bacow/GPT-SoVITS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Refactor: Format Code with Ruff and Update Deprecated G2PW Link (#2255)

Browse Source 
* ruff check --fix

* ruff format --line-length 120 --target-version py39

* Change the link for G2PW Model

* update pytorch version and colab
This commit is contained in:
XXXXRT666
2025-04-07 09:42:47 +01:00
committed by GitHub
parent 9da7e17efe
commit 53cac93589
132 changed files with 8185 additions and 6648 deletions
Download Patch File Download Diff File Expand all files Collapse all files

93
GPT_SoVITS/AR/modules/activation.py
View File

@@ -1,17 +1,14 @@
# modified from https://github.com/lifeiteng/vall-e/blob/main/valle/modules/activation.py
from typing import Optional
from typing import Tuple
from typing import Optional, Tuple
import torch
from torch import Tensor
from torch.nn import Linear
from torch.nn import Module
from torch.nn.init import constant_
from torch.nn.init import xavier_normal_
from torch.nn.init import xavier_uniform_
from torch.nn import Linear, Module
from torch.nn import functional as F
from torch.nn.init import constant_, xavier_normal_, xavier_uniform_
from torch.nn.modules.linear import NonDynamicallyQuantizableLinear
from torch.nn.parameter import Parameter
from torch.nn import functional as F
from AR.modules.patched_mha_with_cache import multi_head_attention_forward_patched
F.multi_head_attention_forward = multi_head_attention_forward_patched
@@ -73,6 +70,7 @@ class MultiheadAttention(Module):
>>> attn_output, attn_output_weights = multihead_attn(query, key, value)
"""
__constants__ = ["batch_first"]
bias_k: Optional[torch.Tensor]
bias_v: Optional[torch.Tensor]
@@ -104,9 +102,7 @@ class MultiheadAttention(Module):
self.dropout = dropout
self.batch_first = batch_first
self.head_dim = embed_dim // num_heads
assert (
self.head_dim * num_heads == self.embed_dim
), "embed_dim must be divisible by num_heads"
assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
if add_bias_kv:
self.bias_k = Parameter(torch.empty((1, 1, embed_dim), **factory_kwargs))
@@ -117,31 +113,32 @@ class MultiheadAttention(Module):
if linear1_cls == Linear:
if not self._qkv_same_embed_dim:
self.q_proj_weight = Parameter(
torch.empty((embed_dim, embed_dim), **factory_kwargs)
torch.empty((embed_dim, embed_dim), **factory_kwargs),
)
self.k_proj_weight = Parameter(
torch.empty((embed_dim, self.kdim), **factory_kwargs)
torch.empty((embed_dim, self.kdim), **factory_kwargs),
)
self.v_proj_weight = Parameter(
torch.empty((embed_dim, self.vdim), **factory_kwargs)
torch.empty((embed_dim, self.vdim), **factory_kwargs),
)
self.register_parameter("in_proj_weight", None)
else:
self.in_proj_weight = Parameter(
torch.empty((3 * embed_dim, embed_dim), **factory_kwargs)
torch.empty((3 * embed_dim, embed_dim), **factory_kwargs),
)
self.register_parameter("q_proj_weight", None)
self.register_parameter("k_proj_weight", None)
self.register_parameter("v_proj_weight", None)
if bias:
self.in_proj_bias = Parameter(
torch.empty(3 * embed_dim, **factory_kwargs)
)
self.in_proj_bias = Parameter(torch.empty(3 * embed_dim, **factory_kwargs))
else:
self.register_parameter("in_proj_bias", None)
self.out_proj = NonDynamicallyQuantizableLinear(
embed_dim, embed_dim, bias=bias, **factory_kwargs
embed_dim,
embed_dim,
bias=bias,
**factory_kwargs,
)
self._reset_parameters()
@@ -150,7 +147,10 @@ class MultiheadAttention(Module):
raise NotImplementedError
else:
self.in_proj_linear = linear1_cls(
embed_dim, 3 * embed_dim, bias=bias, **factory_kwargs
embed_dim,
3 * embed_dim,
bias=bias,
**factory_kwargs,
)
self.in_proj_weight = self.in_proj_linear.weight
@@ -164,7 +164,10 @@ class MultiheadAttention(Module):
self.register_parameter("in_proj_bias", None)
