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
Attention
import torch from torch import nn import torch.nn.utils class Attention(nn.Module): def __init__(self, hidden_dim): super(Attention, self).__init__() self.hidden_dim = hidden_dim self.ff = nn.Linear(in_features=hidden_dim, out_features=1) self.softmax = nn.Softmax(dim=-1) def forward(self, contexts, context_masks=None): """ :param contexts: (batch_size, seq_len, n_hid) :param context_masks: (batch_size, seq_len) :return: (batch_size, n_hid), (batch_size, seq_len) """ out = self.ff(contexts) out = out.view(contexts.size(0), contexts.size(1)) if context_masks is not None: masked_out = out.masked_fill(context_masks, float('-inf')) else: masked_out = out attn_weights = self.softmax(masked_out) out = attn_weights.unsqueeze(1).bmm(contexts) out = out.squeeze(1) return out, attn_weights def get_inputs(): return [torch.rand([4, 4, 4])] def get_init_inputs(): return [[], {'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 from torch._inductor.runtime.triton_helpers import math as tl_math from torch import nn import torch.nn.utils 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 = 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_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 = 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 = 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), (16, 4, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf1 = empty_strided_cuda((16, 1), (1, 1), torch.float32) extern_kernels.addmm(primals_2, reinterpret_tensor(primals_3, (16, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 1), (1, 4), 0 ), alpha=1, beta=1, out=buf1) del primals_1 del primals_2 buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32) get_raw_stream(0) triton_poi_fused__softmax_0[grid(16)](buf1, buf2, 16, XBLOCK=16, num_warps=1, num_stages=1) buf3 = reinterpret_tensor(buf1, (4, 4), (4, 1), 0) del buf1 triton_poi_fused__softmax_1[grid(16)](buf2, buf3, 16, XBLOCK=16, num_warps=1, num_stages=1) buf4 = reinterpret_tensor(buf2, (4, 1, 4), (4, 4, 1), 0) del buf2 extern_kernels.bmm(reinterpret_tensor(buf3, (4, 1, 4), (4, 4, 1), 0 ), primals_3, out=buf4) return reinterpret_tensor(buf4, (4, 4), (4, 1), 0), buf3, primals_3, buf3 class AttentionNew(nn.Module): def __init__(self, hidden_dim): super(AttentionNew, self).__init__() self.hidden_dim = hidden_dim self.ff = nn.Linear(in_features=hidden_dim, out_features=1) self.softmax = nn.Softmax(dim=-1) def forward(self, input_0): primals_1 = self.ff.weight primals_2 = self.ff.bias primals_3 = input_0 output = call([primals_1, primals_2, primals_3]) return output[0], output[1]
bstee615/ReVeal
Attention
false
14,982
[ "MIT" ]
63
fc22d0d54a3a23d4e0bc45a249b7eea22749685e
https://github.com/bstee615/ReVeal/tree/fc22d0d54a3a23d4e0bc45a249b7eea22749685e
TriangularSylvester
import torch from torch import nn class TriangularSylvester(nn.Module): """ Sylvester normalizing flow with Q=P or Q=I. """ def __init__(self, z_size): super(TriangularSylvester, self).__init__() self.z_size = z_size self.h = nn.Tanh() def der_h(self, x): return self.der_tanh(x) def der_tanh(self, x): return 1 - self.h(x) ** 2 def forward(self, zk, r1, r2, b, permute_z=None, sum_ldj=True): """ All flow parameters are amortized. conditions on diagonals of R1 and R2 need to be satisfied outside of this function. Computes the following transformation: z' = z + QR1 h( R2Q^T z + b) or actually z'^T = z^T + h(z^T Q R2^T + b^T)R1^T Q^T with Q = P a permutation matrix (equal to identity matrix if permute_z=None) :param zk: shape: (batch_size, z_size) :param r1: shape: (batch_size, num_ortho_vecs, num_ortho_vecs). :param r2: shape: (batch_size, num_ortho_vecs, num_ortho_vecs). :param b: shape: (batch_size, 1, self.z_size) :return: z, log_det_j """ zk = zk.unsqueeze(1) diag_r1 = torch.diagonal(r1, 0, -1, -2) diag_r2 = torch.diagonal(r2, 0, -1, -2) if permute_z is not None: z_per = zk[:, :, permute_z] else: z_per = zk r2qzb = z_per @ r2.transpose(2, 1) + b z = self.h(r2qzb) @ r1.transpose(2, 1) if permute_z is not None: z = z[:, :, permute_z] z += zk z = z.squeeze(1) diag_j = diag_r1 * diag_r2 diag_j = self.der_h(r2qzb).squeeze(1) * diag_j diag_j += 1.0 log_diag_j = (diag_j.abs() + 1e-08).log() if sum_ldj: log_det_j = log_diag_j.sum(-1) else: log_det_j = log_diag_j return z, log_det_j def get_inputs(): return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand( [4, 4, 4, 4]), torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'z_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, 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_clone_0(in_ptr0, 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 x3 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tl.store(out_ptr0 + x3, tmp0, xmask) @triton.jit def triton_poi_fused_clone_1(in_ptr0, 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 % 4 x2 = xindex // 16 % 4 x3 = xindex // 64 % 4 x5 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 4 * x2 + 16 * x1 + 64 * x3), xmask, eviction_policy='evict_last') tl.store(out_ptr0 + x5, tmp0, xmask) @triton.jit def triton_poi_fused_add_tanh_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 x2 = xindex x0 = xindex % 256 tmp0 = tl.load(in_ptr0 + x2, xmask) tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = libdevice.tanh(tmp2) tl.store(out_ptr0 + x2, tmp3, xmask) @triton.jit def triton_poi_fused_add_3(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 x5 = xindex x0 = xindex % 64 x2 = xindex // 256 tmp0 = tl.load(in_out_ptr0 + x5, xmask) tmp1 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp2 = tmp0 + tmp1 tl.store(in_out_ptr0 + x5, tmp2, xmask) @triton.jit def triton_poi_fused_abs_add_log_mul_pow_rsub_sum_tanh_4(in_ptr0, in_ptr1, in_ptr2, in_ptr3, 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 x4 = xindex % 64 x0 = xindex % 16 tmp0 = tl.load(in_ptr0 + 4 * x3, xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr1 + 4 * x4, xmask, eviction_policy='evict_last') tmp7 = tl.load(in_ptr2 + 16 * x0, xmask, eviction_policy='evict_last') tmp8 = tl.load(in_ptr3 + 16 * x0, xmask, eviction_policy='evict_last') tmp16 = tl.load(in_ptr0 + (1 + 4 * x3), xmask, eviction_policy='evict_last' ) tmp17 = tl.load(in_ptr1 + (1 + 4 * x4), xmask, eviction_policy='evict_last' ) tmp22 = tl.load(in_ptr2 + (5 + 16 * x0), xmask, eviction_policy= 'evict_last') tmp23 = tl.load(in_ptr3 + (5 + 16 * x0), xmask, eviction_policy= 'evict_last') tmp31 = tl.load(in_ptr0 + (2 + 4 * x3), xmask, eviction_policy='evict_last' ) tmp32 = tl.load(in_ptr1 + (2 + 4 * x4), xmask, eviction_policy='evict_last' ) tmp37 = tl.load(in_ptr2 + (10 + 16 * x0), xmask, eviction_policy= 'evict_last') tmp38 = tl.load(in_ptr3 + (10 + 16 * x0), xmask, eviction_policy= 'evict_last') tmp46 = tl.load(in_ptr0 + (3 + 4 * x3), xmask, eviction_policy='evict_last' ) tmp47 = tl.load(in_ptr1 + (3 + 4 * x4), xmask, eviction_policy='evict_last' ) tmp52 = tl.load(in_ptr2 + (15 + 16 * x0), xmask, eviction_policy= 'evict_last') tmp53 = tl.load(in_ptr3 + (15 + 16 * x0), xmask, eviction_policy= 'evict_last') tmp2 = tmp0 + tmp1 tmp3 = libdevice.tanh(tmp2) tmp4 = tmp3 * tmp3 tmp5 = 1.0 tmp6 = tmp5 - tmp4 tmp9 = tmp7 * tmp8 tmp10 = tmp6 * tmp9 tmp11 = tmp10 + tmp5 tmp12 = tl_math.abs(tmp11) tmp13 = 1e-08 tmp14 = tmp12 + tmp13 tmp15 = tl_math.log(tmp14) tmp18 = tmp16 + tmp17 tmp19 = libdevice.tanh(tmp18) tmp20 = tmp19 * tmp19 tmp21 = tmp5 - tmp20 tmp24 = tmp22 * tmp23 tmp25 = tmp21 * tmp24 tmp26 = tmp25 + tmp5 tmp27 = tl_math.abs(tmp26) tmp28 = tmp27 + tmp13 tmp29 = tl_math.log(tmp28) tmp30 = tmp15 + tmp29 tmp33 = tmp31 + tmp32 tmp34 = libdevice.tanh(tmp33) tmp35 = tmp34 * tmp34 tmp36 = tmp5 - tmp35 tmp39 = tmp37 * tmp38 tmp40 = tmp36 * tmp39 tmp41 = tmp40 + tmp5 tmp42 = tl_math.abs(tmp41) tmp43 = tmp42 + tmp13 tmp44 = tl_math.log(tmp43) tmp45 = tmp30 + tmp44 tmp48 = tmp46 + tmp47 tmp49 = libdevice.tanh(tmp48) tmp50 = tmp49 * tmp49 tmp51 = tmp5 - tmp50 tmp54 = tmp52 * tmp53 tmp55 = tmp51 * tmp54 tmp56 = tmp55 + tmp5 tmp57 = tl_math.abs(tmp56) tmp58 = tmp57 + tmp13 tmp59 = tl_math.log(tmp58) tmp60 = tmp45 + tmp59 tl.store(out_ptr0 + x3, tmp60, xmask) def call(args): arg0_1, arg1_1, arg2_1, arg3_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)) assert_size_stride(arg3_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, 4), (256, 64, 16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_clone_0[grid(1024)](arg0_1, buf0, 1024, XBLOCK=256, num_warps=4, num_stages=1) buf1 = empty_strided_cuda((4, 4, 4, 4, 4), (256, 64, 16, 4, 1), torch.float32) triton_poi_fused_clone_1[grid(1024)](arg2_1, buf1, 1024, XBLOCK=128, num_warps=4, num_stages=1) buf2 = empty_strided_cuda((64, 4, 4), (16, 4, 1), torch.float32) extern_kernels.bmm(reinterpret_tensor(buf0, (64, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf1, (64, 4, 4), (16, 4, 1), 0), out=buf2) buf3 = buf1 del buf1 triton_poi_fused_add_tanh_2[grid(1024)](buf2, arg3_1, buf3, 1024, XBLOCK=256, num_warps=4, num_stages=1) buf4 = buf0 del buf0 triton_poi_fused_clone_1[grid(1024)](arg1_1, buf4, 1024, XBLOCK=128, num_warps=4, num_stages=1) buf5 = empty_strided_cuda((64, 4, 4), (16, 4, 1), torch.float32) extern_kernels.bmm(reinterpret_tensor(buf3, (64, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf4, (64, 4, 4), (16, 4, 1), 0), out=buf5) del buf3 del buf4 buf6 = reinterpret_tensor(buf5, (4, 4, 4, 4, 4), (256, 64, 16, 4, 1), 0 ) del buf5 buf7 = buf6 del buf6 triton_poi_fused_add_3[grid(1024)](buf7, arg0_1, 1024, XBLOCK=128, num_warps=4, num_stages=1) del arg0_1 buf8 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused_abs_add_log_mul_pow_rsub_sum_tanh_4[grid(256)](buf2, arg3_1, arg1_1, arg2_1, buf8, 256, XBLOCK=128, num_warps=4, num_stages=1) del arg1_1 del arg2_1 del arg3_1 del buf2 return buf7, buf8 class TriangularSylvesterNew(nn.Module): """ Sylvester normalizing flow with Q=P or Q=I. """ def __init__(self, z_size): super(TriangularSylvesterNew, self).__init__() self.z_size = z_size self.h = nn.Tanh() def der_h(self, x): return self.der_tanh(x) def der_tanh(self, x): return 1 - self.h(x) ** 2 def forward(self, input_0, input_1, input_2, input_3): arg0_1 = input_0 arg1_1 = input_1 arg2_1 = input_2 arg3_1 = input_3 output = call([arg0_1, arg1_1, arg2_1, arg3_1]) return output[0], output[1]
boldsort/NeuralDX7
TriangularSylvester
false
14,983
[ "MIT" ]
119
327844cea18a6dfe35e0dc8f5de0832343487366
https://github.com/boldsort/NeuralDX7/tree/327844cea18a6dfe35e0dc8f5de0832343487366
MSE
import torch import torch.nn as nn import torch.utils.checkpoint class MSE(nn.Module): def __init__(self): super(MSE, self).__init__() def forward(self, pred, real): diffs = torch.add(real, -pred) n = torch.numel(diffs.data) mse = torch.sum(diffs.pow(2)) / n return mse 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 import torch.nn as nn import torch.utils.checkpoint 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_neg_pow_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) tmp1 = tl.load(in_ptr1 + r0, None) tmp2 = -tmp1 tmp3 = tmp0 + tmp2 tmp4 = tmp3 * tmp3 tmp5 = tl.broadcast_to(tmp4, [RBLOCK]) tmp7 = triton_helpers.promote_to_tensor(tl.sum(tmp5, 0)) tmp8 = 0.00390625 tmp9 = tmp7 * tmp8 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp9, 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_div_neg_pow_sum_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 MSENew(nn.Module): def __init__(self): super(MSENew, 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]
byamao1/MMSA
MSE
false
14,984
[ "MIT" ]
198
1a894d042144c9ac75b3465d38871ce8c2987251
https://github.com/byamao1/MMSA/tree/1a894d042144c9ac75b3465d38871ce8c2987251
LRN
import torch import torch.nn as nn class LRN(nn.Module): def __init__(self, local_size=1, alpha=1.0, beta=0.75, ACROSS_CHANNELS=True ): super(LRN, self).__init__() self.ACROSS_CHANNELS = ACROSS_CHANNELS if ACROSS_CHANNELS: self.average = nn.AvgPool3d(kernel_size=(local_size, 1, 1), stride=1, padding=(int((local_size - 1.0) / 2), 0, 0)) else: self.average = nn.AvgPool2d(kernel_size=local_size, stride=1, padding=int((local_size - 1.0) / 2)) self.alpha = alpha self.beta = beta def forward(self, x): if self.ACROSS_CHANNELS: div = x.pow(2).unsqueeze(1) div = self.average(div).squeeze(1) div = div.mul(self.alpha).add(1.0).pow(self.beta) else: div = x.pow(2) div = self.average(div) div = div.mul(self.alpha).add(1.0).pow(self.beta) x = x.div(div) 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 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 @triton.jit def triton_poi_fused_add_div_mul_pow_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 * tmp0 tmp2 = 1.0 tmp3 = tmp1 * tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 + tmp2 tmp6 = 0.75 tmp7 = libdevice.pow(tmp5, tmp6) tmp8 = tmp0 / 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_div_mul_pow_0[grid(256)](arg0_1, buf0, 256, XBLOCK=128, num_warps=4, num_stages=1) del arg0_1 return buf0, class LRNNew(nn.Module): def __init__(self, local_size=1, alpha=1.0, beta=0.75, ACROSS_CHANNELS=True ): super(LRNNew, self).__init__() self.ACROSS_CHANNELS = ACROSS_CHANNELS if ACROSS_CHANNELS: self.average = nn.AvgPool3d(kernel_size=(local_size, 1, 1), stride=1, padding=(int((local_size - 1.0) / 2), 0, 0)) else: self.average = nn.AvgPool2d(kernel_size=local_size, stride=1, padding=int((local_size - 1.0) / 2)) self.alpha = alpha self.beta = beta def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0]
bruinxiong/BNM
LRN
false
14,985
[ "MIT" ]
252
71d4b8c9beca00e77fcbc62a12b69bb093736a82
https://github.com/bruinxiong/BNM/tree/71d4b8c9beca00e77fcbc62a12b69bb093736a82
SIMSE
import torch import torch.nn as nn import torch.utils.checkpoint class SIMSE(nn.Module): def __init__(self): super(SIMSE, self).__init__() def forward(self, pred, real): diffs = torch.add(real, -pred) n = torch.numel(diffs.data) simse = torch.sum(diffs).pow(2) / n ** 2 return simse 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 import torch.nn as nn import torch.utils.checkpoint 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_neg_pow_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) tmp1 = tl.load(in_ptr1 + r0, None) tmp2 = -tmp1 tmp3 = tmp0 + tmp2 tmp4 = tl.broadcast_to(tmp3, [RBLOCK]) tmp6 = triton_helpers.promote_to_tensor(tl.sum(tmp4, 0)) tmp7 = tmp6 * tmp6 tmp8 = 1.52587890625e-05 tmp9 = tmp7 * tmp8 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp9, 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_div_neg_pow_sum_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 SIMSENew(nn.Module): def __init__(self): super(SIMSENew, 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]
byamao1/MMSA
SIMSE
false
14,986
[ "MIT" ]
198
1a894d042144c9ac75b3465d38871ce8c2987251
https://github.com/byamao1/MMSA/tree/1a894d042144c9ac75b3465d38871ce8c2987251
ActorCriticMLP
import torch from torch import Tensor from torch import nn from typing import Tuple from torch.nn import functional as F class ActorCriticMLP(nn.Module): """MLP network with heads for actor and critic.""" def __init__(self, input_shape: 'Tuple[int]', n_actions: 'int', hidden_size: 'int'=128): """ Args: input_shape: observation shape of the environment n_actions: number of discrete actions available in the environment hidden_size: size of hidden layers """ super().__init__() self.fc1 = nn.Linear(input_shape[0], hidden_size) self.actor_head = nn.Linear(hidden_size, n_actions) self.critic_head = nn.Linear(hidden_size, 1) def forward(self, x) ->Tuple[Tensor, Tensor]: """Forward pass through network. Calculates the action logits and the value. Args: x: input to network Returns: action log probs (logits), value """ x = F.relu(self.fc1(x.float())) a = F.log_softmax(self.actor_head(x), dim=-1) c = self.critic_head(x) return a, c def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'input_shape': [4, 4], 'n_actions': 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 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_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__log_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 tl.store(out_ptr0 + x2, tmp8, xmask) @triton.jit def triton_poi_fused__log_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') tmp3 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last') tmp6 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last') tmp9 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last') tmp2 = tl_math.exp(tmp1) tmp4 = tl_math.exp(tmp3) tmp5 = tmp2 + tmp4 tmp7 = tl_math.exp(tmp6) tmp8 = tmp5 + tmp7 tmp10 = tl_math.exp(tmp9) tmp11 = tmp8 + tmp10 tmp12 = tl_math.log(tmp11) tmp13 = tmp0 - 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) = args args.clear() assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_2, (128, 4), (4, 1)) assert_size_stride(primals_3, (128,), (1,)) assert_size_stride(primals_4, (4, 128), (128, 1)) assert_size_stride(primals_5, (4,), (1,)) assert_size_stride(primals_6, (1, 128), (128, 1)) assert_size_stride(primals_7, (1,), (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_1, (64, 4), (4, 1), 0), reinterpret_tensor(primals_2, (4, 128), (1, 4), 0), out=buf0) del primals_2 buf1 = reinterpret_tensor(buf0, (4, 4, 4, 128), (2048, 512, 128, 1), 0) del buf0 buf7 = 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_3, buf7, 8192, XBLOCK=256, num_warps=4, num_stages=1) del primals_3 buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32) extern_kernels.addmm(primals_5, reinterpret_tensor(buf1, (64, 128), (128, 1), 0), reinterpret_tensor(primals_4, (128, 4), (1, 128), 0), alpha=1, beta=1, out=buf2) del primals_5 buf3 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused__log_softmax_1[grid(256)](buf2, buf3, 256, XBLOCK= 256, num_warps=4, num_stages=1) buf4 = reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0) del buf2 triton_poi_fused__log_softmax_2[grid(256)](buf3, buf4, 256, XBLOCK= 256, num_warps=4, num_stages=1) del buf3 buf6 = empty_strided_cuda((64, 1), (1, 1), torch.float32) extern_kernels.addmm(primals_7, reinterpret_tensor(buf1, (64, 128), (128, 1), 0), reinterpret_tensor(primals_6, (128, 1), (1, 128), 0), alpha=1, beta=1, out=buf6) del primals_7 return buf4, reinterpret_tensor(buf6, (4, 4, 4, 1), (16, 4, 1, 1), 0 ), reinterpret_tensor(primals_1, (64, 4), (4, 1), 0 ), reinterpret_tensor(buf1, (64, 128), (128, 1), 0 ), buf4, primals_6, primals_4, buf7 class ActorCriticMLPNew(nn.Module): """MLP network with heads for actor and critic.""" def __init__(self, input_shape: 'Tuple[int]', n_actions: 'int', hidden_size: 'int'=128): """ Args: input_shape: observation shape of the environment n_actions: number of discrete actions available in the environment hidden_size: size of hidden layers """ super().__init__() self.fc1 = nn.Linear(input_shape[0], hidden_size) self.actor_head = nn.Linear(hidden_size, n_actions) self.critic_head = nn.Linear(hidden_size, 1) def forward(self, input_0): primals_2 = self.fc1.weight primals_3 = self.fc1.bias primals_4 = self.actor_head.weight primals_5 = self.actor_head.bias primals_6 = self.critic_head.weight primals_7 = self.critic_head.bias primals_1 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7]) return output[0], output[1]
bzrry/lightning-bolts
ActorCriticMLP
false
14,987
[ "Apache-2.0" ]
822
bd392ad858039290c72c20cc3f10df39384e90b9
https://github.com/bzrry/lightning-bolts/tree/bd392ad858039290c72c20cc3f10df39384e90b9
Mlp
import torch import numpy as np from torch import nn import torch.nn.functional as F class GELU(nn.Module): def __init__(self): super(GELU, self).__init__() def forward(self, x): return 0.5 * x * (1 + F.tanh(np.sqrt(2 / np.pi) * (x + 0.044715 * torch.pow(x, 3)))) class Mlp(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=GELU, drop=0.0): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features drop_probs = drop, drop self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() self.drop1 = nn.Dropout(drop_probs[0]) self.fc2 = nn.Linear(hidden_features, out_features) self.drop2 = nn.Dropout(drop_probs[1]) def forward(self, x): x = self.fc1(x) x = self.act(x) x = self.drop1(x) x = self.fc2(x) x = self.drop2(x) return x def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_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 numpy as np from torch import 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_add_mul_pow_sqrt_tanh_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 = 0.5 tmp2 = tmp0 * tmp1 tmp3 = tmp0 * tmp0 tmp4 = tmp3 * tmp0 tmp5 = 0.044715 tmp6 = tmp4 * tmp5 tmp7 = tmp0 + tmp6 tmp8 = 0.7978845608028654 tmp9 = tmp8 * tmp7 tmp10 = libdevice.tanh(tmp9) tmp11 = 1.0 tmp12 = tmp10 + tmp11 tmp13 = tmp2 * tmp12 tl.store(out_ptr0 + x0, tmp13, 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)) 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,)) 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((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_mul_pow_sqrt_tanh_0[grid(256)](buf0, buf1, 256, 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, 4), ( 4, 1), 0), reinterpret_tensor(primals_4, (4, 4), (1, 4), 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, 4), (4, 1), 0), primals_4 class GELU(nn.Module): def __init__(self): super(GELU, self).__init__() def forward(self, x): return 0.5 * x * (1 + F.tanh(np.sqrt(2 / np.pi) * (x + 0.044715 * torch.pow(x, 3)))) class MlpNew(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=GELU, drop=0.0): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features drop_probs = drop, drop self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() self.drop1 = nn.Dropout(drop_probs[0]) self.fc2 = nn.Linear(hidden_features, out_features) self.drop2 = nn.Dropout(drop_probs[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_3 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5]) return output[0]
bubbliiiing/classification-pytorch
Mlp
false
14,988
[ "MIT" ]
88
ee62c05bd3094c3fab48bada5a57cb2ed8b61c11
https://github.com/bubbliiiing/classification-pytorch/tree/ee62c05bd3094c3fab48bada5a57cb2ed8b61c11
PatchEmbed
import torch from torch import nn class PatchEmbed(nn.Module): def __init__(self, input_shape=[224, 224], patch_size=16, in_chans=3, num_features=768, norm_layer=None, flatten=True): super().__init__() self.num_patches = input_shape[0] // patch_size * (input_shape[1] // patch_size) self.flatten = flatten self.proj = nn.Conv2d(in_chans, num_features, kernel_size= patch_size, stride=patch_size) self.norm = norm_layer(num_features) if norm_layer else nn.Identity() def forward(self, x): x = self.proj(x) if self.flatten: x = x.flatten(2).transpose(1, 2) x = self.norm(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 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 = 2304 xnumel = 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] xmask = xindex < xnumel x2 = xindex y3 = yindex y0 = yindex % 3 y1 = yindex // 3 tmp0 = tl.load(in_ptr0 + (x2 + 256 * y3), xmask & ymask, eviction_policy='evict_last') tl.store(out_ptr0 + (y0 + 3 * x2 + 768 * 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_convolution_2(in_ptr0, in_ptr1, out_ptr0, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): xnumel = 16 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 % 768 y1 = yindex // 768 y3 = yindex tmp0 = tl.load(in_ptr0 + (y0 + 768 * x2 + 12288 * y1), xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr1 + y0, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tl.store(out_ptr0 + (x2 + 16 * y3), tmp2, xmask) def call(args): primals_1, primals_2, primals_3 = args args.clear() assert_size_stride(primals_1, (768, 3, 16, 16), (768, 256, 16, 1)) assert_size_stride(primals_2, (768,), (1,)) assert_size_stride(primals_3, (4, 3, 64, 64), (12288, 4096, 64, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((768, 3, 16, 16), (768, 1, 48, 3), torch. float32) get_raw_stream(0) triton_poi_fused_0[grid(2304, 256)](primals_1, buf0, 2304, 256, 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 = extern_kernels.convolution(buf1, buf0, stride=(16, 16), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf2, (4, 768, 4, 4), (12288, 1, 3072, 768)) buf3 = empty_strided_cuda((4, 768, 4, 4), (12288, 16, 4, 1), torch. float32) triton_poi_fused_convolution_2[grid(3072, 16)](buf2, primals_2, buf3, 3072, 16, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del buf2 del primals_2 return reinterpret_tensor(buf3, (4, 16, 768), (12288, 1, 16), 0 ), buf0, buf1 class PatchEmbedNew(nn.Module): def __init__(self, input_shape=[224, 224], patch_size=16, in_chans=3, num_features=768, norm_layer=None, flatten=True): super().__init__() self.num_patches = input_shape[0] // patch_size * (input_shape[1] // patch_size) self.flatten = flatten self.proj = nn.Conv2d(in_chans, num_features, kernel_size= patch_size, stride=patch_size) self.norm = norm_layer(num_features) if norm_layer else nn.Identity() def forward(self, input_0): primals_1 = self.proj.weight primals_2 = self.proj.bias primals_3 = input_0 output = call([primals_1, primals_2, primals_3]) return output[0]
bubbliiiing/classification-pytorch
PatchEmbed
false
14,989
[ "MIT" ]
88
ee62c05bd3094c3fab48bada5a57cb2ed8b61c11
https://github.com/bubbliiiing/classification-pytorch/tree/ee62c05bd3094c3fab48bada5a57cb2ed8b61c11
QRNNLayer
import torch import torch.nn as nn from torch.optim import * class ForgetMult(torch.nn.Module): """ForgetMult computes a simple recurrent equation: h_t = f_t * x_t + (1 - f_t) * h_{t-1} This equation is equivalent to dynamic weighted averaging. Inputs: X, hidden - X (seq_len, batch, input_size): tensor containing the features of the input sequence. - F (seq_len, batch, input_size): tensor containing the forget gate values, assumed in range [0, 1]. - hidden_init (batch, input_size): tensor containing the initial hidden state for the recurrence (h_{t-1}). """ def __init__(self): super(ForgetMult, self).__init__() def forward(self, f, x, hidden_init=None): result = [] forgets = f.split(1, dim=0) prev_h = hidden_init for i, h in enumerate((f * x).split(1, dim=0)): if prev_h is not None: h = h + (1 - forgets[i]) * prev_h h = h.view(h.size()[1:]) result.append(h) prev_h = h return torch.stack(result) class QRNNLayer(nn.Module): """Applies a single layer Quasi-Recurrent Neural Network (QRNN) to an input sequence. Args: input_size: The number of expected features in the input x. hidden_size: The number of features in the hidden state h. If not specified, the input size is used. save_prev_x: Whether to store previous inputs for use in future convolutional windows (i.e. for a continuing sequence such as in language modeling). If true, you must call reset to remove cached previous values of x. Default: False. window: Defines the size of the convolutional window (how many previous tokens to look when computing the QRNN values). Supports 1 and 2. Default: 1. zoneout: Whether to apply zoneout (i.e. failing to update elements in the hidden state) to the hidden state updates. Default: 0. output_gate: If True, performs QRNN-fo (applying an output gate to the output). If False, performs QRNN-f. Default: True. Inputs: X, hidden - X (seq_len, batch, input_size): tensor containing the features of the input sequence. - hidden (batch, hidden_size): tensor containing the initial hidden state for the QRNN. Outputs: output, h_n - output (seq_len, batch, hidden_size): tensor containing the output of the QRNN for each timestep. - h_n (1, batch, hidden_size): tensor containing the hidden state for t=seq_len """ def __init__(self, input_size, hidden_size=None, save_prev_x=False, zoneout=0, window=1, output_gate=True): super(QRNNLayer, self).__init__() assert window in [1, 2 ], 'This QRNN implementation currently only handles convolutional window of size 1 or size 2' self.window = window self.input_size = input_size self.hidden_size = hidden_size if hidden_size else input_size self.zoneout = zoneout self.save_prev_x = save_prev_x self.prevX = None self.output_gate = output_gate self.linear = nn.Linear(self.window * self.input_size, 3 * self. hidden_size if self.output_gate else 2 * self.hidden_size) def reset(self): self.prevX = None def forward(self, X, hidden=None): seq_len, batch_size, _ = X.size() source = None if self.window == 1: source = X elif self.window == 2: Xm1 = [] Xm1.append(self.prevX if self.prevX is not None else X[:1, :, : ] * 0) if len(X) > 1: Xm1.append(X[:-1, :, :]) Xm1 = torch.cat(Xm1, 0) source = torch.cat([X, Xm1], 2) Y = self.linear(source) if self.output_gate: Y = Y.view(seq_len, batch_size, 3 * self.hidden_size) Z, F, O = Y.chunk(3, dim=2) else: Y = Y.view(seq_len, batch_size, 2 * self.hidden_size) Z, F = Y.chunk(2, dim=2) Z = torch.tanh(Z) F = torch.sigmoid(F) if self.zoneout: if self.training: mask = F.new_empty(F.size(), requires_grad=False).bernoulli_( 1 - self.zoneout) F = F * mask else: F *= 1 - self.zoneout C = ForgetMult()(F, Z, hidden) if self.output_gate: H = torch.sigmoid(O) * C else: H = C if self.window > 1 and self.save_prev_x: self.prevX = X[-1:, :, :].detach() return H, C[-1:, :, :] def get_inputs(): return [torch.rand([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.triton_helpers import libdevice import torch.nn as nn from torch.optim import * 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_tanh_0(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 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') tmp4 = tl.load(in_ptr0 + (4 + x0 + 12 * x1), xmask) tmp5 = tl.load(in_ptr1 + (4 + x0), xmask, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = libdevice.tanh(tmp2) tmp6 = tmp4 + tmp5 tmp7 = tl.sigmoid(tmp6) tmp8 = tmp7 * tmp3 tl.store(out_ptr0 + x2, tmp3, xmask) tl.store(out_ptr1 + x2, tmp7, xmask) tl.store(out_ptr2 + x2, tmp8, xmask) @triton.jit def triton_poi_fused_add_mul_rsub_1(in_ptr0, in_ptr1, out_ptr0, out_ptr1, out_ptr2, out_ptr3, 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 + (16 + x0), xmask) tmp3 = tl.load(in_ptr1 + (16 + x0), xmask) tmp4 = tl.load(in_ptr1 + x0, xmask) tmp7 = tl.load(in_ptr0 + (32 + x0), xmask) tmp9 = tl.load(in_ptr1 + (32 + x0), xmask) tmp1 = 1.0 tmp2 = tmp1 - tmp0 tmp5 = tmp2 * tmp4 tmp6 = tmp3 + tmp5 tmp8 = tmp1 - tmp7 tmp10 = tmp8 * tmp6 tmp11 = tmp9 + tmp10 tl.store(out_ptr0 + x0, tmp2, xmask) tl.store(out_ptr1 + x0, tmp6, xmask) tl.store(out_ptr2 + x0, tmp8, xmask) tl.store(out_ptr3 + x0, tmp11, xmask) @triton.jit def triton_poi_fused_rsub_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 x0 = xindex tmp0 = tl.load(in_ptr0 + (48 + x0), xmask) tmp1 = 1.0 tmp2 = tmp1 - tmp0 tl.store(out_ptr0 + x0, tmp2, xmask) @triton.jit def triton_poi_fused_mul_sigmoid_stack_3(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, 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 x1 = xindex // 4 x0 = xindex % 4 x2 = xindex tmp29 = tl.load(in_ptr4 + (8 + x0 + 12 * x1), xmask) tmp30 = tl.load(in_ptr5 + (8 + x0), xmask, eviction_policy='evict_last') tmp0 = x1 tl.full([1], 0, tl.int64) tmp3 = tl.full([1], 4, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tl.load(in_ptr0 + (x0 + 4 * x1), tmp4 & xmask, other=0.0) tmp6 = tmp0 >= tmp3 tmp7 = tl.full([1], 8, tl.int64) tmp8 = tmp0 < tmp7 tmp9 = tmp6 & tmp8 tmp10 = tl.load(in_ptr1 + (x0 + 4 * (-4 + x1)), tmp9 & xmask, other=0.0) tmp11 = tmp0 >= tmp7 tmp12 = tl.full([1], 12, tl.int64) tmp13 = tmp0 < tmp12 tmp14 = tmp11 & tmp13 tmp15 = tl.load(in_ptr2 + (x0 + 4 * (-8 + x1)), tmp14 & xmask, other=0.0) tmp16 = tmp0 >= tmp12 tl.full([1], 16, tl.int64) tmp19 = tl.load(in_ptr0 + (48 + x0 + 4 * (-12 + x1)), tmp16 & xmask, other=0.0) tmp20 = tl.load(in_ptr3 + (x0 + 4 * (-12 + x1)), tmp16 & xmask, other=0.0) tmp21 = tl.load(in_ptr2 + (x0 + 4 * (-12 + x1)), tmp16 & xmask, other=0.0) tmp22 = tmp20 * tmp21 tmp23 = tmp19 + tmp22 tmp24 = tl.full(tmp23.shape, 0.0, tmp23.dtype) tmp25 = tl.where(tmp16, tmp23, tmp24) tmp26 = tl.where(tmp14, tmp15, tmp25) tmp27 = tl.where(tmp9, tmp10, tmp26) tmp28 = tl.where(tmp4, tmp5, tmp27) tmp31 = tmp29 + tmp30 tmp32 = tl.sigmoid(tmp31) tmp33 = tmp32 * tmp28 tl.store(out_ptr0 + x2, tmp28, xmask) tl.store(out_ptr1 + x2, tmp32, xmask) tl.store(out_ptr2 + x2, tmp33, 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, (12, 4), (4, 1)) assert_size_stride(primals_3, (12,), (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_1, (16, 4), (4, 1), 0), reinterpret_tensor(primals_2, (4, 12), (1, 4), 0), out=buf0) del primals_2 buf1 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) buf2 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) buf3 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_mul_sigmoid_tanh_0[grid(64)](buf0, primals_3, buf1, buf2, buf3, 64, XBLOCK=64, num_warps=1, num_stages=1) buf4 = empty_strided_cuda((1, 4, 4), (16, 4, 1), torch.float32) buf5 = empty_strided_cuda((1, 4, 4), (16, 4, 1), torch.float32) buf6 = empty_strided_cuda((1, 4, 4), (16, 4, 1), torch.float32) buf7 = empty_strided_cuda((1, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused_add_mul_rsub_1[grid(16)](buf2, buf3, buf4, buf5, buf6, buf7, 16, XBLOCK=16, num_warps=1, num_stages=1) buf8 = empty_strided_cuda((1, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused_rsub_2[grid(16)](buf2, buf8, 16, XBLOCK=16, num_warps=1, num_stages=1) buf9 = empty_strided_cuda((16, 4), (4, 1), torch.float32) buf10 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) buf11 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused_mul_sigmoid_stack_3[grid(64)](buf3, buf5, buf7, buf8, buf0, primals_3, buf9, buf10, buf11, 64, XBLOCK=64, num_warps=1, num_stages=1) del buf0 del primals_3 return buf11, reinterpret_tensor(buf9, (1, 4, 4), (16, 4, 1), 48 ), reinterpret_tensor(primals_1, (16, 4), (4, 1), 0 ), buf1, buf2, reinterpret_tensor(buf3, (4, 4), (4, 1), 0 ), buf4, reinterpret_tensor(buf5, (4, 4), (4, 1), 0 ), buf6, reinterpret_tensor(buf7, (4, 4), (4, 1), 0), buf8, buf9, buf10 class ForgetMult(torch.nn.Module): """ForgetMult computes a simple recurrent equation: h_t = f_t * x_t + (1 - f_t) * h_{t-1} This equation is equivalent to dynamic weighted averaging. Inputs: X, hidden - X (seq_len, batch, input_size): tensor containing the features of the input sequence. - F (seq_len, batch, input_size): tensor containing the forget gate values, assumed in range [0, 1]. - hidden_init (batch, input_size): tensor containing the initial hidden state for the recurrence (h_{t-1}). """ def __init__(self): super(ForgetMult, self).__init__() def forward(self, f, x, hidden_init=None): result = [] forgets = f.split(1, dim=0) prev_h = hidden_init for i, h in enumerate((f * x).split(1, dim=0)): if prev_h is not None: h = h + (1 - forgets[i]) * prev_h h = h.view(h.size()[1:]) result.append(h) prev_h = h return torch.stack(result) class QRNNLayerNew(nn.Module): """Applies a single layer Quasi-Recurrent Neural Network (QRNN) to an input sequence. Args: input_size: The number of expected features in the input x. hidden_size: The number of features in the hidden state h. If not specified, the input size is used. save_prev_x: Whether to store previous inputs for use in future convolutional windows (i.e. for a continuing sequence such as in language modeling). If true, you must call reset to remove cached previous values of x. Default: False. window: Defines the size of the convolutional window (how many previous tokens to look when computing the QRNN values). Supports 1 and 2. Default: 1. zoneout: Whether to apply zoneout (i.e. failing to update elements in the hidden state) to the hidden state updates. Default: 0. output_gate: If True, performs QRNN-fo (applying an output gate to the output). If False, performs QRNN-f. Default: True. Inputs: X, hidden - X (seq_len, batch, input_size): tensor containing the features of the input sequence. - hidden (batch, hidden_size): tensor containing the initial hidden state for the QRNN. Outputs: output, h_n - output (seq_len, batch, hidden_size): tensor containing the output of the QRNN for each timestep. - h_n (1, batch, hidden_size): tensor containing the hidden state for t=seq_len """ def __init__(self, input_size, hidden_size=None, save_prev_x=False, zoneout=0, window=1, output_gate=True): super(QRNNLayerNew, self).__init__() assert window in [1, 2 ], 'This QRNN implementation currently only handles convolutional window of size 1 or size 2' self.window = window self.input_size = input_size self.hidden_size = hidden_size if hidden_size else input_size self.zoneout = zoneout self.save_prev_x = save_prev_x self.prevX = None self.output_gate = output_gate self.linear = nn.Linear(self.window * self.input_size, 3 * self. hidden_size if self.output_gate else 2 * self.hidden_size) def reset(self): self.prevX = None def forward(self, input_0): primals_2 = self.linear.weight primals_3 = self.linear.bias primals_1 = input_0 output = call([primals_1, primals_2, primals_3]) return output[0], output[1]
boshining/NeuronBlocks
QRNNLayer
false
14,990
[ "MIT" ]
1,257
74fbb8658fb3f1cffea5c9bc84b2a1da59c20dd9
https://github.com/boshining/NeuronBlocks/tree/74fbb8658fb3f1cffea5c9bc84b2a1da59c20dd9
DiffLoss
import torch import torch.nn as nn import torch.utils.checkpoint class DiffLoss(nn.Module): def __init__(self): super(DiffLoss, self).__init__() def forward(self, input1, input2): batch_size = input1.size(0) input1 = input1.view(batch_size, -1) input2 = input2.view(batch_size, -1) input1_mean = torch.mean(input1, dim=0, keepdims=True) input2_mean = torch.mean(input2, dim=0, keepdims=True) input1 = input1 - input1_mean input2 = input2 - input2_mean input1_l2_norm = torch.norm(input1, p=2, dim=1, keepdim=True).detach() input1_l2 = input1.div(input1_l2_norm.expand_as(input1) + 1e-06) input2_l2_norm = torch.norm(input2, p=2, dim=1, keepdim=True).detach() input2_l2 = input2.div(input2_l2_norm.expand_as(input2) + 1e-06) diff_loss = torch.mean(input1_l2.t().mm(input2_l2).pow(2)) return diff_loss def get_inputs(): return [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.triton_helpers import libdevice import torch.nn as nn import torch.utils.checkpoint 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_div_linalg_vector_norm_mean_sub_0(in_ptr0, out_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr): xnumel = 4 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) r1 = rindex x0 = xindex tmp0 = tl.load(in_ptr0 + (r1 + 64 * x0), xmask, other=0.0) tmp1 = tl.load(in_ptr0 + r1, None, eviction_policy='evict_last') tmp2 = tl.load(in_ptr0 + (64 + r1), None, eviction_policy='evict_last') tmp4 = tl.load(in_ptr0 + (128 + r1), None, eviction_policy='evict_last') tmp6 = tl.load(in_ptr0 + (192 + r1), None, eviction_policy='evict_last') tmp3 = tmp1 + tmp2 tmp5 = tmp3 + tmp4 tmp7 = tmp5 + tmp6 tmp8 = 4.0 tmp9 = tmp7 / tmp8 tmp10 = tmp0 - tmp9 tmp11 = tmp10 * tmp10 tmp12 = tl.broadcast_to(tmp11, [XBLOCK, RBLOCK]) tmp14 = tl.where(xmask, tmp12, 0) tmp15 = tl.sum(tmp14, 1)[:, None] tmp16 = libdevice.sqrt(tmp15) tmp17 = 1e-06 tmp18 = tmp16 + tmp17 tmp19 = tmp10 / tmp18 tl.store(out_ptr2 + (r1 + 64 * x0), tmp19, xmask) @triton.jit def triton_red_fused_mean_pow_1(in_out_ptr0, in_ptr0, xnumel, rnumel, XBLOCK: tl.constexpr, RBLOCK: tl.constexpr): rnumel = 4096 xoffset = tl.program_id(0) * XBLOCK xoffset + tl.arange(0, XBLOCK)[:, None] tl.full([XBLOCK, RBLOCK], True, tl.int1) rbase = tl.arange(0, RBLOCK)[None, :] _tmp3 = tl.full([XBLOCK, RBLOCK], 0, tl.float32) for roffset in range(0, rnumel, RBLOCK): rindex = roffset + rbase rmask = rindex < rnumel r0 = rindex tmp0 = tl.load(in_ptr0 + r0, rmask, eviction_policy='evict_first', other=0.0) tmp1 = tmp0 * tmp0 tmp2 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK]) tmp4 = _tmp3 + tmp2 _tmp3 = tl.where(rmask, tmp4, _tmp3) tmp3 = tl.sum(_tmp3, 1)[:, None] tmp5 = 4096.0 tmp6 = tmp3 / 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) buf4 = empty_strided_cuda((4, 64), (64, 1), torch.float32) get_raw_stream(0) triton_per_fused_add_div_linalg_vector_norm_mean_sub_0[grid(4)](arg0_1, buf4, 4, 64, XBLOCK=1, num_warps=2, num_stages=1) del arg0_1 buf5 = empty_strided_cuda((4, 64), (64, 1), torch.float32) triton_per_fused_add_div_linalg_vector_norm_mean_sub_0[grid(4)](arg1_1, buf5, 4, 64, XBLOCK=1, num_warps=2, num_stages=1) del arg1_1 buf6 = empty_strided_cuda((64, 64), (64, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(buf4, (64, 4), (1, 64), 0), buf5, out=buf6) del buf4 del buf5 buf7 = empty_strided_cuda((), (), torch.float32) buf8 = buf7 del buf7 triton_red_fused_mean_pow_1[grid(1)](buf8, buf6, 1, 4096, XBLOCK=1, RBLOCK=2048, num_warps=16, num_stages=1) del buf6 return buf8, class DiffLossNew(nn.Module): def __init__(self): super(DiffLossNew, 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]
byamao1/MMSA
DiffLoss
false
14,991
[ "MIT" ]
198
1a894d042144c9ac75b3465d38871ce8c2987251
https://github.com/byamao1/MMSA/tree/1a894d042144c9ac75b3465d38871ce8c2987251
MultiheadAttention
import torch import torch.nn as nn import torch.nn.functional as F from torch.nn.parameter import Parameter import torch.utils.checkpoint from torch.nn import Parameter class MultiheadAttention(nn.Module): """Multi-headed attention. See "Attention Is All You Need" for more details. """ def __init__(self, embed_dim, num_heads, attn_dropout=0.0, bias=True, add_bias_kv=False, add_zero_attn=False): super().__init__() self.embed_dim = embed_dim self.num_heads = num_heads self.attn_dropout = attn_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.in_proj_weight = Parameter(torch.Tensor(3 * embed_dim, embed_dim)) self.register_parameter('in_proj_bias', None) if bias: self.in_proj_bias = Parameter(torch.Tensor(3 * embed_dim)) 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() def reset_parameters(self): nn.init.xavier_uniform_(self.in_proj_weight) nn.init.xavier_uniform_(self.out_proj.weight) if self.in_proj_bias is not None: nn.init.constant_(self.in_proj_bias, 0.0) 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, value, attn_mask=None): """Input shape: Time x Batch x Channel Self-attention can be implemented by passing in the same arguments for query, key and value. Timesteps can be masked by supplying a T x T mask in the `attn_mask` argument. Padding elements can be excluded from the key by passing a binary ByteTensor (`key_padding_mask`) with shape: batch x src_len, where padding elements are indicated by 1s. """ qkv_same = query.data_ptr() == key.data_ptr() == value.data_ptr() kv_same = key.data_ptr() == value.data_ptr() tgt_len, bsz, embed_dim = query.size() assert embed_dim == self.embed_dim assert list(query.size()) == [tgt_len, bsz, embed_dim] assert key.size() == value.size() if qkv_same: q, k, v = self.in_proj_qkv(query) elif kv_same: q = self.in_proj_q(query) if key is None: assert value is None k = v = None else: k, v = self.in_proj_kv(key) else: q = self.in_proj_q(query) k = self.in_proj_k(key) v = self.in_proj_v(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) 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) src_len = k.size(1) if self.add_zero_attn: 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) attn_weights = torch.bmm(q, k.transpose(1, 2)) assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len] if attn_mask is not None: try: attn_weights += attn_mask.unsqueeze(0) except: None None assert False attn_weights = F.softmax(attn_weights.float(), dim=-1).type_as( attn_weights) attn_weights = F.dropout(attn_weights, p=self.attn_dropout, training=self.training) attn = torch.bmm(attn_weights, v) assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self. head_dim] attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim) attn = self.out_proj(attn) attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) attn_weights = attn_weights.sum(dim=1) / self.num_heads return attn, attn_weights def in_proj_qkv(self, query): return self._in_proj(query).chunk(3, dim=-1) def in_proj_kv(self, key): return self._in_proj(key, start=self.embed_dim).chunk(2, dim=-1) def in_proj_q(self, query, **kwargs): return self._in_proj(query, end=self.embed_dim, **kwargs) def in_proj_k(self, key): return self._in_proj(key, start=self.embed_dim, end=2 * self.embed_dim) def in_proj_v(self, value): return self._in_proj(value, start=2 * self.embed_dim) def _in_proj(self, input, start=0, end=None, **kwargs): weight = kwargs.get('weight', self.in_proj_weight) bias = kwargs.get('bias', self.in_proj_bias) weight = weight[start:end, :] if bias is not None: bias = bias[start:end] return F.linear(input, weight, bias) def get_inputs(): return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'embed_dim': 4, 'num_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 torch.nn as nn import torch.nn.functional as F from torch.nn.parameter import Parameter import torch.utils.checkpoint 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_mul_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 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 = 1.0 tmp4 = tmp2 * tmp3 tl.store(in_out_ptr0 + x2, tmp4, xmask) @triton.jit def triton_per_fused__softmax_1(in_ptr0, out_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr): xnumel = 64 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_poi_fused_clone_2(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_div_sum_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 x0 = xindex % 64 x1 = xindex // 64 x2 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 256 * x1), xmask) tmp1 = tl.load(in_ptr0 + (64 + x0 + 256 * x1), xmask) tmp3 = tl.load(in_ptr0 + (128 + x0 + 256 * x1), xmask) tmp5 = tl.load(in_ptr0 + (192 + x0 + 256 * x1), xmask) tmp2 = tmp0 + tmp1 tmp4 = tmp2 + tmp3 tmp6 = tmp4 + tmp5 tmp7 = 0.25 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) = args args.clear() assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1)) assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_4, (12, 4), (4, 1)) assert_size_stride(primals_5, (12,), (1,)) assert_size_stride(primals_6, (4, 4), (4, 1)) assert_size_stride(primals_7, (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_4, (4, 4), (1, 4), 0), out=buf0) buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32) extern_kernels.addmm(reinterpret_tensor(primals_5, (4,), (1,), 4), reinterpret_tensor(primals_2, (64, 4), (4, 1), 0), reinterpret_tensor(primals_4, (4, 4), (1, 4), 16), alpha=1, beta=1, out=buf1) buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32) extern_kernels.addmm(reinterpret_tensor(primals_5, (4,), (1,), 8), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(primals_4, (4, 4), (1, 4), 32), alpha=1, beta=1, out=buf2) del primals_4 buf3 = reinterpret_tensor(buf0, (4, 4, 4), (16, 4, 1), 0) del buf0 get_raw_stream(0) triton_poi_fused_mul_0[grid(64)](buf3, primals_5, 64, XBLOCK=64, num_warps=1, num_stages=1) del primals_5 buf4 = empty_strided_cuda((16, 4, 16), (64, 16, 1), torch.float32) extern_kernels.bmm(reinterpret_tensor(buf3, (16, 4, 1), (1, 16, 0), 0), reinterpret_tensor(buf1, (16, 1, 16), (1, 1, 16), 0), out=buf4) buf7 = empty_strided_cuda((16, 4, 16), (64, 16, 1), torch.float32) triton_per_fused__softmax_1[grid(64)](buf4, buf7, 64, 16, XBLOCK=1, num_warps=2, num_stages=1) del buf4 buf8 = empty_strided_cuda((16, 4, 1), (4, 1, 1), torch.float32) extern_kernels.bmm(buf7, reinterpret_tensor(buf2, (16, 16, 1), (1, 16, 1), 0), out=buf8) buf9 = empty_strided_cuda((4, 16, 1), (16, 1, 1), torch.float32) triton_poi_fused_clone_2[grid(4, 16)](buf8, buf9, 4, 16, XBLOCK=16, YBLOCK=4, num_warps=1, num_stages=1) buf10 = reinterpret_tensor(buf8, (16, 4), (4, 1), 0) del buf8 extern_kernels.addmm(primals_7, reinterpret_tensor(buf9, (16, 4), ( 4, 1), 0), reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf10) del primals_7 buf11 = empty_strided_cuda((4, 4, 16), (64, 16, 1), torch.float32) triton_poi_fused_div_sum_3[grid(256)](buf7, buf11, 256, XBLOCK=128, num_warps=4, num_stages=1) return reinterpret_tensor(buf10, (4, 4, 4), (16, 4, 1), 0 ), buf11, reinterpret_tensor(primals_1, (16, 4), (4, 1), 0 ), reinterpret_tensor(primals_2, (64, 4), (4, 1), 0 ), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0 ), buf7, reinterpret_tensor(buf9, (16, 4), (4, 1), 0 ), primals_6, reinterpret_tensor(buf2, (16, 1, 16), (1, 1, 16), 0 ), reinterpret_tensor(buf3, (16, 1, 4), (1, 1, 16), 0 ), reinterpret_tensor(buf1, (16, 16, 1), (1, 16, 1), 0) class MultiheadAttentionNew(nn.Module): """Multi-headed attention. See "Attention Is All You Need" for more details. """ def __init__(self, embed_dim, num_heads, attn_dropout=0.0, bias=True, add_bias_kv=False, add_zero_attn=False): super().__init__() self.embed_dim = embed_dim self.num_heads = num_heads self.attn_dropout = attn_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.in_proj_weight = Parameter(torch.Tensor(3 * embed_dim, embed_dim)) self.register_parameter('in_proj_bias', None) if bias: self.in_proj_bias = Parameter(torch.Tensor(3 * embed_dim)) 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() def reset_parameters(self): nn.init.xavier_uniform_(self.in_proj_weight) nn.init.xavier_uniform_(self.out_proj.weight) if self.in_proj_bias is not None: nn.init.constant_(self.in_proj_bias, 0.0) 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 in_proj_qkv(self, query): return self._in_proj(query).chunk(3, dim=-1) def in_proj_kv(self, key): return self._in_proj(key, start=self.embed_dim).chunk(2, dim=-1) def in_proj_q(self, query, **kwargs): return self._in_proj(query, end=self.embed_dim, **kwargs) def in_proj_k(self, key): return self._in_proj(key, start=self.embed_dim, end=2 * self.embed_dim) def in_proj_v(self, value): return self._in_proj(value, start=2 * self.embed_dim) def _in_proj(self, input, start=0, end=None, **kwargs): weight = kwargs.get('weight', self.in_proj_weight) bias = kwargs.get('bias', self.in_proj_bias) weight = weight[start:end, :] if bias is not None: bias = bias[start:end] return F.linear(input, weight, bias) def forward(self, input_0, input_1, input_2): primals_4 = self.in_proj_weight primals_5 = self.in_proj_bias primals_6 = self.out_proj.weight primals_7 = self.out_proj.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, primals_6, primals_7]) return output[0], output[1]
byamao1/MMSA
MultiheadAttention
false
14,992
[ "MIT" ]
198
1a894d042144c9ac75b3465d38871ce8c2987251
https://github.com/byamao1/MMSA/tree/1a894d042144c9ac75b3465d38871ce8c2987251
Discriminator
import torch import numpy as np from torch import nn from torch.nn import functional as F class Discriminator(nn.Module): def __init__(self, img_shape, hidden_dim=1024): super().__init__() in_dim = int(np.prod(img_shape)) self.fc1 = nn.Linear(in_dim, hidden_dim) self.fc2 = nn.Linear(self.fc1.out_features, self.fc1.out_features // 2) self.fc3 = nn.Linear(self.fc2.out_features, self.fc2.out_features // 2) self.fc4 = nn.Linear(self.fc3.out_features, 1) def forward(self, img): x = img.view(img.size(0), -1) x = F.leaky_relu(self.fc1(x), 0.2) x = F.dropout(x, 0.3) x = F.leaky_relu(self.fc2(x), 0.2) x = F.dropout(x, 0.3) x = F.leaky_relu(self.fc3(x), 0.2) x = F.dropout(x, 0.3) return torch.sigmoid(self.fc4(x)) def get_inputs(): return [torch.rand([4, 4])] def get_init_inputs(): return [[], {'img_shape': 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 numpy as np 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_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 % 1024 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 = 0.2 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_leaky_relu_1(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 % 512 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 = 0.2 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_leaky_relu_2(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 % 256 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 = 0.2 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_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) = args args.clear() assert_size_stride(primals_1, (4, 4), (4, 1)) assert_size_stride(primals_2, (1024, 4), (4, 1)) assert_size_stride(primals_3, (1024,), (1,)) assert_size_stride(primals_4, (512, 1024), (1024, 1)) assert_size_stride(primals_5, (512,), (1,)) assert_size_stride(primals_6, (256, 512), (512, 1)) assert_size_stride(primals_7, (256,), (1,)) assert_size_stride(primals_8, (1, 256), (256, 1)) assert_size_stride(primals_9, (1,), (1,)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 1024), (1024, 1), torch.float32) extern_kernels.mm(primals_1, reinterpret_tensor(primals_2, (4, 1024 ), (1, 4), 0), out=buf0) del primals_2 buf1 = empty_strided_cuda((4, 1024), (1024, 1), torch.bool) buf2 = empty_strided_cuda((4, 1024), (1024, 1), torch.float32) get_raw_stream(0) triton_poi_fused_leaky_relu_0[grid(4096)](buf0, primals_3, buf1, buf2, 4096, XBLOCK=256, num_warps=4, num_stages=1) del buf0 del primals_3 buf3 = torch.ops.aten.native_dropout.default(buf2, 0.3, True) del buf2 buf4 = buf3[0] buf5 = buf3[1] del buf3 buf6 = empty_strided_cuda((4, 512), (512, 1), torch.float32) extern_kernels.mm(buf4, reinterpret_tensor(primals_4, (1024, 512), (1, 1024), 0), out=buf6) buf7 = empty_strided_cuda((4, 512), (512, 1), torch.bool) buf8 = empty_strided_cuda((4, 512), (512, 1), torch.float32) triton_poi_fused_leaky_relu_1[grid(2048)](buf6, primals_5, buf7, buf8, 2048, XBLOCK=128, num_warps=4, num_stages=1) del buf6 del primals_5 buf9 = torch.ops.aten.native_dropout.default(buf8, 0.3, True) del buf8 buf10 = buf9[0] buf11 = buf9[1] del buf9 buf12 = empty_strided_cuda((4, 256), (256, 1), torch.float32) extern_kernels.mm(buf10, reinterpret_tensor(primals_6, (512, 256), (1, 512), 0), out=buf12) buf13 = empty_strided_cuda((4, 256), (256, 1), torch.bool) buf14 = empty_strided_cuda((4, 256), (256, 1), torch.float32) triton_poi_fused_leaky_relu_2[grid(1024)](buf12, primals_7, buf13, buf14, 1024, XBLOCK=256, num_warps=4, num_stages=1) del buf12 del primals_7 buf15 = torch.ops.aten.native_dropout.default(buf14, 0.3, True) del buf14 buf16 = buf15[0] buf17 = buf15[1] del buf15 buf18 = empty_strided_cuda((4, 1), (1, 1), torch.float32) extern_kernels.mm(buf16, reinterpret_tensor(primals_8, (256, 1), (1, 256), 0), out=buf18) buf19 = buf18 del buf18 triton_poi_fused_sigmoid_3[grid(4)](buf19, primals_9, 4, XBLOCK=4, num_warps=1, num_stages=1) del primals_9 return (buf19, primals_1, buf1, buf4, buf5, buf7, buf10, buf11, buf13, buf16, buf17, buf19, primals_8, primals_6, primals_4) class DiscriminatorNew(nn.Module): def __init__(self, img_shape, hidden_dim=1024): super().__init__() in_dim = int(np.prod(img_shape)) self.fc1 = nn.Linear(in_dim, hidden_dim) self.fc2 = nn.Linear(self.fc1.out_features, self.fc1.out_features // 2) self.fc3 = nn.Linear(self.fc2.out_features, self.fc2.out_features // 2) self.fc4 = nn.Linear(self.fc3.out_features, 1) def forward(self, input_0): primals_2 = self.fc1.weight primals_3 = 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.fc4.weight primals_9 = self.fc4.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]
bzrry/lightning-bolts
Discriminator
false
14,993
[ "Apache-2.0" ]
822
bd392ad858039290c72c20cc3f10df39384e90b9
https://github.com/bzrry/lightning-bolts/tree/bd392ad858039290c72c20cc3f10df39384e90b9
SchedulerTestNet
import torch from torch.nn import functional as F class SchedulerTestNet(torch.nn.Module): """adapted from: https://github.com/pytorch/pytorch/blob/master/test/test_optim.py.""" def __init__(self): super().__init__() self.conv1 = torch.nn.Conv2d(1, 1, 1) self.conv2 = torch.nn.Conv2d(1, 1, 1) def forward(self, x): return self.conv2(F.relu(self.conv1(x))) def get_inputs(): return [torch.rand([4, 1, 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 assert_size_stride = torch._C._dynamo.guards.assert_size_stride @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) 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.full([1], 0, tl.int32) tmp5 = triton_helpers.maximum(tmp4, tmp3) tl.store(in_out_ptr0 + x0, tmp5, None) @triton.jit def triton_poi_fused_convolution_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) 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 tl.store(in_out_ptr0 + x0, tmp3, None) def call(args): primals_1, primals_2, primals_3, primals_4, primals_5 = args args.clear() assert_size_stride(primals_1, (1, 1, 1, 1), (1, 1, 1, 1)) assert_size_stride(primals_2, (1,), (1,)) assert_size_stride(primals_3, (4, 1, 64, 64), (4096, 4096, 64, 1)) assert_size_stride(primals_4, (1, 1, 1, 1), (1, 1, 1, 1)) assert_size_stride(primals_5, (1,), (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=1, bias=None) assert_size_stride(buf0, (4, 1, 64, 64), (4096, 4096, 64, 1)) buf1 = buf0 del buf0 get_raw_stream(0) triton_poi_fused_convolution_relu_0[grid(16384)](buf1, primals_2, 16384, XBLOCK=256, num_warps=4, num_stages=1) del primals_2 buf2 = extern_kernels.convolution(buf1, primals_4, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf2, (4, 1, 64, 64), (4096, 4096, 64, 1)) buf3 = buf2 del buf2 triton_poi_fused_convolution_1[grid(16384)](buf3, primals_5, 16384, XBLOCK=256, num_warps=4, num_stages=1) del primals_5 return buf3, primals_1, primals_3, primals_4, buf1 class SchedulerTestNetNew(torch.nn.Module): """adapted from: https://github.com/pytorch/pytorch/blob/master/test/test_optim.py.""" def __init__(self): super().__init__() self.conv1 = torch.nn.Conv2d(1, 1, 1) self.conv2 = torch.nn.Conv2d(1, 1, 1) 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_3 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5]) return output[0]
bzrry/lightning-bolts
SchedulerTestNet
false
14,994
[ "Apache-2.0" ]
822
bd392ad858039290c72c20cc3f10df39384e90b9
https://github.com/bzrry/lightning-bolts/tree/bd392ad858039290c72c20cc3f10df39384e90b9
SELoss
import torch from torch import Tensor from torch import nn class SELoss(nn.MSELoss): def __init__(self): super().__init__(reduction='none') def forward(self, inputs: 'Tensor', target: 'Tensor') ->Tensor: return super().forward(inputs, target).sum(1) 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 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_mse_loss_sum_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 % 16 x1 = xindex // 16 x2 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask) tmp1 = tl.load(in_ptr1 + (x0 + 64 * x1), xmask) tmp4 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask) tmp5 = tl.load(in_ptr1 + (16 + x0 + 64 * x1), xmask) tmp9 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask) tmp10 = tl.load(in_ptr1 + (32 + x0 + 64 * x1), xmask) tmp14 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask) tmp15 = tl.load(in_ptr1 + (48 + x0 + 64 * x1), xmask) tmp2 = tmp0 - tmp1 tmp3 = tmp2 * tmp2 tmp6 = tmp4 - tmp5 tmp7 = tmp6 * tmp6 tmp8 = tmp3 + tmp7 tmp11 = tmp9 - tmp10 tmp12 = tmp11 * tmp11 tmp13 = tmp8 + tmp12 tmp16 = tmp14 - tmp15 tmp17 = tmp16 * tmp16 tmp18 = tmp13 + tmp17 tl.store(out_ptr0 + x2, tmp18, xmask) 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, 4), (16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_mse_loss_sum_0[grid(64)](arg1_1, arg0_1, buf0, 64, XBLOCK=64, num_warps=1, num_stages=1) del arg0_1 del arg1_1 return buf0, class SELossNew(nn.MSELoss): def __init__(self): super().__init__(reduction='none') def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0]
bzrry/lightning-bolts
SELoss
false
14,995
[ "Apache-2.0" ]
822
bd392ad858039290c72c20cc3f10df39384e90b9
https://github.com/bzrry/lightning-bolts/tree/bd392ad858039290c72c20cc3f10df39384e90b9
Discriminator
import torch import torch.nn as nn class Discriminator(nn.Module): def __init__(self, n_in, n_out): super(Discriminator, self).__init__() self.f_k = nn.Bilinear(n_in, n_out, 1) self.sigm = nn.Sigmoid() for m in self.modules(): self.weights_init(m) def weights_init(self, m): if isinstance(m, nn.Bilinear): torch.nn.init.xavier_uniform_(m.weight.data) if m.bias is not None: m.bias.data.fill_(0.0) def forward(self, S, node, s_bias=None): S = S.expand_as(node) score = torch.squeeze(self.f_k(node, S), 1) if s_bias is not None: score += s_bias return self.sigm(score) def get_inputs(): return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'n_in': 4, 'n_out': 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 reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor @triton.jit def triton_poi_fused_add_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 = 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, (1, 4, 4), (16, 4, 1)) assert_size_stride(primals_4, (1,), (1,)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = torch.ops.aten._trilinear.default(reinterpret_tensor( primals_2, (64, 4), (4, 1), 0), primals_3, reinterpret_tensor( primals_1, (64, 4), (4, 1), 0), [1, 3], [0], [1, 2], [2, 3]) del primals_3 buf1 = buf0 del buf0 buf2 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 1), 0) del buf1 get_raw_stream(0) triton_poi_fused_add_sigmoid_0[grid(64)](buf2, primals_4, 64, XBLOCK=64, num_warps=1, num_stages=1) del primals_4 return buf2, reinterpret_tensor(primals_2, (64, 4), (4, 1), 0 ), reinterpret_tensor(primals_1, (64, 4), (4, 1), 0), buf2 class DiscriminatorNew(nn.Module): def __init__(self, n_in, n_out): super(DiscriminatorNew, self).__init__() self.f_k = nn.Bilinear(n_in, n_out, 1) self.sigm = nn.Sigmoid() for m in self.modules(): self.weights_init(m) def weights_init(self, m): if isinstance(m, nn.Bilinear): torch.nn.init.xavier_uniform_(m.weight.data) if m.bias is not None: m.bias.data.fill_(0.0) def forward(self, input_0, input_1): primals_3 = self.f_k.weight primals_4 = self.f_k.bias primals_1 = input_0 primals_2 = input_1 output = call([primals_1, primals_2, primals_3, primals_4]) return output[0]
caojiangxia/BiGI
Discriminator
false
14,996
[ "MIT" ]
57
ed54c20523a5b3f295b90a9c08f7c54e8258d04a
https://github.com/caojiangxia/BiGI/tree/ed54c20523a5b3f295b90a9c08f7c54e8258d04a
GCN
from torch.nn import Module import math import torch import torch.nn as nn from torch.nn.modules.module import Module class GraphConvolution(Module): def __init__(self, in_features, out_features, bias=True): super(GraphConvolution, self).__init__() self.in_features = in_features self.out_features = out_features self.weight = nn.Parameter(torch.FloatTensor(in_features, out_features) ) if bias: self.bias = nn.Parameter(torch.FloatTensor(out_features)) else: self.register_parameter('bias', None) self.reset_parameters() def reset_parameters(self): stdv = 1.0 / math.sqrt(self.weight.size(1)) self.weight.data.uniform_(-stdv, stdv) if self.bias is not None: self.bias.data.uniform_(-stdv, stdv) def forward(self, input, adj): support = torch.mm(input, self.weight) output = torch.spmm(adj, support) if self.bias is not None: return output + self.bias else: return output def __repr__(self): return self.__class__.__name__ + ' (' + str(self.in_features ) + ' -> ' + str(self.out_features) + ')' class GCN(nn.Module): def __init__(self, nfeat, nhid, dropout, alpha): super(GCN, self).__init__() self.gc1 = GraphConvolution(nfeat, nhid) self.dropout = dropout self.leakyrelu = nn.LeakyReLU(alpha) def forward(self, x, adj): x = self.leakyrelu(self.gc1(x, adj)) return x def get_inputs(): return [torch.rand([4, 4]), torch.rand([4, 4])] def get_init_inputs(): return [[], {'nfeat': 4, 'nhid': 4, 'dropout': 0.5, 'alpha': 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.nn import Module import math import torch.nn as nn from torch.nn.modules.module import Module 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_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 = 4.0 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 = 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, 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, 1), torch.float32) extern_kernels.mm(primals_2, primals_1, out=buf0) del primals_1 buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.mm(primals_3, buf0, out=buf1) buf2 = empty_strided_cuda((4, 4), (4, 1), torch.bool) buf3 = buf0 del buf0 get_raw_stream(0) triton_poi_fused_add_leaky_relu_0[grid(16)](buf1, primals_4, buf2, buf3, 16, XBLOCK=16, num_warps=1, num_stages=1) del buf1 del primals_4 return buf3, buf2, reinterpret_tensor(primals_3, (4, 4), (1, 4), 0 ), reinterpret_tensor(primals_2, (4, 4), (1, 4), 0) class GraphConvolution(Module): def __init__(self, in_features, out_features, bias=True): super(GraphConvolution, self).__init__() self.in_features = in_features self.out_features = out_features self.weight = nn.Parameter(torch.FloatTensor(in_features, out_features) ) if bias: self.bias = nn.Parameter(torch.FloatTensor(out_features)) else: self.register_parameter('bias', None) self.reset_parameters() def reset_parameters(self): stdv = 1.0 / math.sqrt(self.weight.size(1)) self.weight.data.uniform_(-stdv, stdv) if self.bias is not None: self.bias.data.uniform_(-stdv, stdv) def forward(self, input, adj): support = torch.mm(input, self.weight) output = torch.spmm(adj, support) if self.bias is not None: return output + self.bias else: return output def __repr__(self): return self.__class__.__name__ + ' (' + str(self.in_features ) + ' -> ' + str(self.out_features) + ')' class GCNNew(nn.Module): def __init__(self, nfeat, nhid, dropout, alpha): super(GCNNew, self).__init__() self.gc1 = GraphConvolution(nfeat, nhid) self.dropout = dropout self.leakyrelu = nn.LeakyReLU(alpha) def forward(self, input_0, input_1): primals_1 = self.gc1.weight primals_4 = self.gc1.bias primals_2 = input_0 primals_3 = input_1 output = call([primals_1, primals_2, primals_3, primals_4]) return output[0]
caojiangxia/BiGI
GCN
false
14,997
[ "MIT" ]
57
ed54c20523a5b3f295b90a9c08f7c54e8258d04a
https://github.com/caojiangxia/BiGI/tree/ed54c20523a5b3f295b90a9c08f7c54e8258d04a
Block
import torch import numpy as np from torch import nn import torch.nn.functional as F def drop_path(x, drop_prob: 'float'=0.0, training: 'bool'=False): 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 + torch.rand(shape, dtype=x.dtype, device=x. device) random_tensor.floor_() output = x.div(keep_prob) * random_tensor return output class GELU(nn.Module): def __init__(self): super(GELU, self).__init__() def forward(self, x): return 0.5 * x * (1 + F.tanh(np.sqrt(2 / np.pi) * (x + 0.044715 * torch.pow(x, 3)))) class DropPath(nn.Module): 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 Attention(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0): super().__init__() self.num_heads = num_heads self.scale = (dim // num_heads) ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads ).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn = q @ k.transpose(-2, -1) * self.scale attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class Mlp(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=GELU, drop=0.0): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features drop_probs = drop, drop self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() self.drop1 = nn.Dropout(drop_probs[0]) self.fc2 = nn.Linear(hidden_features, out_features) self.drop2 = nn.Dropout(drop_probs[1]) def forward(self, x): x = self.fc1(x) x = self.act(x) x = self.drop1(x) x = self.fc2(x) x = self.drop2(x) return x class Block(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4.0, qkv_bias=False, drop= 0.0, attn_drop=0.0, drop_path=0.0, act_layer=GELU, norm_layer=nn. LayerNorm): super().__init__() self.norm1 = norm_layer(dim) self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop) self.norm2 = norm_layer(dim) self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio ), act_layer=act_layer, drop=drop) self.drop_path = DropPath(drop_path ) if drop_path > 0.0 else nn.Identity() 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, 4])] def get_init_inputs(): return [[], {'dim': 4, 'num_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 numpy as np from torch import 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_native_layer_norm_0(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_1(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) @triton.jit def triton_poi_fused_clone_2(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 + 12 * x2 + 48 * y1), xmask & ymask, eviction_policy='evict_last') tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask) @triton.jit def triton_poi_fused_clone_3(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 + (4 + y0 + 12 * x2 + 48 * y1), xmask & ymask, eviction_policy='evict_last') tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask) @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) 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_5(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_6(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 + (8 + y0 + 12 * x2 + 48 * y1), xmask & ymask, eviction_policy='evict_last') tl.store(out_ptr0 + (x2 + 4 * y3), tmp0, xmask & ymask) @triton.jit def triton_poi_fused_clone_7(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_8(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_9(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 = 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_add_mul_pow_sqrt_tanh_10(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 = 0.5 tmp2 = tmp0 * tmp1 tmp3 = tmp0 * tmp0 tmp4 = tmp3 * tmp0 tmp5 = 0.044715 tmp6 = tmp4 * tmp5 tmp7 = tmp0 + tmp6 tmp8 = 0.7978845608028654 tmp9 = tmp8 * tmp7 tmp10 = libdevice.tanh(tmp9) tmp11 = 1.0 tmp12 = tmp10 + tmp11 tmp13 = tmp2 * tmp12 tl.store(out_ptr0 + x0, tmp13, xmask) @triton.jit def triton_poi_fused_add_11(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 x2 = xindex x0 = 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_out_ptr0 + x2, xmask) tmp6 = tl.load(in_ptr3 + x0, xmask, eviction_policy='evict_last') tmp3 = tmp1 + tmp2 tmp4 = tmp0 + tmp3 tmp7 = tmp5 + tmp6 tmp8 = tmp4 + tmp7 tl.store(in_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,), (1,)) assert_size_stride(primals_2, (4,), (1,)) assert_size_stride(primals_3, (4, 4, 4), (16, 4, 1)) assert_size_stride(primals_4, (12, 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,), (1,)) assert_size_stride(primals_8, (4,), (1,)) assert_size_stride(primals_9, (16, 4), (4, 1)) assert_size_stride(primals_10, (16,), (1,)) assert_size_stride(primals_11, (4, 16), (16, 1)) assert_size_stride(primals_12, (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_native_layer_norm_0[grid(16)](primals_3, 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_native_layer_norm_1[grid(64)](primals_3, buf0, buf1, primals_1, primals_2, buf2, 64, XBLOCK=64, num_warps=1, num_stages=1) del primals_1 del primals_2 buf3 = empty_strided_cuda((16, 12), (12, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(buf2, (16, 4), (4, 1), 0), reinterpret_tensor(primals_4, (4, 12), (1, 4), 0), out=buf3) buf4 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32) triton_poi_fused_clone_2[grid(16, 4)](buf3, buf4, 16, 4, XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1) buf5 = empty_strided_cuda((4, 4, 1, 4), (16, 4, 4, 1), torch.float32) triton_poi_fused_clone_3[grid(16, 4)](buf3, buf5, 16, 4, XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1) 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, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused__softmax_4[grid(256)](buf6, buf7, 256, XBLOCK=128, num_warps=4, num_stages=1) buf8 = reinterpret_tensor(buf6, (4, 4, 4, 4), (64, 16, 4, 1), 0) del buf6 triton_poi_fused__softmax_5[grid(256)](buf7, buf8, 256, XBLOCK=128, num_warps=4, num_stages=1) buf9 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32) triton_poi_fused_clone_6[grid(16, 4)](buf3, buf9, 16, 4, XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1) del buf3 buf10 = empty_strided_cuda((16, 4, 1), (4, 1, 1), torch.float32) extern_kernels.bmm(reinterpret_tensor(buf8, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf9, (16, 4, 1), (4, 1, 0), 0), out=buf10) buf11 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused_clone_7[grid(16, 4)](buf10, buf11, 16, 4, XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1) buf12 = reinterpret_tensor(buf10, (16, 4), (4, 1), 0) del buf10 extern_kernels.mm(reinterpret_tensor(buf11, (16, 4), (4, 1), 0), reinterpret_tensor(primals_5, (4, 4), (1, 4), 0), out=buf12) buf13 = buf1 del buf1 buf14 = buf0 del buf0 triton_poi_fused_add_native_layer_norm_8[grid(16)](primals_3, buf12, primals_6, buf13, buf14, 16, XBLOCK=16, num_warps=1, num_stages=1) buf15 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused_add_native_layer_norm_9[grid(64)](primals_3, buf12, primals_6, buf13, buf14, primals_7, primals_8, buf15, 64, XBLOCK=64, num_warps=1, num_stages=1) del buf13 del buf14 del primals_8 buf16 = reinterpret_tensor(buf7, (16, 16), (16, 1), 0) del buf7 extern_kernels.addmm(primals_10, reinterpret_tensor(buf15, (16, 4), (4, 1), 0), reinterpret_tensor(primals_9, (4, 16), (1, 4), 0), alpha=1, beta=1, out=buf16) del primals_10 buf17 = empty_strided_cuda((4, 4, 16), (64, 16, 1), torch.float32) triton_poi_fused_add_mul_pow_sqrt_tanh_10[grid(256)](buf16, buf17, 256, XBLOCK=256, num_warps=4, num_stages=1) buf18 = empty_strided_cuda((16, 4), (4, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(buf17, (16, 16), (16, 1), 0), reinterpret_tensor(primals_11, (16, 4), (1, 16), 0), out=buf18) buf19 = reinterpret_tensor(buf18, (4, 4, 4), (16, 4, 1), 0) del buf18 triton_poi_fused_add_11[grid(64)](buf19, primals_3, buf12, primals_6, primals_12, 64, XBLOCK=64, num_warps=1, num_stages=1) del primals_12 return buf19, primals_3, primals_6, primals_7, reinterpret_tensor(buf2, (16, 4), (4, 1), 0), buf8, reinterpret_tensor(buf11, (16, 4), (4, 1), 0 ), buf12, reinterpret_tensor(buf15, (16, 4), (4, 1), 0 ), buf16, reinterpret_tensor(buf17, (16, 16), (16, 1), 0 ), primals_11, primals_9, primals_5, reinterpret_tensor(buf9, (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_4 def drop_path(x, drop_prob: 'float'=0.0, training: 'bool'=False): 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 + torch.rand(shape, dtype=x.dtype, device=x. device) random_tensor.floor_() output = x.div(keep_prob) * random_tensor return output class GELU(nn.Module): def __init__(self): super(GELU, self).__init__() def forward(self, x): return 0.5 * x * (1 + F.tanh(np.sqrt(2 / np.pi) * (x + 0.044715 * torch.pow(x, 3)))) class DropPath(nn.Module): 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 Attention(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, attn_drop=0.0, proj_drop=0.0): super().__init__() self.num_heads = num_heads self.scale = (dim // num_heads) ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads ).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn = q @ k.transpose(-2, -1) * self.scale attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class Mlp(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=GELU, drop=0.0): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features drop_probs = drop, drop self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() self.drop1 = nn.Dropout(drop_probs[0]) self.fc2 = nn.Linear(hidden_features, out_features) self.drop2 = nn.Dropout(drop_probs[1]) def forward(self, x): x = self.fc1(x) x = self.act(x) x = self.drop1(x) x = self.fc2(x) x = self.drop2(x) return x class BlockNew(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4.0, qkv_bias=False, drop= 0.0, attn_drop=0.0, drop_path=0.0, act_layer=GELU, norm_layer=nn. LayerNorm): super().__init__() self.norm1 = norm_layer(dim) self.attn = Attention(dim, num_heads=num_heads, qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop) self.norm2 = norm_layer(dim) self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio ), act_layer=act_layer, drop=drop) self.drop_path = DropPath(drop_path ) if drop_path > 0.0 else nn.Identity() def forward(self, input_0): primals_1 = self.norm1.weight primals_2 = self.norm1.bias primals_4 = self.attn.qkv.weight primals_5 = self.attn.proj.weight primals_6 = self.attn.proj.bias primals_7 = self.norm2.weight primals_8 = self.norm2.bias primals_9 = self.mlp.fc1.weight primals_10 = self.mlp.fc1.bias primals_11 = self.mlp.fc2.weight primals_12 = 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, primals_12]) return output[0]
