File size: 8,777 Bytes
fbb20ff |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 |
from copy import deepcopy
import torch
import torch.nn as nn
from hydra.utils import instantiate
from genmo.diffusion_utils.model_util import create_gaussian_diffusion
from genmo.diffusion_utils.resample import create_named_schedule_sampler
from genmo.utils.net_utils import length_to_mask
from .genmo_cfg_sampler import ClassifierFreeSampleModel
class GENMODiffusion(nn.Module):
def __init__(
self,
model_cfg,
max_len=120,
# condition
cliffcam_dim=3,
cam_angvel_dim=6,
cam_t_vel_dim=3,
imgseq_dim=1024,
observed_motion_3d_dim=151,
encoded_music_dim=438,
encoded_audio_dim=128,
latent_dim=512,
dropout=0.1,
args=None,
cond_merge_strategy="add",
cond_exists_dim=512,
music_mask_prob=0.1,
img_process_modules=None,
img_process_modules_enable_grad={},
multi_text_module_cfg={},
**kwargs,
):
super().__init__()
self.model_cfg = model_cfg
self.args = args
self.max_len = max_len
self.regression_input_type = self.args.get("regression_input_type", "zero")
self.denoiser = instantiate(self.model_cfg.denoiser)
self.init_diffusion()
self.text_encoder, self.tokenizer = None, None
def init_diffusion(self):
self.train_diffusion = create_gaussian_diffusion(
self.model_cfg.diffusion, training=True
)
self.test_diffusion = create_gaussian_diffusion(
self.model_cfg.diffusion, training=False
)
gen_only_diffusion = deepcopy(self.model_cfg.diffusion)
gen_only_diffusion.test_timestep_respacing = self.model_cfg.diffusion.get(
"gen_only_test_timestep_respacing", "50"
)
print(
f"Gen only test timestep respacing: {gen_only_diffusion.test_timestep_respacing}"
)
self.test_gen_only_diffusion = create_gaussian_diffusion(
gen_only_diffusion, training=False
)
self.schedule_sampler = create_named_schedule_sampler(
self.model_cfg.diffusion.schedule_sampler_type, self.train_diffusion
)
return
def forward_train(self, inputs, mode):
assert self.training, "forward_train should only be called during training"
diffusion = self.train_diffusion if self.training else self.test_diffusion
length = inputs["length"]
# target_x = inputs["target_x"]
motion = inputs["motion"]
f_cond = inputs["f_cond"]
B, L, _ = motion.shape
vis_mask = length_to_mask(length, L) # (B, L)
valid_mask = inputs["mask"]["valid"]
assert (vis_mask == valid_mask).all()
denoiser_kwargs = {
"y": {
"text": inputs.get("caption", [""] * B),
"f_cond": f_cond,
"mask": vis_mask,
"length": length,
},
"inputs": inputs,
}
if "encoded_text" in inputs:
denoiser_kwargs["y"]["encoded_text"] = inputs["encoded_text"]
if "observed_motion_3d" in inputs:
denoiser_kwargs["observed_motion_3d"] = inputs["observed_motion_3d"]
denoiser_kwargs["motion_mask_3d"] = inputs["motion_mask_3d"]
denoiser_kwargs["rm_text_flag"] = inputs["rm_text_flag"]
if mode == "regression":
t = (
(torch.ones(B) * (diffusion.original_num_steps - 1))
.long()
.to(motion.device)
)
t_weights = torch.ones(B).to(motion.device)
x_start = motion
if self.regression_input_type == "zero":
x_t = torch.zeros_like(motion)
elif self.regression_input_type == "normal":
x_t = torch.randn_like(motion)
else:
raise ValueError(
f"Unsupported regression_input_type: {self.regression_input_type}"
)
elif mode == "diffusion":
t, t_weights = self.schedule_sampler.sample(motion.shape[0], motion.device)
if "regression_outputs" in inputs:
pred_x_start_regression = inputs["regression_outputs"]["model_output"][