self.out_proj = linear2_cls(
embed_dim, embed_dim, bias=bias, **factory_kwargs
embed_dim,
embed_dim,
bias=bias,
**factory_kwargs,
)
if self.bias_k is not None:
@@ -261,28 +264,26 @@ class MultiheadAttention(Module):
if key_padding_mask is not None:
_kpm_dtype = key_padding_mask.dtype
if _kpm_dtype != torch.bool and not torch.is_floating_point(
key_padding_mask
key_padding_mask,
):
raise AssertionError(
"only bool and floating types of key_padding_mask are supported"
)
raise AssertionError("only bool and floating types of key_padding_mask are supported")
why_not_fast_path = ""
if not is_batched:
why_not_fast_path = (
f"input not batched; expected query.dim() of 3 but got {query.dim()}"
)
why_not_fast_path = f"input not batched; expected query.dim() of 3 but got {query.dim()}"
elif query is not key or key is not value:
# When lifting this restriction, don't forget to either
# enforce that the dtypes all match or test cases where
# they don't!
why_not_fast_path = "non-self attention was used (query, key, and value are not the same Tensor)"
elif self.in_proj_bias is not None and query.dtype != self.in_proj_bias.dtype:
why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_bias ({self.in_proj_bias.dtype}) don't match"
elif (
self.in_proj_weight is not None and query.dtype != self.in_proj_weight.dtype
):
why_not_fast_path = (
f"dtypes of query ({query.dtype}) and self.in_proj_bias ({self.in_proj_bias.dtype}) don't match"
)
elif self.in_proj_weight is not None and query.dtype != self.in_proj_weight.dtype:
# this case will fail anyway, but at least they'll get a useful error message.
why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_weight ({self.in_proj_weight.dtype}) don't match"
why_not_fast_path = (
f"dtypes of query ({query.dtype}) and self.in_proj_weight ({self.in_proj_weight.dtype}) don't match"
)
elif self.training:
why_not_fast_path = "training is enabled"
elif not self.batch_first:
@@ -300,9 +301,7 @@ class MultiheadAttention(Module):
elif attn_mask is not None:
why_not_fast_path = "attn_mask was not None"
elif query.is_nested and key_padding_mask is not None:
why_not_fast_path = (
"key_padding_mask is not supported with NestedTensor input"
)
why_not_fast_path = "key_padding_mask is not supported with NestedTensor input"
elif self.num_heads % 2 == 1:
why_not_fast_path = "num_heads is odd"
elif torch.is_autocast_enabled():
@@ -322,20 +321,10 @@ class MultiheadAttention(Module):
# generator expressions.
if torch.overrides.has_torch_function(tensor_args):
why_not_fast_path = "some Tensor argument has_torch_function"
elif not all(
[
(x is None or x.is_cuda or "cpu" in str(x.device))
for x in tensor_args
]
):
elif not all([(x is None or x.is_cuda or "cpu" in str(x.device)) for x in tensor_args]):
why_not_fast_path = "some Tensor argument is neither CUDA nor CPU"
elif torch.is_grad_enabled() and any(
[x is not None and x.requires_grad for x in tensor_args]
):
why_not_fast_path = (
"grad is enabled and at least one of query or the "
"input/output projection weights or biases requires_grad"
)
elif torch.is_grad_enabled() and any([x is not None and x.requires_grad for x in tensor_args]):
why_not_fast_path = "grad is enabled and at least one of query or the input/output projection weights or biases requires_grad"
if not why_not_fast_path:
return torch._native_multi_head_attention(
query,
@@ -350,11 +339,7 @@ class MultiheadAttention(Module):
key_padding_mask if key_padding_mask is not None else attn_mask,
need_weights,
average_attn_weights,
1
if key_padding_mask is not None
else 0
if attn_mask is not None
else None,
1 if key_padding_mask is not None else 0 if attn_mask is not None else None,
)
any_nested = query.is_nested or key.is_nested or value.is_nested

