std::cout << "Forwarding into jit module" << std::endl;
std::cout << "Forward code:" << std::endl;
std::cout << *grad.f.get() << std::endl;
std::cout << "Backward code:" << std::endl;
std::cout << *grad.df.get() << std::endl;
std::cout << "End print !" << std::endl;
Printing during the evaluation: Forwarding into jit module
Forward code:
graph(%0 : Double(*, *),
%1 : Double(*),
%2 : Double(*, *),
%3 : Double(*, *),
%4 : Double(*)):
%5 : int = prim::Constant[value=1]() # <string>:193:92
%6 : int = prim::Constant[value=2]() # <string>:384:35
%7 : Double(*, *) = aten::t(%3)
%8 : Double(*, *) = aten::mm(%2, %7) # <string>:8:25
%9 : Double(*, *) = aten::add(%4, %8, %5) # <string>:8:18
%x.2 : Tensor = aten::linear(%9, %0, %1), scope: VAE/Linear[fc4] # /Users/albandes/workspace/pytorch_dev/torch/nn/functional.py:1415:0
%x : Tensor = aten::pow(%x.2, %6), scope: VAE # foo.py:16:0
return (%x, %9, %6, %x.2, %5)
Backward code:
graph(%0 : Tensor,
%1 : Tensor,
%2 : Tensor,
%3 : Double(*, *),
%4 : Double(*),
%5 : Double(*, *),
%6 : Double(*, *),
%7 : Double(*),
%input : Tensor,
%9 : int,
%x.1 : Tensor,
%11 : int):
%12 : None = prim::Constant() # <string>:380:32
%13 : int = prim::Constant[value=0]() # <string>:349:24
%14 : bool = prim::Constant[value=1]() # <string>:349:16
%15 : int = prim::Constant[value=-2]() # <string>:340:23
%16 : int = prim::Constant[value=-1]() # <string>:339:23
%17 : int = prim::Constant[value=1]() # <string>:333:23
%18 : float = prim::Constant[value=0]() # <string>:190:39
%19 : int = prim::Constant[value=2]() # <string>:384:35
%grad_self : Tensor = prim::GradOf[name="aten::pow"](%0)
block0():
%21 : float = prim::Float(%19) # <string>:190:20
%22 : bool = aten::eq(%21, %18) # <string>:190:20
%grad_self.3 : Tensor = prim::If(%22) # <string>:190:17
block0():
%grad_self.4 : Tensor = aten::zeros_like(%x.1) # <string>:191:33
-> (%grad_self.4)
block1():
%25 : Tensor = aten::mul(%0, %9) # <string>:193:33
%26 : float = prim::Float(%9) # <string>:193:74
%27 : float = aten::sub(%26, %11) # <string>:193:74
%28 : Tensor = aten::pow(%x.1, %27) # <string>:193:58
%grad_self.2 : Tensor = aten::mul(%25, %28) # <string>:193:33
-> (%grad_self.2)
-> (%grad_self.3)
%30 : Tensor = prim::AutogradAdd(%2, %grad_self)
%grad_input.3 : Tensor, %grad_weight.3 : Tensor, %grad_bias.3 : Tensor? = prim::GradOf[name="aten::linear"](%30)
block0():
%34 : bool = aten::__isnot__(%4, %12) # <string>:380:20
%grad_bias.2 : Tensor? = prim::If(%34) # <string>:380:17
block0():
%bias.6 : Tensor = prim::unchecked_unwrap_optional(%4)
%37 : int[] = aten::size(%bias.6) # <string>:381:63
%grad_bias.4 : Tensor = aten::_grad_sum_to_size(%30, %37) # <string>:381:33
-> (%grad_bias.4)
block1():
-> (%12)
%39 : int = aten::dim(%input) # <string>:384:20
%40 : bool = aten::eq(%39, %19) # <string>:384:20
%grad_input.2 : Tensor, %grad_weight.2 : Tensor = prim::If(%40) # <string>:384:17
block0():
%grad_input.4 : Tensor = aten::mm(%30, %3) # <string>:385:34
%44 : Tensor = aten::t(%30) # <string>:386:35
%grad_weight.5 : Tensor = aten::mm(%44, %input) # <string>:386:35
-> (%grad_input.4, %grad_weight.5)
block1():
%46 : Tensor = aten::t(%3) # <string>:388:86
%self_size.2 : int[] = aten::size(%input) # <string>:366:25
%other_size.2 : int[] = aten::size(%46) # <string>:367:26
%dim.3 : int = aten::dim(%46) # <string>:332:19
%50 : bool = aten::eq(%dim.3, %17) # <string>:333:16
%out.20 : Tensor = prim::If(%50) # <string>:333:13
block0():
-> (%46)
block1():
%52 : bool = aten::eq(%dim.3, %19) # <string>:335:18
%out.21 : Tensor = prim::If(%52) # <string>:335:18
block0():
%out.22 : Tensor = aten::t(%46) # <string>:336:23
-> (%out.22)
block1():
%dims.3 : int[] = prim::rangelist(%dim.3) # <string>:338:24
%56 : int = aten::sub(%dim.3, %19) # <string>:339:28
%57 : int[] = aten::_set_item(%dims.3, %16, %56) # <string>:339:17
%58 : int = aten::sub(%dim.3, %17) # <string>:340:28
%59 : int[] = aten::_set_item(%dims.3, %15, %58) # <string>:340:17
%out.29 : Tensor = aten::permute(%46, %dims.3) # <string>:341:23
-> (%out.29)
-> (%out.21)
%dim1.3 : int = aten::dim(%30) # <string>:346:20
%dim2.3 : int = aten::dim(%out.20) # <string>:347:20
%dim_out.3 : int = aten::len(%self_size.2) # <string>:348:23
%64 : bool = aten::eq(%dim1.3, %13) # <string>:349:16
%65 : bool = prim::If(%64) # <string>:349:16
block0():
-> (%14)
block1():
%66 : bool = aten::eq(%dim2.3, %13) # <string>:349:29
-> (%66)
%out.30 : Tensor = prim::If(%65) # <string>:349:13
block0():
%out.31 : Tensor = aten::mul(%30, %out.20) # <string>:350:23
-> (%out.31)
block1():
%69 : int = aten::add(%dim1.3, %dim2.3) # <string>:351:18
%70 : bool = aten::eq(%69, %dim_out.3) # <string>:351:18
