(buf2, (4, 512, 1, 49), (25088, 1, 25088, 512))
buf3 = buf2
del buf2
triton_poi_fused_convolution_2[grid(100352)](buf3, primals_3,
100352, XBLOCK=1024, num_warps=4, num_stages=1)
del primals_3
buf4 = empty_strided_cuda((4, 512, 512), (262144, 512, 1), torch.
float32)
extern_kernels.bmm(reinterpret_tensor(buf3, (4, 512, 49), (25088, 1,
512), 0), reinterpret_tensor(buf3, (4, 49, 512), (25088, 512, 1
), 0), out=buf4)
buf7 = empty_strided_cuda((4, 1, 512, 512), (262144, 1, 512, 1),
torch.float32)
triton_per_fused__softmax_sqrt_3[grid(2048)](buf4, buf7, 2048, 512,
num_warps=4, num_stages=1)
del buf4
buf8 = empty_strided_cuda((4, 512, 49), (25088, 49, 1), torch.float32)
extern_kernels.bmm(reinterpret_tensor(buf7, (4, 512, 512), (262144,
512, 1), 0), reinterpret_tensor(buf3, (4, 512, 49), (25088, 1,
512), 0), out=buf8)
buf9 = empty_strided_cuda((2048, 49), (49, 1), torch.float32)
extern_kernels.addmm(primals_5, reinterpret_tensor(buf8, (2048, 49),
(49, 1), 0), reinterpret_tensor(primals_4, (49, 49), (1, 49), 0
), alpha=1, beta=1, out=buf9)
del primals_5
return reinterpret_tensor(buf9, (4, 512, 49), (25088, 49, 1), 0
), buf0, buf1, buf3, buf7, reinterpret_tensor(buf8, (2048, 49), (49,
1), 0), primals_4
class SimplifiedScaledDotProductAttention(nn.Module):
"""
Scaled dot-product attention
"""
def __init__(self, d_model, h, dropout=0.1):
"""
:param d_model: Output dimensionality of the model
:param d_k: Dimensionality of queries and keys
:param d_v: Dimensionality of values
:param h: Number of heads
"""
super(SimplifiedScaledDotProductAttention, self).__init__()
self.d_model = d_model
self.d_k = d_model // h
self.d_v = d_model // h
self.h = h
self.fc_o = nn.Linear(h * self.d_v, d_model)
self.dropout = nn.Dropout(dropout)
self.init_weights()
def init_weights(self):
for m in self.modules():
if isinstance(m, nn.Conv2d):
init.kaiming_normal_(m.weight, mode='fan_out')
if m.bias is not None:
init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
init.constant_(m.weight, 1)
init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal_(m.weight, std=0.001)
if m.bias is not None:
init.constant_(m.bias, 0)
def forward(self, queries, keys, values, attention_mask=None,
attention_weights=None):
"""
Computes
:param queries: Queries (b_s, nq, d_model)
:param keys: Keys (b_s, nk, d_model)
:param values: Values (b_s, nk, d_model)
:param attention_mask: Mask over attention values (b_s, h, nq, nk). True indicates masking.
:param attention_weights: Multiplicative weights for attention values (b_s, h, nq, nk).
:return:
"""
b_s, nq = queries.shape[:2]
nk = keys.shape[1]
q = queries.view(b_s, nq, self.h, self.d_k).permute(0, 2, 1, 3)
k = keys.view(b_s, nk, self.h, self.d_k).permute(0, 2, 3, 1)
v = values.view(b_s, nk, self.h, self.d_v).permute(0, 2, 1, 3)
att = torch.matmul(q, k) / np.sqrt(self.d_k)
if attention_weights is not None:
att = att * attention_weights
if attention_mask is not None:
att = att.masked_fill(attention_mask, -np.inf)
att = torch.softmax(att, -1)
att = self.dropout(att)
out = torch.matmul(att, v).permute(0, 2, 1, 3).contiguous().view(b_s,
nq, self.h * self.d_v)
out = self.fc_o(out)
return out
class ChannelAttentionModuleNew(nn.Module):
def __init__(self, d_model=512, kernel_size=3, H=7, W=7):
super().__init__()
self.cnn = nn.Conv2d(d_model, d_model, kernel_size=kernel_size,
padding=(kernel_size - 1) // 2)
self.pa = SimplifiedScaledDotProductAttention(H * W, h=1)
def forward(self, input_0):
primals_2 = self.cnn.weight
primals_3 = self.cnn.bias
primals_4 = self.pa.fc_o.weight
primals_5 = self.pa.fc_o.bias
primals_1 = input_0
output = call([primals_1, primals_2, primals_3, primals_4, primals_5])
return output[0]
|
rushirajsherlocked/External-Attention-pytorch
|
ChannelAttentionModule
| false
| 4,233
|
[
"MIT"
] | 0
|
7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
https://github.com/rushirajsherlocked/External-Attention-pytorch/tree/7d6814b2d90909adf81c62f3f8a89e30a59d6481
|
LDEPooling
|
import torch
import torch.nn
class LDEPooling(torch.nn.Module):
"""A novel learnable dictionary encoding layer.