52
GPT_SoVITS/AR/modules/activation_onnx.py
View File

@@ -1,17 +1,13 @@
# modified from https://github.com/lifeiteng/vall-e/blob/main/valle/modules/activation.py
from typing import Optional
from typing import Tuple
from typing import Optional, Tuple
import torch
from torch import Tensor
from torch.nn import Linear
from torch.nn import Module
from torch.nn.init import constant_
from torch.nn.init import xavier_normal_
from torch.nn.init import xavier_uniform_
from torch.nn import Linear, Module
from torch.nn.init import constant_, xavier_normal_, xavier_uniform_
from torch.nn.modules.linear import NonDynamicallyQuantizableLinear
from torch.nn.parameter import Parameter
from torch.nn import functional as F
from AR.modules.patched_mha_with_cache_onnx import multi_head_attention_forward_patched
@@ -47,9 +43,7 @@ class MultiheadAttention(Module):
self.dropout = dropout
self.batch_first = batch_first
self.head_dim = embed_dim // num_heads
assert (
self.head_dim * num_heads == self.embed_dim
), "embed_dim must be divisible by num_heads"
assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
if add_bias_kv:
self.bias_k = Parameter(torch.empty((1, 1, embed_dim), **factory_kwargs))
@@ -60,18 +54,30 @@ class MultiheadAttention(Module):
if linear1_cls == Linear:
if not self._qkv_same_embed_dim:
self.q_proj_weight = Parameter(
torch.empty((embed_dim, embed_dim), **factory_kwargs)
torch.empty(
(embed_dim, embed_dim),
**factory_kwargs,
)
)
self.k_proj_weight = Parameter(
torch.empty((embed_dim, self.kdim), **factory_kwargs)
torch.empty(
(embed_dim, self.kdim),
**factory_kwargs,
)
)
self.v_proj_weight = Parameter(
torch.empty((embed_dim, self.vdim), **factory_kwargs)
torch.empty(
(embed_dim, self.vdim),
**factory_kwargs,
)
)
self.register_parameter("in_proj_weight", None)
else:
self.in_proj_weight = Parameter(
torch.empty((3 * embed_dim, embed_dim), **factory_kwargs)
torch.empty(
(3 * embed_dim, embed_dim),
**factory_kwargs,
)
)
self.register_parameter("q_proj_weight", None)
self.register_parameter("k_proj_weight", None)
@@ -79,13 +85,11 @@ class MultiheadAttention(Module):
if bias:
self.in_proj_bias = Parameter(
torch.empty(3 * embed_dim, **factory_kwargs)
torch.empty(3 * embed_dim, **factory_kwargs),
)
else:
self.register_parameter("in_proj_bias", None)
self.out_proj = NonDynamicallyQuantizableLinear(
embed_dim, embed_dim, bias=bias, **factory_kwargs
)
self.out_proj = NonDynamicallyQuantizableLinear(embed_dim, embed_dim, bias=bias, **factory_kwargs)
self._reset_parameters()
else:
@@ -93,7 +97,10 @@ class MultiheadAttention(Module):
raise NotImplementedError
else:
self.in_proj_linear = linear1_cls(
embed_dim, 3 * embed_dim, bias=bias, **factory_kwargs
embed_dim,
3 * embed_dim,
bias=bias,
**factory_kwargs,
)
self.in_proj_weight = self.in_proj_linear.weight
@@ -107,7 +114,10 @@ class MultiheadAttention(Module):
self.register_parameter("in_proj_bias", None)
self.out_proj = linear2_cls(
embed_dim, embed_dim, bias=bias, **factory_kwargs
embed_dim,
embed_dim,
bias=bias,
**factory_kwargs,
)
if self.bias_k is not None:

7
GPT_SoVITS/AR/modules/embedding.py
View File

@@ -60,14 +60,11 @@ class SinePositionalEmbedding(nn.Module):
return
pe = torch.zeros(x.size(1), self.embedding_dim)
if self.reverse:
position = torch.arange(
x.size(1) - 1, -1, -1.0, dtype=torch.float32
).unsqueeze(1)
position = torch.arange(x.size(1) - 1, -1, -1.0, dtype=torch.float32).unsqueeze(1)
else:
position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
div_term = torch.exp(
torch.arange(0, self.embedding_dim, 2, dtype=torch.float32)
* -(math.log(10000.0) / self.embedding_dim)
torch.arange(0, self.embedding_dim, 2, dtype=torch.float32) * -(math.log(10000.0) / self.embedding_dim)
)
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)

2
GPT_SoVITS/AR/modules/embedding_onnx.py
View File

@@ -50,7 +50,7 @@ class SinePositionalEmbedding(nn.Module):
self.div_term = torch.exp(torch.arange(0, self.embedding_dim, 2) * -(math.log(10000.0) / self.embedding_dim))
def extend_pe(self, x):
position = torch.cumsum(torch.ones_like(x[:,:,0]), dim=1).transpose(0, 1)
position = torch.cumsum(torch.ones_like(x[:, :, 0]), dim=1).transpose(0, 1)
scpe = (position * self.div_term).unsqueeze(0)
pe = torch.cat([torch.sin(scpe), torch.cos(scpe)]).permute(1, 2, 0)
pe = pe.contiguous().view(1, -1, self.embedding_dim)

16
GPT_SoVITS/AR/modules/lr_schedulers.py
View File

@@ -49,13 +49,9 @@ class WarmupCosineLRSchedule(torch.optim.lr_scheduler._LRScheduler):
lr = self.end_lr
else:
decay_ratio = (self._current_step - self.warmup_steps) / (
self.total_steps - self.warmup_steps
)
decay_ratio = (self._current_step - self.warmup_steps) / (self.total_steps - self.warmup_steps)
if decay_ratio < 0.0 or decay_ratio > 1.0:
raise RuntimeError(
"Decay ratio must be in [0.0, 1.0]. Fix LR scheduler settings."
)
raise RuntimeError("Decay ratio must be in [0.0, 1.0]. Fix LR scheduler settings.")
coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))
lr = self.end_lr + coeff * (self.peak_lr - self.end_lr)
@@ -70,7 +66,13 @@ if __name__ == "__main__":
m = nn.Linear(10, 10)
opt = Adam(m.parameters(), lr=1e-4)
s = WarmupCosineLRSchedule(
opt, 1e-6, 2e-4, 1e-6, warmup_steps=2000, total_steps=20000, current_step=0
opt,
1e-6,
2e-4,
1e-6,
warmup_steps=2000,
total_steps=20000,
current_step=0,
)
lrs = []
for i in range(25000):