%out.32 : Tensor = prim::If(%70) # <string>:351:18
block0():
%72 : bool = aten::eq(%dim2.3, %17) # <string>:352:20
%target_dim2.2 : int = prim::If(%72) # <string>:352:17
block0():
-> (%13)
block1():
-> (%15)
%74 : Tensor = aten::unsqueeze(%30, %dim1.3) # <string>:356:36
%75 : Tensor = aten::unsqueeze(%out.20, %target_dim2.2) # <string>:356:58
%out.33 : Tensor = aten::matmul(%74, %75) # <string>:356:23
-> (%out.33)
block1():
%77 : int = aten::sub(%dim1.3, %dim2.3) # <string>:357:29
%78 : bool = aten::eq(%dim_out.3, %77) # <string>:357:18
%out.34 : Tensor = prim::If(%78) # <string>:357:18
block0():
%80 : Tensor = aten::unsqueeze(%out.20, %dim2.3) # <string>:358:42
%81 : Tensor = aten::matmul(%30, %80) # <string>:358:23
%out.35 : Tensor = aten::squeeze(%81, %16) # <string>:358:23
-> (%out.35)
block1():
%83 : int = aten::sub(%dim2.3, %dim1.3) # <string>:359:29
%84 : bool = aten::eq(%dim_out.3, %83) # <string>:359:18
%out.36 : Tensor = prim::If(%84) # <string>:359:18
block0():
%86 : Tensor = aten::unsqueeze(%30, %15) # <string>:360:36
%87 : Tensor = aten::matmul(%86, %out.20) # <string>:360:23
%out.37 : Tensor = aten::squeeze(%87, %15) # <string>:360:23
-> (%out.37)
block1():
%out.38 : Tensor = aten::matmul(%30, %out.20) # <string>:362:23
-> (%out.38)
-> (%out.36)
-> (%out.34)
-> (%out.32)
%grad_self.5 : Tensor = aten::_grad_sum_to_size(%out.30, %self_size.2) # <string>:368:25
%dim.4 : int = aten::dim(%input) # <string>:332:19
%92 : bool = aten::eq(%dim.4, %17) # <string>:333:16
%out.39 : Tensor = prim::If(%92) # <string>:333:13
block0():
-> (%input)
block1():
%94 : bool = aten::eq(%dim.4, %19) # <string>:335:18
%out.40 : Tensor = prim::If(%94) # <string>:335:18
block0():
%out.41 : Tensor = aten::t(%input) # <string>:336:23
-> (%out.41)
block1():
%dims.4 : int[] = prim::rangelist(%dim.4) # <string>:338:24
%98 : int = aten::sub(%dim.4, %19) # <string>:339:28
%99 : int[] = aten::_set_item(%dims.4, %16, %98) # <string>:339:17
%100 : int = aten::sub(%dim.4, %17) # <string>:340:28
%101 : int[] = aten::_set_item(%dims.4, %15, %100) # <string>:340:17
%out.42 : Tensor = aten::permute(%input, %dims.4) # <string>:341:23
-> (%out.42)
-> (%out.40)
%dim1.4 : int = aten::dim(%out.39) # <string>:346:20
%dim2.4 : int = aten::dim(%30) # <string>:347:20
%dim_out.4 : int = aten::len(%other_size.2) # <string>:348:23
%106 : bool = aten::eq(%dim1.4, %13) # <string>:349:16
%107 : bool = prim::If(%106) # <string>:349:16
block0():
-> (%14)
block1():
%108 : bool = aten::eq(%dim2.4, %13) # <string>:349:29
-> (%108)
%out.43 : Tensor = prim::If(%107) # <string>:349:13
block0():
%out.44 : Tensor = aten::mul(%out.39, %30) # <string>:350:23
-> (%out.44)
block1():
%111 : int = aten::add(%dim1.4, %dim2.4) # <string>:351:18
%112 : bool = aten::eq(%111, %dim_out.4) # <string>:351:18
%out.45 : Tensor = prim::If(%112) # <string>:351:18
block0():
%114 : bool = aten::eq(%dim2.4, %17) # <string>:352:20
%target_dim2.3 : int = prim::If(%114) # <string>:352:17
block0():
-> (%13)
block1():
-> (%15)
%116 : Tensor = aten::unsqueeze(%out.39, %dim1.4) # <string>:356:36
%117 : Tensor = aten::unsqueeze(%30, %target_dim2.3) # <string>:356:58
%out.46 : Tensor = aten::matmul(%116, %117) # <string>:356:23
-> (%out.46)
block1():
%119 : int = aten::sub(%dim1.4, %dim2.4) # <string>:357:29
%120 : bool = aten::eq(%dim_out.4, %119) # <string>:357:18
%out.47 : Tensor = prim::If(%120) # <string>:357:18
block0():
%122 : Tensor = aten::unsqueeze(%30, %dim2.4) # <string>:358:42
%123 : Tensor = aten::matmul(%out.39, %122) # <string>:358:23
%out.48 : Tensor = aten::squeeze(%123, %16) # <string>:358:23
-> (%out.48)
block1():
%125 : int = aten::sub(%dim2.4, %dim1.4) # <string>:359:29
%126 : bool = aten::eq(%dim_out.4, %125) # <string>:359:18
%out.49 : Tensor = prim::If(%126) # <string>:359:18
block0():
%128 : Tensor = aten::unsqueeze(%out.39, %15) # <string>:360:36
%129 : Tensor = aten::matmul(%128, %30) # <string>:360:23
%out.50 : Tensor = aten::squeeze(%129, %15) # <string>:360:23
-> (%out.50)
block1():
%out.51 : Tensor = aten::matmul(%out.39, %30) # <string>:362:23
-> (%out.51)
-> (%out.49)
-> (%out.47)
-> (%out.45)
%grad_other.2 : Tensor = aten::_grad_sum_to_size(%out.43, %other_size.2) # <string>:369:26
%grad_weight.6 : Tensor = aten::t(%grad_other.2) # <string>:389:35
-> (%grad_self.5, %grad_weight.6)
-> (%grad_input.2, %grad_weight.2, %grad_bias.2)
%134 : Tensor = prim::AutogradAdd(%1, %grad_input.3)
%grad_weight : Tensor, %grad_bias : Tensor? = prim::GradOf[name="aten::linear"](%134)
block0():
%137 : bool = aten::__isnot__(%7, %12) # <string>:380:20
%grad_bias.5 : Tensor? = prim::If(%137) # <string>:380:17
block0():
%bias.5 : Tensor = prim::unchecked_unwrap_optional(%7)