bubbliiiing/classification-pytorch
Block
false
14,998
[ "MIT" ]
88
ee62c05bd3094c3fab48bada5a57cb2ed8b61c11
https://github.com/bubbliiiing/classification-pytorch/tree/ee62c05bd3094c3fab48bada5a57cb2ed8b61c11
Ln_distance
import torch from torch import nn import torch.utils.data class Ln_distance(nn.Module): """If dims is None Compute across all dimensions except first""" def __init__(self, n, dim=None): super(Ln_distance, self).__init__() self.n = n self.dim = dim def forward(self, x, y): d = x - y if self.dim is None: self.dim = list(range(1, len(d.shape))) return torch.abs(d).pow(self.n).sum(dim=self.dim).pow(1.0 / float( self.n)) def get_inputs(): return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'n': 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, math as tl_math from torch import 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_pow_sub_sum_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, rnumel, XBLOCK: tl.constexpr): xnumel = 4 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) r1 = rindex x0 = xindex tmp0 = tl.load(in_ptr0 + (r1 + 64 * x0), xmask, other=0.0) tmp1 = tl.load(in_ptr1 + (r1 + 64 * x0), xmask, other=0.0) tmp2 = tmp0 - tmp1 tmp3 = tl_math.abs(tmp2) tmp4 = tmp3 * tmp3 tmp5 = tmp4 * tmp4 tmp6 = tl.broadcast_to(tmp5, [XBLOCK, RBLOCK]) tmp8 = tl.where(xmask, tmp6, 0) tmp9 = tl.sum(tmp8, 1)[:, None] tmp10 = 0.25 tmp11 = libdevice.pow(tmp9, tmp10) tl.debug_barrier() tl.store(in_out_ptr0 + x0, tmp11, xmask) 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,), (1,), torch.float32) buf1 = buf0 del buf0 get_raw_stream(0) triton_per_fused_abs_pow_sub_sum_0[grid(4)](buf1, arg0_1, arg1_1, 4, 64, XBLOCK=1, num_warps=2, num_stages=1) del arg0_1 del arg1_1 return buf1, class Ln_distanceNew(nn.Module): """If dims is None Compute across all dimensions except first""" def __init__(self, n, dim=None): super(Ln_distanceNew, self).__init__() self.n = n self.dim = dim def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0]
carla-recourse/CARLA
Ln_distance
false
15,000
[ "MIT" ]
140
e9bb3152598a94e700c38d7377825054959dcf48
https://github.com/carla-recourse/CARLA/tree/e9bb3152598a94e700c38d7377825054959dcf48
CombinedTargetMSELoss
import torch import torch.nn as nn class CombinedTargetMSELoss(nn.Module): """MSE loss for combined target. CombinedTarget: The combination of classification target (response map) and regression target (offset map). Paper ref: Huang et al. The Devil is in the Details: Delving into Unbiased Data Processing for Human Pose Estimation (CVPR 2020). Args: use_target_weight (bool): Option to use weighted MSE loss. Different joint types may have different target weights. loss_weight (float): Weight of the loss. Default: 1.0. """ def __init__(self, use_target_weight, loss_weight=1.0): super().__init__() self.criterion = nn.MSELoss(reduction='mean') self.use_target_weight = use_target_weight self.loss_weight = loss_weight def forward(self, output, target, target_weight): batch_size = output.size(0) num_channels = output.size(1) heatmaps_pred = output.reshape((batch_size, num_channels, -1)).split( 1, 1) heatmaps_gt = target.reshape((batch_size, num_channels, -1)).split(1, 1 ) loss = 0.0 num_joints = num_channels // 3 for idx in range(num_joints): heatmap_pred = heatmaps_pred[idx * 3].squeeze() heatmap_gt = heatmaps_gt[idx * 3].squeeze() offset_x_pred = heatmaps_pred[idx * 3 + 1].squeeze() offset_x_gt = heatmaps_gt[idx * 3 + 1].squeeze() offset_y_pred = heatmaps_pred[idx * 3 + 2].squeeze() offset_y_gt = heatmaps_gt[idx * 3 + 2].squeeze() if self.use_target_weight: heatmap_pred = heatmap_pred * target_weight[:, idx] heatmap_gt = heatmap_gt * target_weight[:, idx] loss += 0.5 * self.criterion(heatmap_pred, heatmap_gt) loss += 0.5 * self.criterion(heatmap_gt * offset_x_pred, heatmap_gt * offset_x_gt) loss += 0.5 * self.criterion(heatmap_gt * offset_y_pred, heatmap_gt * offset_y_gt) return loss / num_joints * self.loss_weight def get_inputs(): return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])] def get_init_inputs(): return [[], {'use_target_weight': 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_per_fused_add_div_mse_loss_mul_0(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr): RBLOCK: tl.constexpr = 4 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 + 4 * r0, None, eviction_policy='evict_last') tmp1 = tl.load(in_ptr1 + 4 * r0, None, eviction_policy='evict_last') tmp3 = tl.load(in_ptr2 + 4 * r0, None, eviction_policy='evict_last') tmp10 = tl.load(in_ptr0 + (1 + 4 * r0), None, eviction_policy='evict_last') tmp12 = tl.load(in_ptr2 + (1 + 4 * r0), None, eviction_policy='evict_last') tmp19 = tl.load(in_ptr0 + (2 + 4 * r0), None, eviction_policy='evict_last') tmp21 = tl.load(in_ptr2 + (2 + 4 * r0), None, eviction_policy='evict_last') tmp2 = tmp0 * tmp1 tmp4 = tmp3 * tmp1 tmp5 = tmp2 - tmp4 tmp6 = tmp5 * tmp5 tmp7 = tl.broadcast_to(tmp6, [XBLOCK, RBLOCK]) tmp9 = tl.sum(tmp7, 1)[:, None] tmp11 = tmp4 * tmp10 tmp13 = tmp4 * tmp12 tmp14 = tmp11 - tmp13 tmp15 = tmp14 * tmp14 tmp16 = tl.broadcast_to(tmp15, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp16, 1)[:, None] tmp20 = tmp4 * tmp19 tmp22 = tmp4 * tmp21 tmp23 = tmp20 - tmp22 tmp24 = tmp23 * tmp23 tmp25 = tl.broadcast_to(tmp24, [XBLOCK, RBLOCK]) tmp27 = tl.sum(tmp25, 1)[:, None] tmp28 = 4.0 tmp29 = tmp9 / tmp28 tmp30 = 0.5 tmp31 = tmp29 * tmp30 tmp32 = 0.0 tmp33 = tmp31 + tmp32 tmp34 = tmp18 / tmp28 tmp35 = tmp34 * tmp30 tmp36 = tmp33 + tmp35 tmp37 = tmp27 / tmp28 tmp38 = tmp37 * tmp30 tmp39 = tmp36 + tmp38 tmp40 = 1.0 tmp41 = tmp39 * tmp40 tmp42 = tmp41 * tmp40 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp42, None) def call(args): arg0_1, arg1_1, arg2_1 = args args.clear() assert_size_stride(arg0_1, (4, 4), (4, 1)) assert_size_stride(arg1_1, (4, 4), (4, 1)) assert_size_stride(arg2_1, (4, 4), (4, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((), (), torch.float32) buf3 = buf0 del buf0 get_raw_stream(0) triton_per_fused_add_div_mse_loss_mul_0[grid(1)](buf3, arg0_1, arg2_1, arg1_1, 1, 4, XBLOCK=1, num_warps=2, num_stages=1) del arg0_1 del arg1_1 del arg2_1 return buf3, class CombinedTargetMSELossNew(nn.Module): """MSE loss for combined target. CombinedTarget: The combination of classification target (response map) and regression target (offset map). Paper ref: Huang et al. The Devil is in the Details: Delving into Unbiased Data Processing for Human Pose Estimation (CVPR 2020). Args: use_target_weight (bool): Option to use weighted MSE loss. Different joint types may have different target weights. loss_weight (float): Weight of the loss. Default: 1.0. """ def __init__(self, use_target_weight, loss_weight=1.0): super().__init__() self.criterion = nn.MSELoss(reduction='mean') self.use_target_weight = use_target_weight self.loss_weight = loss_weight 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]
carolchenyx/mmpose
CombinedTargetMSELoss
false
15,001
[ "Apache-2.0" ]
367
cd74bf1d0b13954188cc678415fd0ef98a74b46b
https://github.com/carolchenyx/mmpose/tree/cd74bf1d0b13954188cc678415fd0ef98a74b46b
Square
import torch import torch.nn as nn class Square(nn.Module): def __init__(self): super(Square, self).__init__() def forward(self, x): return torch.mul(x, 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 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_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 = tmp0 * 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_mul_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256, num_warps=4, num_stages=1) del arg0_1 return buf0, class SquareNew(nn.Module): def __init__(self): super(SquareNew, self).__init__() def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0]
carlzhangweiwen/gazelle_mpc
Square
false
15,002
[ "MIT" ]
50
45818ccf6375100a8fe2680f44f37d713380aa5c
https://github.com/carlzhangweiwen/gazelle_mpc/tree/45818ccf6375100a8fe2680f44f37d713380aa5c
Attention
from _paritybench_helpers import _mock_config import math import torch import torch.nn as nn import torch.nn.functional as F class Attention(nn.Module): def __init__(self, opt): super(Attention, self).__init__() self.lin_u = nn.Linear(opt['hidden_dim'], opt['hidden_dim']) self.lin_v = nn.Linear(opt['hidden_dim'], opt['hidden_dim']) self.opt = opt def forward(self, user, item, UV_adj, VU_adj): user = self.lin_u(user) item = self.lin_v(item) query = user key = item value = torch.mm(query, key.transpose(0, 1)) value = UV_adj.to_dense() * value value /= math.sqrt(self.opt['hidden_dim']) value = F.softmax(value, dim=1) learn_user = torch.matmul(value, key) + user query = item key = user value = torch.mm(query, key.transpose(0, 1)) value = VU_adj.to_dense() * value value /= math.sqrt(self.opt['hidden_dim']) value = F.softmax(value, dim=1) learn_item = torch.matmul(value, key) + item return learn_user, learn_item def get_inputs(): return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])] def get_init_inputs(): return [[], {'opt': _mock_config(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 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__softmax_mul_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, 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 + 4 * x0, xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr1 + 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') tmp10 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp11 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp15 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp16 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp2 = tmp0 * tmp1 tmp3 = 1.0 tmp4 = tmp2 * tmp3 tmp7 = tmp5 * tmp6 tmp8 = tmp7 * tmp3 tmp9 = triton_helpers.maximum(tmp4, tmp8) tmp12 = tmp10 * tmp11 tmp13 = tmp12 * tmp3 tmp14 = triton_helpers.maximum(tmp9, tmp13) tmp17 = tmp15 * tmp16 tmp18 = tmp17 * tmp3 tmp19 = triton_helpers.maximum(tmp14, tmp18) tmp20 = tmp4 - tmp19 tmp21 = 0.5 tmp22 = tmp20 * tmp21 tmp23 = tl_math.exp(tmp22) tmp24 = tmp8 - tmp19 tmp25 = tmp24 * tmp21 tmp26 = tl_math.exp(tmp25) tmp27 = tmp23 + tmp26 tmp28 = tmp13 - tmp19 tmp29 = tmp28 * tmp21 tmp30 = tl_math.exp(tmp29) tmp31 = tmp27 + tmp30 tmp32 = tmp18 - tmp19 tmp33 = tmp32 * tmp21 tmp34 = tl_math.exp(tmp33) tmp35 = tmp31 + tmp34 tl.store(out_ptr0 + x0, tmp19, xmask) tl.store(out_ptr1 + x0, tmp35, xmask) @triton.jit def triton_poi_fused__softmax_mul_1(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, 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_out_ptr0 + x2, xmask) tmp5 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last') tmp10 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last') tmp2 = tmp0 * tmp1 tmp3 = 1.0 tmp4 = tmp2 * tmp3 tmp6 = tmp4 - tmp5 tmp7 = 0.5 tmp8 = tmp6 * tmp7 tmp9 = tl_math.exp(tmp8) tmp11 = tmp9 / tmp10 tl.store(in_out_ptr0 + x2, tmp11, 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, (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, 4), (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.mm(buf0, reinterpret_tensor(buf1, (4, 4), (1, 4), 0), out=buf2) buf3 = empty_strided_cuda((4, 1), (1, 4), torch.float32) buf4 = empty_strided_cuda((4, 1), (1, 4), torch.float32) get_raw_stream(0) triton_poi_fused__softmax_mul_0[grid(4)](primals_7, buf2, buf3, buf4, 4, XBLOCK=4, num_warps=1, num_stages=1) buf5 = buf2 del buf2 triton_poi_fused__softmax_mul_1[grid(16)](buf5, primals_7, buf3, buf4, 16, XBLOCK=16, num_warps=1, num_stages=1) buf6 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.addmm(buf0, buf5, buf1, alpha=1, beta=1, out=buf6) buf7 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.mm(buf1, reinterpret_tensor(buf0, (4, 4), (1, 4), 0), out=buf7) buf8 = buf4 del buf4 buf9 = buf3 del buf3 triton_poi_fused__softmax_mul_0[grid(4)](primals_8, buf7, buf8, buf9, 4, XBLOCK=4, num_warps=1, num_stages=1) buf10 = buf7 del buf7 triton_poi_fused__softmax_mul_1[grid(16)](buf10, primals_8, buf8, buf9, 16, XBLOCK=16, num_warps=1, num_stages=1) del buf8 del buf9 buf11 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.addmm(buf1, buf10, buf0, alpha=1, beta=1, out=buf11) return (buf6, buf11, primals_3, primals_6, primals_7, primals_8, buf0, buf1, buf5, buf10) class AttentionNew(nn.Module): def __init__(self, opt): super(AttentionNew, self).__init__() self.lin_u = nn.Linear(opt['hidden_dim'], opt['hidden_dim']) self.lin_v = nn.Linear(opt['hidden_dim'], opt['hidden_dim']) self.opt = opt def forward(self, input_0, input_1, input_2, input_3): primals_1 = self.lin_u.weight primals_2 = self.lin_u.bias primals_3 = self.lin_v.weight primals_5 = self.lin_v.bias primals_4 = input_0 primals_6 = input_1 primals_7 = input_2 primals_8 = input_3 output = call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8]) return output[0], output[1]
caojiangxia/BiGI
Attention
false
15,003
[ "MIT" ]
57
ed54c20523a5b3f295b90a9c08f7c54e8258d04a
https://github.com/caojiangxia/BiGI/tree/ed54c20523a5b3f295b90a9c08f7c54e8258d04a
SplitChannels
import torch class SplitChannels(torch.nn.Module): def __init__(self, split_location): super(SplitChannels, self).__init__() self.split_location = split_location def forward(self, x): a, b = x[:, :self.split_location], x[:, self.split_location:] a, b = a.clone(), b.clone() del x return a, b def inverse(self, x, y): return torch.cat([x, y], dim=1) def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'split_location': 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 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_clone_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) tl.store(out_ptr0 + x0, tmp0, 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_clone_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256, num_warps=4, num_stages=1) del arg0_1 buf1 = empty_strided_cuda((4, 0, 4, 4), (0, 16, 4, 1), torch.float32) return buf0, buf1 class SplitChannelsNew(torch.nn.Module): def __init__(self, split_location): super(SplitChannelsNew, self).__init__() self.split_location = split_location def inverse(self, x, y): return torch.cat([x, y], dim=1) def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0], output[1]
cetmann/iunets
SplitChannels
false
15,004
[ "MIT" ]
86
80ed7cce0e505a0396c42359eaf27819222d71f6
https://github.com/cetmann/iunets/tree/80ed7cce0e505a0396c42359eaf27819222d71f6
SoftTargetCrossEntropy
import torch import torch.nn as nn import torch.nn.functional as F import torch.distributed class SoftTargetCrossEntropy(nn.Module): def forward(self, x, target): loss = torch.sum(-target * F.log_softmax(x, dim=-1), dim=-1) return loss.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 math as tl_math import torch.nn as nn import torch.distributed 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__log_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) 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 tl.store(out_ptr0 + x2, tmp8, xmask) @triton.jit def triton_per_fused__log_softmax_mean_mul_neg_sum_1(in_out_ptr0, in_ptr0, in_ptr1, xnumel, rnumel, XBLOCK: tl.constexpr): RBLOCK: tl.constexpr = 64 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 + 4 * r0, None, eviction_policy='evict_last') tmp2 = tl.load(in_ptr1 + 4 * r0, None, eviction_policy='evict_last') tmp4 = tl.load(in_ptr1 + (1 + 4 * r0), None, eviction_policy='evict_last') tmp7 = tl.load(in_ptr1 + (2 + 4 * r0), None, eviction_policy='evict_last') tmp10 = tl.load(in_ptr1 + (3 + 4 * r0), None, eviction_policy='evict_last') tmp16 = tl.load(in_ptr0 + (1 + 4 * r0), None, eviction_policy='evict_last') tmp21 = tl.load(in_ptr0 + (2 + 4 * r0), None, eviction_policy='evict_last') tmp26 = tl.load(in_ptr0 + (3 + 4 * r0), None, eviction_policy='evict_last') tmp1 = -tmp0 tmp3 = tl_math.exp(tmp2) tmp5 = tl_math.exp(tmp4) tmp6 = tmp3 + tmp5 tmp8 = tl_math.exp(tmp7) tmp9 = tmp6 + tmp8 tmp11 = tl_math.exp(tmp10) tmp12 = tmp9 + tmp11 tmp13 = tl_math.log(tmp12) tmp14 = tmp2 - tmp13 tmp15 = tmp1 * tmp14 tmp17 = -tmp16 tmp18 = tmp4 - tmp13 tmp19 = tmp17 * tmp18 tmp20 = tmp15 + tmp19 tmp22 = -tmp21 tmp23 = tmp7 - tmp13 tmp24 = tmp22 * tmp23 tmp25 = tmp20 + tmp24 tmp27 = -tmp26 tmp28 = tmp10 - tmp13 tmp29 = tmp27 * tmp28 tmp30 = tmp25 + tmp29 tmp31 = tl.broadcast_to(tmp30, [XBLOCK, RBLOCK]) tmp33 = tl.sum(tmp31, 1)[:, None] tmp34 = 64.0 tmp35 = tmp33 / tmp34 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp35, 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, 4, 4), (64, 16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused__log_softmax_0[grid(256)](arg1_1, buf0, 256, XBLOCK=256, num_warps=4, num_stages=1) del arg1_1 buf2 = empty_strided_cuda((), (), torch.float32) buf3 = buf2 del buf2 triton_per_fused__log_softmax_mean_mul_neg_sum_1[grid(1)](buf3, arg0_1, buf0, 1, 64, XBLOCK=1, num_warps=2, num_stages=1) del arg0_1 del buf0 return buf3, class SoftTargetCrossEntropyNew(nn.Module): def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0]
ccjlovewsy/relabel_imagenet
SoftTargetCrossEntropy
false
15,005
[ "Apache-2.0" ]
344
6cd84dffe4ce8005395970b2938b3196d0958351
https://github.com/ccjlovewsy/relabel_imagenet/tree/6cd84dffe4ce8005395970b2938b3196d0958351
SmoothCrossEntropyLoss
import torch import torch.nn.functional as F from torch.nn.modules.loss import _WeightedLoss class SmoothCrossEntropyLoss(_WeightedLoss): def __init__(self, weight=None, reduction='mean', smoothing=0.0): super().__init__(weight=weight, reduction=reduction) self.smoothing = smoothing self.weight = weight self.reduction = reduction @staticmethod def _smooth_one_hot(targets: 'torch.Tensor', n_classes: 'int', smoothing=0.0): assert 0 <= smoothing < 1 with torch.no_grad(): targets = torch.empty(size=(targets.size(0), n_classes), device =targets.device).fill_(smoothing / (n_classes - 1)).scatter_( 1, targets.data.unsqueeze(1), 1.0 - smoothing) return targets def forward(self, inputs, targets): targets = SmoothCrossEntropyLoss._smooth_one_hot(targets, inputs. size(-1), self.smoothing) lsm = F.log_softmax(inputs, -1) if self.weight is not None: lsm = lsm * self.weight.unsqueeze(0) loss = -(targets * lsm).sum(-1) if self.reduction == 'sum': loss = loss.sum() elif self.reduction == 'mean': loss = loss.mean() return loss def get_inputs(): return [torch.rand([4, 4, 4, 4]), torch.ones([4], dtype=torch.int64)] 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 from torch.nn.modules.loss import _WeightedLoss 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__log_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) 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 tl.store(out_ptr0 + x2, tmp8, xmask) @triton.jit def triton_per_fused__log_softmax_mean_mul_neg_scatter_sum_1(in_out_ptr0, in_ptr0, in_ptr1, xnumel, rnumel, XBLOCK: tl.constexpr): RBLOCK: tl.constexpr = 64 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 % 4 r2 = rindex tmp0 = tl.load(in_ptr0 + r0, None, eviction_policy='evict_last') tmp6 = tl.load(in_ptr1 + 4 * r2, None, eviction_policy='evict_last') tmp8 = tl.load(in_ptr1 + (1 + 4 * r2), None, eviction_policy='evict_last') tmp11 = tl.load(in_ptr1 + (2 + 4 * r2), None, eviction_policy='evict_last') tmp14 = tl.load(in_ptr1 + (3 + 4 * r2), None, eviction_policy='evict_last') tmp1 = tl.full([1, 1], 0, tl.int64) tmp2 = tmp0 == tmp1 tmp3 = 1.0 tmp4 = 0.0 tmp5 = tl.where(tmp2, tmp3, tmp4) tmp7 = tl_math.exp(tmp6) tmp9 = tl_math.exp(tmp8) tmp10 = tmp7 + tmp9 tmp12 = tl_math.exp(tmp11) tmp13 = tmp10 + tmp12 tmp15 = tl_math.exp(tmp14) tmp16 = tmp13 + tmp15 tmp17 = tl_math.log(tmp16) tmp18 = tmp6 - tmp17 tmp19 = tmp5 * tmp18 tmp20 = tl.full([1, 1], 1, tl.int64) tmp21 = tmp0 == tmp20 tmp22 = tl.where(tmp21, tmp3, tmp4) tmp23 = tmp8 - tmp17 tmp24 = tmp22 * tmp23 tmp25 = tmp19 + tmp24 tmp26 = tl.full([1, 1], 2, tl.int64) tmp27 = tmp0 == tmp26 tmp28 = tl.where(tmp27, tmp3, tmp4) tmp29 = tmp11 - tmp17 tmp30 = tmp28 * tmp29 tmp31 = tmp25 + tmp30 tmp32 = tl.full([1, 1], 3, tl.int64) tmp33 = tmp0 == tmp32 tmp34 = tl.where(tmp33, tmp3, tmp4) tmp35 = tmp14 - tmp17 tmp36 = tmp34 * tmp35 tmp37 = tmp31 + tmp36 tmp38 = -tmp37 tmp39 = tl.broadcast_to(tmp38, [XBLOCK, RBLOCK]) tmp41 = tl.sum(tmp39, 1)[:, None] tmp42 = 64.0 tmp43 = tmp41 / tmp42 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp43, 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,), (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__log_softmax_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256, num_warps=4, num_stages=1) del arg0_1 buf2 = empty_strided_cuda((), (), torch.float32) buf3 = buf2 del buf2 triton_per_fused__log_softmax_mean_mul_neg_scatter_sum_1[grid(1)](buf3, arg1_1, buf0, 1, 64, XBLOCK=1, num_warps=2, num_stages=1) del arg1_1 del buf0 return buf3, class SmoothCrossEntropyLossNew(_WeightedLoss): def __init__(self, weight=None, reduction='mean', smoothing=0.0): super().__init__(weight=weight, reduction=reduction) self.smoothing = smoothing self.weight = weight self.reduction = reduction @staticmethod def _smooth_one_hot(targets: 'torch.Tensor', n_classes: 'int', smoothing=0.0): assert 0 <= smoothing < 1 with torch.no_grad(): targets = torch.empty(size=(targets.size(0), n_classes), device =targets.device).fill_(smoothing / (n_classes - 1)).scatter_( 1, targets.data.unsqueeze(1), 1.0 - smoothing) return targets def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0]
cclauss/archai
SmoothCrossEntropyLoss
false
15,006
[ "MIT" ]
344
a5fb8f937f7f1319e3204120803b2a045e9f768b
https://github.com/cclauss/archai/tree/a5fb8f937f7f1319e3204120803b2a045e9f768b
GAT
from _paritybench_helpers import _mock_config import math import torch import torch.nn as nn import torch.nn.functional as F class Attention(nn.Module): def __init__(self, opt): super(Attention, self).__init__() self.lin_u = nn.Linear(opt['hidden_dim'], opt['hidden_dim']) self.lin_v = nn.Linear(opt['hidden_dim'], opt['hidden_dim']) self.opt = opt def forward(self, user, item, UV_adj, VU_adj): user = self.lin_u(user) item = self.lin_v(item) query = user key = item value = torch.mm(query, key.transpose(0, 1)) value = UV_adj.to_dense() * value value /= math.sqrt(self.opt['hidden_dim']) value = F.softmax(value, dim=1) learn_user = torch.matmul(value, key) + user query = item key = user value = torch.mm(query, key.transpose(0, 1)) value = VU_adj.to_dense() * value value /= math.sqrt(self.opt['hidden_dim']) value = F.softmax(value, dim=1) learn_item = torch.matmul(value, key) + item return learn_user, learn_item class GAT(nn.Module): def __init__(self, opt): super(GAT, self).__init__() self.att = Attention(opt) self.dropout = opt['dropout'] self.leakyrelu = nn.LeakyReLU(opt['leakey']) def forward(self, ufea, vfea, UV_adj, VU_adj, adj=None): learn_user = ufea learn_item = vfea learn_user = F.dropout(learn_user, self.dropout, training=self.training ) learn_item = F.dropout(learn_item, self.dropout, training=self.training ) learn_user, learn_item = self.att(learn_user, learn_item, UV_adj, VU_adj) return learn_user, learn_item def get_inputs(): return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])] def get_init_inputs(): return [[], {'opt': _mock_config(hidden_dim=4, dropout=0.5, leakey=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__softmax_mul_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, 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 + 4 * x0, xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr1 + 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') tmp10 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp11 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp15 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp16 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp2 = tmp0 * tmp1 tmp3 = 1.0 tmp4 = tmp2 * tmp3 tmp7 = tmp5 * tmp6 tmp8 = tmp7 * tmp3 tmp9 = triton_helpers.maximum(tmp4, tmp8) tmp12 = tmp10 * tmp11 tmp13 = tmp12 * tmp3 tmp14 = triton_helpers.maximum(tmp9, tmp13) tmp17 = tmp15 * tmp16 tmp18 = tmp17 * tmp3 tmp19 = triton_helpers.maximum(tmp14, tmp18) tmp20 = tmp4 - tmp19 tmp21 = 0.5 tmp22 = tmp20 * tmp21 tmp23 = tl_math.exp(tmp22) tmp24 = tmp8 - tmp19 tmp25 = tmp24 * tmp21 tmp26 = tl_math.exp(tmp25) tmp27 = tmp23 + tmp26 tmp28 = tmp13 - tmp19 tmp29 = tmp28 * tmp21 tmp30 = tl_math.exp(tmp29) tmp31 = tmp27 + tmp30 tmp32 = tmp18 - tmp19 tmp33 = tmp32 * tmp21 tmp34 = tl_math.exp(tmp33) tmp35 = tmp31 + tmp34 tl.store(out_ptr0 + x0, tmp19, xmask) tl.store(out_ptr1 + x0, tmp35, xmask) @triton.jit def triton_poi_fused__softmax_mul_1(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, 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_out_ptr0 + x2, xmask) tmp5 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last') tmp10 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last') tmp2 = tmp0 * tmp1 tmp3 = 1.0 tmp4 = tmp2 * tmp3 tmp6 = tmp4 - tmp5 tmp7 = 0.5 tmp8 = tmp6 * tmp7 tmp9 = tl_math.exp(tmp8) tmp11 = tmp9 / tmp10 tl.store(in_out_ptr0 + x2, tmp11, 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, (4, 4), (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,), (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, 4), (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_4, primals_1, reinterpret_tensor( primals_3, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf0) del primals_3 del primals_4 buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.addmm(primals_6, primals_2, reinterpret_tensor( primals_5, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf1) del primals_5 del primals_6 buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.mm(buf0, reinterpret_tensor(buf1, (4, 4), (1, 4), 0), out=buf2) buf3 = empty_strided_cuda((4, 1), (1, 4), torch.float32) buf4 = empty_strided_cuda((4, 1), (1, 4), torch.float32) get_raw_stream(0) triton_poi_fused__softmax_mul_0[grid(4)](primals_7, buf2, buf3, buf4, 4, XBLOCK=4, num_warps=1, num_stages=1) buf5 = buf2 del buf2 triton_poi_fused__softmax_mul_1[grid(16)](buf5, primals_7, buf3, buf4, 16, XBLOCK=16, num_warps=1, num_stages=1) buf6 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.addmm(buf0, buf5, buf1, alpha=1, beta=1, out=buf6) buf7 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.mm(buf1, reinterpret_tensor(buf0, (4, 4), (1, 4), 0), out=buf7) buf8 = buf4 del buf4 buf9 = buf3 del buf3 triton_poi_fused__softmax_mul_0[grid(4)](primals_8, buf7, buf8, buf9, 4, XBLOCK=4, num_warps=1, num_stages=1) buf10 = buf7 del buf7 triton_poi_fused__softmax_mul_1[grid(16)](buf10, primals_8, buf8, buf9, 16, XBLOCK=16, num_warps=1, num_stages=1) del buf8 del buf9 buf11 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.addmm(buf1, buf10, buf0, alpha=1, beta=1, out=buf11) return (buf6, buf11, primals_1, primals_2, primals_7, primals_8, buf0, buf1, buf5, buf10) class Attention(nn.Module): def __init__(self, opt): super(Attention, self).__init__() self.lin_u = nn.Linear(opt['hidden_dim'], opt['hidden_dim']) self.lin_v = nn.Linear(opt['hidden_dim'], opt['hidden_dim']) self.opt = opt def forward(self, user, item, UV_adj, VU_adj): user = self.lin_u(user) item = self.lin_v(item) query = user key = item value = torch.mm(query, key.transpose(0, 1)) value = UV_adj.to_dense() * value value /= math.sqrt(self.opt['hidden_dim']) value = F.softmax(value, dim=1) learn_user = torch.matmul(value, key) + user query = item key = user value = torch.mm(query, key.transpose(0, 1)) value = VU_adj.to_dense() * value value /= math.sqrt(self.opt['hidden_dim']) value = F.softmax(value, dim=1) learn_item = torch.matmul(value, key) + item return learn_user, learn_item class GATNew(nn.Module): def __init__(self, opt): super(GATNew, self).__init__() self.att = Attention(opt) self.dropout = opt['dropout'] self.leakyrelu = nn.LeakyReLU(opt['leakey']) def forward(self, input_0, input_1, input_2, input_3): primals_1 = self.att.lin_u.weight primals_4 = self.att.lin_u.bias primals_2 = self.att.lin_v.weight primals_6 = self.att.lin_v.bias primals_3 = input_0 primals_5 = input_1 primals_7 = input_2 primals_8 = input_3 output = call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7, primals_8]) return output[0], output[1]
caojiangxia/BiGI
GAT
false
15,007
[ "MIT" ]
57
ed54c20523a5b3f295b90a9c08f7c54e8258d04a
https://github.com/caojiangxia/BiGI/tree/ed54c20523a5b3f295b90a9c08f7c54e8258d04a
LinfDistance
import torch from torch import nn import torch.autograd class LinfDistance(nn.Module): def forward(self, img1, img2): return (img1 - img2).reshape(img1.shape[0], -1).abs().max(dim=1)[0] 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 math as tl_math from torch import nn import torch.autograd 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_max_0(in_ptr0, in_ptr1, out_ptr0, xnumel, rnumel, XBLOCK: tl.constexpr): xnumel = 4 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) r1 = rindex x0 = xindex tmp0 = tl.load(in_ptr0 + (r1 + 64 * x0), xmask, other=0.0) tmp1 = tl.load(in_ptr1 + (r1 + 64 * x0), xmask, other=0.0) tmp2 = tmp0 - tmp1 tmp3 = tl_math.abs(tmp2) tmp4 = tl.broadcast_to(tmp3, [XBLOCK, RBLOCK]) tmp6 = tl.where(xmask, tmp4, float('-inf')) tmp7 = triton_helpers.max2(tmp6, 1)[:, None] tl.store(out_ptr0 + x0, tmp7, xmask) 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,), (1,), torch.float32) get_raw_stream(0) triton_per_fused_abs_max_0[grid(4)](arg0_1, arg1_1, buf0, 4, 64, XBLOCK=1, num_warps=2, num_stages=1) del arg0_1 del arg1_1 return buf0, class LinfDistanceNew(nn.Module): def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0]
cassidylaidlaw/perceptual-advex
LinfDistance
false
15,008
[ "MIT" ]
45
d39136eb5b5e950442456ddade6b4f4fba3dd8f6
https://github.com/cassidylaidlaw/perceptual-advex/tree/d39136eb5b5e950442456ddade6b4f4fba3dd8f6
ImageNetNormalizer
import torch from torch import nn import torch.autograd class ImageNetNormalizer(nn.Module): def __init__(self, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]): super().__init__() self.mean = mean self.std = std def forward(self, x): mean = torch.tensor(self.mean, device=x.device) std = torch.tensor(self.std, device=x.device) return (x - mean[None, :, None, None]) / std[None, :, None, None] def get_inputs(): return [torch.rand([4, 3, 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 import nn import torch.autograd 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_div_sub_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 x3 = xindex x1 = xindex // 16 % 3 tmp0 = tl.load(in_ptr0 + x3, xmask) tmp1 = x1 tmp2 = tl.full([1], 1, tl.int64) tmp3 = tmp1 < tmp2 tmp4 = tl.full([1], 2, tl.int64) tmp5 = tmp1 < tmp4 tmp6 = 0.4560000002384186 tmp7 = 0.4059999883174896 tmp8 = tl.where(tmp5, tmp6, tmp7) tmp9 = 0.48500001430511475 tmp10 = tl.where(tmp3, tmp9, tmp8) tmp11 = tmp0 - tmp10 tmp12 = 0.2240000069141388 tmp13 = 0.22499999403953552 tmp14 = tl.where(tmp5, tmp12, tmp13) tmp15 = 0.2290000021457672 tmp16 = tl.where(tmp3, tmp15, tmp14) tmp17 = tmp11 / tmp16 tl.store(out_ptr0 + x3, tmp17, xmask) def call(args): arg0_1, = args args.clear() assert_size_stride(arg0_1, (4, 3, 4, 4), (48, 16, 4, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 3, 4, 4), (48, 16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_div_sub_0[grid(192)](arg0_1, buf0, 192, XBLOCK=128, num_warps=4, num_stages=1) del arg0_1 return buf0, class ImageNetNormalizerNew(nn.Module): def __init__(self, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]): super().__init__() self.mean = mean self.std = std def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0]
cassidylaidlaw/perceptual-advex
ImageNetNormalizer
false
15,009
[ "MIT" ]
45
d39136eb5b5e950442456ddade6b4f4fba3dd8f6
https://github.com/cassidylaidlaw/perceptual-advex/tree/d39136eb5b5e950442456ddade6b4f4fba3dd8f6
L2Distance
import torch from torch import nn import torch.autograd class L2Distance(nn.Module): def forward(self, img1, img2): return (img1 - img2).reshape(img1.shape[0], -1).norm(dim=1) 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.triton_helpers import libdevice from torch import nn import torch.autograd 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_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, rnumel, XBLOCK: tl.constexpr): xnumel = 4 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) r1 = rindex x0 = xindex tmp0 = tl.load(in_ptr0 + (r1 + 64 * x0), xmask, other=0.0) tmp1 = tl.load(in_ptr1 + (r1 + 64 * x0), xmask, other=0.0) tmp2 = tmp0 - tmp1 tmp3 = tmp2 * tmp2 tmp4 = tl.broadcast_to(tmp3, [XBLOCK, RBLOCK]) tmp6 = tl.where(xmask, tmp4, 0) tmp7 = tl.sum(tmp6, 1)[:, None] tmp8 = libdevice.sqrt(tmp7) tl.debug_barrier() tl.store(in_out_ptr0 + x0, tmp8, xmask) 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,), (1,), torch.float32) buf1 = buf0 del buf0 get_raw_stream(0) triton_per_fused_linalg_vector_norm_0[grid(4)](buf1, arg0_1, arg1_1, 4, 64, XBLOCK=1, num_warps=2, num_stages=1) del arg0_1 del arg1_1 return buf1, class L2DistanceNew(nn.Module): def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0]
cassidylaidlaw/perceptual-advex
L2Distance
false
15,010
[ "MIT" ]
45
d39136eb5b5e950442456ddade6b4f4fba3dd8f6
https://github.com/cassidylaidlaw/perceptual-advex/tree/d39136eb5b5e950442456ddade6b4f4fba3dd8f6
GaussianConv2d
import torch import numpy as np import torch.nn.functional as F import torch.nn as nn import torch.utils.data from torch.nn.parameter import Parameter class GaussianConv2d(nn.Module): def __init__(self, in_channels, out_channels, ksize=5): """Applies 2-D Gaussian Blur. Args: in_channels: An integer indicates input channel dimension. out_channels: An integer indicates output channel dimension. ksize: An integer indicates Gaussian kernel size. """ super(GaussianConv2d, self).__init__() weight = (np.arange(ksize, dtype=np.float32) - ksize // 2) ** 2 weight = np.sqrt(weight[None, :] + weight[:, None]) weight = np.reshape(weight, (1, 1, ksize, ksize)) / weight.sum() self.weight = Parameter(torch.Tensor(weight).expand(out_channels, - 1, -1, -1)) self._in_channels = in_channels self._out_channels = out_channels def forward(self, x): with torch.no_grad(): return F.conv2d(x, self.weight, groups=self._in_channels) def get_inputs(): return [torch.rand([4, 4, 64, 64])] 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 import numpy as np import torch.nn as nn import torch.utils.data from torch.nn.parameter 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_0(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 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] tl.full([XBLOCK, YBLOCK], True, tl.int1) x2 = xindex y3 = yindex y0 = yindex % 4 y1 = yindex // 4 tmp0 = tl.load(in_ptr0 + (x2 + 4096 * y3), ymask, eviction_policy= 'evict_last') tl.store(out_ptr0 + (y0 + 4 * x2 + 16384 * y1), tmp0, ymask) @triton.jit def triton_poi_fused_convolution_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 16 xnumel = 3600 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 + 14400 * y1), xmask & ymask, eviction_policy='evict_last') tl.store(out_ptr0 + (x2 + 3600 * y3), tmp0, xmask & ymask) def call(args): arg0_1, arg1_1 = args args.clear() assert_size_stride(arg0_1, (4, 1, 5, 5), (25, 25, 5, 1)) assert_size_stride(arg1_1, (4, 4, 64, 64), (16384, 4096, 64, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 4, 64, 64), (16384, 1, 256, 4), torch .float32) get_raw_stream(0) triton_poi_fused_convolution_0[grid(16, 4096)](arg1_1, buf0, 16, 4096, XBLOCK=64, YBLOCK=16, num_warps=4, num_stages=1) del arg1_1 buf1 = extern_kernels.convolution(buf0, arg0_1, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=4, bias=None) assert_size_stride(buf1, (4, 4, 60, 60), (14400, 1, 240, 4)) del arg0_1 del buf0 buf2 = empty_strided_cuda((4, 4, 60, 60), (14400, 3600, 60, 1), torch.float32) triton_poi_fused_convolution_1[grid(16, 3600)](buf1, buf2, 16, 3600, XBLOCK=64, YBLOCK=16, num_warps=4, num_stages=1) del buf1 return buf2, class GaussianConv2dNew(nn.Module): def __init__(self, in_channels, out_channels, ksize=5): """Applies 2-D Gaussian Blur. Args: in_channels: An integer indicates input channel dimension. out_channels: An integer indicates output channel dimension. ksize: An integer indicates Gaussian kernel size. """ super(GaussianConv2dNew, self).__init__() weight = (np.arange(ksize, dtype=np.float32) - ksize // 2) ** 2 weight = np.sqrt(weight[None, :] + weight[:, None]) weight = np.reshape(weight, (1, 1, ksize, ksize)) / weight.sum() self.weight = Parameter(torch.Tensor(weight).expand(out_channels, - 1, -1, -1)) self._in_channels = in_channels self._out_channels = out_channels def forward(self, input_0): arg0_1 = self.weight arg1_1 = input_0 output = call([arg0_1, arg1_1]) return output[0]
cenkbircanoglu/SPML
GaussianConv2d
false
15,011
[ "MIT" ]
68
f09e4c30ecf2030d42ac70b2c35e7fdeee9bf468
https://github.com/cenkbircanoglu/SPML/tree/f09e4c30ecf2030d42ac70b2c35e7fdeee9bf468
ContrastiveLoss
import torch from torch import nn from torch.nn import functional as F from torch.utils.data import * from torch.distributions import * import torch.nn.parallel import torch.optim import torch.utils.data import torch.utils.data.distributed class ContrastiveLoss(nn.Module): """ Contrastive loss Takes embeddings of two samples and a target label == 1 if samples are from the same class and label == 0 otherwise """ def __init__(self, margin): super(ContrastiveLoss, self).__init__() self.margin = margin self.eps = 1e-09 def forward(self, output1, output2, target, size_average=True): distances = (output2 - output1).pow(2).sum(1) losses = 0.5 * (target.float() * distances + (1 + -1 * target). float() * F.relu(self.margin - (distances + self.eps).sqrt()). pow(2)) return losses.mean() if size_average else losses.sum() 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 [[], {'margin': 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 from torch import nn from torch.utils.data import * from torch.distributions import * 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 @triton.jit def triton_poi_fused_pow_sub_sum_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 % 16 x1 = xindex // 16 x2 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask) tmp1 = tl.load(in_ptr1 + (x0 + 64 * x1), xmask) tmp4 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask) tmp5 = tl.load(in_ptr1 + (16 + x0 + 64 * x1), xmask) tmp9 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask) tmp10 = tl.load(in_ptr1 + (32 + x0 + 64 * x1), xmask) tmp14 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask) tmp15 = tl.load(in_ptr1 + (48 + x0 + 64 * x1), xmask) tmp2 = tmp0 - tmp1 tmp3 = tmp2 * tmp2 tmp6 = tmp4 - tmp5 tmp7 = tmp6 * tmp6 tmp8 = tmp3 + tmp7 tmp11 = tmp9 - tmp10 tmp12 = tmp11 * tmp11 tmp13 = tmp8 + tmp12 tmp16 = tmp14 - tmp15 tmp17 = tmp16 * tmp16 tmp18 = tmp13 + tmp17 tl.store(out_ptr0 + x2, tmp18, xmask) @triton.jit def triton_per_fused_add_mean_mul_pow_relu_rsub_sqrt_1(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) r2 = rindex r0 = rindex % 64 tmp0 = tl.load(in_ptr0 + r2, None) tmp1 = tl.load(in_ptr1 + r0, None, eviction_policy='evict_last') tmp2 = tmp0 * tmp1 tmp3 = -1.0 tmp4 = tmp0 * tmp3 tmp5 = 1.0 tmp6 = tmp4 + tmp5 tmp7 = 1e-09 tmp8 = tmp1 + tmp7 tmp9 = libdevice.sqrt(tmp8) tmp10 = 4.0 tmp11 = tmp10 - tmp9 tmp12 = tl.full([1], 0, tl.int32) tmp13 = triton_helpers.maximum(tmp12, tmp11) tmp14 = tmp13 * tmp13 tmp15 = tmp6 * tmp14 tmp16 = tmp2 + tmp15 tmp17 = 0.5 tmp18 = tmp16 * tmp17 tmp19 = tl.broadcast_to(tmp18, [RBLOCK]) tmp21 = triton_helpers.promote_to_tensor(tl.sum(tmp19, 0)) tmp22 = 256.0 tmp23 = tmp21 / tmp22 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp23, None) 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((4, 4, 4), (16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_pow_sub_sum_0[grid(64)](arg0_1, arg1_1, buf0, 64, XBLOCK=64, num_warps=1, num_stages=1) del arg0_1 del arg1_1 buf1 = empty_strided_cuda((), (), torch.float32) buf2 = buf1 del buf1 triton_per_fused_add_mean_mul_pow_relu_rsub_sqrt_1[grid(1)](buf2, arg2_1, buf0, 1, 256, num_warps=2, num_stages=1) del arg2_1 del buf0 return buf2, class ContrastiveLossNew(nn.Module): """ Contrastive loss Takes embeddings of two samples and a target label == 1 if samples are from the same class and label == 0 otherwise """ def __init__(self, margin): super(ContrastiveLossNew, self).__init__() self.margin = margin self.eps = 1e-09 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]
cgsas/LOB
ContrastiveLoss
false
15,012
[ "MIT" ]
97
4175912194c2a066b2d7df038a376484b57ed76c
https://github.com/cgsas/LOB/tree/4175912194c2a066b2d7df038a376484b57ed76c
CombinedTargetMSELoss
import torch import torch.nn as nn class CombinedTargetMSELoss(nn.Module): """MSE loss for combined target. CombinedTarget: The combination of classification target (response map) and regression target (offset map). Paper ref: Huang et al. The Devil is in the Details: Delving into Unbiased Data Processing for Human Pose Estimation (CVPR 2020). Args: use_target_weight (bool): Option to use weighted MSE loss. Different joint types may have different target weights. """ def __init__(self, use_target_weight): super().__init__() self.criterion = nn.MSELoss(reduction='mean') self.use_target_weight = use_target_weight def forward(self, output, target, target_weight): batch_size = output.size(0) num_channels = output.size(1) heatmaps_pred = output.reshape((batch_size, num_channels, -1)).split( 1, 1) heatmaps_gt = target.reshape((batch_size, num_channels, -1)).split(1, 1 ) loss = 0.0 num_joints = num_channels // 3 for idx in range(num_joints): heatmap_pred = heatmaps_pred[idx * 3].squeeze() heatmap_gt = heatmaps_gt[idx * 3].squeeze() offset_x_pred = heatmaps_pred[idx * 3 + 1].squeeze() offset_x_gt = heatmaps_gt[idx * 3 + 1].squeeze() offset_y_pred = heatmaps_pred[idx * 3 + 2].squeeze() offset_y_gt = heatmaps_gt[idx * 3 + 2].squeeze() if self.use_target_weight: heatmap_pred = heatmap_pred.mul(target_weight[:, idx]) heatmap_gt = heatmap_gt.mul(target_weight[:, idx]) loss += 0.5 * self.criterion(heatmap_pred, heatmap_gt) loss += 0.5 * self.criterion(heatmap_gt * offset_x_pred, heatmap_gt * offset_x_gt) loss += 0.5 * self.criterion(heatmap_gt * offset_y_pred, heatmap_gt * offset_y_gt) return loss / num_joints def get_inputs(): return [torch.rand([4, 4]), torch.rand([4, 4]), torch.rand([4, 4])] def get_init_inputs(): return [[], {'use_target_weight': 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_per_fused_add_div_mse_loss_mul_0(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr): RBLOCK: tl.constexpr = 4 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 + 4 * r0, None, eviction_policy='evict_last') tmp1 = tl.load(in_ptr1 + 4 * r0, None, eviction_policy='evict_last') tmp3 = tl.load(in_ptr2 + 4 * r0, None, eviction_policy='evict_last') tmp10 = tl.load(in_ptr0 + (1 + 4 * r0), None, eviction_policy='evict_last') tmp12 = tl.load(in_ptr2 + (1 + 4 * r0), None, eviction_policy='evict_last') tmp19 = tl.load(in_ptr0 + (2 + 4 * r0), None, eviction_policy='evict_last') tmp21 = tl.load(in_ptr2 + (2 + 4 * r0), None, eviction_policy='evict_last') tmp2 = tmp0 * tmp1 tmp4 = tmp3 * tmp1 tmp5 = tmp2 - tmp4 tmp6 = tmp5 * tmp5 tmp7 = tl.broadcast_to(tmp6, [XBLOCK, RBLOCK]) tmp9 = tl.sum(tmp7, 1)[:, None] tmp11 = tmp4 * tmp10 tmp13 = tmp4 * tmp12 tmp14 = tmp11 - tmp13 tmp15 = tmp14 * tmp14 tmp16 = tl.broadcast_to(tmp15, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp16, 1)[:, None] tmp20 = tmp4 * tmp19 tmp22 = tmp4 * tmp21 tmp23 = tmp20 - tmp22 tmp24 = tmp23 * tmp23 tmp25 = tl.broadcast_to(tmp24, [XBLOCK, RBLOCK]) tmp27 = tl.sum(tmp25, 1)[:, None] tmp28 = 4.0 tmp29 = tmp9 / tmp28 tmp30 = 0.5 tmp31 = tmp29 * tmp30 tmp32 = 0.0 tmp33 = tmp31 + tmp32 tmp34 = tmp18 / tmp28 tmp35 = tmp34 * tmp30 tmp36 = tmp33 + tmp35 tmp37 = tmp27 / tmp28 tmp38 = tmp37 * tmp30 tmp39 = tmp36 + tmp38 tmp40 = 1.0 tmp41 = tmp39 * tmp40 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp41, None) def call(args): arg0_1, arg1_1, arg2_1 = args args.clear() assert_size_stride(arg0_1, (4, 4), (4, 1)) assert_size_stride(arg1_1, (4, 4), (4, 1)) assert_size_stride(arg2_1, (4, 4), (4, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((), (), torch.float32) buf3 = buf0 del buf0 get_raw_stream(0) triton_per_fused_add_div_mse_loss_mul_0[grid(1)](buf3, arg0_1, arg2_1, arg1_1, 1, 4, XBLOCK=1, num_warps=2, num_stages=1) del arg0_1 del arg1_1 del arg2_1 return buf3, class CombinedTargetMSELossNew(nn.Module): """MSE loss for combined target. CombinedTarget: The combination of classification target (response map) and regression target (offset map). Paper ref: Huang et al. The Devil is in the Details: Delving into Unbiased Data Processing for Human Pose Estimation (CVPR 2020). Args: use_target_weight (bool): Option to use weighted MSE loss. Different joint types may have different target weights. """ def __init__(self, use_target_weight): super().__init__() self.criterion = nn.MSELoss(reduction='mean') self.use_target_weight = use_target_weight 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]
chaowentao/mmpose
CombinedTargetMSELoss
false
15,013
[ "Apache-2.0" ]
367
b528c60ef4fab56d35d1ed7e187023794639be26
https://github.com/chaowentao/mmpose/tree/b528c60ef4fab56d35d1ed7e187023794639be26
MoEHead
import math import torch from torch.nn import functional as F from torch.autograd import Variable from torch import nn def softmax(x): if x.dim() == 3: return F.softmax(x.transpose(0, 2)).transpose(0, 2) return F.softmax(x) def gumbel_softmax(input, beta=0.5, tau=1.0): noise = input.data.new(*input.size()).uniform_() noise.add_(TINY).log_().neg_().add_(TINY).log_().neg_() return softmax((input + beta * Variable(noise)) / tau) def matmul(x, y): if x.dim() == y.dim(): return x @ y if x.dim() == y.dim() - 1: return (x.unsqueeze(-2) @ y).squeeze(-2) return (x @ y.unsqueeze(-1)).squeeze(-1) def mask(targets, out, input_mask=None, return_mask=False): if input_mask is None: input_mask = targets != 1 out_mask = input_mask.unsqueeze(-1).expand_as(out) if return_mask: return targets[input_mask], out[out_mask].view(-1, out.size(-1) ), the_mask return targets[input_mask], out[out_mask].view(-1, out.size(-1)) class Linear(nn.Linear): def forward(self, x): size = x.size() return super().forward(x.contiguous().view(-1, size[-1])).view(* size[:-1], -1) class Attention(nn.Module): def __init__(self, d_key, drop_ratio, causal, diag=False, window=-1, noisy=False): super().__init__() self.scale = math.sqrt(d_key) self.dropout = nn.Dropout(drop_ratio) self.causal = causal self.diag = diag self.window = window self.noisy = noisy def forward(self, query, key, value=None, mask=None, feedback=None, beta=0, tau=1, weights=None): dot_products = matmul(query, key.transpose(1, 2)) if weights is not None: dot_products = dot_products + weights if query.dim() == 3 and self.causal and query.size(1) == key.size(1): tri = key.data.new(key.size(1), key.size(1)).fill_(1).triu(1) * INF dot_products.data.sub_(tri.unsqueeze(0)) if self.window > 0: window_mask = key.data.new(key.size(1), key.size(1)).fill_(1) window_mask = (window_mask.triu(self.window + 1) + window_mask. tril(-self.window - 1)) * INF dot_products.data.sub_(window_mask.unsqueeze(0)) if self.diag: inds = torch.arange(0, key.size(1)).long().view(1, 1, -1) if key.is_cuda: inds = inds dot_products.data.scatter_(1, inds.expand(dot_products.size(0), 1, inds.size(-1)), -INF) if mask is not None: if dot_products.dim() == 2: dot_products.data -= (1 - mask) * INF else: dot_products.data -= (1 - mask[:, None, :]) * INF if value is None: return dot_products logits = dot_products / self.scale if not self.noisy: probs = softmax(logits) else: probs = gumbel_softmax(logits, beta=beta, tau=tau) if feedback is not None: feedback.append(probs.contiguous()) return matmul(self.dropout(probs), value) class MoEHead(nn.Module): def __init__(self, d_key, d_value, n_heads, drop_ratio, causal=False, diag=False, window=-1, noisy=False, use_wo=True): super().__init__() self.attention = Attention(d_key, drop_ratio, causal=causal, diag= diag, window=window, noisy=noisy) self.wq = Linear(d_key, d_key, bias=use_wo) self.wk = Linear(d_key, d_key, bias=use_wo) self.wv = Linear(d_value, d_value, bias=use_wo) self.wo = Linear(d_value, d_key, bias=use_wo) self.gate = Linear(d_value // n_heads, 1) self.use_wo = use_wo self.n_heads = n_heads def forward(self, query, key, inputs, mask=None, feedback=None, weights =None, beta=0, tau=1): query, key, value = self.wq(query), self.wk(key), self.wv(inputs) B, Tq, D = query.size() _, Tk, _ = key.size() N = self.n_heads probs = [] query, key, value = (x.contiguous().view(B, -1, N, D // N). transpose(2, 1).contiguous().view(B * N, -1, D // N) for x in ( query, key, value)) if mask is not None: mask = mask[:, None, :].expand(B, N, Tk).contiguous().view(B * N, -1) probs = self.attention(query, key, None, mask, probs, beta, tau, weights) mix = matmul(self.attention.dropout(probs), value).contiguous().view(B, N, -1, D // N).transpose(2, 1).contiguous() mix = softmax(self.gate(mix)) probs = (probs.contiguous().view(B, N, Tq, Tk).transpose(2, 1) * mix ).sum(-2) outputs = matmul(probs, inputs) return self.wo(outputs) 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_key': 4, 'd_value': 4, 'n_heads': 4, 'drop_ratio': 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 math as tl_math import math from torch.nn import functional as F from torch.autograd import Variable 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_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_clone_1(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_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 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__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 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 = tmp1 + tmp2 tmp5 = tmp3 + tmp4 tmp7 = tmp5 + tmp6 tmp8 = tmp0 / tmp7 tl.store(out_ptr0 + x3, tmp8, xmask) @triton.jit def triton_poi_fused__softmax_mul_sum_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 x2 = xindex // 16 x3 = xindex % 16 x4 = xindex // 4 x5 = xindex tmp0 = tl.load(in_ptr0 + (x3 + 64 * x2), xmask) tmp1 = tl.load(in_ptr1 + 4 * x4, xmask, eviction_policy='evict_last') tmp3 = tl.load(in_ptr0 + (16 + x3 + 64 * x2), xmask) tmp4 = tl.load(in_ptr1 + (1 + 4 * x4), xmask, eviction_policy='evict_last') tmp7 = tl.load(in_ptr0 + (32 + x3 + 64 * x2), xmask) tmp8 = tl.load(in_ptr1 + (2 + 4 * x4), xmask, eviction_policy='evict_last') tmp11 = tl.load(in_ptr0 + (48 + x3 + 64 * x2), xmask) tmp12 = tl.load(in_ptr1 + (3 + 4 * x4), 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 + x5, tmp14, 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, (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), (16, 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), (16, 4, 1)) assert_size_stride(primals_8, (4, 4), (4, 1)) assert_size_stride(primals_9, (4,), (1,)) assert_size_stride(primals_10, (1, 1), (1, 1)) assert_size_stride(primals_11, (1,), (1,)) assert_size_stride(primals_12, (4, 4), (4, 1)) assert_size_stride(primals_13, (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_4, (16, 4), (4, 1), 0), reinterpret_tensor(primals_5, (4, 4), (1, 4), 0), out=buf1) del primals_5 buf2 = empty_strided_cuda((16, 4), (4, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(primals_7, (16, 4), (4, 1), 0), reinterpret_tensor(primals_8, (4, 4), (1, 4), 0), out=buf2) del primals_8 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, 4, 1), (16, 4, 1, 1), 0) del buf0 triton_poi_fused_clone_0[grid(16, 4)](buf1, primals_6, buf4, 16, 4, XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1) del primals_6 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, 1), 0) del buf1 triton_poi_fused_clone_0[grid(16, 4)](buf2, primals_9, buf6, 16, 4, XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1) del primals_9 buf7 = reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 1), 0) del buf2 extern_kernels.bmm(buf5, reinterpret_tensor(buf6, (16, 4, 1), (4, 1, 0), 0), out=buf7) buf8 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 1), torch.float32) triton_poi_fused_clone_1[grid(16, 4)](buf7, buf8, 16, 4, XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1) buf10 = reinterpret_tensor(buf7, (64, 1), (1, 1), 0) del buf7 extern_kernels.addmm(primals_11, reinterpret_tensor(buf8, (64, 1), (1, 0), 0), primals_10, alpha=1, beta=1, out=buf10) del primals_11 buf11 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32) triton_poi_fused__softmax_2[grid(64)](buf10, buf11, 64, XBLOCK=64, num_warps=1, num_stages=1) buf12 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32) triton_poi_fused__softmax_3[grid(64)](buf11, buf12, 64, XBLOCK=64, num_warps=1, num_stages=1) buf13 = reinterpret_tensor(buf11, (4, 4, 4), (16, 4, 1), 0) del buf11 triton_poi_fused__softmax_mul_sum_4[grid(64)](buf5, buf12, buf13, 64, XBLOCK=64, num_warps=1, num_stages=1) buf14 = reinterpret_tensor(buf12, (4, 4, 4), (16, 4, 1), 0) del buf12 extern_kernels.bmm(buf13, primals_7, out=buf14) buf15 = reinterpret_tensor(buf13, (16, 4), (4, 1), 0) del buf13 extern_kernels.addmm(primals_13, reinterpret_tensor(buf14, (16, 4), (4, 1), 0), reinterpret_tensor(primals_12, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf15) del primals_13 return reinterpret_tensor(buf15, (4, 4, 4), (16, 4, 1), 0 ), primals_7, reinterpret_tensor(primals_1, (16, 4), (4, 1), 0 ), reinterpret_tensor(primals_4, (16, 4), (4, 1), 0 ), buf5, reinterpret_tensor(buf8, (64, 1), (1, 1), 0 ), buf10, reinterpret_tensor(buf14, (16, 4), (4, 1), 0 ), primals_12, primals_10, reinterpret_tensor(buf6, (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, 1), 0) def softmax(x): if x.dim() == 3: return F.softmax(x.transpose(0, 2)).transpose(0, 2) return F.softmax(x) def gumbel_softmax(input, beta=0.5, tau=1.0): noise = input.data.new(*input.size()).uniform_() noise.add_(TINY).log_().neg_().add_(TINY).log_().neg_() return softmax((input + beta * Variable(noise)) / tau) def matmul(x, y): if x.dim() == y.dim(): return x @ y if x.dim() == y.dim() - 1: return (x.unsqueeze(-2) @ y).squeeze(-2) return (x @ y.unsqueeze(-1)).squeeze(-1) def mask(targets, out, input_mask=None, return_mask=False): if input_mask is None: input_mask = targets != 1 out_mask = input_mask.unsqueeze(-1).expand_as(out) if return_mask: return targets[input_mask], out[out_mask].view(-1, out.size(-1) ), the_mask return targets[input_mask], out[out_mask].view(-1, out.size(-1)) class Linear(nn.Linear): def forward(self, x): size = x.size() return super().forward(x.contiguous().view(-1, size[-1])).view(* size[:-1], -1) class Attention(nn.Module): def __init__(self, d_key, drop_ratio, causal, diag=False, window=-1, noisy=False): super().__init__() self.scale = math.sqrt(d_key) self.dropout = nn.Dropout(drop_ratio) self.causal = causal self.diag = diag self.window = window self.noisy = noisy def forward(self, query, key, value=None, mask=None, feedback=None, beta=0, tau=1, weights=None): dot_products = matmul(query, key.transpose(1, 2)) if weights is not None: dot_products = dot_products + weights if query.dim() == 3 and self.causal and query.size(1) == key.size(1): tri = key.data.new(key.size(1), key.size(1)).fill_(1).triu(1) * INF dot_products.data.sub_(tri.unsqueeze(0)) if self.window > 0: window_mask = key.data.new(key.size(1), key.size(1)).fill_(1) window_mask = (window_mask.triu(self.window + 1) + window_mask. tril(-self.window - 1)) * INF dot_products.data.sub_(window_mask.unsqueeze(0)) if self.diag: inds = torch.arange(0, key.size(1)).long().view(1, 1, -1) if key.is_cuda: inds = inds dot_products.data.scatter_(1, inds.expand(dot_products.size(0), 1, inds.size(-1)), -INF) if mask is not None: if dot_products.dim() == 2: dot_products.data -= (1 - mask) * INF else: dot_products.data -= (1 - mask[:, None, :]) * INF if value is None: return dot_products logits = dot_products / self.scale if not self.noisy: probs = softmax(logits) else: probs = gumbel_softmax(logits, beta=beta, tau=tau) if feedback is not None: feedback.append(probs.contiguous()) return matmul(self.dropout(probs), value) class MoEHeadNew(nn.Module): def __init__(self, d_key, d_value, n_heads, drop_ratio, causal=False, diag=False, window=-1, noisy=False, use_wo=True): super().__init__() self.attention = Attention(d_key, drop_ratio, causal=causal, diag= diag, window=window, noisy=noisy) self.wq = Linear(d_key, d_key, bias=use_wo) self.wk = Linear(d_key, d_key, bias=use_wo) self.wv = Linear(d_value, d_value, bias=use_wo) self.wo = Linear(d_value, d_key, bias=use_wo) self.gate = Linear(d_value // n_heads, 1) self.use_wo = use_wo self.n_heads = n_heads def forward(self, input_0, input_1, input_2): primals_2 = self.wq.weight primals_3 = self.wq.bias primals_5 = self.wk.weight primals_6 = self.wk.bias primals_8 = self.wv.weight primals_9 = self.wv.bias primals_12 = self.wo.weight primals_13 = self.wo.bias primals_10 = self.gate.weight primals_11 = self.gate.bias primals_1 = input_0 primals_4 = input_1 primals_7 = 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, primals_12, primals_13]) return output[0]
cclauss/nonauto-nmt
MoEHead
false
15,014
[ "BSD-3-Clause" ]
262
efcbe4f2329b140ac3ce06abb6409457cebc8e49
https://github.com/cclauss/nonauto-nmt/tree/efcbe4f2329b140ac3ce06abb6409457cebc8e49
InvertibleChannelMixing1D
from torch.autograd import Function import torch from torch import nn from warnings import warn def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalChannelMixing(nn.Module): """Base class for all orthogonal channel mixing layers. """ def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(OrthogonalChannelMixing, self).__init__() self.in_channels = in_channels self.weight = nn.Parameter(torch.zeros((in_channels, in_channels)), requires_grad=learnable) assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self.kwargs = kwargs @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel_matrix_transposed(self): """The orthogonal matrix created by the chosen parametrisation method. """ return torch.transpose(self.kernel_matrix, -1, -2) class InvertibleChannelMixing1D(OrthogonalChannelMixing): """Orthogonal (and hence invertible) channel mixing layer for 1D data. This layer linearly combines the input channels to each output channel. Here, the number of output channels is the same as the number of input channels, and the matrix specifying the connectivity between the channels is orthogonal. :param in_channels: The number of input (and output) channels. :param method: The chosen method for parametrizing the orthogonal matrix which determines the orthogonal channel mixing. Either ``"exp"``, ``"cayley"`` or ``"householder"``. """ def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(InvertibleChannelMixing1D, self).__init__(in_channels= in_channels, method=method, learnable=learnable, **kwargs) self.kwargs = kwargs @property def kernel(self): return self.kernel_matrix.view(self.in_channels, self.in_channels, 1) def forward(self, x): return nn.functional.conv1d(x, self.kernel) def inverse(self, x): return nn.functional.conv_transpose1d(x, self.kernel) def get_inputs(): return [torch.rand([4, 4])] def get_init_inputs(): return [[], {'in_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.autograd import Function from torch import nn from warnings import warn 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_eye_sub_0(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 4 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 x1 = xindex tmp6 = tl.load(in_ptr0 + (x1 + 4 * y0), xmask & ymask) tmp7 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask) tmp0 = y0 tmp1 = x1 tmp2 = tmp0 == tmp1 tmp3 = 1.0 tmp4 = 0.0 tmp5 = tl.where(tmp2, tmp3, tmp4) tmp8 = tmp6 - tmp7 tmp9 = tmp5 + tmp8 tmp10 = tmp5 - tmp8 tl.store(out_ptr0 + (x1 + 4 * y0), tmp9, xmask & ymask) tl.store(out_ptr1 + (x1 + 4 * y0), tmp10, xmask & ymask) @triton.jit def triton_poi_fused_convolution_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 4 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 x1 = xindex y0 = yindex tmp0 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask) tl.store(out_ptr0 + (x1 + 4 * y0), 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, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32) buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_eye_sub_0[grid(4, 4)](primals_1, buf0, buf5, 4, 4, XBLOCK=4, YBLOCK=4, num_warps=1, num_stages=1) del primals_1 buf1 = torch.ops.aten.linalg_lu_factor_ex.default(buf0) buf2 = buf1[0] buf3 = buf1[1] del buf1 buf6 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf5) buf7 = buf6 del buf6 buf8 = reinterpret_tensor(buf0, (4, 4, 1), (4, 1, 1), 0) del buf0 triton_poi_fused_convolution_1[grid(4, 4)](buf7, buf8, 4, 4, XBLOCK =4, YBLOCK=4, num_warps=1, num_stages=1) buf9 = extern_kernels.convolution(reinterpret_tensor(primals_2, (1, 4, 4), (16, 4, 1), 0), buf8, stride=(1,), padding=(0,), dilation=(1,), transposed=False, output_padding=(0,), groups=1, bias=None) assert_size_stride(buf9, (1, 4, 4), (16, 4, 1)) del buf8 buf10 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf7) del buf2 del buf3 buf11 = buf10 del buf10 return reinterpret_tensor(buf9, (4, 4), (4, 1), 0 ), buf5, reinterpret_tensor(buf7, (4, 4, 1), (1, 4, 4), 0 ), reinterpret_tensor(primals_2, (1, 4, 4), (16, 4, 1), 0), buf11 def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalChannelMixing(nn.Module): """Base class for all orthogonal channel mixing layers. """ def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(OrthogonalChannelMixing, self).__init__() self.in_channels = in_channels self.weight = nn.Parameter(torch.zeros((in_channels, in_channels)), requires_grad=learnable) assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self.kwargs = kwargs @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel_matrix_transposed(self): """The orthogonal matrix created by the chosen parametrisation method. """ return torch.transpose(self.kernel_matrix, -1, -2) class InvertibleChannelMixing1DNew(OrthogonalChannelMixing): """Orthogonal (and hence invertible) channel mixing layer for 1D data. This layer linearly combines the input channels to each output channel. Here, the number of output channels is the same as the number of input channels, and the matrix specifying the connectivity between the channels is orthogonal. :param in_channels: The number of input (and output) channels. :param method: The chosen method for parametrizing the orthogonal matrix which determines the orthogonal channel mixing. Either ``"exp"``, ``"cayley"`` or ``"householder"``. """ def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(InvertibleChannelMixing1DNew, self).__init__(in_channels= in_channels, method=method, learnable=learnable, **kwargs) self.kwargs = kwargs @property def kernel(self): return self.kernel_matrix.view(self.in_channels, self.in_channels, 1) def inverse(self, x): return nn.functional.conv_transpose1d(x, self.kernel) def forward(self, input_0): primals_1 = self.weight primals_2 = input_0 output = call([primals_1, primals_2]) return output[0]
cetmann/iunets
InvertibleChannelMixing1D
false
15,015
[ "MIT" ]
86
80ed7cce0e505a0396c42359eaf27819222d71f6
https://github.com/cetmann/iunets/tree/80ed7cce0e505a0396c42359eaf27819222d71f6
InvertibleChannelMixing3D
from torch.autograd import Function import torch from torch import nn from warnings import warn def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalChannelMixing(nn.Module): """Base class for all orthogonal channel mixing layers. """ def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(OrthogonalChannelMixing, self).__init__() self.in_channels = in_channels self.weight = nn.Parameter(torch.zeros((in_channels, in_channels)), requires_grad=learnable) assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self.kwargs = kwargs @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel_matrix_transposed(self): """The orthogonal matrix created by the chosen parametrisation method. """ return torch.transpose(self.kernel_matrix, -1, -2) class InvertibleChannelMixing3D(OrthogonalChannelMixing): def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(InvertibleChannelMixing3D, self).__init__(in_channels= in_channels, method=method, learnable=learnable, **kwargs) self.kwargs = kwargs @property def kernel(self): return self.kernel_matrix.view(self.in_channels, self.in_channels, 1, 1, 1) def forward(self, x): return nn.functional.conv3d(x, self.kernel) def inverse(self, x): return nn.functional.conv_transpose3d(x, self.kernel) def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_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.autograd import Function from torch import nn from warnings import warn 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_eye_sub_0(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 4 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 x1 = xindex tmp6 = tl.load(in_ptr0 + (x1 + 4 * y0), xmask & ymask) tmp7 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask) tmp0 = y0 tmp1 = x1 tmp2 = tmp0 == tmp1 tmp3 = 1.0 tmp4 = 0.0 tmp5 = tl.where(tmp2, tmp3, tmp4) tmp8 = tmp6 - tmp7 tmp9 = tmp5 + tmp8 tmp10 = tmp5 - tmp8 tl.store(out_ptr0 + (x1 + 4 * y0), tmp9, xmask & ymask) tl.store(out_ptr1 + (x1 + 4 * y0), tmp10, xmask & ymask) @triton.jit def triton_poi_fused_convolution_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 4 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 x1 = xindex y0 = yindex tmp0 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask) tl.store(out_ptr0 + (x1 + 4 * y0), 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, 1), torch.float32) buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_eye_sub_0[grid(4, 4)](primals_1, buf0, buf5, 4, 4, XBLOCK=4, YBLOCK=4, num_warps=1, num_stages=1) del primals_1 buf1 = torch.ops.aten.linalg_lu_factor_ex.default(buf0) buf2 = buf1[0] buf3 = buf1[1] del buf1 buf6 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf5) buf7 = buf6 del buf6 buf8 = reinterpret_tensor(buf0, (4, 4, 1, 1, 1), (4, 1, 1, 1, 1), 0) del buf0 triton_poi_fused_convolution_1[grid(4, 4)](buf7, buf8, 4, 4, XBLOCK =4, YBLOCK=4, num_warps=1, num_stages=1) buf9 = extern_kernels.convolution(reinterpret_tensor(primals_2, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1), 0), buf8, stride=(1, 1, 1), padding=(0, 0, 0), dilation=(1, 1, 1), transposed=False, output_padding=(0, 0, 0), groups=1, bias=None) assert_size_stride(buf9, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1)) del buf8 buf10 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf7) del buf2 del buf3 buf11 = buf10 del buf10 return reinterpret_tensor(buf9, (4, 4, 4, 4), (64, 16, 4, 1), 0 ), buf5, reinterpret_tensor(buf7, (4, 4, 1, 1, 1), (1, 4, 4, 4, 4), 0 ), reinterpret_tensor(primals_2, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1), 0), buf11 def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalChannelMixing(nn.Module): """Base class for all orthogonal channel mixing layers. """ def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(OrthogonalChannelMixing, self).__init__() self.in_channels = in_channels self.weight = nn.Parameter(torch.zeros((in_channels, in_channels)), requires_grad=learnable) assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self.kwargs = kwargs @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel_matrix_transposed(self): """The orthogonal matrix created by the chosen parametrisation method. """ return torch.transpose(self.kernel_matrix, -1, -2) class InvertibleChannelMixing3DNew(OrthogonalChannelMixing): def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(InvertibleChannelMixing3DNew, self).__init__(in_channels= in_channels, method=method, learnable=learnable, **kwargs) self.kwargs = kwargs @property def kernel(self): return self.kernel_matrix.view(self.in_channels, self.in_channels, 1, 1, 1) def inverse(self, x): return nn.functional.conv_transpose3d(x, self.kernel) def forward(self, input_0): primals_1 = self.weight primals_2 = input_0 output = call([primals_1, primals_2]) return output[0]
cetmann/iunets
InvertibleChannelMixing3D
false
15,016
[ "MIT" ]
86
80ed7cce0e505a0396c42359eaf27819222d71f6
https://github.com/cetmann/iunets/tree/80ed7cce0e505a0396c42359eaf27819222d71f6
InvertibleChannelMixing2D
from torch.autograd import Function import torch from torch import nn from warnings import warn def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalChannelMixing(nn.Module): """Base class for all orthogonal channel mixing layers. """ def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(OrthogonalChannelMixing, self).__init__() self.in_channels = in_channels self.weight = nn.Parameter(torch.zeros((in_channels, in_channels)), requires_grad=learnable) assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self.kwargs = kwargs @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel_matrix_transposed(self): """The orthogonal matrix created by the chosen parametrisation method. """ return torch.transpose(self.kernel_matrix, -1, -2) class InvertibleChannelMixing2D(OrthogonalChannelMixing): def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(InvertibleChannelMixing2D, self).__init__(in_channels= in_channels, method=method, learnable=learnable, **kwargs) self.kwargs = kwargs @property def kernel(self): return self.kernel_matrix.view(self.in_channels, self.in_channels, 1, 1 ) def forward(self, x): return nn.functional.conv2d(x, self.kernel) def inverse(self, x): return nn.functional.conv_transpose2d(x, self.kernel) def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_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.autograd import Function from torch import nn from warnings import warn 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_eye_sub_0(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 4 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 x1 = xindex tmp6 = tl.load(in_ptr0 + (x1 + 4 * y0), xmask & ymask) tmp7 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask) tmp0 = y0 tmp1 = x1 tmp2 = tmp0 == tmp1 tmp3 = 1.0 tmp4 = 0.0 tmp5 = tl.where(tmp2, tmp3, tmp4) tmp8 = tmp6 - tmp7 tmp9 = tmp5 + tmp8 tmp10 = tmp5 - tmp8 tl.store(out_ptr0 + (x1 + 4 * y0), tmp9, xmask & ymask) tl.store(out_ptr1 + (x1 + 4 * y0), tmp10, xmask & ymask) @triton.jit def triton_poi_fused_convolution_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 4 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 x1 = xindex y0 = yindex tmp0 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask) tl.store(out_ptr0 + (x1 + 4 * y0), 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, 1), torch.float32) buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_eye_sub_0[grid(4, 4)](primals_1, buf0, buf5, 4, 4, XBLOCK=4, YBLOCK=4, num_warps=1, num_stages=1) del primals_1 buf1 = torch.ops.aten.linalg_lu_factor_ex.default(buf0) buf2 = buf1[0] buf3 = buf1[1] del buf1 buf6 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf5) buf7 = buf6 del buf6 buf8 = reinterpret_tensor(buf0, (4, 4, 1, 1), (4, 1, 1, 1), 0) del buf0 triton_poi_fused_convolution_1[grid(4, 4)](buf7, buf8, 4, 4, XBLOCK =4, YBLOCK=4, num_warps=1, num_stages=1) buf9 = extern_kernels.convolution(primals_2, buf8, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf9, (4, 4, 4, 4), (64, 16, 4, 1)) del buf8 buf10 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf7) del buf2 del buf3 buf11 = buf10 del buf10 return buf9, primals_2, buf5, reinterpret_tensor(buf7, (4, 4, 1, 1), (1, 4, 4, 4), 0), buf11 def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalChannelMixing(nn.Module): """Base class for all orthogonal channel mixing layers. """ def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(OrthogonalChannelMixing, self).__init__() self.in_channels = in_channels self.weight = nn.Parameter(torch.zeros((in_channels, in_channels)), requires_grad=learnable) assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self.kwargs = kwargs @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel_matrix_transposed(self): """The orthogonal matrix created by the chosen parametrisation method. """ return torch.transpose(self.kernel_matrix, -1, -2) class InvertibleChannelMixing2DNew(OrthogonalChannelMixing): def __init__(self, in_channels: 'int', method: 'str'='cayley', learnable: 'bool'=True, **kwargs): super(InvertibleChannelMixing2DNew, self).__init__(in_channels= in_channels, method=method, learnable=learnable, **kwargs) self.kwargs = kwargs @property def kernel(self): return self.kernel_matrix.view(self.in_channels, self.in_channels, 1, 1 ) def inverse(self, x): return nn.functional.conv_transpose2d(x, self.kernel) def forward(self, input_0): primals_1 = self.weight primals_2 = input_0 output = call([primals_1, primals_2]) return output[0]
cetmann/iunets
InvertibleChannelMixing2D
false
15,017
[ "MIT" ]
86
80ed7cce0e505a0396c42359eaf27819222d71f6
https://github.com/cetmann/iunets/tree/80ed7cce0e505a0396c42359eaf27819222d71f6
homo_Gauss_mloglike