211
GPT_SoVITS/AR/modules/optim.py
View File

@@ -16,8 +16,7 @@
import contextlib
import logging
from collections import defaultdict
from typing import List
from typing import Tuple
from typing import List, Tuple
import torch
from torch import Tensor
@@ -71,12 +70,8 @@ class BatchedOptimizer(Optimizer):
group_params_names: name for each parameter in group,
which is List[str].
"""
batches = defaultdict(
list
) # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter
batches_names = defaultdict(
list
) # `batches` maps from tuple (dtype_as_str,*shape) to list of str
batches = defaultdict(list) # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter
batches_names = defaultdict(list) # `batches` maps from tuple (dtype_as_str,*shape) to list of str
assert len(param_group) == len(group_params_names)
for p, named_p in zip(param_group, group_params_names):
@@ -85,11 +80,8 @@ class BatchedOptimizer(Optimizer):
batches_names[key].append(named_p)
batches_names_keys = list(batches_names.keys())
sorted_idx = sorted(
range(len(batches_names)), key=lambda i: batches_names_keys[i])
batches_names = [
batches_names[batches_names_keys[idx]] for idx in sorted_idx
]
sorted_idx = sorted(range(len(batches_names)), key=lambda i: batches_names_keys[i])
batches_names = [batches_names[batches_names_keys[idx]] for idx in sorted_idx]
batches = [batches[batches_names_keys[idx]] for idx in sorted_idx]
stacked_params_dict = dict()
@@ -106,16 +98,14 @@ class BatchedOptimizer(Optimizer):
# group. class Optimizer will take care of saving/loading state.
state = self.state[p]
p_stacked = torch.stack(batch)
grad = torch.stack([
torch.zeros_like(p) if p.grad is None else p.grad for p in batch
])
grad = torch.stack([torch.zeros_like(p) if p.grad is None else p.grad for p in batch])
p_stacked.grad = grad
stacked_params_dict[key] = p_stacked
tuples.append((p_stacked, state, batch_names))
yield tuples # <-- calling code will do the actual optimization here!
for ((stacked_params, _state, _names), batch) in zip(tuples, batches):
for (stacked_params, _state, _names), batch in zip(tuples, batches):
for i, p in enumerate(batch): # batch is list of Parameter
p.copy_(stacked_params[i])
@@ -164,25 +154,24 @@ class ScaledAdam(BatchedOptimizer):
"""
def __init__(
self,
params,
lr=3e-02,
clipping_scale=None,
betas=(0.9, 0.98),
scalar_lr_scale=0.1,
eps=1.0e-08,
param_min_rms=1.0e-05,
param_max_rms=3.0,
scalar_max=10.0,
size_update_period=4,
clipping_update_period=100,
parameters_names=None,
show_dominant_parameters=True, ):
self,
params,
lr=3e-02,
clipping_scale=None,
betas=(0.9, 0.98),
scalar_lr_scale=0.1,
eps=1.0e-08,
param_min_rms=1.0e-05,
param_max_rms=3.0,
scalar_max=10.0,
size_update_period=4,
clipping_update_period=100,
parameters_names=None,
show_dominant_parameters=True,
):
assert parameters_names is not None, (
"Please prepare parameters_names,"
"which is a List[List[str]]. Each List[str] is for a group"
"and each str is for a parameter")
"Please prepare parameters_names,which is a List[List[str]]. Each List[str] is for a groupand each str is for a parameter"
)
defaults = dict(
lr=lr,
clipping_scale=clipping_scale,
@@ -193,7 +182,8 @@ class ScaledAdam(BatchedOptimizer):
param_max_rms=param_max_rms,
scalar_max=scalar_max,
size_update_period=size_update_period,
clipping_update_period=clipping_update_period, )
clipping_update_period=clipping_update_period,
)
super(ScaledAdam, self).__init__(params, defaults)
assert len(self.param_groups) == len(parameters_names)
@@ -218,18 +208,13 @@ class ScaledAdam(BatchedOptimizer):
batch = True
for group, group_params_names in zip(self.param_groups,
self.parameters_names):
with self.batched_params(group["params"],
group_params_names) as batches:
for group, group_params_names in zip(self.param_groups, self.parameters_names):
with self.batched_params(group["params"], group_params_names) as batches:
# batches is list of pairs (stacked_param, state). stacked_param is like
# a regular parameter, and will have a .grad, but the 1st dim corresponds to
# a stacking dim, it is not a real dim.
if (len(batches[0][1]) ==
0): # if len(first state) == 0: not yet initialized
if len(batches[0][1]) == 0: # if len(first state) == 0: not yet initialized
clipping_scale = 1
else:
clipping_scale = self._get_clipping_scale(group, batches)
@@ -239,9 +224,7 @@ class ScaledAdam(BatchedOptimizer):
# grad is not going to be None, we handled that when creating the batches.
grad = p.grad
if grad.is_sparse:
raise RuntimeError(
"ScaledAdam optimizer does not support sparse gradients"
)
raise RuntimeError("ScaledAdam optimizer does not support sparse gradients")
# State initialization
if len(state) == 0:
self._init_state(group, p, state)
@@ -274,8 +257,7 @@ class ScaledAdam(BatchedOptimizer):
# parameter-change "delta", which combines all forms of
# update. this is equivalent to how it's done in Adam,
# except for the first few steps.
state["delta"] = torch.zeros_like(
p, memory_format=torch.preserve_format)
state["delta"] = torch.zeros_like(p, memory_format=torch.preserve_format)
batch_size = p.shape[0]
numel = p.numel() // batch_size
@@ -285,22 +267,16 @@ class ScaledAdam(BatchedOptimizer):
# "param_rms" just periodically records the scalar root-mean-square value of
# the parameter tensor.
# it has a shape like (batch_size, 1, 1, 1, 1)
param_rms = (
(p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt())
param_rms = (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt()
state["param_rms"] = param_rms
state["scale_exp_avg_sq"] = torch.zeros_like(param_rms)
state["scale_grads"] = torch.zeros(size_update_period,
*param_rms.shape, **kwargs)
state["scale_grads"] = torch.zeros(size_update_period, *param_rms.shape, **kwargs)
# exp_avg_sq is the weighted sum of scaled gradients. as in Adam.
state["exp_avg_sq"] = torch.zeros_like(
p, memory_format=torch.preserve_format)
state["exp_avg_sq"] = torch.zeros_like(p, memory_format=torch.preserve_format)
def _get_clipping_scale(self,
group: dict,
tuples: List[Tuple[Tensor, dict, List[str]]]
) -> float:
def _get_clipping_scale(self, group: dict, tuples: List[Tuple[Tensor, dict, List[str]]]) -> float:
"""
Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients
by this amount before applying the rest of the update.
@@ -325,20 +301,18 @@ class ScaledAdam(BatchedOptimizer):
clipping_update_period = group["clipping_update_period"]
tot_sumsq = torch.tensor(0.0, device=first_p.device)
for (p, state, param_names) in tuples:
for p, state, param_names in tuples:
grad = p.grad
if grad.is_sparse:
raise RuntimeError(
"ScaledAdam optimizer does not support sparse gradients")
raise RuntimeError("ScaledAdam optimizer does not support sparse gradients")
if p.numel() == p.shape[0]: # a batch of scalars
tot_sumsq += (grad**2).sum() # sum() to change shape [1] to []
else:
tot_sumsq += ((grad * state["param_rms"])**2).sum()
tot_sumsq += ((grad * state["param_rms"]) ** 2).sum()
tot_norm = tot_sumsq.sqrt()
if "model_norms" not in first_state:
first_state["model_norms"] = torch.zeros(
clipping_update_period, device=p.device)
first_state["model_norms"] = torch.zeros(clipping_update_period, device=p.device)
first_state["model_norms"][step % clipping_update_period] = tot_norm
if step % clipping_update_period == 0:
@@ -350,20 +324,20 @@ class ScaledAdam(BatchedOptimizer):
for n in range(0, 5):
index = min(
clipping_update_period - 1,
(clipping_update_period // 4) * n, )
(clipping_update_period // 4) * n,
)
quartiles.append(sorted_norms[index].item())
median = quartiles[2]
threshold = clipping_scale * median
first_state["model_norm_threshold"] = threshold
percent_clipped = (first_state["num_clipped"] * 100.0 /
clipping_update_period
if "num_clipped" in first_state else 0.0)
percent_clipped = (