%140 : int[] = aten::size(%bias.5) # <string>:381:63
%grad_bias.1 : Tensor = aten::_grad_sum_to_size(%134, %140) # <string>:381:33
-> (%grad_bias.1)
block1():
-> (%12)
%142 : int = aten::dim(%5) # <string>:384:20
%143 : bool = aten::eq(%142, %19) # <string>:384:20
%grad_weight.7 : Tensor = prim::If(%143) # <string>:384:17
block0():
%145 : Tensor = aten::t(%134) # <string>:386:35
%grad_weight.1 : Tensor = aten::mm(%145, %5) # <string>:386:35
-> (%grad_weight.1)
block1():
%147 : Tensor = aten::t(%6) # <string>:388:86
%other_size.1 : int[] = aten::size(%147) # <string>:367:26
%dim.2 : int = aten::dim(%147) # <string>:332:19
%150 : bool = aten::eq(%dim.2, %17) # <string>:333:16
= prim::If(%150) # <string>:333:13
block0():
-> ()
block1():
%151 : bool = aten::eq(%dim.2, %19) # <string>:335:18
= prim::If(%151) # <string>:335:18
block0():
-> ()
block1():
%dims.2 : int[] = prim::rangelist(%dim.2) # <string>:338:24
%153 : int = aten::sub(%dim.2, %19) # <string>:339:28
%154 : int[] = aten::_set_item(%dims.2, %16, %153) # <string>:339:17
%155 : int = aten::sub(%dim.2, %17) # <string>:340:28
%156 : int[] = aten::_set_item(%dims.2, %15, %155) # <string>:340:17
-> ()
-> ()
%dim.1 : int = aten::dim(%5) # <string>:332:19
%158 : bool = aten::eq(%dim.1, %17) # <string>:333:16
%out.14 : Tensor = prim::If(%158) # <string>:333:13
block0():
-> (%5)
block1():
%160 : bool = aten::eq(%dim.1, %19) # <string>:335:18
%out.12 : Tensor = prim::If(%160) # <string>:335:18
block0():
%out.15 : Tensor = aten::t(%5) # <string>:336:23
-> (%out.15)
block1():
%dims.1 : int[] = prim::rangelist(%dim.1) # <string>:338:24
%164 : int = aten::sub(%dim.1, %19) # <string>:339:28
%165 : int[] = aten::_set_item(%dims.1, %16, %164) # <string>:339:17
%166 : int = aten::sub(%dim.1, %17) # <string>:340:28
%167 : int[] = aten::_set_item(%dims.1, %15, %166) # <string>:340:17
%out.16 : Tensor = aten::permute(%5, %dims.1) # <string>:341:23
-> (%out.16)
-> (%out.12)
%dim1.1 : int = aten::dim(%out.14) # <string>:346:20
%dim2.1 : int = aten::dim(%134) # <string>:347:20
%dim_out.1 : int = aten::len(%other_size.1) # <string>:348:23
%172 : bool = aten::eq(%dim1.1, %13) # <string>:349:16
%173 : bool = prim::If(%172) # <string>:349:16
block0():
-> (%14)
block1():
%174 : bool = aten::eq(%dim2.1, %13) # <string>:349:29
-> (%174)
%out : Tensor = prim::If(%173) # <string>:349:13
block0():
%out.1 : Tensor = aten::mul(%out.14, %134) # <string>:350:23
-> (%out.1)
block1():
%177 : int = aten::add(%dim1.1, %dim2.1) # <string>:351:18
%178 : bool = aten::eq(%177, %dim_out.1) # <string>:351:18
%out.25 : Tensor = prim::If(%178) # <string>:351:18
block0():
%180 : bool = aten::eq(%dim2.1, %17) # <string>:352:20
%target_dim2 : int = prim::If(%180) # <string>:352:17
block0():
-> (%13)
block1():
-> (%15)
%182 : Tensor = aten::unsqueeze(%out.14, %dim1.1) # <string>:356:36
%183 : Tensor = aten::unsqueeze(%134, %target_dim2) # <string>:356:58
%out.2 : Tensor = aten::matmul(%182, %183) # <string>:356:23
-> (%out.2)
block1():
%185 : int = aten::sub(%dim1.1, %dim2.1) # <string>:357:29
%186 : bool = aten::eq(%dim_out.1, %185) # <string>:357:18
%out.24 : Tensor = prim::If(%186) # <string>:357:18
block0():
%188 : Tensor = aten::unsqueeze(%134, %dim2.1) # <string>:358:42
%189 : Tensor = aten::matmul(%out.14, %188) # <string>:358:23
%out.3 : Tensor = aten::squeeze(%189, %16) # <string>:358:23
-> (%out.3)
block1():
%191 : int = aten::sub(%dim2.1, %dim1.1) # <string>:359:29
%192 : bool = aten::eq(%dim_out.1, %191) # <string>:359:18
%out.23 : Tensor = prim::If(%192) # <string>:359:18
block0():
%194 : Tensor = aten::unsqueeze(%out.14, %15) # <string>:360:36
%195 : Tensor = aten::matmul(%194, %134) # <string>:360:23
%out.4 : Tensor = aten::squeeze(%195, %15) # <string>:360:23
-> (%out.4)
block1():
%out.5 : Tensor = aten::matmul(%out.14, %134) # <string>:362:23
-> (%out.5)
-> (%out.23)
-> (%out.24)
-> (%out.25)
%grad_other.1 : Tensor = aten::_grad_sum_to_size(%out, %other_size.1) # <string>:369:26
%grad_weight.4 : Tensor = aten::t(%grad_other.1) # <string>:389:35
-> (%grad_weight.4)
-> (%grad_weight.7, %grad_bias.5)
return (%grad_weight.3, %grad_bias.3, %grad_weight, %grad_bias)
End print !
Printing during the backward pass Forwarding into jit module
Forward code:
graph(%0 : int[],
%1 : Double(*, *),
%2 : int[],
%3 : Double(*, *),
%4 : Double(*, *),
%5 : Double(*, *),
%6 : float,
%7 : Double(*, *),
%8 : int):
%9 : None = prim::Constant() # <string>:274:28
%10 : int = prim::Constant[value=1]() # <string>:193:92
%11 : Double(*, *) = aten::mul(%7, %8) # <string>:193:33
%12 : Double(*, *) = aten::pow(%5, %6) # <string>:193:58
%grad_self.13 : Double(*, *) = aten::mul(%11, %12) # <string>:193:33