import torch import numpy as np import torch.nn.parallel import torch.utils.data import torch.utils.data.distributed import torch.nn as nn import torch.optim from torch.distributions import Normal class homo_Gauss_mloglike(nn.Module): def __init__(self, Ndims=1, sig=None): super(homo_Gauss_mloglike, self).__init__() if sig is None: self.log_std = nn.Parameter(torch.zeros(Ndims)) else: self.log_std = nn.Parameter(torch.ones(Ndims) * np.log(sig), requires_grad=False) def forward(self, mu, y, model_std=None): sig = self.log_std.exp().clamp(min=0.0001) if model_std is not None: sig = (sig ** 2 + model_std ** 2) ** 0.5 dist = Normal(mu, sig) return -dist.log_prob(y) 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 math as tl_math import numpy as np import torch.nn.parallel import torch.utils.data import torch.utils.data.distributed import torch.nn as nn import torch.optim 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_div_log_mul_neg_pow_sub_0(in_ptr0, in_ptr1, in_ptr2, 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 = tl.load(in_ptr1 + x0, xmask) tmp5 = tl.load(in_ptr2 + 0) tmp6 = tl.broadcast_to(tmp5, [XBLOCK]) tmp2 = tmp0 - tmp1 tmp3 = tmp2 * tmp2 tmp4 = -tmp3 tmp7 = tl_math.exp(tmp6) tmp8 = 0.0001 tmp9 = triton_helpers.maximum(tmp7, tmp8) tmp10 = tmp9 * tmp9 tmp11 = 2.0 tmp12 = tmp10 * tmp11 tmp13 = tmp4 / tmp12 tmp14 = tl_math.log(tmp9) tmp15 = tmp13 - tmp14 tmp16 = 0.9189385332046727 tmp17 = tmp15 - tmp16 tmp18 = -tmp17 tmp19 = tmp13 / tmp12 tl.store(out_ptr0 + x0, tmp18, xmask) tl.store(out_ptr1 + x0, tmp19, xmask) def call(args): primals_1, primals_2, primals_3 = args args.clear() assert_size_stride(primals_1, (1,), (1,)) assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 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((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_div_log_mul_neg_pow_sub_0[grid(256)](primals_3, primals_2, primals_1, buf0, buf1, 256, XBLOCK=128, num_warps=4, num_stages=1) del primals_2 del primals_3 return buf0, primals_1, buf1 class homo_Gauss_mloglikeNew(nn.Module): def __init__(self, Ndims=1, sig=None): super(homo_Gauss_mloglikeNew, self).__init__() if sig is None: self.log_std = nn.Parameter(torch.zeros(Ndims)) else: self.log_std = nn.Parameter(torch.ones(Ndims) * np.log(sig), requires_grad=False) def forward(self, input_0, input_1): primals_1 = self.log_std primals_2 = input_0 primals_3 = input_1 output = call([primals_1, primals_2, primals_3]) return output[0]
chelsealuisa/DUN
homo_Gauss_mloglike
false
15,018
[ "MIT" ]
58
1ccd9bc49b91b13089350f003a25bdb11003d843
https://github.com/chelsealuisa/DUN/tree/1ccd9bc49b91b13089350f003a25bdb11003d843
ContrastiveLoss
import torch import torchvision.transforms.functional as F import torch.nn.functional as F import torch.nn as nn class ContrastiveLoss(nn.Module): """ Contrastive loss function. Based on: http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf """ def __init__(self, margin=2.0): super(ContrastiveLoss, self).__init__() self.margin = margin def forward(self, output1, output2, label): euclidean_distance = F.pairwise_distance(output1, output2) loss_contrastive = torch.mean(label * torch.pow(euclidean_distance, 2) + (1 - label) * torch.pow(torch.clamp(self.margin - euclidean_distance, min=0.0), 2)) return loss_contrastive 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 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 @triton.jit def triton_poi_fused_add_norm_sub_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 tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last') 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') tmp12 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp13 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp18 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp19 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp2 = tmp0 - tmp1 tmp3 = 1e-06 tmp4 = tmp2 + tmp3 tmp5 = tmp4 * tmp4 tmp8 = tmp6 - tmp7 tmp9 = tmp8 + tmp3 tmp10 = tmp9 * tmp9 tmp11 = tmp5 + tmp10 tmp14 = tmp12 - tmp13 tmp15 = tmp14 + tmp3 tmp16 = tmp15 * tmp15 tmp17 = tmp11 + tmp16 tmp20 = tmp18 - tmp19 tmp21 = tmp20 + tmp3 tmp22 = tmp21 * tmp21 tmp23 = tmp17 + tmp22 tmp24 = libdevice.sqrt(tmp23) tl.store(out_ptr0 + x0, tmp24, xmask) @triton.jit def triton_per_fused_add_clamp_mean_mul_pow_rsub_1(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) r2 = rindex r0 = rindex % 64 tmp0 = tl.load(in_ptr0 + r2, None) tmp1 = tl.load(in_ptr1 + r0, None, eviction_policy='evict_last') tmp2 = tmp1 * tmp1 tmp3 = tmp0 * tmp2 tmp4 = 1.0 tmp5 = tmp4 - tmp0 tmp6 = 2.0 tmp7 = tmp6 - tmp1 tmp8 = 0.0 tmp9 = triton_helpers.maximum(tmp7, tmp8) tmp10 = tmp9 * tmp9 tmp11 = tmp5 * tmp10 tmp12 = tmp3 + tmp11 tmp13 = tl.broadcast_to(tmp12, [RBLOCK]) tmp15 = triton_helpers.promote_to_tensor(tl.sum(tmp13, 0)) tmp16 = 256.0 tmp17 = tmp15 / tmp16 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp17, None) 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((4, 4, 4), (16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_norm_sub_0[grid(64)](arg1_1, arg0_1, buf0, 64, XBLOCK=64, num_warps=1, num_stages=1) del arg0_1 del arg1_1 buf1 = empty_strided_cuda((), (), torch.float32) buf2 = buf1 del buf1 triton_per_fused_add_clamp_mean_mul_pow_rsub_1[grid(1)](buf2, arg2_1, buf0, 1, 256, num_warps=2, num_stages=1) del arg2_1 del buf0 return buf2, class ContrastiveLossNew(nn.Module): """ Contrastive loss function. Based on: http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf """ def __init__(self, margin=2.0): super(ContrastiveLossNew, self).__init__() self.margin = margin 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]
chenyanghungry/person-reid-lib
ContrastiveLoss
false
15,019
[ "MIT" ]
81
783e66c9bfedf582e2cf935b9f5be960b543ac3c
https://github.com/chenyanghungry/person-reid-lib/tree/783e66c9bfedf582e2cf935b9f5be960b543ac3c
MLP
import torch from torch import nn import torch.utils.data class MLP(nn.Module): def __init__(self, input_size, output_size, hidden_size=None, dropout=0.1): super().__init__() if hidden_size is None: hidden_size = input_size * 4 self.w_1 = nn.Linear(input_size * 2, hidden_size) self.w_2 = nn.Linear(hidden_size, output_size) self.dropout = nn.Dropout(dropout) def forward(self, x1, x2): x = torch.cat([x1, x2], -1) return self.w_2(self.dropout(torch.nn.functional.relu(self.w_1(x)))) def get_inputs(): return [torch.rand([4, 4, 4, 4]), 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 import triton_helpers from torch import 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_poi_fused_cat_0(in_ptr0, in_ptr1, 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 = 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_relu_threshold_backward_1(in_out_ptr0, in_ptr0, out_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 % 16 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) 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, (16, 8), (8, 1)) assert_size_stride(primals_4, (16,), (1,)) assert_size_stride(primals_5, (4, 16), (16, 1)) assert_size_stride(primals_6, (4,), (1,)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 4, 4, 8), (128, 32, 8, 1), torch.float32) get_raw_stream(0) triton_poi_fused_cat_0[grid(512)](primals_1, primals_2, buf0, 512, XBLOCK=256, num_warps=4, num_stages=1) del primals_1 del primals_2 buf1 = empty_strided_cuda((64, 16), (16, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(buf0, (64, 8), (8, 1), 0), reinterpret_tensor(primals_3, (8, 16), (1, 8), 0), out=buf1) del primals_3 buf2 = reinterpret_tensor(buf1, (4, 4, 4, 16), (256, 64, 16, 1), 0) del buf1 buf4 = empty_strided_cuda((4, 4, 4, 16), (256, 64, 16, 1), torch.bool) triton_poi_fused_relu_threshold_backward_1[grid(1024)](buf2, primals_4, buf4, 1024, XBLOCK=256, num_warps=4, num_stages=1) del primals_4 buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32) extern_kernels.addmm(primals_6, reinterpret_tensor(buf2, (64, 16), (16, 1), 0), reinterpret_tensor(primals_5, (16, 4), (1, 16), 0), alpha=1, beta=1, out=buf3) del primals_6 return reinterpret_tensor(buf3, (4, 4, 4, 4), (64, 16, 4, 1), 0 ), reinterpret_tensor(buf0, (64, 8), (8, 1), 0), reinterpret_tensor( buf2, (64, 16), (16, 1), 0), primals_5, buf4 class MLPNew(nn.Module): def __init__(self, input_size, output_size, hidden_size=None, dropout=0.1): super().__init__() if hidden_size is None: hidden_size = input_size * 4 self.w_1 = nn.Linear(input_size * 2, hidden_size) self.w_2 = nn.Linear(hidden_size, output_size) self.dropout = nn.Dropout(dropout) def forward(self, input_0, input_1): primals_3 = self.w_1.weight primals_4 = self.w_1.bias primals_5 = self.w_2.weight primals_6 = self.w_2.bias primals_1 = input_0 primals_2 = input_1 output = call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6]) return output[0]
chenyangh/tensor2struct-public
MLP
false
15,020
[ "MIT" ]
69
d3257cba6d76d3c658a58a78f687d986bdc755cf
https://github.com/chenyangh/tensor2struct-public/tree/d3257cba6d76d3c658a58a78f687d986bdc755cf
InvertibleDownsampling2D
from torch.autograd import Function import torch import numpy as np from warnings import warn from typing import Union from typing import Tuple from torch.nn.common_types import _size_2_t from torch.nn.modules.utils import _pair import torch.nn.functional as F def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) def __initialize_weight__(kernel_matrix_shape: 'Tuple[int, ...]', stride: 'Tuple[int, ...]', method: 'str'='cayley', init: 'str'='haar', dtype: 'str'='float32', *args, **kwargs): """Function which computes specific orthogonal matrices. For some chosen method of parametrizing orthogonal matrices, this function outputs the required weights necessary to represent a chosen initialization as a Pytorch tensor of matrices. Args: kernel_matrix_shape : The output shape of the orthogonal matrices. Should be (num_matrices, height, width). stride : The stride for the invertible up- or downsampling for which this matrix is to be used. The length of ``stride`` should match the dimensionality of the data. method : The method for parametrising orthogonal matrices. Should be 'exp' or 'cayley' init : The matrix which should be represented. Should be 'squeeze', 'pixel_shuffle', 'haar' or 'random'. 'haar' is only possible if ``stride`` is only 2. dtype : Numpy dtype which should be used for the matrix. *args: Variable length argument iterable. **kwargs: Arbitrary keyword arguments. Returns: Tensor : Orthogonal matrices of shape ``kernel_matrix_shape``. """ dim = len(stride) num_matrices = kernel_matrix_shape[0] assert method in ['exp', 'cayley', 'householder'] if method == 'householder': warn( 'Householder parametrization not fully implemented yet. Only random initialization currently working.' ) init = 'random' if init == 'random': return torch.randn(kernel_matrix_shape) if init == 'haar' and set(stride) != {2}: None None init = 'squeeze' if init == 'haar' and set(stride) == {2}: if method == 'exp': p = np.pi / 4 if dim == 1: weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: weight = np.array([[[0, p, p, 0, p, 0, 0, 0], [0, 0, 0, p, 0, p, 0, 0], [0, 0, 0, p, 0, 0, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, p, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) elif method == 'cayley': if dim == 1: p = -np.sqrt(2) / (2 - np.sqrt(2)) weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: p = 0.5 weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: p = 1 / np.sqrt(2) weight = np.array([[[0, -p, -p, 0, -p, 0, 0, 1 - p], [0, 0, 0, -p, 0, -p, p - 1, 0], [0, 0, 0, -p, 0, p - 1, -p, 0], [0, 0, 0, 0, 1 - p, 0, 0, -p], [0, 0, 0, 0, 0, -p, -p, 0], [0, 0, 0, 0, 0, 0, 0, -p], [0, 0, 0, 0, 0, 0, 0, -p ], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) if init in ['squeeze', 'pixel_shuffle', 'zeros']: if method == 'exp' or method == 'cayley': return torch.zeros(*kernel_matrix_shape) if type(init) is np.ndarray: init = torch.tensor(init.astype(dtype)) if torch.is_tensor(init): if len(init.shape) == 2: init = init.reshape(1, *init.shape) if init.shape[0] == 1: init = init.repeat(num_matrices, 1, 1) assert init.shape == kernel_matrix_shape return init else: raise NotImplementedError('Unknown initialization.') class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalResamplingLayer(torch.nn.Module): """Base class for orthogonal up- and downsampling operators. :param low_channel_number: Lower number of channels. These are the input channels in the case of downsampling ops, and the output channels in the case of upsampling ops. :param stride: The downsampling / upsampling factor for each dimension. :param channel_multiplier: The channel multiplier, i.e. the number by which the number of channels are multiplied (downsampling) or divided (upsampling). :param method: Which method to use for parametrizing orthogonal matrices which are used as convolutional kernels. """ def __init__(self, low_channel_number: 'int', stride: 'Union[int, Tuple[int, ...]]', method: 'str'='cayley', init: 'Union[str, np.ndarray, torch.Tensor]'='haar', learnable: 'bool'= True, init_kwargs: 'dict'=None, **kwargs): super(OrthogonalResamplingLayer, self).__init__() self.low_channel_number = low_channel_number self.method = method self.stride = stride self.channel_multiplier = int(np.prod(stride)) self.high_channel_number = self.channel_multiplier * low_channel_number if init_kwargs is None: init_kwargs = {} self.init_kwargs = init_kwargs self.kwargs = kwargs assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self._kernel_matrix_shape = (self.low_channel_number,) + (self. channel_multiplier,) * 2 self._kernel_shape = (self.high_channel_number, 1) + self.stride self.weight = torch.nn.Parameter(__initialize_weight__( kernel_matrix_shape=self._kernel_matrix_shape, stride=self. stride, method=self.method, init=init, **self.init_kwargs)) self.weight.requires_grad = learnable @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel(self): """The kernel associated with the invertible up-/downsampling. """ return self.kernel_matrix.reshape(*self._kernel_shape) class InvertibleDownsampling2D(OrthogonalResamplingLayer): def __init__(self, in_channels: 'int', stride: '_size_2_t'=2, method: 'str'='cayley', init: 'str'='haar', learnable: 'bool'=True, *args, **kwargs): stride = tuple(_pair(stride)) channel_multiplier = int(np.prod(stride)) self.in_channels = in_channels self.out_channels = in_channels * channel_multiplier super(InvertibleDownsampling2D, self).__init__(*args, low_channel_number=self.in_channels, stride=stride, method= method, init=init, learnable=learnable, **kwargs) def forward(self, x): return F.conv2d(x, self.kernel, stride=self.stride, groups=self. low_channel_number) def inverse(self, x): return F.conv_transpose2d(x, self.kernel, stride=self.stride, groups=self.low_channel_number) def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_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.autograd import Function import numpy as np from warnings import warn from typing import Union from typing import Tuple from torch.nn.common_types import _size_2_t from torch.nn.modules.utils import _pair 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_add_eye_sub_0(in_ptr0, 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 tmp6 = tl.load(in_ptr0 + (x2 + 4 * y3), xmask & ymask, eviction_policy= 'evict_last') tmp7 = tl.load(in_ptr0 + (y0 + 4 * x2 + 16 * y1), xmask & ymask, eviction_policy='evict_last') tmp0 = y0 tmp1 = x2 tmp2 = tmp0 == tmp1 tmp3 = 1.0 tmp4 = 0.0 tmp5 = tl.where(tmp2, tmp3, tmp4) tmp8 = tmp6 - tmp7 tmp9 = tmp5 + tmp8 tmp10 = tmp5 - tmp8 tl.store(out_ptr0 + (x2 + 4 * y3), tmp9, xmask & ymask) tl.store(out_ptr1 + (x2 + 4 * y3), tmp10, xmask & ymask) @triton.jit def triton_poi_fused_clone_1(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 = args args.clear() assert_size_stride(primals_1, (4, 4, 4), (16, 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), (16, 4, 1), torch.float32) buf5 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_eye_sub_0[grid(16, 4)](primals_1, buf0, buf5, 16, 4, XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1) del primals_1 buf1 = torch.ops.aten.linalg_lu_factor_ex.default(buf0) buf2 = buf1[0] buf3 = buf1[1] del buf1 buf6 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf5) buf7 = buf6 del buf6 buf8 = buf0 del buf0 triton_poi_fused_clone_1[grid(16, 4)](buf7, buf8, 16, 4, XBLOCK=4, YBLOCK=16, num_warps=1, num_stages=1) buf9 = extern_kernels.convolution(primals_2, reinterpret_tensor( buf8, (16, 1, 2, 2), (4, 0, 2, 1), 0), stride=(2, 2), padding=( 0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=4, bias=None) assert_size_stride(buf9, (4, 16, 2, 2), (64, 4, 2, 1)) buf10 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf7) del buf2 del buf3 del buf7 buf11 = buf10 del buf10 return buf9, primals_2, buf5, reinterpret_tensor(buf8, (16, 1, 2, 2), ( 4, 4, 2, 1), 0), buf11 def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) def __initialize_weight__(kernel_matrix_shape: 'Tuple[int, ...]', stride: 'Tuple[int, ...]', method: 'str'='cayley', init: 'str'='haar', dtype: 'str'='float32', *args, **kwargs): """Function which computes specific orthogonal matrices. For some chosen method of parametrizing orthogonal matrices, this function outputs the required weights necessary to represent a chosen initialization as a Pytorch tensor of matrices. Args: kernel_matrix_shape : The output shape of the orthogonal matrices. Should be (num_matrices, height, width). stride : The stride for the invertible up- or downsampling for which this matrix is to be used. The length of ``stride`` should match the dimensionality of the data. method : The method for parametrising orthogonal matrices. Should be 'exp' or 'cayley' init : The matrix which should be represented. Should be 'squeeze', 'pixel_shuffle', 'haar' or 'random'. 'haar' is only possible if ``stride`` is only 2. dtype : Numpy dtype which should be used for the matrix. *args: Variable length argument iterable. **kwargs: Arbitrary keyword arguments. Returns: Tensor : Orthogonal matrices of shape ``kernel_matrix_shape``. """ dim = len(stride) num_matrices = kernel_matrix_shape[0] assert method in ['exp', 'cayley', 'householder'] if method == 'householder': warn( 'Householder parametrization not fully implemented yet. Only random initialization currently working.' ) init = 'random' if init == 'random': return torch.randn(kernel_matrix_shape) if init == 'haar' and set(stride) != {2}: None None init = 'squeeze' if init == 'haar' and set(stride) == {2}: if method == 'exp': p = np.pi / 4 if dim == 1: weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: weight = np.array([[[0, p, p, 0, p, 0, 0, 0], [0, 0, 0, p, 0, p, 0, 0], [0, 0, 0, p, 0, 0, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, p, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) elif method == 'cayley': if dim == 1: p = -np.sqrt(2) / (2 - np.sqrt(2)) weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: p = 0.5 weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: p = 1 / np.sqrt(2) weight = np.array([[[0, -p, -p, 0, -p, 0, 0, 1 - p], [0, 0, 0, -p, 0, -p, p - 1, 0], [0, 0, 0, -p, 0, p - 1, -p, 0], [0, 0, 0, 0, 1 - p, 0, 0, -p], [0, 0, 0, 0, 0, -p, -p, 0], [0, 0, 0, 0, 0, 0, 0, -p], [0, 0, 0, 0, 0, 0, 0, -p ], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) if init in ['squeeze', 'pixel_shuffle', 'zeros']: if method == 'exp' or method == 'cayley': return torch.zeros(*kernel_matrix_shape) if type(init) is np.ndarray: init = torch.tensor(init.astype(dtype)) if torch.is_tensor(init): if len(init.shape) == 2: init = init.reshape(1, *init.shape) if init.shape[0] == 1: init = init.repeat(num_matrices, 1, 1) assert init.shape == kernel_matrix_shape return init else: raise NotImplementedError('Unknown initialization.') class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalResamplingLayer(torch.nn.Module): """Base class for orthogonal up- and downsampling operators. :param low_channel_number: Lower number of channels. These are the input channels in the case of downsampling ops, and the output channels in the case of upsampling ops. :param stride: The downsampling / upsampling factor for each dimension. :param channel_multiplier: The channel multiplier, i.e. the number by which the number of channels are multiplied (downsampling) or divided (upsampling). :param method: Which method to use for parametrizing orthogonal matrices which are used as convolutional kernels. """ def __init__(self, low_channel_number: 'int', stride: 'Union[int, Tuple[int, ...]]', method: 'str'='cayley', init: 'Union[str, np.ndarray, torch.Tensor]'='haar', learnable: 'bool'= True, init_kwargs: 'dict'=None, **kwargs): super(OrthogonalResamplingLayer, self).__init__() self.low_channel_number = low_channel_number self.method = method self.stride = stride self.channel_multiplier = int(np.prod(stride)) self.high_channel_number = self.channel_multiplier * low_channel_number if init_kwargs is None: init_kwargs = {} self.init_kwargs = init_kwargs self.kwargs = kwargs assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self._kernel_matrix_shape = (self.low_channel_number,) + (self. channel_multiplier,) * 2 self._kernel_shape = (self.high_channel_number, 1) + self.stride self.weight = torch.nn.Parameter(__initialize_weight__( kernel_matrix_shape=self._kernel_matrix_shape, stride=self. stride, method=self.method, init=init, **self.init_kwargs)) self.weight.requires_grad = learnable @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel(self): """The kernel associated with the invertible up-/downsampling. """ return self.kernel_matrix.reshape(*self._kernel_shape) class InvertibleDownsampling2DNew(OrthogonalResamplingLayer): def __init__(self, in_channels: 'int', stride: '_size_2_t'=2, method: 'str'='cayley', init: 'str'='haar', learnable: 'bool'=True, *args, **kwargs): stride = tuple(_pair(stride)) channel_multiplier = int(np.prod(stride)) self.in_channels = in_channels self.out_channels = in_channels * channel_multiplier super(InvertibleDownsampling2DNew, self).__init__(*args, low_channel_number=self.in_channels, stride=stride, method= method, init=init, learnable=learnable, **kwargs) def inverse(self, x): return F.conv_transpose2d(x, self.kernel, stride=self.stride, groups=self.low_channel_number) def forward(self, input_0): primals_1 = self.weight primals_2 = input_0 output = call([primals_1, primals_2]) return output[0]
cetmann/iunets
InvertibleDownsampling2D
false
15,021
[ "MIT" ]
86
80ed7cce0e505a0396c42359eaf27819222d71f6
https://github.com/cetmann/iunets/tree/80ed7cce0e505a0396c42359eaf27819222d71f6
EncoderLayer
from _paritybench_helpers import _mock_config import math import torch from torch.nn import functional as F from torch.autograd import Variable from torch import nn def softmax(x): if x.dim() == 3: return F.softmax(x.transpose(0, 2)).transpose(0, 2) return F.softmax(x) def gumbel_softmax(input, beta=0.5, tau=1.0): noise = input.data.new(*input.size()).uniform_() noise.add_(TINY).log_().neg_().add_(TINY).log_().neg_() return softmax((input + beta * Variable(noise)) / tau) def matmul(x, y): if x.dim() == y.dim(): return x @ y if x.dim() == y.dim() - 1: return (x.unsqueeze(-2) @ y).squeeze(-2) return (x @ y.unsqueeze(-1)).squeeze(-1) def mask(targets, out, input_mask=None, return_mask=False): if input_mask is None: input_mask = targets != 1 out_mask = input_mask.unsqueeze(-1).expand_as(out) if return_mask: return targets[input_mask], out[out_mask].view(-1, out.size(-1) ), the_mask return targets[input_mask], out[out_mask].view(-1, out.size(-1)) class Linear(nn.Linear): def forward(self, x): size = x.size() return super().forward(x.contiguous().view(-1, size[-1])).view(* size[:-1], -1) class LayerNorm(nn.Module): def __init__(self, d_model, eps=1e-06): super().__init__() self.gamma = nn.Parameter(torch.ones(d_model)) self.beta = nn.Parameter(torch.zeros(d_model)) self.eps = eps def forward(self, x): mean = x.mean(-1, keepdim=True) std = x.std(-1, keepdim=True) return self.gamma * (x - mean) / (std + self.eps) + self.beta class ResidualBlock(nn.Module): def __init__(self, layer, d_model, drop_ratio, pos=0): super().__init__() self.layer = layer self.dropout = nn.Dropout(drop_ratio) self.layernorm = LayerNorm(d_model) self.pos = pos def forward(self, *x): return self.layernorm(x[self.pos] + self.dropout(self.layer(*x))) class Attention(nn.Module): def __init__(self, d_key, drop_ratio, causal, diag=False, window=-1, noisy=False): super().__init__() self.scale = math.sqrt(d_key) self.dropout = nn.Dropout(drop_ratio) self.causal = causal self.diag = diag self.window = window self.noisy = noisy def forward(self, query, key, value=None, mask=None, feedback=None, beta=0, tau=1, weights=None): dot_products = matmul(query, key.transpose(1, 2)) if weights is not None: dot_products = dot_products + weights if query.dim() == 3 and self.causal and query.size(1) == key.size(1): tri = key.data.new(key.size(1), key.size(1)).fill_(1).triu(1) * INF dot_products.data.sub_(tri.unsqueeze(0)) if self.window > 0: window_mask = key.data.new(key.size(1), key.size(1)).fill_(1) window_mask = (window_mask.triu(self.window + 1) + window_mask. tril(-self.window - 1)) * INF dot_products.data.sub_(window_mask.unsqueeze(0)) if self.diag: inds = torch.arange(0, key.size(1)).long().view(1, 1, -1) if key.is_cuda: inds = inds dot_products.data.scatter_(1, inds.expand(dot_products.size(0), 1, inds.size(-1)), -INF) if mask is not None: if dot_products.dim() == 2: dot_products.data -= (1 - mask) * INF else: dot_products.data -= (1 - mask[:, None, :]) * INF if value is None: return dot_products logits = dot_products / self.scale if not self.noisy: probs = softmax(logits) else: probs = gumbel_softmax(logits, beta=beta, tau=tau) if feedback is not None: feedback.append(probs.contiguous()) return matmul(self.dropout(probs), value) class MultiHead2(nn.Module): def __init__(self, d_key, d_value, n_heads, drop_ratio, causal=False, diag=False, window=-1, noisy=False, use_wo=True): super().__init__() self.attention = Attention(d_key, drop_ratio, causal=causal, diag= diag, window=window, noisy=noisy) self.wq = Linear(d_key, d_key, bias=use_wo) self.wk = Linear(d_key, d_key, bias=use_wo) self.wv = Linear(d_value, d_value, bias=use_wo) if use_wo: self.wo = Linear(d_value, d_key, bias=use_wo) self.use_wo = use_wo self.n_heads = n_heads def forward(self, query, key, value, mask=None, feedback=None, weights= None, beta=0, tau=1): query, key, value = self.wq(query), self.wk(key), self.wv(value) B, Tq, D = query.size() _, Tk, _ = key.size() N = self.n_heads probs = [] query, key, value = (x.contiguous().view(B, -1, N, D // N). transpose(2, 1).contiguous().view(B * N, -1, D // N) for x in ( query, key, value)) if mask is not None: mask = mask[:, None, :].expand(B, N, Tk).contiguous().view(B * N, -1) outputs = self.attention(query, key, value, mask, probs, beta, tau, weights) outputs = outputs.contiguous().view(B, N, -1, D // N).transpose(2, 1 ).contiguous().view(B, -1, D) if feedback is not None: feedback.append(probs[0].view(B, N, Tq, Tk)) if self.use_wo: return self.wo(outputs) return outputs class FeedForward(nn.Module): def __init__(self, d_model, d_hidden): super().__init__() self.linear1 = Linear(d_model, d_hidden) self.linear2 = Linear(d_hidden, d_model) def forward(self, x): return self.linear2(F.relu(self.linear1(x))) class EncoderLayer(nn.Module): def __init__(self, args): super().__init__() self.selfattn = ResidualBlock(MultiHead2(args.d_model, args.d_model, args.n_heads, args.drop_ratio, use_wo=args.use_wo), args. d_model, args.drop_ratio) self.feedforward = ResidualBlock(FeedForward(args.d_model, args. d_hidden), args.d_model, args.drop_ratio) def forward(self, x, mask=None): return self.feedforward(self.selfattn(x, x, x, mask)) def get_inputs(): return [torch.rand([4, 4, 4])] def get_init_inputs(): return [[], {'args': _mock_config(d_model=4, n_heads=4, drop_ratio=0.5, use_wo=4, d_hidden=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 from torch.nn import functional as F from torch.autograd import Variable 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_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__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) 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 = 0.5 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 = tl_math.exp(tmp14) tl.store(out_ptr0 + x2, tmp15, xmask) @triton.jit def triton_poi_fused__softmax_2(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 % 16 y1 = yindex // 16 tmp0 = tl.load(in_ptr0 + (y1 + 4 * x2 + 16 * y0), xmask & ymask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr0 + (4 * x2 + 16 * y0), xmask & ymask, eviction_policy='evict_last') tmp2 = tl.load(in_ptr0 + (1 + 4 * x2 + 16 * y0), xmask & ymask, eviction_policy='evict_last') tmp4 = tl.load(in_ptr0 + (2 + 4 * x2 + 16 * y0), xmask & ymask, eviction_policy='evict_last') tmp6 = tl.load(in_ptr0 + (3 + 4 * x2 + 16 * y0), xmask & ymask, eviction_policy='evict_last') tmp3 = tmp1 + tmp2 tmp5 = tmp3 + tmp4 tmp7 = tmp5 + tmp6 tmp8 = tmp0 / tmp7 tl.store(out_ptr0 + (y0 + 16 * x2 + 64 * y1), tmp8, xmask & ymask) @triton.jit def triton_poi_fused_clone_3(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_4(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(in_out_ptr0 + x0, tmp29, xmask) tl.store(out_ptr0 + x0, tmp16, xmask) @triton.jit def triton_poi_fused_add_div_mean_mul_std_sub_5(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_6(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_7(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_add_div_mean_mul_std_sub_8(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) 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) = 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, 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,), (1,)) assert_size_stride(primals_11, (4,), (1,)) assert_size_stride(primals_12, (4, 4), (4, 1)) assert_size_stride(primals_13, (4,), (1,)) assert_size_stride(primals_14, (4, 4), (4, 1)) assert_size_stride(primals_15, (4,), (1,)) assert_size_stride(primals_16, (4,), (1,)) assert_size_stride(primals_17, (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_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_1, (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_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, 4, 1), (16, 4, 1, 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, 4, 16), (1, 4, 16), torch.float32) triton_poi_fused__softmax_1[grid(256)](buf5, buf6, 256, XBLOCK=256, num_warps=4, num_stages=1) buf7 = reinterpret_tensor(buf5, (4, 4, 16), (64, 16, 1), 0) del buf5 triton_poi_fused__softmax_2[grid(64, 4)](buf6, buf7, 64, 4, XBLOCK= 4, YBLOCK=32, num_warps=4, num_stages=1) del buf6 buf8 = reinterpret_tensor(buf1, (4, 4, 4, 1), (16, 4, 1, 1), 0) del buf1 triton_poi_fused_clone_0[grid(16, 4)](buf2, primals_7, buf8, 16, 4, XBLOCK=2, YBLOCK=16, num_warps=1, num_stages=1) del primals_7 buf9 = reinterpret_tensor(buf2, (16, 4, 1), (4, 1, 1), 0) del buf2 extern_kernels.bmm(reinterpret_tensor(buf7, (16, 4, 4), (1, 16, 64), 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_3[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_9, reinterpret_tensor(buf10, (16, 4), (4, 1), 0), reinterpret_tensor(primals_8, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf11) del primals_9 buf12 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32) buf13 = buf12 del buf12 buf14 = empty_strided_cuda((4, 4, 1), (4, 1, 16), torch.float32) triton_poi_fused_add_mean_std_4[grid(16)](buf13, primals_1, buf11, buf14, 16, XBLOCK=16, num_warps=1, num_stages=1) buf15 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused_add_div_mean_mul_std_sub_5[grid(64)](primals_10, primals_1, buf11, buf14, buf13, primals_11, buf15, 64, XBLOCK= 64, num_warps=1, num_stages=1) del buf13 del buf14 del primals_11 buf16 = empty_strided_cuda((16, 4), (4, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(buf15, (16, 4), (4, 1), 0), reinterpret_tensor(primals_12, (4, 4), (1, 4), 0), out=buf16) buf17 = reinterpret_tensor(buf16, (4, 4, 4), (16, 4, 1), 0) del buf16 buf21 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.bool) triton_poi_fused_relu_threshold_backward_6[grid(64)](buf17, primals_13, buf21, 64, XBLOCK=64, num_warps=1, num_stages=1) del primals_13 buf18 = empty_strided_cuda((16, 4), (4, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(buf17, (16, 4), (4, 1), 0), reinterpret_tensor(primals_14, (4, 4), (1, 4), 0), out=buf18) buf19 = reinterpret_tensor(buf18, (4, 4, 4), (16, 4, 1), 0) del buf18 triton_poi_fused_add_7[grid(64)](buf19, buf15, primals_15, 64, XBLOCK=64, num_warps=1, num_stages=1) del primals_15 buf20 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused_add_div_mean_mul_std_sub_8[grid(64)](primals_16, buf19, primals_17, buf20, 64, XBLOCK=64, num_warps=1, num_stages=1) del primals_17 return buf20, primals_1, primals_10, primals_16, buf7, reinterpret_tensor( buf10, (16, 4), (4, 1), 0), buf11, reinterpret_tensor(buf15, (16, 4 ), (4, 1), 0), reinterpret_tensor(buf17, (16, 4), (4, 1), 0 ), buf19, primals_14, buf21, primals_12, primals_8, 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, 1), 0) def softmax(x): if x.dim() == 3: return F.softmax(x.transpose(0, 2)).transpose(0, 2) return F.softmax(x) def gumbel_softmax(input, beta=0.5, tau=1.0): noise = input.data.new(*input.size()).uniform_() noise.add_(TINY).log_().neg_().add_(TINY).log_().neg_() return softmax((input + beta * Variable(noise)) / tau) def matmul(x, y): if x.dim() == y.dim(): return x @ y if x.dim() == y.dim() - 1: return (x.unsqueeze(-2) @ y).squeeze(-2) return (x @ y.unsqueeze(-1)).squeeze(-1) def mask(targets, out, input_mask=None, return_mask=False): if input_mask is None: input_mask = targets != 1 out_mask = input_mask.unsqueeze(-1).expand_as(out) if return_mask: return targets[input_mask], out[out_mask].view(-1, out.size(-1) ), the_mask return targets[input_mask], out[out_mask].view(-1, out.size(-1)) class Linear(nn.Linear): def forward(self, x): size = x.size() return super().forward(x.contiguous().view(-1, size[-1])).view(* size[:-1], -1) class LayerNorm(nn.Module): def __init__(self, d_model, eps=1e-06): super().__init__() self.gamma = nn.Parameter(torch.ones(d_model)) self.beta = nn.Parameter(torch.zeros(d_model)) self.eps = eps def forward(self, x): mean = x.mean(-1, keepdim=True) std = x.std(-1, keepdim=True) return self.gamma * (x - mean) / (std + self.eps) + self.beta class ResidualBlock(nn.Module): def __init__(self, layer, d_model, drop_ratio, pos=0): super().__init__() self.layer = layer self.dropout = nn.Dropout(drop_ratio) self.layernorm = LayerNorm(d_model) self.pos = pos def forward(self, *x): return self.layernorm(x[self.pos] + self.dropout(self.layer(*x))) class Attention(nn.Module): def __init__(self, d_key, drop_ratio, causal, diag=False, window=-1, noisy=False): super().__init__() self.scale = math.sqrt(d_key) self.dropout = nn.Dropout(drop_ratio) self.causal = causal self.diag = diag self.window = window self.noisy = noisy def forward(self, query, key, value=None, mask=None, feedback=None, beta=0, tau=1, weights=None): dot_products = matmul(query, key.transpose(1, 2)) if weights is not None: dot_products = dot_products + weights if query.dim() == 3 and self.causal and query.size(1) == key.size(1): tri = key.data.new(key.size(1), key.size(1)).fill_(1).triu(1) * INF dot_products.data.sub_(tri.unsqueeze(0)) if self.window > 0: window_mask = key.data.new(key.size(1), key.size(1)).fill_(1) window_mask = (window_mask.triu(self.window + 1) + window_mask. tril(-self.window - 1)) * INF dot_products.data.sub_(window_mask.unsqueeze(0)) if self.diag: inds = torch.arange(0, key.size(1)).long().view(1, 1, -1) if key.is_cuda: inds = inds dot_products.data.scatter_(1, inds.expand(dot_products.size(0), 1, inds.size(-1)), -INF) if mask is not None: if dot_products.dim() == 2: dot_products.data -= (1 - mask) * INF else: dot_products.data -= (1 - mask[:, None, :]) * INF if value is None: return dot_products logits = dot_products / self.scale if not self.noisy: probs = softmax(logits) else: probs = gumbel_softmax(logits, beta=beta, tau=tau) if feedback is not None: feedback.append(probs.contiguous()) return matmul(self.dropout(probs), value) class MultiHead2(nn.Module): def __init__(self, d_key, d_value, n_heads, drop_ratio, causal=False, diag=False, window=-1, noisy=False, use_wo=True): super().__init__() self.attention = Attention(d_key, drop_ratio, causal=causal, diag= diag, window=window, noisy=noisy) self.wq = Linear(d_key, d_key, bias=use_wo) self.wk = Linear(d_key, d_key, bias=use_wo) self.wv = Linear(d_value, d_value, bias=use_wo) if use_wo: self.wo = Linear(d_value, d_key, bias=use_wo) self.use_wo = use_wo self.n_heads = n_heads def forward(self, query, key, value, mask=None, feedback=None, weights= None, beta=0, tau=1): query, key, value = self.wq(query), self.wk(key), self.wv(value) B, Tq, D = query.size() _, Tk, _ = key.size() N = self.n_heads probs = [] query, key, value = (x.contiguous().view(B, -1, N, D // N). transpose(2, 1).contiguous().view(B * N, -1, D // N) for x in ( query, key, value)) if mask is not None: mask = mask[:, None, :].expand(B, N, Tk).contiguous().view(B * N, -1) outputs = self.attention(query, key, value, mask, probs, beta, tau, weights) outputs = outputs.contiguous().view(B, N, -1, D // N).transpose(2, 1 ).contiguous().view(B, -1, D) if feedback is not None: feedback.append(probs[0].view(B, N, Tq, Tk)) if self.use_wo: return self.wo(outputs) return outputs class FeedForward(nn.Module): def __init__(self, d_model, d_hidden): super().__init__() self.linear1 = Linear(d_model, d_hidden) self.linear2 = Linear(d_hidden, d_model) def forward(self, x): return self.linear2(F.relu(self.linear1(x))) class EncoderLayerNew(nn.Module): def __init__(self, args): super().__init__() self.selfattn = ResidualBlock(MultiHead2(args.d_model, args.d_model, args.n_heads, args.drop_ratio, use_wo=args.use_wo), args. d_model, args.drop_ratio) self.feedforward = ResidualBlock(FeedForward(args.d_model, args. d_hidden), args.d_model, args.drop_ratio) def forward(self, input_0): primals_2 = self.selfattn.layer.wq.weight primals_3 = self.selfattn.layer.wq.bias primals_4 = self.selfattn.layer.wk.weight primals_5 = self.selfattn.layer.wk.bias primals_6 = self.selfattn.layer.wv.weight primals_7 = self.selfattn.layer.wv.bias primals_8 = self.selfattn.layer.wo.weight primals_9 = self.selfattn.layer.wo.bias primals_10 = self.selfattn.layernorm.gamma primals_11 = self.selfattn.layernorm.beta primals_12 = self.feedforward.layer.linear1.weight primals_13 = self.feedforward.layer.linear1.bias primals_14 = self.feedforward.layer.linear2.weight primals_15 = self.feedforward.layer.linear2.bias primals_16 = self.feedforward.layernorm.gamma primals_17 = self.feedforward.layernorm.beta 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]) return output[0]
cclauss/nonauto-nmt
EncoderLayer
false
15,022
[ "BSD-3-Clause" ]
262
efcbe4f2329b140ac3ce06abb6409457cebc8e49
https://github.com/cclauss/nonauto-nmt/tree/efcbe4f2329b140ac3ce06abb6409457cebc8e49
BatchHardTripletLoss
import torch import torch.nn as nn class BatchHardTripletLoss(nn.Module): def __init__(self, margin=0): super(BatchHardTripletLoss, self).__init__() self.margin = margin self.ranking_loss = nn.MarginRankingLoss(margin=margin) def forward(self, inputs, targets): batch_size = inputs.size(0) dist = torch.pow(inputs, 2).sum(dim=1, keepdim=True).expand(batch_size, batch_size) dist = dist + dist.t() dist.addmm_(1, -2, inputs, inputs.t()) dist = dist.clamp(min=1e-12).sqrt() mask = targets.expand(batch_size, batch_size).eq(targets.expand( batch_size, batch_size).t()) dist_ap = dist[mask == 1] dist_ap = dist_ap.view(batch_size, -1) dist_an = dist[mask == 0] dist_an = dist_an.view(batch_size, -1) dist_ap, _ = torch.max(dist_ap, dim=1) dist_an, _ = torch.min(dist_an, dim=1) y = torch.ones(dist_an.size(), dtype=dist_an.dtype, device=dist_an. device) loss = self.ranking_loss(dist_an, dist_ap, y) return loss def get_inputs(): return [torch.rand([4, 4]), torch.rand([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 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_clamp_sqrt_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 x1 = xindex // 4 x0 = xindex % 4 tmp0 = tl.load(in_out_ptr0 + x2, xmask) tmp1 = tl.load(in_ptr0 + 4 * x1, xmask, eviction_policy='evict_last') tmp3 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last') tmp6 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last') tmp9 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last') tmp12 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last') tmp14 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp17 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp20 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp2 = tmp1 * tmp1 tmp4 = tmp3 * tmp3 tmp5 = tmp2 + tmp4 tmp7 = tmp6 * tmp6 tmp8 = tmp5 + tmp7 tmp10 = tmp9 * tmp9 tmp11 = tmp8 + tmp10 tmp13 = tmp12 * tmp12 tmp15 = tmp14 * tmp14 tmp16 = tmp13 + tmp15 tmp18 = tmp17 * tmp17 tmp19 = tmp16 + tmp18 tmp21 = tmp20 * tmp20 tmp22 = tmp19 + tmp21 tmp23 = tmp11 + tmp22 tmp24 = tmp0 + tmp23 tmp25 = 1e-12 tmp26 = triton_helpers.maximum(tmp24, tmp25) tmp27 = libdevice.sqrt(tmp26) tl.store(in_out_ptr0 + x2, tmp27, xmask) @triton.jit def triton_poi_fused_eq_1(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 4 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 x1 = xindex y0 = yindex tmp0 = tl.load(in_ptr0 + (x1 + 4 * y0), xmask & ymask) tmp1 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask) tmp2 = tmp0 == tmp1 tmp3 = tmp2.to(tl.int64) tmp4 = tl.full([1, 1], 1, tl.int64) tmp5 = tmp3 == tmp4 tl.store(out_ptr0 + (x1 + 4 * y0), tmp2, xmask & ymask) tl.store(out_ptr1 + (x1 + 4 * y0), tmp5, xmask & ymask) def call(args): arg0_1, arg1_1 = args args.clear() assert_size_stride(arg0_1, (4, 4), (4, 1)) assert_size_stride(arg1_1, (4, 4), (4, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.mm(arg0_1, reinterpret_tensor(arg0_1, (4, 4), (1, 4), 0), out=buf0) buf1 = buf0 del buf0 buf2 = buf1 del buf1 get_raw_stream(0) triton_poi_fused_add_clamp_sqrt_0[grid(16)](buf2, arg0_1, 16, XBLOCK=16, num_warps=1, num_stages=1) del arg0_1 buf3 = empty_strided_cuda((4, 4), (4, 1), torch.bool) buf4 = empty_strided_cuda((4, 4), (4, 1), torch.bool) triton_poi_fused_eq_1[grid(4, 4)](arg1_1, buf3, buf4, 4, 4, XBLOCK= 4, YBLOCK=4, num_warps=1, num_stages=1) del arg1_1 return buf2, buf4, buf3 class BatchHardTripletLossNew(nn.Module): def __init__(self, margin=0): super(BatchHardTripletLossNew, self).__init__() self.margin = margin self.ranking_loss = nn.MarginRankingLoss(margin=margin) def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0]
chenyanghungry/person-reid-lib
BatchHardTripletLoss
false
15,023
[ "MIT" ]
81
783e66c9bfedf582e2cf935b9f5be960b543ac3c
https://github.com/chenyanghungry/person-reid-lib/tree/783e66c9bfedf582e2cf935b9f5be960b543ac3c
SelfAttentive
import torch import torch.nn as nn class SelfAttentive(nn.Module): def __init__(self, hidden_size, att_hops=1, att_unit=200, dropout=0.2): super(SelfAttentive, self).__init__() self.drop = nn.Dropout(dropout) self.ws1 = nn.Linear(hidden_size, att_unit, bias=False) self.ws2 = nn.Linear(att_unit, att_hops, bias=False) self.tanh = nn.Tanh() self.softmax = nn.Softmax() self.attention_hops = att_hops def forward(self, rnn_out, mask=None): outp = rnn_out size = outp.size() compressed_embeddings = outp.reshape(-1, size[2]) hbar = self.tanh(self.ws1(self.drop(compressed_embeddings))) alphas = self.ws2(hbar).view(size[0], size[1], -1) alphas = torch.transpose(alphas, 1, 2).contiguous() if mask is not None: mask = mask.squeeze(2) concatenated_mask = [mask for i in range(self.attention_hops)] concatenated_mask = torch.cat(concatenated_mask, 1) penalized_alphas = alphas + concatenated_mask else: penalized_alphas = alphas alphas = self.softmax(penalized_alphas.view(-1, size[1])) alphas = alphas.view(size[0], self.attention_hops, size[1]) return torch.bmm(alphas, outp), alphas def get_inputs(): return [torch.rand([4, 4, 4])] def get_init_inputs(): return [[], {'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 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_tanh_0(in_out_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 3200 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) @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 = args args.clear() assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1)) assert_size_stride(primals_2, (200, 4), (4, 1)) assert_size_stride(primals_3, (1, 200), (200, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((16, 200), (200, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(primals_1, (16, 4), (4, 1), 0), reinterpret_tensor(primals_2, (4, 200), (1, 4), 0), out=buf0) del primals_2 buf1 = buf0 del buf0 get_raw_stream(0) triton_poi_fused_tanh_0[grid(3200)](buf1, 3200, XBLOCK=128, num_warps=4, num_stages=1) buf2 = empty_strided_cuda((16, 1), (1, 1), torch.float32) extern_kernels.mm(buf1, reinterpret_tensor(primals_3, (200, 1), (1, 200), 0), out=buf2) buf3 = empty_strided_cuda((4, 4), (4, 1), torch.float32) triton_poi_fused__softmax_1[grid(16)](buf2, buf3, 16, XBLOCK=16, num_warps=1, num_stages=1) buf4 = reinterpret_tensor(buf2, (4, 4), (4, 1), 0) del buf2 triton_poi_fused__softmax_2[grid(16)](buf3, buf4, 16, XBLOCK=16, num_warps=1, num_stages=1) buf5 = reinterpret_tensor(buf3, (4, 1, 4), (4, 4, 1), 0) del buf3 extern_kernels.bmm(reinterpret_tensor(buf4, (4, 1, 4), (4, 4, 1), 0 ), primals_1, out=buf5) return buf5, reinterpret_tensor(buf4, (4, 1, 4), (4, 4, 1), 0 ), primals_1, buf1, buf4, primals_3 class SelfAttentiveNew(nn.Module): def __init__(self, hidden_size, att_hops=1, att_unit=200, dropout=0.2): super(SelfAttentiveNew, self).__init__() self.drop = nn.Dropout(dropout) self.ws1 = nn.Linear(hidden_size, att_unit, bias=False) self.ws2 = nn.Linear(att_unit, att_hops, bias=False) self.tanh = nn.Tanh() self.softmax = nn.Softmax() self.attention_hops = att_hops def forward(self, input_0): primals_2 = self.ws1.weight primals_3 = self.ws2.weight primals_1 = input_0 output = call([primals_1, primals_2, primals_3]) return output[0], output[1]
chenyangh/SemEval2019-Task3
SelfAttentive
false
15,024
[ "MIT" ]
50
c6204797b4b6cc08cb4d2d88108405f959d63ee9
https://github.com/chenyangh/SemEval2019-Task3/tree/c6204797b4b6cc08cb4d2d88108405f959d63ee9
Attention
import torch import torch.nn as nn import torch.nn class Attention(nn.Module): def __init__(self, dim_i, dim_o): """ build the target-aware attention input schema: dim_i: the dimension of the input feature vector dim_o: the dimension of the output feature vector output schema: return a aggregated vector from the context k, v of q """ super(Attention, self).__init__() self.Q = nn.Linear(dim_i, dim_o) self.K = nn.Linear(dim_i, dim_o) self.V = nn.Linear(dim_i, dim_o) def forward(self, hist_seq_emb, hist_seq_mask, cand_emb): q, k, v = self.Q(cand_emb), self.K(hist_seq_emb), self.V(hist_seq_emb) logits = torch.sum(q.unsqueeze(1) * k, dim=2) logits = logits * hist_seq_mask + logits * (1 - hist_seq_mask ) * -2 ** 32.0 scores = torch.softmax(logits, dim=1) output = torch.sum(scores.unsqueeze(2) * v, dim=1) return output 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 [[], {'dim_i': 4, 'dim_o': 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.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_sum_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 % 16 x2 = xindex // 64 x1 = xindex // 16 % 4 x3 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr1 + (x0 + 64 * x1), xmask, eviction_policy= 'evict_last') tmp3 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp4 = tl.load(in_ptr1 + (16 + x0 + 64 * x1), xmask, eviction_policy= 'evict_last') tmp7 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp8 = tl.load(in_ptr1 + (32 + x0 + 64 * x1), xmask, eviction_policy= 'evict_last') tmp11 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp12 = tl.load(in_ptr1 + (48 + x0 + 64 * x1), 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 + x3, tmp14, xmask) @triton.jit def triton_poi_fused__softmax_add_mul_rsub_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 x0 = xindex % 16 x1 = xindex // 16 x2 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask) tmp1 = tl.load(in_ptr1 + (x0 + 64 * x1), xmask) tmp9 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask) tmp10 = tl.load(in_ptr1 + (16 + x0 + 64 * x1), xmask) tmp17 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask) tmp18 = tl.load(in_ptr1 + (32 + x0 + 64 * x1), xmask) tmp25 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask) tmp26 = tl.load(in_ptr1 + (48 + x0 + 64 * x1), xmask) tmp2 = tmp0 * tmp1 tmp3 = 1.0 tmp4 = tmp3 - tmp1 tmp5 = tmp0 * tmp4 tmp6 = -4294967296.0 tmp7 = tmp5 * tmp6 tmp8 = tmp2 + tmp7 tmp11 = tmp9 * tmp10 tmp12 = tmp3 - tmp10 tmp13 = tmp9 * tmp12 tmp14 = tmp13 * tmp6 tmp15 = tmp11 + tmp14 tmp16 = triton_helpers.maximum(tmp8, tmp15) tmp19 = tmp17 * tmp18 tmp20 = tmp3 - tmp18 tmp21 = tmp17 * tmp20 tmp22 = tmp21 * tmp6 tmp23 = tmp19 + tmp22 tmp24 = triton_helpers.maximum(tmp16, tmp23) tmp27 = tmp25 * tmp26 tmp28 = tmp3 - tmp26 tmp29 = tmp25 * tmp28 tmp30 = tmp29 * tmp6 tmp31 = tmp27 + tmp30 tmp32 = triton_helpers.maximum(tmp24, tmp31) tmp33 = tmp8 - tmp32 tmp34 = tl_math.exp(tmp33) tmp35 = tmp15 - tmp32 tmp36 = tl_math.exp(tmp35) tmp37 = tmp34 + tmp36 tmp38 = tmp23 - tmp32 tmp39 = tl_math.exp(tmp38) tmp40 = tmp37 + tmp39 tmp41 = tmp31 - tmp32 tmp42 = tl_math.exp(tmp41) tmp43 = tmp40 + tmp42 tl.store(out_ptr0 + x2, tmp32, xmask) tl.store(out_ptr1 + x2, tmp43, xmask) @triton.jit def triton_poi_fused_mul_sum_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 x0 = xindex % 16 x2 = xindex // 64 x3 = xindex % 64 x4 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr1 + (x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp9 = tl.load(in_ptr2 + (x0 + 16 * x2), xmask, eviction_policy= 'evict_last') tmp12 = tl.load(in_ptr3 + (x0 + 16 * x2), xmask, eviction_policy= 'evict_last') tmp14 = tl.load(in_ptr4 + x3, xmask, eviction_policy='evict_last') tmp16 = tl.load(in_ptr0 + (16 + x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp17 = tl.load(in_ptr1 + (16 + x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp26 = tl.load(in_ptr4 + (64 + x3), xmask, eviction_policy='evict_last') tmp29 = tl.load(in_ptr0 + (32 + x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp30 = tl.load(in_ptr1 + (32 + x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp39 = tl.load(in_ptr4 + (128 + x3), xmask, eviction_policy='evict_last') tmp42 = tl.load(in_ptr0 + (48 + x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp43 = tl.load(in_ptr1 + (48 + x0 + 64 * x2), xmask, eviction_policy= 'evict_last') tmp52 = tl.load(in_ptr4 + (192 + x3), xmask, eviction_policy='evict_last') tmp2 = tmp0 * tmp1 tmp3 = 1.0 tmp4 = tmp3 - tmp1 tmp5 = tmp0 * tmp4 tmp6 = -4294967296.0 tmp7 = tmp5 * tmp6 tmp8 = tmp2 + tmp7 tmp10 = tmp8 - tmp9 tmp11 = tl_math.exp(tmp10) tmp13 = tmp11 / tmp12 tmp15 = tmp13 * tmp14 tmp18 = tmp16 * tmp17 tmp19 = tmp3 - tmp17 tmp20 = tmp16 * tmp19 tmp21 = tmp20 * tmp6 tmp22 = tmp18 + tmp21 tmp23 = tmp22 - tmp9 tmp24 = tl_math.exp(tmp23) tmp25 = tmp24 / tmp12 tmp27 = tmp25 * tmp26 tmp28 = tmp15 + tmp27 tmp31 = tmp29 * tmp30 tmp32 = tmp3 - tmp30 tmp33 = tmp29 * tmp32 tmp34 = tmp33 * tmp6 tmp35 = tmp31 + tmp34 tmp36 = tmp35 - tmp9 tmp37 = tl_math.exp(tmp36) tmp38 = tmp37 / tmp12 tmp40 = tmp38 * tmp39 tmp41 = tmp28 + tmp40 tmp44 = tmp42 * tmp43 tmp45 = tmp3 - tmp43 tmp46 = tmp42 * tmp45 tmp47 = tmp46 * tmp6 tmp48 = tmp44 + tmp47 tmp49 = tmp48 - tmp9 tmp50 = tl_math.exp(tmp49) tmp51 = tmp50 / tmp12 tmp53 = tmp51 * tmp52 tmp54 = tmp41 + tmp53 tl.store(out_ptr0 + x4, tmp54, 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, 4), (64, 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, 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_1 del primals_2 buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32) extern_kernels.addmm(primals_5, reinterpret_tensor(primals_6, (64, 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((64, 4), (4, 1), torch.float32) extern_kernels.addmm(primals_8, reinterpret_tensor(primals_6, (64, 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, 4), (64, 16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_mul_sum_0[grid(256)](buf0, buf1, buf3, 256, XBLOCK =128, num_warps=4, num_stages=1) buf4 = empty_strided_cuda((4, 1, 4, 4), (16, 64, 4, 1), torch.float32) buf5 = empty_strided_cuda((4, 1, 4, 4), (16, 64, 4, 1), torch.float32) triton_poi_fused__softmax_add_mul_rsub_1[grid(64)](buf3, primals_9, buf4, buf5, 64, XBLOCK=64, num_warps=1, num_stages=1) buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused_mul_sum_2[grid(256)](buf3, primals_9, buf4, buf5, buf2, buf6, 256, XBLOCK=128, num_warps=4, num_stages=1) del buf3 del buf4 del buf5 return buf6, primals_9, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0 ), buf0, reinterpret_tensor(primals_6, (64, 4), (4, 1), 0), buf1, buf2 class AttentionNew(nn.Module): def __init__(self, dim_i, dim_o): """ build the target-aware attention input schema: dim_i: the dimension of the input feature vector dim_o: the dimension of the output feature vector output schema: return a aggregated vector from the context k, v of q """ super(AttentionNew, self).__init__() self.Q = nn.Linear(dim_i, dim_o) self.K = nn.Linear(dim_i, dim_o) self.V = nn.Linear(dim_i, dim_o) 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]
chencsgit/luoxi_models
Attention
false
15,025
[ "Apache-2.0" ]
58
ea9e69dfb81b29f41ed92c75faacf81114c69a2f
https://github.com/chencsgit/luoxi_models/tree/ea9e69dfb81b29f41ed92c75faacf81114c69a2f
PoseNormalize
import torch import torch.nn as nn class PoseNormalize(nn.Module): @torch.no_grad() def forward(self, x): return x * 2 - 1 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 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_mul_sub_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 = 2.0 tmp2 = tmp0 * tmp1 tmp3 = 1.0 tmp4 = tmp2 - tmp3 tl.store(out_ptr0 + x0, tmp4, 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_sub_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256, num_warps=4, num_stages=1) del arg0_1 return buf0, class PoseNormalizeNew(nn.Module): def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0]
chinitaberrio/DeepPrivacy
PoseNormalize
false
15,026
[ "MIT" ]
1,128
d50e1b5ae762b47ab5a8f54cb90e66465057bd25
https://github.com/chinitaberrio/DeepPrivacy/tree/d50e1b5ae762b47ab5a8f54cb90e66465057bd25
InvertibleDownsampling3D
from torch.autograd import Function import torch import numpy as np from warnings import warn from typing import Union from typing import Tuple from torch.nn.common_types import _size_3_t from torch.nn.modules.utils import _triple import torch.nn.functional as F def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) def __initialize_weight__(kernel_matrix_shape: 'Tuple[int, ...]', stride: 'Tuple[int, ...]', method: 'str'='cayley', init: 'str'='haar', dtype: 'str'='float32', *args, **kwargs): """Function which computes specific orthogonal matrices. For some chosen method of parametrizing orthogonal matrices, this function outputs the required weights necessary to represent a chosen initialization as a Pytorch tensor of matrices. Args: kernel_matrix_shape : The output shape of the orthogonal matrices. Should be (num_matrices, height, width). stride : The stride for the invertible up- or downsampling for which this matrix is to be used. The length of ``stride`` should match the dimensionality of the data. method : The method for parametrising orthogonal matrices. Should be 'exp' or 'cayley' init : The matrix which should be represented. Should be 'squeeze', 'pixel_shuffle', 'haar' or 'random'. 'haar' is only possible if ``stride`` is only 2. dtype : Numpy dtype which should be used for the matrix. *args: Variable length argument iterable. **kwargs: Arbitrary keyword arguments. Returns: Tensor : Orthogonal matrices of shape ``kernel_matrix_shape``. """ dim = len(stride) num_matrices = kernel_matrix_shape[0] assert method in ['exp', 'cayley', 'householder'] if method == 'householder': warn( 'Householder parametrization not fully implemented yet. Only random initialization currently working.' ) init = 'random' if init == 'random': return torch.randn(kernel_matrix_shape) if init == 'haar' and set(stride) != {2}: None None init = 'squeeze' if init == 'haar' and set(stride) == {2}: if method == 'exp': p = np.pi / 4 if dim == 1: weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: weight = np.array([[[0, p, p, 0, p, 0, 0, 0], [0, 0, 0, p, 0, p, 0, 0], [0, 0, 0, p, 0, 0, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, p, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) elif method == 'cayley': if dim == 1: p = -np.sqrt(2) / (2 - np.sqrt(2)) weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: p = 0.5 weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: p = 1 / np.sqrt(2) weight = np.array([[[0, -p, -p, 0, -p, 0, 0, 1 - p], [0, 0, 0, -p, 0, -p, p - 1, 0], [0, 0, 0, -p, 0, p - 1, -p, 0], [0, 0, 0, 0, 1 - p, 0, 0, -p], [0, 0, 0, 0, 0, -p, -p, 0], [0, 0, 0, 0, 0, 0, 0, -p], [0, 0, 0, 0, 0, 0, 0, -p ], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) if init in ['squeeze', 'pixel_shuffle', 'zeros']: if method == 'exp' or method == 'cayley': return torch.zeros(*kernel_matrix_shape) if type(init) is np.ndarray: init = torch.tensor(init.astype(dtype)) if torch.is_tensor(init): if len(init.shape) == 2: init = init.reshape(1, *init.shape) if init.shape[0] == 1: init = init.repeat(num_matrices, 1, 1) assert init.shape == kernel_matrix_shape return init else: raise NotImplementedError('Unknown initialization.') class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalResamplingLayer(torch.nn.Module): """Base class for orthogonal up- and downsampling operators. :param low_channel_number: Lower number of channels. These are the input channels in the case of downsampling ops, and the output channels in the case of upsampling ops. :param stride: The downsampling / upsampling factor for each dimension. :param channel_multiplier: The channel multiplier, i.e. the number by which the number of channels are multiplied (downsampling) or divided (upsampling). :param method: Which method to use for parametrizing orthogonal matrices which are used as convolutional kernels. """ def __init__(self, low_channel_number: 'int', stride: 'Union[int, Tuple[int, ...]]', method: 'str'='cayley', init: 'Union[str, np.ndarray, torch.Tensor]'='haar', learnable: 'bool'= True, init_kwargs: 'dict'=None, **kwargs): super(OrthogonalResamplingLayer, self).__init__() self.low_channel_number = low_channel_number self.method = method self.stride = stride self.channel_multiplier = int(np.prod(stride)) self.high_channel_number = self.channel_multiplier * low_channel_number if init_kwargs is None: init_kwargs = {} self.init_kwargs = init_kwargs self.kwargs = kwargs assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self._kernel_matrix_shape = (self.low_channel_number,) + (self. channel_multiplier,) * 2 self._kernel_shape = (self.high_channel_number, 1) + self.stride self.weight = torch.nn.Parameter(__initialize_weight__( kernel_matrix_shape=self._kernel_matrix_shape, stride=self. stride, method=self.method, init=init, **self.init_kwargs)) self.weight.requires_grad = learnable @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel(self): """The kernel associated with the invertible up-/downsampling. """ return self.kernel_matrix.reshape(*self._kernel_shape) class InvertibleDownsampling3D(OrthogonalResamplingLayer): def __init__(self, in_channels: 'int', stride: '_size_3_t'=2, method: 'str'='cayley', init: 'str'='haar', learnable: 'bool'=True, *args, **kwargs): stride = tuple(_triple(stride)) channel_multiplier = int(np.prod(stride)) self.in_channels = in_channels self.out_channels = in_channels * channel_multiplier super(InvertibleDownsampling3D, self).__init__(*args, low_channel_number=self.in_channels, stride=stride, method= method, init=init, learnable=learnable, **kwargs) def forward(self, x): return F.conv3d(x, self.kernel, stride=self.stride, groups=self. low_channel_number) def inverse(self, x): return F.conv_transpose3d(x, self.kernel, stride=self.stride, groups=self.low_channel_number) def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_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.autograd import Function import numpy as np from warnings import warn from typing import Union from typing import Tuple from torch.nn.common_types import _size_3_t from torch.nn.modules.utils import _triple 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_add_eye_sub_0(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 32 xnumel = 8 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 % 8 x2 = xindex y3 = yindex y1 = yindex // 8 tmp6 = tl.load(in_ptr0 + (x2 + 8 * y3), xmask & ymask, eviction_policy= 'evict_last') tmp7 = tl.load(in_ptr0 + (y0 + 8 * x2 + 64 * y1), xmask & ymask, eviction_policy='evict_last') tmp0 = y0 tmp1 = x2 tmp2 = tmp0 == tmp1 tmp3 = 1.0 tmp4 = 0.0 tmp5 = tl.where(tmp2, tmp3, tmp4) tmp8 = tmp6 - tmp7 tmp9 = tmp5 + tmp8 tmp10 = tmp5 - tmp8 tl.store(out_ptr0 + (x2 + 8 * y3), tmp9, xmask & ymask) tl.store(out_ptr1 + (x2 + 8 * y3), tmp10, xmask & ymask) @triton.jit def triton_poi_fused_clone_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl. constexpr, XBLOCK: tl.constexpr): ynumel = 32 xnumel = 8 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 % 8 y1 = yindex // 8 y3 = yindex tmp0 = tl.load(in_ptr0 + (y0 + 8 * x2 + 64 * y1), xmask & ymask, eviction_policy='evict_last') tl.store(out_ptr0 + (x2 + 8 * y3), tmp0, xmask & ymask) def call(args): primals_1, primals_2 = args args.clear() assert_size_stride(primals_1, (4, 8, 8), (64, 8, 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, 8, 8), (64, 8, 1), torch.float32) buf5 = empty_strided_cuda((4, 8, 8), (64, 8, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_eye_sub_0[grid(32, 8)](primals_1, buf0, buf5, 32, 8, XBLOCK=8, YBLOCK=32, num_warps=4, num_stages=1) del primals_1 buf1 = torch.ops.aten.linalg_lu_factor_ex.default(buf0) buf2 = buf1[0] buf3 = buf1[1] del buf1 buf6 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf5) buf7 = buf6 del buf6 buf8 = buf0 del buf0 triton_poi_fused_clone_1[grid(32, 8)](buf7, buf8, 32, 8, XBLOCK=8, YBLOCK=32, num_warps=4, num_stages=1) buf9 = extern_kernels.convolution(reinterpret_tensor(primals_2, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1), 0), reinterpret_tensor(buf8, (32, 1, 2, 2, 2), (8, 0, 4, 2, 1), 0), stride=(2, 2, 2), padding=(0, 0, 0), dilation=(1, 1, 1), transposed=False, output_padding=(0, 0, 0), groups=4, bias=None) assert_size_stride(buf9, (1, 32, 2, 2, 2), (256, 8, 4, 2, 1)) buf10 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf7) del buf2 del buf3 del buf7 buf11 = buf10 del buf10 return reinterpret_tensor(buf9, (32, 2, 2, 2), (8, 4, 2, 1), 0 ), buf5, reinterpret_tensor(buf8, (32, 1, 2, 2, 2), (8, 8, 4, 2, 1), 0 ), reinterpret_tensor(primals_2, (1, 4, 4, 4, 4), (256, 64, 16, 4, 1), 0), buf11 def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) def __initialize_weight__(kernel_matrix_shape: 'Tuple[int, ...]', stride: 'Tuple[int, ...]', method: 'str'='cayley', init: 'str'='haar', dtype: 'str'='float32', *args, **kwargs): """Function which computes specific orthogonal matrices. For some chosen method of parametrizing orthogonal matrices, this function outputs the required weights necessary to represent a chosen initialization as a Pytorch tensor of matrices. Args: kernel_matrix_shape : The output shape of the orthogonal matrices. Should be (num_matrices, height, width). stride : The stride for the invertible up- or downsampling for which this matrix is to be used. The length of ``stride`` should match the dimensionality of the data. method : The method for parametrising orthogonal matrices. Should be 'exp' or 'cayley' init : The matrix which should be represented. Should be 'squeeze', 'pixel_shuffle', 'haar' or 'random'. 'haar' is only possible if ``stride`` is only 2. dtype : Numpy dtype which should be used for the matrix. *args: Variable length argument iterable. **kwargs: Arbitrary keyword arguments. Returns: Tensor : Orthogonal matrices of shape ``kernel_matrix_shape``. """ dim = len(stride) num_matrices = kernel_matrix_shape[0] assert method in ['exp', 'cayley', 'householder'] if method == 'householder': warn( 'Householder parametrization not fully implemented yet. Only random initialization currently working.' ) init = 'random' if init == 'random': return torch.randn(kernel_matrix_shape) if init == 'haar' and set(stride) != {2}: None None init = 'squeeze' if init == 'haar' and set(stride) == {2}: if method == 'exp': p = np.pi / 4 if dim == 1: weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: weight = np.array([[[0, p, p, 0, p, 0, 0, 0], [0, 0, 0, p, 0, p, 0, 0], [0, 0, 0, p, 0, 0, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, p, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) elif method == 'cayley': if dim == 1: p = -np.sqrt(2) / (2 - np.sqrt(2)) weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: p = 0.5 weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: p = 1 / np.sqrt(2) weight = np.array([[[0, -p, -p, 0, -p, 0, 0, 1 - p], [0, 0, 0, -p, 0, -p, p - 1, 0], [0, 0, 0, -p, 0, p - 1, -p, 0], [0, 0, 0, 0, 1 - p, 0, 0, -p], [0, 0, 0, 0, 0, -p, -p, 0], [0, 0, 0, 0, 0, 0, 0, -p], [0, 0, 0, 0, 0, 0, 0, -p ], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) if init in ['squeeze', 'pixel_shuffle', 'zeros']: if method == 'exp' or method == 'cayley': return torch.zeros(*kernel_matrix_shape) if type(init) is np.ndarray: init = torch.tensor(init.astype(dtype)) if torch.is_tensor(init): if len(init.shape) == 2: init = init.reshape(1, *init.shape) if init.shape[0] == 1: init = init.repeat(num_matrices, 1, 1) assert init.shape == kernel_matrix_shape return init else: raise NotImplementedError('Unknown initialization.') class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalResamplingLayer(torch.nn.Module): """Base class for orthogonal up- and downsampling operators. :param low_channel_number: Lower number of channels. These are the input channels in the case of downsampling ops, and the output channels in the case of upsampling ops. :param stride: The downsampling / upsampling factor for each dimension. :param channel_multiplier: The channel multiplier, i.e. the number by which the number of channels are multiplied (downsampling) or divided (upsampling). :param method: Which method to use for parametrizing orthogonal matrices which are used as convolutional kernels. """ def __init__(self, low_channel_number: 'int', stride: 'Union[int, Tuple[int, ...]]', method: 'str'='cayley', init: 'Union[str, np.ndarray, torch.Tensor]'='haar', learnable: 'bool'= True, init_kwargs: 'dict'=None, **kwargs): super(OrthogonalResamplingLayer, self).__init__() self.low_channel_number = low_channel_number self.method = method self.stride = stride self.channel_multiplier = int(np.prod(stride)) self.high_channel_number = self.channel_multiplier * low_channel_number if init_kwargs is None: init_kwargs = {} self.init_kwargs = init_kwargs self.kwargs = kwargs assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self._kernel_matrix_shape = (self.low_channel_number,) + (self. channel_multiplier,) * 2 self._kernel_shape = (self.high_channel_number, 1) + self.stride self.weight = torch.nn.Parameter(__initialize_weight__( kernel_matrix_shape=self._kernel_matrix_shape, stride=self. stride, method=self.method, init=init, **self.init_kwargs)) self.weight.requires_grad = learnable @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel(self): """The kernel associated with the invertible up-/downsampling. """ return self.kernel_matrix.reshape(*self._kernel_shape) class InvertibleDownsampling3DNew(OrthogonalResamplingLayer): def __init__(self, in_channels: 'int', stride: '_size_3_t'=2, method: 'str'='cayley', init: 'str'='haar', learnable: 'bool'=True, *args, **kwargs): stride = tuple(_triple(stride)) channel_multiplier = int(np.prod(stride)) self.in_channels = in_channels self.out_channels = in_channels * channel_multiplier super(InvertibleDownsampling3DNew, self).__init__(*args, low_channel_number=self.in_channels, stride=stride, method= method, init=init, learnable=learnable, **kwargs) def inverse(self, x): return F.conv_transpose3d(x, self.kernel, stride=self.stride, groups=self.low_channel_number) def forward(self, input_0): primals_1 = self.weight primals_2 = input_0 output = call([primals_1, primals_2]) return output[0]
cetmann/iunets
InvertibleDownsampling3D
false
15,027
[ "MIT" ]
86
80ed7cce0e505a0396c42359eaf27819222d71f6
https://github.com/cetmann/iunets/tree/80ed7cce0e505a0396c42359eaf27819222d71f6
ToyNet