"pred_x_start"
].detach()
else:
raise ValueError("No regression outputs found")
# pred_x_start_regression = torch.zeros_like(motion)
x_start_reg = pred_x_start_regression.clone()
x_start = motion.clone()
x_start[inputs["mask"]["2d_only"]] = x_start_reg[inputs["mask"]["2d_only"]]
# regression_mask = (
# torch.rand(B).to(motion.device) < self.args.use_regression_outputs_prob
# ).float()
# if "gen_only" in inputs and self.args.get("use_gt_for_gen_only", True):
# regression_mask[inputs["gen_only"]] = 0
# x_start = x_start_reg * regression_mask[:, None, None] + x_start_gt * (
# 1 - regression_mask[:, None, None]
# )
noise = torch.randn_like(x_start)
x_t = self.train_diffusion.q_sample(x_start.clone(), t, noise=noise)
denoise_out = self.denoiser(
x_t, diffusion._scale_timesteps(t), return_aux=False, **denoiser_kwargs
)
output = {
"target_x_start": x_start,
"t_weights": t_weights,
}
output.update(denoise_out)
for x in self.args.out_attr:
assert x in output, f"Output {x} not found in denoise_out"
return output
def forward_test(self, inputs, progress=False):
assert not self.training, "forward_test should only be called during inference"
diffusion = self.test_gen_only_diffusion
denoiser = self.denoiser
length = inputs["length"]
B, L = inputs["B"], inputs["L"]
motion = inputs["motion"]
f_cond, f_uncond = inputs["f_cond"], inputs["f_uncond"]
vis_mask = length_to_mask(length, L) # (B, L)
denoiser_kwargs = {
"y": {
"text": inputs.get("caption", [""] * B),
"f_cond": f_cond,
"f_uncond": f_uncond,
"mask": vis_mask,
"length": length,
},
"inputs": inputs,
}
if "encoded_text" in inputs:
denoiser_kwargs["y"]["encoded_text"] = inputs["encoded_text"]
if "meta" in inputs and "multi_text_data" in inputs["meta"][0]:
denoiser_kwargs["y"]["multi_text_data"] = inputs["meta"][0][
"multi_text_data"
]
if "observed_motion_3d" in inputs:
denoiser_kwargs["observed_motion_3d"] = inputs["observed_motion_3d"]
denoiser_kwargs["motion_mask_3d"] = inputs["motion_mask_3d"]
denoiser_kwargs["rm_text_flag"] = inputs.get("rm_text_flag", None)
if self.args.get("use_cfg_sampler_for_gen", False):
denoiser = ClassifierFreeSampleModel(denoiser)
denoiser_kwargs["y"]["scale"] = self.model_cfg.diffusion.guidance_param
diff_sampler = self.model_cfg.diffusion.get("sampler", "ddim")
if diff_sampler == "ddim":
sample_fn = diffusion.ddim_sample_loop_with_aux
kwargs = {"eta": self.model_cfg.diffusion.ddim_eta}
else:
raise NotImplementedError(f"Sampler {diff_sampler} not implemented")
if self.args.get("force_zero_noise", False):
noise = torch.zeros_like(motion)
elif self.args.get("force_rand_noise", False):
noise = torch.randn_like(motion)
else:
noise = torch.randn_like(motion)
if self.args.get("return_mid", False):
kwargs["return_mid"] = True
denoise_out = sample_fn(
denoiser,
motion.shape,
clip_denoised=False,
model_kwargs=denoiser_kwargs,
skip_timesteps=0, # 0 is the default value - i.e. don't skip any step
init_image=None,
progress=progress,
dump_steps=None,
noise=noise,
const_noise=False,
**kwargs,
)
output = denoise_out.copy()
for x in self.args.out_attr:
assert x in output, f"Output {x} not found in denoise_out"
return output
def forward(
self,
inputs,
train=False,
postproc=False,
static_cam=False,
mode=None,
test_mode=None,
normalizer_stats=None,
):
if train:
return self.forward_train(inputs, mode=mode)
else:
return self.forward_test(inputs)
|