first_state["num_clipped"] * 100.0 / clipping_update_period if "num_clipped" in first_state else 0.0
)
first_state["num_clipped"] = 0
quartiles = " ".join(["%.3e" % x for x in quartiles])
logging.info(
f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, "
f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}"
f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}"
)
if step < clipping_update_period:
@@ -373,25 +347,20 @@ class ScaledAdam(BatchedOptimizer):
model_norm_threshold = first_state["model_norm_threshold"]
except KeyError:
logging.info(
"Warning: model_norm_threshold not in state: possibly "
"you changed config when restarting, adding clipping_scale option?"
"Warning: model_norm_threshold not in state: possibly you changed config when restarting, adding clipping_scale option?"
)
return 1.0
ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item())
if ans < 1.0:
first_state["num_clipped"] += 1
if ans < 0.1:
logging.warn(
f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}"
)
logging.warn(f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}")
if self.show_dominant_parameters:
assert p.shape[0] == len(param_names)
self._show_gradient_dominating_parameter(tuples, tot_sumsq)
return ans
def _show_gradient_dominating_parameter(
self, tuples: List[Tuple[Tensor, dict, List[str]]],
tot_sumsq: Tensor):
def _show_gradient_dominating_parameter(self, tuples: List[Tuple[Tensor, dict, List[str]]], tot_sumsq: Tensor):
"""
Show information of parameter wihch dominanting tot_sumsq.
@@ -406,7 +375,7 @@ class ScaledAdam(BatchedOptimizer):
from tuples, we still pass it to save some time.
"""
all_sumsq_orig = {}
for (p, state, batch_param_names) in tuples:
for p, state, batch_param_names in tuples:
# p is a stacked batch parameters.
batch_grad = p.grad
if p.numel() == p.shape[0]: # a batch of scalars
@@ -415,41 +384,46 @@ class ScaledAdam(BatchedOptimizer):
batch_rms_orig = torch.ones(p.shape[0])
else:
batch_rms_orig = state["param_rms"]
batch_sumsq_orig = ((batch_grad * batch_rms_orig)**2).sum(
dim=list(range(1, batch_grad.ndim)))
for name, sumsq_orig, rms, grad in zip(batch_param_names,
batch_sumsq_orig,
batch_rms_orig, batch_grad):
batch_sumsq_orig = ((batch_grad * batch_rms_orig) ** 2).sum(dim=list(range(1, batch_grad.ndim)))
for name, sumsq_orig, rms, grad in zip(
batch_param_names,
batch_sumsq_orig,
batch_rms_orig,
batch_grad,
):
proportion_orig = sumsq_orig / tot_sumsq
all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad)
assert torch.isclose(
sum([value[0] for value in all_sumsq_orig.values()]).cpu(),
torch.tensor(1.0), )
torch.tensor(1.0),
)
sorted_by_proportion = {
k: v
for k, v in sorted(
all_sumsq_orig.items(),
key=lambda item: item[1][0],
reverse=True, )
reverse=True,
)
}
dominant_param_name = next(iter(sorted_by_proportion))
(dominant_proportion, dominant_sumsq, dominant_rms,
dominant_grad, ) = sorted_by_proportion[dominant_param_name]
logging.info(f"Parameter Dominanting tot_sumsq {dominant_param_name}"
f" with proportion {dominant_proportion:.2f},"
f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)"
f"={dominant_sumsq:.3e},"
f" grad_sumsq = {(dominant_grad**2).sum():.3e},"
f" orig_rms_sq={(dominant_rms**2).item():.3e}")
(
dominant_proportion,
dominant_sumsq,
dominant_rms,
dominant_grad,
) = sorted_by_proportion[dominant_param_name]
logging.info(
f"Parameter Dominanting tot_sumsq {dominant_param_name}"
f" with proportion {dominant_proportion:.2f},"
f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)"
f"={dominant_sumsq:.3e},"
f" grad_sumsq = {(dominant_grad**2).sum():.3e},"
f" orig_rms_sq={(dominant_rms**2).item():.3e}"
)
def _step_one_batch(self,
group: dict,
p: Tensor,
state: dict,
clipping_scale: float):
def _step_one_batch(self, group: dict, p: Tensor, state: dict, clipping_scale: float):
"""
Do the step for one parameter, which is actually going to be a batch of