import torch import torch.nn as nn import torch.nn.functional as F class ToyNet(nn.Module): def __init__(self): super(ToyNet, self).__init__() self.conv1 = nn.Conv2d(3, 6, 5) self.pool = nn.MaxPool2d(2, 2) self.conv2 = nn.Conv2d(6, 16, 5) self.conv3 = nn.Conv2d(16, 64, 3) self.conv4 = nn.Conv2d(64, 256, 3) self.fc0 = nn.Linear(256 * 11 * 11, 2048) self.fc1 = nn.Linear(2048, 512) self.fc2 = nn.Linear(512, 128) self.fc3 = nn.Linear(128, 2) 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 = self.pool(F.relu(self.conv4(x))) x = x.view(-1, 256 * 11 * 11) x = F.relu(self.fc0(x)) x = F.relu(self.fc1(x)) x = F.relu(self.fc2(x)) x = self.fc3(x) return x def get_inputs(): return [torch.rand([4, 3, 216, 216])] 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 = 18 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 = 46656 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 + 46656 * y3), xmask & ymask, eviction_policy='evict_last') tl.store(out_ptr0 + (y0 + 3 * x2 + 139968 * y1), tmp0, xmask & ymask) @triton.jit def triton_poi_fused_2(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 % 6 y1 = yindex // 6 tmp0 = tl.load(in_ptr0 + (x2 + 25 * y3), xmask & ymask, eviction_policy ='evict_last') tl.store(out_ptr0 + (y0 + 6 * x2 + 150 * y1), tmp0, xmask & ymask) @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 % 16 y1 = yindex // 16 tmp0 = tl.load(in_ptr0 + (x2 + 9 * y3), xmask, eviction_policy='evict_last' ) tl.store(out_ptr0 + (y0 + 16 * x2 + 144 * 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 % 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_convolution_relu_5(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 1078656 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x2 = xindex x0 = xindex % 6 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_max_pool2d_with_indices_6(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr): xnumel = 269664 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex % 6 x1 = xindex // 6 % 106 x2 = xindex // 636 x3 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 12 * x1 + 2544 * x2), xmask) tmp1 = tl.load(in_ptr0 + (6 + x0 + 12 * x1 + 2544 * x2), xmask) tmp3 = tl.load(in_ptr0 + (1272 + x0 + 12 * x1 + 2544 * x2), xmask) tmp5 = tl.load(in_ptr0 + (1278 + x0 + 12 * x1 + 2544 * x2), xmask) 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, xmask) tl.store(out_ptr1 + x3, tmp16, xmask) @triton.jit def triton_poi_fused_convolution_relu_7(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 665856 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 = 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_max_pool2d_with_indices_8(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr): xnumel = 166464 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex % 16 x1 = xindex // 16 % 51 x2 = xindex // 816 x3 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 32 * x1 + 3264 * x2), xmask) tmp1 = tl.load(in_ptr0 + (16 + x0 + 32 * x1 + 3264 * x2), xmask) tmp3 = tl.load(in_ptr0 + (1632 + x0 + 32 * x1 + 3264 * x2), xmask) tmp5 = tl.load(in_ptr0 + (1648 + x0 + 32 * x1 + 3264 * x2), xmask) 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, xmask) tl.store(out_ptr1 + x3, tmp16, xmask) @triton.jit def triton_poi_fused_convolution_relu_9(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 614656 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_max_pool2d_with_indices_10(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 % 24 x2 = xindex // 1536 % 24 x3 = xindex // 36864 x4 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 128 * x1 + 6272 * x2 + 153664 * x3), None) tmp1 = tl.load(in_ptr0 + (64 + x0 + 128 * x1 + 6272 * x2 + 153664 * x3), None) tmp3 = tl.load(in_ptr0 + (3136 + x0 + 128 * x1 + 6272 * x2 + 153664 * x3), None) tmp5 = tl.load(in_ptr0 + (3200 + x0 + 128 * x1 + 6272 * x2 + 153664 * x3), 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 + x4, tmp6, None) tl.store(out_ptr1 + x4, tmp16, None) @triton.jit def triton_poi_fused_convolution_relu_11(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 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_12(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 484 xnumel = 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] xmask = xindex < xnumel x2 = xindex y0 = yindex % 11 y1 = yindex // 11 y5 = yindex y4 = yindex // 121 y6 = yindex % 121 tmp0 = tl.load(in_ptr0 + (x2 + 512 * y0 + 11264 * y1), xmask & ymask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr0 + (256 + x2 + 512 * y0 + 11264 * y1), xmask & ymask, eviction_policy='evict_last') tmp7 = tl.load(in_ptr0 + (5632 + x2 + 512 * y0 + 11264 * y1), xmask & ymask, eviction_policy='evict_last') tmp12 = tl.load(in_ptr0 + (5888 + x2 + 512 * y0 + 11264 * y1), xmask & ymask, eviction_policy='evict_last') tmp2 = tmp1 > tmp0 tmp3 = tl.full([1, 1], 1, tl.int8) tmp4 = tl.full([1, 1], 0, tl.int8) tmp5 = tl.where(tmp2, tmp3, tmp4) tmp6 = triton_helpers.maximum(tmp1, tmp0) tmp8 = tmp7 > tmp6 tmp9 = tl.full([1, 1], 2, tl.int8) tmp10 = tl.where(tmp8, tmp9, tmp5) tmp11 = triton_helpers.maximum(tmp7, tmp6) tmp13 = tmp12 > tmp11 tmp14 = tl.full([1, 1], 3, tl.int8) tmp15 = tl.where(tmp13, tmp14, tmp10) tmp16 = triton_helpers.maximum(tmp12, tmp11) tl.store(out_ptr0 + (x2 + 256 * y5), tmp15, xmask & ymask) tl.store(out_ptr1 + (y6 + 121 * x2 + 30976 * y4), tmp16, xmask & ymask) @triton.jit def triton_poi_fused_relu_13(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 % 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) tl.store(in_out_ptr0 + x2, tmp4, None) @triton.jit def triton_poi_fused_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 % 512 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_15(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr ): xnumel = 512 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) 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) = args args.clear() assert_size_stride(primals_1, (6, 3, 5, 5), (75, 25, 5, 1)) assert_size_stride(primals_2, (6,), (1,)) assert_size_stride(primals_3, (4, 3, 216, 216), (139968, 46656, 216, 1)) assert_size_stride(primals_4, (16, 6, 5, 5), (150, 25, 5, 1)) assert_size_stride(primals_5, (16,), (1,)) assert_size_stride(primals_6, (64, 16, 3, 3), (144, 9, 3, 1)) assert_size_stride(primals_7, (64,), (1,)) assert_size_stride(primals_8, (256, 64, 3, 3), (576, 9, 3, 1)) assert_size_stride(primals_9, (256,), (1,)) assert_size_stride(primals_10, (2048, 30976), (30976, 1)) assert_size_stride(primals_11, (2048,), (1,)) assert_size_stride(primals_12, (512, 2048), (2048, 1)) assert_size_stride(primals_13, (512,), (1,)) assert_size_stride(primals_14, (128, 512), (512, 1)) assert_size_stride(primals_15, (128,), (1,)) assert_size_stride(primals_16, (2, 128), (128, 1)) assert_size_stride(primals_17, (2,), (1,)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((6, 3, 5, 5), (75, 1, 15, 3), torch.float32) get_raw_stream(0) triton_poi_fused_0[grid(18, 25)](primals_1, buf0, 18, 25, XBLOCK=32, YBLOCK=32, num_warps=4, num_stages=1) del primals_1 buf1 = empty_strided_cuda((4, 3, 216, 216), (139968, 1, 648, 3), torch.float32) triton_poi_fused_1[grid(12, 46656)](primals_3, buf1, 12, 46656, XBLOCK=64, YBLOCK=16, num_warps=4, num_stages=1) del primals_3 buf2 = empty_strided_cuda((16, 6, 5, 5), (150, 1, 30, 6), torch.float32 ) triton_poi_fused_2[grid(96, 25)](primals_4, buf2, 96, 25, XBLOCK=32, YBLOCK=32, num_warps=4, num_stages=1) del primals_4 buf3 = empty_strided_cuda((64, 16, 3, 3), (144, 1, 48, 16), torch. float32) triton_poi_fused_3[grid(1024, 9)](primals_6, buf3, 1024, 9, XBLOCK= 16, YBLOCK=64, num_warps=4, num_stages=1) del primals_6 buf4 = empty_strided_cuda((256, 64, 3, 3), (576, 1, 192, 64), torch .float32) triton_poi_fused_4[grid(16384, 9)](primals_8, buf4, 16384, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_8 buf5 = extern_kernels.convolution(buf1, buf0, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf5, (4, 6, 212, 212), (269664, 1, 1272, 6)) buf6 = buf5 del buf5 triton_poi_fused_convolution_relu_5[grid(1078656)](buf6, primals_2, 1078656, XBLOCK=1024, num_warps=4, num_stages=1) del primals_2 buf7 = empty_strided_cuda((4, 6, 106, 106), (67416, 1, 636, 6), torch.float32) buf8 = empty_strided_cuda((4, 6, 106, 106), (67416, 1, 636, 6), torch.int8) triton_poi_fused_max_pool2d_with_indices_6[grid(269664)](buf6, buf7, buf8, 269664, XBLOCK=512, num_warps=8, num_stages=1) buf9 = extern_kernels.convolution(buf7, buf2, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf9, (4, 16, 102, 102), (166464, 1, 1632, 16)) buf10 = buf9 del buf9 triton_poi_fused_convolution_relu_7[grid(665856)](buf10, primals_5, 665856, XBLOCK=1024, num_warps=4, num_stages=1) del primals_5 buf11 = empty_strided_cuda((4, 16, 51, 51), (41616, 1, 816, 16), torch.float32) buf12 = empty_strided_cuda((4, 16, 51, 51), (41616, 1, 816, 16), torch.int8) triton_poi_fused_max_pool2d_with_indices_8[grid(166464)](buf10, buf11, buf12, 166464, XBLOCK=512, num_warps=8, num_stages=1) buf13 = extern_kernels.convolution(buf11, buf3, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf13, (4, 64, 49, 49), (153664, 1, 3136, 64)) buf14 = buf13 del buf13 triton_poi_fused_convolution_relu_9[grid(614656)](buf14, primals_7, 614656, XBLOCK=1024, num_warps=4, num_stages=1) del primals_7 buf15 = empty_strided_cuda((4, 64, 24, 24), (36864, 1, 1536, 64), torch.float32) buf16 = empty_strided_cuda((4, 64, 24, 24), (36864, 1, 1536, 64), torch.int8) triton_poi_fused_max_pool2d_with_indices_10[grid(147456)](buf14, buf15, buf16, 147456, XBLOCK=1024, num_warps=4, num_stages=1) buf17 = extern_kernels.convolution(buf15, buf4, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf17, (4, 256, 22, 22), (123904, 1, 5632, 256)) buf18 = buf17 del buf17 triton_poi_fused_convolution_relu_11[grid(495616)](buf18, primals_9, 495616, XBLOCK=512, num_warps=8, num_stages=1) del primals_9 buf19 = empty_strided_cuda((4, 256, 11, 11), (30976, 1, 2816, 256), torch.int8) buf20 = empty_strided_cuda((4, 256, 11, 11), (30976, 121, 11, 1), torch.float32) triton_poi_fused_max_pool2d_with_indices_12[grid(484, 256)](buf18, buf19, buf20, 484, 256, XBLOCK=256, YBLOCK=4, num_warps=4, num_stages=1) buf21 = empty_strided_cuda((4, 2048), (2048, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(buf20, (4, 30976), (30976, 1), 0), reinterpret_tensor(primals_10, (30976, 2048), (1, 30976), 0 ), out=buf21) buf22 = buf21 del buf21 triton_poi_fused_relu_13[grid(8192)](buf22, primals_11, 8192, XBLOCK=256, num_warps=4, num_stages=1) del primals_11 buf23 = empty_strided_cuda((4, 512), (512, 1), torch.float32) extern_kernels.mm(buf22, reinterpret_tensor(primals_12, (2048, 512), (1, 2048), 0), out=buf23) buf24 = buf23 del buf23 triton_poi_fused_relu_14[grid(2048)](buf24, primals_13, 2048, XBLOCK=256, num_warps=4, num_stages=1) del primals_13 buf25 = empty_strided_cuda((4, 128), (128, 1), torch.float32) extern_kernels.mm(buf24, reinterpret_tensor(primals_14, (512, 128), (1, 512), 0), out=buf25) buf26 = buf25 del buf25 triton_poi_fused_relu_15[grid(512)](buf26, primals_15, 512, XBLOCK= 256, num_warps=4, num_stages=1) del primals_15 buf27 = empty_strided_cuda((4, 2), (2, 1), torch.float32) extern_kernels.addmm(primals_17, buf26, reinterpret_tensor( primals_16, (128, 2), (1, 128), 0), alpha=1, beta=1, out=buf27) del primals_17 return (buf27, buf0, buf1, buf2, buf3, buf4, buf6, buf7, buf8, buf10, buf11, buf12, buf14, buf15, buf16, buf18, buf19, reinterpret_tensor (buf20, (4, 30976), (30976, 1), 0), buf22, buf24, buf26, primals_16, primals_14, primals_12, primals_10) class ToyNetNew(nn.Module): def __init__(self): super(ToyNetNew, self).__init__() self.conv1 = nn.Conv2d(3, 6, 5) self.pool = nn.MaxPool2d(2, 2) self.conv2 = nn.Conv2d(6, 16, 5) self.conv3 = nn.Conv2d(16, 64, 3) self.conv4 = nn.Conv2d(64, 256, 3) self.fc0 = nn.Linear(256 * 11 * 11, 2048) self.fc1 = nn.Linear(2048, 512) self.fc2 = nn.Linear(512, 128) self.fc3 = nn.Linear(128, 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.conv4.weight primals_9 = self.conv4.bias primals_10 = self.fc0.weight primals_11 = self.fc0.bias primals_12 = self.fc1.weight primals_13 = self.fc1.bias primals_14 = self.fc2.weight primals_15 = self.fc2.bias primals_16 = self.fc3.weight primals_17 = 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, primals_14, primals_15, primals_16, primals_17]) return output[0]
asalmanp/MIVisionX
ToyNet
false
15,028
[ "MIT" ]
153
a964774944331827c8d6e9bb1ffbb2578f335056
https://github.com/asalmanp/MIVisionX/tree/a964774944331827c8d6e9bb1ffbb2578f335056
MS_Block
import torch import torch.nn as nn import torch.multiprocessing class MS_Block(nn.Module): def __init__(self, input_feature, out_feature, d=[1, 2, 4], group=1): super(MS_Block, self).__init__() self.l1 = nn.Conv2d(input_feature, out_feature, 3, padding=d[0], dilation=d[0], bias=False, groups=group) self.l2 = nn.Conv2d(input_feature, out_feature, 3, padding=d[1], dilation=d[1], bias=False, groups=group) self.l3 = nn.Conv2d(input_feature, out_feature, 3, padding=d[2], dilation=d[2], bias=False, groups=group) def forward(self, x): out = self.l1(x) + self.l2(x) + self.l3(x) return out def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'input_feature': 4, 'out_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 import torch.multiprocessing assert_size_stride = torch._C._dynamo.guards.assert_size_stride @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 x0 = xindex tmp0 = tl.load(in_out_ptr0 + x0, xmask) tmp1 = tl.load(in_ptr0 + x0, xmask) tmp3 = tl.load(in_ptr1 + x0, xmask) tmp2 = tmp0 + tmp1 tmp4 = tmp2 + tmp3 tl.store(in_out_ptr0 + x0, tmp4, xmask) def call(args): primals_1, primals_2, primals_3, primals_4 = args args.clear() assert_size_stride(primals_1, (4, 4, 3, 3), (36, 9, 3, 1)) assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_3, (4, 4, 3, 3), (36, 9, 3, 1)) assert_size_stride(primals_4, (4, 4, 3, 3), (36, 9, 3, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = extern_kernels.convolution(primals_2, primals_1, stride=(1, 1), padding=(1, 1), 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_2, primals_3, stride=(1, 1), padding=(2, 2), dilation=(2, 2), 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_2, primals_4, stride=(1, 1), padding=(4, 4), dilation=(4, 4), 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_add_0[grid(256)](buf3, buf1, buf2, 256, XBLOCK=256, num_warps=4, num_stages=1) del buf1 del buf2 return buf3, primals_1, primals_2, primals_3, primals_4 class MS_BlockNew(nn.Module): def __init__(self, input_feature, out_feature, d=[1, 2, 4], group=1): super(MS_BlockNew, self).__init__() self.l1 = nn.Conv2d(input_feature, out_feature, 3, padding=d[0], dilation=d[0], bias=False, groups=group) self.l2 = nn.Conv2d(input_feature, out_feature, 3, padding=d[1], dilation=d[1], bias=False, groups=group) self.l3 = nn.Conv2d(input_feature, out_feature, 3, padding=d[2], dilation=d[2], bias=False, groups=group) def forward(self, input_0): primals_1 = self.l1.weight primals_3 = self.l2.weight primals_4 = self.l3.weight primals_2 = input_0 output = call([primals_1, primals_2, primals_3, primals_4]) return output[0]
chiukin/RANet
MS_Block
false
15,029
[ "Apache-2.0" ]
267
681a47d9b1f114653290678f02f2d3ecdf4010bc
https://github.com/chiukin/RANet/tree/681a47d9b1f114653290678f02f2d3ecdf4010bc
InvertibleUpsampling2D
from torch.autograd import Function import torch import numpy as np from warnings import warn from typing import Union from typing import Tuple from torch.nn.common_types import _size_2_t from torch.nn.modules.utils import _pair import torch.nn.functional as F def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) def __initialize_weight__(kernel_matrix_shape: 'Tuple[int, ...]', stride: 'Tuple[int, ...]', method: 'str'='cayley', init: 'str'='haar', dtype: 'str'='float32', *args, **kwargs): """Function which computes specific orthogonal matrices. For some chosen method of parametrizing orthogonal matrices, this function outputs the required weights necessary to represent a chosen initialization as a Pytorch tensor of matrices. Args: kernel_matrix_shape : The output shape of the orthogonal matrices. Should be (num_matrices, height, width). stride : The stride for the invertible up- or downsampling for which this matrix is to be used. The length of ``stride`` should match the dimensionality of the data. method : The method for parametrising orthogonal matrices. Should be 'exp' or 'cayley' init : The matrix which should be represented. Should be 'squeeze', 'pixel_shuffle', 'haar' or 'random'. 'haar' is only possible if ``stride`` is only 2. dtype : Numpy dtype which should be used for the matrix. *args: Variable length argument iterable. **kwargs: Arbitrary keyword arguments. Returns: Tensor : Orthogonal matrices of shape ``kernel_matrix_shape``. """ dim = len(stride) num_matrices = kernel_matrix_shape[0] assert method in ['exp', 'cayley', 'householder'] if method == 'householder': warn( 'Householder parametrization not fully implemented yet. Only random initialization currently working.' ) init = 'random' if init == 'random': return torch.randn(kernel_matrix_shape) if init == 'haar' and set(stride) != {2}: None None init = 'squeeze' if init == 'haar' and set(stride) == {2}: if method == 'exp': p = np.pi / 4 if dim == 1: weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: weight = np.array([[[0, p, p, 0, p, 0, 0, 0], [0, 0, 0, p, 0, p, 0, 0], [0, 0, 0, p, 0, 0, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, p, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) elif method == 'cayley': if dim == 1: p = -np.sqrt(2) / (2 - np.sqrt(2)) weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: p = 0.5 weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: p = 1 / np.sqrt(2) weight = np.array([[[0, -p, -p, 0, -p, 0, 0, 1 - p], [0, 0, 0, -p, 0, -p, p - 1, 0], [0, 0, 0, -p, 0, p - 1, -p, 0], [0, 0, 0, 0, 1 - p, 0, 0, -p], [0, 0, 0, 0, 0, -p, -p, 0], [0, 0, 0, 0, 0, 0, 0, -p], [0, 0, 0, 0, 0, 0, 0, -p ], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) if init in ['squeeze', 'pixel_shuffle', 'zeros']: if method == 'exp' or method == 'cayley': return torch.zeros(*kernel_matrix_shape) if type(init) is np.ndarray: init = torch.tensor(init.astype(dtype)) if torch.is_tensor(init): if len(init.shape) == 2: init = init.reshape(1, *init.shape) if init.shape[0] == 1: init = init.repeat(num_matrices, 1, 1) assert init.shape == kernel_matrix_shape return init else: raise NotImplementedError('Unknown initialization.') class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalResamplingLayer(torch.nn.Module): """Base class for orthogonal up- and downsampling operators. :param low_channel_number: Lower number of channels. These are the input channels in the case of downsampling ops, and the output channels in the case of upsampling ops. :param stride: The downsampling / upsampling factor for each dimension. :param channel_multiplier: The channel multiplier, i.e. the number by which the number of channels are multiplied (downsampling) or divided (upsampling). :param method: Which method to use for parametrizing orthogonal matrices which are used as convolutional kernels. """ def __init__(self, low_channel_number: 'int', stride: 'Union[int, Tuple[int, ...]]', method: 'str'='cayley', init: 'Union[str, np.ndarray, torch.Tensor]'='haar', learnable: 'bool'= True, init_kwargs: 'dict'=None, **kwargs): super(OrthogonalResamplingLayer, self).__init__() self.low_channel_number = low_channel_number self.method = method self.stride = stride self.channel_multiplier = int(np.prod(stride)) self.high_channel_number = self.channel_multiplier * low_channel_number if init_kwargs is None: init_kwargs = {} self.init_kwargs = init_kwargs self.kwargs = kwargs assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self._kernel_matrix_shape = (self.low_channel_number,) + (self. channel_multiplier,) * 2 self._kernel_shape = (self.high_channel_number, 1) + self.stride self.weight = torch.nn.Parameter(__initialize_weight__( kernel_matrix_shape=self._kernel_matrix_shape, stride=self. stride, method=self.method, init=init, **self.init_kwargs)) self.weight.requires_grad = learnable @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel(self): """The kernel associated with the invertible up-/downsampling. """ return self.kernel_matrix.reshape(*self._kernel_shape) class InvertibleUpsampling2D(OrthogonalResamplingLayer): def __init__(self, in_channels: 'int', stride: '_size_2_t'=2, method: 'str'='cayley', init: 'str'='haar', learnable: 'bool'=True, *args, **kwargs): stride = tuple(_pair(stride)) channel_multiplier = int(np.prod(stride)) self.in_channels = in_channels self.out_channels = in_channels // channel_multiplier super(InvertibleUpsampling2D, self).__init__(*args, low_channel_number=self.out_channels, stride=stride, method= method, init=init, learnable=learnable, **kwargs) def forward(self, x): return F.conv_transpose2d(x, self.kernel, stride=self.stride, groups=self.low_channel_number) def inverse(self, x): return F.conv2d(x, self.kernel, stride=self.stride, groups=self. low_channel_number) def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_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.autograd import Function import numpy as np from warnings import warn from typing import Union from typing import Tuple from torch.nn.common_types import _size_2_t from torch.nn.modules.utils import _pair 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_add_eye_sub_0(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 4 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 x1 = xindex tmp6 = tl.load(in_ptr0 + (x1 + 4 * y0), xmask & ymask) tmp7 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask) tmp0 = y0 tmp1 = x1 tmp2 = tmp0 == tmp1 tmp3 = 1.0 tmp4 = 0.0 tmp5 = tl.where(tmp2, tmp3, tmp4) tmp8 = tmp6 - tmp7 tmp9 = tmp5 + tmp8 tmp10 = tmp5 - tmp8 tl.store(out_ptr0 + (x1 + 4 * y0), tmp9, xmask & ymask) tl.store(out_ptr1 + (x1 + 4 * y0), tmp10, xmask & ymask) @triton.jit def triton_poi_fused_convolution_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 4 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 x1 = xindex y0 = yindex tmp0 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask) tl.store(out_ptr0 + (x1 + 4 * y0), tmp0, xmask & ymask) def call(args): primals_1, primals_2 = args args.clear() assert_size_stride(primals_1, (1, 4, 4), (16, 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((1, 4, 4), (16, 4, 1), torch.float32) buf5 = empty_strided_cuda((1, 4, 4), (16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_eye_sub_0[grid(4, 4)](primals_1, buf0, buf5, 4, 4, XBLOCK=4, YBLOCK=4, num_warps=1, num_stages=1) del primals_1 buf1 = torch.ops.aten.linalg_lu_factor_ex.default(buf0) buf2 = buf1[0] buf3 = buf1[1] del buf1 buf6 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf5) buf7 = buf6 del buf6 buf8 = reinterpret_tensor(buf0, (4, 1, 2, 2), (4, 4, 2, 1), 0) del buf0 triton_poi_fused_convolution_1[grid(4, 4)](buf7, buf8, 4, 4, XBLOCK =4, YBLOCK=4, num_warps=1, num_stages=1) buf9 = extern_kernels.convolution(primals_2, buf8, stride=(2, 2), padding=(0, 0), dilation=(1, 1), transposed=True, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf9, (4, 1, 8, 8), (64, 64, 8, 1)) del buf8 buf10 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf7) del buf2 del buf3 buf11 = buf10 del buf10 return buf9, primals_2, buf5, reinterpret_tensor(buf7, (4, 1, 2, 2), (1, 16, 8, 4), 0), buf11 def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) def __initialize_weight__(kernel_matrix_shape: 'Tuple[int, ...]', stride: 'Tuple[int, ...]', method: 'str'='cayley', init: 'str'='haar', dtype: 'str'='float32', *args, **kwargs): """Function which computes specific orthogonal matrices. For some chosen method of parametrizing orthogonal matrices, this function outputs the required weights necessary to represent a chosen initialization as a Pytorch tensor of matrices. Args: kernel_matrix_shape : The output shape of the orthogonal matrices. Should be (num_matrices, height, width). stride : The stride for the invertible up- or downsampling for which this matrix is to be used. The length of ``stride`` should match the dimensionality of the data. method : The method for parametrising orthogonal matrices. Should be 'exp' or 'cayley' init : The matrix which should be represented. Should be 'squeeze', 'pixel_shuffle', 'haar' or 'random'. 'haar' is only possible if ``stride`` is only 2. dtype : Numpy dtype which should be used for the matrix. *args: Variable length argument iterable. **kwargs: Arbitrary keyword arguments. Returns: Tensor : Orthogonal matrices of shape ``kernel_matrix_shape``. """ dim = len(stride) num_matrices = kernel_matrix_shape[0] assert method in ['exp', 'cayley', 'householder'] if method == 'householder': warn( 'Householder parametrization not fully implemented yet. Only random initialization currently working.' ) init = 'random' if init == 'random': return torch.randn(kernel_matrix_shape) if init == 'haar' and set(stride) != {2}: None None init = 'squeeze' if init == 'haar' and set(stride) == {2}: if method == 'exp': p = np.pi / 4 if dim == 1: weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: weight = np.array([[[0, p, p, 0, p, 0, 0, 0], [0, 0, 0, p, 0, p, 0, 0], [0, 0, 0, p, 0, 0, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, p, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) elif method == 'cayley': if dim == 1: p = -np.sqrt(2) / (2 - np.sqrt(2)) weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: p = 0.5 weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: p = 1 / np.sqrt(2) weight = np.array([[[0, -p, -p, 0, -p, 0, 0, 1 - p], [0, 0, 0, -p, 0, -p, p - 1, 0], [0, 0, 0, -p, 0, p - 1, -p, 0], [0, 0, 0, 0, 1 - p, 0, 0, -p], [0, 0, 0, 0, 0, -p, -p, 0], [0, 0, 0, 0, 0, 0, 0, -p], [0, 0, 0, 0, 0, 0, 0, -p ], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) if init in ['squeeze', 'pixel_shuffle', 'zeros']: if method == 'exp' or method == 'cayley': return torch.zeros(*kernel_matrix_shape) if type(init) is np.ndarray: init = torch.tensor(init.astype(dtype)) if torch.is_tensor(init): if len(init.shape) == 2: init = init.reshape(1, *init.shape) if init.shape[0] == 1: init = init.repeat(num_matrices, 1, 1) assert init.shape == kernel_matrix_shape return init else: raise NotImplementedError('Unknown initialization.') class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalResamplingLayer(torch.nn.Module): """Base class for orthogonal up- and downsampling operators. :param low_channel_number: Lower number of channels. These are the input channels in the case of downsampling ops, and the output channels in the case of upsampling ops. :param stride: The downsampling / upsampling factor for each dimension. :param channel_multiplier: The channel multiplier, i.e. the number by which the number of channels are multiplied (downsampling) or divided (upsampling). :param method: Which method to use for parametrizing orthogonal matrices which are used as convolutional kernels. """ def __init__(self, low_channel_number: 'int', stride: 'Union[int, Tuple[int, ...]]', method: 'str'='cayley', init: 'Union[str, np.ndarray, torch.Tensor]'='haar', learnable: 'bool'= True, init_kwargs: 'dict'=None, **kwargs): super(OrthogonalResamplingLayer, self).__init__() self.low_channel_number = low_channel_number self.method = method self.stride = stride self.channel_multiplier = int(np.prod(stride)) self.high_channel_number = self.channel_multiplier * low_channel_number if init_kwargs is None: init_kwargs = {} self.init_kwargs = init_kwargs self.kwargs = kwargs assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self._kernel_matrix_shape = (self.low_channel_number,) + (self. channel_multiplier,) * 2 self._kernel_shape = (self.high_channel_number, 1) + self.stride self.weight = torch.nn.Parameter(__initialize_weight__( kernel_matrix_shape=self._kernel_matrix_shape, stride=self. stride, method=self.method, init=init, **self.init_kwargs)) self.weight.requires_grad = learnable @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel(self): """The kernel associated with the invertible up-/downsampling. """ return self.kernel_matrix.reshape(*self._kernel_shape) class InvertibleUpsampling2DNew(OrthogonalResamplingLayer): def __init__(self, in_channels: 'int', stride: '_size_2_t'=2, method: 'str'='cayley', init: 'str'='haar', learnable: 'bool'=True, *args, **kwargs): stride = tuple(_pair(stride)) channel_multiplier = int(np.prod(stride)) self.in_channels = in_channels self.out_channels = in_channels // channel_multiplier super(InvertibleUpsampling2DNew, self).__init__(*args, low_channel_number=self.out_channels, stride=stride, method= method, init=init, learnable=learnable, **kwargs) def inverse(self, x): return F.conv2d(x, self.kernel, stride=self.stride, groups=self. low_channel_number) def forward(self, input_0): primals_1 = self.weight primals_2 = input_0 output = call([primals_1, primals_2]) return output[0]
cetmann/iunets
InvertibleUpsampling2D
false
15,030
[ "MIT" ]
86
80ed7cce0e505a0396c42359eaf27819222d71f6
https://github.com/cetmann/iunets/tree/80ed7cce0e505a0396c42359eaf27819222d71f6
EPE
import torch from torch import nn import torch.utils.cpp_extension class EPE(nn.Module): def __init__(self): super(EPE, self).__init__() def forward(self, flow, gt, loss_mask): loss_map = (flow - gt.detach()) ** 2 loss_map = (loss_map.sum(1, True) + 1e-06) ** 0.5 return loss_map * loss_mask 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 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 import torch.utils.cpp_extension 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_pow_sub_sum_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 % 16 x1 = xindex // 16 x2 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask) tmp1 = tl.load(in_ptr1 + (x0 + 64 * x1), xmask) tmp4 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask) tmp5 = tl.load(in_ptr1 + (16 + x0 + 64 * x1), xmask) tmp9 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask) tmp10 = tl.load(in_ptr1 + (32 + x0 + 64 * x1), xmask) tmp14 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask) tmp15 = tl.load(in_ptr1 + (48 + x0 + 64 * x1), xmask) tmp2 = tmp0 - tmp1 tmp3 = tmp2 * tmp2 tmp6 = tmp4 - tmp5 tmp7 = tmp6 * tmp6 tmp8 = tmp3 + tmp7 tmp11 = tmp9 - tmp10 tmp12 = tmp11 * tmp11 tmp13 = tmp8 + tmp12 tmp16 = tmp14 - tmp15 tmp17 = tmp16 * tmp16 tmp18 = tmp13 + tmp17 tmp19 = 1e-06 tmp20 = tmp18 + tmp19 tmp21 = libdevice.sqrt(tmp20) tl.store(out_ptr0 + x2, tmp21, xmask) @triton.jit def triton_poi_fused_add_mul_pow_sub_sum_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') tmp1 = tl.load(in_ptr1 + x3, xmask) tmp2 = tmp0 * tmp1 tl.store(out_ptr0 + x3, tmp2, 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((4, 1, 4, 4), (16, 64, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_pow_sub_sum_0[grid(64)](arg1_1, arg0_1, buf0, 64, XBLOCK=64, num_warps=1, num_stages=1) del arg0_1 del arg1_1 buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused_add_mul_pow_sub_sum_1[grid(256)](buf0, arg2_1, buf1, 256, XBLOCK=256, num_warps=4, num_stages=1) del arg2_1 del buf0 return buf1, class EPENew(nn.Module): def __init__(self): super(EPENew, self).__init__() 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]
P2Oileen/oh-my-face
EPE
false
15,031
[ "MIT" ]
45
b73cb8ea713205bbf2bc1408145fa668c715359b
https://github.com/P2Oileen/oh-my-face/tree/b73cb8ea713205bbf2bc1408145fa668c715359b
InvertibleDownsampling1D
from torch.autograd import Function import torch import numpy as np from warnings import warn from typing import Union from typing import Tuple from torch.nn.common_types import _size_1_t from torch.nn.modules.utils import _single import torch.nn.functional as F def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) def __initialize_weight__(kernel_matrix_shape: 'Tuple[int, ...]', stride: 'Tuple[int, ...]', method: 'str'='cayley', init: 'str'='haar', dtype: 'str'='float32', *args, **kwargs): """Function which computes specific orthogonal matrices. For some chosen method of parametrizing orthogonal matrices, this function outputs the required weights necessary to represent a chosen initialization as a Pytorch tensor of matrices. Args: kernel_matrix_shape : The output shape of the orthogonal matrices. Should be (num_matrices, height, width). stride : The stride for the invertible up- or downsampling for which this matrix is to be used. The length of ``stride`` should match the dimensionality of the data. method : The method for parametrising orthogonal matrices. Should be 'exp' or 'cayley' init : The matrix which should be represented. Should be 'squeeze', 'pixel_shuffle', 'haar' or 'random'. 'haar' is only possible if ``stride`` is only 2. dtype : Numpy dtype which should be used for the matrix. *args: Variable length argument iterable. **kwargs: Arbitrary keyword arguments. Returns: Tensor : Orthogonal matrices of shape ``kernel_matrix_shape``. """ dim = len(stride) num_matrices = kernel_matrix_shape[0] assert method in ['exp', 'cayley', 'householder'] if method == 'householder': warn( 'Householder parametrization not fully implemented yet. Only random initialization currently working.' ) init = 'random' if init == 'random': return torch.randn(kernel_matrix_shape) if init == 'haar' and set(stride) != {2}: None None init = 'squeeze' if init == 'haar' and set(stride) == {2}: if method == 'exp': p = np.pi / 4 if dim == 1: weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: weight = np.array([[[0, p, p, 0, p, 0, 0, 0], [0, 0, 0, p, 0, p, 0, 0], [0, 0, 0, p, 0, 0, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, p, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) elif method == 'cayley': if dim == 1: p = -np.sqrt(2) / (2 - np.sqrt(2)) weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: p = 0.5 weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: p = 1 / np.sqrt(2) weight = np.array([[[0, -p, -p, 0, -p, 0, 0, 1 - p], [0, 0, 0, -p, 0, -p, p - 1, 0], [0, 0, 0, -p, 0, p - 1, -p, 0], [0, 0, 0, 0, 1 - p, 0, 0, -p], [0, 0, 0, 0, 0, -p, -p, 0], [0, 0, 0, 0, 0, 0, 0, -p], [0, 0, 0, 0, 0, 0, 0, -p ], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) if init in ['squeeze', 'pixel_shuffle', 'zeros']: if method == 'exp' or method == 'cayley': return torch.zeros(*kernel_matrix_shape) if type(init) is np.ndarray: init = torch.tensor(init.astype(dtype)) if torch.is_tensor(init): if len(init.shape) == 2: init = init.reshape(1, *init.shape) if init.shape[0] == 1: init = init.repeat(num_matrices, 1, 1) assert init.shape == kernel_matrix_shape return init else: raise NotImplementedError('Unknown initialization.') class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalResamplingLayer(torch.nn.Module): """Base class for orthogonal up- and downsampling operators. :param low_channel_number: Lower number of channels. These are the input channels in the case of downsampling ops, and the output channels in the case of upsampling ops. :param stride: The downsampling / upsampling factor for each dimension. :param channel_multiplier: The channel multiplier, i.e. the number by which the number of channels are multiplied (downsampling) or divided (upsampling). :param method: Which method to use for parametrizing orthogonal matrices which are used as convolutional kernels. """ def __init__(self, low_channel_number: 'int', stride: 'Union[int, Tuple[int, ...]]', method: 'str'='cayley', init: 'Union[str, np.ndarray, torch.Tensor]'='haar', learnable: 'bool'= True, init_kwargs: 'dict'=None, **kwargs): super(OrthogonalResamplingLayer, self).__init__() self.low_channel_number = low_channel_number self.method = method self.stride = stride self.channel_multiplier = int(np.prod(stride)) self.high_channel_number = self.channel_multiplier * low_channel_number if init_kwargs is None: init_kwargs = {} self.init_kwargs = init_kwargs self.kwargs = kwargs assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self._kernel_matrix_shape = (self.low_channel_number,) + (self. channel_multiplier,) * 2 self._kernel_shape = (self.high_channel_number, 1) + self.stride self.weight = torch.nn.Parameter(__initialize_weight__( kernel_matrix_shape=self._kernel_matrix_shape, stride=self. stride, method=self.method, init=init, **self.init_kwargs)) self.weight.requires_grad = learnable @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel(self): """The kernel associated with the invertible up-/downsampling. """ return self.kernel_matrix.reshape(*self._kernel_shape) class InvertibleDownsampling1D(OrthogonalResamplingLayer): def __init__(self, in_channels: 'int', stride: '_size_1_t'=2, method: 'str'='cayley', init: 'str'='haar', learnable: 'bool'=True, *args, **kwargs): stride = tuple(_single(stride)) channel_multiplier = int(np.prod(stride)) self.in_channels = in_channels self.out_channels = in_channels * channel_multiplier super(InvertibleDownsampling1D, self).__init__(*args, low_channel_number=self.in_channels, stride=stride, method= method, init=init, learnable=learnable, **kwargs) def forward(self, x): return F.conv1d(x, self.kernel, stride=self.stride, groups=self. low_channel_number) def inverse(self, x): return F.conv_transpose1d(x, self.kernel, stride=self.stride, groups=self.low_channel_number) def get_inputs(): return [torch.rand([4, 4])] def get_init_inputs(): return [[], {'in_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.autograd import Function import numpy as np from warnings import warn from typing import Union from typing import Tuple from torch.nn.common_types import _size_1_t from torch.nn.modules.utils import _single 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_add_eye_sub_0(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 8 xnumel = 2 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 % 2 x2 = xindex y3 = yindex y1 = yindex // 2 tmp6 = tl.load(in_ptr0 + (x2 + 2 * y3), xmask & ymask, eviction_policy= 'evict_last') tmp7 = tl.load(in_ptr0 + (y0 + 2 * x2 + 4 * y1), xmask & ymask, eviction_policy='evict_last') tmp0 = y0 tmp1 = x2 tmp2 = tmp0 == tmp1 tmp3 = 1.0 tmp4 = 0.0 tmp5 = tl.where(tmp2, tmp3, tmp4) tmp8 = tmp6 - tmp7 tmp9 = tmp5 + tmp8 tmp10 = tmp5 - tmp8 tl.store(out_ptr0 + (x2 + 2 * y3), tmp9, xmask & ymask) tl.store(out_ptr1 + (x2 + 2 * y3), tmp10, xmask & ymask) @triton.jit def triton_poi_fused_clone_1(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl. constexpr, XBLOCK: tl.constexpr): ynumel = 8 xnumel = 2 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 % 2 y1 = yindex // 2 y3 = yindex tmp0 = tl.load(in_ptr0 + (y0 + 2 * x2 + 4 * y1), xmask & ymask, eviction_policy='evict_last') tl.store(out_ptr0 + (x2 + 2 * y3), tmp0, xmask & ymask) def call(args): primals_1, primals_2 = args args.clear() assert_size_stride(primals_1, (4, 2, 2), (4, 2, 1)) assert_size_stride(primals_2, (4, 4), (4, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 2, 2), (4, 2, 1), torch.float32) buf5 = empty_strided_cuda((4, 2, 2), (4, 2, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_eye_sub_0[grid(8, 2)](primals_1, buf0, buf5, 8, 2, XBLOCK=2, YBLOCK=8, num_warps=1, num_stages=1) del primals_1 buf1 = torch.ops.aten.linalg_lu_factor_ex.default(buf0) buf2 = buf1[0] buf3 = buf1[1] del buf1 buf6 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf5) buf7 = buf6 del buf6 buf8 = buf0 del buf0 triton_poi_fused_clone_1[grid(8, 2)](buf7, buf8, 8, 2, XBLOCK=2, YBLOCK=8, num_warps=1, num_stages=1) buf9 = extern_kernels.convolution(reinterpret_tensor(primals_2, (1, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf8, (8, 1, 2), (2, 0, 1), 0), stride=(2,), padding=(0,), dilation=(1,), transposed =False, output_padding=(0,), groups=4, bias=None) assert_size_stride(buf9, (1, 8, 2), (16, 2, 1)) buf10 = torch.ops.aten.linalg_lu_solve.default(buf2, buf3, buf7) del buf2 del buf3 del buf7 buf11 = buf10 del buf10 return reinterpret_tensor(buf9, (8, 2), (2, 1), 0 ), buf5, reinterpret_tensor(buf8, (8, 1, 2), (2, 2, 1), 0 ), reinterpret_tensor(primals_2, (1, 4, 4), (16, 4, 1), 0), buf11 def _cayley(A): I = torch.eye(A.shape[-1], device=A.device) LU = torch.lu(I + A, pivot=True) return torch.lu_solve(I - A, *LU) def _cayley_frechet(A, H, Q=None): I = torch.eye(A.shape[-1], device=A.device) if Q is None: Q = _cayley(A) _LU = torch.lu(I + A, pivot=True) p = torch.lu_solve(Q, *_LU) _LU = torch.lu(I - A, pivot=True) q = torch.lu_solve(H, *_LU) return 2.0 * q @ p def __calculate_kernel_matrix_cayley__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return cayley.apply(skew_symmetric_matrix) def matrix_1_norm(A): """Calculates the 1-norm of a matrix or a batch of matrices. Args: A (torch.Tensor): Can be either of size (n,n) or (m,n,n). Returns: torch.Tensor : The 1-norm of A. """ norm, _indices = torch.max(torch.sum(torch.abs(A), axis=-2), axis=-1) return norm def _compute_scales(A): """Compute optimal parameters for scaling-and-squaring algorithm. The constants used in this function are determined by the MATLAB function found in https://github.com/cetmann/pytorch_expm/blob/master/determine_frechet_scaling_constant.m """ norm = matrix_1_norm(A) max_norm = torch.max(norm) s = torch.zeros_like(norm) if A.dtype == torch.float64: if A.requires_grad: ell = {(3): 0.010813385777848, (5): 0.199806320697895, (7): 0.783460847296204, (9): 1.782448623969279, (13): 4.740307543765127} else: ell = {(3): 0.014955852179582, (5): 0.253939833006323, (7): 0.950417899616293, (9): 2.097847961257068, (13): 5.371920351148152} if max_norm >= ell[9]: m = 13 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5, 7, 9]: if max_norm < ell[m]: magic_number = ell[m] break elif A.dtype == torch.float32: if A.requires_grad: ell = {(3): 0.30803304184533, (5): 1.482532614793145, (7): 3.248671755200478} else: ell = {(3): 0.425873001692283, (5): 1.880152677804762, (7): 3.92572478313866} if max_norm >= ell[5]: m = 7 magic_number = ell[m] s = torch.relu_(torch.ceil(torch.log2_(norm / magic_number))) else: for m in [3, 5]: if max_norm < ell[m]: magic_number = ell[m] break return s, m def _eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def _expm_frechet_pade(A, E, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: if m >= 3: M_2 = A @ E + E @ A A_2 = A @ A U = b[3] * A_2 V = b[2] * A_2 L_U = b[3] * M_2 L_V = b[2] * M_2 if m >= 5: M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 L_U = L_U + b[5] * M_4 L_V = L_V + b[4] * M_4 if m >= 7: M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 L_U = L_U + b[7] * M_6 L_V = L_V + b[6] * M_6 if m == 9: M_8 = A_4 @ M_4 + M_4 @ A_4 A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 L_U = L_U + b[9] * M_8 L_V = L_V + b[8] * M_8 U = U + b[1] * I V = U + b[0] * I del I L_U = A @ L_U L_U = L_U + E @ U U = A @ U else: M_2 = A @ E + E @ A A_2 = A @ A M_4 = A_2 @ M_2 + M_2 @ A_2 A_4 = A_2 @ A_2 M_6 = A_4 @ M_2 + M_4 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 L_W1 = b[13] * M_6 + b[11] * M_4 + b[9] * M_2 L_W2 = b[7] * M_6 + b[5] * M_4 + b[3] * M_2 L_Z1 = b[12] * M_6 + b[10] * M_4 + b[8] * M_2 L_Z2 = b[6] * M_6 + b[4] * M_4 + b[2] * M_2 L_W = A_6 @ L_W1 + M_6 @ W_1 + L_W2 L_U = A @ L_W + E @ W L_V = A_6 @ L_Z1 + M_6 @ Z_1 + L_Z2 lu_decom = torch.lu(-U + V) exp_A = torch.lu_solve(U + V, *lu_decom) dexp_A = torch.lu_solve(L_U + L_V + (L_U - L_V) @ exp_A, *lu_decom) return exp_A, dexp_A def _square(s, R, L=None): """The `squaring` part of the `scaling-and-squaring` algorithm. This works both for the forward as well as the derivative of the matrix exponential. """ s_max = torch.max(s).int() if s_max > 0: I = _eye_like(R) if L is not None: O = torch.zeros_like(R) indices = [(1) for k in range(len(R.shape) - 1)] for i in range(s_max): mask = i >= s matrices_mask = mask.view(-1, *indices) temp_eye = torch.clone(R).masked_scatter(matrices_mask, I) if L is not None: temp_zeros = torch.clone(R).masked_scatter(matrices_mask, O) L = temp_eye @ L + temp_zeros @ L R = R @ temp_eye del temp_eye, mask if L is not None: return R, L else: return R def _expm_frechet_scaling_squaring(A, E, adjoint=False): """Numerical Fréchet derivative of matrix exponentiation. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False if adjoint is True: A = torch.transpose(A, -1, -2) s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] scaling_factors = torch.pow(2, -s).view(-1, *indices) A = A * scaling_factors E = E * scaling_factors exp_A, dexp_A = _expm_frechet_pade(A, E, m) exp_A, dexp_A = _square(s, exp_A, dexp_A) return dexp_A def _expm_pade(A, m=7): assert m in [3, 5, 7, 9, 13] if m == 3: b = [120.0, 60.0, 12.0, 1.0] elif m == 5: b = [30240.0, 15120.0, 3360.0, 420.0, 30.0, 1.0] elif m == 7: b = [17297280.0, 8648640.0, 1995840.0, 277200.0, 25200.0, 1512.0, 56.0, 1.0] elif m == 9: b = [17643225600.0, 8821612800.0, 2075673600.0, 302702400.0, 30270240.0, 2162160.0, 110880.0, 3960.0, 90.0, 1.0] elif m == 13: b = [6.476475253248e+16, 3.238237626624e+16, 7771770303897600.0, 1187353796428800.0, 129060195264000.0, 10559470521600.0, 670442572800.0, 33522128640.0, 1323241920.0, 40840800.0, 960960.0, 16380.0, 182.0, 1.0] I = _eye_like(A) if m != 13: U = b[1] * I V = b[0] * I if m >= 3: A_2 = A @ A U = U + b[3] * A_2 V = V + b[2] * A_2 if m >= 5: A_4 = A_2 @ A_2 U = U + b[5] * A_4 V = V + b[4] * A_4 if m >= 7: A_6 = A_4 @ A_2 U = U + b[7] * A_6 V = V + b[6] * A_6 if m == 9: A_8 = A_4 @ A_4 U = U + b[9] * A_8 V = V + b[8] * A_8 U = A @ U else: A_2 = A @ A A_4 = A_2 @ A_2 A_6 = A_4 @ A_2 W_1 = b[13] * A_6 + b[11] * A_4 + b[9] * A_2 W_2 = b[7] * A_6 + b[5] * A_4 + b[3] * A_2 + b[1] * I W = A_6 @ W_1 + W_2 Z_1 = b[12] * A_6 + b[10] * A_4 + b[8] * A_2 Z_2 = b[6] * A_6 + b[4] * A_4 + b[2] * A_2 + b[0] * I U = A @ W V = A_6 @ Z_1 + Z_2 del A_2 if m >= 5: del A_4 if m >= 7: del A_6 if m == 9: del A_8 R = torch.lu_solve(U + V, *torch.lu(-U + V)) del U, V return R def _expm_scaling_squaring(A): """Scaling-and-squaring algorithm for matrix eponentiation. This is based on the observation that exp(A) = exp(A/k)^k, where e.g. k=2^s. The exponential exp(A/(2^s)) is calculated by a diagonal Padé approximation, where s is chosen based on the 1-norm of A, such that certain approximation guarantees can be given. exp(A) is then calculated by repeated squaring via exp(A/(2^s))^(2^s). This function works both for (n,n)-tensors as well as batchwise for (m,n,n)-tensors. """ assert A.shape[-1] == A.shape[-2] and len(A.shape) in [2, 3] True if len(A.shape) == 3 else False s, m = _compute_scales(A) if torch.max(s) > 0: indices = [(1) for k in range(len(A.shape) - 1)] A = A * torch.pow(2, -s).view(-1, *indices) exp_A = _expm_pade(A, m) exp_A = _square(s, exp_A) return exp_A def __calculate_kernel_matrix_exp__(weight, **kwargs): skew_symmetric_matrix = weight - torch.transpose(weight, -1, -2) return expm.apply(skew_symmetric_matrix) def eye_like(M, device=None, dtype=None): """Creates an identity matrix of the same shape as another matrix. For matrix M, the output is same shape as M, if M is a (n,n)-matrix. If M is a batch of m matrices (i.e. a (m,n,n)-tensor), create a batch of (n,n)-identity-matrices. Args: M (torch.Tensor) : A tensor of either shape (n,n) or (m,n,n), for which either an identity matrix or a batch of identity matrices of the same shape will be created. device (torch.device, optional) : The device on which the output will be placed. By default, it is placed on the same device as M. dtype (torch.dtype, optional) : The dtype of the output. By default, it is the same dtype as M. Returns: torch.Tensor : Identity matrix or batch of identity matrices. """ assert len(M.shape) in [2, 3] assert M.shape[-1] == M.shape[-2] n = M.shape[-1] if device is None: device = M.device if dtype is None: dtype = M.dtype eye = torch.eye(M.shape[-1], device=device, dtype=dtype) if len(M.shape) == 2: return eye else: m = M.shape[0] return eye.view(-1, n, n).expand(m, -1, -1) def householder_matrix(unit_vector): if unit_vector.shape[-1] != 1: if len(unit_vector.shape) == 1: return torch.ones_like(unit_vector) unit_vector = unit_vector.view(*tuple(unit_vector.shape), 1) transform = 2 * unit_vector @ torch.transpose(unit_vector, -1, -2) return eye_like(transform) - transform def normalize_matrix_rows(matrix, eps=1e-06): norms = torch.sqrt(torch.sum(matrix ** 2, dim=-2, keepdim=True) + eps) return matrix / norms def householder_transform(matrix, n_reflections=-1, eps=1e-06): """Implements a product of Householder transforms. """ if n_reflections == -1: n_reflections = matrix.shape[-1] if n_reflections > matrix.shape[-1]: warn('n_reflections is set higher than the number of rows.') n_reflections = matrix.shape[-1] matrix = normalize_matrix_rows(matrix, eps) if n_reflections == 0: output = torch.eye(matrix.shape[-2], dtype=matrix.dtype, device= matrix.device) if len(matrix.shape) == 3: output = output.view(1, matrix.shape[1], matrix.shape[1]) output = output.expand(matrix.shape[0], -1, -1) for i in range(n_reflections): unit_vector = matrix[..., i:i + 1] householder = householder_matrix(unit_vector) if i == 0: output = householder else: output = output @ householder return output def __calculate_kernel_matrix_householder__(weight, **kwargs): n_reflections = kwargs.get('n_reflections', -1) eps = kwargs.get('eps', 1e-06) weight.shape[-1] weight = weight[..., n_reflections:] return householder_transform(weight, n_reflections, eps) def __initialize_weight__(kernel_matrix_shape: 'Tuple[int, ...]', stride: 'Tuple[int, ...]', method: 'str'='cayley', init: 'str'='haar', dtype: 'str'='float32', *args, **kwargs): """Function which computes specific orthogonal matrices. For some chosen method of parametrizing orthogonal matrices, this function outputs the required weights necessary to represent a chosen initialization as a Pytorch tensor of matrices. Args: kernel_matrix_shape : The output shape of the orthogonal matrices. Should be (num_matrices, height, width). stride : The stride for the invertible up- or downsampling for which this matrix is to be used. The length of ``stride`` should match the dimensionality of the data. method : The method for parametrising orthogonal matrices. Should be 'exp' or 'cayley' init : The matrix which should be represented. Should be 'squeeze', 'pixel_shuffle', 'haar' or 'random'. 'haar' is only possible if ``stride`` is only 2. dtype : Numpy dtype which should be used for the matrix. *args: Variable length argument iterable. **kwargs: Arbitrary keyword arguments. Returns: Tensor : Orthogonal matrices of shape ``kernel_matrix_shape``. """ dim = len(stride) num_matrices = kernel_matrix_shape[0] assert method in ['exp', 'cayley', 'householder'] if method == 'householder': warn( 'Householder parametrization not fully implemented yet. Only random initialization currently working.' ) init = 'random' if init == 'random': return torch.randn(kernel_matrix_shape) if init == 'haar' and set(stride) != {2}: None None init = 'squeeze' if init == 'haar' and set(stride) == {2}: if method == 'exp': p = np.pi / 4 if dim == 1: weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: weight = np.array([[[0, p, p, 0, p, 0, 0, 0], [0, 0, 0, p, 0, p, 0, 0], [0, 0, 0, p, 0, 0, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, p, p, 0], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, p], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) elif method == 'cayley': if dim == 1: p = -np.sqrt(2) / (2 - np.sqrt(2)) weight = np.array([[[0, p], [0, 0]]], dtype=dtype) if dim == 2: p = 0.5 weight = np.array([[[0, 0, p, p], [0, 0, -p, -p], [0, 0, 0, 0], [0, 0, 0, 0]]], dtype=dtype) if dim == 3: p = 1 / np.sqrt(2) weight = np.array([[[0, -p, -p, 0, -p, 0, 0, 1 - p], [0, 0, 0, -p, 0, -p, p - 1, 0], [0, 0, 0, -p, 0, p - 1, -p, 0], [0, 0, 0, 0, 1 - p, 0, 0, -p], [0, 0, 0, 0, 0, -p, -p, 0], [0, 0, 0, 0, 0, 0, 0, -p], [0, 0, 0, 0, 0, 0, 0, -p ], [0, 0, 0, 0, 0, 0, 0, 0]]], dtype=dtype) return torch.tensor(weight).repeat(num_matrices, 1, 1) if init in ['squeeze', 'pixel_shuffle', 'zeros']: if method == 'exp' or method == 'cayley': return torch.zeros(*kernel_matrix_shape) if type(init) is np.ndarray: init = torch.tensor(init.astype(dtype)) if torch.is_tensor(init): if len(init.shape) == 2: init = init.reshape(1, *init.shape) if init.shape[0] == 1: init = init.repeat(num_matrices, 1, 1) assert init.shape == kernel_matrix_shape return init else: raise NotImplementedError('Unknown initialization.') class cayley(Function): """Computes the Cayley transform. """ @staticmethod def forward(ctx, M): cayley_M = _cayley(M) ctx.save_for_backward(M, cayley_M) return cayley_M @staticmethod def backward(ctx, grad_out): M, cayley_M = ctx.saved_tensors dcayley_M = _cayley_frechet(M, grad_out, Q=cayley_M) return dcayley_M class expm(Function): """Computes the matrix exponential. """ @staticmethod def forward(ctx, M): expm_M = _expm_scaling_squaring(M) ctx.save_for_backward(M) return expm_M @staticmethod def backward(ctx, grad_out): M = ctx.saved_tensors[0] dexpm = _expm_frechet_scaling_squaring(M, grad_out, adjoint=True) return dexpm class OrthogonalResamplingLayer(torch.nn.Module): """Base class for orthogonal up- and downsampling operators. :param low_channel_number: Lower number of channels. These are the input channels in the case of downsampling ops, and the output channels in the case of upsampling ops. :param stride: The downsampling / upsampling factor for each dimension. :param channel_multiplier: The channel multiplier, i.e. the number by which the number of channels are multiplied (downsampling) or divided (upsampling). :param method: Which method to use for parametrizing orthogonal matrices which are used as convolutional kernels. """ def __init__(self, low_channel_number: 'int', stride: 'Union[int, Tuple[int, ...]]', method: 'str'='cayley', init: 'Union[str, np.ndarray, torch.Tensor]'='haar', learnable: 'bool'= True, init_kwargs: 'dict'=None, **kwargs): super(OrthogonalResamplingLayer, self).__init__() self.low_channel_number = low_channel_number self.method = method self.stride = stride self.channel_multiplier = int(np.prod(stride)) self.high_channel_number = self.channel_multiplier * low_channel_number if init_kwargs is None: init_kwargs = {} self.init_kwargs = init_kwargs self.kwargs = kwargs assert method in ['exp', 'cayley', 'householder'] if method == 'exp': self.__calculate_kernel_matrix__ = __calculate_kernel_matrix_exp__ elif method == 'cayley': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_cayley__) elif method == 'householder': self.__calculate_kernel_matrix__ = ( __calculate_kernel_matrix_householder__) self._kernel_matrix_shape = (self.low_channel_number,) + (self. channel_multiplier,) * 2 self._kernel_shape = (self.high_channel_number, 1) + self.stride self.weight = torch.nn.Parameter(__initialize_weight__( kernel_matrix_shape=self._kernel_matrix_shape, stride=self. stride, method=self.method, init=init, **self.init_kwargs)) self.weight.requires_grad = learnable @property def kernel_matrix(self): """The orthogonal matrix created by the chosen parametrisation method. """ return self.__calculate_kernel_matrix__(self.weight, **self.kwargs) @property def kernel(self): """The kernel associated with the invertible up-/downsampling. """ return self.kernel_matrix.reshape(*self._kernel_shape) class InvertibleDownsampling1DNew(OrthogonalResamplingLayer): def __init__(self, in_channels: 'int', stride: '_size_1_t'=2, method: 'str'='cayley', init: 'str'='haar', learnable: 'bool'=True, *args, **kwargs): stride = tuple(_single(stride)) channel_multiplier = int(np.prod(stride)) self.in_channels = in_channels self.out_channels = in_channels * channel_multiplier super(InvertibleDownsampling1DNew, self).__init__(*args, low_channel_number=self.in_channels, stride=stride, method= method, init=init, learnable=learnable, **kwargs) def inverse(self, x): return F.conv_transpose1d(x, self.kernel, stride=self.stride, groups=self.low_channel_number) def forward(self, input_0): primals_1 = self.weight primals_2 = input_0 output = call([primals_1, primals_2]) return output[0]
cetmann/iunets
InvertibleDownsampling1D
false
15,032
[ "MIT" ]
86
80ed7cce0e505a0396c42359eaf27819222d71f6
https://github.com/cetmann/iunets/tree/80ed7cce0e505a0396c42359eaf27819222d71f6
MLP_G
from _paritybench_helpers import _mock_config import torch import torch.nn as nn def weights_init(m): classname = m.__class__.__name__ if classname.find('Linear') != -1: m.weight.data.normal_(0.0, 0.02) m.bias.data.fill_(0) elif classname.find('BatchNorm') != -1: m.weight.data.normal_(1.0, 0.02) m.bias.data.fill_(0) class MLP_G(nn.Module): def __init__(self, opt): super(MLP_G, self).__init__() self.fc1 = nn.Linear(opt.attSize + opt.nz, opt.ngh) self.fc2 = nn.Linear(opt.ngh, opt.resSize) self.lrelu = nn.LeakyReLU(0.2, True) self.relu = nn.ReLU(True) self.apply(weights_init) def forward(self, noise, att): h = torch.cat((noise, att), 1) h = self.lrelu(self.fc1(h)) h = self.relu(self.fc2(h)) return h def get_inputs(): return [torch.rand([4, 4]), torch.rand([4, 4])] def get_init_inputs(): return [[], {'opt': _mock_config(attSize=4, nz=4, ngh=4, resSize=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, 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_leaky_relu_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 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tmp5 = 0.2 tmp6 = tmp2 * tmp5 tmp7 = tl.where(tmp4, tmp2, tmp6) tl.store(in_out_ptr0 + x2, tmp7, xmask) @triton.jit def triton_poi_fused_relu_threshold_backward_2(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 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) 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, 4), (4, 1)) assert_size_stride(primals_3, (4, 8), (8, 1)) assert_size_stride(primals_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((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, 4), (4, 1), torch.float32) extern_kernels.mm(buf0, reinterpret_tensor(primals_3, (8, 4), (1, 8 ), 0), out=buf1) del primals_3 buf2 = buf1 del buf1 triton_poi_fused_leaky_relu_1[grid(16)](buf2, primals_4, 16, XBLOCK =16, num_warps=1, num_stages=1) del primals_4 buf3 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.mm(buf2, reinterpret_tensor(primals_5, (4, 4), (1, 4 ), 0), out=buf3) buf4 = buf3 del buf3 buf5 = empty_strided_cuda((4, 4), (4, 1), torch.bool) triton_poi_fused_relu_threshold_backward_2[grid(16)](buf4, primals_6, buf5, 16, XBLOCK=16, num_warps=1, num_stages=1) del primals_6 return buf4, buf0, buf2, buf5, primals_5 def weights_init(m): classname = m.__class__.__name__ if classname.find('Linear') != -1: m.weight.data.normal_(0.0, 0.02) m.bias.data.fill_(0) elif classname.find('BatchNorm') != -1: m.weight.data.normal_(1.0, 0.02) m.bias.data.fill_(0) class MLP_GNew(nn.Module): def __init__(self, opt): super(MLP_GNew, self).__init__() self.fc1 = nn.Linear(opt.attSize + opt.nz, opt.ngh) self.fc2 = nn.Linear(opt.ngh, opt.resSize) self.lrelu = nn.LeakyReLU(0.2, True) self.relu = nn.ReLU(True) self.apply(weights_init) def forward(self, input_0, input_1): primals_3 = self.fc1.weight primals_4 = self.fc1.bias primals_1 = self.fc2.weight primals_6 = self.fc2.bias primals_2 = input_0 primals_5 = input_1 output = call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6]) return output[0]
Huihui-z/CE-GZSL
MLP_G
false
15,033
[ "MIT" ]
58
7bf5358ac4727ea1dc2dc9dec2f453b014500bd8
https://github.com/Huihui-z/CE-GZSL/tree/7bf5358ac4727ea1dc2dc9dec2f453b014500bd8
SqueezeExcite
import torch import torch.nn as nn import torch.nn.functional as F from itertools import product as product def _make_divisible(v, divisor, min_value=None): """ This function is taken from the original tf repo. It ensures that all layers have a channel number that is divisible by 8 It can be seen here: https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py :param v: :param divisor: :param min_value: :return: """ if min_value is None: min_value = divisor new_v = max(min_value, int(v + divisor / 2) // divisor * divisor) if new_v < 0.9 * v: new_v += divisor return new_v def hard_sigmoid(x, inplace: 'bool'=False): if inplace: return x.add_(3.0).clamp_(0.0, 6.0).div_(6.0) else: return F.relu6(x + 3.0) / 6.0 class SqueezeExcite(nn.Module): def __init__(self, in_chs, se_ratio=0.25, reduced_base_chs=None, act_layer=nn.ReLU, gate_fn=hard_sigmoid, divisor=4, **_): super(SqueezeExcite, self).__init__() self.gate_fn = gate_fn reduced_chs = _make_divisible((reduced_base_chs or in_chs) * se_ratio, divisor) self.avg_pool = nn.AdaptiveAvgPool2d(1) self.conv_reduce = nn.Conv2d(in_chs, reduced_chs, 1, bias=True) self.act1 = act_layer(inplace=True) self.conv_expand = nn.Conv2d(reduced_chs, in_chs, 1, bias=True) def forward(self, x): x_se = self.avg_pool(x) x_se = self.conv_reduce(x_se) x_se = self.act1(x_se) x_se = self.conv_expand(x_se) x = x * self.gate_fn(x_se) return x def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_chs': 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 import torch.nn.functional as F from itertools import product as product 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_relu_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 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_add_convolution_div_hardtanh_mul_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 x3 = xindex x4 = xindex // 16 x1 = xindex // 16 % 4 tmp0 = tl.load(in_ptr0 + x3, xmask) tmp1 = tl.load(in_ptr1 + x4, xmask, eviction_policy='evict_last') tmp2 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last') tmp3 = tmp1 + tmp2 tmp4 = 3.0 tmp5 = tmp3 + tmp4 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = 6.0 tmp9 = triton_helpers.minimum(tmp7, tmp8) tmp10 = 0.16666666666666666 tmp11 = tmp9 * tmp10 tmp12 = tmp0 * tmp11 tl.store(out_ptr0 + x3, tmp12, xmask) @triton.jit def triton_poi_fused_add_convolution_hardtanh_backward_3(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 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 = 3.0 tmp4 = tmp2 + tmp3 tmp5 = 0.0 tmp6 = tmp4 <= tmp5 tmp7 = 6.0 tmp8 = tmp4 >= tmp7 tmp9 = tmp6 | tmp8 tl.store(out_ptr0 + x2, tmp9, 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, 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,)) 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 = extern_kernels.convolution(buf1, primals_2, 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, 1, 1), (4, 1, 1, 1)) buf3 = buf2 del buf2 triton_poi_fused_convolution_relu_1[grid(16)](buf3, primals_3, 16, XBLOCK=16, num_warps=1, num_stages=1) del primals_3 buf4 = extern_kernels.convolution(buf3, primals_4, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf4, (4, 4, 1, 1), (4, 1, 1, 1)) buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused_add_convolution_div_hardtanh_mul_2[grid(256)]( primals_1, buf4, primals_5, buf5, 256, XBLOCK=256, num_warps=4, num_stages=1) buf6 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.bool) triton_poi_fused_add_convolution_hardtanh_backward_3[grid(16)](buf4, primals_5, buf6, 16, XBLOCK=16, num_warps=1, num_stages=1) del buf4 del primals_5 return buf5, primals_1, primals_2, primals_4, buf1, buf3, buf6 def _make_divisible(v, divisor, min_value=None): """ This function is taken from the original tf repo. It ensures that all layers have a channel number that is divisible by 8 It can be seen here: https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py :param v: :param divisor: :param min_value: :return: """ if min_value is None: min_value = divisor new_v = max(min_value, int(v + divisor / 2) // divisor * divisor) if new_v < 0.9 * v: new_v += divisor return new_v def hard_sigmoid(x, inplace: 'bool'=False): if inplace: return x.add_(3.0).clamp_(0.0, 6.0).div_(6.0) else: return F.relu6(x + 3.0) / 6.0 class SqueezeExciteNew(nn.Module): def __init__(self, in_chs, se_ratio=0.25, reduced_base_chs=None, act_layer=nn.ReLU, gate_fn=hard_sigmoid, divisor=4, **_): super(SqueezeExciteNew, self).__init__() self.gate_fn = gate_fn reduced_chs = _make_divisible((reduced_base_chs or in_chs) * se_ratio, divisor) self.avg_pool = nn.AdaptiveAvgPool2d(1) self.conv_reduce = nn.Conv2d(in_chs, reduced_chs, 1, bias=True) self.act1 = act_layer(inplace=True) self.conv_expand = nn.Conv2d(reduced_chs, in_chs, 1, bias=True) def forward(self, input_0): primals_2 = self.conv_reduce.weight primals_3 = self.conv_reduce.bias primals_4 = self.conv_expand.weight primals_5 = self.conv_expand.bias primals_1 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5]) return output[0]
chuanli11/SynergyNet
SqueezeExcite
false
15,034
[ "MIT" ]
82
a8044d8dabbfb811d4299f59e64e0fb749027e86
https://github.com/chuanli11/SynergyNet/tree/a8044d8dabbfb811d4299f59e64e0fb749027e86
BasicBlock_ins
import torch import torch.nn as nn import torch.multiprocessing def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock_ins(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock_ins, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.InstanceNorm2d(planes, affine=True, track_running_stats=False) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.InstanceNorm2d(planes, affine=True, track_running_stats=False) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) 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 libdevice import torch.nn as nn import torch.multiprocessing 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_relu_repeat_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1, out_ptr3, out_ptr4, 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) x0 = xindex r1 = rindex x2 = xindex % 4 tmp0 = tl.load(in_ptr0 + x0 % 4, xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr1 + (r1 + 16 * x0), xmask, other=0.0) tmp26 = tl.load(in_ptr2 + x2, xmask, eviction_policy='evict_last') 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 = tmp1 - tmp11 tmp19 = 16.0 tmp20 = tmp17 / tmp19 tmp21 = 1e-05 tmp22 = tmp20 + tmp21 tmp23 = libdevice.rsqrt(tmp22) tmp24 = tmp18 * tmp23 tmp25 = tmp24 * tmp0 tmp27 = tmp25 + tmp26 tmp28 = tl.full([1, 1], 0, tl.int32) tmp29 = triton_helpers.maximum(tmp28, tmp27) tl.store(out_ptr0 + x0, tmp0, xmask) tl.store(out_ptr3 + (r1 + 16 * x0), tmp29, xmask) tl.store(out_ptr4 + x0, tmp23, xmask) tl.store(out_ptr1 + x0, tmp11, xmask) @triton.jit def triton_per_fused__native_batch_norm_legit_relu_repeat_threshold_backward_1( in_ptr0, in_ptr1, in_ptr2, in_ptr3, out_ptr0, out_ptr1, out_ptr3, out_ptr4, out_ptr5, 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) x0 = xindex r1 = rindex x2 = xindex % 4 tmp0 = tl.load(in_ptr0 + x0 % 4, xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr1 + (r1 + 16 * x0), xmask, other=0.0) tmp26 = tl.load(in_ptr2 + x2, xmask, eviction_policy='evict_last') tmp28 = 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 = tmp1 - tmp11 tmp19 = 16.0 tmp20 = tmp17 / tmp19 tmp21 = 1e-05 tmp22 = tmp20 + tmp21 tmp23 = libdevice.rsqrt(tmp22) tmp24 = tmp18 * tmp23 tmp25 = tmp24 * tmp0 tmp27 = tmp25 + tmp26 tmp29 = tmp27 + tmp28 tmp30 = tl.full([1, 1], 0, tl.int32) tmp31 = triton_helpers.maximum(tmp30, tmp29) tmp32 = 0.0 tmp33 = tmp31 <= tmp32 tl.store(out_ptr0 + x0, tmp0, xmask) tl.store(out_ptr3 + (r1 + 16 * x0), tmp31, xmask) tl.store(out_ptr4 + (r1 + 16 * x0), tmp33, xmask) tl.store(out_ptr5 + x0, tmp23, 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) = args args.clear() assert_size_stride(primals_1, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_2, (4, 4, 3, 3), (36, 9, 3, 1)) assert_size_stride(primals_3, (4,), (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,)) assert_size_stride(primals_7, (4,), (1,)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = extern_kernels.convolution(primals_1, primals_2, stride=(1, 1), padding=(1, 1), 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 = empty_strided_cuda((16,), (1,), torch.float32) buf2 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch.float32 ) buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) buf5 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch.float32 ) get_raw_stream(0) triton_per_fused__native_batch_norm_legit_relu_repeat_0[grid(16)]( primals_3, buf0, primals_4, buf1, buf2, buf6, buf5, 16, 16, XBLOCK=8, num_warps=2, num_stages=1) del primals_3 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((16,), (1,), torch.float32) buf9 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch.float32 ) buf13 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) buf14 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool) buf12 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch. float32) triton_per_fused__native_batch_norm_legit_relu_repeat_threshold_backward_1[ grid(16)](primals_6, buf7, primals_7, primals_1, buf8, buf9, buf13, buf14, buf12, 16, 16, XBLOCK=1, num_warps=2, num_stages=1) del primals_6 del primals_7 return (buf13, primals_1, primals_2, primals_5, buf0, buf1, reinterpret_tensor(buf5, (16,), (1,), 0), buf6, buf7, buf8, reinterpret_tensor(buf12, (16,), (1,), 0), buf14, reinterpret_tensor(buf9, (1, 16, 1, 1), (16, 1, 1, 1), 0), reinterpret_tensor(buf2, (1, 16, 1, 1), (16, 1, 1, 1), 0)) def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock_insNew(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock_insNew, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.InstanceNorm2d(planes, affine=True, track_running_stats=False) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.InstanceNorm2d(planes, affine=True, track_running_stats=False) self.downsample = downsample self.stride = stride def forward(self, input_0): primals_2 = self.conv1.weight primals_3 = self.bn1.weight primals_4 = self.bn1.bias primals_5 = self.conv2.weight primals_6 = self.bn2.weight primals_7 = self.bn2.bias primals_1 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7]) return output[0]
chiukin/RANet
BasicBlock_ins
false
15,035
[ "Apache-2.0" ]
267
681a47d9b1f114653290678f02f2d3ecdf4010bc
https://github.com/chiukin/RANet/tree/681a47d9b1f114653290678f02f2d3ecdf4010bc
ResBlock2
import torch import torch.nn as nn import torch.multiprocessing class ResBlock2(nn.Module): def __init__(self, input_feature, planes, dilated=1, group=1): super(ResBlock2, self).__init__() self.conv1 = nn.Conv2d(input_feature, planes, kernel_size=1, bias= False, groups=group) self.bn1 = nn.InstanceNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1 * dilated, bias=False, dilation=dilated, groups=group) self.bn2 = nn.InstanceNorm2d(planes) self.conv3 = nn.Conv2d(planes, input_feature, kernel_size=1, bias= False, groups=group) self.bn3 = nn.InstanceNorm2d(input_feature) self.relu = nn.ReLU(inplace=True) def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) out += residual out = self.relu(out) return out def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'input_feature': 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 libdevice import torch.nn as nn import torch.multiprocessing 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_relu_0(in_ptr0, out_ptr0, out_ptr2, out_ptr3, 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]) 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 = tmp0 - tmp10 tmp18 = 16.0 tmp19 = tmp16 / tmp18 tmp20 = 1e-05 tmp21 = tmp19 + tmp20 tmp22 = libdevice.rsqrt(tmp21) tmp23 = tmp17 * tmp22 tmp24 = tl.full([1, 1], 0, tl.int32) tmp25 = triton_helpers.maximum(tmp24, tmp23) tl.store(out_ptr2 + (r1 + 16 * x0), tmp25, xmask) tl.store(out_ptr3 + x0, tmp22, xmask) tl.store(out_ptr0 + x0, tmp10, xmask) @triton.jit def triton_per_fused__native_batch_norm_legit_relu_threshold_backward_1(in_ptr0 , in_ptr1, out_ptr0, out_ptr2, out_ptr3, out_ptr4, 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) tmp24 = tl.load(in_ptr1 + (r1 + 16 * x0), 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 = tmp0 - tmp10 tmp18 = 16.0 tmp19 = tmp16 / tmp18 tmp20 = 1e-05 tmp21 = tmp19 + tmp20 tmp22 = libdevice.rsqrt(tmp21) tmp23 = tmp17 * tmp22 tmp25 = tmp23 + tmp24 tmp26 = tl.full([1, 1], 0, tl.int32) tmp27 = triton_helpers.maximum(tmp26, tmp25) tmp28 = 0.0 tmp29 = tmp27 <= tmp28 tl.store(out_ptr2 + (r1 + 16 * x0), tmp27, xmask) tl.store(out_ptr3 + (r1 + 16 * x0), tmp29, xmask) tl.store(out_ptr4 + x0, tmp22, xmask) tl.store(out_ptr0 + x0, tmp10, 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, 1, 1), (4, 1, 1, 1)) assert_size_stride(primals_3, (4, 4, 3, 3), (36, 9, 3, 1)) assert_size_stride(primals_4, (4, 4, 1, 1), (4, 1, 1, 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 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch.float32 ) buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) buf4 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch.float32 ) get_raw_stream(0) triton_per_fused__native_batch_norm_legit_relu_0[grid(16)](buf0, buf1, buf5, buf4, 16, 16, XBLOCK=8, num_warps=2, num_stages=1) buf6 = extern_kernels.convolution(buf5, primals_3, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf6, (4, 4, 4, 4), (64, 16, 4, 1)) buf7 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch.float32 ) buf11 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) buf10 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch. float32) triton_per_fused__native_batch_norm_legit_relu_0[grid(16)](buf6, buf7, buf11, buf10, 16, 16, XBLOCK=8, num_warps=2, num_stages=1) buf12 = extern_kernels.convolution(buf11, primals_4, 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 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch. float32) buf17 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) buf18 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool) buf16 = empty_strided_cuda((1, 16, 1, 1), (16, 1, 16, 16), torch. float32) triton_per_fused__native_batch_norm_legit_relu_threshold_backward_1[ grid(16)](buf12, primals_1, buf13, buf17, buf18, buf16, 16, 16, XBLOCK=1, num_warps=2, num_stages=1) return (buf17, primals_1, primals_2, primals_3, primals_4, buf0, reinterpret_tensor(buf4, (16,), (1,), 0), buf5, buf6, reinterpret_tensor(buf10, (16,), (1,), 0), buf11, buf12, reinterpret_tensor(buf16, (16,), (1,), 0), buf18, reinterpret_tensor(buf13, (1, 16, 1, 1), (16, 1, 1, 1), 0), reinterpret_tensor(buf7, (1, 16, 1, 1), (16, 1, 1, 1), 0), reinterpret_tensor(buf1, (1, 16, 1, 1), (16, 1, 1, 1), 0)) class ResBlock2New(nn.Module): def __init__(self, input_feature, planes, dilated=1, group=1): super(ResBlock2New, self).__init__() self.conv1 = nn.Conv2d(input_feature, planes, kernel_size=1, bias= False, groups=group) self.bn1 = nn.InstanceNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1 * dilated, bias=False, dilation=dilated, groups=group) self.bn2 = nn.InstanceNorm2d(planes) self.conv3 = nn.Conv2d(planes, input_feature, kernel_size=1, bias= False, groups=group) self.bn3 = nn.InstanceNorm2d(input_feature) self.relu = nn.ReLU(inplace=True) def forward(self, input_0): primals_2 = self.conv1.weight primals_3 = self.conv2.weight primals_4 = self.conv3.weight primals_1 = input_0 output = call([primals_1, primals_2, primals_3, primals_4]) return output[0]
chiukin/RANet
ResBlock2
false
15,036
[ "Apache-2.0" ]
267
681a47d9b1f114653290678f02f2d3ecdf4010bc
https://github.com/chiukin/RANet/tree/681a47d9b1f114653290678f02f2d3ecdf4010bc
FPNHead
import torch import torch.nn as nn import torch.nn.parallel import torch.optim import torch.utils.data class FPNHead(nn.Module): def __init__(self, num_in, num_mid, num_out): super().__init__() self.block0 = nn.Conv2d(num_in, num_mid, kernel_size=3, padding=1, bias=False) self.block1 = nn.Conv2d(num_mid, num_out, kernel_size=3, padding=1, bias=False) def forward(self, x): x = nn.functional.relu(self.block0(x), inplace=True) x = nn.functional.relu(self.block1(x), inplace=True) return x def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'num_in': 4, 'num_mid': 4, 'num_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 from torch._inductor.runtime import triton_helpers import torch.nn as nn import torch.nn.parallel import torch.optim 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_relu_0(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 = tl.full([1], 0, tl.int32) tmp2 = triton_helpers.maximum(tmp1, tmp0) tl.store(in_out_ptr0 + x0, tmp2, xmask) @triton.jit def triton_poi_fused_relu_threshold_backward_1(in_out_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_out_ptr0 + x0, xmask) tmp1 = tl.full([1], 0, tl.int32) tmp2 = triton_helpers.maximum(tmp1, tmp0) tmp3 = 0.0 tmp4 = tmp2 <= tmp3 tl.store(in_out_ptr0 + x0, tmp2, xmask) tl.store(out_ptr0 + x0, tmp4, xmask) def call(args): primals_1, primals_2, primals_3 = args args.clear() assert_size_stride(primals_1, (4, 4, 3, 3), (36, 9, 3, 1)) assert_size_stride(primals_2, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_3, (4, 4, 3, 3), (36, 9, 3, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = extern_kernels.convolution(primals_2, primals_1, stride=(1, 1), padding=(1, 1), 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 get_raw_stream(0) triton_poi_fused_relu_0[grid(256)](buf1, 256, XBLOCK=128, num_warps =4, num_stages=1) buf2 = extern_kernels.convolution(buf1, primals_3, stride=(1, 1), padding=(1, 1), 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 = buf2 del buf2 buf4 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool) triton_poi_fused_relu_threshold_backward_1[grid(256)](buf3, buf4, 256, XBLOCK=128, num_warps=4, num_stages=1) return buf3, primals_1, primals_2, primals_3, buf1, buf4 class FPNHeadNew(nn.Module): def __init__(self, num_in, num_mid, num_out): super().__init__() self.block0 = nn.Conv2d(num_in, num_mid, kernel_size=3, padding=1, bias=False) self.block1 = nn.Conv2d(num_mid, num_out, kernel_size=3, padding=1, bias=False) def forward(self, input_0): primals_1 = self.block0.weight primals_3 = self.block1.weight primals_2 = input_0 output = call([primals_1, primals_2, primals_3]) return output[0]
choprahetarth/DeblurGANv2
FPNHead
false
15,037
[ "BSD-3-Clause" ]
321
e36dc2fef169b8a37036abe62192b6a925fb6c81
https://github.com/choprahetarth/DeblurGANv2/tree/e36dc2fef169b8a37036abe62192b6a925fb6c81
ScaledDotProductAttention
import torch import torch.nn as nn import torch.nn.functional as F class ScaledDotProductAttention(nn.Module): """ Scaled Dot-Product Attention """ def __init__(self, temperature, attn_dropout=0.1): super().__init__() self.temperature = temperature self.dropout = nn.Dropout(attn_dropout) self.softmax = nn.Softmax(dim=2) def forward(self, q, k, v): attn = torch.bmm(q, k.transpose(1, 2)) attn = attn / self.temperature log_attn = F.log_softmax(attn, 2) attn = self.softmax(attn) attn = self.dropout(attn) output = torch.bmm(attn, v) return output, attn, log_attn def get_inputs(): return [torch.rand([4, 4, 4]), torch.rand([4, 4, 4]), torch.rand([4, 4, 4]) ] def get_init_inputs(): return [[], {'temperature': 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__softmax_0(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 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 = 0.25 tmp16 = tmp14 * tmp15 tmp17 = tl_math.exp(tmp16) tl.store(out_ptr0 + x2, tmp17, xmask) tl.store(out_ptr1 + x2, tmp16, 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) @triton.jit def triton_poi_fused__log_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') tmp3 = tl.load(in_ptr0 + (1 + 4 * x1), xmask, eviction_policy='evict_last') tmp6 = tl.load(in_ptr0 + (2 + 4 * x1), xmask, eviction_policy='evict_last') tmp9 = tl.load(in_ptr0 + (3 + 4 * x1), xmask, eviction_policy='evict_last') tmp2 = tl_math.exp(tmp1) tmp4 = tl_math.exp(tmp3) tmp5 = tmp2 + tmp4 tmp7 = tl_math.exp(tmp6) tmp8 = tmp5 + tmp7 tmp10 = tl_math.exp(tmp9) tmp11 = tmp8 + tmp10 tmp12 = tl_math.log(tmp11) tmp13 = tmp0 - tmp12 tl.store(out_ptr0 + x2, tmp13, xmask) def call(args): arg0_1, arg1_1, arg2_1 = args args.clear() assert_size_stride(arg0_1, (4, 4, 4), (16, 4, 1)) assert_size_stride(arg1_1, (4, 4, 4), (16, 4, 1)) assert_size_stride(arg2_1, (4, 4, 4), (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) extern_kernels.bmm(arg1_1, reinterpret_tensor(arg0_1, (4, 4, 4), ( 16, 1, 4), 0), out=buf0) del arg0_1 del arg1_1 buf1 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) buf4 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused__softmax_0[grid(64)](buf0, buf1, buf4, 64, XBLOCK= 64, num_warps=1, num_stages=1) buf2 = buf0 del buf0 triton_poi_fused__softmax_1[grid(64)](buf1, buf2, 64, XBLOCK=64, num_warps=1, num_stages=1) buf3 = buf1 del buf1 extern_kernels.bmm(buf2, arg2_1, out=buf3) del arg2_1 buf5 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused__log_softmax_2[grid(64)](buf4, buf5, 64, XBLOCK=64, num_warps=1, num_stages=1) del buf4 return buf3, buf2, buf5 class ScaledDotProductAttentionNew(nn.Module): """ Scaled Dot-Product Attention """ def __init__(self, temperature, attn_dropout=0.1): super().__init__() self.temperature = temperature self.dropout = nn.Dropout(attn_dropout) self.softmax = nn.Softmax(dim=2) 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], output[2]
