Trying to use torch.compile with mode=max-autotune on this model code:
import torch
from torch import nn
import torch.nn.functional as F
from src.facerender.modules.util import SameBlock2d, DownBlock2d, ResBlock3d, SPADEResnetBlock
from src.facerender.modules.dense_motion import DenseMotionNetwork
class SPADEDecoder(nn.Module):
def __init__(self):
super().__init__()
ic = 256
oc = 64
norm_G = 'spadespectralinstance'
label_nc = 256
self.fc = nn.Conv2d(ic, 2 * ic, 3, padding=1)
self.G_middle_0 = SPADEResnetBlock(2 * ic, 2 * ic, norm_G, label_nc)
self.G_middle_1 = SPADEResnetBlock(2 * ic, 2 * ic, norm_G, label_nc)
self.G_middle_2 = SPADEResnetBlock(2 * ic, 2 * ic, norm_G, label_nc)
self.G_middle_3 = SPADEResnetBlock(2 * ic, 2 * ic, norm_G, label_nc)
self.G_middle_4 = SPADEResnetBlock(2 * ic, 2 * ic, norm_G, label_nc)
self.G_middle_5 = SPADEResnetBlock(2 * ic, 2 * ic, norm_G, label_nc)
self.up_0 = SPADEResnetBlock(2 * ic, ic, norm_G, label_nc)
self.up_1 = SPADEResnetBlock(ic, oc, norm_G, label_nc)
self.conv_img = nn.Conv2d(oc, 3, 3, padding=1)
self.up = nn.Upsample(scale_factor=2)
def forward(self, feature):
seg = feature
x = self.fc(feature)
x = self.G_middle_0(x, seg)
x = self.G_middle_1(x, seg)
x = self.G_middle_2(x, seg)
x = self.G_middle_3(x, seg)
x = self.G_middle_4(x, seg)
x = self.G_middle_5(x, seg)
x = self.up(x)
x = self.up_0(x, seg) # 256, 128, 128
x = self.up(x)
x = self.up_1(x, seg) # 64, 256, 256
x = self.conv_img(F.leaky_relu(x, 2e-1))
# x = torch.tanh(x)
x = F.sigmoid(x)
return x
class OcclusionAwareSPADEGenerator(nn.Module):
def __init__(self, image_channel, feature_channel, num_kp, block_expansion, max_features, num_down_blocks, reshape_channel, reshape_depth,
num_resblocks, estimate_occlusion_map=False, dense_motion_params=None, estimate_jacobian=False):
super(OcclusionAwareSPADEGenerator, self).__init__()
if dense_motion_params is not None:
self.dense_motion_network = DenseMotionNetwork(num_kp=num_kp, feature_channel=feature_channel,
estimate_occlusion_map=estimate_occlusion_map,
**dense_motion_params)
else:
self.dense_motion_network = None
self.first = SameBlock2d(image_channel, block_expansion, kernel_size=(3, 3), padding=(1, 1))
down_blocks = []
for i in range(num_down_blocks):
in_features = min(max_features, block_expansion * (2 ** i))
out_features = min(max_features, block_expansion * (2 ** (i + 1)))
down_blocks.append(DownBlock2d(in_features, out_features, kernel_size=(3, 3), padding=(1, 1)))
self.down_blocks = nn.ModuleList(down_blocks)
self.second = nn.Conv2d(in_channels=out_features, out_channels=max_features, kernel_size=1, stride=1)
self.reshape_channel = reshape_channel
self.reshape_depth = reshape_depth
self.resblocks_3d = torch.nn.Sequential()
for i in range(num_resblocks):
self.resblocks_3d.add_module('3dr' + str(i), ResBlock3d(reshape_channel, kernel_size=3, padding=1))
out_features = block_expansion * (2 ** (num_down_blocks))
self.third = SameBlock2d(max_features, out_features, kernel_size=(3, 3), padding=(1, 1), lrelu=True)
self.fourth = nn.Conv2d(in_channels=out_features, out_channels=out_features, kernel_size=1, stride=1)
self.estimate_occlusion_map = estimate_occlusion_map
self.image_channel = image_channel
self.decoder = SPADEDecoder()
def deform_input(self, inp, deformation):
_, d_old, h_old, w_old, _ = deformation.shape
_, _, d, h, w = inp.shape
if d_old != d or h_old != h or w_old != w:
deformation = deformation.permute(0, 4, 1, 2, 3)
deformation = F.interpolate(deformation, size=(d, h, w), mode='trilinear')
