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

Get rid of unused codes in vq

Browse Source
This commit is contained in:
James Betker 2022-03-03 13:41:38 -07:00
parent 619da9ea28
commit 3cd6c7f428
1 changed files with 1 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
codes/models/vqvae/vqvae.py
View File

@ -29,7 +29,7 @@ from utils.util import checkpoint, opt_get
class Quantize(nn.Module): class Quantize(nn.Module):
def __init__(self, dim, n_embed, decay=0.99, eps=1e-5, balancing_heuristic=False, new_return_order=False): def __init__(self, dim, n_embed, decay=0.99, eps=1e-5, new_return_order=False):
super().__init__() super().__init__()
self.dim = dim self.dim = dim
@ -37,10 +37,7 @@ class Quantize(nn.Module):
self.decay = decay self.decay = decay
self.eps = eps self.eps = eps
self.balancing_heuristic = balancing_heuristic
self.codes = None self.codes = None
self.max_codes = 64000
self.codes_full = False
self.new_return_order = new_return_order self.new_return_order = new_return_order
embed = torch.randn(dim, n_embed) embed = torch.randn(dim, n_embed)
@ -49,20 +46,6 @@ class Quantize(nn.Module):
self.register_buffer("embed_avg", embed.clone()) self.register_buffer("embed_avg", embed.clone())
def forward(self, input, return_soft_codes=False): def forward(self, input, return_soft_codes=False):
if self.balancing_heuristic and self.codes_full:
h = torch.histc(self.codes, bins=self.n_embed, min=0, max=self.n_embed) / len(self.codes)
mask = torch.logical_or(h > .9, h < .01).unsqueeze(1)
ep = self.embed.permute(1,0)
ea = self.embed_avg.permute(1,0)
rand_embed = torch.randn_like(ep) * mask
self.embed = (ep * ~mask + rand_embed).permute(1,0)
self.embed_avg = (ea * ~mask + rand_embed).permute(1,0)
self.cluster_size = self.cluster_size * ~mask.squeeze()
if torch.any(mask):
print(f"Reset {torch.sum(mask)} embedding codes.")
self.codes = None
self.codes_full = False
flatten = input.reshape(-1, self.dim) flatten = input.reshape(-1, self.dim)
dist = ( dist = (
flatten.pow(2).sum(1, keepdim=True) flatten.pow(2).sum(1, keepdim=True)
@ -75,15 +58,6 @@ class Quantize(nn.Module):
embed_ind = embed_ind.view(*input.shape[:-1]) embed_ind = embed_ind.view(*input.shape[:-1])
quantize = self.embed_code(embed_ind) quantize = self.embed_code(embed_ind)
if self.balancing_heuristic:
if self.codes is None:
self.codes = embed_ind.flatten()
else:
self.codes = torch.cat([self.codes, embed_ind.flatten()])
if len(self.codes) > self.max_codes:
self.codes = self.codes[-self.max_codes:]
self.codes_full = True
if self.training: if self.training:
embed_onehot_sum = embed_onehot.sum(0) embed_onehot_sum = embed_onehot.sum(0)
embed_sum = flatten.transpose(0, 1) @ embed_onehot embed_sum = flatten.transpose(0, 1) @ embed_onehot