cjy97/FEAT
ScaledDotProductAttention
false
15,038
[ "MIT" ]
330
9d48b254bc5f0a2211c2aad0a60388a8a2c8081c
https://github.com/cjy97/FEAT/tree/9d48b254bc5f0a2211c2aad0a60388a8a2c8081c
MFBFusion
from _paritybench_helpers import _mock_config import time import torch from torch import nn class BaseModel(nn.Module): def __init__(self): super(BaseModel, self).__init__() self.model_name = str(type(self)) def load(self, path): self.load_state_dict(torch.load(path)) def save(self, name=None): if name is None: prefix = 'checkpoints/' + self.model_name + '_' name = time.strftime(prefix + '%m%d_%H:%M:%S.pth') torch.save(self.state_dict(), name) return name def forward(self, *input): raise NotImplementedError class MFBFusion(BaseModel): def __init__(self, mfb_fusion_option: 'MFBFusionOption'): super(MFBFusion, self).__init__() self.out_dim = mfb_fusion_option.outdim self.linear1 = nn.Linear(mfb_fusion_option.text_size, mfb_fusion_option.joint_emb_size) self.linear2 = nn.Linear(mfb_fusion_option.image_size, mfb_fusion_option.joint_emb_size) self.dropout = nn.Dropout(p=mfb_fusion_option.dropout) def forward(self, text_feat, image_feat): batch_size = text_feat.size(0) text_proj = self.linear1(text_feat) image_proj = self.linear2(image_feat) mm_eltwise = torch.mul(text_proj, image_proj) mm_drop = self.dropout(mm_eltwise) mm_resh = mm_drop.view(batch_size, 1, self.out_dim, -1) mm_sumpool = torch.sum(mm_resh, 3, keepdim=True) mfb_out = torch.squeeze(mm_sumpool) return mfb_out def get_inputs(): return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'mfb_fusion_option': _mock_config(outdim=4, text_size=4, joint_emb_size=4, image_size=4, dropout=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 import time 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_sum_0(in_ptr0, in_ptr1, 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.load(in_ptr1 + (r1 + 16 * x0), xmask, other=0.0) tmp2 = tmp0 * tmp1 tmp3 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK]) tmp5 = tl.where(xmask, tmp3, 0) tmp6 = tl.sum(tmp5, 1)[:, None] tl.store(out_ptr0 + x0, tmp6, 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, (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, 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_3, reinterpret_tensor(primals_1, (64, 4), (4, 1), 0), reinterpret_tensor(primals_2, (4, 4), (1, 4), 0 ), alpha=1, beta=1, out=buf0) del primals_2 del primals_3 buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32) extern_kernels.addmm(primals_5, reinterpret_tensor(primals_6, (64, 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((4, 1, 4, 1), (4, 16, 1, 16), torch.float32) get_raw_stream(0) triton_per_fused_sum_0[grid(16)](buf0, buf1, buf2, 16, 16, XBLOCK=1, num_warps=2, num_stages=1) return reinterpret_tensor(buf2, (4, 4), (4, 1), 0), reinterpret_tensor( primals_1, (64, 4), (4, 1), 0), buf0, reinterpret_tensor(primals_6, (64, 4), (4, 1), 0), buf1 class BaseModel(nn.Module): def __init__(self): super(BaseModel, self).__init__() self.model_name = str(type(self)) def load(self, path): self.load_state_dict(torch.load(path)) def save(self, name=None): if name is None: prefix = 'checkpoints/' + self.model_name + '_' name = time.strftime(prefix + '%m%d_%H:%M:%S.pth') torch.save(self.state_dict(), name) return name def forward(self, *input): raise NotImplementedError class MFBFusionNew(BaseModel): def __init__(self, mfb_fusion_option: 'MFBFusionOption'): super(MFBFusionNew, self).__init__() self.out_dim = mfb_fusion_option.outdim self.linear1 = nn.Linear(mfb_fusion_option.text_size, mfb_fusion_option.joint_emb_size) self.linear2 = nn.Linear(mfb_fusion_option.image_size, mfb_fusion_option.joint_emb_size) self.dropout = nn.Dropout(p=mfb_fusion_option.dropout) def forward(self, input_0, input_1): primals_2 = self.linear1.weight primals_3 = self.linear1.bias primals_4 = self.linear2.weight primals_5 = self.linear2.bias primals_1 = input_0 primals_6 = input_1 output = call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6]) return output[0]
chorseng/UMD
MFBFusion
false
15,039
[ "MIT" ]
48
680681fea76abcea02ff5f351727bcbb468c372a
https://github.com/chorseng/UMD/tree/680681fea76abcea02ff5f351727bcbb468c372a
SimpleConvNetBlock
import torch import torch.nn as nn class SimpleConvNetBlock(nn.Module): def __init__(self, in_channels, out_channels, kernel): nn.Module.__init__(self) self.conv = nn.Conv2d(in_channels=in_channels, out_channels= out_channels, kernel_size=kernel, padding=1) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(2) def forward(self, x): y = self.conv(x) y = self.relu(y) y = self.maxpool(y) return y def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_channels': 4, 'out_channels': 4, 'kernel': 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 @triton.jit def triton_poi_fused_convolution_relu_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 144 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x3 = xindex x1 = xindex // 9 % 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_max_pool2d_with_indices_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 x0 = xindex tmp0 = tl.load(in_ptr0 + 9 * x0, xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr0 + (1 + 9 * x0), xmask, eviction_policy='evict_last') tmp3 = tl.load(in_ptr0 + (3 + 9 * x0), xmask, eviction_policy='evict_last') tmp5 = tl.load(in_ptr0 + (4 + 9 * x0), 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 + x0, tmp6, xmask) tl.store(out_ptr1 + x0, tmp16, 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,), (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 = extern_kernels.convolution(primals_3, primals_1, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf0, (4, 4, 3, 3), (36, 9, 3, 1)) buf1 = buf0 del buf0 get_raw_stream(0) triton_poi_fused_convolution_relu_0[grid(144)](buf1, primals_2, 144, XBLOCK=128, num_warps=4, num_stages=1) del primals_2 buf2 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.float32) buf3 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 1, 1), torch.int8) triton_poi_fused_max_pool2d_with_indices_1[grid(16)](buf1, buf2, buf3, 16, XBLOCK=16, num_warps=1, num_stages=1) return buf2, primals_1, primals_3, buf1, buf3 class SimpleConvNetBlockNew(nn.Module): def __init__(self, in_channels, out_channels, kernel): nn.Module.__init__(self) self.conv = nn.Conv2d(in_channels=in_channels, out_channels= out_channels, kernel_size=kernel, padding=1) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(2) def forward(self, input_0): primals_1 = self.conv.weight primals_2 = self.conv.bias primals_3 = input_0 output = call([primals_1, primals_2, primals_3]) return output[0]
cle-ros/RoutingNetworks
SimpleConvNetBlock
false
15,040
[ "Apache-2.0" ]
63
0f1fe1221c67a224a02bca6247d3c4488ede0a04
https://github.com/cle-ros/RoutingNetworks/tree/0f1fe1221c67a224a02bca6247d3c4488ede0a04
PositionwiseFeedForward
import math import torch from torch import nn def gelu(x): return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))) class PositionwiseFeedForward(nn.Module): """ A two-layer Feed-Forward-Network with residual layer norm. Args: d_model (int): the size of input for the first-layer of the FFN. d_ff (int): the hidden layer size of the second-layer of the FNN. dropout (float): dropout probability in :math:`[0, 1)`. """ def __init__(self, d_model, d_ff, dropout=0.1): super(PositionwiseFeedForward, self).__init__() self.w_1 = nn.Linear(d_model, d_ff) self.w_2 = nn.Linear(d_ff, d_model) self.layer_norm = nn.LayerNorm(d_model, eps=1e-06) self.actv = gelu self.dropout_1 = nn.Dropout(dropout) self.dropout_2 = nn.Dropout(dropout) def forward(self, x): inter = self.dropout_1(self.actv(self.w_1(self.layer_norm(x)))) output = self.dropout_2(self.w_2(inter)) return output + x def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'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.triton_helpers import libdevice 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_native_layer_norm_0(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-06 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_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 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) @triton.jit def triton_poi_fused_add_mul_pow_tanh_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 tmp0 = tl.load(in_ptr0 + x0, xmask) tmp1 = 0.5 tmp2 = tmp0 * tmp1 tmp3 = tmp0 * tmp0 tmp4 = tmp3 * tmp0 tmp5 = 0.044715 tmp6 = tmp4 * tmp5 tmp7 = tmp0 + tmp6 tmp8 = 0.7978845608028654 tmp9 = tmp7 * tmp8 tmp10 = libdevice.tanh(tmp9) tmp11 = 1.0 tmp12 = tmp10 + tmp11 tmp13 = tmp2 * tmp12 tl.store(out_ptr0 + x0, tmp13, xmask) @triton.jit def triton_poi_fused_add_3(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) 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,), (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,)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32) buf1 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32) get_raw_stream(0) triton_poi_fused_native_layer_norm_0[grid(64)](primals_3, buf0, buf1, 64, XBLOCK=64, num_warps=1, num_stages=1) buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused_native_layer_norm_1[grid(256)](primals_3, buf0, buf1, primals_1, primals_2, buf2, 256, XBLOCK=256, num_warps=4, num_stages=1) del buf0 del buf1 del primals_1 del primals_2 buf3 = empty_strided_cuda((64, 4), (4, 1), torch.float32) extern_kernels.addmm(primals_5, reinterpret_tensor(buf2, (64, 4), ( 4, 1), 0), reinterpret_tensor(primals_4, (4, 4), (1, 4), 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_add_mul_pow_tanh_2[grid(256)](buf3, buf4, 256, XBLOCK=256, num_warps=4, num_stages=1) buf5 = empty_strided_cuda((64, 4), (4, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(buf4, (64, 4), (4, 1), 0), reinterpret_tensor(primals_6, (4, 4), (1, 4), 0), out=buf5) buf6 = reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0) del buf5 triton_poi_fused_add_3[grid(256)](buf6, primals_7, primals_3, 256, XBLOCK=256, num_warps=4, num_stages=1) del primals_7 return buf6, primals_3, reinterpret_tensor(buf2, (64, 4), (4, 1), 0 ), buf3, reinterpret_tensor(buf4, (64, 4), (4, 1), 0 ), primals_6, primals_4 def gelu(x): return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))) class PositionwiseFeedForwardNew(nn.Module): """ A two-layer Feed-Forward-Network with residual layer norm. Args: d_model (int): the size of input for the first-layer of the FFN. d_ff (int): the hidden layer size of the second-layer of the FNN. dropout (float): dropout probability in :math:`[0, 1)`. """ def __init__(self, d_model, d_ff, dropout=0.1): super(PositionwiseFeedForwardNew, self).__init__() self.w_1 = nn.Linear(d_model, d_ff) self.w_2 = nn.Linear(d_ff, d_model) self.layer_norm = nn.LayerNorm(d_model, eps=1e-06) self.actv = gelu self.dropout_1 = nn.Dropout(dropout) self.dropout_2 = nn.Dropout(dropout) def forward(self, input_0): primals_4 = self.w_1.weight primals_1 = self.w_1.bias primals_6 = self.w_2.weight primals_2 = self.w_2.bias primals_5 = self.layer_norm.weight primals_7 = self.layer_norm.bias primals_3 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7]) return output[0]
clairett/fast-bert
PositionwiseFeedForward
false
15,041
[ "Apache-2.0" ]
1,542
506771b930aa70e7ca2852e5e8ebb14656d97bfa
https://github.com/clairett/fast-bert/tree/506771b930aa70e7ca2852e5e8ebb14656d97bfa
SparsemaxBisect
from torch.autograd import Function import torch import torch.nn as nn def sparsemax_bisect(X, dim=-1, n_iter=50, ensure_sum_one=True): """sparsemax: normalizing sparse transform (a la softmax), via bisection. Solves the projection: min_p ||x - p||_2 s.t. p >= 0, sum(p) == 1. Parameters ---------- X : torch.Tensor The input tensor. dim : int The dimension along which to apply sparsemax. n_iter : int Number of bisection iterations. For float32, 24 iterations should suffice for machine precision. ensure_sum_one : bool, Whether to divide the result by its sum. If false, the result might sum to close but not exactly 1, which might cause downstream problems. Note: This function does not yet support normalizing along anything except the last dimension. Please use transposing and views to achieve more general behavior. Returns ------- P : torch tensor, same shape as X The projection result, such that P.sum(dim=dim) == 1 elementwise. """ return SparsemaxBisectFunction.apply(X, dim, n_iter, ensure_sum_one) class EntmaxBisectFunction(Function): @classmethod def _gp(cls, x, alpha): return x ** (alpha - 1) @classmethod def _gp_inv(cls, y, alpha): return y ** (1 / (alpha - 1)) @classmethod def _p(cls, X, alpha): return cls._gp_inv(torch.clamp(X, min=0), alpha) @classmethod def forward(cls, ctx, X, alpha=1.5, dim=-1, n_iter=50, ensure_sum_one=True ): if not isinstance(alpha, torch.Tensor): alpha = torch.tensor(alpha, dtype=X.dtype, device=X.device) alpha_shape = list(X.shape) alpha_shape[dim] = 1 alpha = alpha.expand(*alpha_shape) ctx.alpha = alpha ctx.dim = dim d = X.shape[dim] max_val, _ = X.max(dim=dim, keepdim=True) X = X * (alpha - 1) max_val = max_val * (alpha - 1) tau_lo = max_val - cls._gp(1, alpha) tau_hi = max_val - cls._gp(1 / d, alpha) f_lo = cls._p(X - tau_lo, alpha).sum(dim) - 1 dm = tau_hi - tau_lo for it in range(n_iter): dm /= 2 tau_m = tau_lo + dm p_m = cls._p(X - tau_m, alpha) f_m = p_m.sum(dim) - 1 mask = (f_m * f_lo >= 0).unsqueeze(dim) tau_lo = torch.where(mask, tau_m, tau_lo) if ensure_sum_one: p_m /= p_m.sum(dim=dim).unsqueeze(dim=dim) ctx.save_for_backward(p_m) return p_m @classmethod def backward(cls, ctx, dY): Y, = ctx.saved_tensors gppr = torch.where(Y > 0, Y ** (2 - ctx.alpha), Y.new_zeros(1)) dX = dY * gppr q = dX.sum(ctx.dim) / gppr.sum(ctx.dim) q = q.unsqueeze(ctx.dim) dX -= q * gppr d_alpha = None if ctx.needs_input_grad[1]: S = torch.where(Y > 0, Y * torch.log(Y), Y.new_zeros(1)) ent = S.sum(ctx.dim).unsqueeze(ctx.dim) Y_skewed = gppr / gppr.sum(ctx.dim).unsqueeze(ctx.dim) d_alpha = dY * (Y - Y_skewed) / (ctx.alpha - 1) ** 2 d_alpha -= dY * (S - Y_skewed * ent) / (ctx.alpha - 1) d_alpha = d_alpha.sum(ctx.dim).unsqueeze(ctx.dim) return dX, d_alpha, None, None, None class SparsemaxBisectFunction(EntmaxBisectFunction): @classmethod def _gp(cls, x, alpha): return x @classmethod def _gp_inv(cls, y, alpha): return y @classmethod def _p(cls, x, alpha): return torch.clamp(x, min=0) @classmethod def forward(cls, ctx, X, dim=-1, n_iter=50, ensure_sum_one=True): return super().forward(ctx, X, alpha=2, dim=dim, n_iter=50, ensure_sum_one=True) @classmethod def backward(cls, ctx, dY): Y, = ctx.saved_tensors gppr = Y > 0 dX = dY * gppr q = dX.sum(ctx.dim) / gppr.sum(ctx.dim) q = q.unsqueeze(ctx.dim) dX -= q * gppr return dX, None, None, None class SparsemaxBisect(nn.Module): def __init__(self, dim=-1, n_iter=None): """sparsemax: normalizing sparse transform (a la softmax) via bisection Solves the projection: min_p ||x - p||_2 s.t. p >= 0, sum(p) == 1. Parameters ---------- dim : int The dimension along which to apply sparsemax. n_iter : int Number of bisection iterations. For float32, 24 iterations should suffice for machine precision. """ self.dim = dim self.n_iter = n_iter super().__init__() def forward(self, X): return sparsemax_bisect(X, dim=self.dim, n_iter=self.n_iter) 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.autograd import Function 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_clamp_div_max_mul_sub_sum_where_0(in_out_ptr8, 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_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 = triton_helpers.maximum(tmp0, tmp1) tmp4 = triton_helpers.maximum(tmp2, tmp3) tmp6 = triton_helpers.maximum(tmp4, tmp5) tmp7 = 1.0 tmp8 = tmp6 - tmp7 tmp9 = 0.25 tmp10 = tmp6 - tmp9 tmp11 = tmp10 - tmp8 tmp12 = 0.5 tmp13 = tmp11 * tmp12 tmp14 = tmp8 + tmp13 tmp15 = tmp0 - tmp14 tmp16 = 0.0 tmp17 = triton_helpers.maximum(tmp15, tmp16) tmp18 = tmp1 - tmp14 tmp19 = triton_helpers.maximum(tmp18, tmp16) tmp20 = tmp17 + tmp19 tmp21 = tmp3 - tmp14 tmp22 = triton_helpers.maximum(tmp21, tmp16) tmp23 = tmp20 + tmp22 tmp24 = tmp5 - tmp14 tmp25 = triton_helpers.maximum(tmp24, tmp16) tmp26 = tmp23 + tmp25 tmp27 = tmp26 - tmp7 tmp28 = tmp0 - tmp8 tmp29 = triton_helpers.maximum(tmp28, tmp16) tmp30 = tmp1 - tmp8 tmp31 = triton_helpers.maximum(tmp30, tmp16) tmp32 = tmp29 + tmp31 tmp33 = tmp3 - tmp8 tmp34 = triton_helpers.maximum(tmp33, tmp16) tmp35 = tmp32 + tmp34 tmp36 = tmp5 - tmp8 tmp37 = triton_helpers.maximum(tmp36, tmp16) tmp38 = tmp35 + tmp37 tmp39 = tmp38 - tmp7 tmp40 = tmp27 * tmp39 tmp41 = tmp40 >= tmp16 tmp42 = tl.where(tmp41, tmp14, tmp8) tmp43 = tmp13 * tmp12 tmp44 = tmp42 + tmp43 tmp45 = tmp0 - tmp44 tmp46 = triton_helpers.maximum(tmp45, tmp16) tmp47 = tmp1 - tmp44 tmp48 = triton_helpers.maximum(tmp47, tmp16) tmp49 = tmp46 + tmp48 tmp50 = tmp3 - tmp44 tmp51 = triton_helpers.maximum(tmp50, tmp16) tmp52 = tmp49 + tmp51 tmp53 = tmp5 - tmp44 tmp54 = triton_helpers.maximum(tmp53, tmp16) tmp55 = tmp52 + tmp54 tmp56 = tmp55 - tmp7 tmp57 = tmp56 * tmp39 tmp58 = tmp57 >= tmp16 tmp59 = tl.where(tmp58, tmp44, tmp42) tmp60 = tmp43 * tmp12 tmp61 = tmp59 + tmp60 tmp62 = tmp0 - tmp61 tmp63 = triton_helpers.maximum(tmp62, tmp16) tmp64 = tmp1 - tmp61 tmp65 = triton_helpers.maximum(tmp64, tmp16) tmp66 = tmp63 + tmp65 tmp67 = tmp3 - tmp61 tmp68 = triton_helpers.maximum(tmp67, tmp16) tmp69 = tmp66 + tmp68 tmp70 = tmp5 - tmp61 tmp71 = triton_helpers.maximum(tmp70, tmp16) tmp72 = tmp69 + tmp71 tmp73 = tmp72 - tmp7 tmp74 = tmp73 * tmp39 tmp75 = tmp74 >= tmp16 tmp76 = tl.where(tmp75, tmp61, tmp59) tmp77 = tmp60 * tmp12 tmp78 = tmp76 + tmp77 tmp79 = tmp0 - tmp78 tmp80 = triton_helpers.maximum(tmp79, tmp16) tmp81 = tmp1 - tmp78 tmp82 = triton_helpers.maximum(tmp81, tmp16) tmp83 = tmp80 + tmp82 tmp84 = tmp3 - tmp78 tmp85 = triton_helpers.maximum(tmp84, tmp16) tmp86 = tmp83 + tmp85 tmp87 = tmp5 - tmp78 tmp88 = triton_helpers.maximum(tmp87, tmp16) tmp89 = tmp86 + tmp88 tmp90 = tmp89 - tmp7 tmp91 = tmp90 * tmp39 tmp92 = tmp91 >= tmp16 tmp93 = tl.where(tmp92, tmp78, tmp76) tmp94 = tmp77 * tmp12 tmp95 = tmp93 + tmp94 tmp96 = tmp0 - tmp95 tmp97 = triton_helpers.maximum(tmp96, tmp16) tmp98 = tmp1 - tmp95 tmp99 = triton_helpers.maximum(tmp98, tmp16) tmp100 = tmp97 + tmp99 tmp101 = tmp3 - tmp95 tmp102 = triton_helpers.maximum(tmp101, tmp16) tmp103 = tmp100 + tmp102 tmp104 = tmp5 - tmp95 tmp105 = triton_helpers.maximum(tmp104, tmp16) tmp106 = tmp103 + tmp105 tmp107 = tmp106 - tmp7 tmp108 = tmp107 * tmp39 tmp109 = tmp108 >= tmp16 tmp110 = tl.where(tmp109, tmp95, tmp93) tmp111 = tmp94 * tmp12 tmp112 = tmp110 + tmp111 tmp113 = tmp0 - tmp112 tmp114 = triton_helpers.maximum(tmp113, tmp16) tmp115 = tmp1 - tmp112 tmp116 = triton_helpers.maximum(tmp115, tmp16) tmp117 = tmp114 + tmp116 tmp118 = tmp3 - tmp112 tmp119 = triton_helpers.maximum(tmp118, tmp16) tmp120 = tmp117 + tmp119 tmp121 = tmp5 - tmp112 tmp122 = triton_helpers.maximum(tmp121, tmp16) tmp123 = tmp120 + tmp122 tmp124 = tmp123 - tmp7 tmp125 = tmp124 * tmp39 tmp126 = tmp125 >= tmp16 tmp127 = tl.where(tmp126, tmp112, tmp110) tmp128 = tmp111 * tmp12 tmp129 = tmp127 + tmp128 tmp130 = tmp0 - tmp129 tmp131 = triton_helpers.maximum(tmp130, tmp16) tmp132 = tmp1 - tmp129 tmp133 = triton_helpers.maximum(tmp132, tmp16) tmp134 = tmp131 + tmp133 tmp135 = tmp3 - tmp129 tmp136 = triton_helpers.maximum(tmp135, tmp16) tmp137 = tmp134 + tmp136 tmp138 = tmp5 - tmp129 tmp139 = triton_helpers.maximum(tmp138, tmp16) tmp140 = tmp137 + tmp139 tmp141 = tmp140 - tmp7 tmp142 = tmp141 * tmp39 tmp143 = tmp142 >= tmp16 tmp144 = tl.where(tmp143, tmp129, tmp127) tmp145 = tmp128 * tmp12 tmp146 = tmp144 + tmp145 tmp147 = tmp0 - tmp146 tmp148 = triton_helpers.maximum(tmp147, tmp16) tmp149 = tmp1 - tmp146 tmp150 = triton_helpers.maximum(tmp149, tmp16) tmp151 = tmp148 + tmp150 tmp152 = tmp3 - tmp146 tmp153 = triton_helpers.maximum(tmp152, tmp16) tmp154 = tmp151 + tmp153 tmp155 = tmp5 - tmp146 tmp156 = triton_helpers.maximum(tmp155, tmp16) tmp157 = tmp154 + tmp156 tmp158 = tmp157 - tmp7 tmp159 = tmp158 * tmp39 tmp160 = tmp159 >= tmp16 tmp161 = tl.where(tmp160, tmp146, tmp144) tmp162 = tmp145 * tmp12 tmp163 = tmp161 + tmp162 tmp164 = tmp0 - tmp163 tmp165 = triton_helpers.maximum(tmp164, tmp16) tmp166 = tmp1 - tmp163 tmp167 = triton_helpers.maximum(tmp166, tmp16) tmp168 = tmp165 + tmp167 tmp169 = tmp3 - tmp163 tmp170 = triton_helpers.maximum(tmp169, tmp16) tmp171 = tmp168 + tmp170 tmp172 = tmp5 - tmp163 tmp173 = triton_helpers.maximum(tmp172, tmp16) tmp174 = tmp171 + tmp173 tmp175 = tmp174 - tmp7 tmp176 = tmp175 * tmp39 tmp177 = tmp176 >= tmp16 tmp178 = tl.where(tmp177, tmp163, tmp161) tmp179 = tmp162 * tmp12 tmp180 = tmp178 + tmp179 tmp181 = tmp0 - tmp180 tmp182 = triton_helpers.maximum(tmp181, tmp16) tmp183 = tmp1 - tmp180 tmp184 = triton_helpers.maximum(tmp183, tmp16) tmp185 = tmp182 + tmp184 tmp186 = tmp3 - tmp180 tmp187 = triton_helpers.maximum(tmp186, tmp16) tmp188 = tmp185 + tmp187 tmp189 = tmp5 - tmp180 tmp190 = triton_helpers.maximum(tmp189, tmp16) tmp191 = tmp188 + tmp190 tmp192 = tmp191 - tmp7 tmp193 = tmp192 * tmp39 tmp194 = tmp193 >= tmp16 tmp195 = tl.where(tmp194, tmp180, tmp178) tmp196 = tmp179 * tmp12 tmp197 = tmp195 + tmp196 tmp198 = tmp0 - tmp197 tmp199 = triton_helpers.maximum(tmp198, tmp16) tmp200 = tmp1 - tmp197 tmp201 = triton_helpers.maximum(tmp200, tmp16) tmp202 = tmp199 + tmp201 tmp203 = tmp3 - tmp197 tmp204 = triton_helpers.maximum(tmp203, tmp16) tmp205 = tmp202 + tmp204 tmp206 = tmp5 - tmp197 tmp207 = triton_helpers.maximum(tmp206, tmp16) tmp208 = tmp205 + tmp207 tmp209 = tmp208 - tmp7 tmp210 = tmp209 * tmp39 tmp211 = tmp210 >= tmp16 tmp212 = tl.where(tmp211, tmp197, tmp195) tmp213 = tmp196 * tmp12 tmp214 = tmp212 + tmp213 tmp215 = tmp0 - tmp214 tmp216 = triton_helpers.maximum(tmp215, tmp16) tmp217 = tmp1 - tmp214 tmp218 = triton_helpers.maximum(tmp217, tmp16) tmp219 = tmp216 + tmp218 tmp220 = tmp3 - tmp214 tmp221 = triton_helpers.maximum(tmp220, tmp16) tmp222 = tmp219 + tmp221 tmp223 = tmp5 - tmp214 tmp224 = triton_helpers.maximum(tmp223, tmp16) tmp225 = tmp222 + tmp224 tmp226 = tmp225 - tmp7 tmp227 = tmp226 * tmp39 tmp228 = tmp227 >= tmp16 tmp229 = tl.where(tmp228, tmp214, tmp212) tmp230 = tmp213 * tmp12 tmp231 = tmp229 + tmp230 tmp232 = tmp0 - tmp231 tmp233 = triton_helpers.maximum(tmp232, tmp16) tmp234 = tmp1 - tmp231 tmp235 = triton_helpers.maximum(tmp234, tmp16) tmp236 = tmp233 + tmp235 tmp237 = tmp3 - tmp231 tmp238 = triton_helpers.maximum(tmp237, tmp16) tmp239 = tmp236 + tmp238 tmp240 = tmp5 - tmp231 tmp241 = triton_helpers.maximum(tmp240, tmp16) tmp242 = tmp239 + tmp241 tmp243 = tmp242 - tmp7 tmp244 = tmp243 * tmp39 tmp245 = tmp244 >= tmp16 tmp246 = tl.where(tmp245, tmp231, tmp229) tmp247 = tmp230 * tmp12 tmp248 = tmp246 + tmp247 tmp249 = tmp0 - tmp248 tmp250 = triton_helpers.maximum(tmp249, tmp16) tmp251 = tmp1 - tmp248 tmp252 = triton_helpers.maximum(tmp251, tmp16) tmp253 = tmp250 + tmp252 tmp254 = tmp3 - tmp248 tmp255 = triton_helpers.maximum(tmp254, tmp16) tmp256 = tmp253 + tmp255 tmp257 = tmp5 - tmp248 tmp258 = triton_helpers.maximum(tmp257, tmp16) tmp259 = tmp256 + tmp258 tmp260 = tmp259 - tmp7 tmp261 = tmp260 * tmp39 tmp262 = tmp261 >= tmp16 tmp263 = tl.where(tmp262, tmp248, tmp246) tmp264 = tmp247 * tmp12 tmp265 = tmp263 + tmp264 tmp266 = tmp0 - tmp265 tmp267 = triton_helpers.maximum(tmp266, tmp16) tmp268 = tmp1 - tmp265 tmp269 = triton_helpers.maximum(tmp268, tmp16) tmp270 = tmp267 + tmp269 tmp271 = tmp3 - tmp265 tmp272 = triton_helpers.maximum(tmp271, tmp16) tmp273 = tmp270 + tmp272 tmp274 = tmp5 - tmp265 tmp275 = triton_helpers.maximum(tmp274, tmp16) tmp276 = tmp273 + tmp275 tmp277 = tmp276 - tmp7 tmp278 = tmp277 * tmp39 tmp279 = tmp278 >= tmp16 tmp280 = tl.where(tmp279, tmp265, tmp263) tmp281 = tmp264 * tmp12 tmp282 = tmp280 + tmp281 tmp283 = tmp0 - tmp282 tmp284 = triton_helpers.maximum(tmp283, tmp16) tmp285 = tmp1 - tmp282 tmp286 = triton_helpers.maximum(tmp285, tmp16) tmp287 = tmp284 + tmp286 tmp288 = tmp3 - tmp282 tmp289 = triton_helpers.maximum(tmp288, tmp16) tmp290 = tmp287 + tmp289 tmp291 = tmp5 - tmp282 tmp292 = triton_helpers.maximum(tmp291, tmp16) tmp293 = tmp290 + tmp292 tmp294 = tmp293 - tmp7 tmp295 = tmp294 * tmp39 tmp296 = tmp295 >= tmp16 tmp297 = tl.where(tmp296, tmp282, tmp280) tmp298 = tmp281 * tmp12 tmp299 = tmp297 + tmp298 tmp300 = tmp0 - tmp299 tmp301 = triton_helpers.maximum(tmp300, tmp16) tmp302 = tmp1 - tmp299 tmp303 = triton_helpers.maximum(tmp302, tmp16) tmp304 = tmp301 + tmp303 tmp305 = tmp3 - tmp299 tmp306 = triton_helpers.maximum(tmp305, tmp16) tmp307 = tmp304 + tmp306 tmp308 = tmp5 - tmp299 tmp309 = triton_helpers.maximum(tmp308, tmp16) tmp310 = tmp307 + tmp309 tmp311 = tmp310 - tmp7 tmp312 = tmp311 * tmp39 tmp313 = tmp312 >= tmp16 tmp314 = tl.where(tmp313, tmp299, tmp297) tl.store(in_out_ptr8 + x0, tmp314, xmask) @triton.jit def triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_1(in_out_ptr16, 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 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') tmp31 = tl.load(in_ptr1 + x0, xmask) tmp2 = triton_helpers.maximum(tmp0, tmp1) tmp4 = triton_helpers.maximum(tmp2, tmp3) tmp6 = triton_helpers.maximum(tmp4, tmp5) tmp7 = 0.25 tmp8 = tmp6 - tmp7 tmp9 = 1.0 tmp10 = tmp6 - tmp9 tmp11 = tmp8 - tmp10 tmp12 = 0.5 tmp13 = tmp11 * tmp12 tmp14 = tmp13 * tmp12 tmp15 = tmp14 * tmp12 tmp16 = tmp15 * tmp12 tmp17 = tmp16 * tmp12 tmp18 = tmp17 * tmp12 tmp19 = tmp18 * tmp12 tmp20 = tmp19 * tmp12 tmp21 = tmp20 * tmp12 tmp22 = tmp21 * tmp12 tmp23 = tmp22 * tmp12 tmp24 = tmp23 * tmp12 tmp25 = tmp24 * tmp12 tmp26 = tmp25 * tmp12 tmp27 = tmp26 * tmp12 tmp28 = tmp27 * tmp12 tmp29 = tmp28 * tmp12 tmp30 = tmp29 * tmp12 tmp32 = tmp31 + tmp30 tmp33 = tmp0 - tmp32 tmp34 = 0.0 tmp35 = triton_helpers.maximum(tmp33, tmp34) tmp36 = tmp1 - tmp32 tmp37 = triton_helpers.maximum(tmp36, tmp34) tmp38 = tmp35 + tmp37 tmp39 = tmp3 - tmp32 tmp40 = triton_helpers.maximum(tmp39, tmp34) tmp41 = tmp38 + tmp40 tmp42 = tmp5 - tmp32 tmp43 = triton_helpers.maximum(tmp42, tmp34) tmp44 = tmp41 + tmp43 tmp45 = tmp44 - tmp9 tmp46 = tmp0 - tmp10 tmp47 = triton_helpers.maximum(tmp46, tmp34) tmp48 = tmp1 - tmp10 tmp49 = triton_helpers.maximum(tmp48, tmp34) tmp50 = tmp47 + tmp49 tmp51 = tmp3 - tmp10 tmp52 = triton_helpers.maximum(tmp51, tmp34) tmp53 = tmp50 + tmp52 tmp54 = tmp5 - tmp10 tmp55 = triton_helpers.maximum(tmp54, tmp34) tmp56 = tmp53 + tmp55 tmp57 = tmp56 - tmp9 tmp58 = tmp45 * tmp57 tmp59 = tmp58 >= tmp34 tmp60 = tl.where(tmp59, tmp32, tmp31) tmp61 = tmp30 * tmp12 tmp62 = tmp60 + tmp61 tmp63 = tmp0 - tmp62 tmp64 = triton_helpers.maximum(tmp63, tmp34) tmp65 = tmp1 - tmp62 tmp66 = triton_helpers.maximum(tmp65, tmp34) tmp67 = tmp64 + tmp66 tmp68 = tmp3 - tmp62 tmp69 = triton_helpers.maximum(tmp68, tmp34) tmp70 = tmp67 + tmp69 tmp71 = tmp5 - tmp62 tmp72 = triton_helpers.maximum(tmp71, tmp34) tmp73 = tmp70 + tmp72 tmp74 = tmp73 - tmp9 tmp75 = tmp74 * tmp57 tmp76 = tmp75 >= tmp34 tmp77 = tl.where(tmp76, tmp62, tmp60) tmp78 = tmp61 * tmp12 tmp79 = tmp77 + tmp78 tmp80 = tmp0 - tmp79 tmp81 = triton_helpers.maximum(tmp80, tmp34) tmp82 = tmp1 - tmp79 tmp83 = triton_helpers.maximum(tmp82, tmp34) tmp84 = tmp81 + tmp83 tmp85 = tmp3 - tmp79 tmp86 = triton_helpers.maximum(tmp85, tmp34) tmp87 = tmp84 + tmp86 tmp88 = tmp5 - tmp79 tmp89 = triton_helpers.maximum(tmp88, tmp34) tmp90 = tmp87 + tmp89 tmp91 = tmp90 - tmp9 tmp92 = tmp91 * tmp57 tmp93 = tmp92 >= tmp34 tmp94 = tl.where(tmp93, tmp79, tmp77) tmp95 = tmp78 * tmp12 tmp96 = tmp94 + tmp95 tmp97 = tmp0 - tmp96 tmp98 = triton_helpers.maximum(tmp97, tmp34) tmp99 = tmp1 - tmp96 tmp100 = triton_helpers.maximum(tmp99, tmp34) tmp101 = tmp98 + tmp100 tmp102 = tmp3 - tmp96 tmp103 = triton_helpers.maximum(tmp102, tmp34) tmp104 = tmp101 + tmp103 tmp105 = tmp5 - tmp96 tmp106 = triton_helpers.maximum(tmp105, tmp34) tmp107 = tmp104 + tmp106 tmp108 = tmp107 - tmp9 tmp109 = tmp108 * tmp57 tmp110 = tmp109 >= tmp34 tmp111 = tl.where(tmp110, tmp96, tmp94) tmp112 = tmp95 * tmp12 tmp113 = tmp111 + tmp112 tmp114 = tmp0 - tmp113 tmp115 = triton_helpers.maximum(tmp114, tmp34) tmp116 = tmp1 - tmp113 tmp117 = triton_helpers.maximum(tmp116, tmp34) tmp118 = tmp115 + tmp117 tmp119 = tmp3 - tmp113 tmp120 = triton_helpers.maximum(tmp119, tmp34) tmp121 = tmp118 + tmp120 tmp122 = tmp5 - tmp113 tmp123 = triton_helpers.maximum(tmp122, tmp34) tmp124 = tmp121 + tmp123 tmp125 = tmp124 - tmp9 tmp126 = tmp125 * tmp57 tmp127 = tmp126 >= tmp34 tmp128 = tl.where(tmp127, tmp113, tmp111) tmp129 = tmp112 * tmp12 tmp130 = tmp128 + tmp129 tmp131 = tmp0 - tmp130 tmp132 = triton_helpers.maximum(tmp131, tmp34) tmp133 = tmp1 - tmp130 tmp134 = triton_helpers.maximum(tmp133, tmp34) tmp135 = tmp132 + tmp134 tmp136 = tmp3 - tmp130 tmp137 = triton_helpers.maximum(tmp136, tmp34) tmp138 = tmp135 + tmp137 tmp139 = tmp5 - tmp130 tmp140 = triton_helpers.maximum(tmp139, tmp34) tmp141 = tmp138 + tmp140 tmp142 = tmp141 - tmp9 tmp143 = tmp142 * tmp57 tmp144 = tmp143 >= tmp34 tmp145 = tl.where(tmp144, tmp130, tmp128) tmp146 = tmp129 * tmp12 tmp147 = tmp145 + tmp146 tmp148 = tmp0 - tmp147 tmp149 = triton_helpers.maximum(tmp148, tmp34) tmp150 = tmp1 - tmp147 tmp151 = triton_helpers.maximum(tmp150, tmp34) tmp152 = tmp149 + tmp151 tmp153 = tmp3 - tmp147 tmp154 = triton_helpers.maximum(tmp153, tmp34) tmp155 = tmp152 + tmp154 tmp156 = tmp5 - tmp147 tmp157 = triton_helpers.maximum(tmp156, tmp34) tmp158 = tmp155 + tmp157 tmp159 = tmp158 - tmp9 tmp160 = tmp159 * tmp57 tmp161 = tmp160 >= tmp34 tmp162 = tl.where(tmp161, tmp147, tmp145) tmp163 = tmp146 * tmp12 tmp164 = tmp162 + tmp163 tmp165 = tmp0 - tmp164 tmp166 = triton_helpers.maximum(tmp165, tmp34) tmp167 = tmp1 - tmp164 tmp168 = triton_helpers.maximum(tmp167, tmp34) tmp169 = tmp166 + tmp168 tmp170 = tmp3 - tmp164 tmp171 = triton_helpers.maximum(tmp170, tmp34) tmp172 = tmp169 + tmp171 tmp173 = tmp5 - tmp164 tmp174 = triton_helpers.maximum(tmp173, tmp34) tmp175 = tmp172 + tmp174 tmp176 = tmp175 - tmp9 tmp177 = tmp176 * tmp57 tmp178 = tmp177 >= tmp34 tmp179 = tl.where(tmp178, tmp164, tmp162) tmp180 = tmp163 * tmp12 tmp181 = tmp179 + tmp180 tmp182 = tmp0 - tmp181 tmp183 = triton_helpers.maximum(tmp182, tmp34) tmp184 = tmp1 - tmp181 tmp185 = triton_helpers.maximum(tmp184, tmp34) tmp186 = tmp183 + tmp185 tmp187 = tmp3 - tmp181 tmp188 = triton_helpers.maximum(tmp187, tmp34) tmp189 = tmp186 + tmp188 tmp190 = tmp5 - tmp181 tmp191 = triton_helpers.maximum(tmp190, tmp34) tmp192 = tmp189 + tmp191 tmp193 = tmp192 - tmp9 tmp194 = tmp193 * tmp57 tmp195 = tmp194 >= tmp34 tmp196 = tl.where(tmp195, tmp181, tmp179) tmp197 = tmp180 * tmp12 tmp198 = tmp196 + tmp197 tmp199 = tmp0 - tmp198 tmp200 = triton_helpers.maximum(tmp199, tmp34) tmp201 = tmp1 - tmp198 tmp202 = triton_helpers.maximum(tmp201, tmp34) tmp203 = tmp200 + tmp202 tmp204 = tmp3 - tmp198 tmp205 = triton_helpers.maximum(tmp204, tmp34) tmp206 = tmp203 + tmp205 tmp207 = tmp5 - tmp198 tmp208 = triton_helpers.maximum(tmp207, tmp34) tmp209 = tmp206 + tmp208 tmp210 = tmp209 - tmp9 tmp211 = tmp210 * tmp57 tmp212 = tmp211 >= tmp34 tmp213 = tl.where(tmp212, tmp198, tmp196) tmp214 = tmp197 * tmp12 tmp215 = tmp213 + tmp214 tmp216 = tmp0 - tmp215 tmp217 = triton_helpers.maximum(tmp216, tmp34) tmp218 = tmp1 - tmp215 tmp219 = triton_helpers.maximum(tmp218, tmp34) tmp220 = tmp217 + tmp219 tmp221 = tmp3 - tmp215 tmp222 = triton_helpers.maximum(tmp221, tmp34) tmp223 = tmp220 + tmp222 tmp224 = tmp5 - tmp215 tmp225 = triton_helpers.maximum(tmp224, tmp34) tmp226 = tmp223 + tmp225 tmp227 = tmp226 - tmp9 tmp228 = tmp227 * tmp57 tmp229 = tmp228 >= tmp34 tmp230 = tl.where(tmp229, tmp215, tmp213) tmp231 = tmp214 * tmp12 tmp232 = tmp230 + tmp231 tmp233 = tmp0 - tmp232 tmp234 = triton_helpers.maximum(tmp233, tmp34) tmp235 = tmp1 - tmp232 tmp236 = triton_helpers.maximum(tmp235, tmp34) tmp237 = tmp234 + tmp236 tmp238 = tmp3 - tmp232 tmp239 = triton_helpers.maximum(tmp238, tmp34) tmp240 = tmp237 + tmp239 tmp241 = tmp5 - tmp232 tmp242 = triton_helpers.maximum(tmp241, tmp34) tmp243 = tmp240 + tmp242 tmp244 = tmp243 - tmp9 tmp245 = tmp244 * tmp57 tmp246 = tmp245 >= tmp34 tmp247 = tl.where(tmp246, tmp232, tmp230) tmp248 = tmp231 * tmp12 tmp249 = tmp247 + tmp248 tmp250 = tmp0 - tmp249 tmp251 = triton_helpers.maximum(tmp250, tmp34) tmp252 = tmp1 - tmp249 tmp253 = triton_helpers.maximum(tmp252, tmp34) tmp254 = tmp251 + tmp253 tmp255 = tmp3 - tmp249 tmp256 = triton_helpers.maximum(tmp255, tmp34) tmp257 = tmp254 + tmp256 tmp258 = tmp5 - tmp249 tmp259 = triton_helpers.maximum(tmp258, tmp34) tmp260 = tmp257 + tmp259 tmp261 = tmp260 - tmp9 tmp262 = tmp261 * tmp57 tmp263 = tmp262 >= tmp34 tmp264 = tl.where(tmp263, tmp249, tmp247) tmp265 = tmp248 * tmp12 tmp266 = tmp264 + tmp265 tmp267 = tmp0 - tmp266 tmp268 = triton_helpers.maximum(tmp267, tmp34) tmp269 = tmp1 - tmp266 tmp270 = triton_helpers.maximum(tmp269, tmp34) tmp271 = tmp268 + tmp270 tmp272 = tmp3 - tmp266 tmp273 = triton_helpers.maximum(tmp272, tmp34) tmp274 = tmp271 + tmp273 tmp275 = tmp5 - tmp266 tmp276 = triton_helpers.maximum(tmp275, tmp34) tmp277 = tmp274 + tmp276 tmp278 = tmp277 - tmp9 tmp279 = tmp278 * tmp57 tmp280 = tmp279 >= tmp34 tmp281 = tl.where(tmp280, tmp266, tmp264) tmp282 = tmp265 * tmp12 tmp283 = tmp281 + tmp282 tmp284 = tmp0 - tmp283 tmp285 = triton_helpers.maximum(tmp284, tmp34) tmp286 = tmp1 - tmp283 tmp287 = triton_helpers.maximum(tmp286, tmp34) tmp288 = tmp285 + tmp287 tmp289 = tmp3 - tmp283 tmp290 = triton_helpers.maximum(tmp289, tmp34) tmp291 = tmp288 + tmp290 tmp292 = tmp5 - tmp283 tmp293 = triton_helpers.maximum(tmp292, tmp34) tmp294 = tmp291 + tmp293 tmp295 = tmp294 - tmp9 tmp296 = tmp295 * tmp57 tmp297 = tmp296 >= tmp34 tmp298 = tl.where(tmp297, tmp283, tmp281) tmp299 = tmp282 * tmp12 tmp300 = tmp298 + tmp299 tmp301 = tmp0 - tmp300 tmp302 = triton_helpers.maximum(tmp301, tmp34) tmp303 = tmp1 - tmp300 tmp304 = triton_helpers.maximum(tmp303, tmp34) tmp305 = tmp302 + tmp304 tmp306 = tmp3 - tmp300 tmp307 = triton_helpers.maximum(tmp306, tmp34) tmp308 = tmp305 + tmp307 tmp309 = tmp5 - tmp300 tmp310 = triton_helpers.maximum(tmp309, tmp34) tmp311 = tmp308 + tmp310 tmp312 = tmp311 - tmp9 tmp313 = tmp312 * tmp57 tmp314 = tmp313 >= tmp34 tmp315 = tl.where(tmp314, tmp300, tmp298) tmp316 = tmp299 * tmp12 tmp317 = tmp315 + tmp316 tmp318 = tmp0 - tmp317 tmp319 = triton_helpers.maximum(tmp318, tmp34) tmp320 = tmp1 - tmp317 tmp321 = triton_helpers.maximum(tmp320, tmp34) tmp322 = tmp319 + tmp321 tmp323 = tmp3 - tmp317 tmp324 = triton_helpers.maximum(tmp323, tmp34) tmp325 = tmp322 + tmp324 tmp326 = tmp5 - tmp317 tmp327 = triton_helpers.maximum(tmp326, tmp34) tmp328 = tmp325 + tmp327 tmp329 = tmp328 - tmp9 tmp330 = tmp329 * tmp57 tmp331 = tmp330 >= tmp34 tmp332 = tl.where(tmp331, tmp317, tmp315) tmp333 = tmp316 * tmp12 tmp334 = tmp332 + tmp333 tmp335 = tmp0 - tmp334 tmp336 = triton_helpers.maximum(tmp335, tmp34) tmp337 = tmp1 - tmp334 tmp338 = triton_helpers.maximum(tmp337, tmp34) tmp339 = tmp336 + tmp338 tmp340 = tmp3 - tmp334 tmp341 = triton_helpers.maximum(tmp340, tmp34) tmp342 = tmp339 + tmp341 tmp343 = tmp5 - tmp334 tmp344 = triton_helpers.maximum(tmp343, tmp34) tmp345 = tmp342 + tmp344 tmp346 = tmp345 - tmp9 tmp347 = tmp346 * tmp57 tmp348 = tmp347 >= tmp34 tmp349 = tl.where(tmp348, tmp334, tmp332) tmp350 = tmp333 * tmp12 tmp351 = tmp349 + tmp350 tmp352 = tmp0 - tmp351 tmp353 = triton_helpers.maximum(tmp352, tmp34) tmp354 = tmp1 - tmp351 tmp355 = triton_helpers.maximum(tmp354, tmp34) tmp356 = tmp353 + tmp355 tmp357 = tmp3 - tmp351 tmp358 = triton_helpers.maximum(tmp357, tmp34) tmp359 = tmp356 + tmp358 tmp360 = tmp5 - tmp351 tmp361 = triton_helpers.maximum(tmp360, tmp34) tmp362 = tmp359 + tmp361 tmp363 = tmp362 - tmp9 tmp364 = tmp363 * tmp57 tmp365 = tmp364 >= tmp34 tmp366 = tl.where(tmp365, tmp351, tmp349) tmp367 = tmp350 * tmp12 tmp368 = tmp366 + tmp367 tmp369 = tmp0 - tmp368 tmp370 = triton_helpers.maximum(tmp369, tmp34) tmp371 = tmp1 - tmp368 tmp372 = triton_helpers.maximum(tmp371, tmp34) tmp373 = tmp370 + tmp372 tmp374 = tmp3 - tmp368 tmp375 = triton_helpers.maximum(tmp374, tmp34) tmp376 = tmp373 + tmp375 tmp377 = tmp5 - tmp368 tmp378 = triton_helpers.maximum(tmp377, tmp34) tmp379 = tmp376 + tmp378 tmp380 = tmp379 - tmp9 tmp381 = tmp380 * tmp57 tmp382 = tmp381 >= tmp34 tmp383 = tl.where(tmp382, tmp368, tmp366) tmp384 = tmp367 * tmp12 tmp385 = tmp383 + tmp384 tmp386 = tmp0 - tmp385 tmp387 = triton_helpers.maximum(tmp386, tmp34) tmp388 = tmp1 - tmp385 tmp389 = triton_helpers.maximum(tmp388, tmp34) tmp390 = tmp387 + tmp389 tmp391 = tmp3 - tmp385 tmp392 = triton_helpers.maximum(tmp391, tmp34) tmp393 = tmp390 + tmp392 tmp394 = tmp5 - tmp385 tmp395 = triton_helpers.maximum(tmp394, tmp34) tmp396 = tmp393 + tmp395 tmp397 = tmp396 - tmp9 tmp398 = tmp397 * tmp57 tmp399 = tmp398 >= tmp34 tmp400 = tl.where(tmp399, tmp385, tmp383) tmp401 = tmp384 * tmp12 tmp402 = tmp400 + tmp401 tmp403 = tmp0 - tmp402 tmp404 = triton_helpers.maximum(tmp403, tmp34) tmp405 = tmp1 - tmp402 tmp406 = triton_helpers.maximum(tmp405, tmp34) tmp407 = tmp404 + tmp406 tmp408 = tmp3 - tmp402 tmp409 = triton_helpers.maximum(tmp408, tmp34) tmp410 = tmp407 + tmp409 tmp411 = tmp5 - tmp402 tmp412 = triton_helpers.maximum(tmp411, tmp34) tmp413 = tmp410 + tmp412 tmp414 = tmp413 - tmp9 tmp415 = tmp414 * tmp57 tmp416 = tmp415 >= tmp34 tmp417 = tl.where(tmp416, tmp402, tmp400) tmp418 = tmp401 * tmp12 tmp419 = tmp417 + tmp418 tmp420 = tmp0 - tmp419 tmp421 = triton_helpers.maximum(tmp420, tmp34) tmp422 = tmp1 - tmp419 tmp423 = triton_helpers.maximum(tmp422, tmp34) tmp424 = tmp421 + tmp423 tmp425 = tmp3 - tmp419 tmp426 = triton_helpers.maximum(tmp425, tmp34) tmp427 = tmp424 + tmp426 tmp428 = tmp5 - tmp419 tmp429 = triton_helpers.maximum(tmp428, tmp34) tmp430 = tmp427 + tmp429 tmp431 = tmp430 - tmp9 tmp432 = tmp431 * tmp57 tmp433 = tmp432 >= tmp34 tmp434 = tl.where(tmp433, tmp419, tmp417) tl.store(out_ptr0 + x0, tmp30, xmask) tl.store(in_out_ptr16 + x0, tmp434, xmask) @triton.jit def triton_poi_fused_add_clamp_div_sub_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 x2 = xindex x1 = xindex // 4 tmp0 = tl.load(in_ptr0 + x2, xmask) tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last') tmp2 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last') tmp3 = 0.5 tmp4 = tmp2 * tmp3 tmp5 = tmp4 * tmp3 tmp6 = tmp5 * tmp3 tmp7 = tmp6 * tmp3 tmp8 = tmp7 * tmp3 tmp9 = tmp8 * tmp3 tmp10 = tmp9 * tmp3 tmp11 = tmp10 * tmp3 tmp12 = tmp11 * tmp3 tmp13 = tmp12 * tmp3 tmp14 = tmp13 * tmp3 tmp15 = tmp14 * tmp3 tmp16 = tmp15 * tmp3 tmp17 = tmp16 * tmp3 tmp18 = tmp17 * tmp3 tmp19 = tmp18 * tmp3 tmp20 = tmp19 * tmp3 tmp21 = tmp20 * tmp3 tmp22 = tmp21 * tmp3 tmp23 = tmp22 * tmp3 tmp24 = tmp23 * tmp3 tmp25 = tmp24 * tmp3 tmp26 = tmp25 * tmp3 tmp27 = tmp1 + tmp26 tmp28 = tmp0 - tmp27 tmp29 = 0.0 tmp30 = triton_helpers.maximum(tmp28, tmp29) tl.store(out_ptr0 + x2, tmp30, xmask) @triton.jit def triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_3(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 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') tmp9 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last') tmp10 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp12 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp14 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp32 = tl.load(in_out_ptr0 + x0, xmask) tmp33 = tl.load(in_ptr2 + x0, xmask) tmp2 = tmp0 + tmp1 tmp4 = tmp2 + tmp3 tmp6 = tmp4 + tmp5 tmp7 = 1.0 tmp8 = tmp6 - tmp7 tmp11 = triton_helpers.maximum(tmp9, tmp10) tmp13 = triton_helpers.maximum(tmp11, tmp12) tmp15 = triton_helpers.maximum(tmp13, tmp14) tmp16 = tmp15 - tmp7 tmp17 = tmp9 - tmp16 tmp18 = 0.0 tmp19 = triton_helpers.maximum(tmp17, tmp18) tmp20 = tmp10 - tmp16 tmp21 = triton_helpers.maximum(tmp20, tmp18) tmp22 = tmp19 + tmp21 tmp23 = tmp12 - tmp16 tmp24 = triton_helpers.maximum(tmp23, tmp18) tmp25 = tmp22 + tmp24 tmp26 = tmp14 - tmp16 tmp27 = triton_helpers.maximum(tmp26, tmp18) tmp28 = tmp25 + tmp27 tmp29 = tmp28 - tmp7 tmp30 = tmp8 * tmp29 tmp31 = tmp30 >= tmp18 tmp34 = 0.5 tmp35 = tmp33 * tmp34 tmp36 = tmp35 * tmp34 tmp37 = tmp36 * tmp34 tmp38 = tmp37 * tmp34 tmp39 = tmp38 * tmp34 tmp40 = tmp39 * tmp34 tmp41 = tmp40 * tmp34 tmp42 = tmp41 * tmp34 tmp43 = tmp42 * tmp34 tmp44 = tmp43 * tmp34 tmp45 = tmp44 * tmp34 tmp46 = tmp45 * tmp34 tmp47 = tmp46 * tmp34 tmp48 = tmp47 * tmp34 tmp49 = tmp48 * tmp34 tmp50 = tmp49 * tmp34 tmp51 = tmp50 * tmp34 tmp52 = tmp51 * tmp34 tmp53 = tmp52 * tmp34 tmp54 = tmp53 * tmp34 tmp55 = tmp54 * tmp34 tmp56 = tmp55 * tmp34 tmp57 = tmp56 * tmp34 tmp58 = tmp32 + tmp57 tmp59 = tl.where(tmp31, tmp58, tmp32) tl.store(in_out_ptr0 + x0, tmp59, xmask) @triton.jit def triton_poi_fused_add_clamp_div_sub_4(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 x1 = xindex // 4 tmp0 = tl.load(in_ptr0 + x2, xmask) tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last') tmp2 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last') tmp3 = 0.5 tmp4 = tmp2 * tmp3 tmp5 = tmp4 * tmp3 tmp6 = tmp5 * tmp3 tmp7 = tmp6 * tmp3 tmp8 = tmp7 * tmp3 tmp9 = tmp8 * tmp3 tmp10 = tmp9 * tmp3 tmp11 = tmp10 * tmp3 tmp12 = tmp11 * tmp3 tmp13 = tmp12 * tmp3 tmp14 = tmp13 * tmp3 tmp15 = tmp14 * tmp3 tmp16 = tmp15 * tmp3 tmp17 = tmp16 * tmp3 tmp18 = tmp17 * tmp3 tmp19 = tmp18 * tmp3 tmp20 = tmp19 * tmp3 tmp21 = tmp20 * tmp3 tmp22 = tmp21 * tmp3 tmp23 = tmp22 * tmp3 tmp24 = tmp23 * tmp3 tmp25 = tmp24 * tmp3 tmp26 = tmp25 * tmp3 tmp27 = tmp26 * tmp3 tmp28 = tmp1 + tmp27 tmp29 = tmp0 - tmp28 tmp30 = 0.0 tmp31 = triton_helpers.maximum(tmp29, tmp30) tl.store(out_ptr0 + x2, tmp31, xmask) @triton.jit def triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_5(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 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') tmp9 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last') tmp10 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp12 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp14 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp32 = tl.load(in_out_ptr0 + x0, xmask) tmp33 = tl.load(in_ptr2 + x0, xmask) tmp2 = tmp0 + tmp1 tmp4 = tmp2 + tmp3 tmp6 = tmp4 + tmp5 tmp7 = 1.0 tmp8 = tmp6 - tmp7 tmp11 = triton_helpers.maximum(tmp9, tmp10) tmp13 = triton_helpers.maximum(tmp11, tmp12) tmp15 = triton_helpers.maximum(tmp13, tmp14) tmp16 = tmp15 - tmp7 tmp17 = tmp9 - tmp16 tmp18 = 0.0 tmp19 = triton_helpers.maximum(tmp17, tmp18) tmp20 = tmp10 - tmp16 tmp21 = triton_helpers.maximum(tmp20, tmp18) tmp22 = tmp19 + tmp21 tmp23 = tmp12 - tmp16 tmp24 = triton_helpers.maximum(tmp23, tmp18) tmp25 = tmp22 + tmp24 tmp26 = tmp14 - tmp16 tmp27 = triton_helpers.maximum(tmp26, tmp18) tmp28 = tmp25 + tmp27 tmp29 = tmp28 - tmp7 tmp30 = tmp8 * tmp29 tmp31 = tmp30 >= tmp18 tmp34 = 0.5 tmp35 = tmp33 * tmp34 tmp36 = tmp35 * tmp34 tmp37 = tmp36 * tmp34 tmp38 = tmp37 * tmp34 tmp39 = tmp38 * tmp34 tmp40 = tmp39 * tmp34 tmp41 = tmp40 * tmp34 tmp42 = tmp41 * tmp34 tmp43 = tmp42 * tmp34 tmp44 = tmp43 * tmp34 tmp45 = tmp44 * tmp34 tmp46 = tmp45 * tmp34 tmp47 = tmp46 * tmp34 tmp48 = tmp47 * tmp34 tmp49 = tmp48 * tmp34 tmp50 = tmp49 * tmp34 tmp51 = tmp50 * tmp34 tmp52 = tmp51 * tmp34 tmp53 = tmp52 * tmp34 tmp54 = tmp53 * tmp34 tmp55 = tmp54 * tmp34 tmp56 = tmp55 * tmp34 tmp57 = tmp56 * tmp34 tmp58 = tmp57 * tmp34 tmp59 = tmp32 + tmp58 tmp60 = tl.where(tmp31, tmp59, tmp32) tl.store(in_out_ptr0 + x0, tmp60, xmask) @triton.jit def triton_poi_fused_add_div_sub_6(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 x1 = xindex // 4 tmp0 = tl.load(in_ptr0 + x2, xmask) tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last') tmp2 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last') tmp3 = 0.5 tmp4 = tmp2 * tmp3 tmp5 = tmp4 * tmp3 tmp6 = tmp5 * tmp3 tmp7 = tmp6 * tmp3 tmp8 = tmp7 * tmp3 tmp9 = tmp8 * tmp3 tmp10 = tmp9 * tmp3 tmp11 = tmp10 * tmp3 tmp12 = tmp11 * tmp3 tmp13 = tmp12 * tmp3 tmp14 = tmp13 * tmp3 tmp15 = tmp14 * tmp3 tmp16 = tmp15 * tmp3 tmp17 = tmp16 * tmp3 tmp18 = tmp17 * tmp3 tmp19 = tmp18 * tmp3 tmp20 = tmp19 * tmp3 tmp21 = tmp20 * tmp3 tmp22 = tmp21 * tmp3 tmp23 = tmp22 * tmp3 tmp24 = tmp23 * tmp3 tmp25 = tmp24 * tmp3 tmp26 = tmp25 * tmp3 tmp27 = tmp26 * tmp3 tmp28 = tmp27 * tmp3 tmp29 = tmp1 + tmp28 tmp30 = tmp0 - tmp29 tl.store(out_ptr0 + x2, tmp30, xmask) @triton.jit def triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_7(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 x0 = xindex tmp0 = tl.load(in_ptr0 + 4 * x0, xmask, eviction_policy='evict_last') tmp3 = tl.load(in_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last') tmp6 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last') tmp9 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last') tmp14 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last') tmp15 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp17 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp19 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp36 = tl.load(in_out_ptr0 + x0, xmask) tmp37 = tl.load(in_ptr2 + x0, xmask) tmp1 = 0.0 tmp2 = triton_helpers.maximum(tmp0, tmp1) tmp4 = triton_helpers.maximum(tmp3, tmp1) tmp5 = tmp2 + tmp4 tmp7 = triton_helpers.maximum(tmp6, tmp1) tmp8 = tmp5 + tmp7 tmp10 = triton_helpers.maximum(tmp9, tmp1) tmp11 = tmp8 + tmp10 tmp12 = 1.0 tmp13 = tmp11 - tmp12 tmp16 = triton_helpers.maximum(tmp14, tmp15) tmp18 = triton_helpers.maximum(tmp16, tmp17) tmp20 = triton_helpers.maximum(tmp18, tmp19) tmp21 = tmp20 - tmp12 tmp22 = tmp14 - tmp21 tmp23 = triton_helpers.maximum(tmp22, tmp1) tmp24 = tmp15 - tmp21 tmp25 = triton_helpers.maximum(tmp24, tmp1) tmp26 = tmp23 + tmp25 tmp27 = tmp17 - tmp21 tmp28 = triton_helpers.maximum(tmp27, tmp1) tmp29 = tmp26 + tmp28 tmp30 = tmp19 - tmp21 tmp31 = triton_helpers.maximum(tmp30, tmp1) tmp32 = tmp29 + tmp31 tmp33 = tmp32 - tmp12 tmp34 = tmp13 * tmp33 tmp35 = tmp34 >= tmp1 tmp38 = 0.5 tmp39 = tmp37 * tmp38 tmp40 = tmp39 * tmp38 tmp41 = tmp40 * tmp38 tmp42 = tmp41 * tmp38 tmp43 = tmp42 * tmp38 tmp44 = tmp43 * tmp38 tmp45 = tmp44 * tmp38 tmp46 = tmp45 * tmp38 tmp47 = tmp46 * tmp38 tmp48 = tmp47 * tmp38 tmp49 = tmp48 * tmp38 tmp50 = tmp49 * tmp38 tmp51 = tmp50 * tmp38 tmp52 = tmp51 * tmp38 tmp53 = tmp52 * tmp38 tmp54 = tmp53 * tmp38 tmp55 = tmp54 * tmp38 tmp56 = tmp55 * tmp38 tmp57 = tmp56 * tmp38 tmp58 = tmp57 * tmp38 tmp59 = tmp58 * tmp38 tmp60 = tmp59 * tmp38 tmp61 = tmp60 * tmp38 tmp62 = tmp61 * tmp38 tmp63 = tmp62 * tmp38 tmp64 = tmp36 + tmp63 tmp65 = tl.where(tmp35, tmp64, tmp36) tl.store(in_out_ptr0 + x0, tmp65, xmask) @triton.jit def triton_poi_fused_add_div_sub_8(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 x1 = xindex // 4 tmp0 = tl.load(in_ptr0 + x2, xmask) tmp1 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last') tmp2 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last') tmp3 = 0.5 tmp4 = tmp2 * tmp3 tmp5 = tmp4 * tmp3 tmp6 = tmp5 * tmp3 tmp7 = tmp6 * tmp3 tmp8 = tmp7 * tmp3 tmp9 = tmp8 * tmp3 tmp10 = tmp9 * tmp3 tmp11 = tmp10 * tmp3 tmp12 = tmp11 * tmp3 tmp13 = tmp12 * tmp3 tmp14 = tmp13 * tmp3 tmp15 = tmp14 * tmp3 tmp16 = tmp15 * tmp3 tmp17 = tmp16 * tmp3 tmp18 = tmp17 * tmp3 tmp19 = tmp18 * tmp3 tmp20 = tmp19 * tmp3 tmp21 = tmp20 * tmp3 tmp22 = tmp21 * tmp3 tmp23 = tmp22 * tmp3 tmp24 = tmp23 * tmp3 tmp25 = tmp24 * tmp3 tmp26 = tmp25 * tmp3 tmp27 = tmp26 * tmp3 tmp28 = tmp27 * tmp3 tmp29 = tmp28 * tmp3 tmp30 = tmp1 + tmp29 tmp31 = tmp0 - tmp30 tl.store(out_ptr0 + x2, tmp31, xmask) @triton.jit def triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_9(in_out_ptr0, in_out_ptr2, in_ptr0, in_ptr1, in_ptr2, out_ptr4, out_ptr7, 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_ptr0 + (1 + 4 * x0), xmask, eviction_policy='evict_last') tmp6 = tl.load(in_ptr0 + (2 + 4 * x0), xmask, eviction_policy='evict_last') tmp9 = tl.load(in_ptr0 + (3 + 4 * x0), xmask, eviction_policy='evict_last') tmp14 = tl.load(in_ptr1 + 4 * x0, xmask, eviction_policy='evict_last') tmp15 = tl.load(in_ptr1 + (1 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp17 = tl.load(in_ptr1 + (2 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp19 = tl.load(in_ptr1 + (3 + 4 * x0), xmask, eviction_policy='evict_last' ) tmp36 = tl.load(in_out_ptr0 + x0, xmask) tmp37 = tl.load(in_ptr2 + x0, xmask) tmp1 = 0.0 tmp2 = triton_helpers.maximum(tmp0, tmp1) tmp4 = triton_helpers.maximum(tmp3, tmp1) tmp5 = tmp2 + tmp4 tmp7 = triton_helpers.maximum(tmp6, tmp1) tmp8 = tmp5 + tmp7 tmp10 = triton_helpers.maximum(tmp9, tmp1) tmp11 = tmp8 + tmp10 tmp12 = 1.0 tmp13 = tmp11 - tmp12 tmp16 = triton_helpers.maximum(tmp14, tmp15) tmp18 = triton_helpers.maximum(tmp16, tmp17) tmp20 = triton_helpers.maximum(tmp18, tmp19) tmp21 = tmp20 - tmp12 tmp22 = tmp14 - tmp21 tmp23 = triton_helpers.maximum(tmp22, tmp1) tmp24 = tmp15 - tmp21 tmp25 = triton_helpers.maximum(tmp24, tmp1) tmp26 = tmp23 + tmp25 tmp27 = tmp17 - tmp21 tmp28 = triton_helpers.maximum(tmp27, tmp1) tmp29 = tmp26 + tmp28 tmp30 = tmp19 - tmp21 tmp31 = triton_helpers.maximum(tmp30, tmp1) tmp32 = tmp29 + tmp31 tmp33 = tmp32 - tmp12 tmp34 = tmp13 * tmp33 tmp35 = tmp34 >= tmp1 tmp38 = 0.5 tmp39 = tmp37 * tmp38 tmp40 = tmp39 * tmp38 tmp41 = tmp40 * tmp38 tmp42 = tmp41 * tmp38 tmp43 = tmp42 * tmp38 tmp44 = tmp43 * tmp38 tmp45 = tmp44 * tmp38 tmp46 = tmp45 * tmp38 tmp47 = tmp46 * tmp38 tmp48 = tmp47 * tmp38 tmp49 = tmp48 * tmp38 tmp50 = tmp49 * tmp38 tmp51 = tmp50 * tmp38 tmp52 = tmp51 * tmp38 tmp53 = tmp52 * tmp38 tmp54 = tmp53 * tmp38 tmp55 = tmp54 * tmp38 tmp56 = tmp55 * tmp38 tmp57 = tmp56 * tmp38 tmp58 = tmp57 * tmp38 tmp59 = tmp58 * tmp38 tmp60 = tmp59 * tmp38 tmp61 = tmp60 * tmp38 tmp62 = tmp61 * tmp38 tmp63 = tmp62 * tmp38 tmp64 = tmp63 * tmp38 tmp65 = tmp36 + tmp64 tmp66 = tl.where(tmp35, tmp65, tmp36) tmp67 = tmp64 * tmp38 tmp68 = tmp66 + tmp67 tmp69 = tmp14 - tmp68 tmp70 = triton_helpers.maximum(tmp69, tmp1) tmp71 = tmp15 - tmp68 tmp72 = triton_helpers.maximum(tmp71, tmp1) tmp73 = tmp70 + tmp72 tmp74 = tmp17 - tmp68 tmp75 = triton_helpers.maximum(tmp74, tmp1) tmp76 = tmp73 + tmp75 tmp77 = tmp19 - tmp68 tmp78 = triton_helpers.maximum(tmp77, tmp1) tmp79 = tmp76 + tmp78 tmp80 = tmp79 - tmp12 tmp81 = tmp80 * tmp33 tmp82 = tmp81 >= tmp1 tmp83 = tl.where(tmp82, tmp68, tmp66) tmp84 = tmp67 * tmp38 tmp85 = tmp83 + tmp84 tmp86 = tmp14 - tmp85 tmp87 = triton_helpers.maximum(tmp86, tmp1) tmp88 = tmp15 - tmp85 tmp89 = triton_helpers.maximum(tmp88, tmp1) tmp90 = tmp87 + tmp89 tmp91 = tmp17 - tmp85 tmp92 = triton_helpers.maximum(tmp91, tmp1) tmp93 = tmp90 + tmp92 tmp94 = tmp19 - tmp85 tmp95 = triton_helpers.maximum(tmp94, tmp1) tmp96 = tmp93 + tmp95 tmp97 = tmp96 - tmp12 tmp98 = tmp97 * tmp33 tmp99 = tmp98 >= tmp1 tmp100 = tl.where(tmp99, tmp85, tmp83) tmp101 = tmp84 * tmp38 tmp102 = tmp100 + tmp101 tmp103 = tmp14 - tmp102 tmp104 = triton_helpers.maximum(tmp103, tmp1) tmp105 = tmp15 - tmp102 tmp106 = triton_helpers.maximum(tmp105, tmp1) tmp107 = tmp104 + tmp106 tmp108 = tmp17 - tmp102 tmp109 = triton_helpers.maximum(tmp108, tmp1) tmp110 = tmp107 + tmp109 tmp111 = tmp19 - tmp102 tmp112 = triton_helpers.maximum(tmp111, tmp1) tmp113 = tmp110 + tmp112 tmp114 = tmp113 - tmp12 tmp115 = tmp114 * tmp33 tmp116 = tmp115 >= tmp1 tmp117 = tl.where(tmp116, tmp102, tmp100) tmp118 = tmp101 * tmp38 tmp119 = tmp117 + tmp118 tmp120 = tmp14 - tmp119 tmp121 = triton_helpers.maximum(tmp120, tmp1) tmp122 = tmp15 - tmp119 tmp123 = triton_helpers.maximum(tmp122, tmp1) tmp124 = tmp121 + tmp123 tmp125 = tmp17 - tmp119 tmp126 = triton_helpers.maximum(tmp125, tmp1) tmp127 = tmp124 + tmp126 tmp128 = tmp19 - tmp119 tmp129 = triton_helpers.maximum(tmp128, tmp1) tmp130 = tmp127 + tmp129 tmp131 = tmp130 - tmp12 tmp132 = tmp131 * tmp33 tmp133 = tmp132 >= tmp1 tmp134 = tl.where(tmp133, tmp119, tmp117) tmp135 = tmp118 * tmp38 tmp136 = tmp134 + tmp135 tmp137 = tmp14 - tmp136 tmp138 = triton_helpers.maximum(tmp137, tmp1) tmp139 = tmp15 - tmp136 tmp140 = triton_helpers.maximum(tmp139, tmp1) tmp141 = tmp138 + tmp140 tmp142 = tmp17 - tmp136 tmp143 = triton_helpers.maximum(tmp142, tmp1) tmp144 = tmp141 + tmp143 tmp145 = tmp19 - tmp136 tmp146 = triton_helpers.maximum(tmp145, tmp1) tmp147 = tmp144 + tmp146 tmp148 = tmp147 - tmp12 tmp149 = tmp148 * tmp33 tmp150 = tmp149 >= tmp1 tmp151 = tl.where(tmp150, tmp136, tmp134) tmp152 = tmp135 * tmp38 tmp153 = tmp151 + tmp152 tmp154 = tmp14 - tmp153 tmp155 = triton_helpers.maximum(tmp154, tmp1) tmp156 = tmp15 - tmp153 tmp157 = triton_helpers.maximum(tmp156, tmp1) tmp158 = tmp155 + tmp157 tmp159 = tmp17 - tmp153 tmp160 = triton_helpers.maximum(tmp159, tmp1) tmp161 = tmp158 + tmp160 tmp162 = tmp19 - tmp153 tmp163 = triton_helpers.maximum(tmp162, tmp1) tmp164 = tmp161 + tmp163 tl.store(out_ptr4 + x0, tmp101, xmask) tl.store(in_out_ptr2 + x0, tmp151, xmask) tl.store(out_ptr7 + x0, tmp164, xmask) @triton.jit def triton_poi_fused_add_clamp_div_sub_10(in_ptr0, in_ptr1, in_ptr2, in_ptr3, 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_ptr1 + x1, xmask, eviction_policy='evict_last') tmp2 = tl.load(in_ptr2 + x1, xmask, eviction_policy='evict_last') tmp11 = tl.load(in_ptr3 + x1, xmask, eviction_policy='evict_last') tmp3 = 0.5 tmp4 = tmp2 * tmp3 tmp5 = tmp4 * tmp3 tmp6 = tmp5 * tmp3 tmp7 = tmp1 + tmp6 tmp8 = tmp0 - tmp7 tmp9 = 0.0 tmp10 = triton_helpers.maximum(tmp8, tmp9) tmp12 = tmp10 / tmp11 tl.store(out_ptr0 + x2, 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) buf40 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) buf41 = reinterpret_tensor(buf40, (4, 4, 4, 1), (16, 4, 1, 64), 0) del buf40 get_raw_stream(0) triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_0[grid(64)](buf41, arg0_1, 64, XBLOCK=64, num_warps=1, num_stages=1) buf42 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32) buf81 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) buf82 = reinterpret_tensor(buf81, (4, 4, 4, 1), (16, 4, 1, 64), 0) del buf81 triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_1[grid(64)](buf82, arg0_1, buf41, buf42, 64, XBLOCK=64, num_warps=1, num_stages=1) buf83 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused_add_clamp_div_sub_2[grid(256)](arg0_1, buf82, buf42, buf83, 256, XBLOCK=256, num_warps=4, num_stages=1) buf85 = buf82 del buf82 triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_3[grid(64)](buf85, buf83, arg0_1, buf42, 64, XBLOCK=64, num_warps=1, num_stages=1) buf86 = buf83 del buf83 triton_poi_fused_add_clamp_div_sub_4[grid(256)](arg0_1, buf85, buf42, buf86, 256, XBLOCK=256, num_warps=4, num_stages=1) buf88 = buf85 del buf85 triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_5[grid(64)](buf88, buf86, arg0_1, buf42, 64, XBLOCK=64, num_warps=1, num_stages=1) buf89 = buf86 del buf86 triton_poi_fused_add_div_sub_6[grid(256)](arg0_1, buf88, buf42, buf89, 256, XBLOCK=256, num_warps=4, num_stages=1) buf91 = buf88 del buf88 triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_7[grid(64)](buf91, buf89, arg0_1, buf42, 64, XBLOCK=64, num_warps=1, num_stages=1) buf92 = buf89 del buf89 triton_poi_fused_add_div_sub_8[grid(256)](arg0_1, buf91, buf42, buf92, 256, XBLOCK=128, num_warps=4, num_stages=1) buf94 = buf91 del buf91 buf100 = buf41 del buf41 buf103 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) buf104 = reinterpret_tensor(buf103, (4, 4, 4, 1), (16, 4, 1, 64), 0) del buf103 buf105 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused_add_clamp_div_max_mul_sub_sum_where_9[grid(64)](buf94, buf104, buf92, arg0_1, buf42, buf100, buf105, 64, XBLOCK=64, num_warps=1, num_stages=1) del buf42 del buf94 buf106 = buf92 del buf92 triton_poi_fused_add_clamp_div_sub_10[grid(256)](arg0_1, buf104, buf100, buf105, buf106, 256, XBLOCK=128, num_warps=4, num_stages=1) del arg0_1 del buf100 del buf104 del buf105 return buf106, def sparsemax_bisect(X, dim=-1, n_iter=50, ensure_sum_one=True): """sparsemax: normalizing sparse transform (a la softmax), via bisection. Solves the projection: min_p ||x - p||_2 s.t. p >= 0, sum(p) == 1. Parameters ---------- X : torch.Tensor The input tensor. dim : int The dimension along which to apply sparsemax. n_iter : int Number of bisection iterations. For float32, 24 iterations should suffice for machine precision. ensure_sum_one : bool, Whether to divide the result by its sum. If false, the result might sum to close but not exactly 1, which might cause downstream problems. Note: This function does not yet support normalizing along anything except the last dimension. Please use transposing and views to achieve more general behavior. Returns ------- P : torch tensor, same shape as X The projection result, such that P.sum(dim=dim) == 1 elementwise. """ return SparsemaxBisectFunction.apply(X, dim, n_iter, ensure_sum_one) class EntmaxBisectFunction(Function): @classmethod def _gp(cls, x, alpha): return x ** (alpha - 1) @classmethod def _gp_inv(cls, y, alpha): return y ** (1 / (alpha - 1)) @classmethod def _p(cls, X, alpha): return cls._gp_inv(torch.clamp(X, min=0), alpha) @classmethod def forward(cls, ctx, X, alpha=1.5, dim=-1, n_iter=50, ensure_sum_one=True ): if not isinstance(alpha, torch.Tensor): alpha = torch.tensor(alpha, dtype=X.dtype, device=X.device) alpha_shape = list(X.shape) alpha_shape[dim] = 1 alpha = alpha.expand(*alpha_shape) ctx.alpha = alpha ctx.dim = dim d = X.shape[dim] max_val, _ = X.max(dim=dim, keepdim=True) X = X * (alpha - 1) max_val = max_val * (alpha - 1) tau_lo = max_val - cls._gp(1, alpha) tau_hi = max_val - cls._gp(1 / d, alpha) f_lo = cls._p(X - tau_lo, alpha).sum(dim) - 1 dm = tau_hi - tau_lo for it in range(n_iter): dm /= 2 tau_m = tau_lo + dm p_m = cls._p(X - tau_m, alpha) f_m = p_m.sum(dim) - 1 mask = (f_m * f_lo >= 0).unsqueeze(dim) tau_lo = torch.where(mask, tau_m, tau_lo) if ensure_sum_one: p_m /= p_m.sum(dim=dim).unsqueeze(dim=dim) ctx.save_for_backward(p_m) return p_m @classmethod def backward(cls, ctx, dY): Y, = ctx.saved_tensors gppr = torch.where(Y > 0, Y ** (2 - ctx.alpha), Y.new_zeros(1)) dX = dY * gppr q = dX.sum(ctx.dim) / gppr.sum(ctx.dim) q = q.unsqueeze(ctx.dim) dX -= q * gppr d_alpha = None if ctx.needs_input_grad[1]: S = torch.where(Y > 0, Y * torch.log(Y), Y.new_zeros(1)) ent = S.sum(ctx.dim).unsqueeze(ctx.dim) Y_skewed = gppr / gppr.sum(ctx.dim).unsqueeze(ctx.dim) d_alpha = dY * (Y - Y_skewed) / (ctx.alpha - 1) ** 2 d_alpha -= dY * (S - Y_skewed * ent) / (ctx.alpha - 1) d_alpha = d_alpha.sum(ctx.dim).unsqueeze(ctx.dim) return dX, d_alpha, None, None, None class SparsemaxBisectFunction(EntmaxBisectFunction): @classmethod def _gp(cls, x, alpha): return x @classmethod def _gp_inv(cls, y, alpha): return y @classmethod def _p(cls, x, alpha): return torch.clamp(x, min=0) @classmethod def forward(cls, ctx, X, dim=-1, n_iter=50, ensure_sum_one=True): return super().forward(ctx, X, alpha=2, dim=dim, n_iter=50, ensure_sum_one=True) @classmethod def backward(cls, ctx, dY): Y, = ctx.saved_tensors gppr = Y > 0 dX = dY * gppr q = dX.sum(ctx.dim) / gppr.sum(ctx.dim) q = q.unsqueeze(ctx.dim) dX -= q * gppr return dX, None, None, None class SparsemaxBisectNew(nn.Module): def __init__(self, dim=-1, n_iter=None): """sparsemax: normalizing sparse transform (a la softmax) via bisection Solves the projection: min_p ||x - p||_2 s.t. p >= 0, sum(p) == 1. Parameters ---------- dim : int The dimension along which to apply sparsemax. n_iter : int Number of bisection iterations. For float32, 24 iterations should suffice for machine precision. """ self.dim = dim self.n_iter = n_iter super().__init__() def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0]
cifkao/entmax
SparsemaxBisect
false
15,042
[ "MIT" ]
298
f18bab9318f9d2471a36545ee0b4c97be6d48a87
https://github.com/cifkao/entmax/tree/f18bab9318f9d2471a36545ee0b4c97be6d48a87
AttentionModule
from _paritybench_helpers import _mock_config import torch class AttentionModule(torch.nn.Module): """ SimGNN Attention Module to make a pass on graph. """ def __init__(self, args): """ :param args: Arguments object. """ super(AttentionModule, self).__init__() self.args = args self.setup_weights() self.init_parameters() def setup_weights(self): """ Defining weights. """ self.weight_matrix = torch.nn.Parameter(torch.Tensor(self.args. filters_3, self.args.filters_3)) def init_parameters(self): """ Initializing weights. """ torch.nn.init.xavier_uniform_(self.weight_matrix) def forward(self, embedding): """ Making a forward propagation pass to create a graph level representation. :param embedding: Result of the GCN. :return representation: A graph level representation vector. """ batch_size = embedding.shape[0] global_context = torch.mean(torch.matmul(embedding, self. weight_matrix), dim=1) transformed_global = torch.tanh(global_context) sigmoid_scores = torch.sigmoid(torch.matmul(embedding, transformed_global.view(batch_size, -1, 1))) representation = torch.matmul(embedding.permute(0, 2, 1), sigmoid_scores) return representation, sigmoid_scores def get_inputs(): return [torch.rand([4, 4, 4])] def get_init_inputs(): return [[], {'args': _mock_config(filters_3=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 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_tanh_tanh_backward_0(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 % 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 tmp9 = libdevice.tanh(tmp8) tmp10 = tmp9 * tmp9 tmp11 = 1.0 tmp12 = tmp11 - tmp10 tl.store(out_ptr0 + x2, tmp9, xmask) tl.store(out_ptr1 + x2, tmp12, xmask) @triton.jit def triton_poi_fused_sigmoid_1(in_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.sigmoid(tmp0) tl.store(in_out_ptr0 + x0, tmp1, xmask) def call(args): primals_1, primals_2 = args args.clear() assert_size_stride(primals_1, (4, 4, 4), (16, 4, 1)) assert_size_stride(primals_2, (4, 4), (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), primals_2, out=buf0) del primals_2 buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32) buf5 = empty_strided_cuda((4, 4), (4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_mean_tanh_tanh_backward_0[grid(16)](buf0, buf1, buf5, 16, XBLOCK=16, num_warps=1, num_stages=1) del buf0 buf2 = empty_strided_cuda((4, 4, 1), (4, 1, 1), torch.float32) extern_kernels.bmm(primals_1, reinterpret_tensor(buf1, (4, 4, 1), ( 4, 1, 0), 0), out=buf2) buf3 = buf2 del buf2 triton_poi_fused_sigmoid_1[grid(16)](buf3, 16, XBLOCK=16, num_warps =1, num_stages=1) buf4 = reinterpret_tensor(buf1, (4, 4, 1), (4, 1, 1), 0) del buf1 extern_kernels.bmm(reinterpret_tensor(primals_1, (4, 4, 4), (16, 1, 4), 0), buf3, out=buf4) return buf4, buf3, primals_1, buf3, buf5 class AttentionModuleNew(torch.nn.Module): """ SimGNN Attention Module to make a pass on graph. """ def __init__(self, args): """ :param args: Arguments object. """ super(AttentionModuleNew, self).__init__() self.args = args self.setup_weights() self.init_parameters() def setup_weights(self): """ Defining weights. """ self.weight_matrix = torch.nn.Parameter(torch.Tensor(self.args. filters_3, self.args.filters_3)) def init_parameters(self): """ Initializing weights. """ torch.nn.init.xavier_uniform_(self.weight_matrix) def forward(self, input_0): primals_2 = self.weight_matrix primals_1 = input_0 output = call([primals_1, primals_2]) return output[0], output[1]
cloudcjf/SG_PR
AttentionModule
false
15,043
[ "MIT" ]
105
1339d00811ea3c4c18963efa24bf6fc778e15794
https://github.com/cloudcjf/SG_PR/tree/1339d00811ea3c4c18963efa24bf6fc778e15794
GraphConvolution
from torch.nn import Module import torch from torch.nn import Parameter from torch.nn import functional as F import torch.multiprocessing import torch.utils.data from torch.nn.modules.module import Module from torch.nn.parameter import Parameter import torch.nn.modules.loss class GraphConvolution(Module): """ Simple GCN layer, similar to https://arxiv.org/abs/1609.02907 """ def __init__(self, in_features, out_features, dropout=0.0, act=F.relu): super(GraphConvolution, self).__init__() self.in_features = in_features self.out_features = out_features self.dropout = dropout self.act = act self.weight = Parameter(torch.FloatTensor(in_features, out_features)) self.reset_parameters() def reset_parameters(self): torch.nn.init.xavier_uniform_(self.weight) def forward(self, input, adj): input = F.dropout(input, self.dropout, self.training) support = torch.mm(input, self.weight) output = torch.spmm(adj, support) output = self.act(output) return output def __repr__(self): return self.__class__.__name__ + ' (' + str(self.in_features ) + ' -> ' + str(self.out_features) + ')' def get_inputs(): return [torch.rand([4, 4]), torch.rand([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 from torch.nn import Module from torch.nn import Parameter from torch.nn import functional as F import torch.multiprocessing import torch.utils.data from torch.nn.modules.module import Module from torch.nn.parameter import Parameter import torch.nn.modules.loss 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, 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.full([1], 0, tl.int32) tmp2 = triton_helpers.maximum(tmp1, tmp0) tmp3 = 0.0 tmp4 = tmp2 <= tmp3 tl.store(in_out_ptr0 + x0, tmp2, xmask) tl.store(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, 1)) assert_size_stride(primals_2, (4, 4), (4, 1)) assert_size_stride(primals_3, (4, 4), (4, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.mm(primals_1, primals_2, out=buf0) del primals_2 buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.mm(primals_3, buf0, out=buf1) del buf0 buf2 = buf1 del buf1 buf3 = empty_strided_cuda((4, 4), (4, 1), torch.bool) get_raw_stream(0) triton_poi_fused_relu_threshold_backward_0[grid(16)](buf2, buf3, 16, XBLOCK=16, num_warps=1, num_stages=1) return buf2, buf3, reinterpret_tensor(primals_3, (4, 4), (1, 4), 0 ), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0) class GraphConvolutionNew(Module): """ Simple GCN layer, similar to https://arxiv.org/abs/1609.02907 """ def __init__(self, in_features, out_features, dropout=0.0, act=F.relu): super(GraphConvolutionNew, self).__init__() self.in_features = in_features self.out_features = out_features self.dropout = dropout self.act = act self.weight = Parameter(torch.FloatTensor(in_features, out_features)) self.reset_parameters() def reset_parameters(self): torch.nn.init.xavier_uniform_(self.weight) def __repr__(self): return self.__class__.__name__ + ' (' + str(self.in_features ) + ' -> ' + str(self.out_features) + ')' def forward(self, input_0, input_1): primals_1 = self.weight primals_2 = input_0 primals_3 = input_1 output = call([primals_1, primals_2, primals_3]) return output[0]
cminusQAQ/graph4nlp
GraphConvolution
false
15,044
[ "Apache-2.0" ]
1,269
d980e897131f1b9d3766750c06316d94749904fa
https://github.com/cminusQAQ/graph4nlp/tree/d980e897131f1b9d3766750c06316d94749904fa
DilatedResidualLayer
import torch import torch.nn as nn import torch.nn.functional as F class DilatedResidualLayer(nn.Module): def __init__(self, dilation, in_channels, out_channels): super(DilatedResidualLayer, self).__init__() self.conv_dilated = nn.Conv1d(in_channels, out_channels, 3, padding =dilation, dilation=dilation) self.conv_1x1 = nn.Conv1d(out_channels, out_channels, 1) self.dropout = nn.Dropout() def forward(self, x): out = F.relu(self.conv_dilated(x)) out = self.conv_1x1(out) out = self.dropout(out) return x + out def get_inputs(): return [torch.rand([4, 4])] def get_init_inputs(): return [[], {'dilation': 1, '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 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 = 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_out_ptr0 + x2, 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 + x2, tmp4, xmask) tl.store(out_ptr0 + x2, tmp6, xmask) @triton.jit def triton_poi_fused_add_1(in_out_ptr0, in_ptr0, in_ptr1, 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_out_ptr0 + x2, xmask) tmp2 = tl.load(in_ptr1 + x1, xmask, eviction_policy='evict_last') tmp3 = tmp1 + tmp2 tmp4 = tmp0 + tmp3 tl.store(in_out_ptr0 + x2, tmp4, 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), (12, 3, 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=(1,), dilation=(1,), transposed=False, output_padding=(0,), groups=1, bias=None) assert_size_stride(buf0, (1, 4, 4), (16, 4, 1)) buf1 = reinterpret_tensor(buf0, (4, 4), (4, 1), 0) del buf0 buf4 = empty_strided_cuda((4, 4), (4, 1), torch.bool) get_raw_stream(0) triton_poi_fused_relu_threshold_backward_0[grid(16)](buf1, primals_2, buf4, 16, XBLOCK=16, num_warps=1, num_stages=1) del primals_2 buf2 = extern_kernels.convolution(reinterpret_tensor(buf1, (1, 4, 4 ), (0, 4, 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, 4), (16, 4, 1)) buf3 = reinterpret_tensor(buf2, (4, 4), (4, 1), 0) del buf2 triton_poi_fused_add_1[grid(16)](buf3, primals_3, primals_5, 16, XBLOCK=16, num_warps=1, num_stages=1) del primals_5 return buf3, primals_1, primals_4, reinterpret_tensor(primals_3, (1, 4, 4), (16, 4, 1), 0), reinterpret_tensor(buf1, (1, 4, 4), (16, 4, 1), 0 ), buf4 class DilatedResidualLayerNew(nn.Module): def __init__(self, dilation, in_channels, out_channels): super(DilatedResidualLayerNew, self).__init__() self.conv_dilated = nn.Conv1d(in_channels, out_channels, 3, padding =dilation, dilation=dilation) self.conv_1x1 = nn.Conv1d(out_channels, out_channels, 1) self.dropout = nn.Dropout() def forward(self, input_0): primals_1 = self.conv_dilated.weight primals_2 = self.conv_dilated.bias primals_4 = self.conv_1x1.weight primals_5 = self.conv_1x1.bias primals_3 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5]) return output[0]
cmhungsteve/SSTDA
DilatedResidualLayer
false
15,045
[ "MIT" ]
154
9c5e1df952bd122ea474046d91e3ac6fa79ec312
https://github.com/cmhungsteve/SSTDA/tree/9c5e1df952bd122ea474046d91e3ac6fa79ec312
MeanEmbedding
import torch import torch.nn as nn import torch.multiprocessing import torch.utils.data import torch.nn.modules.loss class MeanEmbedding(nn.Module): """Mean embedding class.""" def __init__(self): super(MeanEmbedding, self).__init__() def forward(self, emb, len_): """Compute average embeddings. Parameters ---------- emb : torch.Tensor The input embedding tensor. len_ : torch.Tensor The sequence length tensor. Returns ------- torch.Tensor The average embedding tensor. """ summed = torch.sum(emb, dim=-2) len_ = len_.unsqueeze(-1).expand_as(summed).float() return summed / len_ def get_inputs(): return [torch.rand([4, 4, 4, 4]), torch.rand([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 import torch.nn as nn import torch.multiprocessing import torch.utils.data import torch.nn.modules.loss 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_div_sum_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 + 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) tmp7 = tl.load(in_ptr1 + x1, 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): 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, 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_div_sum_0[grid(64)](arg0_1, arg1_1, buf0, 64, XBLOCK=64, num_warps=1, num_stages=1) del arg0_1 del arg1_1 return buf0, class MeanEmbeddingNew(nn.Module): """Mean embedding class.""" def __init__(self): super(MeanEmbeddingNew, 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]
cminusQAQ/graph4nlp
MeanEmbedding
false
15,046
[ "Apache-2.0" ]
1,269
d980e897131f1b9d3766750c06316d94749904fa
https://github.com/cminusQAQ/graph4nlp/tree/d980e897131f1b9d3766750c06316d94749904fa
SoftDiceLoss
import torch import torch.nn.functional as F import torch.nn as nn class SoftDiceLoss(nn.Module): def __init__(self, weight=None, size_average=True): super(SoftDiceLoss, self).__init__() def forward(self, logits, targets): num = targets.size(0) probs = F.sigmoid(logits) m1 = probs.view(num, -1) m2 = targets.view(num, -1) intersection = m1 * m2 score = 2.0 * (intersection.sum(1) + 1) / (m1.sum(1) + m2.sum(1) + 1) score = 1 - score.sum() / num return score 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 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_mul_sum_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, out_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr): xnumel = 4 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) r1 = rindex x0 = xindex tmp0 = tl.load(in_ptr0 + (r1 + 64 * x0), xmask, other=0.0) tmp2 = tl.load(in_ptr1 + (r1 + 64 * x0), xmask, other=0.0) tmp1 = tl.sigmoid(tmp0) tmp3 = tmp1 * tmp2 tmp4 = tl.broadcast_to(tmp3, [XBLOCK, RBLOCK]) tmp6 = tl.where(xmask, tmp4, 0) tmp7 = tl.sum(tmp6, 1)[:, None] tmp8 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK]) tmp10 = tl.where(xmask, tmp8, 0) tmp11 = tl.sum(tmp10, 1)[:, None] tmp12 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK]) tmp14 = tl.where(xmask, tmp12, 0) tmp15 = tl.sum(tmp14, 1)[:, None] tl.store(out_ptr0 + x0, tmp7, xmask) tl.store(out_ptr1 + x0, tmp11, xmask) tl.store(out_ptr2 + x0, tmp15, xmask) @triton.jit def triton_per_fused_add_div_mul_rsub_sum_1(in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, xnumel, rnumel, XBLOCK: tl.constexpr): RBLOCK: tl.constexpr = 4 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) tmp5 = tl.load(in_ptr1 + r0, None) tmp6 = tl.load(in_ptr2 + r0, None) tmp1 = 1.0 tmp2 = tmp0 + tmp1 tmp3 = 2.0 tmp4 = tmp2 * tmp3 tmp7 = tmp5 + tmp6 tmp8 = tmp7 + tmp1 tmp9 = tmp4 / tmp8 tmp10 = tl.broadcast_to(tmp9, [XBLOCK, RBLOCK]) tmp12 = tl.sum(tmp10, 1)[:, None] tmp13 = 0.25 tmp14 = tmp12 * tmp13 tmp15 = tmp1 - tmp14 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([XBLOCK, 1], 0, tl.int32), tmp15, 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,), (1,), torch.float32) buf1 = empty_strided_cuda((4,), (1,), torch.float32) buf2 = empty_strided_cuda((4,), (1,), torch.float32) get_raw_stream(0) triton_per_fused_mul_sum_0[grid(4)](arg1_1, arg0_1, buf0, buf1, buf2, 4, 64, XBLOCK=1, num_warps=2, num_stages=1) del arg0_1 del arg1_1 buf3 = empty_strided_cuda((), (), torch.float32) buf4 = buf3 del buf3 triton_per_fused_add_div_mul_rsub_sum_1[grid(1)](buf4, buf0, buf1, buf2, 1, 4, XBLOCK=1, num_warps=2, num_stages=1) del buf0 del buf1 del buf2 return buf4, class SoftDiceLossNew(nn.Module): def __init__(self, weight=None, size_average=True): super(SoftDiceLossNew, 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]