Reference: Weicheng Cai, etc., "A NOVEL LEARNABLE DICTIONARY ENCODING LAYER FOR END-TO-END
LANGUAGE IDENTIFICATION", icassp, 2018
"""
def __init__(self, input_dim, c_num=64, eps=1e-10):
super(LDEPooling, self).__init__()
self.input_dim = input_dim
self.output_dim = input_dim * c_num
self.eps = eps
self.mu = torch.nn.Parameter(torch.randn(input_dim, c_num))
self.s = torch.nn.Parameter(torch.ones(c_num))
self.softmax_for_w = torch.nn.Softmax(dim=3)
def forward(self, inputs):
"""
@inputs: a 3-dimensional tensor (a batch), including [samples-index, frames-dim-index, frames-index]
"""
assert len(inputs.shape) == 3
assert inputs.shape[1] == self.input_dim
r = inputs.transpose(1, 2).unsqueeze(3) - self.mu
w = self.softmax_for_w(-(self.s ** 2 + self.eps) * torch.sum(r ** 2,
dim=2, keepdim=True))
e = torch.mean(w * r, dim=1)
return e.reshape(-1, self.output_dim, 1)
def get_output_dim(self):
return self.output_dim
def get_inputs():
return [torch.rand([4, 4, 4])]
def get_init_inputs():
return [[], {'input_dim': 4}]
|
import torch
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._inductor.runtime import triton_helpers
from torch._inductor.runtime.triton_helpers import math as tl_math
import torch.nn
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
@triton.jit
def triton_per_fused__softmax_add_mul_neg_pow_sub_sum_0(in_ptr0, in_ptr1,
in_ptr2, out_ptr0, out_ptr1, out_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr
):
xnumel = 16
RBLOCK: tl.constexpr = 64
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
xmask = xindex < xnumel
rindex = tl.arange(0, RBLOCK)[None, :]
tl.full([XBLOCK, RBLOCK], True, tl.int1)
r2 = rindex
x0 = xindex % 4
x1 = xindex // 4
x3 = xindex
tmp0 = tl.load(in_ptr0 + r2, None, eviction_policy='evict_last')
tmp5 = tl.load(in_ptr1 + (x0 + 16 * x1), xmask, eviction_policy=
'evict_last')
tmp6 = tl.load(in_ptr2 + r2, None, eviction_policy='evict_last')
tmp9 = tl.load(in_ptr1 + (4 + x0 + 16 * x1), xmask, eviction_policy=
'evict_last')
tmp10 = tl.load(in_ptr2 + (64 + r2), None, eviction_policy='evict_last')
tmp14 = tl.load(in_ptr1 + (8 + x0 + 16 * x1), xmask, eviction_policy=
'evict_last')
tmp15 = tl.load(in_ptr2 + (128 + r2), None, eviction_policy='evict_last')
tmp19 = tl.load(in_ptr1 + (12 + x0 + 16 * x1), xmask, eviction_policy=
'evict_last')
tmp20 = tl.load(in_ptr2 + (192 + r2), None, eviction_policy='evict_last')
tmp1 = tmp0 * tmp0
tmp2 = 1e-10
tmp3 = tmp1 + tmp2
tmp4 = -tmp3
tmp7 = tmp5 - tmp6
tmp8 = tmp7 * tmp7
tmp11 = tmp9 - tmp10
tmp12 = tmp11 * tmp11
tmp13 = tmp8 + tmp12
tmp16 = tmp14 - tmp15
tmp17 = tmp16 * tmp16
tmp18 = tmp13 + tmp17
tmp21 = tmp19 - tmp20
tmp22 = tmp21 * tmp21
tmp23 = tmp18 + tmp22
tmp24 = tmp4 * tmp23
tmp25 = tl.broadcast_to(tmp24, [XBLOCK, RBLOCK])
tmp27 = tl.where(xmask, tmp25, float('-inf'))
tmp28 = triton_helpers.max2(tmp27, 1)[:, None]
tmp29 = tmp24 - tmp28
tmp30 = tl_math.exp(tmp29)
tmp31 = tl.broadcast_to(tmp30, [XBLOCK, RBLOCK])
tmp33 = tl.where(xmask, tmp31, 0)
tmp34 = tl.sum(tmp33, 1)[:, None]
tl.store(out_ptr0 + (r2 + 64 * x3), tmp24, xmask)
tl.store(out_ptr1 + x3, tmp28, xmask)
tl.store(out_ptr2 + x3, tmp34, xmask)
@triton.jit
def triton_poi_fused__softmax_mean_mul_sub_1(in_ptr0, in_ptr1, in_ptr2,
in_ptr3, in_ptr4, out_ptr0, xnumel, XBLOCK: tl.constexpr):
xnumel = 1024
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[:]
xmask = xindex < xnumel
x0 = xindex % 64
x2 = xindex // 256
x4 = xindex // 64
x3 = xindex % 256
x5 = xindex
tmp0 = tl.load(in_ptr0 + (x0 + 256 * x2), xmask, eviction_policy=
'evict_last')
tmp1 = tl.load(in_ptr1 + 4 * x2, xmask, eviction_policy='evict_last')
tmp4 = tl.load(in_ptr2 + 4 * x2, xmask, eviction_policy='evict_last')
