kxngluca/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

More mods

Browse Source
This commit is contained in:
James Betker 2022-06-19 22:30:46 -06:00
parent 691ed196da
commit c5ea2bee52
1 changed files with 6 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
codes/models/audio/music/unet_diffusion_waveform_gen3.py
View File

@ -150,8 +150,6 @@ class DiffusionWaveformGen(nn.Module):
:param dropout: the dropout probability.
:param channel_mult: channel multiplier for each level of the UNet.
:param dims: determines if the signal is 1D, 2D, or 3D.
:param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
:param resblock_updown: use residual blocks for up/downsampling.
"""
def __init__(
@ -166,7 +164,6 @@ class DiffusionWaveformGen(nn.Module):
num_res_blocks=(1,1,0),
token_conditioning_resolutions=(1,4),
mid_resnet_depth=10,
dims=1,
use_fp16=False,
time_embed_dim_multiplier=1,
# Parameters for regularization.
@ -179,7 +176,6 @@ class DiffusionWaveformGen(nn.Module):
self.out_channels = out_channels
self.dropout = dropout
self.channel_mult = channel_mult
self.dims = dims
self.unconditioned_percentage = unconditioned_percentage
self.enable_fp16 = use_fp16
self.alignment_size = 2 ** (len(channel_mult)+1)
@ -203,7 +199,7 @@ class DiffusionWaveformGen(nn.Module):
self.input_blocks = nn.ModuleList(
[
TimestepEmbedSequential(
conv_nd(dims, in_channels, model_channels, 3, padding=1)
conv_nd(1, in_channels, model_channels, 3, padding=1)
)
]
)
@ -227,7 +223,7 @@ class DiffusionWaveformGen(nn.Module):
time_embed_dim,
dropout,
out_channels=int(mult * model_channels),
dims=dims,
dims=1,
kernel_size=3,
use_scale_shift_norm=True,
)
@ -241,7 +237,7 @@ class DiffusionWaveformGen(nn.Module):
self.input_blocks.append(
TimestepEmbedSequential(
Downsample(
ch, True, dims=dims, out_channels=out_ch, factor=2, ksize=3, pad=1
ch, True, dims=1, out_channels=out_ch, factor=2, ksize=3, pad=1
)
)
)
@ -264,7 +260,7 @@ class DiffusionWaveformGen(nn.Module):
time_embed_dim,
dropout,
out_channels=int(model_channels * mult),
dims=dims,
dims=1,
kernel_size=3,
use_scale_shift_norm=True,
)
@ -273,7 +269,7 @@ class DiffusionWaveformGen(nn.Module):
if level and i == num_blocks:
out_ch = ch
layers.append(
Upsample(ch, True, dims=dims, out_channels=out_ch, factor=2)
Upsample(ch, True, dims=1, out_channels=out_ch, factor=2)
)
ds //= 2
self.output_blocks.append(TimestepEmbedSequential(*layers))
@ -282,7 +278,7 @@ class DiffusionWaveformGen(nn.Module):
self.out = nn.Sequential(
normalization(ch),
nn.SiLU(),
zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
zero_module(conv_nd(1, model_channels, out_channels, 3, padding=1)),
)
def get_grad_norm_parameter_groups(self):
@ -355,13 +351,6 @@ class DiffusionWaveformGen(nn.Module):
return out[:, :, :orig_x_shape]
def before_step(self, step):
# The middle block traditionally gets really small gradients; scale them up by an order of magnitude.
scaled_grad_parameters = self.middle_block.parameters()
for p in scaled_grad_parameters:
if hasattr(p, 'grad') and p.grad is not None:
p.grad *= 10
@register_model
def register_unet_diffusion_waveform_gen3(opt_net, opt):