cmarasinou/carvana-challenge
SoftDiceLoss
false
15,047
[ "MIT" ]
93
4e1c43f306cfbef1df267acfce59bdcf19504850
https://github.com/cmarasinou/carvana-challenge/tree/4e1c43f306cfbef1df267acfce59bdcf19504850
InnerProductDecoder
import torch import torch.nn as nn from torch.nn import functional as F import torch.multiprocessing import torch.utils.data import torch.nn.modules.loss class InnerProductDecoder(nn.Module): """Decoder for using inner product for prediction.""" def __init__(self, dropout, act=torch.sigmoid): super(InnerProductDecoder, self).__init__() self.dropout = dropout self.act = act def forward(self, z): z = F.dropout(z, self.dropout, training=self.training) adj = self.act(torch.mm(z, z.t())) return adj def get_inputs(): return [torch.rand([4, 4])] def get_init_inputs(): return [[], {'dropout': 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 import torch.nn as nn import torch.multiprocessing import torch.utils.data import torch.nn.modules.loss 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, 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.sigmoid(tmp0) tl.store(in_out_ptr0 + x0, tmp1, xmask) def call(args): arg0_1, = args args.clear() assert_size_stride(arg0_1, (4, 4), (4, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.mm(arg0_1, reinterpret_tensor(arg0_1, (4, 4), (1, 4), 0), out=buf0) del arg0_1 buf1 = buf0 del buf0 get_raw_stream(0) triton_poi_fused_sigmoid_0[grid(16)](buf1, 16, XBLOCK=16, num_warps =1, num_stages=1) return buf1, class InnerProductDecoderNew(nn.Module): """Decoder for using inner product for prediction.""" def __init__(self, dropout, act=torch.sigmoid): super(InnerProductDecoderNew, self).__init__() self.dropout = dropout self.act = act def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0]
cminusQAQ/graph4nlp
InnerProductDecoder
false
15,048
[ "Apache-2.0" ]
1,269
d980e897131f1b9d3766750c06316d94749904fa
https://github.com/cminusQAQ/graph4nlp/tree/d980e897131f1b9d3766750c06316d94749904fa
BCELoss2d
import torch import torch.nn.functional as F import torch.nn as nn class BCELoss2d(nn.Module): def __init__(self, weight=None, size_average=True): super(BCELoss2d, self).__init__() self.bce_loss = nn.BCELoss(weight, size_average) def forward(self, logits, targets): probs = F.sigmoid(logits) probs_flat = probs.view(-1) targets_flat = targets.view(-1) return self.bce_loss(probs_flat, targets_flat) 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 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_binary_cross_entropy_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) tmp3 = tl.load(in_ptr1 + r0, None) tmp1 = 1.0 tmp2 = tmp0 - tmp1 tmp4 = tl.sigmoid(tmp3) tmp5 = -tmp4 tmp6 = libdevice.log1p(tmp5) tmp7 = -100.0 tmp8 = triton_helpers.maximum(tmp6, tmp7) tmp9 = tmp2 * tmp8 tmp10 = tl_math.log(tmp4) tmp11 = triton_helpers.maximum(tmp10, tmp7) tmp12 = tmp0 * tmp11 tmp13 = tmp9 - tmp12 tmp14 = tl.broadcast_to(tmp13, [RBLOCK]) tmp16 = triton_helpers.promote_to_tensor(tl.sum(tmp14, 0)) tmp17 = 256.0 tmp18 = tmp16 / tmp17 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp18, 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_binary_cross_entropy_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 BCELoss2dNew(nn.Module): def __init__(self, weight=None, size_average=True): super(BCELoss2dNew, self).__init__() self.bce_loss = nn.BCELoss(weight, size_average) def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0]
cmarasinou/carvana-challenge
BCELoss2d
false
15,049
[ "MIT" ]
93
4e1c43f306cfbef1df267acfce59bdcf19504850
https://github.com/cmarasinou/carvana-challenge/tree/4e1c43f306cfbef1df267acfce59bdcf19504850
mlpblock
import torch import torch.nn as nn import torch.nn.functional as F class linearblock(nn.Module): def __init__(self, in_features, out_features, bias=True, dropout='none'): super(linearblock, self).__init__() self.conv = nn.Linear(in_features, out_features, bias=bias) self.relu = nn.ReLU(inplace=True) self.dropout = nn.Dropout(inplace=True) self.dropoutoption = dropout def forward(self, x): x = self.conv(x) x = self.relu(x) if self.dropoutoption == 'normal': x = self.dropout(x) return x class mlpblock(nn.Module): def __init__(self, inputdim, outputdim, nlayer=2, hiddendim=10, activation='ReLU'): super(mlpblock, self).__init__() self.nlayer = nlayer self.hiddendim = hiddendim self.inputdim = inputdim self.outputdim = outputdim if activation == 'ReLU': self.act = F.relu else: raise NotImplementedError self.fc1 = nn.Linear(self.inputdim, self.hiddendim) fc_iter = [] for n in range(self.nlayer - 2): fc_iter.append(linearblock(self.hiddendim, self.hiddendim)) self.fc_iter = nn.Sequential(*fc_iter) self.fc_final = nn.Linear(self.hiddendim, self.outputdim) def forward(self, x): x = self.act(self.fc1(x)) x = self.fc_iter(x) x = self.fc_final(x) return x def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'inputdim': 4, 'outputdim': 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 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_relu_threshold_backward_0(in_out_ptr0, in_ptr0, out_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 640 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x2 = xindex x0 = xindex % 10 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) def call(args): primals_1, primals_2, primals_3, primals_4, primals_5 = args args.clear() assert_size_stride(primals_1, (10, 4), (4, 1)) assert_size_stride(primals_2, (10,), (1,)) assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_4, (4, 10), (10, 1)) assert_size_stride(primals_5, (4,), (1,)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((64, 10), (10, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 10), (1, 4), 0), out=buf0) del primals_1 buf1 = reinterpret_tensor(buf0, (4, 4, 4, 10), (160, 40, 10, 1), 0) del buf0 buf3 = empty_strided_cuda((4, 4, 4, 10), (160, 40, 10, 1), torch.bool) get_raw_stream(0) triton_poi_fused_relu_threshold_backward_0[grid(640)](buf1, primals_2, buf3, 640, XBLOCK=256, num_warps=4, 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, 10), (10, 1), 0), reinterpret_tensor(primals_4, (10, 4), (1, 10), 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 ), reinterpret_tensor(buf1, (64, 10), (10, 1), 0), primals_4, buf3 class linearblock(nn.Module): def __init__(self, in_features, out_features, bias=True, dropout='none'): super(linearblock, self).__init__() self.conv = nn.Linear(in_features, out_features, bias=bias) self.relu = nn.ReLU(inplace=True) self.dropout = nn.Dropout(inplace=True) self.dropoutoption = dropout def forward(self, x): x = self.conv(x) x = self.relu(x) if self.dropoutoption == 'normal': x = self.dropout(x) return x class mlpblockNew(nn.Module): def __init__(self, inputdim, outputdim, nlayer=2, hiddendim=10, activation='ReLU'): super(mlpblockNew, self).__init__() self.nlayer = nlayer self.hiddendim = hiddendim self.inputdim = inputdim self.outputdim = outputdim if activation == 'ReLU': self.act = F.relu else: raise NotImplementedError self.fc1 = nn.Linear(self.inputdim, self.hiddendim) fc_iter = [] for n in range(self.nlayer - 2): fc_iter.append(linearblock(self.hiddendim, self.hiddendim)) self.fc_iter = nn.Sequential(*fc_iter) self.fc_final = nn.Linear(self.hiddendim, self.outputdim) def forward(self, input_0): primals_1 = self.fc1.weight primals_2 = self.fc1.bias primals_4 = self.fc_final.weight primals_5 = self.fc_final.bias primals_3 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5]) return output[0]
coallaoh/WhitenBlackBox
mlpblock
false
15,050
[ "MIT" ]
46
816363c59a11248e79ffed70f1a14510b0967dab
https://github.com/coallaoh/WhitenBlackBox/tree/816363c59a11248e79ffed70f1a14510b0967dab
ShiftedSoftplus
import torch import torch.nn.functional as F from torch import nn class ShiftedSoftplus(nn.Module): __constants__ = ['beta', 'threshold'] beta: 'int' threshold: 'int' def __init__(self, beta: 'int'=1, threshold: 'int'=20) ->None: super(ShiftedSoftplus, self).__init__() self.beta = beta self.threshold = threshold def forward(self, x: 'torch.Tensor') ->torch.Tensor: return F.softplus(x - 1, self.beta, self.threshold) def extra_repr(self) ->str: return 'beta={}, threshold={}'.format(self.beta, self.threshold) 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 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_softplus_sub_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 = 20.0 tmp4 = tmp2 > tmp3 tmp5 = tl_math.exp(tmp2) tmp6 = libdevice.log1p(tmp5) tmp7 = tl.where(tmp4, tmp2, tmp6) tl.store(out_ptr0 + x0, tmp7, 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_softplus_sub_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256, num_warps=4, num_stages=1) del arg0_1 return buf0, class ShiftedSoftplusNew(nn.Module): __constants__ = ['beta', 'threshold'] beta: 'int' threshold: 'int' def __init__(self, beta: 'int'=1, threshold: 'int'=20) ->None: super(ShiftedSoftplusNew, self).__init__() self.beta = beta self.threshold = threshold def extra_repr(self) ->str: return 'beta={}, threshold={}'.format(self.beta, self.threshold) def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0]
cmusatyalab/mega-nerf
ShiftedSoftplus
false
15,051
[ "MIT" ]
107
306e06cc316dd4f5c84d0610308bcbc208228fc3
https://github.com/cmusatyalab/mega-nerf/tree/306e06cc316dd4f5c84d0610308bcbc208228fc3
Accuracy
import torch from torch import nn class Accuracy(nn.Module): label = 'Accuracy' def forward(self, prediction, truth): prediction = prediction.argmax(dim=1) correct = prediction == truth accuracy = correct.float().mean() return accuracy 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 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_argmax_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) tmp17 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask) tmp32 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask) tmp2 = tmp0 > tmp1 tmp3 = tmp0 == tmp1 tmp4 = tmp0 != tmp0 tmp5 = tmp1 != tmp1 tmp6 = tmp4 > tmp5 tmp7 = tmp2 | tmp6 tmp8 = tmp4 & tmp5 tmp9 = tmp3 | tmp8 tmp10 = tl.full([1], 0, tl.int64) tmp11 = tl.full([1], 1, tl.int64) tmp12 = tmp10 < tmp11 tmp13 = tmp9 & tmp12 tmp14 = tmp7 | tmp13 tmp15 = tl.where(tmp14, tmp0, tmp1) tmp16 = tl.where(tmp14, tmp10, tmp11) tmp18 = tmp15 > tmp17 tmp19 = tmp15 == tmp17 tmp20 = tmp15 != tmp15 tmp21 = tmp17 != tmp17 tmp22 = tmp20 > tmp21 tmp23 = tmp18 | tmp22 tmp24 = tmp20 & tmp21 tmp25 = tmp19 | tmp24 tmp26 = tl.full([1], 2, tl.int64) tmp27 = tmp16 < tmp26 tmp28 = tmp25 & tmp27 tmp29 = tmp23 | tmp28 tmp30 = tl.where(tmp29, tmp15, tmp17) tmp31 = tl.where(tmp29, tmp16, tmp26) tmp33 = tmp30 > tmp32 tmp34 = tmp30 == tmp32 tmp35 = tmp30 != tmp30 tmp36 = tmp32 != tmp32 tmp37 = tmp35 > tmp36 tmp38 = tmp33 | tmp37 tmp39 = tmp35 & tmp36 tmp40 = tmp34 | tmp39 tmp41 = tl.full([1], 3, tl.int64) tmp42 = tmp31 < tmp41 tmp43 = tmp40 & tmp42 tmp44 = tmp38 | tmp43 tl.where(tmp44, tmp30, tmp32) tmp46 = tl.where(tmp44, tmp31, tmp41) tl.store(out_ptr0 + x2, tmp46, xmask) @triton.jit def triton_per_fused__to_copy_eq_mean_1(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 % 64 r2 = rindex tmp0 = tl.load(in_ptr0 + r0, None, eviction_policy='evict_last') tmp2 = tl.load(in_ptr1 + r2, None) tmp1 = tmp0.to(tl.float32) tmp3 = tmp1 == tmp2 tmp4 = tmp3.to(tl.float32) tmp5 = tl.broadcast_to(tmp4, [RBLOCK]) tmp7 = triton_helpers.promote_to_tensor(tl.sum(tmp5, 0)) tmp8 = 256.0 tmp9 = tmp7 / tmp8 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp9, 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, 4), (16, 4, 1), torch.int64) get_raw_stream(0) triton_poi_fused_argmax_0[grid(64)](arg0_1, buf0, 64, XBLOCK=64, num_warps=1, num_stages=1) del arg0_1 buf1 = empty_strided_cuda((), (), torch.float32) buf2 = buf1 del buf1 triton_per_fused__to_copy_eq_mean_1[grid(1)](buf2, buf0, arg1_1, 1, 256, num_warps=2, num_stages=1) del arg1_1 del buf0 return buf2, class AccuracyNew(nn.Module): label = 'Accuracy' def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0]
cms-flash/beauty-net
Accuracy
false
15,052
[ "MIT" ]
155
668210a95ccb4462d7beff10505e4e83532682f2
https://github.com/cms-flash/beauty-net/tree/668210a95ccb4462d7beff10505e4e83532682f2
ConvBlock
import torch import torch.nn.functional as F import torch.nn as nn class ConvBlock(nn.Module): def __init__(self): super(ConvBlock, self).__init__() self.conv1 = nn.Conv2d(3, 6, 5) self.pool = nn.MaxPool2d(2, 2) self.conv2 = nn.Conv2d(6, 16, 5) def forward(self, x): x = self.pool(F.relu(self.conv1(x))) x = self.pool(F.relu(self.conv2(x))) x = x.view(-1, 16 * 5 * 5) return x def get_inputs(): return [torch.rand([4, 3, 32, 32])] 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_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 18816 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x3 = xindex x1 = xindex // 784 % 6 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_1(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr): xnumel = 4704 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex % 14 x3 = xindex // 14 x2 = xindex // 1176 x4 = xindex % 1176 tmp0 = tl.load(in_ptr0 + (2 * x0 + 56 * x3), xmask, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 56 * x3), xmask, eviction_policy ='evict_last') tmp3 = tl.load(in_ptr0 + (28 + 2 * x0 + 56 * x3), xmask, eviction_policy='evict_last') tmp5 = tl.load(in_ptr0 + (29 + 2 * x0 + 56 * x3), 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 + 1184 * x2), tmp6, xmask) tl.store(out_ptr1 + (x4 + 1280 * x2), tmp16, xmask) @triton.jit def triton_poi_fused_convolution_relu_2(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 6400 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x3 = xindex x1 = xindex // 100 % 16 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 = 1600 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex % 5 x1 = xindex // 5 x2 = xindex tmp0 = tl.load(in_ptr0 + (2 * x0 + 20 * x1), xmask, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 20 * x1), xmask, eviction_policy ='evict_last') tmp7 = tl.load(in_ptr0 + (10 + 2 * x0 + 20 * x1), xmask, eviction_policy='evict_last') tmp12 = tl.load(in_ptr0 + (11 + 2 * x0 + 20 * x1), 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 + x2, tmp15, xmask) tl.store(out_ptr1 + x2, tmp16, xmask) def call(args): primals_1, primals_2, primals_3, primals_4, primals_5 = args args.clear() assert_size_stride(primals_1, (6, 3, 5, 5), (75, 25, 5, 1)) assert_size_stride(primals_2, (6,), (1,)) assert_size_stride(primals_3, (4, 3, 32, 32), (3072, 1024, 32, 1)) assert_size_stride(primals_4, (16, 6, 5, 5), (150, 25, 5, 1)) assert_size_stride(primals_5, (16,), (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=1, bias=None) assert_size_stride(buf0, (4, 6, 28, 28), (4704, 784, 28, 1)) buf1 = buf0 del buf0 get_raw_stream(0) triton_poi_fused_convolution_relu_0[grid(18816)](buf1, primals_2, 18816, XBLOCK=256, num_warps=4, num_stages=1) del primals_2 buf2 = empty_strided_cuda((4, 6, 14, 14), (1184, 196, 14, 1), torch .float32) buf3 = empty_strided_cuda((4, 6, 14, 14), (1280, 196, 14, 1), torch .int8) triton_poi_fused_max_pool2d_with_indices_1[grid(4704)](buf1, buf2, buf3, 4704, XBLOCK=256, num_warps=4, num_stages=1) buf4 = 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(buf4, (4, 16, 10, 10), (1600, 100, 10, 1)) buf5 = buf4 del buf4 triton_poi_fused_convolution_relu_2[grid(6400)](buf5, primals_5, 6400, XBLOCK=256, num_warps=4, num_stages=1) del primals_5 buf6 = empty_strided_cuda((4, 16, 5, 5), (400, 25, 5, 1), torch.int8) buf7 = empty_strided_cuda((4, 16, 5, 5), (400, 25, 5, 1), torch.float32 ) triton_poi_fused_max_pool2d_with_indices_3[grid(1600)](buf5, buf6, buf7, 1600, XBLOCK=128, num_warps=4, num_stages=1) return reinterpret_tensor(buf7, (4, 400), (400, 1), 0 ), primals_1, primals_3, primals_4, buf1, buf2, buf3, buf5, buf6 class ConvBlockNew(nn.Module): def __init__(self): super(ConvBlockNew, self).__init__() self.conv1 = nn.Conv2d(3, 6, 5) self.pool = nn.MaxPool2d(2, 2) self.conv2 = nn.Conv2d(6, 16, 5) 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_3 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5]) return output[0]
coasxu/FedMA
ConvBlock
false
15,053
[ "MIT" ]
254
21f4d32338fd2563ebd97c737e3b9f4f470029d9
https://github.com/coasxu/FedMA/tree/21f4d32338fd2563ebd97c737e3b9f4f470029d9
HirarchicalAttention
from torch.nn import Module import torch from typing import * import torch.utils.data import torch.nn as nn import torch.onnx.operators import torch.optim class HirarchicalAttention(Module): """ ref: Hierarchical Attention Networks for Document Classification """ def __init__(self, hidden_size: 'int'): super(HirarchicalAttention, self).__init__() self.w_linear = nn.Linear(hidden_size, hidden_size) self.u_w = nn.Linear(hidden_size, 1, bias=False) def forward(self, input: 'torch.Tensor') ->torch.Tensor: u_it = torch.tanh(self.w_linear(input)) a_it = torch.softmax(self.u_w(u_it), dim=1) s_i = (input * a_it).sum(dim=1) return s_i def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'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 from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math from torch.nn import Module from typing import * import torch.utils.data import torch.nn as nn import torch.onnx.operators import torch.optim 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) @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__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 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 = tmp1 + tmp2 tmp5 = tmp3 + tmp4 tmp7 = tmp5 + tmp6 tmp8 = tmp0 / tmp7 tl.store(out_ptr0 + x3, tmp8, xmask) @triton.jit def triton_poi_fused__softmax_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 x2 = xindex // 16 x3 = xindex % 16 x1 = xindex // 4 % 4 x4 = xindex tmp0 = tl.load(in_ptr0 + (x3 + 64 * x2), xmask) tmp1 = tl.load(in_ptr1 + (x1 + 16 * x2), xmask, eviction_policy= 'evict_last') tmp3 = tl.load(in_ptr0 + (16 + x3 + 64 * x2), xmask) tmp4 = tl.load(in_ptr1 + (4 + x1 + 16 * x2), xmask, eviction_policy= 'evict_last') tmp7 = tl.load(in_ptr0 + (32 + x3 + 64 * x2), xmask) tmp8 = tl.load(in_ptr1 + (8 + x1 + 16 * x2), xmask, eviction_policy= 'evict_last') tmp11 = tl.load(in_ptr0 + (48 + x3 + 64 * x2), xmask) tmp12 = tl.load(in_ptr1 + (12 + x1 + 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) 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,), (1,)) assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_4, (1, 4), (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_tanh_0[grid(256)](buf1, primals_2, 256, XBLOCK=256, num_warps=4, num_stages=1) del primals_2 buf2 = empty_strided_cuda((64, 1), (1, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(buf1, (64, 4), (4, 1), 0), reinterpret_tensor(primals_4, (4, 1), (1, 4), 0), out=buf2) buf3 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32) triton_poi_fused__softmax_1[grid(64)](buf2, buf3, 64, XBLOCK=64, num_warps=1, num_stages=1) buf4 = empty_strided_cuda((4, 4, 4, 1), (16, 4, 1, 64), torch.float32) triton_poi_fused__softmax_2[grid(64)](buf3, buf4, 64, XBLOCK=64, num_warps=1, num_stages=1) buf5 = reinterpret_tensor(buf3, (4, 4, 4), (16, 4, 1), 0) del buf3 triton_poi_fused__softmax_mul_sum_3[grid(64)](primals_3, buf4, buf5, 64, XBLOCK=64, num_warps=1, num_stages=1) del buf4 return buf5, primals_3, buf1, buf2, primals_4 class HirarchicalAttentionNew(Module): """ ref: Hierarchical Attention Networks for Document Classification """ def __init__(self, hidden_size: 'int'): super(HirarchicalAttentionNew, self).__init__() self.w_linear = nn.Linear(hidden_size, hidden_size) self.u_w = nn.Linear(hidden_size, 1, bias=False) def forward(self, input_0): primals_1 = self.w_linear.weight primals_2 = self.w_linear.bias primals_4 = self.u_w.weight primals_3 = input_0 output = call([primals_1, primals_2, primals_3, primals_4]) return output[0]
code-backdoor/code-backdoor
HirarchicalAttention
false
15,054
[ "MIT" ]
71
1eeb3d79aa8a54c8f08e8d0156b569de5edd974e
https://github.com/code-backdoor/code-backdoor/tree/1eeb3d79aa8a54c8f08e8d0156b569de5edd974e
ConvLayer
import torch import torch.nn.functional as F from typing import * import torch.utils.data import torch.nn as nn import torch.onnx.operators import torch.optim class ConvLayer(nn.Module): def __init__(self, in_channels, out_channels): super(ConvLayer, self).__init__() self.in_channels = in_channels self.out_channels = out_channels self.weight = nn.Parameter(torch.Tensor(out_channels, in_channels, 1, 1)) nn.init.constant_(self.weight, 0.1) def forward(self, edges): edges = (edges * F.softmax(self.weight, dim=1)).sum(dim=1) return edges def extra_repr(self) ->str: return 'ConV {}'.format(self.weight.size()) def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_channels': 4, 'out_channels': 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 from typing import * import torch.utils.data import torch.nn as nn import torch.onnx.operators import torch.optim 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__softmax_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 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_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 = tmp1 + tmp2 tmp5 = tmp3 + tmp4 tmp7 = tmp5 + tmp6 tmp8 = tmp0 / tmp7 tl.store(out_ptr0 + x2, tmp8, xmask) @triton.jit def triton_poi_fused__softmax_mul_sum_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 % 16 x1 = xindex // 16 x2 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask) tmp1 = tl.load(in_ptr1 + 4 * x1, xmask, eviction_policy='evict_last') tmp3 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask) tmp4 = tl.load(in_ptr1 + (1 + 4 * x1), xmask, eviction_policy='evict_last') tmp7 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask) tmp8 = tl.load(in_ptr1 + (2 + 4 * x1), xmask, eviction_policy='evict_last') tmp11 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask) tmp12 = tl.load(in_ptr1 + (3 + 4 * x1), 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 + x2, tmp14, xmask) def call(args): primals_1, primals_2 = args args.clear() assert_size_stride(primals_1, (4, 4, 1, 1), (4, 1, 1, 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, 1, 1), (4, 1, 16, 16), torch.float32) get_raw_stream(0) triton_poi_fused__softmax_0[grid(16)](primals_1, buf0, 16, XBLOCK= 16, num_warps=1, num_stages=1) buf1 = empty_strided_cuda((4, 4, 1, 1), (4, 1, 16, 16), torch.float32) triton_poi_fused__softmax_1[grid(16)](buf0, buf1, 16, XBLOCK=16, num_warps=1, num_stages=1) del buf0 buf2 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused__softmax_mul_sum_2[grid(64)](primals_2, buf1, buf2, 64, XBLOCK=64, num_warps=1, num_stages=1) del buf1 return buf2, primals_1, primals_2 class ConvLayerNew(nn.Module): def __init__(self, in_channels, out_channels): super(ConvLayerNew, self).__init__() self.in_channels = in_channels self.out_channels = out_channels self.weight = nn.Parameter(torch.Tensor(out_channels, in_channels, 1, 1)) nn.init.constant_(self.weight, 0.1) def extra_repr(self) ->str: return 'ConV {}'.format(self.weight.size()) def forward(self, input_0): primals_1 = self.weight primals_2 = input_0 output = call([primals_1, primals_2]) return output[0]
code-backdoor/code-backdoor
ConvLayer
false
15,055
[ "MIT" ]
71
1eeb3d79aa8a54c8f08e8d0156b569de5edd974e
https://github.com/code-backdoor/code-backdoor/tree/1eeb3d79aa8a54c8f08e8d0156b569de5edd974e
SimpleCNNContainerConvBlocks
import torch import torch.nn.functional as F import torch.nn as nn class SimpleCNNContainerConvBlocks(nn.Module): def __init__(self, input_channel, num_filters, kernel_size, output_dim=10): super(SimpleCNNContainerConvBlocks, self).__init__() """ A testing cnn container, which allows initializing a CNN with given dims We use this one to estimate matched output of conv blocks num_filters (list) :: number of convolution filters hidden_dims (list) :: number of neurons in hidden layers Assumptions: i) we use only two conv layers and three hidden layers (including the output layer) ii) kernel size in the two conv layers are identical """ self.conv1 = nn.Conv2d(input_channel, num_filters[0], kernel_size) self.pool = nn.MaxPool2d(2, 2) self.conv2 = nn.Conv2d(num_filters[0], num_filters[1], kernel_size) def forward(self, x): x = self.pool(F.relu(self.conv1(x))) x = self.pool(F.relu(self.conv2(x))) return x def get_inputs(): return [torch.rand([4, 4, 64, 64])] def get_init_inputs(): return [[], {'input_channel': 4, 'num_filters': [4, 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 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_convolution_relu_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 59536 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x3 = xindex x1 = xindex // 3721 % 4 x0 = xindex % 3721 x4 = xindex // 3721 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 + 3744 * 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 = 14400 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex % 30 x1 = xindex // 30 % 30 x4 = xindex // 900 x3 = xindex // 3600 x5 = xindex % 3600 tmp0 = tl.load(in_ptr0 + (2 * x0 + 122 * x1 + 3744 * x4), xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 122 * x1 + 3744 * x4), xmask, eviction_policy='evict_last') tmp3 = tl.load(in_ptr0 + (61 + 2 * x0 + 122 * x1 + 3744 * x4), xmask, eviction_policy='evict_last') tmp5 = tl.load(in_ptr0 + (62 + 2 * x0 + 122 * x1 + 3744 * x4), 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 + (x5 + 3616 * x3), tmp6, xmask) tl.store(out_ptr1 + (x5 + 3712 * x3), tmp16, xmask) @triton.jit def triton_poi_fused_convolution_relu_2(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 11664 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x3 = xindex x1 = xindex // 729 % 4 x2 = xindex // 2916 x4 = xindex % 2916 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 + (x4 + 2944 * x2), tmp4, xmask) @triton.jit def triton_poi_fused_max_pool2d_with_indices_3(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr): xnumel = 2704 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex % 13 x1 = xindex // 13 % 13 x2 = xindex // 169 % 4 x3 = xindex // 676 x4 = xindex tmp0 = tl.load(in_ptr0 + (2 * x0 + 54 * x1 + 729 * x2 + 2944 * x3), xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 54 * x1 + 729 * x2 + 2944 * x3), xmask, eviction_policy='evict_last') tmp3 = tl.load(in_ptr0 + (27 + 2 * x0 + 54 * x1 + 729 * x2 + 2944 * x3), xmask, eviction_policy='evict_last') tmp5 = tl.load(in_ptr0 + (28 + 2 * x0 + 54 * x1 + 729 * x2 + 2944 * x3), 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 + x4, tmp16, 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,), (1,)) assert_size_stride(primals_3, (4, 4, 64, 64), (16384, 4096, 64, 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 = extern_kernels.convolution(primals_3, 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, 4, 61, 61), (14884, 3721, 61, 1)) buf1 = empty_strided_cuda((4, 4, 61, 61), (14976, 3744, 61, 1), torch.float32) get_raw_stream(0) triton_poi_fused_convolution_relu_0[grid(59536)](buf0, primals_2, buf1, 59536, XBLOCK=256, num_warps=4, num_stages=1) del buf0 del primals_2 buf2 = empty_strided_cuda((4, 4, 30, 30), (3616, 900, 30, 1), torch .float32) buf3 = empty_strided_cuda((4, 4, 30, 30), (3712, 900, 30, 1), torch .int8) triton_poi_fused_max_pool2d_with_indices_1[grid(14400)](buf1, buf2, buf3, 14400, XBLOCK=256, num_warps=4, num_stages=1) buf4 = 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(buf4, (4, 4, 27, 27), (2916, 729, 27, 1)) buf5 = empty_strided_cuda((4, 4, 27, 27), (2944, 729, 27, 1), torch .float32) triton_poi_fused_convolution_relu_2[grid(11664)](buf4, primals_5, buf5, 11664, XBLOCK=256, num_warps=4, num_stages=1) del buf4 del primals_5 buf6 = empty_strided_cuda((4, 4, 13, 13), (676, 169, 13, 1), torch. float32) buf7 = empty_strided_cuda((4, 4, 13, 13), (676, 169, 13, 1), torch.int8 ) triton_poi_fused_max_pool2d_with_indices_3[grid(2704)](buf5, buf6, buf7, 2704, XBLOCK=128, num_warps=4, num_stages=1) return buf6, primals_1, primals_3, primals_4, buf1, buf2, buf3, buf5, buf7 class SimpleCNNContainerConvBlocksNew(nn.Module): def __init__(self, input_channel, num_filters, kernel_size, output_dim=10): super(SimpleCNNContainerConvBlocksNew, self).__init__() """ A testing cnn container, which allows initializing a CNN with given dims We use this one to estimate matched output of conv blocks num_filters (list) :: number of convolution filters hidden_dims (list) :: number of neurons in hidden layers Assumptions: i) we use only two conv layers and three hidden layers (including the output layer) ii) kernel size in the two conv layers are identical """ self.conv1 = nn.Conv2d(input_channel, num_filters[0], kernel_size) self.pool = nn.MaxPool2d(2, 2) self.conv2 = nn.Conv2d(num_filters[0], num_filters[1], kernel_size) 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_3 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5]) return output[0]
coasxu/FedMA
SimpleCNNContainerConvBlocks
false
15,056
[ "MIT" ]
254
21f4d32338fd2563ebd97c737e3b9f4f470029d9
https://github.com/coasxu/FedMA/tree/21f4d32338fd2563ebd97c737e3b9f4f470029d9
UNet
import torch import torch.nn as nn import torch.nn.functional as F class down(nn.Module): """ A class for creating neural network blocks containing layers: Average Pooling --> Convlution + Leaky ReLU --> Convolution + Leaky ReLU This is used in the UNet Class to create a UNet like NN architecture. ... Methods ------- forward(x) Returns output tensor after passing input `x` to the neural network block. """ def __init__(self, inChannels, outChannels, filterSize): """ Parameters ---------- inChannels : int number of input channels for the first convolutional layer. outChannels : int number of output channels for the first convolutional layer. This is also used as input and output channels for the second convolutional layer. filterSize : int filter size for the convolution filter. input N would create a N x N filter. """ super(down, self).__init__() self.conv1 = nn.Conv2d(inChannels, outChannels, filterSize, stride= 1, padding=int((filterSize - 1) / 2)) self.conv2 = nn.Conv2d(outChannels, outChannels, filterSize, stride =1, padding=int((filterSize - 1) / 2)) def forward(self, x): """ Returns output tensor after passing input `x` to the neural network block. Parameters ---------- x : tensor input to the NN block. Returns ------- tensor output of the NN block. """ x = F.avg_pool2d(x, 2) x = F.leaky_relu(self.conv1(x), negative_slope=0.1) x = F.leaky_relu(self.conv2(x), negative_slope=0.1) return x class up(nn.Module): """ A class for creating neural network blocks containing layers: Bilinear interpolation --> Convlution + Leaky ReLU --> Convolution + Leaky ReLU This is used in the UNet Class to create a UNet like NN architecture. ... Methods ------- forward(x, skpCn) Returns output tensor after passing input `x` to the neural network block. """ def __init__(self, inChannels, outChannels): """ Parameters ---------- inChannels : int number of input channels for the first convolutional layer. outChannels : int number of output channels for the first convolutional layer. This is also used for setting input and output channels for the second convolutional layer. """ super(up, self).__init__() self.conv1 = nn.Conv2d(inChannels, outChannels, 3, stride=1, padding=1) self.conv2 = nn.Conv2d(2 * outChannels, outChannels, 3, stride=1, padding=1) def forward(self, x, skpCn): """ Returns output tensor after passing input `x` to the neural network block. Parameters ---------- x : tensor input to the NN block. skpCn : tensor skip connection input to the NN block. Returns ------- tensor output of the NN block. """ x = F.interpolate(x, scale_factor=2, mode='bilinear') x = F.leaky_relu(self.conv1(x), negative_slope=0.1) x = F.leaky_relu(self.conv2(torch.cat((x, skpCn), 1)), negative_slope=0.1) return x class UNet(nn.Module): """ A class for creating UNet architecture with skip connections. ... Methods ------- forward(x) Returns output tensor after passing input `x` to the neural network block. """ def __init__(self, inChannels, outChannels): """ Parameters ---------- inChannels : int number of input channels for the UNet. outChannels : int number of output channels for the UNet. """ super(UNet, self).__init__() self.conv1 = nn.Conv2d(inChannels, 32, 9, stride=1, padding=4) self.conv2 = nn.Conv2d(32, 32, 7, stride=1, padding=3) self.down1 = down(32, 64, 5) self.down2 = down(64, 128, 3) self.down3 = down(128, 256, 3) self.down4 = down(256, 512, 3) self.down5 = down(512, 512, 3) self.up1 = up(512, 512) self.up2 = up(512, 256) self.up3 = up(256, 128) self.up4 = up(128, 64) self.up5 = up(64, 32) self.conv3 = nn.Conv2d(32, outChannels, 3, stride=1, padding=1) def forward(self, x): """ Returns output tensor after passing input `x` to the neural network. Parameters ---------- x : tensor input to the UNet. Returns ------- tensor output of the UNet. """ x = F.leaky_relu(self.conv1(x), negative_slope=0.1) s1 = F.leaky_relu(self.conv2(x), negative_slope=0.1) s2 = self.down1(s1) s3 = self.down2(s2) s4 = self.down3(s3) s5 = self.down4(s4) x = self.down5(s5) x = self.up1(x, s5) x = self.up2(x, s4) x = self.up3(x, s3) x = self.up4(x, s2) x = self.up5(x, s1) x = F.leaky_relu(self.conv3(x), negative_slope=0.1) return x def get_inputs(): return [torch.rand([4, 4, 64, 64])] def get_init_inputs(): return [[], {'inChannels': 4, 'outChannels': 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 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_convolution_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) x3 = xindex x1 = xindex // 4096 % 32 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tmp5 = 0.1 tmp6 = tmp2 * tmp5 tmp7 = tl.where(tmp4, tmp2, tmp6) tl.store(out_ptr0 + x3, tmp4, None) tl.store(out_ptr1 + x3, tmp7, None) @triton.jit def triton_poi_fused_avg_pool2d_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) x0 = xindex % 32 x1 = xindex // 32 x2 = xindex tmp0 = tl.load(in_ptr0 + (2 * x0 + 128 * x1), None, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 128 * x1), None, eviction_policy ='evict_last') tmp3 = tl.load(in_ptr0 + (64 + 2 * x0 + 128 * x1), None, eviction_policy='evict_last') tmp5 = tl.load(in_ptr0 + (65 + 2 * x0 + 128 * x1), None, eviction_policy='evict_last') tmp2 = tmp1 + tmp0 tmp4 = tmp3 + tmp2 tmp6 = tmp5 + tmp4 tmp7 = 0.25 tmp8 = tmp6 * tmp7 tl.store(out_ptr0 + x2, tmp8, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_2(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) x3 = xindex x1 = xindex // 1024 % 64 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tmp5 = 0.1 tmp6 = tmp2 * tmp5 tmp7 = tl.where(tmp4, tmp2, tmp6) tl.store(out_ptr0 + x3, tmp4, None) tl.store(out_ptr1 + x3, tmp7, None) @triton.jit def triton_poi_fused_avg_pool2d_3(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 % 16 x1 = xindex // 16 x2 = xindex tmp0 = tl.load(in_ptr0 + (2 * x0 + 64 * x1), None, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 64 * x1), None, eviction_policy= 'evict_last') tmp3 = tl.load(in_ptr0 + (32 + 2 * x0 + 64 * x1), None, eviction_policy ='evict_last') tmp5 = tl.load(in_ptr0 + (33 + 2 * x0 + 64 * x1), None, eviction_policy ='evict_last') tmp2 = tmp1 + tmp0 tmp4 = tmp3 + tmp2 tmp6 = tmp5 + tmp4 tmp7 = 0.25 tmp8 = tmp6 * tmp7 tl.store(out_ptr0 + x2, tmp8, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_4(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) x3 = xindex x1 = xindex // 256 % 128 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tmp5 = 0.1 tmp6 = tmp2 * tmp5 tmp7 = tl.where(tmp4, tmp2, tmp6) tl.store(out_ptr0 + x3, tmp4, None) tl.store(out_ptr1 + x3, tmp7, None) @triton.jit def triton_poi_fused_avg_pool2d_5(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 % 8 x1 = xindex // 8 x2 = xindex tmp0 = tl.load(in_ptr0 + (2 * x0 + 32 * x1), None, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 32 * x1), None, eviction_policy= 'evict_last') tmp3 = tl.load(in_ptr0 + (16 + 2 * x0 + 32 * x1), None, eviction_policy ='evict_last') tmp5 = tl.load(in_ptr0 + (17 + 2 * x0 + 32 * x1), None, eviction_policy ='evict_last') tmp2 = tmp1 + tmp0 tmp4 = tmp3 + tmp2 tmp6 = tmp5 + tmp4 tmp7 = 0.25 tmp8 = tmp6 * tmp7 tl.store(out_ptr0 + x2, tmp8, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_6(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) x3 = xindex x1 = xindex // 64 % 256 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tmp5 = 0.1 tmp6 = tmp2 * tmp5 tmp7 = tl.where(tmp4, tmp2, tmp6) tl.store(out_ptr0 + x3, tmp4, None) tl.store(out_ptr1 + x3, tmp7, None) @triton.jit def triton_poi_fused_avg_pool2d_7(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 % 4 x1 = xindex // 4 x2 = xindex tmp0 = tl.load(in_ptr0 + (2 * x0 + 16 * x1), None, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 16 * x1), None, eviction_policy= 'evict_last') tmp3 = tl.load(in_ptr0 + (8 + 2 * x0 + 16 * x1), None, eviction_policy= 'evict_last') tmp5 = tl.load(in_ptr0 + (9 + 2 * x0 + 16 * x1), None, eviction_policy= 'evict_last') tmp2 = tmp1 + tmp0 tmp4 = tmp3 + tmp2 tmp6 = tmp5 + tmp4 tmp7 = 0.25 tmp8 = tmp6 * tmp7 tl.store(out_ptr0 + x2, tmp8, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_8(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) x3 = xindex x1 = xindex // 16 % 512 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tmp5 = 0.1 tmp6 = tmp2 * tmp5 tmp7 = tl.where(tmp4, tmp2, tmp6) tl.store(out_ptr0 + x3, tmp4, None) tl.store(out_ptr1 + x3, tmp7, None) @triton.jit def triton_poi_fused_avg_pool2d_9(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 % 2 x1 = xindex // 2 x2 = xindex tmp0 = tl.load(in_ptr0 + (2 * x0 + 8 * x1), None, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr0 + (1 + 2 * x0 + 8 * x1), None, eviction_policy= 'evict_last') tmp3 = tl.load(in_ptr0 + (4 + 2 * x0 + 8 * x1), None, eviction_policy= 'evict_last') tmp5 = tl.load(in_ptr0 + (5 + 2 * x0 + 8 * x1), None, eviction_policy= 'evict_last') tmp2 = tmp1 + tmp0 tmp4 = tmp3 + tmp2 tmp6 = tmp5 + tmp4 tmp7 = 0.25 tmp8 = tmp6 * tmp7 tl.store(out_ptr0 + x2, tmp8, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_10(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) x3 = xindex x1 = xindex // 4 % 512 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tmp5 = 0.1 tmp6 = tmp2 * tmp5 tmp7 = tl.where(tmp4, tmp2, tmp6) tl.store(out_ptr0 + x3, tmp4, None) tl.store(out_ptr1 + x3, tmp7, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_11(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) x3 = xindex x1 = xindex // 4 % 512 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tl.store(out_ptr0 + x3, tmp4, None) @triton.jit def triton_poi_fused__to_copy_12(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 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tl.store(out_ptr0 + x0, tmp8, xmask) @triton.jit def triton_poi_fused_add_clamp_13(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 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tmp9 = tl.full([1], 1, tl.int64) tmp10 = tmp8 + tmp9 tmp11 = triton_helpers.minimum(tmp10, tmp9) tl.store(out_ptr0 + x0, tmp11, xmask) @triton.jit def triton_poi_fused__to_copy_add_arange_clamp_mul_sub_14(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 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tmp9 = tmp8.to(tl.float32) tmp10 = tmp7 - tmp9 tmp11 = triton_helpers.maximum(tmp10, tmp6) tmp12 = 1.0 tmp13 = triton_helpers.minimum(tmp11, tmp12) tl.store(out_ptr0 + x0, tmp13, xmask) @triton.jit def triton_poi_fused__unsafe_index_add_convolution_leaky_relu_mul_sub_15( in_out_ptr1, in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, in_ptr6, in_ptr7, in_ptr8, xnumel, XBLOCK: tl.constexpr): xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] tl.full([XBLOCK], True, tl.int1) x1 = xindex // 4 % 4 x0 = xindex % 4 x6 = xindex // 16 x2 = xindex // 16 % 512 x4 = xindex tmp0 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last') tmp5 = tl.load(in_ptr1 + x0, None, eviction_policy='evict_last') tmp11 = tl.load(in_ptr4 + x2, None, eviction_policy='evict_last') tmp16 = tl.load(in_ptr5 + x1, None, eviction_policy='evict_last') tmp25 = tl.load(in_ptr6 + x0, None, eviction_policy='evict_last') tmp35 = tl.load(in_ptr7 + x0, None, eviction_policy='evict_last') tmp47 = tl.load(in_ptr8 + x1, None, eviction_policy='evict_last') tmp1 = tl.full([XBLOCK], 2, tl.int32) tmp2 = tmp0 + tmp1 tmp3 = tmp0 < 0 tmp4 = tl.where(tmp3, tmp2, tmp0) tmp6 = tmp5 + tmp1 tmp7 = tmp5 < 0 tmp8 = tl.where(tmp7, tmp6, tmp5) tmp9 = tl.load(in_ptr2 + (tmp8 + 2 * tmp4 + 4 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp10 = tl.load(in_ptr3 + (tmp8 + 2 * tmp4 + 4 * x6), None, eviction_policy='evict_last') tmp12 = tmp10 + tmp11 tmp13 = 0.1 tmp14 = tmp12 * tmp13 tmp15 = tl.where(tmp9, tmp12, tmp14) tmp17 = tmp16 + tmp1 tmp18 = tmp16 < 0 tmp19 = tl.where(tmp18, tmp17, tmp16) tmp20 = tl.load(in_ptr2 + (tmp8 + 2 * tmp19 + 4 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp21 = tl.load(in_ptr3 + (tmp8 + 2 * tmp19 + 4 * x6), None, eviction_policy='evict_last') tmp22 = tmp21 + tmp11 tmp23 = tmp22 * tmp13 tmp24 = tl.where(tmp20, tmp22, tmp23) tmp26 = tmp25 + tmp1 tmp27 = tmp25 < 0 tmp28 = tl.where(tmp27, tmp26, tmp25) tmp29 = tl.load(in_ptr2 + (tmp28 + 2 * tmp19 + 4 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp30 = tl.load(in_ptr3 + (tmp28 + 2 * tmp19 + 4 * x6), None, eviction_policy='evict_last') tmp31 = tmp30 + tmp11 tmp32 = tmp31 * tmp13 tmp33 = tl.where(tmp29, tmp31, tmp32) tmp34 = tmp33 - tmp24 tmp36 = tmp34 * tmp35 tmp37 = tmp24 + tmp36 tmp38 = tl.load(in_ptr2 + (tmp28 + 2 * tmp4 + 4 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp39 = tl.load(in_ptr3 + (tmp28 + 2 * tmp4 + 4 * x6), None, eviction_policy='evict_last') tmp40 = tmp39 + tmp11 tmp41 = tmp40 * tmp13 tmp42 = tl.where(tmp38, tmp40, tmp41) tmp43 = tmp42 - tmp15 tmp44 = tmp43 * tmp35 tmp45 = tmp15 + tmp44 tmp46 = tmp45 - tmp37 tmp48 = tmp46 * tmp47 tmp49 = tmp37 + tmp48 tl.store(in_out_ptr1 + x4, tmp49, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_16(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) x3 = xindex x1 = xindex // 16 % 512 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tl.store(out_ptr0 + x3, tmp4, None) @triton.jit def triton_poi_fused_cat_17(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) x1 = xindex // 16 % 1024 x0 = xindex % 16 x2 = xindex // 16384 x3 = xindex tmp0 = x1 tl.full([1], 0, tl.int64) tmp3 = tl.full([1], 512, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tl.load(in_ptr0 + (x0 + 16 * x1 + 8192 * x2), tmp4, other=0.0).to(tl .int1) tmp6 = tl.load(in_ptr1 + (x0 + 16 * x1 + 8192 * x2), tmp4, other=0.0) tmp7 = tl.load(in_ptr2 + x1, tmp4, eviction_policy='evict_last', other=0.0) tmp8 = tmp6 + tmp7 tmp9 = 0.1 tmp10 = tmp8 * tmp9 tmp11 = tl.where(tmp5, tmp8, tmp10) tmp12 = tl.full(tmp11.shape, 0.0, tmp11.dtype) tmp13 = tl.where(tmp4, tmp11, tmp12) tmp14 = tmp0 >= tmp3 tl.full([1], 1024, tl.int64) tmp17 = tl.load(in_ptr3 + (x0 + 16 * (-512 + x1) + 8192 * x2), tmp14, other=0.0) tmp18 = tl.where(tmp4, tmp13, tmp17) tl.store(out_ptr0 + x3, tmp18, None) @triton.jit def triton_poi_fused__to_copy_18(out_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 8 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex tmp0 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tl.store(out_ptr0 + x0, tmp8, xmask) @triton.jit def triton_poi_fused_add_clamp_19(out_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 8 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex tmp0 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tmp9 = tl.full([1], 1, tl.int64) tmp10 = tmp8 + tmp9 tmp11 = tl.full([1], 3, tl.int64) tmp12 = triton_helpers.minimum(tmp10, tmp11) tl.store(out_ptr0 + x0, tmp12, xmask) @triton.jit def triton_poi_fused__to_copy_add_arange_clamp_mul_sub_20(out_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 8 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex tmp0 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tmp9 = tmp8.to(tl.float32) tmp10 = tmp7 - tmp9 tmp11 = triton_helpers.maximum(tmp10, tmp6) tmp12 = 1.0 tmp13 = triton_helpers.minimum(tmp11, tmp12) tl.store(out_ptr0 + x0, tmp13, xmask) @triton.jit def triton_poi_fused__unsafe_index_add_convolution_leaky_relu_mul_sub_21( in_out_ptr1, in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, in_ptr6, in_ptr7, in_ptr8, xnumel, XBLOCK: tl.constexpr): xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] tl.full([XBLOCK], True, tl.int1) x1 = xindex // 8 % 8 x0 = xindex % 8 x6 = xindex // 64 x2 = xindex // 64 % 512 x4 = xindex tmp0 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last') tmp5 = tl.load(in_ptr1 + x0, None, eviction_policy='evict_last') tmp11 = tl.load(in_ptr4 + x2, None, eviction_policy='evict_last') tmp16 = tl.load(in_ptr5 + x1, None, eviction_policy='evict_last') tmp25 = tl.load(in_ptr6 + x0, None, eviction_policy='evict_last') tmp35 = tl.load(in_ptr7 + x0, None, eviction_policy='evict_last') tmp47 = tl.load(in_ptr8 + x1, None, eviction_policy='evict_last') tmp1 = tl.full([XBLOCK], 4, tl.int32) tmp2 = tmp0 + tmp1 tmp3 = tmp0 < 0 tmp4 = tl.where(tmp3, tmp2, tmp0) tmp6 = tmp5 + tmp1 tmp7 = tmp5 < 0 tmp8 = tl.where(tmp7, tmp6, tmp5) tmp9 = tl.load(in_ptr2 + (tmp8 + 4 * tmp4 + 16 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp10 = tl.load(in_ptr3 + (tmp8 + 4 * tmp4 + 16 * x6), None, eviction_policy='evict_last') tmp12 = tmp10 + tmp11 tmp13 = 0.1 tmp14 = tmp12 * tmp13 tmp15 = tl.where(tmp9, tmp12, tmp14) tmp17 = tmp16 + tmp1 tmp18 = tmp16 < 0 tmp19 = tl.where(tmp18, tmp17, tmp16) tmp20 = tl.load(in_ptr2 + (tmp8 + 4 * tmp19 + 16 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp21 = tl.load(in_ptr3 + (tmp8 + 4 * tmp19 + 16 * x6), None, eviction_policy='evict_last') tmp22 = tmp21 + tmp11 tmp23 = tmp22 * tmp13 tmp24 = tl.where(tmp20, tmp22, tmp23) tmp26 = tmp25 + tmp1 tmp27 = tmp25 < 0 tmp28 = tl.where(tmp27, tmp26, tmp25) tmp29 = tl.load(in_ptr2 + (tmp28 + 4 * tmp19 + 16 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp30 = tl.load(in_ptr3 + (tmp28 + 4 * tmp19 + 16 * x6), None, eviction_policy='evict_last') tmp31 = tmp30 + tmp11 tmp32 = tmp31 * tmp13 tmp33 = tl.where(tmp29, tmp31, tmp32) tmp34 = tmp33 - tmp24 tmp36 = tmp34 * tmp35 tmp37 = tmp24 + tmp36 tmp38 = tl.load(in_ptr2 + (tmp28 + 4 * tmp4 + 16 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp39 = tl.load(in_ptr3 + (tmp28 + 4 * tmp4 + 16 * x6), None, eviction_policy='evict_last') tmp40 = tmp39 + tmp11 tmp41 = tmp40 * tmp13 tmp42 = tl.where(tmp38, tmp40, tmp41) tmp43 = tmp42 - tmp15 tmp44 = tmp43 * tmp35 tmp45 = tmp15 + tmp44 tmp46 = tmp45 - tmp37 tmp48 = tmp46 * tmp47 tmp49 = tmp37 + tmp48 tl.store(in_out_ptr1 + x4, tmp49, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_22(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) x3 = xindex x1 = xindex // 64 % 256 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tl.store(out_ptr0 + x3, tmp4, None) @triton.jit def triton_poi_fused_cat_23(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) x1 = xindex // 64 % 512 x0 = xindex % 64 x2 = xindex // 32768 x3 = xindex tmp0 = x1 tl.full([1], 0, tl.int64) tmp3 = tl.full([1], 256, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tl.load(in_ptr0 + (x0 + 64 * x1 + 16384 * x2), tmp4, other=0.0).to( tl.int1) tmp6 = tl.load(in_ptr1 + (x0 + 64 * x1 + 16384 * x2), tmp4, other=0.0) tmp7 = tl.load(in_ptr2 + x1, tmp4, eviction_policy='evict_last', other=0.0) tmp8 = tmp6 + tmp7 tmp9 = 0.1 tmp10 = tmp8 * tmp9 tmp11 = tl.where(tmp5, tmp8, tmp10) tmp12 = tl.full(tmp11.shape, 0.0, tmp11.dtype) tmp13 = tl.where(tmp4, tmp11, tmp12) tmp14 = tmp0 >= tmp3 tl.full([1], 512, tl.int64) tmp17 = tl.load(in_ptr3 + (x0 + 64 * (-256 + x1) + 16384 * x2), tmp14, other=0.0) tmp18 = tl.where(tmp4, tmp13, tmp17) tl.store(out_ptr0 + x3, tmp18, None) @triton.jit def triton_poi_fused__to_copy_24(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 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tl.store(out_ptr0 + x0, tmp8, xmask) @triton.jit def triton_poi_fused_add_clamp_25(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 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tmp9 = tl.full([1], 1, tl.int64) tmp10 = tmp8 + tmp9 tmp11 = tl.full([1], 7, tl.int64) tmp12 = triton_helpers.minimum(tmp10, tmp11) tl.store(out_ptr0 + x0, tmp12, xmask) @triton.jit def triton_poi_fused__to_copy_add_arange_clamp_mul_sub_26(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 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tmp9 = tmp8.to(tl.float32) tmp10 = tmp7 - tmp9 tmp11 = triton_helpers.maximum(tmp10, tmp6) tmp12 = 1.0 tmp13 = triton_helpers.minimum(tmp11, tmp12) tl.store(out_ptr0 + x0, tmp13, xmask) @triton.jit def triton_poi_fused__unsafe_index_add_convolution_leaky_relu_mul_sub_27( in_out_ptr1, in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, in_ptr6, in_ptr7, in_ptr8, xnumel, XBLOCK: tl.constexpr): xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] tl.full([XBLOCK], True, tl.int1) x1 = xindex // 16 % 16 x0 = xindex % 16 x6 = xindex // 256 x2 = xindex // 256 % 256 x4 = xindex tmp0 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last') tmp5 = tl.load(in_ptr1 + x0, None, eviction_policy='evict_last') tmp11 = tl.load(in_ptr4 + x2, None, eviction_policy='evict_last') tmp16 = tl.load(in_ptr5 + x1, None, eviction_policy='evict_last') tmp25 = tl.load(in_ptr6 + x0, None, eviction_policy='evict_last') tmp35 = tl.load(in_ptr7 + x0, None, eviction_policy='evict_last') tmp47 = tl.load(in_ptr8 + x1, None, eviction_policy='evict_last') tmp1 = tl.full([XBLOCK], 8, tl.int32) tmp2 = tmp0 + tmp1 tmp3 = tmp0 < 0 tmp4 = tl.where(tmp3, tmp2, tmp0) tmp6 = tmp5 + tmp1 tmp7 = tmp5 < 0 tmp8 = tl.where(tmp7, tmp6, tmp5) tmp9 = tl.load(in_ptr2 + (tmp8 + 8 * tmp4 + 64 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp10 = tl.load(in_ptr3 + (tmp8 + 8 * tmp4 + 64 * x6), None, eviction_policy='evict_last') tmp12 = tmp10 + tmp11 tmp13 = 0.1 tmp14 = tmp12 * tmp13 tmp15 = tl.where(tmp9, tmp12, tmp14) tmp17 = tmp16 + tmp1 tmp18 = tmp16 < 0 tmp19 = tl.where(tmp18, tmp17, tmp16) tmp20 = tl.load(in_ptr2 + (tmp8 + 8 * tmp19 + 64 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp21 = tl.load(in_ptr3 + (tmp8 + 8 * tmp19 + 64 * x6), None, eviction_policy='evict_last') tmp22 = tmp21 + tmp11 tmp23 = tmp22 * tmp13 tmp24 = tl.where(tmp20, tmp22, tmp23) tmp26 = tmp25 + tmp1 tmp27 = tmp25 < 0 tmp28 = tl.where(tmp27, tmp26, tmp25) tmp29 = tl.load(in_ptr2 + (tmp28 + 8 * tmp19 + 64 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp30 = tl.load(in_ptr3 + (tmp28 + 8 * tmp19 + 64 * x6), None, eviction_policy='evict_last') tmp31 = tmp30 + tmp11 tmp32 = tmp31 * tmp13 tmp33 = tl.where(tmp29, tmp31, tmp32) tmp34 = tmp33 - tmp24 tmp36 = tmp34 * tmp35 tmp37 = tmp24 + tmp36 tmp38 = tl.load(in_ptr2 + (tmp28 + 8 * tmp4 + 64 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp39 = tl.load(in_ptr3 + (tmp28 + 8 * tmp4 + 64 * x6), None, eviction_policy='evict_last') tmp40 = tmp39 + tmp11 tmp41 = tmp40 * tmp13 tmp42 = tl.where(tmp38, tmp40, tmp41) tmp43 = tmp42 - tmp15 tmp44 = tmp43 * tmp35 tmp45 = tmp15 + tmp44 tmp46 = tmp45 - tmp37 tmp48 = tmp46 * tmp47 tmp49 = tmp37 + tmp48 tl.store(in_out_ptr1 + x4, tmp49, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_28(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) x3 = xindex x1 = xindex // 256 % 128 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tl.store(out_ptr0 + x3, tmp4, None) @triton.jit def triton_poi_fused_cat_29(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) x1 = xindex // 256 % 256 x0 = xindex % 256 x2 = xindex // 65536 x3 = xindex tmp0 = x1 tl.full([1], 0, tl.int64) tmp3 = tl.full([1], 128, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tl.load(in_ptr0 + (x0 + 256 * x1 + 32768 * x2), tmp4, other=0.0).to( tl.int1) tmp6 = tl.load(in_ptr1 + (x0 + 256 * x1 + 32768 * x2), tmp4, other=0.0) tmp7 = tl.load(in_ptr2 + x1, tmp4, eviction_policy='evict_last', other=0.0) tmp8 = tmp6 + tmp7 tmp9 = 0.1 tmp10 = tmp8 * tmp9 tmp11 = tl.where(tmp5, tmp8, tmp10) tmp12 = tl.full(tmp11.shape, 0.0, tmp11.dtype) tmp13 = tl.where(tmp4, tmp11, tmp12) tmp14 = tmp0 >= tmp3 tl.full([1], 256, tl.int64) tmp17 = tl.load(in_ptr3 + (x0 + 256 * (-128 + x1) + 32768 * x2), tmp14, other=0.0) tmp18 = tl.where(tmp4, tmp13, tmp17) tl.store(out_ptr0 + x3, tmp18, None) @triton.jit def triton_poi_fused__to_copy_30(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 tmp0 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tl.store(out_ptr0 + x0, tmp8, xmask) @triton.jit def triton_poi_fused_add_clamp_31(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 tmp0 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tmp9 = tl.full([1], 1, tl.int64) tmp10 = tmp8 + tmp9 tmp11 = tl.full([1], 15, tl.int64) tmp12 = triton_helpers.minimum(tmp10, tmp11) tl.store(out_ptr0 + x0, tmp12, xmask) @triton.jit def triton_poi_fused__to_copy_add_arange_clamp_mul_sub_32(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 tmp0 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tmp9 = tmp8.to(tl.float32) tmp10 = tmp7 - tmp9 tmp11 = triton_helpers.maximum(tmp10, tmp6) tmp12 = 1.0 tmp13 = triton_helpers.minimum(tmp11, tmp12) tl.store(out_ptr0 + x0, tmp13, xmask) @triton.jit def triton_poi_fused__unsafe_index_add_convolution_leaky_relu_mul_sub_33( in_out_ptr1, in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, in_ptr6, in_ptr7, in_ptr8, xnumel, XBLOCK: tl.constexpr): xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] tl.full([XBLOCK], True, tl.int1) x1 = xindex // 32 % 32 x0 = xindex % 32 x6 = xindex // 1024 x2 = xindex // 1024 % 128 x4 = xindex tmp0 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last') tmp5 = tl.load(in_ptr1 + x0, None, eviction_policy='evict_last') tmp11 = tl.load(in_ptr4 + x2, None, eviction_policy='evict_last') tmp16 = tl.load(in_ptr5 + x1, None, eviction_policy='evict_last') tmp25 = tl.load(in_ptr6 + x0, None, eviction_policy='evict_last') tmp35 = tl.load(in_ptr7 + x0, None, eviction_policy='evict_last') tmp47 = tl.load(in_ptr8 + x1, None, eviction_policy='evict_last') tmp1 = tl.full([XBLOCK], 16, tl.int32) tmp2 = tmp0 + tmp1 tmp3 = tmp0 < 0 tmp4 = tl.where(tmp3, tmp2, tmp0) tmp6 = tmp5 + tmp1 tmp7 = tmp5 < 0 tmp8 = tl.where(tmp7, tmp6, tmp5) tmp9 = tl.load(in_ptr2 + (tmp8 + 16 * tmp4 + 256 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp10 = tl.load(in_ptr3 + (tmp8 + 16 * tmp4 + 256 * x6), None, eviction_policy='evict_last') tmp12 = tmp10 + tmp11 tmp13 = 0.1 tmp14 = tmp12 * tmp13 tmp15 = tl.where(tmp9, tmp12, tmp14) tmp17 = tmp16 + tmp1 tmp18 = tmp16 < 0 tmp19 = tl.where(tmp18, tmp17, tmp16) tmp20 = tl.load(in_ptr2 + (tmp8 + 16 * tmp19 + 256 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp21 = tl.load(in_ptr3 + (tmp8 + 16 * tmp19 + 256 * x6), None, eviction_policy='evict_last') tmp22 = tmp21 + tmp11 tmp23 = tmp22 * tmp13 tmp24 = tl.where(tmp20, tmp22, tmp23) tmp26 = tmp25 + tmp1 tmp27 = tmp25 < 0 tmp28 = tl.where(tmp27, tmp26, tmp25) tmp29 = tl.load(in_ptr2 + (tmp28 + 16 * tmp19 + 256 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp30 = tl.load(in_ptr3 + (tmp28 + 16 * tmp19 + 256 * x6), None, eviction_policy='evict_last') tmp31 = tmp30 + tmp11 tmp32 = tmp31 * tmp13 tmp33 = tl.where(tmp29, tmp31, tmp32) tmp34 = tmp33 - tmp24 tmp36 = tmp34 * tmp35 tmp37 = tmp24 + tmp36 tmp38 = tl.load(in_ptr2 + (tmp28 + 16 * tmp4 + 256 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp39 = tl.load(in_ptr3 + (tmp28 + 16 * tmp4 + 256 * x6), None, eviction_policy='evict_last') tmp40 = tmp39 + tmp11 tmp41 = tmp40 * tmp13 tmp42 = tl.where(tmp38, tmp40, tmp41) tmp43 = tmp42 - tmp15 tmp44 = tmp43 * tmp35 tmp45 = tmp15 + tmp44 tmp46 = tmp45 - tmp37 tmp48 = tmp46 * tmp47 tmp49 = tmp37 + tmp48 tl.store(in_out_ptr1 + x4, tmp49, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_34(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) x3 = xindex x1 = xindex // 1024 % 64 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tl.store(out_ptr0 + x3, tmp4, None) @triton.jit def triton_poi_fused_cat_35(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) x1 = xindex // 1024 % 128 x0 = xindex % 1024 x2 = xindex // 131072 x3 = xindex tmp0 = x1 tl.full([1], 0, tl.int64) tmp3 = tl.full([1], 64, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tl.load(in_ptr0 + (x0 + 1024 * x1 + 65536 * x2), tmp4, other=0.0 ).to(tl.int1) tmp6 = tl.load(in_ptr1 + (x0 + 1024 * x1 + 65536 * x2), tmp4, other=0.0) tmp7 = tl.load(in_ptr2 + x1, tmp4, eviction_policy='evict_last', other=0.0) tmp8 = tmp6 + tmp7 tmp9 = 0.1 tmp10 = tmp8 * tmp9 tmp11 = tl.where(tmp5, tmp8, tmp10) tmp12 = tl.full(tmp11.shape, 0.0, tmp11.dtype) tmp13 = tl.where(tmp4, tmp11, tmp12) tmp14 = tmp0 >= tmp3 tl.full([1], 128, tl.int64) tmp17 = tl.load(in_ptr3 + (x0 + 1024 * (-64 + x1) + 65536 * x2), tmp14, other=0.0) tmp18 = tl.where(tmp4, tmp13, tmp17) tl.store(out_ptr0 + x3, tmp18, None) @triton.jit def triton_poi_fused__to_copy_36(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 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tl.store(out_ptr0 + x0, tmp8, xmask) @triton.jit def triton_poi_fused_add_clamp_37(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 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tmp9 = tl.full([1], 1, tl.int64) tmp10 = tmp8 + tmp9 tmp11 = tl.full([1], 31, tl.int64) tmp12 = triton_helpers.minimum(tmp10, tmp11) tl.store(out_ptr0 + x0, tmp12, xmask) @triton.jit def triton_poi_fused__to_copy_add_arange_clamp_mul_sub_38(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 = x0 tmp1 = tmp0.to(tl.float32) tmp2 = 0.5 tmp3 = tmp1 + tmp2 tmp4 = tmp3 * tmp2 tmp5 = tmp4 - tmp2 tmp6 = 0.0 tmp7 = triton_helpers.maximum(tmp5, tmp6) tmp8 = tmp7.to(tl.int32) tmp9 = tmp8.to(tl.float32) tmp10 = tmp7 - tmp9 tmp11 = triton_helpers.maximum(tmp10, tmp6) tmp12 = 1.0 tmp13 = triton_helpers.minimum(tmp11, tmp12) tl.store(out_ptr0 + x0, tmp13, xmask) @triton.jit def triton_poi_fused__unsafe_index_add_convolution_leaky_relu_mul_sub_39( in_out_ptr1, in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, in_ptr6, in_ptr7, in_ptr8, xnumel, XBLOCK: tl.constexpr): xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] tl.full([XBLOCK], True, tl.int1) x1 = xindex // 64 % 64 x0 = xindex % 64 x6 = xindex // 4096 x2 = xindex // 4096 % 64 x4 = xindex tmp0 = tl.load(in_ptr0 + x1, None, eviction_policy='evict_last') tmp5 = tl.load(in_ptr1 + x0, None, eviction_policy='evict_last') tmp11 = tl.load(in_ptr4 + x2, None, eviction_policy='evict_last') tmp16 = tl.load(in_ptr5 + x1, None, eviction_policy='evict_last') tmp25 = tl.load(in_ptr6 + x0, None, eviction_policy='evict_last') tmp35 = tl.load(in_ptr7 + x0, None, eviction_policy='evict_last') tmp47 = tl.load(in_ptr8 + x1, None, eviction_policy='evict_last') tmp1 = tl.full([XBLOCK], 32, tl.int32) tmp2 = tmp0 + tmp1 tmp3 = tmp0 < 0 tmp4 = tl.where(tmp3, tmp2, tmp0) tmp6 = tmp5 + tmp1 tmp7 = tmp5 < 0 tmp8 = tl.where(tmp7, tmp6, tmp5) tmp9 = tl.load(in_ptr2 + (tmp8 + 32 * tmp4 + 1024 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp10 = tl.load(in_ptr3 + (tmp8 + 32 * tmp4 + 1024 * x6), None, eviction_policy='evict_last') tmp12 = tmp10 + tmp11 tmp13 = 0.1 tmp14 = tmp12 * tmp13 tmp15 = tl.where(tmp9, tmp12, tmp14) tmp17 = tmp16 + tmp1 tmp18 = tmp16 < 0 tmp19 = tl.where(tmp18, tmp17, tmp16) tmp20 = tl.load(in_ptr2 + (tmp8 + 32 * tmp19 + 1024 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp21 = tl.load(in_ptr3 + (tmp8 + 32 * tmp19 + 1024 * x6), None, eviction_policy='evict_last') tmp22 = tmp21 + tmp11 tmp23 = tmp22 * tmp13 tmp24 = tl.where(tmp20, tmp22, tmp23) tmp26 = tmp25 + tmp1 tmp27 = tmp25 < 0 tmp28 = tl.where(tmp27, tmp26, tmp25) tmp29 = tl.load(in_ptr2 + (tmp28 + 32 * tmp19 + 1024 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp30 = tl.load(in_ptr3 + (tmp28 + 32 * tmp19 + 1024 * x6), None, eviction_policy='evict_last') tmp31 = tmp30 + tmp11 tmp32 = tmp31 * tmp13 tmp33 = tl.where(tmp29, tmp31, tmp32) tmp34 = tmp33 - tmp24 tmp36 = tmp34 * tmp35 tmp37 = tmp24 + tmp36 tmp38 = tl.load(in_ptr2 + (tmp28 + 32 * tmp4 + 1024 * x6), None, eviction_policy='evict_last').to(tl.int1) tmp39 = tl.load(in_ptr3 + (tmp28 + 32 * tmp4 + 1024 * x6), None, eviction_policy='evict_last') tmp40 = tmp39 + tmp11 tmp41 = tmp40 * tmp13 tmp42 = tl.where(tmp38, tmp40, tmp41) tmp43 = tmp42 - tmp15 tmp44 = tmp43 * tmp35 tmp45 = tmp15 + tmp44 tmp46 = tmp45 - tmp37 tmp48 = tmp46 * tmp47 tmp49 = tmp37 + tmp48 tl.store(in_out_ptr1 + x4, tmp49, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_40(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) x3 = xindex x1 = xindex // 4096 % 32 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tl.store(out_ptr0 + x3, tmp4, None) @triton.jit def triton_poi_fused_cat_41(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) x1 = xindex // 4096 % 64 x0 = xindex % 4096 x2 = xindex // 262144 x3 = xindex tmp0 = x1 tl.full([1], 0, tl.int64) tmp3 = tl.full([1], 32, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tl.load(in_ptr0 + (x0 + 4096 * x1 + 131072 * x2), tmp4, other=0.0 ).to(tl.int1) tmp6 = tl.load(in_ptr1 + (x0 + 4096 * x1 + 131072 * x2), tmp4, other=0.0) tmp7 = tl.load(in_ptr2 + x1, tmp4, eviction_policy='evict_last', other=0.0) tmp8 = tmp6 + tmp7 tmp9 = 0.1 tmp10 = tmp8 * tmp9 tmp11 = tl.where(tmp5, tmp8, tmp10) tmp12 = tl.full(tmp11.shape, 0.0, tmp11.dtype) tmp13 = tl.where(tmp4, tmp11, tmp12) tmp14 = tmp0 >= tmp3 tl.full([1], 64, tl.int64) tmp17 = tl.load(in_ptr3 + (x0 + 4096 * (-32 + x1) + 131072 * x2), tmp14, other=0.0) tmp18 = tl.where(tmp4, tmp13, tmp17) tl.store(out_ptr0 + x3, tmp18, None) @triton.jit def triton_poi_fused_convolution_leaky_relu_42(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) x3 = xindex x1 = xindex // 4096 % 4 tmp0 = tl.load(in_ptr0 + x3, None) tmp1 = tl.load(in_ptr1 + x1, None, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp3 = 0.0 tmp4 = tmp2 > tmp3 tmp5 = 0.1 tmp6 = tmp2 * tmp5 tmp7 = tl.where(tmp4, tmp2, tmp6) tl.store(out_ptr0 + x3, tmp4, None) tl.store(out_ptr1 + x3, tmp7, 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, primals_32, primals_33, primals_34, primals_35, primals_36, primals_37, primals_38, primals_39, primals_40, primals_41, primals_42, primals_43, primals_44, primals_45, primals_46, primals_47) = args args.clear() assert_size_stride(primals_1, (32, 4, 9, 9), (324, 81, 9, 1)) assert_size_stride(primals_2, (32,), (1,)) assert_size_stride(primals_3, (4, 4, 64, 64), (16384, 4096, 64, 1)) assert_size_stride(primals_4, (32, 32, 7, 7), (1568, 49, 7, 1)) assert_size_stride(primals_5, (32,), (1,)) assert_size_stride(primals_6, (64, 32, 5, 5), (800, 25, 5, 1)) assert_size_stride(primals_7, (64,), (1,)) assert_size_stride(primals_8, (64, 64, 5, 5), (1600, 25, 5, 1)) assert_size_stride(primals_9, (64,), (1,)) assert_size_stride(primals_10, (128, 64, 3, 3), (576, 9, 3, 1)) assert_size_stride(primals_11, (128,), (1,)) assert_size_stride(primals_12, (128, 128, 3, 3), (1152, 9, 3, 1)) assert_size_stride(primals_13, (128,), (1,)) assert_size_stride(primals_14, (256, 128, 3, 3), (1152, 9, 3, 1)) assert_size_stride(primals_15, (256,), (1,)) assert_size_stride(primals_16, (256, 256, 3, 3), (2304, 9, 3, 1)) assert_size_stride(primals_17, (256,), (1,)) assert_size_stride(primals_18, (512, 256, 3, 3), (2304, 9, 3, 1)) assert_size_stride(primals_19, (512,), (1,)) assert_size_stride(primals_20, (512, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_21, (512,), (1,)) assert_size_stride(primals_22, (512, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_23, (512,), (1,)) assert_size_stride(primals_24, (512, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_25, (512,), (1,)) assert_size_stride(primals_26, (512, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_27, (512,), (1,)) assert_size_stride(primals_28, (512, 1024, 3, 3), (9216, 9, 3, 1)) assert_size_stride(primals_29, (512,), (1,)) assert_size_stride(primals_30, (256, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_31, (256,), (1,)) assert_size_stride(primals_32, (256, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_33, (256,), (1,)) assert_size_stride(primals_34, (128, 256, 3, 3), (2304, 9, 3, 1)) assert_size_stride(primals_35, (128,), (1,)) assert_size_stride(primals_36, (128, 256, 3, 3), (2304, 9, 3, 1)) assert_size_stride(primals_37, (128,), (1,)) assert_size_stride(primals_38, (64, 128, 3, 3), (1152, 9, 3, 1)) assert_size_stride(primals_39, (64,), (1,)) assert_size_stride(primals_40, (64, 128, 3, 3), (1152, 9, 3, 1)) assert_size_stride(primals_41, (64,), (1,)) assert_size_stride(primals_42, (32, 64, 3, 3), (576, 9, 3, 1)) assert_size_stride(primals_43, (32,), (1,)) assert_size_stride(primals_44, (32, 64, 3, 3), (576, 9, 3, 1)) assert_size_stride(primals_45, (32,), (1,)) assert_size_stride(primals_46, (4, 32, 3, 3), (288, 9, 3, 1)) assert_size_stride(primals_47, (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=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf0, (4, 32, 64, 64), (131072, 4096, 64, 1)) buf1 = empty_strided_cuda((4, 32, 64, 64), (131072, 4096, 64, 1), torch.bool) buf2 = empty_strided_cuda((4, 32, 64, 64), (131072, 4096, 64, 1), torch.float32) get_raw_stream(0) triton_poi_fused_convolution_leaky_relu_0[grid(524288)](buf0, primals_2, buf1, buf2, 524288, XBLOCK=1024, num_warps=4, num_stages=1) del primals_2 buf3 = extern_kernels.convolution(buf2, primals_4, stride=(1, 1), padding=(3, 3), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf3, (4, 32, 64, 64), (131072, 4096, 64, 1)) buf4 = empty_strided_cuda((4, 32, 64, 64), (131072, 4096, 64, 1), torch.bool) buf5 = buf0 del buf0 triton_poi_fused_convolution_leaky_relu_0[grid(524288)](buf3, primals_5, buf4, buf5, 524288, XBLOCK=1024, num_warps=4, num_stages=1) del primals_5 buf6 = empty_strided_cuda((4, 32, 32, 32), (32768, 1024, 32, 1), torch.float32) triton_poi_fused_avg_pool2d_1[grid(131072)](buf5, buf6, 131072, XBLOCK=512, num_warps=8, num_stages=1) buf7 = extern_kernels.convolution(buf6, primals_6, stride=(1, 1), padding=(2, 2), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf7, (4, 64, 32, 32), (65536, 1024, 32, 1)) buf8 = empty_strided_cuda((4, 64, 32, 32), (65536, 1024, 32, 1), torch.bool) buf9 = empty_strided_cuda((4, 64, 32, 32), (65536, 1024, 32, 1), torch.float32) triton_poi_fused_convolution_leaky_relu_2[grid(262144)](buf7, primals_7, buf8, buf9, 262144, XBLOCK=512, num_warps=8, num_stages=1) del primals_7 buf10 = extern_kernels.convolution(buf9, primals_8, stride=(1, 1), padding=(2, 2), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf10, (4, 64, 32, 32), (65536, 1024, 32, 1)) buf11 = empty_strided_cuda((4, 64, 32, 32), (65536, 1024, 32, 1), torch.bool) buf12 = buf7 del buf7 triton_poi_fused_convolution_leaky_relu_2[grid(262144)](buf10, primals_9, buf11, buf12, 262144, XBLOCK=512, num_warps=8, num_stages=1) del primals_9 buf13 = empty_strided_cuda((4, 64, 16, 16), (16384, 256, 16, 1), torch.float32) triton_poi_fused_avg_pool2d_3[grid(65536)](buf12, buf13, 65536, XBLOCK=256, num_warps=4, num_stages=1) buf14 = extern_kernels.convolution(buf13, primals_10, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf14, (4, 128, 16, 16), (32768, 256, 16, 1)) buf15 = empty_strided_cuda((4, 128, 16, 16), (32768, 256, 16, 1), torch.bool) buf16 = empty_strided_cuda((4, 128, 16, 16), (32768, 256, 16, 1), torch.float32) triton_poi_fused_convolution_leaky_relu_4[grid(131072)](buf14, primals_11, buf15, buf16, 131072, XBLOCK=512, num_warps=8, num_stages=1) del primals_11 buf17 = extern_kernels.convolution(buf16, primals_12, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf17, (4, 128, 16, 16), (32768, 256, 16, 1)) buf18 = empty_strided_cuda((4, 128, 16, 16), (32768, 256, 16, 1), torch.bool) buf19 = buf14 del buf14 triton_poi_fused_convolution_leaky_relu_4[grid(131072)](buf17, primals_13, buf18, buf19, 131072, XBLOCK=512, num_warps=8, num_stages=1) del primals_13 buf20 = empty_strided_cuda((4, 128, 8, 8), (8192, 64, 8, 1), torch. float32) triton_poi_fused_avg_pool2d_5[grid(32768)](buf19, buf20, 32768, XBLOCK=128, num_warps=4, num_stages=1) buf21 = extern_kernels.convolution(buf20, primals_14, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf21, (4, 256, 8, 8), (16384, 64, 8, 1)) buf22 = empty_strided_cuda((4, 256, 8, 8), (16384, 64, 8, 1), torch .bool) buf23 = empty_strided_cuda((4, 256, 8, 8), (16384, 64, 8, 1), torch .float32) triton_poi_fused_convolution_leaky_relu_6[grid(65536)](buf21, primals_15, buf22, buf23, 65536, XBLOCK=256, num_warps=4, num_stages=1) del primals_15 buf24 = extern_kernels.convolution(buf23, primals_16, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf24, (4, 256, 8, 8), (16384, 64, 8, 1)) buf25 = empty_strided_cuda((4, 256, 8, 8), (16384, 64, 8, 1), torch .bool) buf26 = buf21 del buf21 triton_poi_fused_convolution_leaky_relu_6[grid(65536)](buf24, primals_17, buf25, buf26, 65536, XBLOCK=256, num_warps=4, num_stages=1) del primals_17 buf27 = empty_strided_cuda((4, 256, 4, 4), (4096, 16, 4, 1), torch. float32) triton_poi_fused_avg_pool2d_7[grid(16384)](buf26, buf27, 16384, XBLOCK=128, num_warps=4, num_stages=1) buf28 = extern_kernels.convolution(buf27, primals_18, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf28, (4, 512, 4, 4), (8192, 16, 4, 1)) buf29 = empty_strided_cuda((4, 512, 4, 4), (8192, 16, 4, 1), torch.bool ) buf30 = empty_strided_cuda((4, 512, 4, 4), (8192, 16, 4, 1), torch. float32) triton_poi_fused_convolution_leaky_relu_8[grid(32768)](buf28, primals_19, buf29, buf30, 32768, XBLOCK=128, num_warps=4, num_stages=1) del primals_19 buf31 = extern_kernels.convolution(buf30, primals_20, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf31, (4, 512, 4, 4), (8192, 16, 4, 1)) buf32 = empty_strided_cuda((4, 512, 4, 4), (8192, 16, 4, 1), torch.bool ) buf33 = buf28 del buf28 triton_poi_fused_convolution_leaky_relu_8[grid(32768)](buf31, primals_21, buf32, buf33, 32768, XBLOCK=128, num_warps=4, num_stages=1) del primals_21 buf34 = empty_strided_cuda((4, 512, 2, 2), (2048, 4, 2, 1), torch. float32) triton_poi_fused_avg_pool2d_9[grid(8192)](buf33, buf34, 8192, XBLOCK=128, num_warps=4, num_stages=1) buf35 = extern_kernels.convolution(buf34, primals_22, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf35, (4, 512, 2, 2), (2048, 4, 2, 1)) buf36 = empty_strided_cuda((4, 512, 2, 2), (2048, 4, 2, 1), torch.bool) buf37 = empty_strided_cuda((4, 512, 2, 2), (2048, 4, 2, 1), torch. float32) triton_poi_fused_convolution_leaky_relu_10[grid(8192)](buf35, primals_23, buf36, buf37, 8192, XBLOCK=256, num_warps=4, num_stages=1) del buf35 del primals_23 buf38 = extern_kernels.convolution(buf37, primals_24, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf38, (4, 512, 2, 2), (2048, 4, 2, 1)) buf39 = empty_strided_cuda((4, 512, 2, 2), (2048, 4, 2, 1), torch.bool) triton_poi_fused_convolution_leaky_relu_11[grid(8192)](buf38, primals_25, buf39, 8192, XBLOCK=128, num_warps=4, num_stages=1) buf40 = empty_strided_cuda((4, 1), (1, 1), torch.int64) triton_poi_fused__to_copy_12[grid(4)](buf40, 4, XBLOCK=4, num_warps =1, num_stages=1) buf41 = empty_strided_cuda((4, 1), (1, 1), torch.int64) triton_poi_fused_add_clamp_13[grid(4)](buf41, 4, XBLOCK=4, num_warps=1, num_stages=1) buf42 = empty_strided_cuda((4,), (1,), torch.int64) triton_poi_fused__to_copy_12[grid(4)](buf42, 4, XBLOCK=4, num_warps =1, num_stages=1) buf43 = empty_strided_cuda((4,), (1,), torch.int64) triton_poi_fused_add_clamp_13[grid(4)](buf43, 4, XBLOCK=4, num_warps=1, num_stages=1) buf46 = empty_strided_cuda((4,), (1,), torch.float32) triton_poi_fused__to_copy_add_arange_clamp_mul_sub_14[grid(4)](buf46, 4, XBLOCK=4, num_warps=1, num_stages=1) buf48 = empty_strided_cuda((4, 1), (1, 1), torch.float32) triton_poi_fused__to_copy_add_arange_clamp_mul_sub_14[grid(4)](buf48, 4, XBLOCK=4, num_warps=1, num_stages=1) buf45 = buf31 del buf31 buf49 = buf45 del buf45 buf50 = buf49 del buf49 triton_poi_fused__unsafe_index_add_convolution_leaky_relu_mul_sub_15[ grid(32768)](buf50, buf41, buf42, buf39, buf38, primals_25, buf40, buf43, buf46, buf48, 32768, XBLOCK=128, num_warps=4, num_stages=1) del buf38 del primals_25 buf51 = extern_kernels.convolution(buf50, primals_26, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf51, (4, 512, 4, 4), (8192, 16, 4, 1)) buf52 = empty_strided_cuda((4, 512, 4, 4), (8192, 16, 4, 1), torch.bool ) triton_poi_fused_convolution_leaky_relu_16[grid(32768)](buf51, primals_27, buf52, 32768, XBLOCK=128, num_warps=4, num_stages=1) buf53 = reinterpret_tensor(buf24, (4, 1024, 4, 4), (16384, 16, 4, 1), 0 ) del buf24 triton_poi_fused_cat_17[grid(65536)](buf52, buf51, primals_27, buf33, buf53, 65536, XBLOCK=256, num_warps=4, num_stages=1) del buf51 del primals_27 buf54 = extern_kernels.convolution(buf53, primals_28, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf54, (4, 512, 4, 4), (8192, 16, 4, 1)) buf55 = empty_strided_cuda((4, 512, 4, 4), (8192, 16, 4, 1), torch.bool ) triton_poi_fused_convolution_leaky_relu_16[grid(32768)](buf54, primals_29, buf55, 32768, XBLOCK=128, num_warps=4, num_stages=1) buf56 = empty_strided_cuda((8, 1), (1, 1), torch.int64) triton_poi_fused__to_copy_18[grid(8)](buf56, 8, XBLOCK=8, num_warps =1, num_stages=1) buf57 = empty_strided_cuda((8, 1), (1, 1), torch.int64) triton_poi_fused_add_clamp_19[grid(8)](buf57, 8, XBLOCK=8, num_warps=1, num_stages=1) buf58 = empty_strided_cuda((8,), (1,), torch.int64) triton_poi_fused__to_copy_18[grid(8)](buf58, 8, XBLOCK=8, num_warps =1, num_stages=1) buf59 = empty_strided_cuda((8,), (1,), torch.int64) triton_poi_fused_add_clamp_19[grid(8)](buf59, 8, XBLOCK=8, num_warps=1, num_stages=1) buf62 = empty_strided_cuda((8,), (1,), torch.float32) triton_poi_fused__to_copy_add_arange_clamp_mul_sub_20[grid(8)](buf62, 8, XBLOCK=8, num_warps=1, num_stages=1) buf64 = empty_strided_cuda((8, 1), (1, 1), torch.float32) triton_poi_fused__to_copy_add_arange_clamp_mul_sub_20[grid(8)](buf64, 8, XBLOCK=8, num_warps=1, num_stages=1) buf61 = reinterpret_tensor(buf17, (4, 512, 8, 8), (32768, 64, 8, 1), 0) del buf17 buf65 = buf61 del buf61 buf66 = buf65 del buf65 triton_poi_fused__unsafe_index_add_convolution_leaky_relu_mul_sub_21[ grid(131072)](buf66, buf57, buf58, buf55, buf54, primals_29, buf56, buf59, buf62, buf64, 131072, XBLOCK=512, num_warps=8, num_stages=1) del buf54 del primals_29 buf67 = extern_kernels.convolution(buf66, primals_30, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf67, (4, 256, 8, 8), (16384, 64, 8, 1)) buf68 = empty_strided_cuda((4, 256, 8, 8), (16384, 64, 8, 1), torch .bool) triton_poi_fused_convolution_leaky_relu_22[grid(65536)](buf67, primals_31, buf68, 65536, XBLOCK=512, num_warps=4, num_stages=1) buf69 = empty_strided_cuda((4, 512, 8, 8), (32768, 64, 8, 1), torch .float32) triton_poi_fused_cat_23[grid(131072)](buf68, buf67, primals_31, buf26, buf69, 131072, XBLOCK=512, num_warps=8, num_stages=1) del buf67 del primals_31 buf70 = extern_kernels.convolution(buf69, primals_32, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf70, (4, 256, 8, 8), (16384, 64, 8, 1)) buf71 = empty_strided_cuda((4, 256, 8, 8), (16384, 64, 8, 1), torch .bool) triton_poi_fused_convolution_leaky_relu_22[grid(65536)](buf70, primals_33, buf71, 65536, XBLOCK=512, num_warps=4, num_stages=1) buf72 = empty_strided_cuda((16, 1), (1, 1), torch.int64) triton_poi_fused__to_copy_24[grid(16)](buf72, 16, XBLOCK=16, num_warps=1, num_stages=1) buf73 = empty_strided_cuda((16, 1), (1, 1), torch.int64) triton_poi_fused_add_clamp_25[grid(16)](buf73, 16, XBLOCK=16, num_warps=1, num_stages=1) buf74 = empty_strided_cuda((16,), (1,), torch.int64) triton_poi_fused__to_copy_24[grid(16)](buf74, 16, XBLOCK=16, num_warps=1, num_stages=1) buf75 = empty_strided_cuda((16,), (1,), torch.int64) triton_poi_fused_add_clamp_25[grid(16)](buf75, 16, XBLOCK=16, num_warps=1, num_stages=1) buf78 = empty_strided_cuda((16,), (1,), torch.float32) triton_poi_fused__to_copy_add_arange_clamp_mul_sub_26[grid(16)](buf78, 16, XBLOCK=16, num_warps=1, num_stages=1) buf80 = empty_strided_cuda((16, 1), (1, 1), torch.float32) triton_poi_fused__to_copy_add_arange_clamp_mul_sub_26[grid(16)](buf80, 16, XBLOCK=16, num_warps=1, num_stages=1) buf77 = reinterpret_tensor(buf10, (4, 256, 16, 16), (65536, 256, 16, 1), 0) del buf10 buf81 = buf77 del buf77 buf82 = buf81 del buf81 triton_poi_fused__unsafe_index_add_convolution_leaky_relu_mul_sub_27[ grid(262144)](buf82, buf73, buf74, buf71, buf70, primals_33, buf72, buf75, buf78, buf80, 262144, XBLOCK=512, num_warps=8, num_stages=1) del primals_33 buf83 = extern_kernels.convolution(buf82, primals_34, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf83, (4, 128, 16, 16), (32768, 256, 16, 1)) buf84 = empty_strided_cuda((4, 128, 16, 16), (32768, 256, 16, 1), torch.bool) triton_poi_fused_convolution_leaky_relu_28[grid(131072)](buf83, primals_35, buf84, 131072, XBLOCK=1024, num_warps=4, num_stages=1) buf85 = empty_strided_cuda((4, 256, 16, 16), (65536, 256, 16, 1), torch.float32) triton_poi_fused_cat_29[grid(262144)](buf84, buf83, primals_35, buf19, buf85, 262144, XBLOCK=512, num_warps=8, num_stages=1) del buf83 del primals_35 buf86 = extern_kernels.convolution(buf85, primals_36, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf86, (4, 128, 16, 16), (32768, 256, 16, 1)) buf87 = empty_strided_cuda((4, 128, 16, 16), (32768, 256, 16, 1), torch.bool) triton_poi_fused_convolution_leaky_relu_28[grid(131072)](buf86, primals_37, buf87, 131072, XBLOCK=1024, num_warps=4, num_stages=1) buf88 = empty_strided_cuda((32, 1), (1, 1), torch.int64) triton_poi_fused__to_copy_30[grid(32)](buf88, 32, XBLOCK=32, num_warps=1, num_stages=1) buf89 = empty_strided_cuda((32, 1), (1, 1), torch.int64) triton_poi_fused_add_clamp_31[grid(32)](buf89, 32, XBLOCK=32, num_warps=1, num_stages=1) buf90 = empty_strided_cuda((32,), (1,), torch.int64) triton_poi_fused__to_copy_30[grid(32)](buf90, 32, XBLOCK=32, num_warps=1, num_stages=1) buf91 = empty_strided_cuda((32,), (1,), torch.int64) triton_poi_fused_add_clamp_31[grid(32)](buf91, 32, XBLOCK=32, num_warps=1, num_stages=1) buf94 = empty_strided_cuda((32,), (1,), torch.float32) triton_poi_fused__to_copy_add_arange_clamp_mul_sub_32[grid(32)](buf94, 32, XBLOCK=32, num_warps=1, num_stages=1) buf96 = empty_strided_cuda((32, 1), (1, 1), torch.float32) triton_poi_fused__to_copy_add_arange_clamp_mul_sub_32[grid(32)](buf96, 32, XBLOCK=32, num_warps=1, num_stages=1) buf93 = reinterpret_tensor(buf3, (4, 128, 32, 32), (131072, 1024, 32, 1), 0) del buf3 buf97 = buf93 del buf93 buf98 = buf97 del buf97 triton_poi_fused__unsafe_index_add_convolution_leaky_relu_mul_sub_33[ grid(524288)](buf98, buf89, buf90, buf87, buf86, primals_37, buf88, buf91, buf94, buf96, 524288, XBLOCK=512, num_warps=8, num_stages=1) del buf86 del primals_37 buf99 = extern_kernels.convolution(buf98, primals_38, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf99, (4, 64, 32, 32), (65536, 1024, 32, 1)) buf100 = empty_strided_cuda((4, 64, 32, 32), (65536, 1024, 32, 1), torch.bool) triton_poi_fused_convolution_leaky_relu_34[grid(262144)](buf99, primals_39, buf100, 262144, XBLOCK=1024, num_warps=4, num_stages=1) buf101 = empty_strided_cuda((4, 128, 32, 32), (131072, 1024, 32, 1), torch.float32) triton_poi_fused_cat_35[grid(524288)](buf100, buf99, primals_39, buf12, buf101, 524288, XBLOCK=512, num_warps=8, num_stages=1) del buf99 del primals_39 buf102 = extern_kernels.convolution(buf101, primals_40, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf102, (4, 64, 32, 32), (65536, 1024, 32, 1)) buf103 = empty_strided_cuda((4, 64, 32, 32), (65536, 1024, 32, 1), torch.bool) triton_poi_fused_convolution_leaky_relu_34[grid(262144)](buf102, primals_41, buf103, 262144, XBLOCK=1024, num_warps=4, num_stages=1) buf104 = empty_strided_cuda((64, 1), (1, 1), torch.int64) triton_poi_fused__to_copy_36[grid(64)](buf104, 64, XBLOCK=64, num_warps=1, num_stages=1) buf105 = empty_strided_cuda((64, 1), (1, 1), torch.int64) triton_poi_fused_add_clamp_37[grid(64)](buf105, 64, XBLOCK=64, num_warps=1, num_stages=1) buf106 = empty_strided_cuda((64,), (1,), torch.int64) triton_poi_fused__to_copy_36[grid(64)](buf106, 64, XBLOCK=64, num_warps=1, num_stages=1) buf107 = empty_strided_cuda((64,), (1,), torch.int64) triton_poi_fused_add_clamp_37[grid(64)](buf107, 64, XBLOCK=64, num_warps=1, num_stages=1) buf110 = empty_strided_cuda((64,), (1,), torch.float32) triton_poi_fused__to_copy_add_arange_clamp_mul_sub_38[grid(64)](buf110, 64, XBLOCK=64, num_warps=1, num_stages=1) buf112 = empty_strided_cuda((64, 1), (1, 1), torch.float32) triton_poi_fused__to_copy_add_arange_clamp_mul_sub_38[grid(64)](buf112, 64, XBLOCK=64, num_warps=1, num_stages=1) buf109 = empty_strided_cuda((4, 64, 64, 64), (262144, 4096, 64, 1), torch.float32) buf113 = buf109 del buf109 buf114 = buf113 del buf113 triton_poi_fused__unsafe_index_add_convolution_leaky_relu_mul_sub_39[ grid(1048576)](buf114, buf105, buf106, buf103, buf102, primals_41, buf104, buf107, buf110, buf112, 1048576, XBLOCK= 1024, num_warps=4, num_stages=1) del buf102 del primals_41 buf115 = extern_kernels.convolution(buf114, primals_42, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf115, (4, 32, 64, 64), (131072, 4096, 64, 1)) buf116 = empty_strided_cuda((4, 32, 64, 64), (131072, 4096, 64, 1), torch.bool) triton_poi_fused_convolution_leaky_relu_40[grid(524288)](buf115, primals_43, buf116, 524288, XBLOCK=1024, num_warps=4, num_stages=1) buf117 = empty_strided_cuda((4, 64, 64, 64), (262144, 4096, 64, 1), torch.float32) triton_poi_fused_cat_41[grid(1048576)](buf116, buf115, primals_43, buf5, buf117, 1048576, XBLOCK=512, num_warps=8, num_stages=1) del primals_43 buf118 = extern_kernels.convolution(buf117, primals_44, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf118, (4, 32, 64, 64), (131072, 4096, 64, 1)) buf119 = empty_strided_cuda((4, 32, 64, 64), (131072, 4096, 64, 1), torch.bool) buf120 = buf115 del buf115 triton_poi_fused_convolution_leaky_relu_0[grid(524288)](buf118, primals_45, buf119, buf120, 524288, XBLOCK=1024, num_warps=4, num_stages=1) del buf118 del primals_45 buf121 = extern_kernels.convolution(buf120, primals_46, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf121, (4, 4, 64, 64), (16384, 4096, 64, 1)) buf122 = empty_strided_cuda((4, 4, 64, 64), (16384, 4096, 64, 1), torch.bool) buf123 = reinterpret_tensor(buf70, (4, 4, 64, 64), (16384, 4096, 64, 1), 0) del buf70 triton_poi_fused_convolution_leaky_relu_42[grid(65536)](buf121, primals_47, buf122, buf123, 65536, XBLOCK=512, num_warps=4, num_stages=1) del buf121 del primals_47 return (buf123, 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, primals_32, primals_34, primals_36, primals_38, primals_40, primals_42, primals_44, primals_46, buf1, buf2, buf4, buf5, buf6, buf8, buf9, buf11, buf12, buf13, buf15, buf16, buf18, buf19, buf20, buf22, buf23, buf25, buf26, buf27, buf29, buf30, buf32, buf33, buf34, buf36, buf37, buf39, buf40, buf41, buf42, buf43, buf46, buf48, buf50, buf52, buf53, buf55, buf56, buf57, buf58, buf59, buf62, buf64, buf66, buf68, buf69, buf71, buf72, buf73, buf74, buf75, buf78, buf80, buf82, buf84, buf85, buf87, buf88, buf89, buf90, buf91, buf94, buf96, buf98, buf100, buf101, buf103, buf104, buf105, buf106, buf107, buf110, buf112, buf114, buf116, buf117, buf119, buf120, buf122) class down(nn.Module): """ A class for creating neural network blocks containing layers: Average Pooling --> Convlution + Leaky ReLU --> Convolution + Leaky ReLU This is used in the UNet Class to create a UNet like NN architecture. ... Methods ------- forward(x) Returns output tensor after passing input `x` to the neural network block. """ def __init__(self, inChannels, outChannels, filterSize): """ Parameters ---------- inChannels : int number of input channels for the first convolutional layer. outChannels : int number of output channels for the first convolutional layer. This is also used as input and output channels for the second convolutional layer. filterSize : int filter size for the convolution filter. input N would create a N x N filter. """ super(down, self).