%14 : int[] = aten::size(%12) # <string>:2:94
%15 : int[] = aten::size(%grad_self.13) # <string>:2:105
%16 : int[]? = aten::_size_if_not_equal(%14, %15) # <string>:2:69
%grad_input.7 : Double(*, *) = aten::mm(%grad_self.13, %4) # <string>:385:34
%18 : Double(*, *) = aten::t(%grad_input.7) # <string>:247:20
%grad_weight.1 : Double(*, *) = aten::mm(%18, %3) # <string>:386:35
%self_size.3 : int[] = aten::size(%grad_input.7) # <string>:275:29
%result.3 : Tensor = aten::_grad_sum_to_size(%grad_input.7, %2) # <string>:279:22
%22 : Double(*, *) = aten::t(%grad_self.13) # <string>:247:20
%grad_weight.5 : Double(*, *) = aten::mm(%22, %1) # <string>:386:35
%self_size.2 : int[] = aten::size(%grad_self.13) # <string>:275:29
%result.2 : Tensor = aten::_grad_sum_to_size(%grad_self.13, %0) # <string>:279:22
return (%result.2, %grad_weight.5, %result.3, %grad_weight.1, %11, %10, %grad_self.13, %16, %grad_input.7, %9, %self_size.3, %22, %self_size.2)
Backward code:
graph(%0 : Tensor,
%1 : Double(*, *),
%2 : Tensor,
%3 : Double(*, *),
%4 : Double(*, *),
%5 : Double(*, *),
%6 : int[],
%7 : Double(*, *),
%8 : int[],
%9 : Double(*, *),
%10 : Double(*, *),
%11 : Double(*, *),
%12 : float,
%13 : Double(*, *),
%14 : int,
%grad_self.11 : Double(*, *),
%16 : int[]?,
%grad_input.4 : Double(*, *),
%18 : None,
%self_size.20 : int[]?,
%20 : Double(*, *),
%self_size.18 : int[]?):
%22 : None = prim::Constant() # <string>:281:33
%23 : bool = prim::Constant[value=0]()
%24 : float = prim::Constant[value=0]() # <string>:190:39
%grad_input.3 : Tensor = prim::GradOf[name="aten::_grad_sum_to_size"](%0)
block0():
%26 : bool = aten::__is__(%self_size.18, %22) # <string>:281:20
%grad_input.2 : Tensor = prim::If(%26) # <string>:281:17
block0():
-> (%0)
block1():
%self_size.12 : int[] = prim::unchecked_unwrap_optional(%self_size.18)
%grad_input.5 : Tensor = aten::expand(%0, %self_size.12, %23) # <string>:284:34
-> (%grad_input.5)
-> (%grad_input.2)
%grad_self.4 : Tensor, %grad_mat2.3 : Tensor = prim::GradOf[name="aten::mm"](%1)
block0():
%32 : Tensor = aten::t(%7) # <string>:192:28
%grad_self.3 : Tensor = aten::mm(%1, %32) # <string>:192:20
%34 : Tensor = aten::t(%20) # <string>:195:20
%grad_mat2.2 : Tensor = aten::mm(%34, %1) # <string>:195:20
-> (%grad_self.3, %grad_mat2.2)
%36 : Tensor = prim::GradOf[name="aten::t"](%grad_self.4)
block0():
%37 : Tensor = aten::t(%grad_self.4) # <string>:245:24
-> (%37)
%38 : Tensor = prim::AutogradAdd(%grad_input.3, %36)
%grad_input : Tensor = prim::GradOf[name="aten::_grad_sum_to_size"](%2)
block0():
%40 : bool = aten::__is__(%self_size.20, %22) # <string>:281:20
%grad_input.6 : Tensor = prim::If(%40) # <string>:281:17
block0():
-> (%2)
block1():
%self_size.11 : int[] = prim::unchecked_unwrap_optional(%self_size.20)
%grad_input.1 : Tensor = aten::expand(%2, %self_size.11, %23) # <string>:284:34
-> (%grad_input.1)
-> (%grad_input.6)
%grad_self.6 : Tensor = prim::GradOf[name="aten::mm"](%3)
block0():
%45 : Tensor = aten::t(%9) # <string>:192:28
%grad_self.5 : Tensor = aten::mm(%3, %45) # <string>:192:20
-> (%grad_self.5)
%47 : Tensor = prim::GradOf[name="aten::t"](%grad_self.6)
block0():
%48 : Tensor = aten::t(%grad_self.6) # <string>:245:24
-> (%48)
%49 : Tensor = prim::AutogradAdd(%grad_input, %47)
%50 : Tensor = prim::AutogradAdd(%5, %49)
%grad_self.8 : Tensor, %grad_mat2.1 : Tensor = prim::GradOf[name="aten::mm"](%50)
block0():
%53 : Tensor = aten::t(%10) # <string>:192:28
%grad_self.7 : Tensor = aten::mm(%50, %53) # <string>:192:20
%55 : Tensor = aten::t(%grad_self.11) # <string>:195:20
%grad_mat2.6 : Tensor = aten::mm(%55, %50) # <string>:195:20
-> (%grad_self.7, %grad_mat2.6)
%57 : Tensor = prim::AutogradAdd(%38, %grad_self.8)
%58 : Tensor = prim::AutogradAdd(%4, %57)
%grad_other.1 : Tensor = prim::GradOf[name="aten::mul"](%58)
block0():
%60 : Tensor = aten::mul(%58, %13) # <string>:12:31
%grad_other.2 : Tensor = aten::_grad_sum_to_size(%60, %16) # <string>:12:31
-> (%grad_other.2)
%grad_self : Tensor = prim::GradOf[name="aten::pow"](%grad_other.1)
block0():
%63 : float = prim::Float(%12) # <string>:190:20
%64 : bool = aten::eq(%63, %24) # <string>:190:20
%grad_self.12 : Tensor = prim::If(%64) # <string>:190:17
block0():
%grad_self.1 : Tensor = aten::zeros_like(%11) # <string>:191:33
-> (%grad_self.1)
block1():
%67 : Tensor = aten::mul(%grad_other.1, %12) # <string>:193:33
%68 : float = prim::Float(%12) # <string>:193:74
%69 : float = aten::sub(%68, %14) # <string>:193:74
%70 : Tensor = aten::pow(%11, %69) # <string>:193:58
%grad_self.2 : Tensor = aten::mul(%67, %70) # <string>:193:33
-> (%grad_self.2)
-> (%grad_self.12)
return (%grad_mat2.3, %grad_mat2.1, %grad_self)
End print !