tmp6 = tl.load(in_ptr3 + 4 * x4, xmask, eviction_policy='evict_last')
tmp7 = tl.load(in_ptr4 + x3, xmask, eviction_policy='evict_last')
tmp10 = tl.load(in_ptr0 + (64 + x0 + 256 * x2), xmask, eviction_policy=
'evict_last')
tmp11 = tl.load(in_ptr1 + (1 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp14 = tl.load(in_ptr2 + (1 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp16 = tl.load(in_ptr3 + (1 + 4 * x4), xmask, eviction_policy='evict_last'
)
tmp20 = tl.load(in_ptr0 + (128 + x0 + 256 * x2), xmask, eviction_policy
='evict_last')
tmp21 = tl.load(in_ptr1 + (2 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp24 = tl.load(in_ptr2 + (2 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp26 = tl.load(in_ptr3 + (2 + 4 * x4), xmask, eviction_policy='evict_last'
)
tmp30 = tl.load(in_ptr0 + (192 + x0 + 256 * x2), xmask, eviction_policy
='evict_last')
tmp31 = tl.load(in_ptr1 + (3 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp34 = tl.load(in_ptr2 + (3 + 4 * x2), xmask, eviction_policy='evict_last'
)
tmp36 = tl.load(in_ptr3 + (3 + 4 * x4), xmask, eviction_policy='evict_last'
)
tmp2 = tmp0 - tmp1
tmp3 = tl_math.exp(tmp2)
tmp5 = tmp3 / tmp4
tmp8 = tmp6 - tmp7
tmp9 = tmp5 * tmp8
tmp12 = tmp10 - tmp11
tmp13 = tl_math.exp(tmp12)
tmp15 = tmp13 / tmp14
tmp17 = tmp16 - tmp7
tmp18 = tmp15 * tmp17
tmp19 = tmp9 + tmp18
tmp22 = tmp20 - tmp21
tmp23 = tl_math.exp(tmp22)
tmp25 = tmp23 / tmp24
tmp27 = tmp26 - tmp7
tmp28 = tmp25 * tmp27
tmp29 = tmp19 + tmp28
tmp32 = tmp30 - tmp31
tmp33 = tl_math.exp(tmp32)
tmp35 = tmp33 / tmp34
tmp37 = tmp36 - tmp7
tmp38 = tmp35 * tmp37
tmp39 = tmp29 + tmp38
tmp40 = 4.0
tmp41 = tmp39 / tmp40
tl.store(out_ptr0 + x5, tmp41, xmask)
def call(args):
primals_1, primals_2, primals_3 = args
args.clear()
assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1))
assert_size_stride(primals_2, (4, 64), (64, 1))
assert_size_stride(primals_3, (64,), (1,))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((4, 4, 1, 64), (256, 64, 1024, 1), torch.
float32)
buf1 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
buf2 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32)
get_raw_stream(0)
triton_per_fused__softmax_add_mul_neg_pow_sub_sum_0[grid(16)](primals_3
, primals_1, primals_2, buf0, buf1, buf2, 16, 64, XBLOCK=8,
num_warps=4, num_stages=1)
buf3 = empty_strided_cuda((4, 4, 64), (256, 64, 1), torch.float32)
triton_poi_fused__softmax_mean_mul_sub_1[grid(1024)](buf0, buf1,
buf2, primals_1, primals_2, buf3, 1024, XBLOCK=128, num_warps=4,
num_stages=1)
del buf0
return reinterpret_tensor(buf3, (4, 256, 1), (256, 1, 1), 0
), primals_1, primals_2, primals_3, buf1, buf2
class LDEPoolingNew(torch.nn.Module):
"""A novel learnable dictionary encoding layer.
Reference: Weicheng Cai, etc., "A NOVEL LEARNABLE DICTIONARY ENCODING LAYER FOR END-TO-END
LANGUAGE IDENTIFICATION", icassp, 2018
"""
def __init__(self, input_dim, c_num=64, eps=1e-10):
super(LDEPoolingNew, self).__init__()
self.input_dim = input_dim
self.output_dim = input_dim * c_num
self.eps = eps
self.mu = torch.nn.Parameter(torch.randn(input_dim, c_num))
self.s = torch.nn.Parameter(torch.ones(c_num))
self.softmax_for_w = torch.nn.Softmax(dim=3)
def get_output_dim(self):
return self.output_dim
def forward(self, input_0):
primals_2 = self.mu
primals_3 = self.s
primals_1 = input_0
output = call([primals_1, primals_2, primals_3])
return output[0]
|
qlindazm/asv-subtools
|
LDEPooling
| false
| 4,234
|
[
"Apache-2.0"
] | 0
|
fe1d31db9f3268622016babe944201f6ff81ed56
|
https://github.com/qlindazm/asv-subtools/tree/fe1d31db9f3268622016babe944201f6ff81ed56
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.