__init__() self.conv1 = nn.Conv2d(inChannels, outChannels, filterSize, stride= 1, padding=int((filterSize - 1) / 2)) self.conv2 = nn.Conv2d(outChannels, outChannels, filterSize, stride =1, padding=int((filterSize - 1) / 2)) def forward(self, x): """ Returns output tensor after passing input `x` to the neural network block. Parameters ---------- x : tensor input to the NN block. Returns ------- tensor output of the NN block. """ x = F.avg_pool2d(x, 2) x = F.leaky_relu(self.conv1(x), negative_slope=0.1) x = F.leaky_relu(self.conv2(x), negative_slope=0.1) return x class up(nn.Module): """ A class for creating neural network blocks containing layers: Bilinear interpolation --> Convlution + Leaky ReLU --> Convolution + Leaky ReLU This is used in the UNet Class to create a UNet like NN architecture. ... Methods ------- forward(x, skpCn) Returns output tensor after passing input `x` to the neural network block. """ def __init__(self, inChannels, outChannels): """ Parameters ---------- inChannels : int number of input channels for the first convolutional layer. outChannels : int number of output channels for the first convolutional layer. This is also used for setting input and output channels for the second convolutional layer. """ super(up, self).__init__() self.conv1 = nn.Conv2d(inChannels, outChannels, 3, stride=1, padding=1) self.conv2 = nn.Conv2d(2 * outChannels, outChannels, 3, stride=1, padding=1) def forward(self, x, skpCn): """ Returns output tensor after passing input `x` to the neural network block. Parameters ---------- x : tensor input to the NN block. skpCn : tensor skip connection input to the NN block. Returns ------- tensor output of the NN block. """ x = F.interpolate(x, scale_factor=2, mode='bilinear') x = F.leaky_relu(self.conv1(x), negative_slope=0.1) x = F.leaky_relu(self.conv2(torch.cat((x, skpCn), 1)), negative_slope=0.1) return x class UNetNew(nn.Module): """ A class for creating UNet architecture with skip connections. ... Methods ------- forward(x) Returns output tensor after passing input `x` to the neural network block. """ def __init__(self, inChannels, outChannels): """ Parameters ---------- inChannels : int number of input channels for the UNet. outChannels : int number of output channels for the UNet. """ super(UNetNew, self).__init__() self.conv1 = nn.Conv2d(inChannels, 32, 9, stride=1, padding=4) self.conv2 = nn.Conv2d(32, 32, 7, stride=1, padding=3) self.down1 = down(32, 64, 5) self.down2 = down(64, 128, 3) self.down3 = down(128, 256, 3) self.down4 = down(256, 512, 3) self.down5 = down(512, 512, 3) self.up1 = up(512, 512) self.up2 = up(512, 256) self.up3 = up(256, 128) self.up4 = up(128, 64) self.up5 = up(64, 32) self.conv3 = nn.Conv2d(32, outChannels, 3, stride=1, padding=1) 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.down1.conv1.weight primals_7 = self.down1.conv1.bias primals_8 = self.down1.conv2.weight primals_9 = self.down1.conv2.bias primals_10 = self.down2.conv1.weight primals_11 = self.down2.conv1.bias primals_12 = self.down2.conv2.weight primals_13 = self.down2.conv2.bias primals_14 = self.down3.conv1.weight primals_15 = self.down3.conv1.bias primals_16 = self.down3.conv2.weight primals_17 = self.down3.conv2.bias primals_18 = self.down4.conv1.weight primals_19 = self.down4.conv1.bias primals_20 = self.down4.conv2.weight primals_21 = self.down4.conv2.bias primals_22 = self.down5.conv1.weight primals_23 = self.down5.conv1.bias primals_24 = self.down5.conv2.weight primals_25 = self.down5.conv2.bias primals_26 = self.up1.conv1.weight primals_27 = self.up1.conv1.bias primals_28 = self.up1.conv2.weight primals_29 = self.up1.conv2.bias primals_30 = self.up2.conv1.weight primals_31 = self.up2.conv1.bias primals_32 = self.up2.conv2.weight primals_33 = self.up2.conv2.bias primals_34 = self.up3.conv1.weight primals_35 = self.up3.conv1.bias primals_36 = self.up3.conv2.weight primals_37 = self.up3.conv2.bias primals_38 = self.up4.conv1.weight primals_39 = self.up4.conv1.bias primals_40 = self.up4.conv2.weight primals_41 = self.up4.conv2.bias primals_42 = self.up5.conv1.weight primals_43 = self.up5.conv1.bias primals_44 = self.up5.conv2.weight primals_45 = self.up5.conv2.bias primals_46 = self.conv3.weight primals_47 = self.conv3.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, primals_32, primals_33, primals_34, primals_35, primals_36, primals_37, primals_38, primals_39, primals_40, primals_41, primals_42, primals_43, primals_44, primals_45, primals_46, primals_47]) return output[0]
avinashpaliwal/Deep-SloMo
UNet
false
15,057
[ "MIT" ]
76
93373aa3cb9fd384fbf905e235fe6eb4f9cac780
https://github.com/avinashpaliwal/Deep-SloMo/tree/93373aa3cb9fd384fbf905e235fe6eb4f9cac780
NormMLP
import torch import torch.nn as nn import torch.nn.functional as F class NormMLP(nn.Module): def __init__(self, input_size, output_size): super(NormMLP, self).__init__() self.linear = nn.Linear(input_size, output_size) self.layer_norm = nn.LayerNorm(output_size) def forward(self, activations): return self.layer_norm(self.linear(F.relu(activations))) 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 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_poi_fused_relu_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.full([1], 0, tl.int32) tmp2 = triton_helpers.maximum(tmp1, tmp0) tl.store(out_ptr0 + x0, tmp2, 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-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_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 = 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, (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, 4), (64, 16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_relu_0[grid(256)](primals_1, buf0, 256, XBLOCK=128, num_warps=4, num_stages=1) del primals_1 buf1 = empty_strided_cuda((64, 4), (4, 1), torch.float32) extern_kernels.addmm(primals_3, reinterpret_tensor(buf0, (64, 4), ( 4, 1), 0), reinterpret_tensor(primals_2, (4, 4), (1, 4), 0), alpha=1, beta=1, out=buf1) del primals_2 del primals_3 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_4, primals_5, buf4, 256, XBLOCK=256, num_warps=4, num_stages=1) del buf2 del buf3 del primals_5 return buf4, primals_4, reinterpret_tensor(buf0, (64, 4), (4, 1), 0), buf1 class NormMLPNew(nn.Module): def __init__(self, input_size, output_size): super(NormMLPNew, self).__init__() self.linear = nn.Linear(input_size, output_size) self.layer_norm = nn.LayerNorm(output_size) def forward(self, input_0): primals_2 = self.linear.weight primals_3 = self.linear.bias primals_4 = self.layer_norm.weight primals_5 = self.layer_norm.bias primals_1 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5]) return output[0]
cogentlabs/apl
NormMLP
false
15,058
[ "MIT" ]
50
78092b162e019a2df0ab5ea31d4db0b9860090d3
https://github.com/cogentlabs/apl/tree/78092b162e019a2df0ab5ea31d4db0b9860090d3
SoftTargetCrossEntropyLoss
import torch def _convert_to_one_hot(targets: 'torch.Tensor', classes: 'int' ) ->torch.Tensor: """ This function converts target class indices to one-hot vectors, given the number of classes. """ if torch.max(targets).item() >= classes: raise ValueError('Class Index must be less than number of classes') one_hot_targets = torch.zeros((targets.shape[0], classes), dtype=torch. long, device=targets.device) one_hot_targets.scatter_(1, targets.long(), 1) return one_hot_targets class SoftTargetCrossEntropyLoss(torch.nn.CrossEntropyLoss): """This loss allows the targets for the cross entropy loss to be multi-label. Args: reduction (str): specifies reduction to apply to the output. normalize_targets (bool): whether the targets should be normalized to a sum of 1 based on the total count of positive targets for a given sample. """ def __init__(self, reduction: 'str'='mean', normalize_targets: 'bool'=True ) ->None: super().__init__(reduction=reduction) self.normalize_targets = normalize_targets self._eps: 'float' = torch.finfo(torch.float32).eps def forward(self, input: 'torch.Tensor', target: 'torch.Tensor' ) ->torch.Tensor: target = target.detach().clone() if target.ndim == 1: if input.shape[0] != target.shape[0]: raise ValueError( 'SoftTargetCrossEntropyLoss requires input and target to have same batch size!' ) target = _convert_to_one_hot(target.view(-1, 1), input.shape[1]) target = target.float() if self.normalize_targets: target /= self._eps + target.sum(dim=1, keepdim=True) return super().forward(input, target) 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 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__log_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 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 tl.store(out_ptr0 + x3, tmp8, xmask) @triton.jit def triton_per_fused__log_softmax_add_div_mul_neg_sum_1(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) r3 = rindex r0 = rindex % 16 r2 = rindex // 64 tmp0 = tl.load(in_ptr0 + r3, None) tmp1 = tl.load(in_ptr0 + (r0 + 64 * r2), None, eviction_policy='evict_last' ) tmp3 = tl.load(in_ptr0 + (16 + r0 + 64 * r2), None, eviction_policy= 'evict_last') tmp6 = tl.load(in_ptr0 + (32 + r0 + 64 * r2), None, eviction_policy= 'evict_last') tmp9 = tl.load(in_ptr0 + (48 + r0 + 64 * r2), None, eviction_policy= 'evict_last') tmp14 = tl.load(in_ptr1 + r3, None) tmp15 = tl.load(in_ptr1 + (r0 + 64 * r2), None, eviction_policy= 'evict_last') tmp16 = tl.load(in_ptr1 + (16 + r0 + 64 * r2), None, eviction_policy= 'evict_last') tmp18 = tl.load(in_ptr1 + (32 + r0 + 64 * r2), None, eviction_policy= 'evict_last') tmp20 = tl.load(in_ptr1 + (48 + r0 + 64 * r2), None, eviction_policy= 'evict_last') tmp2 = tl_math.exp(tmp1) tmp4 = tl_math.exp(tmp3) tmp5 = tmp2 + tmp4 tmp7 = tl_math.exp(tmp6) tmp8 = tmp5 + tmp7 tmp10 = tl_math.exp(tmp9) tmp11 = tmp8 + tmp10 tmp12 = tl_math.log(tmp11) tmp13 = tmp0 - tmp12 tmp17 = tmp15 + tmp16 tmp19 = tmp17 + tmp18 tmp21 = tmp19 + tmp20 tmp22 = 1.1920928955078125e-07 tmp23 = tmp21 + tmp22 tmp24 = tmp14 / tmp23 tmp25 = tmp13 * tmp24 tmp26 = tl.broadcast_to(tmp25, [RBLOCK]) tmp28 = triton_helpers.promote_to_tensor(tl.sum(tmp26, 0)) tmp29 = -tmp28 tmp30 = 0.015625 tmp31 = tmp29 * tmp30 tl.debug_barrier() tl.store(in_out_ptr0 + tl.full([1], 0, tl.int32), tmp31, 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, 4, 4), (64, 16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused__log_softmax_0[grid(256)](arg1_1, buf0, 256, XBLOCK=128, num_warps=4, num_stages=1) del arg1_1 buf2 = empty_strided_cuda((), (), torch.float32) buf3 = buf2 del buf2 triton_per_fused__log_softmax_add_div_mul_neg_sum_1[grid(1)](buf3, buf0, arg0_1, 1, 256, num_warps=2, num_stages=1) del arg0_1 del buf0 return buf3, def _convert_to_one_hot(targets: 'torch.Tensor', classes: 'int' ) ->torch.Tensor: """ This function converts target class indices to one-hot vectors, given the number of classes. """ if torch.max(targets).item() >= classes: raise ValueError('Class Index must be less than number of classes') one_hot_targets = torch.zeros((targets.shape[0], classes), dtype=torch. long, device=targets.device) one_hot_targets.scatter_(1, targets.long(), 1) return one_hot_targets class SoftTargetCrossEntropyLossNew(torch.nn.CrossEntropyLoss): """This loss allows the targets for the cross entropy loss to be multi-label. Args: reduction (str): specifies reduction to apply to the output. normalize_targets (bool): whether the targets should be normalized to a sum of 1 based on the total count of positive targets for a given sample. """ def __init__(self, reduction: 'str'='mean', normalize_targets: 'bool'=True ) ->None: super().__init__(reduction=reduction) self.normalize_targets = normalize_targets self._eps: 'float' = torch.finfo(torch.float32).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]
colin2328/recipes
SoftTargetCrossEntropyLoss
false
15,059
[ "BSD-3-Clause" ]
161
a6cd0e12c9fcb48749721a6548d0a02319d54bd1
https://github.com/colin2328/recipes/tree/a6cd0e12c9fcb48749721a6548d0a02319d54bd1
LeakyReLU
import torch import numpy as np import torch.nn as nn from numbers import Number def keep_variance_fn(x): return x + 0.001 def normcdf(value, mu=0.0, stddev=1.0): sinv = 1.0 / stddev if isinstance(stddev, Number) else stddev.reciprocal() return 0.5 * (1.0 + torch.erf((value - mu) * sinv / np.sqrt(2.0))) def _normal_log_pdf(value, mu, stddev): var = stddev ** 2 log_scale = np.log(stddev) if isinstance(stddev, Number) else torch.log( stddev) return -(value - mu) ** 2 / (2.0 * var) - log_scale - np.log(np.sqrt( 2.0 * np.pi)) def normpdf(value, mu=0.0, stddev=1.0): return torch.exp(_normal_log_pdf(value, mu, stddev)) class LeakyReLU(nn.Module): def __init__(self, negative_slope=0.01, keep_variance_fn=None): super(LeakyReLU, self).__init__() self._keep_variance_fn = keep_variance_fn self._negative_slope = negative_slope def forward(self, features_mean, features_variance): features_stddev = torch.sqrt(features_variance) div = features_mean / features_stddev pdf = normpdf(div) cdf = normcdf(div) negative_cdf = 1.0 - cdf mu_cdf = features_mean * cdf stddev_pdf = features_stddev * pdf squared_mean_variance = features_mean ** 2 + features_variance mean_stddev_pdf = features_mean * stddev_pdf mean_r = mu_cdf + stddev_pdf variance_r = (squared_mean_variance * cdf + mean_stddev_pdf - mean_r ** 2) mean_n = -features_mean * negative_cdf + stddev_pdf variance_n = (squared_mean_variance * negative_cdf - mean_stddev_pdf - mean_n ** 2) covxy = -mean_r * mean_n outputs_mean = mean_r - self._negative_slope * mean_n outputs_variance = (variance_r + self._negative_slope * self. _negative_slope * variance_n - 2.0 * self._negative_slope * covxy) if self._keep_variance_fn is not None: outputs_variance = self._keep_variance_fn(outputs_variance) return outputs_mean, outputs_variance 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.triton_helpers import libdevice, math as tl_math import numpy as np import torch.nn as nn from numbers import Number 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_div_erf_exp_log_mul_neg_pow_rsub_sqrt_sub_0( in_out_ptr0, 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 tmp0 = tl.load(in_ptr0 + x0, xmask) tmp1 = tl.load(in_ptr1 + x0, xmask) tmp2 = libdevice.sqrt(tmp1) tmp3 = tmp0 / tmp2 tmp4 = 0.0 tmp5 = tmp3 - tmp4 tmp6 = 1.0 tmp7 = tmp5 * tmp6 tmp8 = 1.414213562373095 tmp9 = tmp7 / tmp8 tmp10 = libdevice.erf(tmp9) tmp11 = tmp10 + tmp6 tmp12 = 0.5 tmp13 = tmp11 * tmp12 tmp14 = tmp0 * tmp13 tmp15 = tmp5 * tmp5 tmp16 = -tmp15 tmp17 = tmp16 * tmp12 tmp18 = 0.9189385332046727 tmp19 = tmp17 - tmp18 tmp20 = tl_math.exp(tmp19) tmp21 = tmp2 * tmp20 tmp22 = tmp14 + tmp21 tmp23 = -tmp0 tmp24 = tmp6 - tmp13 tmp25 = tmp23 * tmp24 tmp26 = tmp25 + tmp21 tmp27 = 0.01 tmp28 = tmp26 * tmp27 tmp29 = tmp22 - tmp28 tmp30 = tmp0 * tmp0 tmp31 = tmp30 + tmp1 tmp32 = tmp31 * tmp13 tmp33 = tmp0 * tmp21 tmp34 = tmp32 + tmp33 tmp35 = tmp22 * tmp22 tmp36 = tmp34 - tmp35 tmp37 = tmp31 * tmp24 tmp38 = tmp37 - tmp33 tmp39 = tmp26 * tmp26 tmp40 = tmp38 - tmp39 tmp41 = -tmp22 tmp42 = tmp41 * tmp26 tmp43 = 0.0001 tmp44 = tmp40 * tmp43 tmp45 = tmp36 + tmp44 tmp46 = 0.02 tmp47 = tmp42 * tmp46 tmp48 = tmp45 - tmp47 tl.store(out_ptr0 + x0, tmp29, xmask) tl.store(in_out_ptr0 + x0, tmp48, xmask) 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, 4, 4), (64, 16, 4, 1), torch.float32) buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) buf4 = buf1 del buf1 get_raw_stream(0) triton_poi_fused_add_div_erf_exp_log_mul_neg_pow_rsub_sqrt_sub_0[grid (256)](buf4, arg1_1, arg0_1, buf0, 256, XBLOCK=128, num_warps=4, num_stages=1) del arg0_1 del arg1_1 return buf0, buf4 def keep_variance_fn(x): return x + 0.001 def normcdf(value, mu=0.0, stddev=1.0): sinv = 1.0 / stddev if isinstance(stddev, Number) else stddev.reciprocal() return 0.5 * (1.0 + torch.erf((value - mu) * sinv / np.sqrt(2.0))) def _normal_log_pdf(value, mu, stddev): var = stddev ** 2 log_scale = np.log(stddev) if isinstance(stddev, Number) else torch.log( stddev) return -(value - mu) ** 2 / (2.0 * var) - log_scale - np.log(np.sqrt( 2.0 * np.pi)) def normpdf(value, mu=0.0, stddev=1.0): return torch.exp(_normal_log_pdf(value, mu, stddev)) class LeakyReLUNew(nn.Module): def __init__(self, negative_slope=0.01, keep_variance_fn=None): super(LeakyReLUNew, self).__init__() self._keep_variance_fn = keep_variance_fn self._negative_slope = negative_slope def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0], output[1]
collector-m/LiDAR-MOS
LeakyReLU
false
15,060
[ "MIT" ]
268
7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
https://github.com/collector-m/LiDAR-MOS/tree/7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
ReLU
import torch import numpy as np import torch.nn as nn from numbers import Number def keep_variance_fn(x): return x + 0.001 def normcdf(value, mu=0.0, stddev=1.0): sinv = 1.0 / stddev if isinstance(stddev, Number) else stddev.reciprocal() return 0.5 * (1.0 + torch.erf((value - mu) * sinv / np.sqrt(2.0))) def _normal_log_pdf(value, mu, stddev): var = stddev ** 2 log_scale = np.log(stddev) if isinstance(stddev, Number) else torch.log( stddev) return -(value - mu) ** 2 / (2.0 * var) - log_scale - np.log(np.sqrt( 2.0 * np.pi)) def normpdf(value, mu=0.0, stddev=1.0): return torch.exp(_normal_log_pdf(value, mu, stddev)) class ReLU(nn.Module): def __init__(self, keep_variance_fn=None): super(ReLU, self).__init__() self._keep_variance_fn = keep_variance_fn def forward(self, features_mean, features_variance): features_stddev = torch.sqrt(features_variance) div = features_mean / features_stddev pdf = normpdf(div) cdf = normcdf(div) outputs_mean = features_mean * cdf + features_stddev * pdf outputs_variance = (features_mean ** 2 + features_variance ) * cdf + features_mean * features_stddev * pdf - outputs_mean ** 2 if self._keep_variance_fn is not None: outputs_variance = self._keep_variance_fn(outputs_variance) return outputs_mean, outputs_variance 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.triton_helpers import libdevice, math as tl_math import numpy as np import torch.nn as nn from numbers import Number 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_div_erf_exp_log_mul_neg_pow_sqrt_sub_0(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 x0 = xindex tmp0 = tl.load(in_ptr0 + x0, xmask) tmp1 = tl.load(in_ptr1 + x0, xmask) tmp2 = libdevice.sqrt(tmp1) tmp3 = tmp0 / tmp2 tmp4 = 0.0 tmp5 = tmp3 - tmp4 tmp6 = 1.0 tmp7 = tmp5 * tmp6 tmp8 = 1.414213562373095 tmp9 = tmp7 / tmp8 tmp10 = libdevice.erf(tmp9) tmp11 = tmp10 + tmp6 tmp12 = 0.5 tmp13 = tmp11 * tmp12 tmp14 = tmp0 * tmp13 tmp15 = tmp5 * tmp5 tmp16 = -tmp15 tmp17 = tmp16 * tmp12 tmp18 = 0.9189385332046727 tmp19 = tmp17 - tmp18 tmp20 = tl_math.exp(tmp19) tmp21 = tmp2 * tmp20 tmp22 = tmp14 + tmp21 tmp23 = tmp0 * tmp0 tmp24 = tmp23 + tmp1 tmp25 = tmp24 * tmp13 tmp26 = tmp0 * tmp2 tmp27 = tmp26 * tmp20 tmp28 = tmp25 + tmp27 tmp29 = tmp22 * tmp22 tmp30 = tmp28 - tmp29 tl.store(out_ptr0 + x0, tmp22, xmask) tl.store(out_ptr1 + x0, tmp30, xmask) 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, 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_div_erf_exp_log_mul_neg_pow_sqrt_sub_0[grid(256)]( arg1_1, arg0_1, buf0, buf1, 256, XBLOCK=256, num_warps=4, num_stages=1) del arg0_1 del arg1_1 return buf0, buf1 def keep_variance_fn(x): return x + 0.001 def normcdf(value, mu=0.0, stddev=1.0): sinv = 1.0 / stddev if isinstance(stddev, Number) else stddev.reciprocal() return 0.5 * (1.0 + torch.erf((value - mu) * sinv / np.sqrt(2.0))) def _normal_log_pdf(value, mu, stddev): var = stddev ** 2 log_scale = np.log(stddev) if isinstance(stddev, Number) else torch.log( stddev) return -(value - mu) ** 2 / (2.0 * var) - log_scale - np.log(np.sqrt( 2.0 * np.pi)) def normpdf(value, mu=0.0, stddev=1.0): return torch.exp(_normal_log_pdf(value, mu, stddev)) class ReLUNew(nn.Module): def __init__(self, keep_variance_fn=None): super(ReLUNew, self).__init__() self._keep_variance_fn = keep_variance_fn def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0], output[1]
collector-m/LiDAR-MOS
ReLU
false
15,061
[ "MIT" ]
268
7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
https://github.com/collector-m/LiDAR-MOS/tree/7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
SoftmaxFocalClassificationLoss
import torch import torch.nn as nn import torch.nn.functional as F class SoftmaxFocalClassificationLoss(nn.Module): """Criterion that computes Focal loss. According to [1], the Focal loss is computed as follows: .. math:: \\text{FL}(p_t) = -\\alpha_t (1 - p_t)^{\\gamma} \\, \\text{log}(p_t) where: - :math:`p_t` is the model's estimated probability for each class. Arguments: alpha (float): Weighting factor :math:`\\alpha \\in [0, 1]`. gamma (float): Focusing parameter :math:`\\gamma >= 0`. reduction (str, optional): Specifies the reduction to apply to the output: ‘none’ | ‘mean’ | ‘sum’. ‘none’: no reduction will be applied, ‘mean’: the sum of the output will be divided by the number of elements in the output, ‘sum’: the output will be summed. Default: ‘none’. Shape: - Input: :math:`(N, C, *)` where C = number of classes. - Target: :math:`(N, *)` where each value is :math:`0 ≤ targets[i] ≤ C−1`. Examples: >>> N = 5 # num_classes >>> kwargs = {"alpha": 0.5, "gamma": 2.0, "reduction": 'mean'} >>> loss = kornia.losses.FocalLoss(**kwargs) >>> input = torch.randn(1, N, 3, 5, requires_grad=True) >>> target = torch.empty(1, 3, 5, dtype=torch.long).random_(N) >>> output = loss(input, target) >>> output.backward() References: [1] https://arxiv.org/abs/1708.02002 """ def __init__(self, alpha: 'float'=1.0, gamma: 'float'=2.0, reduction: 'str'='none') ->None: super(SoftmaxFocalClassificationLoss, self).__init__() self.alpha: 'float' = alpha self.gamma: 'float' = gamma self.reduction: 'str' = reduction self.eps: 'float' = 1e-06 def focal_loss(self, input: 'torch.Tensor', target: 'torch.Tensor', weights: 'torch.Tensor', alpha: 'float'=1.0, gamma: 'float'=2.0, reduction: 'str'='none', eps: 'float'=1e-08) ->torch.Tensor: """Function that computes Focal loss. See :class:`~kornia.losses.FocalLoss` for details. """ if not torch.is_tensor(input): raise TypeError('Input type is not a torch.Tensor. Got {}'. format(type(input))) if not len(input.shape) >= 2: raise ValueError('Invalid input shape, we expect BxCx*. Got: {}' .format(input.shape)) if input.size(0) != target.size(0): raise ValueError( 'Expected input batch_size ({}) to match target batch_size ({}).' .format(input.size(0), target.size(0))) if not input.device == target.device: raise ValueError( 'input and target must be in the same device. Got: {} and {}' .format(input.device, target.device)) input_soft: 'torch.Tensor' = F.softmax(input, dim=1) + eps weight = torch.pow(-input_soft + 1.0, gamma) focal = -alpha * weight * torch.log(input_soft) loss_tmp = torch.sum(target * focal, dim=1, keepdims=True) if weights is None: return loss_tmp return loss_tmp * weights def forward(self, input: 'torch.Tensor', target: 'torch.Tensor', weights: 'torch.Tensor') ->torch.Tensor: return self.focal_loss(input, target, weights, self.alpha, self. gamma, self.reduction, self.eps) 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 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.functional as F 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__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 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_add_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 = tmp1 + tmp2 tmp5 = tmp3 + tmp4 tmp7 = tmp5 + tmp6 tmp8 = tmp0 / tmp7 tmp9 = 1e-06 tmp10 = tmp8 + tmp9 tl.store(out_ptr0 + x3, tmp10, xmask) @triton.jit def triton_poi_fused_add_log_mul_neg_pow_sum_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 % 16 x1 = xindex // 16 x2 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask) tmp1 = tl.load(in_ptr1 + (x0 + 64 * x1), xmask) tmp11 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask) tmp12 = tl.load(in_ptr1 + (16 + x0 + 64 * x1), xmask) tmp21 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask) tmp22 = tl.load(in_ptr1 + (32 + x0 + 64 * x1), xmask) tmp31 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask) tmp32 = tl.load(in_ptr1 + (48 + x0 + 64 * x1), xmask) tmp2 = -tmp1 tmp3 = 1.0 tmp4 = tmp2 + tmp3 tmp5 = tmp4 * tmp4 tmp6 = -1.0 tmp7 = tmp5 * tmp6 tmp8 = tl_math.log(tmp1) tmp9 = tmp7 * tmp8 tmp10 = tmp0 * tmp9 tmp13 = -tmp12 tmp14 = tmp13 + tmp3 tmp15 = tmp14 * tmp14 tmp16 = tmp15 * tmp6 tmp17 = tl_math.log(tmp12) tmp18 = tmp16 * tmp17 tmp19 = tmp11 * tmp18 tmp20 = tmp10 + tmp19 tmp23 = -tmp22 tmp24 = tmp23 + tmp3 tmp25 = tmp24 * tmp24 tmp26 = tmp25 * tmp6 tmp27 = tl_math.log(tmp22) tmp28 = tmp26 * tmp27 tmp29 = tmp21 * tmp28 tmp30 = tmp20 + tmp29 tmp33 = -tmp32 tmp34 = tmp33 + tmp3 tmp35 = tmp34 * tmp34 tmp36 = tmp35 * tmp6 tmp37 = tl_math.log(tmp32) tmp38 = tmp36 * tmp37 tmp39 = tmp31 * tmp38 tmp40 = tmp30 + tmp39 tl.store(out_ptr0 + x2, tmp40, xmask) @triton.jit def triton_poi_fused_mul_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 x0 = xindex % 16 x2 = xindex // 64 x3 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 16 * x2), xmask, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr1 + x3, xmask) tmp2 = tmp0 * tmp1 tl.store(out_ptr0 + x3, tmp2, 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((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused__softmax_0[grid(256)](arg0_1, buf0, 256, XBLOCK= 256, num_warps=4, num_stages=1) del arg0_1 buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused__softmax_add_1[grid(256)](buf0, buf1, 256, XBLOCK= 128, num_warps=4, num_stages=1) del buf0 buf2 = empty_strided_cuda((4, 1, 4, 4), (16, 64, 4, 1), torch.float32) triton_poi_fused_add_log_mul_neg_pow_sum_2[grid(64)](arg1_1, buf1, buf2, 64, XBLOCK=64, num_warps=1, num_stages=1) del arg1_1 buf3 = buf1 del buf1 triton_poi_fused_mul_3[grid(256)](buf2, arg2_1, buf3, 256, XBLOCK= 128, num_warps=4, num_stages=1) del arg2_1 del buf2 return buf3, class SoftmaxFocalClassificationLossNew(nn.Module): """Criterion that computes Focal loss. According to [1], the Focal loss is computed as follows: .. math:: \\text{FL}(p_t) = -\\alpha_t (1 - p_t)^{\\gamma} \\, \\text{log}(p_t) where: - :math:`p_t` is the model's estimated probability for each class. Arguments: alpha (float): Weighting factor :math:`\\alpha \\in [0, 1]`. gamma (float): Focusing parameter :math:`\\gamma >= 0`. reduction (str, optional): Specifies the reduction to apply to the output: ‘none’ | ‘mean’ | ‘sum’. ‘none’: no reduction will be applied, ‘mean’: the sum of the output will be divided by the number of elements in the output, ‘sum’: the output will be summed. Default: ‘none’. Shape: - Input: :math:`(N, C, *)` where C = number of classes. - Target: :math:`(N, *)` where each value is :math:`0 ≤ targets[i] ≤ C−1`. Examples: >>> N = 5 # num_classes >>> kwargs = {"alpha": 0.5, "gamma": 2.0, "reduction": 'mean'} >>> loss = kornia.losses.FocalLoss(**kwargs) >>> input = torch.randn(1, N, 3, 5, requires_grad=True) >>> target = torch.empty(1, 3, 5, dtype=torch.long).random_(N) >>> output = loss(input, target) >>> output.backward() References: [1] https://arxiv.org/abs/1708.02002 """ def __init__(self, alpha: 'float'=1.0, gamma: 'float'=2.0, reduction: 'str'='none') ->None: super(SoftmaxFocalClassificationLossNew, self).__init__() self.alpha: 'float' = alpha self.gamma: 'float' = gamma self.reduction: 'str' = reduction self.eps: 'float' = 1e-06 def focal_loss(self, input: 'torch.Tensor', target: 'torch.Tensor', weights: 'torch.Tensor', alpha: 'float'=1.0, gamma: 'float'=2.0, reduction: 'str'='none', eps: 'float'=1e-08) ->torch.Tensor: """Function that computes Focal loss. See :class:`~kornia.losses.FocalLoss` for details. """ if not torch.is_tensor(input): raise TypeError('Input type is not a torch.Tensor. Got {}'. format(type(input))) if not len(input.shape) >= 2: raise ValueError('Invalid input shape, we expect BxCx*. Got: {}' .format(input.shape)) if input.size(0) != target.size(0): raise ValueError( 'Expected input batch_size ({}) to match target batch_size ({}).' .format(input.size(0), target.size(0))) if not input.device == target.device: raise ValueError( 'input and target must be in the same device. Got: {} and {}' .format(input.device, target.device)) input_soft: 'torch.Tensor' = F.softmax(input, dim=1) + eps weight = torch.pow(-input_soft + 1.0, gamma) focal = -alpha * weight * torch.log(input_soft) loss_tmp = torch.sum(target * focal, dim=1, keepdims=True) if weights is None: return loss_tmp return loss_tmp * weights 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]
collector-m/BtcDet
SoftmaxFocalClassificationLoss
false
15,062
[ "Apache-2.0" ]
108
80bee34f2f40931600f812a6edbcb27e51cb7ec3
https://github.com/collector-m/BtcDet/tree/80bee34f2f40931600f812a6edbcb27e51cb7ec3
AvgPool2d
import torch import torch.nn as nn import torch.nn.functional as F def keep_variance_fn(x): return x + 0.001 class AvgPool2d(nn.Module): def __init__(self, keep_variance_fn=None, kernel_size=2): super(AvgPool2d, self).__init__() self._keep_variance_fn = keep_variance_fn self.kernel_size = kernel_size def forward(self, inputs_mean, inputs_variance): outputs_mean = F.avg_pool2d(inputs_mean, self.kernel_size, stride=2, padding=1) outputs_variance = F.avg_pool2d(inputs_variance, self.kernel_size, stride=2, padding=1) outputs_variance = outputs_variance / (inputs_mean.size(2) * inputs_mean.size(3)) if self._keep_variance_fn is not None: outputs_variance = self._keep_variance_fn(outputs_variance) return outputs_mean, outputs_variance / (inputs_mean.shape[2] * inputs_mean.shape[3]) 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 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_avg_pool2d_0(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 x1 = xindex // 3 % 3 x0 = xindex % 3 x2 = xindex // 9 x4 = xindex tmp0 = -1 + 2 * x1 tmp1 = tl.full([1], 0, tl.int64) tmp2 = tmp0 >= tmp1 tmp3 = tl.full([1], 4, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tmp2 & tmp4 tmp6 = -1 + 2 * x0 tmp7 = tmp6 >= tmp1 tmp8 = tmp6 < tmp3 tmp9 = tmp7 & tmp8 tmp10 = tmp5 & tmp9 tmp11 = tl.load(in_ptr0 + (-5 + 2 * x0 + 8 * x1 + 16 * x2), tmp10 & xmask, eviction_policy='evict_last', other=0.0) tmp12 = 2 * x0 tmp13 = tmp12 >= tmp1 tmp14 = tmp12 < tmp3 tmp15 = tmp13 & tmp14 tmp16 = tmp5 & tmp15 tmp17 = tl.load(in_ptr0 + (-4 + 2 * x0 + 8 * x1 + 16 * x2), tmp16 & xmask, eviction_policy='evict_last', other=0.0) tmp18 = tmp17 + tmp11 tmp19 = 2 * x1 tmp20 = tmp19 >= tmp1 tmp21 = tmp19 < tmp3 tmp22 = tmp20 & tmp21 tmp23 = tmp22 & tmp9 tmp24 = tl.load(in_ptr0 + (-1 + 2 * x0 + 8 * x1 + 16 * x2), tmp23 & xmask, eviction_policy='evict_last', other=0.0) tmp25 = tmp24 + tmp18 tmp26 = tmp22 & tmp15 tmp27 = tl.load(in_ptr0 + (2 * x0 + 8 * x1 + 16 * x2), tmp26 & xmask, eviction_policy='evict_last', other=0.0) tmp28 = tmp27 + tmp25 tmp29 = 1 + -2 * x0 + -2 * x1 + (5 * (5 <= 1 + 2 * x0) + (1 + 2 * x0) * (1 + 2 * x0 < 5)) * (5 * (5 <= 1 + 2 * x1) + (1 + 2 * x1) * (1 + 2 * x1 < 5)) + -2 * x0 * (5 * (5 <= 1 + 2 * x1) + (1 + 2 * x1) * (1 + 2 * x1 < 5)) + -2 * x1 * (5 * (5 <= 1 + 2 * x0) + (1 + 2 * x0) * (1 + 2 * x0 < 5)) + 4 * x0 * x1 + (5 * (5 <= 1 + 2 * x0) + (1 + 2 * x0) * (1 + 2 * x0 < 5)) + (5 * (5 <= 1 + 2 * x1) + (1 + 2 * x1) * (1 + 2 * x1 < 5) ) tmp30 = tmp28 / tmp29 tl.store(out_ptr0 + x4, tmp30, xmask) @triton.jit def triton_poi_fused_avg_pool2d_div_1(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 144 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x1 = xindex // 3 % 3 x0 = xindex % 3 x2 = xindex // 9 x3 = xindex tmp0 = -1 + 2 * x1 tmp1 = tl.full([1], 0, tl.int64) tmp2 = tmp0 >= tmp1 tmp3 = tl.full([1], 4, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tmp2 & tmp4 tmp6 = -1 + 2 * x0 tmp7 = tmp6 >= tmp1 tmp8 = tmp6 < tmp3 tmp9 = tmp7 & tmp8 tmp10 = tmp5 & tmp9 tmp11 = tl.load(in_ptr0 + (-5 + 2 * x0 + 8 * x1 + 16 * x2), tmp10 & xmask, eviction_policy='evict_last', other=0.0) tmp12 = 2 * x0 tmp13 = tmp12 >= tmp1 tmp14 = tmp12 < tmp3 tmp15 = tmp13 & tmp14 tmp16 = tmp5 & tmp15 tmp17 = tl.load(in_ptr0 + (-4 + 2 * x0 + 8 * x1 + 16 * x2), tmp16 & xmask, eviction_policy='evict_last', other=0.0) tmp18 = tmp17 + tmp11 tmp19 = 2 * x1 tmp20 = tmp19 >= tmp1 tmp21 = tmp19 < tmp3 tmp22 = tmp20 & tmp21 tmp23 = tmp22 & tmp9 tmp24 = tl.load(in_ptr0 + (-1 + 2 * x0 + 8 * x1 + 16 * x2), tmp23 & xmask, eviction_policy='evict_last', other=0.0) tmp25 = tmp24 + tmp18 tmp26 = tmp22 & tmp15 tmp27 = tl.load(in_ptr0 + (2 * x0 + 8 * x1 + 16 * x2), tmp26 & xmask, eviction_policy='evict_last', other=0.0) tmp28 = tmp27 + tmp25 tmp29 = 1 + -2 * x0 + -2 * x1 + (5 * (5 <= 1 + 2 * x0) + (1 + 2 * x0) * (1 + 2 * x0 < 5)) * (5 * (5 <= 1 + 2 * x1) + (1 + 2 * x1) * (1 + 2 * x1 < 5)) + -2 * x0 * (5 * (5 <= 1 + 2 * x1) + (1 + 2 * x1) * (1 + 2 * x1 < 5)) + -2 * x1 * (5 * (5 <= 1 + 2 * x0) + (1 + 2 * x0) * (1 + 2 * x0 < 5)) + 4 * x0 * x1 + (5 * (5 <= 1 + 2 * x0) + (1 + 2 * x0) * (1 + 2 * x0 < 5)) + (5 * (5 <= 1 + 2 * x1) + (1 + 2 * x1) * (1 + 2 * x1 < 5) ) tmp30 = tmp28 / tmp29 tmp31 = 0.0625 tmp32 = tmp30 * tmp31 tmp33 = tmp32 * tmp31 tl.store(in_out_ptr0 + x3, tmp33, xmask) 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, 3, 3), (36, 9, 3, 1), torch.float32) get_raw_stream(0) triton_poi_fused_avg_pool2d_0[grid(144)](arg0_1, buf0, 144, XBLOCK= 256, num_warps=4, num_stages=1) del arg0_1 buf1 = empty_strided_cuda((4, 4, 3, 3), (36, 9, 3, 1), torch.float32) buf2 = buf1 del buf1 triton_poi_fused_avg_pool2d_div_1[grid(144)](buf2, arg1_1, 144, XBLOCK=256, num_warps=4, num_stages=1) del arg1_1 return buf0, buf2 def keep_variance_fn(x): return x + 0.001 class AvgPool2dNew(nn.Module): def __init__(self, keep_variance_fn=None, kernel_size=2): super(AvgPool2dNew, self).__init__() self._keep_variance_fn = keep_variance_fn self.kernel_size = kernel_size def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0], output[1]
collector-m/LiDAR-MOS
AvgPool2d
false
15,063
[ "MIT" ]
268
7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
https://github.com/collector-m/LiDAR-MOS/tree/7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
Conv
import torch from torch import nn class Conv(nn.Module): """ Convenience class that does padding and convolution for inputs in the format [batch_size, sequence length, hidden size] """ def __init__(self, input_size, output_size, kernel_size, pad_type): """ Parameters: input_size: Input feature size output_size: Output feature size kernel_size: Kernel width pad_type: left -> pad on the left side (to mask future data), both -> pad on both sides """ super(Conv, self).__init__() padding = (kernel_size - 1, 0) if pad_type == 'left' else ( kernel_size // 2, (kernel_size - 1) // 2) self.pad = nn.ConstantPad1d(padding, 0) self.conv = nn.Conv1d(input_size, output_size, kernel_size= kernel_size, padding=0) def forward(self, inputs): inputs = self.pad(inputs.permute(0, 2, 1)) outputs = self.conv(inputs).permute(0, 2, 1) return outputs def get_inputs(): return [torch.rand([4, 4, 4])] def get_init_inputs(): return [[], {'input_size': 4, 'output_size': 4, 'kernel_size': 4, 'pad_type': 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 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, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 16 xnumel = 7 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 = -2 + x2 tmp1 = tl.full([1, 1], 0, tl.int64) tmp2 = tmp0 >= tmp1 tmp3 = tl.full([1, 1], 4, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tmp2 & tmp4 tmp6 = tl.load(in_ptr0 + (-8 + y0 + 4 * x2 + 16 * y1), tmp5 & xmask & ymask, eviction_policy='evict_last', other=0.0) tl.store(out_ptr0 + (x2 + 7 * y3), tmp6, xmask & ymask) @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) 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), (16, 4, 1)) assert_size_stride(primals_3, (4,), (1,)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 4, 7), (28, 7, 1), torch.float32) get_raw_stream(0) triton_poi_fused_constant_pad_nd_0[grid(16, 7)](primals_1, buf0, 16, 7, XBLOCK=8, YBLOCK=16, num_warps=4, num_stages=1) del primals_1 buf1 = extern_kernels.convolution(buf0, primals_2, stride=(1,), padding=(0,), dilation=(1,), transposed=False, output_padding=( 0,), groups=1, bias=None) assert_size_stride(buf1, (4, 4, 4), (16, 4, 1)) buf2 = buf1 del buf1 triton_poi_fused_convolution_1[grid(64)](buf2, primals_3, 64, XBLOCK=64, num_warps=1, num_stages=1) del primals_3 return reinterpret_tensor(buf2, (4, 4, 4), (16, 1, 4), 0), primals_2, buf0 class ConvNew(nn.Module): """ Convenience class that does padding and convolution for inputs in the format [batch_size, sequence length, hidden size] """ def __init__(self, input_size, output_size, kernel_size, pad_type): """ Parameters: input_size: Input feature size output_size: Output feature size kernel_size: Kernel width pad_type: left -> pad on the left side (to mask future data), both -> pad on both sides """ super(ConvNew, self).__init__() padding = (kernel_size - 1, 0) if pad_type == 'left' else ( kernel_size // 2, (kernel_size - 1) // 2) self.pad = nn.ConstantPad1d(padding, 0) self.conv = nn.Conv1d(input_size, output_size, kernel_size= kernel_size, padding=0) def forward(self, input_0): primals_1 = self.conv.weight primals_3 = self.conv.bias primals_2 = input_0 output = call([primals_1, primals_2, primals_3]) return output[0]
colincen/coach
Conv
false
15,064
[ "MIT" ]
72
2b1b543851cc7ba359f48dac6a5c72f1ced9b530
https://github.com/colincen/coach/tree/2b1b543851cc7ba359f48dac6a5c72f1ced9b530
FCN8VGG16
import torch import numpy as np from torch import nn import torch.utils.model_zoo as model_zoo def conv3x3(in_planes, out_planes, stride=1, padding=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=(3, 3), stride=( stride, stride), padding=(padding, padding)) def get_upsampling_weight(in_channels, out_channels, kernel_size): """Make a 2D bilinear kernel suitable for upsampling""" factor = (kernel_size + 1) // 2 if kernel_size % 2 == 1: center = factor - 1 else: center = factor - 0.5 og = np.ogrid[:kernel_size, :kernel_size] filt = (1 - abs(og[0] - center) / factor) * (1 - abs(og[1] - center) / factor) weight = np.zeros((in_channels, out_channels, kernel_size, kernel_size), dtype=np.float64) weight[range(in_channels), range(out_channels), :, :] = filt return torch.from_numpy(weight).float() class FCN8VGG16(nn.Module): def __init__(self, n_classes): super().__init__() self.n_classes = n_classes self.pool = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True) self.relu = nn.ReLU(inplace=True) self.conv1_1 = conv3x3(3, 64, stride=1, padding=100) self.conv1_2 = conv3x3(64, 64) self.conv2_1 = conv3x3(64, 128) self.conv2_2 = conv3x3(128, 128) self.conv3_1 = conv3x3(128, 256) self.conv3_2 = conv3x3(256, 256) self.conv3_3 = conv3x3(256, 256) self.conv4_1 = conv3x3(256, 512) self.conv4_2 = conv3x3(512, 512) self.conv4_3 = conv3x3(512, 512) self.conv5_1 = conv3x3(512, 512) self.conv5_2 = conv3x3(512, 512) self.conv5_3 = conv3x3(512, 512) self.fc6 = nn.Conv2d(512, 4096, kernel_size=7, stride=1, padding=0) self.dropout_f6 = nn.Dropout() self.fc7 = nn.Conv2d(4096, 4096, kernel_size=1, stride=1, padding=0) self.dropout_f7 = nn.Dropout() self.scoring_layer = nn.Conv2d(4096, self.n_classes, kernel_size=1, stride=1, padding=0) self.upscore2 = nn.ConvTranspose2d(self.n_classes, self.n_classes, kernel_size=4, stride=2, bias=False) self.upscore_pool4 = nn.ConvTranspose2d(self.n_classes, self. n_classes, kernel_size=4, stride=2, bias=False) self.upscore8 = nn.ConvTranspose2d(self.n_classes, self.n_classes, kernel_size=16, stride=8, bias=False) self.scoring_layer.weight.data.zero_() self.scoring_layer.bias.data.zero_() self.score_pool3 = nn.Conv2d(256, self.n_classes, kernel_size=1) self.score_pool4 = nn.Conv2d(512, self.n_classes, kernel_size=1) self.score_pool3.weight.data.zero_() self.score_pool3.bias.data.zero_() self.score_pool4.weight.data.zero_() self.score_pool4.bias.data.zero_() self.upscore2.weight.data.copy_(get_upsampling_weight(self. n_classes, self.n_classes, 4)) self.upscore_pool4.weight.data.copy_(get_upsampling_weight(self. n_classes, self.n_classes, 4)) self.upscore8.weight.data.copy_(get_upsampling_weight(self. n_classes, self.n_classes, 16)) pth_url = 'https://download.pytorch.org/models/vgg16-397923af.pth' state_dict = model_zoo.load_url(pth_url) layer_names = [layer_name for layer_name in state_dict] counter = 0 for p in self.parameters(): if counter < 26: p.data = state_dict[layer_names[counter]] elif counter == 26: p.data = state_dict[layer_names[counter]].view(4096, 512, 7, 7) elif counter == 27: p.data = state_dict[layer_names[counter]] elif counter == 28: p.data = state_dict[layer_names[counter]].view(4096, 4096, 1, 1 ) elif counter == 29: p.data = state_dict[layer_names[counter]] counter += 1 def forward(self, x): _n, _c, h, w = x.size() conv1_1 = self.relu(self.conv1_1(x)) conv1_2 = self.relu(self.conv1_2(conv1_1)) pool1 = self.pool(conv1_2) conv2_1 = self.relu(self.conv2_1(pool1)) conv2_2 = self.relu(self.conv2_2(conv2_1)) pool2 = self.pool(conv2_2) conv3_1 = self.relu(self.conv3_1(pool2)) conv3_2 = self.relu(self.conv3_2(conv3_1)) conv3_3 = self.relu(self.conv3_3(conv3_2)) pool3 = self.pool(conv3_3) conv4_1 = self.relu(self.conv4_1(pool3)) conv4_2 = self.relu(self.conv4_2(conv4_1)) conv4_3 = self.relu(self.conv4_3(conv4_2)) pool4 = self.pool(conv4_3) conv5_1 = self.relu(self.conv5_1(pool4)) conv5_2 = self.relu(self.conv5_2(conv5_1)) conv5_3 = self.relu(self.conv5_3(conv5_2)) pool5 = self.pool(conv5_3) fc6 = self.dropout_f6(self.relu(self.fc6(pool5))) fc7 = self.dropout_f7(self.relu(self.fc7(fc6))) scores = self.scoring_layer(fc7) upscore2 = self.upscore2(scores) score_pool4 = self.score_pool4(pool4) score_pool4c = score_pool4[:, :, 5:5 + upscore2.size(2), 5:5 + upscore2.size(3)] upscore_pool4 = self.upscore_pool4(score_pool4c + upscore2) score_pool3 = self.score_pool3(pool3) score_pool3c = score_pool3[:, :, 9:9 + upscore_pool4.size(2), 9:9 + upscore_pool4.size(3)] output = self.upscore8(score_pool3c + upscore_pool4) return output[:, :, 31:31 + h, 31:31 + w].contiguous() def get_inputs(): return [torch.rand([4, 3, 64, 64])] def get_init_inputs(): return [[], {'n_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 from torch._inductor.runtime import triton_helpers import numpy as np from torch import nn import torch.utils.model_zoo as model_zoo 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 = 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 = 192 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 % 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_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) * 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_6(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_7(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_8(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_9(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl. constexpr, XBLOCK: tl.constexpr): xnumel = 49 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 + 49 * y3), xmask, eviction_policy= 'evict_last') tl.store(out_ptr0 + (y0 + 512 * x2 + 25088 * y1), tmp0, xmask) @triton.jit def triton_poi_fused_10(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl. constexpr, XBLOCK: tl.constexpr): ynumel = 16 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 x2 = xindex y3 = yindex y0 = yindex % 4 y1 = yindex // 4 tmp0 = tl.load(in_ptr0 + (x2 + 16 * y3), xmask & ymask) tl.store(out_ptr0 + (y0 + 4 * x2 + 64 * y1), tmp0, xmask & ymask) @triton.jit def triton_poi_fused_11(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl. constexpr, XBLOCK: tl.constexpr): ynumel = 16 xnumel = 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] xmask = xindex < xnumel x2 = xindex y3 = yindex y0 = yindex % 4 y1 = yindex // 4 tmp0 = tl.load(in_ptr0 + (x2 + 256 * y3), xmask & ymask, eviction_policy='evict_last') tl.store(out_ptr0 + (y0 + 4 * x2 + 1024 * y1), tmp0, xmask & ymask) @triton.jit def triton_poi_fused_convolution_relu_12(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 17572864 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_max_pool2d_with_indices_13(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr): xnumel = 4393216 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex % 64 x1 = xindex // 64 % 131 x2 = xindex // 8384 x3 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 128 * x1 + 33536 * x2), xmask) tmp1 = tl.load(in_ptr0 + (64 + x0 + 128 * x1 + 33536 * x2), xmask) tmp3 = tl.load(in_ptr0 + (16768 + x0 + 128 * x1 + 33536 * x2), xmask) tmp5 = tl.load(in_ptr0 + (16832 + x0 + 128 * x1 + 33536 * x2), xmask) 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, xmask) tl.store(out_ptr1 + x3, tmp16, xmask) @triton.jit def triton_poi_fused_convolution_relu_14(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 8786432 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_max_pool2d_with_indices_15(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) x2 = xindex // 8448 % 66 x1 = xindex // 128 % 66 x0 = xindex % 128 x3 = xindex // 557568 x6 = xindex tmp0 = 2 * x2 tmp1 = tl.full([1], 0, tl.int64) tmp2 = tmp0 >= tmp1 tmp3 = tl.full([1], 131, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tmp2 & tmp4 tmp6 = 2 * x1 tmp7 = tmp6 >= tmp1 tmp8 = tmp6 < tmp3 tmp9 = tmp7 & tmp8 tmp10 = tmp5 & tmp9 tmp11 = tl.load(in_ptr0 + (x0 + 256 * x1 + 33536 * x2 + 2196608 * x3), tmp10, other=float('-inf')) tmp12 = 1 + 2 * x1 tmp13 = tmp12 >= tmp1 tmp14 = tmp12 < tmp3 tmp15 = tmp13 & tmp14 tmp16 = tmp5 & tmp15 tmp17 = tl.load(in_ptr0 + (128 + x0 + 256 * x1 + 33536 * x2 + 2196608 * x3), tmp16, other=float('-inf')) tmp18 = triton_helpers.maximum(tmp17, tmp11) tmp19 = 1 + 2 * x2 tmp20 = tmp19 >= tmp1 tmp21 = tmp19 < tmp3 tmp22 = tmp20 & tmp21 tmp23 = tmp22 & tmp9 tmp24 = tl.load(in_ptr0 + (16768 + x0 + 256 * x1 + 33536 * x2 + 2196608 * x3), tmp23, other=float('-inf')) tmp25 = triton_helpers.maximum(tmp24, tmp18) tmp26 = tmp22 & tmp15 tmp27 = tl.load(in_ptr0 + (16896 + x0 + 256 * x1 + 33536 * x2 + 2196608 * x3), tmp26, other=float('-inf')) tmp28 = triton_helpers.maximum(tmp27, tmp25) tmp29 = tmp17 > tmp11 tmp30 = tl.full([1], 1, tl.int8) tmp31 = tl.full([1], 0, tl.int8) tmp32 = tl.where(tmp29, tmp30, tmp31) tmp33 = tmp24 > tmp18 tmp34 = tl.full([1], 2, tl.int8) tmp35 = tl.where(tmp33, tmp34, tmp32) tmp36 = tmp27 > tmp25 tmp37 = tl.full([1], 3, tl.int8) tmp38 = tl.where(tmp36, tmp37, tmp35) tl.store(out_ptr0 + x6, tmp28, None) tl.store(out_ptr1 + x6, tmp38, 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 % 256 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_17(in_ptr0, out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr): xnumel = 1115136 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex % 256 x1 = xindex // 256 % 33 x2 = xindex // 8448 x3 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 512 * x1 + 33792 * x2), xmask) tmp1 = tl.load(in_ptr0 + (256 + x0 + 512 * x1 + 33792 * x2), xmask) tmp3 = tl.load(in_ptr0 + (16896 + x0 + 512 * x1 + 33792 * x2), xmask) tmp5 = tl.load(in_ptr0 + (17152 + x0 + 512 * x1 + 33792 * x2), xmask) 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, xmask) tl.store(out_ptr1 + x3, tmp16, xmask) @triton.jit def triton_poi_fused_convolution_relu_18(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 = 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_19(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) x2 = xindex // 8704 % 17 x1 = xindex // 512 % 17 x0 = xindex % 512 x3 = xindex // 147968 x6 = xindex tmp0 = 2 * x2 tmp1 = tl.full([1], 0, tl.int64) tmp2 = tmp0 >= tmp1 tmp3 = tl.full([1], 33, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tmp2 & tmp4 tmp6 = 2 * x1 tmp7 = tmp6 >= tmp1 tmp8 = tmp6 < tmp3 tmp9 = tmp7 & tmp8 tmp10 = tmp5 & tmp9 tmp11 = tl.load(in_ptr0 + (x0 + 1024 * x1 + 33792 * x2 + 557568 * x3), tmp10, other=float('-inf')) tmp12 = 1 + 2 * x1 tmp13 = tmp12 >= tmp1 tmp14 = tmp12 < tmp3 tmp15 = tmp13 & tmp14 tmp16 = tmp5 & tmp15 tmp17 = tl.load(in_ptr0 + (512 + x0 + 1024 * x1 + 33792 * x2 + 557568 * x3), tmp16, other=float('-inf')) tmp18 = triton_helpers.maximum(tmp17, tmp11) tmp19 = 1 + 2 * x2 tmp20 = tmp19 >= tmp1 tmp21 = tmp19 < tmp3 tmp22 = tmp20 & tmp21 tmp23 = tmp22 & tmp9 tmp24 = tl.load(in_ptr0 + (16896 + x0 + 1024 * x1 + 33792 * x2 + 557568 * x3), tmp23, other=float('-inf')) tmp25 = triton_helpers.maximum(tmp24, tmp18) tmp26 = tmp22 & tmp15 tmp27 = tl.load(in_ptr0 + (17408 + x0 + 1024 * x1 + 33792 * x2 + 557568 * x3), tmp26, other=float('-inf')) tmp28 = triton_helpers.maximum(tmp27, tmp25) tmp29 = tmp17 > tmp11 tmp30 = tl.full([1], 1, tl.int8) tmp31 = tl.full([1], 0, tl.int8) tmp32 = tl.where(tmp29, tmp30, tmp31) tmp33 = tmp24 > tmp18 tmp34 = tl.full([1], 2, tl.int8) tmp35 = tl.where(tmp33, tmp34, tmp32) tmp36 = tmp27 > tmp25 tmp37 = tl.full([1], 3, tl.int8) tmp38 = tl.where(tmp36, tmp37, tmp35) tl.store(out_ptr0 + x6, tmp28, None) tl.store(out_ptr1 + x6, tmp38, None) @triton.jit def triton_poi_fused_convolution_relu_20(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 = 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_21(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) x2 = xindex // 4608 % 9 x1 = xindex // 512 % 9 x0 = xindex % 512 x3 = xindex // 41472 x6 = xindex tmp0 = 2 * x2 tmp1 = tl.full([1], 0, tl.int64) tmp2 = tmp0 >= tmp1 tmp3 = tl.full([1], 17, tl.int64) tmp4 = tmp0 < tmp3 tmp5 = tmp2 & tmp4 tmp6 = 2 * x1 tmp7 = tmp6 >= tmp1 tmp8 = tmp6 < tmp3 tmp9 = tmp7 & tmp8 tmp10 = tmp5 & tmp9 tmp11 = tl.load(in_ptr0 + (x0 + 1024 * x1 + 17408 * x2 + 147968 * x3), tmp10, other=float('-inf')) tmp12 = 1 + 2 * x1 tmp13 = tmp12 >= tmp1 tmp14 = tmp12 < tmp3 tmp15 = tmp13 & tmp14 tmp16 = tmp5 & tmp15 tmp17 = tl.load(in_ptr0 + (512 + x0 + 1024 * x1 + 17408 * x2 + 147968 * x3), tmp16, other=float('-inf')) tmp18 = triton_helpers.maximum(tmp17, tmp11) tmp19 = 1 + 2 * x2 tmp20 = tmp19 >= tmp1 tmp21 = tmp19 < tmp3 tmp22 = tmp20 & tmp21 tmp23 = tmp22 & tmp9 tmp24 = tl.load(in_ptr0 + (8704 + x0 + 1024 * x1 + 17408 * x2 + 147968 * x3), tmp23, other=float('-inf')) tmp25 = triton_helpers.maximum(tmp24, tmp18) tmp26 = tmp22 & tmp15 tmp27 = tl.load(in_ptr0 + (9216 + x0 + 1024 * x1 + 17408 * x2 + 147968 * x3), tmp26, other=float('-inf')) tmp28 = triton_helpers.maximum(tmp27, tmp25) tmp29 = tmp17 > tmp11 tmp30 = tl.full([1], 1, tl.int8) tmp31 = tl.full([1], 0, tl.int8) tmp32 = tl.where(tmp29, tmp30, tmp31) tmp33 = tmp24 > tmp18 tmp34 = tl.full([1], 2, tl.int8) tmp35 = tl.where(tmp33, tmp34, tmp32) tmp36 = tmp27 > tmp25 tmp37 = tl.full([1], 3, tl.int8) tmp38 = tl.where(tmp36, tmp37, tmp35) tl.store(out_ptr0 + x6, tmp28, None) tl.store(out_ptr1 + x6, tmp38, None) @triton.jit def triton_poi_fused_convolution_relu_22(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_convolution_23(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 144 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) @triton.jit def triton_poi_fused_add_24(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK: tl.constexpr): xnumel = 1024 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x2 = xindex // 32 % 8 x3 = xindex // 256 x4 = xindex % 32 x0 = xindex % 4 x5 = xindex tmp0 = tl.load(in_ptr0 + (360 + x4 + 68 * x2 + 1156 * x3), xmask) tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last') tmp3 = tl.load(in_out_ptr0 + x5, xmask) tmp2 = tmp0 + tmp1 tmp4 = tmp2 + tmp3 tl.store(in_out_ptr0 + x5, tmp4, xmask) @triton.jit def triton_poi_fused_add_25(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK: tl.constexpr): xnumel = 5184 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x2 = xindex // 72 % 18 x3 = xindex // 1296 x4 = xindex % 72 x0 = xindex % 4 x5 = xindex tmp0 = tl.load(in_ptr0 + (1224 + x4 + 132 * x2 + 4356 * x3), xmask) tmp1 = tl.load(in_ptr1 + x0, xmask, eviction_policy='evict_last') tmp3 = tl.load(in_out_ptr0 + x5, xmask) tmp2 = tmp0 + tmp1 tmp4 = tmp2 + tmp3 tl.store(in_out_ptr0 + x5, tmp4, xmask) @triton.jit def triton_poi_fused_clone_26(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl .constexpr, XBLOCK: tl.constexpr): ynumel = 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] tl.full([XBLOCK, YBLOCK], True, tl.int1) x2 = xindex % 64 x3 = xindex // 64 y0 = yindex % 4 y1 = yindex // 4 x5 = xindex y4 = yindex tmp0 = tl.load(in_ptr0 + (18972 + y0 + 4 * x2 + 608 * x3 + 92416 * y1), ymask, eviction_policy='evict_last') tl.store(out_ptr0 + (x5 + 4096 * y4), tmp0, 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, 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, primals_32, primals_33, primals_34, primals_35, primals_36, primals_37, primals_38, primals_39, primals_40) = args args.clear() assert_size_stride(primals_1, (4, 3, 64, 64), (12288, 4096, 64, 1)) assert_size_stride(primals_2, (64, 3, 3, 3), (27, 9, 3, 1)) assert_size_stride(primals_3, (64,), (1,)) assert_size_stride(primals_4, (64, 64, 3, 3), (576, 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, 128, 3, 3), (1152, 9, 3, 1)) assert_size_stride(primals_9, (128,), (1,)) assert_size_stride(primals_10, (256, 128, 3, 3), (1152, 9, 3, 1)) assert_size_stride(primals_11, (256,), (1,)) assert_size_stride(primals_12, (256, 256, 3, 3), (2304, 9, 3, 1)) assert_size_stride(primals_13, (256,), (1,)) assert_size_stride(primals_14, (256, 256, 3, 3), (2304, 9, 3, 1)) assert_size_stride(primals_15, (256,), (1,)) assert_size_stride(primals_16, (512, 256, 3, 3), (2304, 9, 3, 1)) assert_size_stride(primals_17, (512,), (1,)) assert_size_stride(primals_18, (512, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_19, (512,), (1,)) assert_size_stride(primals_20, (512, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_21, (512,), (1,)) assert_size_stride(primals_22, (512, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_23, (512,), (1,)) assert_size_stride(primals_24, (512, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_25, (512,), (1,)) assert_size_stride(primals_26, (512, 512, 3, 3), (4608, 9, 3, 1)) assert_size_stride(primals_27, (512,), (1,)) assert_size_stride(primals_28, (4096, 512, 7, 7), (25088, 49, 7, 1)) assert_size_stride(primals_29, (4096,), (1,)) assert_size_stride(primals_30, (4096, 4096, 1, 1), (4096, 1, 1, 1)) assert_size_stride(primals_31, (4096,), (1,)) assert_size_stride(primals_32, (4, 4096, 1, 1), (4096, 1, 1, 1)) assert_size_stride(primals_33, (4,), (1,)) assert_size_stride(primals_34, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_35, (4, 512, 1, 1), (512, 1, 1, 1)) assert_size_stride(primals_36, (4,), (1,)) assert_size_stride(primals_37, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_38, (4, 256, 1, 1), (256, 1, 1, 1)) assert_size_stride(primals_39, (4,), (1,)) assert_size_stride(primals_40, (4, 4, 16, 16), (1024, 256, 16, 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((64, 3, 3, 3), (27, 1, 9, 3), torch.float32) triton_poi_fused_1[grid(192, 9)](primals_2, buf1, 192, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_2 buf2 = empty_strided_cuda((64, 64, 3, 3), (576, 1, 192, 64), torch. float32) triton_poi_fused_2[grid(4096, 9)](primals_4, buf2, 4096, 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, 128, 3, 3), (1152, 1, 384, 128), torch.float32) triton_poi_fused_4[grid(16384, 9)](primals_8, buf4, 16384, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_8 buf5 = empty_strided_cuda((256, 128, 3, 3), (1152, 1, 384, 128), torch.float32) triton_poi_fused_5[grid(32768, 9)](primals_10, buf5, 32768, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_10 buf6 = empty_strided_cuda((256, 256, 3, 3), (2304, 1, 768, 256), torch.float32) triton_poi_fused_6[grid(65536, 9)](primals_12, buf6, 65536, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_12 buf7 = empty_strided_cuda((256, 256, 3, 3), (2304, 1, 768, 256), torch.float32) triton_poi_fused_6[grid(65536, 9)](primals_14, buf7, 65536, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_14 buf8 = empty_strided_cuda((512, 256, 3, 3), (2304, 1, 768, 256), torch.float32) triton_poi_fused_7[grid(131072, 9)](primals_16, buf8, 131072, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_16 buf9 = empty_strided_cuda((512, 512, 3, 3), (4608, 1, 1536, 512), torch.float32) triton_poi_fused_8[grid(262144, 9)](primals_18, buf9, 262144, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_18 buf10 = empty_strided_cuda((512, 512, 3, 3), (4608, 1, 1536, 512), torch.float32) triton_poi_fused_8[grid(262144, 9)](primals_20, buf10, 262144, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_20 buf11 = empty_strided_cuda((512, 512, 3, 3), (4608, 1, 1536, 512), torch.float32) triton_poi_fused_8[grid(262144, 9)](primals_22, buf11, 262144, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_22 buf12 = empty_strided_cuda((512, 512, 3, 3), (4608, 1, 1536, 512), torch.float32) triton_poi_fused_8[grid(262144, 9)](primals_24, buf12, 262144, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_24 buf13 = empty_strided_cuda((512, 512, 3, 3), (4608, 1, 1536, 512), torch.float32) triton_poi_fused_8[grid(262144, 9)](primals_26, buf13, 262144, 9, XBLOCK=16, YBLOCK=64, num_warps=4, num_stages=1) del primals_26 buf14 = empty_strided_cuda((4096, 512, 7, 7), (25088, 1, 3584, 512), torch.float32) triton_poi_fused_9[grid(2097152, 49)](primals_28, buf14, 2097152, 49, XBLOCK=32, YBLOCK=64, num_warps=8, num_stages=1) del primals_28 buf15 = empty_strided_cuda((4, 4, 4, 4), (64, 1, 16, 4), torch.float32) triton_poi_fused_10[grid(16, 16)](primals_34, buf15, 16, 16, XBLOCK =16, YBLOCK=16, num_warps=4, num_stages=1) del primals_34 buf16 = empty_strided_cuda((4, 4, 4, 4), (64, 1, 16, 4), torch.float32) triton_poi_fused_10[grid(16, 16)](primals_37, buf16, 16, 16, XBLOCK =16, YBLOCK=16, num_warps=4, num_stages=1) del primals_37 buf17 = empty_strided_cuda((4, 4, 16, 16), (1024, 1, 64, 4), torch. float32) triton_poi_fused_11[grid(16, 256)](primals_40, buf17, 16, 256, XBLOCK=64, YBLOCK=16, num_warps=4, num_stages=1) del primals_40 buf18 = extern_kernels.convolution(buf0, buf1, stride=(1, 1), padding=(100, 100), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf18, (4, 64, 262, 262), (4393216, 1, 16768, 64)) buf19 = buf18 del buf18 triton_poi_fused_convolution_relu_12[grid(17572864)](buf19, primals_3, 17572864, XBLOCK=512, num_warps=8, num_stages=1) del primals_3 buf20 = extern_kernels.convolution(buf19, buf2, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf20, (4, 64, 262, 262), (4393216, 1, 16768, 64)) buf21 = buf20 del buf20 triton_poi_fused_convolution_relu_12[grid(17572864)](buf21, primals_5, 17572864, XBLOCK=512, num_warps=8, num_stages=1) del primals_5 buf22 = empty_strided_cuda((4, 64, 131, 131), (1098304, 1, 8384, 64 ), torch.float32) buf23 = empty_strided_cuda((4, 64, 131, 131), (1098304, 1, 8384, 64 ), torch.int8) triton_poi_fused_max_pool2d_with_indices_13[grid(4393216)](buf21, buf22, buf23, 4393216, XBLOCK=512, num_warps=8, num_stages=1) buf24 = extern_kernels.convolution(buf22, buf3, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf24, (4, 128, 131, 131), (2196608, 1, 16768, 128)) buf25 = buf24 del buf24 triton_poi_fused_convolution_relu_14[grid(8786432)](buf25, primals_7, 8786432, XBLOCK=1024, num_warps=4, num_stages=1) del primals_7 buf26 = extern_kernels.convolution(buf25, buf4, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf26, (4, 128, 131, 131), (2196608, 1, 16768, 128)) buf27 = buf26 del buf26 triton_poi_fused_convolution_relu_14[grid(8786432)](buf27, primals_9, 8786432, XBLOCK=1024, num_warps=4, num_stages=1) del primals_9 buf28 = empty_strided_cuda((4, 128, 66, 66), (557568, 1, 8448, 128), torch.float32) buf29 = empty_strided_cuda((4, 128, 66, 66), (557568, 1, 8448, 128), torch.int8) triton_poi_fused_max_pool2d_with_indices_15[grid(2230272)](buf27, buf28, buf29, 2230272, XBLOCK=512, num_warps=8, num_stages=1) buf30 = extern_kernels.convolution(buf28, buf5, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf30, (4, 256, 66, 66), (1115136, 1, 16896, 256)) buf31 = buf30 del buf30 triton_poi_fused_convolution_relu_16[grid(4460544)](buf31, primals_11, 4460544, XBLOCK=1024, num_warps=4, num_stages=1) del primals_11 buf32 = extern_kernels.convolution(buf31, buf6, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf32, (4, 256, 66, 66), (1115136, 1, 16896, 256)) buf33 = buf32 del buf32 triton_poi_fused_convolution_relu_16[grid(4460544)](buf33, primals_13, 4460544, XBLOCK=1024, num_warps=4, num_stages=1) del primals_13 buf34 = extern_kernels.convolution(buf33, buf7, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf34, (4, 256, 66, 66), (1115136, 1, 16896, 256)) buf35 = buf34 del buf34 triton_poi_fused_convolution_relu_16[grid(4460544)](buf35, primals_15, 4460544, XBLOCK=1024, num_warps=4, num_stages=1) del primals_15 buf36 = empty_strided_cuda((4, 256, 33, 33), (278784, 1, 8448, 256), torch.float32) buf37 = empty_strided_cuda((4, 256, 33, 33), (278784, 1, 8448, 256), torch.int8) triton_poi_fused_max_pool2d_with_indices_17[grid(1115136)](buf35, buf36, buf37, 1115136, XBLOCK=512, num_warps=8, num_stages=1) buf38 = extern_kernels.convolution(buf36, buf8, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf38, (4, 512, 33, 33), (557568, 1, 16896, 512)) buf39 = buf38 del buf38 triton_poi_fused_convolution_relu_18[grid(2230272)](buf39, primals_17, 2230272, XBLOCK=512, num_warps=8, num_stages=1) del primals_17 buf40 = extern_kernels.convolution(buf39, buf9, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf40, (4, 512, 33, 33), (557568, 1, 16896, 512)) buf41 = buf40 del buf40 triton_poi_fused_convolution_relu_18[grid(2230272)](buf41, primals_19, 2230272, XBLOCK=512, num_warps=8, num_stages=1) del primals_19 buf42 = extern_kernels.convolution(buf41, buf10, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf42, (4, 512, 33, 33), (557568, 1, 16896, 512)) buf43 = buf42 del buf42 triton_poi_fused_convolution_relu_18[grid(2230272)](buf43, primals_21, 2230272, XBLOCK=512, num_warps=8, num_stages=1) del primals_21 buf44 = empty_strided_cuda((4, 512, 17, 17), (147968, 1, 8704, 512), torch.float32) buf45 = empty_strided_cuda((4, 512, 17, 17), (147968, 1, 8704, 512), torch.int8) triton_poi_fused_max_pool2d_with_indices_19[grid(591872)](buf43, buf44, buf45, 591872, XBLOCK=1024, num_warps=4, num_stages=1) buf46 = extern_kernels.convolution(buf44, buf11, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf46, (4, 512, 17, 17), (147968, 1, 8704, 512)) buf47 = buf46 del buf46 triton_poi_fused_convolution_relu_20[grid(591872)](buf47, primals_23, 591872, XBLOCK=1024, num_warps=4, num_stages=1) del primals_23 buf48 = extern_kernels.convolution(buf47, buf12, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf48, (4, 512, 17, 17), (147968, 1, 8704, 512)) buf49 = buf48 del buf48 triton_poi_fused_convolution_relu_20[grid(591872)](buf49, primals_25, 591872, XBLOCK=1024, num_warps=4, num_stages=1) del primals_25 buf50 = extern_kernels.convolution(buf49, buf13, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf50, (4, 512, 17, 17), (147968, 1, 8704, 512)) buf51 = buf50 del buf50 triton_poi_fused_convolution_relu_20[grid(591872)](buf51, primals_27, 591872, XBLOCK=1024, num_warps=4, num_stages=1) del primals_27 buf52 = empty_strided_cuda((4, 512, 9, 9), (41472, 1, 4608, 512), torch.float32) buf53 = empty_strided_cuda((4, 512, 9, 9), (41472, 1, 4608, 512), torch.int8) triton_poi_fused_max_pool2d_with_indices_21[grid(165888)](buf51, buf52, buf53, 165888, XBLOCK=512, num_warps=8, num_stages=1) buf54 = extern_kernels.convolution(buf52, buf14, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf54, (4, 4096, 3, 3), (36864, 1, 12288, 4096)) buf55 = buf54 del buf54 triton_poi_fused_convolution_relu_22[grid(147456)](buf55, primals_29, 147456, XBLOCK=512, num_warps=8, num_stages=1) del primals_29 buf56 = extern_kernels.convolution(buf55, primals_30, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf56, (4, 4096, 3, 3), (36864, 1, 12288, 4096)) buf57 = buf56 del buf56 triton_poi_fused_convolution_relu_22[grid(147456)](buf57, primals_31, 147456, XBLOCK=512, num_warps=8, num_stages=1) del primals_31 buf58 = extern_kernels.convolution(buf57, primals_32, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf58, (4, 4, 3, 3), (36, 1, 12, 4)) buf59 = buf58 del buf58 triton_poi_fused_convolution_23[grid(144)](buf59, primals_33, 144, XBLOCK=128, num_warps=4, num_stages=1) del primals_33 buf60 = extern_kernels.convolution(buf59, buf15, stride=(2, 2), padding=(0, 0), dilation=(1, 1), transposed=True, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf60, (4, 4, 8, 8), (256, 1, 32, 4)) buf61 = extern_kernels.convolution(buf44, primals_35, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf61, (4, 4, 17, 17), (1156, 1, 68, 4)) buf62 = buf60 del buf60 triton_poi_fused_add_24[grid(1024)](buf62, buf61, primals_36, 1024, XBLOCK=128, num_warps=4, num_stages=1) del buf61 del primals_36 buf63 = extern_kernels.convolution(buf62, buf16, stride=(2, 2), padding=(0, 0), dilation=(1, 1), transposed=True, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf63, (4, 4, 18, 18), (1296, 1, 72, 4)) buf64 = extern_kernels.convolution(buf36, primals_38, stride=(1, 1), padding=(0, 0), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf64, (4, 4, 33, 33), (4356, 1, 132, 4)) buf65 = buf63 del buf63 triton_poi_fused_add_25[grid(5184)](buf65, buf64, primals_39, 5184, XBLOCK=128, num_warps=4, num_stages=1) del buf64 del primals_39 buf66 = extern_kernels.convolution(buf65, buf17, stride=(8, 8), padding=(0, 0), dilation=(1, 1), transposed=True, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf66, (4, 4, 152, 152), (92416, 1, 608, 4)) buf67 = empty_strided_cuda((4, 4, 64, 64), (16384, 4096, 64, 1), torch.float32) triton_poi_fused_clone_26[grid(16, 4096)](buf66, buf67, 16, 4096, XBLOCK=64, YBLOCK=16, num_warps=4, num_stages=1) del buf66 return (buf67, buf0, buf1, buf2, buf3, buf4, buf5, buf6, buf7, buf8, buf9, buf10, buf11, buf12, buf13, buf14, primals_30, primals_32, buf15, primals_35, buf16, primals_38, buf17, buf19, buf21, buf22, buf23, buf25, buf27, buf28, buf29, buf31, buf33, buf35, buf36, buf37, buf39, buf41, buf43, buf44, buf45, buf47, buf49, buf51, buf52, buf53, buf55, buf57, buf59, buf62, buf65) def conv3x3(in_planes, out_planes, stride=1, padding=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=(3, 3), stride=( stride, stride), padding=(padding, padding)) def get_upsampling_weight(in_channels, out_channels, kernel_size): """Make a 2D bilinear kernel suitable for upsampling""" factor = (kernel_size + 1) // 2 if kernel_size % 2 == 1: center = factor - 1 else: center = factor - 0.5 og = np.ogrid[:kernel_size, :kernel_size] filt = (1 - abs(og[0] - center) / factor) * (1 - abs(og[1] - center) / factor) weight = np.zeros((in_channels, out_channels, kernel_size, kernel_size), dtype=np.float64) weight[range(in_channels), range(out_channels), :, :] = filt return torch.from_numpy(weight).float() class FCN8VGG16New(nn.Module): def __init__(self, n_classes): super().__init__() self.n_classes = n_classes self.pool = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True) self.relu = nn.ReLU(inplace=True) self.conv1_1 = conv3x3(3, 64, stride=1, padding=100) self.conv1_2 = conv3x3(64, 64) self.conv2_1 = conv3x3(64, 128) self.conv2_2 = conv3x3(128, 128) self.conv3_1 = conv3x3(128, 256) self.conv3_2 = conv3x3(256, 256) self.conv3_3 = conv3x3(256, 256) self.conv4_1 = conv3x3(256, 512) self.conv4_2 = conv3x3(512, 512) self.conv4_3 = conv3x3(512, 512) self.conv5_1 = conv3x3(512, 512) self.conv5_2 = conv3x3(512, 512) self.conv5_3 = conv3x3(512, 512) self.fc6 = nn.Conv2d(512, 4096, kernel_size=7, stride=1, padding=0) self.dropout_f6 = nn.Dropout() self.fc7 = nn.Conv2d(4096, 4096, kernel_size=1, stride=1, padding=0) self.dropout_f7 = nn.Dropout() self.scoring_layer = nn.Conv2d(4096, self.n_classes, kernel_size=1, stride=1, padding=0) self.upscore2 = nn.ConvTranspose2d(self.n_classes, self.n_classes, kernel_size=4, stride=2, bias=False) self.upscore_pool4 = nn.ConvTranspose2d(self.n_classes, self. n_classes, kernel_size=4, stride=2, bias=False) self.upscore8 = nn.ConvTranspose2d(self.n_classes, self.n_classes, kernel_size=16, stride=8, bias=False) self.scoring_layer.weight.data.zero_() self.scoring_layer.bias.data.zero_() self.score_pool3 = nn.Conv2d(256, self.n_classes, kernel_size=1) self.score_pool4 = nn.Conv2d(512, self.n_classes, kernel_size=1) self.score_pool3.weight.data.zero_() self.score_pool3.bias.data.zero_() self.score_pool4.weight.data.zero_() self.score_pool4.bias.data.zero_() self.upscore2.weight.data.copy_(get_upsampling_weight(self. n_classes, self.n_classes, 4)) self.upscore_pool4.weight.data.copy_(get_upsampling_weight(self. n_classes, self.n_classes, 4)) self.upscore8.weight.data.copy_(get_upsampling_weight(self. n_classes, self.n_classes, 16)) pth_url = 'https://download.pytorch.org/models/vgg16-397923af.pth' state_dict = model_zoo.load_url(pth_url) layer_names = [layer_name for layer_name in state_dict] counter = 0 for p in self.parameters(): if counter < 26: p.data = state_dict[layer_names[counter]] elif counter == 26: p.data = state_dict[layer_names[counter]].view(4096, 512, 7, 7) elif counter == 27: p.data = state_dict[layer_names[counter]] elif counter == 28: p.data = state_dict[layer_names[counter]].view(4096, 4096, 1, 1 ) elif counter == 29: p.data = state_dict[layer_names[counter]] counter += 1 def forward(self, input_0): primals_2 = self.conv1_1.weight primals_3 = self.conv1_1.bias primals_4 = self.conv1_2.weight primals_5 = self.conv1_2.bias primals_6 = self.conv2_1.weight primals_7 = self.conv2_1.bias primals_8 = self.conv2_2.weight primals_9 = self.conv2_2.bias primals_10 = self.conv3_1.weight primals_11 = self.conv3_1.bias primals_12 = self.conv3_2.weight primals_13 = self.conv3_2.bias primals_14 = self.conv3_3.weight primals_15 = self.conv3_3.bias primals_16 = self.conv4_1.weight primals_17 = self.conv4_1.bias primals_18 = self.conv4_2.weight primals_19 = self.conv4_2.bias primals_20 = self.conv4_3.weight primals_21 = self.conv4_3.bias primals_22 = self.conv5_1.weight primals_23 = self.conv5_1.bias primals_24 = self.conv5_2.weight primals_25 = self.conv5_2.bias primals_26 = self.conv5_3.weight primals_27 = self.conv5_3.bias primals_28 = self.fc6.weight primals_29 = self.fc6.bias primals_30 = self.fc7.weight primals_31 = self.fc7.bias primals_32 = self.scoring_layer.weight primals_33 = self.scoring_layer.bias primals_34 = self.upscore2.weight primals_37 = self.upscore_pool4.weight primals_40 = self.upscore8.weight primals_38 = self.score_pool3.weight primals_36 = self.score_pool3.bias primals_35 = self.score_pool4.weight primals_39 = self.score_pool4.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, primals_20, primals_21, primals_22, primals_23, primals_24, primals_25, primals_26, primals_27, primals_28, primals_29, primals_30, primals_31, primals_32, primals_33, primals_34, primals_35, primals_36, primals_37, primals_38, primals_39, primals_40]) return output[0]