Full print with python side printing as well: Tracing
foo.py:15: TracerWarning: Converting a tensor to a Python integer might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
print(x.size(), x.grad_fn, x.requires_grad)
torch.Size([1, 4]) <AddmmBackward object at 0x1210ba690> True
torch.Size([1, 4]) None False
torch.Size([1, 4]) None False
Tracing done
Forwarding into jit module
Forward code:
graph(%0 : Double(*, *),
%1 : Double(*),
%2 : Double(*, *),
%3 : Double(*, *),
%4 : Double(*)):
%5 : int = prim::Constant[value=1]() # <string>:193:92
%6 : int = prim::Constant[value=2]() # <string>:384:35
%7 : Double(*, *) = aten::t(%3)
%8 : Double(*, *) = aten::mm(%2, %7) # <string>:8:25
%9 : Double(*, *) = aten::add(%4, %8, %5) # <string>:8:18
%x.2 : Tensor = aten::linear(%9, %0, %1), scope: VAE/Linear[fc4] # /Users/albandes/workspace/pytorch_dev/torch/nn/functional.py:1415:0
%x : Tensor = aten::pow(%x.2, %6), scope: VAE # foo.py:16:0
return (%x, %9, %6, %x.2, %5)
Backward code:
graph(%0 : Tensor,
%1 : Tensor,
%2 : Tensor,
%3 : Double(*, *),
%4 : Double(*),
%5 : Double(*, *),
%6 : Double(*, *),
%7 : Double(*),
%input : Tensor,
%9 : int,
%x.1 : Tensor,
%11 : int):
%12 : None = prim::Constant() # <string>:380:32
%13 : int = prim::Constant[value=0]() # <string>:349:24
%14 : bool = prim::Constant[value=1]() # <string>:349:16
%15 : int = prim::Constant[value=-2]() # <string>:340:23
%16 : int = prim::Constant[value=-1]() # <string>:339:23
%17 : int = prim::Constant[value=1]() # <string>:333:23
%18 : float = prim::Constant[value=0]() # <string>:190:39
%19 : int = prim::Constant[value=2]() # <string>:384:35
%grad_self : Tensor = prim::GradOf[name="aten::pow"](%0)
block0():
%21 : float = prim::Float(%19) # <string>:190:20
%22 : bool = aten::eq(%21, %18) # <string>:190:20
%grad_self.3 : Tensor = prim::If(%22) # <string>:190:17
block0():
%grad_self.4 : Tensor = aten::zeros_like(%x.1) # <string>:191:33
-> (%grad_self.4)
block1():
%25 : Tensor = aten::mul(%0, %9) # <string>:193:33
%26 : float = prim::Float(%9) # <string>:193:74
%27 : float = aten::sub(%26, %11) # <string>:193:74
%28 : Tensor = aten::pow(%x.1, %27) # <string>:193:58
%grad_self.2 : Tensor = aten::mul(%25, %28) # <string>:193:33
-> (%grad_self.2)
-> (%grad_self.3)
%30 : Tensor = prim::AutogradAdd(%2, %grad_self)
%grad_input.3 : Tensor, %grad_weight.3 : Tensor, %grad_bias.3 : Tensor? = prim::GradOf[name="aten::linear"](%30)
block0():
%34 : bool = aten::__isnot__(%4, %12) # <string>:380:20
%grad_bias.2 : Tensor? = prim::If(%34) # <string>:380:17
block0():
%bias.6 : Tensor = prim::unchecked_unwrap_optional(%4)
%37 : int[] = aten::size(%bias.6) # <string>:381:63
%grad_bias.4 : Tensor = aten::_grad_sum_to_size(%30, %37) # <string>:381:33
-> (%grad_bias.4)
block1():
-> (%12)
%39 : int = aten::dim(%input) # <string>:384:20
%40 : bool = aten::eq(%39, %19) # <string>:384:20
%grad_input.2 : Tensor, %grad_weight.2 : Tensor = prim::If(%40) # <string>:384:17
block0():
%grad_input.4 : Tensor = aten::mm(%30, %3) # <string>:385:34
%44 : Tensor = aten::t(%30) # <string>:386:35
%grad_weight.5 : Tensor = aten::mm(%44, %input) # <string>:386:35
-> (%grad_input.4, %grad_weight.5)
block1():
%46 : Tensor = aten::t(%3) # <string>:388:86
%self_size.2 : int[] = aten::size(%input) # <string>:366:25
%other_size.2 : int[] = aten::size(%46) # <string>:367:26
%dim.3 : int = aten::dim(%46) # <string>:332:19
%50 : bool = aten::eq(%dim.3, %17) # <string>:333:16
%out.20 : Tensor = prim::If(%50) # <string>:333:13
block0():
-> (%46)
block1():
%52 : bool = aten::eq(%dim.3, %19) # <string>:335:18
%out.21 : Tensor = prim::If(%52) # <string>:335:18
block0():
%out.22 : Tensor = aten::t(%46) # <string>:336:23
-> (%out.22)
block1():
%dims.3 : int[] = prim::rangelist(%dim.3) # <string>:338:24
%56 : int = aten::sub(%dim.3, %19) # <string>:339:28
%57 : int[] = aten::_set_item(%dims.3, %16, %56) # <string>:339:17
%58 : int = aten::sub(%dim.3, %17) # <string>:340:28
%59 : int[] = aten::_set_item(%dims.3, %15, %58) # <string>:340:17
%out.29 : Tensor = aten::permute(%46, %dims.3) # <string>:341:23
-> (%out.29)
-> (%out.21)
%dim1.3 : int = aten::dim(%30) # <string>:346:20
%dim2.3 : int = aten::dim(%out.20) # <string>:347:20
%dim_out.3 : int = aten::len(%self_size.2) # <string>:348:23
%64 : bool = aten::eq(%dim1.3, %13) # <string>:349:16
%65 : bool = prim::If(%64) # <string>:349:16
block0():
-> (%14)
block1():
%66 : bool = aten::eq(%dim2.3, %13) # <string>:349:29
-> (%66)
%out.30 : Tensor = prim::If(%65) # <string>:349:13
block0():
%out.31 : Tensor = aten::mul(%30, %out.20) # <string>:350:23
-> (%out.31)
block1():
%69 : int = aten::add(%dim1.3, %dim2.3) # <string>:351:18
%70 : bool = aten::eq(%69, %dim_out.3) # <string>:351:18
%out.32 : Tensor = prim::If(%70) # <string>:351:18
block0():
%72 : bool = aten::eq(%dim2.3, %17) # <string>:352:20
%target_dim2.2 : int = prim::If(%72) # <string>:352:17
block0():
-> (%13)
block1():
-> (%15)
%74 : Tensor = aten::unsqueeze(%30, %dim1.3) # <string>:356:36
%75 : Tensor = aten::unsqueeze(%out.20, %target_dim2.2) # <string>:356:58
%out.33 : Tensor = aten::matmul(%74, %75) # <string>:356:23
-> (%out.33)
block1():
%77 : int = aten::sub(%dim1.3, %dim2.3) # <string>:357:29
%78 : bool = aten::eq(%dim_out.3, %77) # <string>:357:18
%out.34 : Tensor = prim::If(%78) # <string>:357:18
block0():
