Distill_torchscript mods

Starts down the path of writing a custom trace that works using torch's hook mechanism.
This commit is contained in:
James Betker 2020-06-27 08:28:09 -06:00
parent db08dedfe2
commit 6f2bc36c61

View File

@ -1,9 +1,91 @@
import argparse
import functools
import torch
import options.options as option
from models.networks import define_G
import torch
import torchvision
import torch.nn.functional as F
class TracedModule:
def __init__(self, idname):
self.idname = idname
self.traced_outputs = []
self.traced_inputs = []
class TorchCustomTrace:
def __init__(self):
self.module_name_counter = {}
self.modules = {}
self.graph = {}
self.module_map_by_inputs = {}
self.module_map_by_outputs = {}
self.inputs_to_func_output_tuple = {}
def add_tracked_module(self, mod: torch.nn.Module):
modname = type(mod).__name__
if modname not in self.module_name_counter.keys():
self.module_name_counter[modname] = 0
self.module_name_counter[modname] += 1
idname = "%s(%03d)" % (modname, self.module_name_counter[modname])
self.modules[idname] = TracedModule(idname)
return idname
# Only called for nn.Modules since those are the only things we can access. Filling in the gaps will be done in
# the backwards pass.
def mem_forward_hook(self, module: torch.nn.Module, inputs, outputs, trace: str, mod_id: str):
mod = self.modules[mod_id]
for li in inputs:
if type(li) == torch.Tensor:
li = [li]
if type(li) == list:
for i in li:
if i.data_ptr() in self.module_map_by_inputs.keys():
self.module_map_by_inputs[i.data_ptr()].append(mod)
else:
self.module_map_by_inputs[i.data_ptr()] = [mod]
for o in outputs:
if o.data_ptr() in self.module_map_by_inputs.keys():
self.module_map_by_inputs[o.data_ptr()].append(mod)
else:
self.module_map_by_inputs[o.data_ptr()] = [mod]
# print(trace, [i.data_ptr() for i in inputs], [o.data_ptr() for o in outputs])
def mem_backward_hook(self, inputs, outputs, op):
if len(inputs) == 0:
print("No inputs.. %s" % (op,))
outs = [o.data_ptr() for o in outputs]
tup = (outs, op)
#print(tup)
for li in inputs:
if type(li) == torch.Tensor:
li = [li]
if type(li) == list:
for i in li:
if i.data_ptr() in self.module_map_by_inputs.keys():
print("%i: [%s] {%s}" % (i.data_ptr(), op, [n.idname for n in self.module_map_by_inputs[i.data_ptr()]]))
if i.data_ptr() in self.inputs_to_func_output_tuple.keys():
self.inputs_to_func_output_tuple[i.data_ptr()].append(tup)
else:
self.inputs_to_func_output_tuple[i.data_ptr()] = [tup]
def install_hooks(self, mod: torch.nn.Module, trace=""):
mod_id = self.add_tracked_module(mod)
my_trace = trace + "->" + mod_id
# If this module has parameters, it also has a state worth tracking.
#if next(mod.parameters(recurse=False), None) is not None:
mod.register_forward_hook(functools.partial(self.mem_forward_hook, trace=my_trace, mod_id=mod_id))
for m in mod.children():
self.install_hooks(m, my_trace)
def install_backward_hooks(self, grad_fn):
# AccumulateGrad simply pushes a gradient into the specified variable, and isn't useful for the purposes of
# tracing the graph.
if grad_fn is None or "AccumulateGrad" in str(grad_fn):
return
grad_fn.register_hook(functools.partial(self.mem_backward_hook, op=str(grad_fn)))
for g, _ in grad_fn.next_functions:
self.install_backward_hooks(g)
if __name__ == "__main__":
@ -13,9 +95,9 @@ if __name__ == "__main__":
opt = option.dict_to_nonedict(opt)
netG = define_G(opt)
dummyInput = torch.rand(1,3,8,8)
dummyInput = torch.rand(1,3,32,32)
mode = 'trace'
mode = 'memtrace'
if mode == 'torchscript':
print("Tracing generator network..")
traced_netG = torch.jit.trace(netG, dummyInput)
@ -31,7 +113,16 @@ if __name__ == "__main__":
dynamic_axes = {'lr_input': {0: 'batch', 1: 'filters', 2: 'h', 3: 'w'}, 'hr_image': {0: 'batch', 1: 'filters', 2: 'h', 3: 'w'}}
torch.onnx.export(netG, dummyInput, "../results/gen.onnx", verbose=True, input_names=input_names,
output_names=output_names, dynamic_axes=dynamic_axes, opset_version=11)
output_names=output_names, dynamic_axes=dynamic_axes, opset_version=12)
elif mode == 'memtrace':
criterion = torch.nn.MSELoss()
tracer = TorchCustomTrace()
tracer.install_hooks(netG)
out, = netG(dummyInput)
tracer.install_backward_hooks(out.grad_fn)
target = torch.zeros_like(out)
loss = criterion(out, target)
loss.backward()
elif mode == 'trace':
out = netG.forward(dummyInput)[0]
print(out.shape)