alzayats/DeepFish
FCN8VGG16
false
15,065
[ "MIT" ]
48
4d9ebfb0474a7e9346c72e2a5411ab6f72e878e2
https://github.com/alzayats/DeepFish/tree/4d9ebfb0474a7e9346c72e2a5411ab6f72e878e2
Softmax
import torch import torch.nn as nn def keep_variance_fn(x): return x + 0.001 class Softmax(nn.Module): def __init__(self, dim=1, keep_variance_fn=None): super(Softmax, self).__init__() self.dim = dim self._keep_variance_fn = keep_variance_fn def forward(self, features_mean, features_variance, eps=1e-05): """Softmax function applied to a multivariate Gaussian distribution. It works under the assumption that features_mean and features_variance are the parameters of a the indepent gaussians that contribute to the multivariate gaussian. Mean and variance of the log-normal distribution are computed following https://en.wikipedia.org/wiki/Log-normal_distribution.""" log_gaussian_mean = features_mean + 0.5 * features_variance log_gaussian_variance = 2 * log_gaussian_mean log_gaussian_mean = torch.exp(log_gaussian_mean) log_gaussian_variance = torch.exp(log_gaussian_variance) log_gaussian_variance = log_gaussian_variance * (torch.exp( features_variance) - 1) constant = torch.sum(log_gaussian_mean, dim=self.dim) + eps constant = constant.unsqueeze(self.dim) outputs_mean = log_gaussian_mean / constant outputs_variance = log_gaussian_variance / constant ** 2 if self._keep_variance_fn is not None: outputs_variance = self._keep_variance_fn(outputs_variance) return outputs_mean, outputs_variance 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.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_poi_fused_add_exp_mul_sum_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 % 16 x1 = xindex // 16 x2 = xindex tmp0 = tl.load(in_ptr0 + (x0 + 64 * x1), xmask) tmp1 = tl.load(in_ptr1 + (x0 + 64 * x1), xmask) tmp6 = tl.load(in_ptr0 + (16 + x0 + 64 * x1), xmask) tmp7 = tl.load(in_ptr1 + (16 + x0 + 64 * x1), xmask) tmp12 = tl.load(in_ptr0 + (32 + x0 + 64 * x1), xmask) tmp13 = tl.load(in_ptr1 + (32 + x0 + 64 * x1), xmask) tmp18 = tl.load(in_ptr0 + (48 + x0 + 64 * x1), xmask) tmp19 = tl.load(in_ptr1 + (48 + x0 + 64 * x1), xmask) tmp2 = 0.5 tmp3 = tmp1 * tmp2 tmp4 = tmp0 + tmp3 tmp5 = tl_math.exp(tmp4) tmp8 = tmp7 * tmp2 tmp9 = tmp6 + tmp8 tmp10 = tl_math.exp(tmp9) tmp11 = tmp5 + tmp10 tmp14 = tmp13 * tmp2 tmp15 = tmp12 + tmp14 tmp16 = tl_math.exp(tmp15) tmp17 = tmp11 + tmp16 tmp20 = tmp19 * tmp2 tmp21 = tmp18 + tmp20 tmp22 = tl_math.exp(tmp21) tmp23 = tmp17 + tmp22 tmp24 = 1e-05 tmp25 = tmp23 + tmp24 tl.store(out_ptr0 + x2, tmp25, xmask) @triton.jit def triton_poi_fused_add_div_exp_mul_pow_sub_1(in_ptr0, in_ptr1, in_ptr2, 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 x0 = xindex % 16 x2 = xindex // 64 tmp0 = tl.load(in_ptr0 + x3, xmask) tmp1 = tl.load(in_ptr1 + x3, xmask) tmp6 = tl.load(in_ptr2 + (x0 + 16 * x2), xmask, eviction_policy= 'evict_last') tmp2 = 0.5 tmp3 = tmp1 * tmp2 tmp4 = tmp0 + tmp3 tmp5 = tl_math.exp(tmp4) tmp7 = tmp5 / tmp6 tmp8 = 2.0 tmp9 = tmp4 * tmp8 tmp10 = tl_math.exp(tmp9) tmp11 = tl_math.exp(tmp1) tmp12 = 1.0 tmp13 = tmp11 - tmp12 tmp14 = tmp10 * tmp13 tmp15 = tmp6 * tmp6 tmp16 = tmp14 / tmp15 tl.store(out_ptr0 + x3, tmp7, xmask) tl.store(out_ptr1 + x3, tmp16, xmask) 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, 4), (16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_add_exp_mul_sum_0[grid(64)](arg1_1, arg0_1, buf0, 64, XBLOCK=64, num_warps=1, num_stages=1) buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused_add_div_exp_mul_pow_sub_1[grid(256)](arg1_1, arg0_1, buf0, buf1, buf2, 256, XBLOCK=128, num_warps=4, num_stages=1) del arg0_1 del arg1_1 del buf0 return buf1, buf2 def keep_variance_fn(x): return x + 0.001 class SoftmaxNew(nn.Module): def __init__(self, dim=1, keep_variance_fn=None): super(SoftmaxNew, self).__init__() self.dim = dim self._keep_variance_fn = keep_variance_fn def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0], output[1]
collector-m/LiDAR-MOS
Softmax
false
15,066
[ "MIT" ]
268
7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
https://github.com/collector-m/LiDAR-MOS/tree/7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
lovasz_hinge
import torch import torch.nn.parallel import torch.utils.data from torchvision.transforms import functional as F import torch.nn.functional as F from torch.autograd import Variable def flatten_binary_scores(scores, labels, ignore=255): """ Flattens predictions in the batch (binary case) Remove labels equal to 'ignore' """ scores = scores.view(-1) labels = labels.view(-1) if ignore is None: return scores, labels valid = labels != ignore vscores = scores[valid] vlabels = labels[valid] return vscores, vlabels def lovasz_grad(gt_sorted): """ Computes gradient of the Lovasz extension w.r.t sorted errors See Alg. 1 in paper """ p = len(gt_sorted) gts = gt_sorted.sum() intersection = gts.float() - gt_sorted.float().cumsum(0) union = gts.float() + (1 - gt_sorted).float().cumsum(0) jaccard = 1.0 - intersection / union if p > 1: jaccard[1:p] = jaccard[1:p] - jaccard[0:-1] return jaccard def isnan(x): return x != x def mean(l, ignore_nan=False, empty=0): """ nanmean compatible with generators. """ l = iter(l) if ignore_nan: l = ifilterfalse(isnan, l) try: n = 1 acc = next(l) except StopIteration: if empty == 'raise': raise ValueError('Empty mean') return empty for n, v in enumerate(l, 2): acc += v if n == 1: return acc return acc / n class lovasz_hinge(torch.nn.Module): def __init__(self, per_img=False, ignore=255): """ :param weight: 1D weight vector to deal with the class-imbalance """ super().__init__() self.per_image = per_img self.ignore = ignore def lovasz_hinge_flat(self, logits, labels): """ Binary Lovasz hinge loss logits: [P] Variable, logits at each prediction (between -\\infty and +\\infty) labels: [P] Tensor, binary ground truth labels (0 or 1) ignore: label to ignore """ if len(labels) == 0: return logits.sum() * 0.0 signs = 2.0 * labels.float() - 1.0 errors = 1.0 - logits * Variable(signs) errors_sorted, perm = torch.sort(errors, dim=0, descending=True) perm = perm.data gt_sorted = labels[perm] grad = lovasz_grad(gt_sorted) loss = torch.dot(F.relu(errors_sorted), Variable(grad)) return loss def forward(self, logits, labels): """ Binary Lovasz hinge loss logits: [B, H, W] Variable, logits at each pixel (between -\\infty and +\\infty) labels: [B, H, W] Tensor, binary ground truth masks (0 or 1) per_image: compute the loss per image instead of per batch ignore: void class id """ if self.per_image: loss = mean(self.lovasz_hinge_flat(*flatten_binary_scores(log. unsqueeze(0), lab.unsqueeze(0), self.ignore)) for log, lab in zip(logits, labels)) else: loss = self.lovasz_hinge_flat(*flatten_binary_scores(logits, labels, self.ignore)) 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 import torch.nn.parallel import torch.utils.data from torchvision.transforms import functional as F import torch.nn.functional as F from torch.autograd import Variable 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_ne_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 = 255.0 tmp2 = tmp0 != tmp1 tl.store(out_ptr0 + x0, tmp2, xmask) 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((256,), (1,), torch.bool) get_raw_stream(0) triton_poi_fused_ne_0[grid(256)](arg1_1, buf0, 256, XBLOCK=256, num_warps=4, num_stages=1) return reinterpret_tensor(arg0_1, (256,), (1,), 0 ), buf0, reinterpret_tensor(arg1_1, (256,), (1,), 0) def flatten_binary_scores(scores, labels, ignore=255): """ Flattens predictions in the batch (binary case) Remove labels equal to 'ignore' """ scores = scores.view(-1) labels = labels.view(-1) if ignore is None: return scores, labels valid = labels != ignore vscores = scores[valid] vlabels = labels[valid] return vscores, vlabels def lovasz_grad(gt_sorted): """ Computes gradient of the Lovasz extension w.r.t sorted errors See Alg. 1 in paper """ p = len(gt_sorted) gts = gt_sorted.sum() intersection = gts.float() - gt_sorted.float().cumsum(0) union = gts.float() + (1 - gt_sorted).float().cumsum(0) jaccard = 1.0 - intersection / union if p > 1: jaccard[1:p] = jaccard[1:p] - jaccard[0:-1] return jaccard def isnan(x): return x != x def mean(l, ignore_nan=False, empty=0): """ nanmean compatible with generators. """ l = iter(l) if ignore_nan: l = ifilterfalse(isnan, l) try: n = 1 acc = next(l) except StopIteration: if empty == 'raise': raise ValueError('Empty mean') return empty for n, v in enumerate(l, 2): acc += v if n == 1: return acc return acc / n class lovasz_hingeNew(torch.nn.Module): def __init__(self, per_img=False, ignore=255): """ :param weight: 1D weight vector to deal with the class-imbalance """ super().__init__() self.per_image = per_img self.ignore = ignore def lovasz_hinge_flat(self, logits, labels): """ Binary Lovasz hinge loss logits: [P] Variable, logits at each prediction (between -\\infty and +\\infty) labels: [P] Tensor, binary ground truth labels (0 or 1) ignore: label to ignore """ if len(labels) == 0: return logits.sum() * 0.0 signs = 2.0 * labels.float() - 1.0 errors = 1.0 - logits * Variable(signs) errors_sorted, perm = torch.sort(errors, dim=0, descending=True) perm = perm.data gt_sorted = labels[perm] grad = lovasz_grad(gt_sorted) loss = torch.dot(F.relu(errors_sorted), Variable(grad)) return loss def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0]
clovaai/ext_portrait_segmentation
lovasz_hinge
false
15,067
[ "MIT" ]
227
9bc1bada1cb7bd17a3a80a2964980f4b4befef5b
https://github.com/clovaai/ext_portrait_segmentation/tree/9bc1bada1cb7bd17a3a80a2964980f4b4befef5b
Linear
import torch import torch.nn as nn import torch.nn.functional as F from torch.nn.parameter import Parameter def keep_variance_fn(x): return x + 0.001 class Linear(nn.Module): def __init__(self, in_features, out_features, bias=True, keep_variance_fn=None): super(Linear, self).__init__() self._keep_variance_fn = keep_variance_fn self.in_features = in_features self.out_features = out_features self.weight = Parameter(torch.Tensor(out_features, in_features)) if bias: self.bias = Parameter(torch.Tensor(out_features)) else: self.register_parameter('bias', None) def forward(self, inputs_mean, inputs_variance): outputs_mean = F.linear(inputs_mean, self.weight, self.bias) outputs_variance = F.linear(inputs_variance, self.weight ** 2, None) if self._keep_variance_fn is not None: outputs_variance = self._keep_variance_fn(outputs_variance) return outputs_mean, outputs_variance def get_inputs(): return [torch.rand([4, 4, 4, 4]), 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 import torch.nn as nn from torch.nn.parameter 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_pow_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 + x0, xmask) tmp1 = tmp0 * tmp0 tl.store(out_ptr0 + x0, tmp1, 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,), (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)) 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((4, 4), (4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_pow_0[grid(16)](primals_1, buf1, 16, XBLOCK=16, num_warps=1, num_stages=1) buf2 = empty_strided_cuda((64, 4), (4, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(primals_4, (64, 4), (4, 1), 0), reinterpret_tensor(buf1, (4, 4), (1, 4), 0), out=buf2) del buf1 return reinterpret_tensor(buf0, (4, 4, 4, 4), (64, 16, 4, 1), 0 ), reinterpret_tensor(buf2, (4, 4, 4, 4), (64, 16, 4, 1), 0 ), primals_1, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0 ), reinterpret_tensor(primals_4, (64, 4), (4, 1), 0) def keep_variance_fn(x): return x + 0.001 class LinearNew(nn.Module): def __init__(self, in_features, out_features, bias=True, keep_variance_fn=None): super(LinearNew, self).__init__() self._keep_variance_fn = keep_variance_fn self.in_features = in_features self.out_features = out_features self.weight = Parameter(torch.Tensor(out_features, in_features)) if bias: self.bias = Parameter(torch.Tensor(out_features)) else: self.register_parameter('bias', None) def forward(self, input_0, input_1): primals_1 = self.weight primals_2 = self.bias primals_3 = input_0 primals_4 = input_1 output = call([primals_1, primals_2, primals_3, primals_4]) return output[0], output[1]
collector-m/LiDAR-MOS
Linear
false
15,068
[ "MIT" ]
268
7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
https://github.com/collector-m/LiDAR-MOS/tree/7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
Conv2d
import torch import torch.nn.functional as F from torch.nn.modules.conv import _ConvNd from torch.nn.modules.utils import _pair def keep_variance_fn(x): return x + 0.001 class Conv2d(_ConvNd): def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, keep_variance_fn=None, padding_mode='zeros'): self._keep_variance_fn = keep_variance_fn kernel_size = _pair(kernel_size) stride = _pair(stride) padding = _pair(padding) dilation = _pair(dilation) super(Conv2d, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, False, _pair(0), groups, bias, padding_mode) def forward(self, inputs_mean, inputs_variance): outputs_mean = F.conv2d(inputs_mean, self.weight, self.bias, self. stride, self.padding, self.dilation, self.groups) outputs_variance = F.conv2d(inputs_variance, self.weight ** 2, None, self.stride, self.padding, self.dilation, self.groups) if self._keep_variance_fn is not None: outputs_variance = self._keep_variance_fn(outputs_variance) return outputs_mean, outputs_variance def get_inputs(): return [torch.rand([4, 4, 4, 4]), 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 from torch.nn.modules.conv import _ConvNd from torch.nn.modules.utils import _pair 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_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 tl.store(in_out_ptr0 + x2, tmp2, xmask) @triton.jit def triton_poi_fused_pow_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 = tmp0 * tmp0 tl.store(out_ptr0 + x0, tmp1, 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,), (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)) 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=1, bias=None) assert_size_stride(buf0, (4, 4, 1, 1), (4, 1, 1, 1)) buf1 = buf0 del buf0 get_raw_stream(0) triton_poi_fused_convolution_0[grid(16)](buf1, primals_2, 16, XBLOCK=16, num_warps=1, num_stages=1) del primals_2 buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused_pow_1[grid(256)](primals_1, buf2, 256, XBLOCK=256, num_warps=4, num_stages=1) buf3 = extern_kernels.convolution(primals_4, buf2, 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)) return buf1, buf3, primals_1, primals_3, primals_4, buf2 def keep_variance_fn(x): return x + 0.001 class Conv2dNew(_ConvNd): def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, keep_variance_fn=None, padding_mode='zeros'): self._keep_variance_fn = keep_variance_fn kernel_size = _pair(kernel_size) stride = _pair(stride) padding = _pair(padding) dilation = _pair(dilation) super(Conv2dNew, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, False, _pair(0), groups, bias, padding_mode) def forward(self, input_0, input_1): primals_1 = self.weight primals_2 = self.bias primals_3 = input_0 primals_4 = input_1 output = call([primals_1, primals_2, primals_3, primals_4]) return output[0], output[1]
collector-m/LiDAR-MOS
Conv2d
false
15,069
[ "MIT" ]
268
7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
https://github.com/collector-m/LiDAR-MOS/tree/7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
MaxPool2d
import torch import numpy as np import torch.nn as nn from numbers import Number def keep_variance_fn(x): return x + 0.001 def normcdf(value, mu=0.0, stddev=1.0): sinv = 1.0 / stddev if isinstance(stddev, Number) else stddev.reciprocal() return 0.5 * (1.0 + torch.erf((value - mu) * sinv / np.sqrt(2.0))) def _normal_log_pdf(value, mu, stddev): var = stddev ** 2 log_scale = np.log(stddev) if isinstance(stddev, Number) else torch.log( stddev) return -(value - mu) ** 2 / (2.0 * var) - log_scale - np.log(np.sqrt( 2.0 * np.pi)) def normpdf(value, mu=0.0, stddev=1.0): return torch.exp(_normal_log_pdf(value, mu, stddev)) class MaxPool2d(nn.Module): def __init__(self, keep_variance_fn=None): super(MaxPool2d, self).__init__() self._keep_variance_fn = keep_variance_fn def _max_pool_internal(self, mu_a, mu_b, var_a, var_b): stddev = torch.sqrt(var_a + var_b) ab = mu_a - mu_b alpha = ab / stddev pdf = normpdf(alpha) cdf = normcdf(alpha) z_mu = stddev * pdf + ab * cdf + mu_b z_var = (mu_a + mu_b) * stddev * pdf + (mu_a ** 2 + var_a) * cdf + ( mu_b ** 2 + var_b) * (1.0 - cdf) - z_mu ** 2 if self._keep_variance_fn is not None: z_var = self._keep_variance_fn(z_var) return z_mu, z_var def _max_pool_1x2(self, inputs_mean, inputs_variance): mu_a = inputs_mean[:, :, :, 0::2] mu_b = inputs_mean[:, :, :, 1::2] var_a = inputs_variance[:, :, :, 0::2] var_b = inputs_variance[:, :, :, 1::2] outputs_mean, outputs_variance = self._max_pool_internal(mu_a, mu_b, var_a, var_b) return outputs_mean, outputs_variance def _max_pool_2x1(self, inputs_mean, inputs_variance): mu_a = inputs_mean[:, :, 0::2, :] mu_b = inputs_mean[:, :, 1::2, :] var_a = inputs_variance[:, :, 0::2, :] var_b = inputs_variance[:, :, 1::2, :] outputs_mean, outputs_variance = self._max_pool_internal(mu_a, mu_b, var_a, var_b) return outputs_mean, outputs_variance def forward(self, inputs_mean, inputs_variance): z_mean, z_variance = self._max_pool_1x2(inputs_mean, inputs_variance) outputs_mean, outputs_variance = self._max_pool_2x1(z_mean, z_variance) return outputs_mean, outputs_variance 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.triton_helpers import libdevice, math as tl_math import numpy as np import torch.nn as nn from numbers import Number 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_div_erf_exp_log_mul_neg_pow_rsub_sqrt_sub_0( in_out_ptr0, in_ptr0, in_ptr1, 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_ptr0 + 2 * x0, xmask, eviction_policy='evict_last') tmp1 = tl.load(in_ptr0 + (1 + 2 * x0), xmask, eviction_policy='evict_last') tmp3 = tl.load(in_ptr1 + 2 * x0, xmask, eviction_policy='evict_last') tmp4 = tl.load(in_ptr1 + (1 + 2 * x0), xmask, eviction_policy='evict_last') tmp2 = tmp0 + tmp1 tmp5 = tmp3 + tmp4 tmp6 = libdevice.sqrt(tmp5) tmp7 = tmp2 * tmp6 tmp8 = tmp0 - tmp1 tmp9 = tmp8 / tmp6 tmp10 = 0.0 tmp11 = tmp9 - tmp10 tmp12 = tmp11 * tmp11 tmp13 = -tmp12 tmp14 = 0.5 tmp15 = tmp13 * tmp14 tmp16 = 0.9189385332046727 tmp17 = tmp15 - tmp16 tmp18 = tl_math.exp(tmp17) tmp19 = tmp7 * tmp18 tmp20 = tmp0 * tmp0 tmp21 = tmp20 + tmp3 tmp22 = 1.0 tmp23 = tmp11 * tmp22 tmp24 = 1.414213562373095 tmp25 = tmp23 / tmp24 tmp26 = libdevice.erf(tmp25) tmp27 = tmp26 + tmp22 tmp28 = tmp27 * tmp14 tmp29 = tmp21 * tmp28 tmp30 = tmp19 + tmp29 tmp31 = tmp6 * tmp18 tmp32 = tmp8 * tmp28 tmp33 = tmp31 + tmp32 tmp34 = tmp33 + tmp1 tmp35 = tmp34 * tmp34 tmp36 = tmp1 * tmp1 tmp37 = tmp36 + tmp4 tmp38 = tmp22 - tmp28 tmp39 = tmp37 * tmp38 tmp40 = tmp30 + tmp39 tmp41 = tmp40 - tmp35 tl.store(in_out_ptr0 + x0, tmp41, xmask) @triton.jit def triton_poi_fused_add_div_erf_exp_log_mul_neg_pow_rsub_sqrt_sub_1( in_out_ptr0, in_ptr0, in_ptr1, in_ptr2, out_ptr2, xnumel, XBLOCK: tl. constexpr): xnumel = 64 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x0 = xindex % 2 x1 = xindex // 2 x2 = xindex tmp0 = tl.load(in_ptr0 + (4 + 2 * x0 + 8 * x1), xmask, eviction_policy= 'evict_last') tmp1 = tl.load(in_ptr0 + (5 + 2 * x0 + 8 * x1), xmask, eviction_policy= 'evict_last') tmp4 = tl.load(in_ptr1 + (4 + 2 * x0 + 8 * x1), xmask, eviction_policy= 'evict_last') tmp5 = tl.load(in_ptr1 + (5 + 2 * x0 + 8 * x1), xmask, eviction_policy= 'evict_last') tmp29 = tl.load(in_ptr0 + (2 * x0 + 8 * x1), xmask, eviction_policy= 'evict_last') tmp30 = tl.load(in_ptr0 + (1 + 2 * x0 + 8 * x1), xmask, eviction_policy ='evict_last') tmp33 = tl.load(in_ptr1 + (2 * x0 + 8 * x1), xmask, eviction_policy= 'evict_last') tmp34 = tl.load(in_ptr1 + (1 + 2 * x0 + 8 * x1), xmask, eviction_policy ='evict_last') tmp54 = tl.load(in_ptr2 + (x0 + 4 * x1), xmask) tmp55 = tl.load(in_ptr2 + (2 + x0 + 4 * x1), xmask) tmp2 = tmp0 + tmp1 tmp3 = libdevice.sqrt(tmp2) tmp6 = tmp4 - tmp5 tmp7 = tmp6 / tmp3 tmp8 = 0.0 tmp9 = tmp7 - tmp8 tmp10 = tmp9 * tmp9 tmp11 = -tmp10 tmp12 = 0.5 tmp13 = tmp11 * tmp12 tmp14 = 0.9189385332046727 tmp15 = tmp13 - tmp14 tmp16 = tl_math.exp(tmp15) tmp17 = tmp3 * tmp16 tmp18 = 1.0 tmp19 = tmp9 * tmp18 tmp20 = 1.414213562373095 tmp21 = tmp19 / tmp20 tmp22 = libdevice.erf(tmp21) tmp23 = tmp22 + tmp18 tmp24 = tmp23 * tmp12 tmp25 = tmp6 * tmp24 tmp26 = tmp17 + tmp25 tmp27 = tmp26 + tmp5 tmp28 = tmp27 * tmp27 tmp31 = tmp29 + tmp30 tmp32 = libdevice.sqrt(tmp31) tmp35 = tmp33 - tmp34 tmp36 = tmp35 / tmp32 tmp37 = tmp36 - tmp8 tmp38 = tmp37 * tmp37 tmp39 = -tmp38 tmp40 = tmp39 * tmp12 tmp41 = tmp40 - tmp14 tmp42 = tl_math.exp(tmp41) tmp43 = tmp32 * tmp42 tmp44 = tmp37 * tmp18 tmp45 = tmp44 / tmp20 tmp46 = libdevice.erf(tmp45) tmp47 = tmp46 + tmp18 tmp48 = tmp47 * tmp12 tmp49 = tmp35 * tmp48 tmp50 = tmp43 + tmp49 tmp51 = tmp50 + tmp34 tmp52 = tmp51 + tmp27 tmp53 = tmp51 - tmp27 tmp56 = tmp54 + tmp55 tmp57 = libdevice.sqrt(tmp56) tmp58 = tmp53 / tmp57 tmp59 = tmp58 - tmp8 tmp60 = tmp59 * tmp59 tmp61 = -tmp60 tmp62 = tmp61 * tmp12 tmp63 = tmp62 - tmp14 tmp64 = tl_math.exp(tmp63) tmp65 = tmp57 * tmp64 tmp66 = tmp59 * tmp18 tmp67 = tmp66 / tmp20 tmp68 = libdevice.erf(tmp67) tmp69 = tmp68 + tmp18 tmp70 = tmp69 * tmp12 tmp71 = tmp53 * tmp70 tmp72 = tmp65 + tmp71 tmp73 = tmp72 + tmp27 tmp74 = tmp51 * tmp51 tmp75 = tmp52 * tmp57 tmp76 = tmp75 * tmp64 tmp77 = tmp74 + tmp54 tmp78 = tmp77 * tmp70 tmp79 = tmp76 + tmp78 tmp80 = tmp28 + tmp55 tmp81 = tmp18 - tmp70 tmp82 = tmp80 * tmp81 tmp83 = tmp79 + tmp82 tmp84 = tmp73 * tmp73 tmp85 = tmp83 - tmp84 tl.store(out_ptr2 + x2, tmp73, xmask) tl.store(in_out_ptr0 + x2, tmp85, xmask) 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) buf1 = empty_strided_cuda((4, 4, 4, 2), (32, 8, 2, 1), torch.float32) buf3 = buf1 del buf1 get_raw_stream(0) triton_poi_fused_add_div_erf_exp_log_mul_neg_pow_rsub_sqrt_sub_0[grid (128)](buf3, arg0_1, arg1_1, 128, XBLOCK=128, num_warps=4, num_stages=1) buf0 = empty_strided_cuda((4, 4, 2, 2), (16, 4, 2, 1), torch.float32) buf8 = empty_strided_cuda((4, 4, 2, 2), (16, 4, 2, 1), torch.float32) buf6 = buf0 del buf0 buf9 = buf6 del buf6 triton_poi_fused_add_div_erf_exp_log_mul_neg_pow_rsub_sqrt_sub_1[grid (64)](buf9, arg1_1, arg0_1, buf3, buf8, 64, XBLOCK=64, num_warps=1, num_stages=1) del arg0_1 del arg1_1 del buf3 return buf8, buf9 def keep_variance_fn(x): return x + 0.001 def normcdf(value, mu=0.0, stddev=1.0): sinv = 1.0 / stddev if isinstance(stddev, Number) else stddev.reciprocal() return 0.5 * (1.0 + torch.erf((value - mu) * sinv / np.sqrt(2.0))) def _normal_log_pdf(value, mu, stddev): var = stddev ** 2 log_scale = np.log(stddev) if isinstance(stddev, Number) else torch.log( stddev) return -(value - mu) ** 2 / (2.0 * var) - log_scale - np.log(np.sqrt( 2.0 * np.pi)) def normpdf(value, mu=0.0, stddev=1.0): return torch.exp(_normal_log_pdf(value, mu, stddev)) class MaxPool2dNew(nn.Module): def __init__(self, keep_variance_fn=None): super(MaxPool2dNew, self).__init__() self._keep_variance_fn = keep_variance_fn def _max_pool_internal(self, mu_a, mu_b, var_a, var_b): stddev = torch.sqrt(var_a + var_b) ab = mu_a - mu_b alpha = ab / stddev pdf = normpdf(alpha) cdf = normcdf(alpha) z_mu = stddev * pdf + ab * cdf + mu_b z_var = (mu_a + mu_b) * stddev * pdf + (mu_a ** 2 + var_a) * cdf + ( mu_b ** 2 + var_b) * (1.0 - cdf) - z_mu ** 2 if self._keep_variance_fn is not None: z_var = self._keep_variance_fn(z_var) return z_mu, z_var def _max_pool_1x2(self, inputs_mean, inputs_variance): mu_a = inputs_mean[:, :, :, 0::2] mu_b = inputs_mean[:, :, :, 1::2] var_a = inputs_variance[:, :, :, 0::2] var_b = inputs_variance[:, :, :, 1::2] outputs_mean, outputs_variance = self._max_pool_internal(mu_a, mu_b, var_a, var_b) return outputs_mean, outputs_variance def _max_pool_2x1(self, inputs_mean, inputs_variance): mu_a = inputs_mean[:, :, 0::2, :] mu_b = inputs_mean[:, :, 1::2, :] var_a = inputs_variance[:, :, 0::2, :] var_b = inputs_variance[:, :, 1::2, :] outputs_mean, outputs_variance = self._max_pool_internal(mu_a, mu_b, var_a, var_b) return outputs_mean, outputs_variance def forward(self, input_0, input_1): arg0_1 = input_0 arg1_1 = input_1 output = call([arg0_1, arg1_1]) return output[0], output[1]
collector-m/LiDAR-MOS
MaxPool2d
false
15,070
[ "MIT" ]
268
7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
https://github.com/collector-m/LiDAR-MOS/tree/7ccbb63b4ee7c40195b35dd0dddd71473fae25b1
TransformerEncoderLayer
import torch import torch.utils.data import torch.nn.functional as F import torch.nn as nn from torch.nn.init import xavier_uniform_ from torch.nn.init import constant_ from torch.nn.init import xavier_normal_ def _get_activation_fn(activation): if activation == 'relu': return F.relu elif activation == 'gelu': return F.gelu raise RuntimeError('activation should be relu/gelu, not {}'.format( activation)) class MultiheadAttention(nn.Module): __annotations__ = {'bias_k': torch._jit_internal.Optional[torch.Tensor], 'bias_v': torch._jit_internal.Optional[torch.Tensor]} __constants__ = ['q_proj_weight', 'k_proj_weight', 'v_proj_weight', 'in_proj_weight'] def __init__(self, embed_dim, num_heads, dropout=0.0, bias=True, add_bias_kv=False, add_zero_attn=False, kdim=None, vdim=None): super(MultiheadAttention, self).__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_embed_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' if self._qkv_same_embed_dim is False: self.q_proj_weight = nn.Parameter(torch.Tensor(embed_dim, embed_dim)) self.k_proj_weight = nn.Parameter(torch.Tensor(embed_dim, self. kdim)) self.v_proj_weight = nn.Parameter(torch.Tensor(embed_dim, self. vdim)) self.register_parameter('in_proj_weight', None) else: self.in_proj_weight = nn.Parameter(torch.empty(3 * embed_dim, embed_dim)) self.register_parameter('q_proj_weight', None) self.register_parameter('k_proj_weight', None) self.register_parameter('v_proj_weight', None) if bias: self.in_proj_bias = nn.Parameter(torch.empty(3 * embed_dim)) else: self.register_parameter('in_proj_bias', None) self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias) if add_bias_kv: self.bias_k = nn.Parameter(torch.empty(1, 1, embed_dim)) self.bias_v = nn.Parameter(torch.empty(1, 1, embed_dim)) else: self.bias_k = self.bias_v = None self.add_zero_attn = add_zero_attn 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 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, key, value, key_padding_mask=None, need_weights=True, attn_mask=None): if not self._qkv_same_embed_dim: return F.multi_head_attention_forward(query, key, value, 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 F.multi_head_attention_forward(query, key, value, 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 TransformerEncoderLayer(nn.Module): def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1, activation='relu'): super(TransformerEncoderLayer, self).__init__() self.self_attn = MultiheadAttention(d_model, nhead, dropout=dropout) self.linear1 = nn.Linear(d_model, dim_feedforward) self.dropout = nn.Dropout(dropout) self.linear2 = nn.Linear(dim_feedforward, d_model) self.norm1 = nn.LayerNorm(d_model) self.norm2 = nn.LayerNorm(d_model) self.dropout1 = nn.Dropout(dropout) self.dropout2 = nn.Dropout(dropout) self.activation = _get_activation_fn(activation) def __setstate__(self, state): if 'activation' not in state: state['activation'] = F.relu super(TransformerEncoderLayer, self).__setstate__(state) def forward(self, src, src_mask=None, src_key_padding_mask=None): src2 = self.self_attn(src, src, src, 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])] 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 torch.utils.data import torch.nn.functional as F import torch.nn as nn from torch.nn.init import xavier_uniform_ from torch.nn.init import constant_ from torch.nn.init import xavier_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_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 = 1.0 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 = 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_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 = 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_3(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl. constexpr, XBLOCK: tl.constexpr): ynumel = 4 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 x1 = xindex y0 = yindex tmp0 = tl.load(in_ptr0 + (y0 + 4 * x1), xmask & ymask) tl.store(out_ptr0 + (x1 + 4 * y0), tmp0, xmask & ymask) @triton.jit def triton_poi_fused_add_native_layer_norm_4(in_ptr0, in_ptr1, out_ptr0, out_ptr1, 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 + 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_5(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, 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 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_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 % 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) tl.store(in_out_ptr0 + x2, tmp4, None) @triton.jit def triton_poi_fused_add_7(in_out_ptr0, in_ptr0, in_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_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_8(in_ptr0, out_ptr0, out_ptr1, 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 + 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_9(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, 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 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, 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((4, 4), (4, 1), torch.float32) extern_kernels.mm(primals_5, reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0) buf1 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.addmm(reinterpret_tensor(primals_2, (4,), (1,), 4), primals_5, reinterpret_tensor(primals_1, (4, 4), (1, 4), 16), alpha=1, beta=1, out=buf1) buf2 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.addmm(reinterpret_tensor(primals_2, (4,), (1,), 8), primals_5, reinterpret_tensor(primals_1, (4, 4), (1, 4), 32), alpha=1, beta=1, out=buf2) del primals_1 buf3 = reinterpret_tensor(buf0, (4, 4, 1), (1, 4, 16), 0) del buf0 get_raw_stream(0) triton_poi_fused_mul_0[grid(16)](buf3, primals_2, 16, XBLOCK=16, num_warps=1, num_stages=1) del primals_2 buf4 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) extern_kernels.bmm(buf3, reinterpret_tensor(buf1, (4, 1, 4), (1, 1, 4), 0), out=buf4) buf5 = empty_strided_cuda((4, 4, 4), (16, 4, 1), torch.float32) triton_poi_fused__softmax_1[grid(64)](buf4, buf5, 64, XBLOCK=64, num_warps=1, num_stages=1) buf6 = buf4 del buf4 triton_poi_fused__softmax_2[grid(64)](buf5, buf6, 64, XBLOCK=64, num_warps=1, num_stages=1) del buf5 buf7 = empty_strided_cuda((4, 4, 1), (4, 1, 1), torch.float32) extern_kernels.bmm(buf6, reinterpret_tensor(buf2, (4, 4, 1), (1, 4, 1), 0), out=buf7) buf8 = empty_strided_cuda((4, 4, 1), (4, 1, 1), torch.float32) triton_poi_fused_clone_3[grid(4, 4)](buf7, buf8, 4, 4, XBLOCK=4, YBLOCK=4, num_warps=1, num_stages=1) buf9 = reinterpret_tensor(buf7, (4, 4), (4, 1), 0) del buf7 extern_kernels.addmm(primals_4, reinterpret_tensor(buf8, (4, 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, 1), (1, 4), torch.float32) buf11 = empty_strided_cuda((4, 1), (1, 4), torch.float32) triton_poi_fused_add_native_layer_norm_4[grid(4)](primals_5, buf9, buf10, buf11, 4, XBLOCK=4, num_warps=1, num_stages=1) buf12 = empty_strided_cuda((4, 4), (4, 1), torch.float32) triton_poi_fused_add_native_layer_norm_5[grid(16)](primals_5, buf9, buf10, buf11, primals_6, primals_7, buf12, 16, XBLOCK=16, num_warps=1, num_stages=1) del primals_7 buf13 = empty_strided_cuda((4, 2048), (2048, 1), torch.float32) extern_kernels.mm(buf12, reinterpret_tensor(primals_8, (4, 2048), ( 1, 4), 0), out=buf13) buf14 = buf13 del buf13 triton_poi_fused_relu_6[grid(8192)](buf14, primals_9, 8192, XBLOCK= 256, num_warps=4, num_stages=1) del primals_9 buf15 = empty_strided_cuda((4, 4), (4, 1), torch.float32) extern_kernels.mm(buf14, reinterpret_tensor(primals_10, (2048, 4), (1, 2048), 0), out=buf15) buf16 = buf15 del buf15 triton_poi_fused_add_7[grid(16)](buf16, buf12, primals_11, 16, XBLOCK=16, num_warps=1, num_stages=1) del primals_11 buf17 = buf11 del buf11 buf18 = buf10 del buf10 triton_poi_fused_native_layer_norm_8[grid(4)](buf16, buf17, buf18, 4, XBLOCK=4, num_warps=1, num_stages=1) buf19 = empty_strided_cuda((4, 4), (4, 1), torch.float32) triton_poi_fused_native_layer_norm_9[grid(16)](buf16, buf17, buf18, primals_12, primals_13, buf19, 16, XBLOCK=16, 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, (4, 4), (4, 1), 0), buf9, buf12, buf14, buf16, primals_10, primals_8, primals_3, reinterpret_tensor(buf2, ( 4, 1, 4), (1, 1, 4), 0), reinterpret_tensor(buf3, (4, 1, 4), (1, 1, 4), 0), reinterpret_tensor(buf1, (4, 4, 1), (1, 4, 1), 0)) def _get_activation_fn(activation): if activation == 'relu': return F.relu elif activation == 'gelu': return F.gelu raise RuntimeError('activation should be relu/gelu, not {}'.format( activation)) class MultiheadAttention(nn.Module): __annotations__ = {'bias_k': torch._jit_internal.Optional[torch.Tensor], 'bias_v': torch._jit_internal.Optional[torch.Tensor]} __constants__ = ['q_proj_weight', 'k_proj_weight', 'v_proj_weight', 'in_proj_weight'] def __init__(self, embed_dim, num_heads, dropout=0.0, bias=True, add_bias_kv=False, add_zero_attn=False, kdim=None, vdim=None): super(MultiheadAttention, self).__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_embed_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' if self._qkv_same_embed_dim is False: self.q_proj_weight = nn.Parameter(torch.Tensor(embed_dim, embed_dim)) self.k_proj_weight = nn.Parameter(torch.Tensor(embed_dim, self. kdim)) self.v_proj_weight = nn.Parameter(torch.Tensor(embed_dim, self. vdim)) self.register_parameter('in_proj_weight', None) else: self.in_proj_weight = nn.Parameter(torch.empty(3 * embed_dim, embed_dim)) self.register_parameter('q_proj_weight', None) self.register_parameter('k_proj_weight', None) self.register_parameter('v_proj_weight', None) if bias: self.in_proj_bias = nn.Parameter(torch.empty(3 * embed_dim)) else: self.register_parameter('in_proj_bias', None) self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias) if add_bias_kv: self.bias_k = nn.Parameter(torch.empty(1, 1, embed_dim)) self.bias_v = nn.Parameter(torch.empty(1, 1, embed_dim)) else: self.bias_k = self.bias_v = None self.add_zero_attn = add_zero_attn 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 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, key, value, key_padding_mask=None, need_weights=True, attn_mask=None): if not self._qkv_same_embed_dim: return F.multi_head_attention_forward(query, key, value, 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 F.multi_head_attention_forward(query, key, value, 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 TransformerEncoderLayerNew(nn.Module): def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1, activation='relu'): super(TransformerEncoderLayerNew, self).__init__() self.self_attn = MultiheadAttention(d_model, nhead, dropout=dropout) self.linear1 = nn.Linear(d_model, dim_feedforward) self.dropout = nn.Dropout(dropout) self.linear2 = nn.Linear(dim_feedforward, d_model) self.norm1 = nn.LayerNorm(d_model) self.norm2 = nn.LayerNorm(d_model) self.dropout1 = nn.Dropout(dropout) self.dropout2 = nn.Dropout(dropout) self.activation = _get_activation_fn(activation) def __setstate__(self, state): if 'activation' not in state: state['activation'] = F.relu super(TransformerEncoderLayerNew, self).__setstate__(state) 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]
codeboy5/cvpr20-scatter-text-recognizer
TransformerEncoderLayer
false
15,071
[ "Apache-2.0" ]
63
4bd6cfbd4d7f64ce11864514f6b6b0646267c285
https://github.com/codeboy5/cvpr20-scatter-text-recognizer/tree/4bd6cfbd4d7f64ce11864514f6b6b0646267c285
_MLP_B
import torch import torch.nn as nn class _MLP_B(nn.Module): """MLP that only use age gender MMSE""" def __init__(self, in_size, drop_rate, fil_num): super(_MLP_B, self).__init__() self.fc1 = nn.Linear(in_size, fil_num) self.fc2 = nn.Linear(fil_num, 2) self.do1 = nn.Dropout(drop_rate) self.do2 = nn.Dropout(drop_rate) self.ac1 = nn.LeakyReLU() def forward(self, X): out = self.do1(X) out = self.fc1(out) out = self.ac1(out) out = self.do2(out) out = self.fc2(out) return out def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_size': 4, 'drop_rate': 0.5, 'fil_num': 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 = 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, (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, (2, 4), (4, 1)) assert_size_stride(primals_5, (2,), (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((4, 4, 4, 4), (64, 16, 4, 1), torch.bool) buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_leaky_relu_0[grid(256)](buf0, primals_3, buf1, buf2, 256, XBLOCK=128, num_warps=4, num_stages=1) del buf0 del primals_3 buf3 = empty_strided_cuda((64, 2), (2, 1), torch.float32) extern_kernels.addmm(primals_5, reinterpret_tensor(buf2, (64, 4), ( 4, 1), 0), reinterpret_tensor(primals_4, (4, 2), (1, 4), 0), alpha=1, beta=1, out=buf3) del primals_5 return reinterpret_tensor(buf3, (4, 4, 4, 2), (32, 8, 2, 1), 0 ), reinterpret_tensor(primals_1, (64, 4), (4, 1), 0 ), buf1, reinterpret_tensor(buf2, (64, 4), (4, 1), 0), primals_4 class _MLP_BNew(nn.Module): """MLP that only use age gender MMSE""" def __init__(self, in_size, drop_rate, fil_num): super(_MLP_BNew, self).__init__() self.fc1 = nn.Linear(in_size, fil_num) self.fc2 = nn.Linear(fil_num, 2) self.do1 = nn.Dropout(drop_rate) self.do2 = nn.Dropout(drop_rate) self.ac1 = nn.LeakyReLU() def forward(self, input_0): primals_2 = self.fc1.weight primals_3 = self.fc1.bias primals_4 = self.fc2.weight primals_5 = self.fc2.bias primals_1 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5]) return output[0]
colorfulbrain/brain2020
_MLP_B
false
15,072
[ "MIT" ]
91
1dde5d34fd2ba1f38bcc38f2c973d167c8c3a168
https://github.com/colorfulbrain/brain2020/tree/1dde5d34fd2ba1f38bcc38f2c973d167c8c3a168
Context2AnswerAttention
import torch import torch.nn as nn import torch.multiprocessing import torch.utils.data import torch.nn.modules.loss class Context2AnswerAttention(nn.Module): def __init__(self, dim, hidden_size): super(Context2AnswerAttention, self).__init__() self.linear_sim = nn.Linear(dim, hidden_size, bias=False) def forward(self, context, answers, out_answers, mask_answers=None): """ Parameters :context, (B, L, dim) :answers, (B, N, dim) :mask, (L, N) Returns :ques_emb, (L, dim) """ context_fc = torch.relu(self.linear_sim(context)) questions_fc = torch.relu(self.linear_sim(answers)) attention = torch.matmul(context_fc, questions_fc.transpose(-1, -2)) if mask_answers is not None: attention = attention.masked_fill(~mask_answers.unsqueeze(1). bool(), -Constants.INF) prob = torch.softmax(attention, dim=-1) emb = torch.matmul(prob, out_answers) return emb 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 [[], {'dim': 4, '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 from torch._inductor.runtime.triton_helpers import math as tl_math import torch.nn as nn import torch.multiprocessing import torch.utils.data import torch.nn.modules.loss 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, 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 = tl.full([1], 0, tl.int32) tmp2 = triton_helpers.maximum(tmp1, tmp0) tmp3 = 0.0 tmp4 = tmp2 <= tmp3 tl.store(in_out_ptr0 + x0, tmp2, xmask) tl.store(out_ptr0 + x0, tmp4, 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 = 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 = 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)) 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)) 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_2, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 4), (1, 4), 0), out=buf0) buf1 = 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=buf1) del primals_1 buf2 = 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)](buf2, buf9, 256, XBLOCK=256, num_warps=4, num_stages=1) buf3 = reinterpret_tensor(buf1, (4, 4, 4, 4), (64, 16, 4, 1), 0) del buf1 buf8 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool) triton_poi_fused_relu_threshold_backward_0[grid(256)](buf3, buf8, 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, 1, 4), 0), out=buf4) buf5 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused__softmax_1[grid(256)](buf4, buf5, 256, XBLOCK=128, num_warps=4, num_stages=1) buf6 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused__softmax_2[grid(256)](buf5, buf6, 256, XBLOCK=128, num_warps=4, num_stages=1) buf7 = reinterpret_tensor(buf5, (16, 4, 4), (16, 4, 1), 0) del buf5 extern_kernels.bmm(reinterpret_tensor(buf6, (16, 4, 4), (16, 4, 1), 0), reinterpret_tensor(primals_4, (16, 4, 4), (16, 4, 1), 0), out=buf7) del buf6 return reinterpret_tensor(buf7, (4, 4, 4, 4), (64, 16, 4, 1), 0 ), reinterpret_tensor(primals_2, (64, 4), (4, 1), 0 ), reinterpret_tensor(primals_3, (64, 4), (4, 1), 0 ), buf4, reinterpret_tensor(primals_4, (16, 4, 4), (16, 1, 4), 0 ), reinterpret_tensor(buf2, (16, 4, 4), (16, 1, 4), 0 ), reinterpret_tensor(buf3, (16, 4, 4), (16, 4, 1), 0), buf8, buf9 class Context2AnswerAttentionNew(nn.Module): def __init__(self, dim, hidden_size): super(Context2AnswerAttentionNew, self).__init__() self.linear_sim = nn.Linear(dim, hidden_size, bias=False) def forward(self, input_0, input_1, input_2): primals_1 = self.linear_sim.weight primals_2 = input_0 primals_3 = input_1 primals_4 = input_2 output = call([primals_1, primals_2, primals_3, primals_4]) return output[0]
cminusQAQ/graph4nlp
Context2AnswerAttention
false
15,073
[ "Apache-2.0" ]
1,269
d980e897131f1b9d3766750c06316d94749904fa
https://github.com/cminusQAQ/graph4nlp/tree/d980e897131f1b9d3766750c06316d94749904fa
HardMish
import torch from torch import nn class HardMish(nn.Module): def __init__(self): super().__init__() def forward(self, x): return x / 2 * torch.clamp(x + 2, min=0, max=2) 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 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_clamp_div_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 = 0.5 tmp2 = tmp0 * tmp1 tmp3 = 2.0 tmp4 = tmp0 + tmp3 tmp5 = 0.0 tmp6 = triton_helpers.maximum(tmp4, tmp5) tmp7 = triton_helpers.minimum(tmp6, tmp3) tmp8 = tmp2 * 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_div_mul_0[grid(256)](arg0_1, buf0, 256, XBLOCK=128, num_warps=4, num_stages=1) del arg0_1 return buf0, class HardMishNew(nn.Module): def __init__(self): super().__init__() def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0]
cooked-sashimi/Yet-Another-YOLOv4-Pytorch
HardMish
false
15,074
[ "MIT" ]
133
c884ef8849987a75b0e17eba1b739c22d3782e90
https://github.com/cooked-sashimi/Yet-Another-YOLOv4-Pytorch/tree/c884ef8849987a75b0e17eba1b739c22d3782e90
_MLP_C
import torch import torch.nn as nn class _MLP_C(nn.Module): """MLP that use DPMs from fcn and age, gender and MMSE""" def __init__(self, in_size, drop_rate, fil_num): super(_MLP_C, self).__init__() self.fc1 = nn.Linear(in_size, fil_num) self.fc2 = nn.Linear(fil_num, 2) self.do1 = nn.Dropout(drop_rate) self.do2 = nn.Dropout(drop_rate) self.ac1 = nn.LeakyReLU() def forward(self, X1, X2): X = torch.cat((X1, X2), 1) out = self.do1(X) out = self.fc1(out) out = self.ac1(out) out = self.do2(out) out = self.fc2(out) return out def get_inputs(): return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_size': 4, 'drop_rate': 0.5, 'fil_num': 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_cat_0(in_ptr0, in_ptr1, out_ptr0, xnumel, XBLOCK: tl. constexpr): xnumel = 512 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x1 = xindex // 16 % 8 x0 = xindex % 16 x2 = xindex // 128 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 + (x0 + 16 * x1 + 64 * x2), tmp4 & xmask, other=0.0) tmp6 = tmp0 >= tmp3 tl.full([1], 8, tl.int64) tmp9 = tl.load(in_ptr1 + (x0 + 16 * (-4 + x1) + 64 * 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_leaky_relu_1(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK: tl.constexpr): xnumel = 512 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, 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, 4), (4, 1)) assert_size_stride(primals_4, (4,), (1,)) assert_size_stride(primals_5, (2, 4), (4, 1)) assert_size_stride(primals_6, (2,), (1,)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = empty_strided_cuda((4, 8, 4, 4), (128, 16, 4, 1), torch.float32) get_raw_stream(0) triton_poi_fused_cat_0[grid(512)](primals_1, primals_2, buf0, 512, XBLOCK=256, num_warps=4, num_stages=1) del primals_1 del primals_2 buf1 = empty_strided_cuda((128, 4), (4, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(buf0, (128, 4), (4, 1), 0), reinterpret_tensor(primals_3, (4, 4), (1, 4), 0), out=buf1) del primals_3 buf2 = empty_strided_cuda((4, 8, 4, 4), (128, 16, 4, 1), torch.bool) buf3 = empty_strided_cuda((4, 8, 4, 4), (128, 16, 4, 1), torch.float32) triton_poi_fused_leaky_relu_1[grid(512)](buf1, primals_4, buf2, buf3, 512, XBLOCK=128, num_warps=4, num_stages=1) del buf1 del primals_4 buf4 = empty_strided_cuda((128, 2), (2, 1), torch.float32) extern_kernels.addmm(primals_6, reinterpret_tensor(buf3, (128, 4), (4, 1), 0), reinterpret_tensor(primals_5, (4, 2), (1, 4), 0), alpha=1, beta=1, out=buf4) del primals_6 return reinterpret_tensor(buf4, (4, 8, 4, 2), (64, 8, 2, 1), 0 ), reinterpret_tensor(buf0, (128, 4), (4, 1), 0 ), buf2, reinterpret_tensor(buf3, (128, 4), (4, 1), 0), primals_5 class _MLP_CNew(nn.Module): """MLP that use DPMs from fcn and age, gender and MMSE""" def __init__(self, in_size, drop_rate, fil_num): super(_MLP_CNew, self).__init__() self.fc1 = nn.Linear(in_size, fil_num) self.fc2 = nn.Linear(fil_num, 2) self.do1 = nn.Dropout(drop_rate) self.do2 = nn.Dropout(drop_rate) self.ac1 = nn.LeakyReLU() def forward(self, input_0, input_1): primals_3 = self.fc1.weight primals_4 = self.fc1.bias primals_5 = self.fc2.weight primals_6 = self.fc2.bias primals_1 = input_0 primals_2 = input_1 output = call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6]) return output[0]
colorfulbrain/brain2020
_MLP_C
false
15,075
[ "MIT" ]
91
1dde5d34fd2ba1f38bcc38f2c973d167c8c3a168
https://github.com/colorfulbrain/brain2020/tree/1dde5d34fd2ba1f38bcc38f2c973d167c8c3a168
DarknetMish
import torch import torch.nn.functional as F from torch import nn class darknet_mish(torch.autograd.Function): """ We can implement our own custom autograd Functions by subclassing torch.autograd.Function and implementing the forward and backward passes which operate on Tensors. """ @staticmethod def forward(ctx, input): """ In the forward pass we receive a Tensor containing the input and return a Tensor containing the output. ctx is a context object that can be used to stash information for backward computation. You can cache arbitrary objects for use in the backward pass using the ctx.save_for_backward method. """ ctx.save_for_backward(input) e = torch.exp(input) n = e * e + 2 * e mask = input <= -0.6 input[mask] = (input * (n / (n + 2)))[mask] input[~mask] = (input - 2 * (input / (n + 2)))[~mask] return input @staticmethod def backward(ctx, grad_output): """ In the backward pass we receive a Tensor containing the gradient of the loss with respect to the output, and we need to compute the gradient of the loss with respect to the input. """ input, = ctx.saved_tensors sp = F.softplus(input) grad_sp = -torch.expm1(sp) tsp = F.tanh(sp) grad_tsp = (1 - tsp * tsp) * grad_sp grad = input * grad_tsp + tsp return grad class DarknetMish(nn.Module): def __init__(self): super().__init__() def forward(self, x): return darknet_mish.apply(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 import torch.nn.functional as F 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_div_exp_le_mul_0(in_ptr0, out_ptr0, out_ptr1, out_ptr2, 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.exp(tmp0) tmp2 = tmp1 * tmp1 tmp3 = 2.0 tmp4 = tmp1 * tmp3 tmp5 = tmp2 + tmp4 tmp6 = tmp5 + tmp3 tmp7 = tmp5 / tmp6 tmp8 = tmp0 * tmp7 tmp9 = -0.6 tmp10 = tmp0 <= tmp9 tl.store(out_ptr0 + x0, tmp5, xmask) tl.store(out_ptr1 + x0, tmp8, xmask) tl.store(out_ptr2 + x0, 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, 4, 4), (64, 16, 4, 1), torch.float32) buf1 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.bool) get_raw_stream(0) triton_poi_fused_add_div_exp_le_mul_0[grid(256)](arg0_1, buf0, buf1, buf2, 256, XBLOCK=256, num_warps=4, num_stages=1) del arg0_1 return buf1, buf2, buf0 class darknet_mish(torch.autograd.Function): """ We can implement our own custom autograd Functions by subclassing torch.autograd.Function and implementing the forward and backward passes which operate on Tensors. """ @staticmethod def forward(ctx, input): """ In the forward pass we receive a Tensor containing the input and return a Tensor containing the output. ctx is a context object that can be used to stash information for backward computation. You can cache arbitrary objects for use in the backward pass using the ctx.save_for_backward method. """ ctx.save_for_backward(input) e = torch.exp(input) n = e * e + 2 * e mask = input <= -0.6 input[mask] = (input * (n / (n + 2)))[mask] input[~mask] = (input - 2 * (input / (n + 2)))[~mask] return input @staticmethod def backward(ctx, grad_output): """ In the backward pass we receive a Tensor containing the gradient of the loss with respect to the output, and we need to compute the gradient of the loss with respect to the input. """ input, = ctx.saved_tensors sp = F.softplus(input) grad_sp = -torch.expm1(sp) tsp = F.tanh(sp) grad_tsp = (1 - tsp * tsp) * grad_sp grad = input * grad_tsp + tsp return grad class DarknetMishNew(nn.Module): def __init__(self): super().__init__() def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0]
cooked-sashimi/Yet-Another-YOLOv4-Pytorch
DarknetMish
false
15,076
[ "MIT" ]
133
c884ef8849987a75b0e17eba1b739c22d3782e90
https://github.com/cooked-sashimi/Yet-Another-YOLOv4-Pytorch/tree/c884ef8849987a75b0e17eba1b739c22d3782e90
Tanh2
import torch import torch.utils.data import torch.nn as nn import torch.nn.parallel import torch.optim class Tanh2(nn.Module): def __init__(self): super(Tanh2, self).__init__() self.tanh = nn.Tanh() def forward(self, x): return (self.tanh(x) + 1) / 2 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 import torch.utils.data import torch.nn as nn import torch.nn.parallel import torch.optim 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_div_tanh_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.tanh(tmp0) tmp2 = 1.0 tmp3 = tmp1 + tmp2 tmp4 = 0.5 tmp5 = tmp3 * tmp4 tl.store(out_ptr0 + x0, tmp5, 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_div_tanh_0[grid(256)](arg0_1, buf0, 256, XBLOCK=256, num_warps=4, num_stages=1) del arg0_1 return buf0, class Tanh2New(nn.Module): def __init__(self): super(Tanh2New, self).__init__() self.tanh = nn.Tanh() def forward(self, input_0): arg0_1 = input_0 output = call([arg0_1]) return output[0]
csyxwei/FFWM
Tanh2
false
15,077
[ "MIT" ]
83
d42c578cabe1b81c6b1bb0c3cb707b190fca3c68
https://github.com/csyxwei/FFWM/tree/d42c578cabe1b81c6b1bb0c3cb707b190fca3c68
SAM
import torch from torch import nn class SAM(nn.Module): def __init__(self, in_channels): super().__init__() self.conv = nn.Conv2d(in_channels, out_channels=1, kernel_size=1) def forward(self, x): spatial_features = self.conv(x) attention = torch.sigmoid(spatial_features) return attention.expand_as(x) * x def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_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 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_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 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_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, primals_3 = args args.clear() assert_size_stride(primals_1, (1, 4, 1, 1), (4, 1, 1, 1)) assert_size_stride(primals_2, (1,), (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 = extern_kernels.convolution(primals_3, 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, 1, 4, 4), (16, 16, 4, 1)) buf1 = buf0 del buf0 get_raw_stream(0) triton_poi_fused_convolution_0[grid(64)](buf1, primals_2, 64, XBLOCK=64, num_warps=1, num_stages=1) del primals_2 buf2 = empty_strided_cuda((4, 4, 4, 4), (64, 16, 4, 1), torch.float32) triton_poi_fused_mul_1[grid(256)](buf1, primals_3, buf2, 256, XBLOCK=256, num_warps=4, num_stages=1) return buf2, primals_1, primals_3, buf1 class SAMNew(nn.Module): def __init__(self, in_channels): super().__init__() self.conv = nn.Conv2d(in_channels, out_channels=1, kernel_size=1) def forward(self, input_0): primals_1 = self.conv.weight primals_2 = self.conv.bias primals_3 = input_0 output = call([primals_1, primals_2, primals_3]) return output[0]
cooked-sashimi/Yet-Another-YOLOv4-Pytorch
SAM
false
15,078
[ "MIT" ]
133
c884ef8849987a75b0e17eba1b739c22d3782e90
https://github.com/cooked-sashimi/Yet-Another-YOLOv4-Pytorch/tree/c884ef8849987a75b0e17eba1b739c22d3782e90
GlobalAttentionGeneral
import torch import torch.nn as nn import torch.nn.parallel def conv1x1(in_planes, out_planes, bias=False): """1x1 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1, padding=0, bias=bias) class GlobalAttentionGeneral(nn.Module): def __init__(self, idf, cdf): super(GlobalAttentionGeneral, self).__init__() self.conv_context = conv1x1(cdf, idf) self.sm = nn.Softmax(dim=1) self.mask = None def applyMask(self, mask): self.mask = mask def forward(self, input, context): """ input: batch x idf x ih x iw (queryL=ihxiw) context: batch x cdf x sourceL """ ih, iw = input.size(2), input.size(3) queryL = ih * iw batch_size, sourceL = context.size(0), context.size(2) target = input.view(batch_size, -1, queryL) targetT = torch.transpose(target, 1, 2).contiguous() sourceT = context.unsqueeze(3) sourceT = self.conv_context(sourceT).squeeze(3) attn = torch.bmm(targetT, sourceT) attn = attn.view(batch_size * queryL, sourceL) if self.mask is not None: mask = self.mask.repeat(queryL, 1) attn.data.masked_fill_(mask.data, -float('inf')) attn = self.sm(attn) attn = attn.view(batch_size, queryL, sourceL) attn = torch.transpose(attn, 1, 2).contiguous() weightedContext = torch.bmm(sourceT, attn) weightedContext = weightedContext.view(batch_size, -1, ih, iw) attn = attn.view(batch_size, -1, ih, iw) return weightedContext, attn def get_inputs(): return [torch.rand([4, 4, 4, 4]), torch.rand([4, 4, 4])] def get_init_inputs(): return [[], {'idf': 4, 'cdf': 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.nn.parallel 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_transpose_0(in_ptr0, out_ptr0, out_ptr1, ynumel, xnumel, YBLOCK: tl.constexpr, XBLOCK: tl.constexpr): ynumel = 16 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 y2 = yindex % 4 y3 = yindex // 4 tmp0 = tl.load(in_ptr0 + (x1 + 16 * y0), xmask & ymask) tl.store(out_ptr0 + (x1 + 16 * y0), tmp0, xmask & ymask) tl.store(out_ptr1 + (y2 + 4 * x1 + 64 * 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_clone_2(in_ptr0, out_ptr0, ynumel, xnumel, YBLOCK: tl. constexpr, XBLOCK: tl.constexpr): ynumel = 16 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 x2 = xindex y0 = yindex % 4 y1 = yindex // 4 y3 = yindex tmp0 = tl.load(in_ptr0 + (y0 + 4 * x2 + 64 * y1), xmask & ymask) tmp1 = tl.load(in_ptr0 + (4 * x2 + 64 * y1), xmask & ymask, eviction_policy='evict_last') tmp2 = tl.load(in_ptr0 + (1 + 4 * x2 + 64 * y1), xmask & ymask, eviction_policy='evict_last') tmp4 = tl.load(in_ptr0 + (2 + 4 * x2 + 64 * y1), xmask & ymask, eviction_policy='evict_last') tmp6 = tl.load(in_ptr0 + (3 + 4 * x2 + 64 * y1), xmask & ymask, eviction_policy='evict_last') tmp3 = tmp1 + tmp2 tmp5 = tmp3 + tmp4 tmp7 = tmp5 + tmp6 tmp8 = tmp0 / tmp7 tl.store(out_ptr0 + (x2 + 16 * y3), tmp8, xmask & ymask) 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, 4), (16, 4, 1)) assert_size_stride(primals_3, (4, 4, 1, 1), (4, 1, 1, 1)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = extern_kernels.convolution(reinterpret_tensor(primals_2, (4, 4, 4, 1), (16, 4, 1, 1), 0), primals_3, 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, 1), (16, 4, 1, 1)) buf1 = empty_strided_cuda((4, 16, 4), (64, 1, 16), torch.float32) buf6 = empty_strided_cuda((4, 4, 16), (64, 1, 4), torch.float32) get_raw_stream(0) triton_poi_fused_clone_transpose_0[grid(16, 16)](primals_1, buf1, buf6, 16, 16, XBLOCK=16, YBLOCK=16, num_warps=4, num_stages=1) del primals_1 buf2 = empty_strided_cuda((4, 16, 4), (64, 4, 1), torch.float32) extern_kernels.bmm(buf1, reinterpret_tensor(buf0, (4, 4, 4), (16, 4, 1), 0), out=buf2) buf3 = reinterpret_tensor(buf1, (64, 4), (4, 1), 0) del buf1 triton_poi_fused__softmax_1[grid(256)](buf2, buf3, 256, XBLOCK=128, num_warps=4, num_stages=1) buf4 = empty_strided_cuda((4, 4, 16), (64, 16, 1), torch.float32) triton_poi_fused_clone_2[grid(16, 16)](buf3, buf4, 16, 16, XBLOCK= 16, YBLOCK=16, num_warps=4, num_stages=1) buf5 = reinterpret_tensor(buf3, (4, 4, 16), (64, 16, 1), 0) del buf3 extern_kernels.bmm(reinterpret_tensor(buf0, (4, 4, 4), (16, 4, 1), 0), buf4, out=buf5) return reinterpret_tensor(buf5, (4, 4, 4, 4), (64, 16, 4, 1), 0 ), reinterpret_tensor(buf4, (4, 4, 4, 4), (64, 16, 4, 1), 0 ), primals_3, reinterpret_tensor(primals_2, (4, 4, 4, 1), (16, 4, 1, 1), 0), buf2, reinterpret_tensor(buf0, (4, 4, 4), (16, 1, 4), 0 ), reinterpret_tensor(buf4, (4, 16, 4), (64, 1, 16), 0), buf6 def conv1x1(in_planes, out_planes, bias=False): """1x1 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=1, padding=0, bias=bias) class GlobalAttentionGeneralNew(nn.Module): def __init__(self, idf, cdf): super(GlobalAttentionGeneralNew, self).__init__() self.conv_context = conv1x1(cdf, idf) self.sm = nn.Softmax(dim=1) self.mask = None def applyMask(self, mask): self.mask = mask def forward(self, input_0, input_1): primals_3 = self.conv_context.weight primals_1 = input_0 primals_2 = input_1 output = call([primals_1, primals_2, primals_3]) return output[0], output[1]
comtalyst/multi-gan-material-defects
GlobalAttentionGeneral
false
15,079
[ "MIT" ]
112
aa1c9d4b918b5b5ad7f5fe03fdceec91a66e1007
https://github.com/comtalyst/multi-gan-material-defects/tree/aa1c9d4b918b5b5ad7f5fe03fdceec91a66e1007
Attention
import torch from torch import nn class Attention(nn.Module): def __init__(self, in_channels): super(Attention, self).__init__() self.out_channels = int(in_channels / 2) self.conv1 = nn.Conv2d(in_channels, self.out_channels, kernel_size= 3, padding=1, stride=1) self.relu1 = nn.ReLU() self.conv2 = nn.Conv2d(self.out_channels, self.out_channels, kernel_size=3, padding=1, stride=1) self.relu2 = nn.ReLU() self.conv3 = nn.Conv2d(self.out_channels, 4, kernel_size=1, padding =0, stride=1) self.sigmod = nn.Sigmoid() def forward(self, x): out = self.conv1(x) out = self.relu1(out) out = self.conv2(out) out = self.relu2(out) out = self.conv3(out) out = self.sigmod(out) return out def get_inputs(): return [torch.rand([4, 4, 4, 4])] def get_init_inputs(): return [[], {'in_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 from torch import nn assert_size_stride = torch._C._dynamo.guards.assert_size_stride @triton.jit def triton_poi_fused_convolution_relu_0(in_out_ptr0, in_ptr0, xnumel, XBLOCK: tl.constexpr): xnumel = 128 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x3 = xindex x1 = xindex // 16 % 2 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_sigmoid_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 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 = tl.sigmoid(tmp2) tl.store(in_out_ptr0 + x3, 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, (2, 4, 3, 3), (36, 9, 3, 1)) assert_size_stride(primals_2, (2,), (1,)) assert_size_stride(primals_3, (4, 4, 4, 4), (64, 16, 4, 1)) assert_size_stride(primals_4, (2, 2, 3, 3), (18, 9, 3, 1)) assert_size_stride(primals_5, (2,), (1,)) assert_size_stride(primals_6, (4, 2, 1, 1), (2, 1, 1, 1)) assert_size_stride(primals_7, (4,), (1,)) with torch.cuda._DeviceGuard(0): torch.cuda.set_device(0) buf0 = extern_kernels.convolution(primals_3, primals_1, stride=(1, 1), padding=(1, 1), 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 = buf0 del buf0 get_raw_stream(0) triton_poi_fused_convolution_relu_0[grid(128)](buf1, primals_2, 128, XBLOCK=128, num_warps=4, num_stages=1) del primals_2 buf2 = extern_kernels.convolution(buf1, primals_4, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf2, (4, 2, 4, 4), (32, 16, 4, 1)) buf3 = buf2 del buf2 triton_poi_fused_convolution_relu_0[grid(128)](buf3, primals_5, 128, XBLOCK=128, num_warps=4, num_stages=1) del primals_5 buf4 = extern_kernels.convolution(buf3, primals_6, stride=(1, 1), padding=(0, 0), 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 = buf4 del buf4 triton_poi_fused_convolution_sigmoid_1[grid(256)](buf5, primals_7, 256, XBLOCK=256, num_warps=4, num_stages=1) del primals_7 return buf5, primals_1, primals_3, primals_4, primals_6, buf1, buf3, buf5 class AttentionNew(nn.Module): def __init__(self, in_channels): super(AttentionNew, self).__init__() self.out_channels = int(in_channels / 2) self.conv1 = nn.Conv2d(in_channels, self.out_channels, kernel_size= 3, padding=1, stride=1) self.relu1 = nn.ReLU() self.conv2 = nn.Conv2d(self.out_channels, self.out_channels, kernel_size=3, padding=1, stride=1) self.relu2 = nn.ReLU() self.conv3 = nn.Conv2d(self.out_channels, 4, kernel_size=1, padding =0, stride=1) self.sigmod = nn.Sigmoid() 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_3 = input_0 output = call([primals_1, primals_2, primals_3, primals_4, primals_5, primals_6, primals_7]) return output[0]
createnewdemo/SPANet
Attention
false
15,080
[ "BSD-3-Clause" ]
177
86cfb05d1778cf30142ef30692e995a5b7b59bb8
https://github.com/createnewdemo/SPANet/tree/86cfb05d1778cf30142ef30692e995a5b7b59bb8
Bottleneck
import torch from torch import nn from collections import OrderedDict class Bottleneck(nn.Module): def __init__(self, in_channels, out_channels): super(Bottleneck, self).__init__() m = OrderedDict() m['conv1'] = nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) m['relu1'] = nn.ReLU(True) m['conv2'] = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=2, bias=False, dilation=2) m['relu2'] = nn.ReLU(True) m['conv3'] = nn.Conv2d(out_channels, out_channels, kernel_size=1, bias=False) self.group1 = nn.Sequential(m) self.relu = nn.Sequential(nn.ReLU(True)) def forward(self, x): out = self.group1(x) return out def get_inputs(): return [torch.rand([4, 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 from torch import nn from collections import OrderedDict assert_size_stride = torch._C._dynamo.guards.assert_size_stride @triton.jit def triton_poi_fused_relu_0(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 = tl.full([1], 0, tl.int32) tmp2 = triton_helpers.maximum(tmp1, tmp0) tl.store(in_out_ptr0 + x0, tmp2, xmask) def call(args): primals_1, primals_2, primals_3, primals_4 = args args.clear() assert_size_stride(primals_1, (4, 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, (4, 4, 3, 3), (36, 9, 3, 1)) assert_size_stride(primals_4, (4, 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, 4, 4, 4), (64, 16, 4, 1)) buf1 = buf0 del buf0 get_raw_stream(0) triton_poi_fused_relu_0[grid(256)](buf1, 256, XBLOCK=128, num_warps =4, num_stages=1) buf2 = extern_kernels.convolution(buf1, primals_3, stride=(1, 1), padding=(2, 2), dilation=(2, 2), transposed=False, output_padding=(0, 0), groups=1, bias=None) assert_size_stride(buf2, (4, 4, 4, 4), (64, 16, 4, 1)) buf3 = buf2 del buf2 triton_poi_fused_relu_0[grid(256)](buf3, 256, XBLOCK=128, num_warps =4, num_stages=1) buf4 = extern_kernels.convolution(buf3, primals_4, stride=(1, 1), padding=(0, 0), 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)) return buf4, primals_1, primals_2, primals_3, primals_4, buf1, buf3 class BottleneckNew(nn.Module): def __init__(self, in_channels, out_channels): super(BottleneckNew, self).__init__() m = OrderedDict() m['conv1'] = nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) m['relu1'] = nn.ReLU(True) m['conv2'] = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=2, bias=False, dilation=2) m['relu2'] = nn.ReLU(True) m['conv3'] = nn.Conv2d(out_channels, out_channels, kernel_size=1, bias=False) self.group1 = nn.Sequential(m) self.relu = nn.Sequential(nn.ReLU(True)) def forward(self, input_0): primals_1 = self.group1.conv1.weight primals_3 = self.group1.conv2.weight primals_4 = self.group1.conv3.weight primals_2 = input_0 output = call([primals_1, primals_2, primals_3, primals_4]) return output[0]
createnewdemo/SPANet
Bottleneck
false
15,081
[ "BSD-3-Clause" ]
177
86cfb05d1778cf30142ef30692e995a5b7b59bb8
https://github.com/createnewdemo/SPANet/tree/86cfb05d1778cf30142ef30692e995a5b7b59bb8
FeatureExtractFF
import torch import torch.utils.data import torch.nn as nn class FeatureExtractFF(nn.Module): def __init__(self, input_dim, hidden_sizes=(15,), activation_fn=nn.ReLU, **activation_args): super(FeatureExtractFF, self).__init__() self._in = input_dim self._hidden_sizes = hidden_sizes self._activation_fn = activation_fn self._activation_args = activation_args self.feature = nn.Sequential() hin = self._in for i, h in enumerate(self._hidden_sizes): self.feature.add_module(f'f_fc{i}', nn.Linear(hin, h)) self.feature.add_module(f'f_{activation_fn.__name__}{i}', activation_fn(**activation_args)) hin = h self._out_features = hin def forward(self, input_data): return self.feature(input_data) def extra_repr(self): return f'FC: {self.hidden_sizes}x{self._activation_fn.__name__}' def hidden_layer(self, index=0): return self.feature[index * 2] def output_size(self): return self._out_features 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 import triton_helpers import torch.utils.data 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 = 960 xoffset = tl.program_id(0) * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:] xmask = xindex < xnumel x2 = xindex x0 = xindex % 15 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) def call(args): primals_1, primals_2, primals_3 = args args.clear() assert_size_stride(primals_1, (15, 4), (4, 1)) assert_size_stride(primals_2, (15,), (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, 15), (15, 1), torch.float32) extern_kernels.mm(reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), reinterpret_tensor(primals_1, (4, 15), (1, 4), 0), out=buf0) del primals_1 buf1 = reinterpret_tensor(buf0, (4, 4, 4, 15), (240, 60, 15, 1), 0) del buf0 buf2 = empty_strided_cuda((4, 4, 4, 15), (240, 60, 15, 1), torch.bool) get_raw_stream(0) triton_poi_fused_relu_threshold_backward_0[grid(960)](buf1, primals_2, buf2, 960, XBLOCK=256, num_warps=4, num_stages=1) del primals_2 return buf1, reinterpret_tensor(primals_3, (64, 4), (4, 1), 0), buf2 class FeatureExtractFFNew(nn.Module): def __init__(self, input_dim, hidden_sizes=(15,), activation_fn=nn.ReLU, **activation_args): super(FeatureExtractFFNew, self).__init__() self._in = input_dim self._hidden_sizes = hidden_sizes self._activation_fn = activation_fn self._activation_args = activation_args self.feature = nn.Sequential() hin = self._in for i, h in enumerate(self._hidden_sizes): self.feature.add_module(f'f_fc{i}', nn.Linear(hin, h)) self.feature.add_module(f'f_{activation_fn.__name__}{i}', activation_fn(**activation_args)) hin = h self._out_features = hin def extra_repr(self): return f'FC: {self.hidden_sizes}x{self._activation_fn.__name__}' def hidden_layer(self, index=0): return self.feature[index * 2] def output_size(self): return self._out_features def forward(self, input_0): primals_1 = self.feature.f_fc0.weight primals_2 = self.feature.f_fc0.bias primals_3 = input_0 output = call([primals_1, primals_2, primals_3]) return output[0]
criteo-research/pytorch-ada
FeatureExtractFF
false
15,082
[ "Apache-2.0" ]
68
4b8861ce1c12fc8a4391eb14a811459e3e8a074a
https://github.com/criteo-research/pytorch-ada/tree/4b8861ce1c12fc8a4391eb14a811459e3e8a074a