%80 : Tensor = aten::unsqueeze(%out.20, %dim2.3) # <string>:358:42
%81 : Tensor = aten::matmul(%30, %80) # <string>:358:23
%out.35 : Tensor = aten::squeeze(%81, %16) # <string>:358:23
-> (%out.35)
block1():
%83 : int = aten::sub(%dim2.3, %dim1.3) # <string>:359:29
%84 : bool = aten::eq(%dim_out.3, %83) # <string>:359:18
%out.36 : Tensor = prim::If(%84) # <string>:359:18
block0():
%86 : Tensor = aten::unsqueeze(%30, %15) # <string>:360:36
%87 : Tensor = aten::matmul(%86, %out.20) # <string>:360:23
%out.37 : Tensor = aten::squeeze(%87, %15) # <string>:360:23
-> (%out.37)
block1():
%out.38 : Tensor = aten::matmul(%30, %out.20) # <string>:362:23
-> (%out.38)
-> (%out.36)
-> (%out.34)
-> (%out.32)
%grad_self.5 : Tensor = aten::_grad_sum_to_size(%out.30, %self_size.2) # <string>:368:25
%dim.4 : int = aten::dim(%input) # <string>:332:19
%92 : bool = aten::eq(%dim.4, %17) # <string>:333:16
%out.39 : Tensor = prim::If(%92) # <string>:333:13
block0():
-> (%input)
block1():
%94 : bool = aten::eq(%dim.4, %19) # <string>:335:18
%out.40 : Tensor = prim::If(%94) # <string>:335:18
block0():
%out.41 : Tensor = aten::t(%input) # <string>:336:23
-> (%out.41)
block1():
%dims.4 : int[] = prim::rangelist(%dim.4) # <string>:338:24
%98 : int = aten::sub(%dim.4, %19) # <string>:339:28
%99 : int[] = aten::_set_item(%dims.4, %16, %98) # <string>:339:17
%100 : int = aten::sub(%dim.4, %17) # <string>:340:28
%101 : int[] = aten::_set_item(%dims.4, %15, %100) # <string>:340:17
%out.42 : Tensor = aten::permute(%input, %dims.4) # <string>:341:23
-> (%out.42)
-> (%out.40)
%dim1.4 : int = aten::dim(%out.39) # <string>:346:20
%dim2.4 : int = aten::dim(%30) # <string>:347:20
%dim_out.4 : int = aten::len(%other_size.2) # <string>:348:23
%106 : bool = aten::eq(%dim1.4, %13) # <string>:349:16
%107 : bool = prim::If(%106) # <string>:349:16
block0():
-> (%14)
block1():
%108 : bool = aten::eq(%dim2.4, %13) # <string>:349:29
-> (%108)
%out.43 : Tensor = prim::If(%107) # <string>:349:13
block0():
%out.44 : Tensor = aten::mul(%out.39, %30) # <string>:350:23
-> (%out.44)
block1():
%111 : int = aten::add(%dim1.4, %dim2.4) # <string>:351:18
%112 : bool = aten::eq(%111, %dim_out.4) # <string>:351:18
%out.45 : Tensor = prim::If(%112) # <string>:351:18
block0():
%114 : bool = aten::eq(%dim2.4, %17) # <string>:352:20
%target_dim2.3 : int = prim::If(%114) # <string>:352:17
block0():
-> (%13)
block1():
-> (%15)
%116 : Tensor = aten::unsqueeze(%out.39, %dim1.4) # <string>:356:36
%117 : Tensor = aten::unsqueeze(%30, %target_dim2.3) # <string>:356:58
%out.46 : Tensor = aten::matmul(%116, %117) # <string>:356:23
-> (%out.46)
block1():
%119 : int = aten::sub(%dim1.4, %dim2.4) # <string>:357:29
%120 : bool = aten::eq(%dim_out.4, %119) # <string>:357:18
%out.47 : Tensor = prim::If(%120) # <string>:357:18
block0():
%122 : Tensor = aten::unsqueeze(%30, %dim2.4) # <string>:358:42
%123 : Tensor = aten::matmul(%out.39, %122) # <string>:358:23
%out.48 : Tensor = aten::squeeze(%123, %16) # <string>:358:23
-> (%out.48)
block1():
%125 : int = aten::sub(%dim2.4, %dim1.4) # <string>:359:29
%126 : bool = aten::eq(%dim_out.4, %125) # <string>:359:18
%out.49 : Tensor = prim::If(%126) # <string>:359:18
block0():
%128 : Tensor = aten::unsqueeze(%out.39, %15) # <string>:360:36
%129 : Tensor = aten::matmul(%128, %30) # <string>:360:23
%out.50 : Tensor = aten::squeeze(%129, %15) # <string>:360:23
-> (%out.50)
block1():
%out.51 : Tensor = aten::matmul(%out.39, %30) # <string>:362:23
-> (%out.51)
-> (%out.49)
-> (%out.47)
-> (%out.45)
%grad_other.2 : Tensor = aten::_grad_sum_to_size(%out.43, %other_size.2) # <string>:369:26
%grad_weight.6 : Tensor = aten::t(%grad_other.2) # <string>:389:35
-> (%grad_self.5, %grad_weight.6)
-> (%grad_input.2, %grad_weight.2, %grad_bias.2)
%134 : Tensor = prim::AutogradAdd(%1, %grad_input.3)
%grad_weight : Tensor, %grad_bias : Tensor? = prim::GradOf[name="aten::linear"](%134)
block0():
%137 : bool = aten::__isnot__(%7, %12) # <string>:380:20
%grad_bias.5 : Tensor? = prim::If(%137) # <string>:380:17
block0():
%bias.5 : Tensor = prim::unchecked_unwrap_optional(%7)
%140 : int[] = aten::size(%bias.5) # <string>:381:63
%grad_bias.1 : Tensor = aten::_grad_sum_to_size(%134, %140) # <string>:381:33
-> (%grad_bias.1)
block1():
-> (%12)
%142 : int = aten::dim(%5) # <string>:384:20
%143 : bool = aten::eq(%142, %19) # <string>:384:20
%grad_weight.7 : Tensor = prim::If(%143) # <string>:384:17
block0():
%145 : Tensor = aten::t(%134) # <string>:386:35
%grad_weight.1 : Tensor = aten::mm(%145, %5) # <string>:386:35
-> (%grad_weight.1)
block1():
%147 : Tensor = aten::t(%6) # <string>:388:86
%other_size.1 : int[] = aten::size(%147) # <string>:367:26
%dim.2 : int = aten::dim(%147) # <string>:332:19
%150 : bool = aten::eq(%dim.2, %17) # <string>:333:16
= prim::If(%150) # <string>:333:13
block0():
-> ()
block1():
%151 : bool = aten::eq(%dim.2, %19) # <string>:335:18
= prim::If(%151) # <string>:335:18
block0():
-> ()
block1():
%dims.2 : int[] = prim::rangelist(%dim.2) # <string>:338:24
%153 : int = aten::sub(%dim.2, %19) # <string>:339:28
%154 : int[] = aten::_set_item(%dims.2, %16, %153) # <string>:339:17
%155 : int = aten::sub(%dim.2, %17) # <string>:340:28
%156 : int[] = aten::_set_item(%dims.2, %15, %155) # <string>:340:17
-> ()
-> ()
%dim.1 : int = aten::dim(%5) # <string>:332:19