deformation = deformation.permute(0, 2, 3, 4, 1)
return F.grid_sample(inp, deformation)
def forward(self, source_image, kp_driving, kp_source):
# Encoding (downsampling) part
out = self.first(source_image)
for i in range(len(self.down_blocks)):
out = self.down_blocks[i](out)
out = self.second(out)
bs, c, h, w = out.shape
# print(out.shape)
feature_3d = out.view(bs, self.reshape_channel, self.reshape_depth, h ,w)
feature_3d = self.resblocks_3d(feature_3d)
# Transforming feature representation according to deformation and occlusion
output_dict = {}
if self.dense_motion_network is not None:
dense_motion = self.dense_motion_network(feature=feature_3d, kp_driving=kp_driving,
kp_source=kp_source)
output_dict['mask'] = dense_motion['mask']
# import pdb; pdb.set_trace()
if 'occlusion_map' in dense_motion:
occlusion_map = dense_motion['occlusion_map']
output_dict['occlusion_map'] = occlusion_map
else:
occlusion_map = None
deformation = dense_motion['deformation']
out = self.deform_input(feature_3d, deformation)
bs, c, d, h, w = out.shape
out = out.view(bs, c*d, h, w)
out = self.third(out)
out = self.fourth(out)
# occlusion_map = torch.where(occlusion_map < 0.95, 0, occlusion_map)
if occlusion_map is not None:
if out.shape[2] != occlusion_map.shape[2] or out.shape[3] != occlusion_map.shape[3]:
occlusion_map = F.interpolate(occlusion_map, size=out.shape[2:], mode='bilinear')
out = out * occlusion_map
# Decoding part
out = self.decoder(out)
output_dict["prediction"] = out
return output_dict
The calling part:
out = generator(source_image, kp_source=kp_source, kp_driving=kp_norm)
outputs.append(out["prediction"])
This results in this error:
File "/home/vardan/Desktop/experiments/marketing.ai/pitch_god/video_creation/SadTalker/inference.py", line 177, in <module>
obj.run()
File "/home/vardan/Desktop/experiments/marketing.ai/pitch_god/video_creation/SadTalker/inference.py", line 113, in run
result = self.animate_from_coeff.generate(data, self.save_dir, self.pic_path, crop_info, \
File "/home/vardan/Desktop/experiments/marketing.ai/pitch_god/video_creation/SadTalker/src/facerender/animate.py", line 173, in generate
predictions_video = make_animation(source_image, source_semantics, target_semantics,
File "/home/vardan/Desktop/experiments/marketing.ai/pitch_god/video_creation/SadTalker/src/facerender/modules/make_animation.py", line 147, in make_animation
predictions_ts = torch.stack(predictions, dim=1)
RuntimeError: Error: accessing tensor output of CUDAGraphs that has been overwritten by a subsequent run. Stack trace: File "/home/vardan/Desktop/experiments/marketing.ai/pitch_god/video_creation/SadTalker/src/facerender/modules/generator.py", line 139, in forward
out = self.decoder(out)
File "/home/vardan/.local/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1520, in _call_impl
return forward_call(*args, **kwargs)
File "/home/vardan/Desktop/experiments/marketing.ai/pitch_god/video_creation/SadTalker/src/facerender/modules/generator.py", line 44, in forward
x = F.sigmoid(x). To prevent overwriting, clone the tensor outside of torch.compile() or call torch.compiler.cudagraph_mark_step_begin() before each model invocation.
On changing the result list append line from outputs.append(out["prediction"]) to outputs.append(out["prediction"].clone()) the code works but the benefit obtained in chosing the mode max-autotune vs default (in which the code was working) is lost.
Is there any other change I can do to keep that inference optimization benefit with max-autotune?