%158 : bool = aten::eq(%dim.1, %17) # <string>:333:16
%out.14 : Tensor = prim::If(%158) # <string>:333:13
block0():
-> (%5)
block1():
%160 : bool = aten::eq(%dim.1, %19) # <string>:335:18
%out.12 : Tensor = prim::If(%160) # <string>:335:18
block0():
%out.15 : Tensor = aten::t(%5) # <string>:336:23
-> (%out.15)
block1():
%dims.1 : int[] = prim::rangelist(%dim.1) # <string>:338:24
%164 : int = aten::sub(%dim.1, %19) # <string>:339:28
%165 : int[] = aten::_set_item(%dims.1, %16, %164) # <string>:339:17
%166 : int = aten::sub(%dim.1, %17) # <string>:340:28
%167 : int[] = aten::_set_item(%dims.1, %15, %166) # <string>:340:17
%out.16 : Tensor = aten::permute(%5, %dims.1) # <string>:341:23
-> (%out.16)
-> (%out.12)
%dim1.1 : int = aten::dim(%out.14) # <string>:346:20
%dim2.1 : int = aten::dim(%134) # <string>:347:20
%dim_out.1 : int = aten::len(%other_size.1) # <string>:348:23
%172 : bool = aten::eq(%dim1.1, %13) # <string>:349:16
%173 : bool = prim::If(%172) # <string>:349:16
block0():
-> (%14)
block1():
%174 : bool = aten::eq(%dim2.1, %13) # <string>:349:29
-> (%174)
%out : Tensor = prim::If(%173) # <string>:349:13
block0():
%out.1 : Tensor = aten::mul(%out.14, %134) # <string>:350:23
-> (%out.1)
block1():
%177 : int = aten::add(%dim1.1, %dim2.1) # <string>:351:18
%178 : bool = aten::eq(%177, %dim_out.1) # <string>:351:18
%out.25 : Tensor = prim::If(%178) # <string>:351:18
block0():
%180 : bool = aten::eq(%dim2.1, %17) # <string>:352:20
%target_dim2 : int = prim::If(%180) # <string>:352:17
block0():
-> (%13)
block1():
-> (%15)
%182 : Tensor = aten::unsqueeze(%out.14, %dim1.1) # <string>:356:36
%183 : Tensor = aten::unsqueeze(%134, %target_dim2) # <string>:356:58
%out.2 : Tensor = aten::matmul(%182, %183) # <string>:356:23
-> (%out.2)
block1():
%185 : int = aten::sub(%dim1.1, %dim2.1) # <string>:357:29
%186 : bool = aten::eq(%dim_out.1, %185) # <string>:357:18
%out.24 : Tensor = prim::If(%186) # <string>:357:18
block0():
%188 : Tensor = aten::unsqueeze(%134, %dim2.1) # <string>:358:42
%189 : Tensor = aten::matmul(%out.14, %188) # <string>:358:23
%out.3 : Tensor = aten::squeeze(%189, %16) # <string>:358:23
-> (%out.3)
block1():
%191 : int = aten::sub(%dim2.1, %dim1.1) # <string>:359:29
%192 : bool = aten::eq(%dim_out.1, %191) # <string>:359:18
%out.23 : Tensor = prim::If(%192) # <string>:359:18
block0():
%194 : Tensor = aten::unsqueeze(%out.14, %15) # <string>:360:36
%195 : Tensor = aten::matmul(%194, %134) # <string>:360:23
%out.4 : Tensor = aten::squeeze(%195, %15) # <string>:360:23
-> (%out.4)
block1():
%out.5 : Tensor = aten::matmul(%out.14, %134) # <string>:362:23
-> (%out.5)
-> (%out.23)
-> (%out.24)
-> (%out.25)
%grad_other.1 : Tensor = aten::_grad_sum_to_size(%out, %other_size.1) # <string>:369:26
%grad_weight.4 : Tensor = aten::t(%grad_other.1) # <string>:389:35
-> (%grad_weight.4)
-> (%grad_weight.7, %grad_bias.5)
return (%grad_weight.3, %grad_bias.3, %grad_weight, %grad_bias)
End print !
Saving [4, 3]||0
Saving [4]||0
Saving [1, 2]||0
Saving [3, 2]||0
Saving [3]||0
Saving [1, 3]||1
Saving [1, 4]||1
def forward(self,
input: Tensor) -> Tensor:
_0 = self.fc1
weight = _0.weight
bias = _0.bias
_1 = self.fc4
weight0 = _1.weight
bias0 = _1.bias
input0 = torch.linear(input, weight, bias)
x = torch.linear(input0, weight0, bias0)
x0 = torch.pow(x, 2)
return torch.sum(x0)
Unpacking tensor...[4, 3]
done
Unpacking tensor...[4]
done
Unpacking tensor...[1, 2]
done
Unpacking tensor...[3, 2]
done
Unpacking tensor...[3]
done
Unpacking tensor...[1, 3]
done
Unpacking tensor...[1, 4]
done
Forwarding into jit module
Forward code:
graph(%0 : int[],
%1 : Double(*, *),
%2 : int[],
%3 : Double(*, *),
%4 : Double(*, *),
%5 : Double(*, *),
%6 : float,
%7 : Double(*, *),
%8 : int):
%9 : None = prim::Constant() # <string>:274:28
%10 : int = prim::Constant[value=1]() # <string>:193:92
%11 : Double(*, *) = aten::mul(%7, %8) # <string>:193:33
%12 : Double(*, *) = aten::pow(%5, %6) # <string>:193:58
%grad_self.13 : Double(*, *) = aten::mul(%11, %12) # <string>:193:33
%14 : int[] = aten::size(%12) # <string>:2:94
%15 : int[] = aten::size(%grad_self.13) # <string>:2:105
%16 : int[]? = aten::_size_if_not_equal(%14, %15) # <string>:2:69
%grad_input.7 : Double(*, *) = aten::mm(%grad_self.13, %4) # <string>:385:34
%18 : Double(*, *) = aten::t(%grad_input.7) # <string>:247:20
%grad_weight.1 : Double(*, *) = aten::mm(%18, %3) # <string>:386:35
%self_size.3 : int[] = aten::size(%grad_input.7) # <string>:275:29
%result.3 : Tensor = aten::_grad_sum_to_size(%grad_input.7, %2) # <string>:279:22
%22 : Double(*, *) = aten::t(%grad_self.13) # <string>:247:20
%grad_weight.5 : Double(*, *) = aten::mm(%22, %1) # <string>:386:35
%self_size.2 : int[] = aten::size(%grad_self.13) # <string>:275:29
%result.2 : Tensor = aten::_grad_sum_to_size(%grad_self.13, %0) # <string>:279:22
return (%result.2, %grad_weight.5, %result.3, %grad_weight.1, %11, %10, %grad_self.13, %16, %grad_input.7, %9, %self_size.3, %22, %self_size.2)
Backward code:
graph(%0 : Tensor,
%1 : Double(*, *),
%2 : Tensor,
%3 : Double(*, *),
%4 : Double(*, *),
%5 : Double(*, *),
%6 : int[],
%7 : Double(*, *),
%8 : int[],
%9 : Double(*, *),
%10 : Double(*, *),
%11 : Double(*, *),
%12 : float,
%13 : Double(*, *),
%14 : int,
%grad_self.11 : Double(*, *),
%16 : int[]?,
%grad_input.4 : Double(*, *),
%18 : None,
%self_size.20 : int[]?,
%20 : Double(*, *),
%self_size.18 : int[]?):
%22 : None = prim::Constant() # <string>:281:33
%23 : bool = prim::Constant[value=0]()
%24 : float = prim::Constant[value=0]() # <string>:190:39
%grad_input.3 : Tensor = prim::GradOf[name="aten::_grad_sum_to_size"](%0)
block0():
%26 : bool = aten::__is__(%self_size.18, %22) # <string>:281:20
%grad_input.2 : Tensor = prim::If(%26) # <string>:281:17
block0():
-> (%0)
block1():
%self_size.12 : int[] = prim::unchecked_unwrap_optional(%self_size.18)
%grad_input.5 : Tensor = aten::expand(%0, %self_size.12, %23) # <string>:284:34
-> (%grad_input.5)
-> (%grad_input.2)
%grad_self.4 : Tensor, %grad_mat2.3 : Tensor = prim::GradOf[name="aten::mm"](%1)
block0():
%32 : Tensor = aten::t(%7) # <string>:192:28
%grad_self.3 : Tensor = aten::mm(%1, %32) # <string>:192:20
%34 : Tensor = aten::t(%20) # <string>:195:20
%grad_mat2.2 : Tensor = aten::mm(%34, %1) # <string>:195:20
-> (%grad_self.3, %grad_mat2.2)
%36 : Tensor = prim::GradOf[name="aten::t"](%grad_self.4)
block0():
%37 : Tensor = aten::t(%grad_self.4) # <string>:245:24
-> (%37)
%38 : Tensor = prim::AutogradAdd(%grad_input.3, %36)
%grad_input : Tensor = prim::GradOf[name="aten::_grad_sum_to_size"](%2)
block0():
%40 : bool = aten::__is__(%self_size.20, %22) # <string>:281:20
%grad_input.6 : Tensor = prim::If(%40) # <string>:281:17
block0():
-> (%2)
block1():
%self_size.11 : int[] = prim::unchecked_unwrap_optional(%self_size.20)
%grad_input.1 : Tensor = aten::expand(%2, %self_size.11, %23) # <string>:284:34
-> (%grad_input.1)
-> (%grad_input.6)
%grad_self.6 : Tensor = prim::GradOf[name="aten::mm"](%3)
block0():
%45 : Tensor = aten::t(%9) # <string>:192:28
%grad_self.5 : Tensor = aten::mm(%3, %45) # <string>:192:20
-> (%grad_self.5)
%47 : Tensor = prim::GradOf[name="aten::t"](%grad_self.6)
block0():
%48 : Tensor = aten::t(%grad_self.6) # <string>:245:24
-> (%48)
%49 : Tensor = prim::AutogradAdd(%grad_input, %47)
%50 : Tensor = prim::AutogradAdd(%5, %49)
%grad_self.8 : Tensor, %grad_mat2.1 : Tensor = prim::GradOf[name="aten::mm"](%50)
block0():
%53 : Tensor = aten::t(%10) # <string>:192:28
%grad_self.7 : Tensor = aten::mm(%50, %53) # <string>:192:20
%55 : Tensor = aten::t(%grad_self.11) # <string>:195:20
%grad_mat2.6 : Tensor = aten::mm(%55, %50) # <string>:195:20
-> (%grad_self.7, %grad_mat2.6)
%57 : Tensor = prim::AutogradAdd(%38, %grad_self.8)
%58 : Tensor = prim::AutogradAdd(%4, %57)
%grad_other.1 : Tensor = prim::GradOf[name="aten::mul"](%58)
block0():
%60 : Tensor = aten::mul(%58, %13) # <string>:12:31
%grad_other.2 : Tensor = aten::_grad_sum_to_size(%60, %16) # <string>:12:31
-> (%grad_other.2)
%grad_self : Tensor = prim::GradOf[name="aten::pow"](%grad_other.1)
block0():
%63 : float = prim::Float(%12) # <string>:190:20
%64 : bool = aten::eq(%63, %24) # <string>:190:20
%grad_self.12 : Tensor = prim::If(%64) # <string>:190:17
block0():
%grad_self.1 : Tensor = aten::zeros_like(%11) # <string>:191:33
-> (%grad_self.1)
block1():
%67 : Tensor = aten::mul(%grad_other.1, %12) # <string>:193:33
%68 : float = prim::Float(%12) # <string>:193:74
%69 : float = aten::sub(%68, %14) # <string>:193:74
%70 : Tensor = aten::pow(%11, %69) # <string>:193:58
%grad_self.2 : Tensor = aten::mul(%67, %70) # <string>:193:33
-> (%grad_self.2)
-> (%grad_self.12)
return (%grad_mat2.3, %grad_mat2.1, %grad_self)
End print !
Saving [1, 3]||0
Saving [1, 2]||0
Saving [4, 3]||0
Saving [1, 4]||0
Saving [1, 4]||1
Saving [1, 4]||1
Saving [1, 3]||1
Saving [4, 1]||1
grad torch.Size([3, 2])
grad torch.Size([3])
grad torch.Size([4, 3])
grad torch.Size([4])
FW/BW ok
Unpacking tensor...[1, 3]
done
Unpacking tensor...[1, 2]
done
Unpacking tensor...[4, 3]
done
Unpacking tensor...[1, 4]
done
Unpacking tensor...[1, 4]
done
Unpacking tensor...[1, 4]
done
Unpacking tensor...[1, 3]
done
Unpacking tensor...[4, 1]
[4, 1]
0 libtorch.dylib 0x0000000118afc004 _ZNK5torch8autograd13SavedVariable6unpackENSt3__110shared_ptrINS0_8FunctionEEE + 2244
1 libtorch.dylib 0x0000000118b566fe _ZN5torch3jit12_GLOBAL__N_127DifferentiableGraphBackward5applyEONSt3__16vectorINS_8autograd8VariableENS3_9allocatorIS6_EEEE + 1630
2 libtorch.dylib 0x0000000118ae73df _ZN5torch8autograd8FunctionclEONSt3__16vectorINS0_8VariableENS2_9allocatorIS4_EEEE + 367
3 libtorch.dylib 0x0000000118adf7fc _ZN5torch8autograd6Engine17evaluate_functionERNS0_12FunctionTaskE + 1100
4 libtorch.dylib 0x0000000118adec79 _ZN5torch8autograd6Engine11thread_mainEPNS0_9GraphTaskE + 217
5 libtorch.dylib 0x0000000118adeb88 _ZN5torch8autograd6Engine11thread_initEi + 136
6 libtorch_python.dylib 0x0000000115c3d05c _ZN5torch8autograd6python12PythonEngine11thread_initEi + 44
7 libtorch.dylib 0x0000000118aead02 _ZNSt3__114__thread_proxyINS_5tupleIJNS_10unique_ptrINS_15__thread_structENS_14default_deleteIS3_EEEEMN5torch8autograd6EngineEFviEPS9_iEEEEEPvSE_ + 66
8 libsystem_pthread.dylib 0x00007fff79a2c2eb _pthread_body + 126
9 libsystem_pthread.dylib 0x00007fff79a2f249 _pthread_start + 66
10 libsystem_pthread.dylib 0x00007fff79a2b40d thread_start + 13
Traceback (most recent call last):
File "foo.py", line 49, in <module>
torch.autograd.grad(s, model.parameters())
File "/Users/albandes/workspace/pytorch_dev/torch/autograd/__init__.py", line 149, in grad
inputs, allow_unused)
RuntimeError: No grad accumulator for a saved leaf!