`real` parameters, with dim 0 as the batch dim.
@@ -475,13 +449,10 @@ class ScaledAdam(BatchedOptimizer):
if numel > 1:
# Update the size/scale of p, and set param_rms
scale_grads = state["scale_grads"]
scale_grads[step % size_update_period] = (p * grad).sum(
dim=list(range(1, p.ndim)), keepdim=True)
scale_grads[step % size_update_period] = (p * grad).sum(dim=list(range(1, p.ndim)), keepdim=True)
if step % size_update_period == size_update_period - 1:
param_rms = state["param_rms"] # shape: (batch_size, 1, 1, ..)
param_rms.copy_((p**2)
.mean(dim=list(range(1, p.ndim)), keepdim=True)
.sqrt())
param_rms.copy_((p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt())
if step > 0:
# self._size_update() learns the overall scale on the
# parameter, by shrinking or expanding it.
@@ -496,11 +467,13 @@ class ScaledAdam(BatchedOptimizer):
state["step"] = step + 1
def _size_update(self,
group: dict,
scale_grads: Tensor,
p: Tensor,
state: dict) -> None:
def _size_update(
self,
group: dict,
scale_grads: Tensor,
p: Tensor,
state: dict,
) -> None:
"""
Called only where p.numel() > 1, this updates the scale of the parameter.
If we imagine: p = underlying_param * scale.exp(), and we are doing
@@ -529,11 +502,11 @@ class ScaledAdam(BatchedOptimizer):
# faster decay at this level.
beta2_corr = beta2**size_update_period
scale_exp_avg_sq = state[
"scale_exp_avg_sq"] # shape: (batch_size, 1, 1, ..)
scale_exp_avg_sq = state["scale_exp_avg_sq"] # shape: (batch_size, 1, 1, ..)
scale_exp_avg_sq.mul_(beta2_corr).add_(
(scale_grads**2).mean(dim=0), # mean over dim `size_update_period`
alpha=1 - beta2_corr, ) # shape is (batch_size, 1, 1, ...)
alpha=1 - beta2_corr,
) # shape is (batch_size, 1, 1, ...)
# The 1st time we reach here is when size_step == 1.
size_step = (step + 1) // size_update_period
@@ -543,8 +516,7 @@ class ScaledAdam(BatchedOptimizer):
denom = scale_exp_avg_sq.sqrt() + eps
scale_step = (-size_lr * (bias_correction2**0.5) *
scale_grads.sum(dim=0) / denom)
scale_step = -size_lr * (bias_correction2**0.5) * scale_grads.sum(dim=0) / denom
is_too_small = param_rms < param_min_rms
is_too_large = param_rms > param_max_rms
@@ -580,9 +552,8 @@ class ScaledAdam(BatchedOptimizer):
exp_avg_sq = state["exp_avg_sq"]
exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2))
this_step = state["step"] - (state["zero_step"]
if "zero_step" in state else 0)
bias_correction2 = 1 - beta2**(this_step + 1)
this_step = state["step"] - (state["zero_step"] if "zero_step" in state else 0)
bias_correction2 = 1 - beta2 ** (this_step + 1)
if bias_correction2 < 0.99:
# note: not in-place.
exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2)
@@ -613,7 +584,7 @@ class ScaledAdam(BatchedOptimizer):
# bias_correction2 is like in Adam. Don't bother with bias_correction1;
# slower update at the start will help stability anyway.
bias_correction2 = 1 - beta2**(state["step"] + 1)
bias_correction2 = 1 - beta2 ** (state["step"] + 1)
denom = (exp_avg_sq / bias_correction2).sqrt() + eps
delta = state["delta"]

121
GPT_SoVITS/AR/modules/patched_mha_with_cache.py
View File

@@ -5,7 +5,6 @@ from torch.nn.functional import (
_none_or_dtype,
_in_projection_packed,
)
from torch.nn import functional as F
import torch
# Tensor = torch.Tensor
# from typing import Callable, List, Optional, Tuple, Union
@@ -25,18 +24,18 @@ def multi_head_attention_forward_patched(
dropout_p: float,
out_proj_weight,
out_proj_bias,
training = True,
key_padding_mask = None,
need_weights = True,
attn_mask = None,
use_separate_proj_weight = False,
q_proj_weight = None,
k_proj_weight = None,
v_proj_weight = None,
static_k = None,
static_v = None,
average_attn_weights = True,
is_causal = False,
training=True,
key_padding_mask=None,
need_weights=True,
attn_mask=None,
use_separate_proj_weight=False,
q_proj_weight=None,
k_proj_weight=None,
v_proj_weight=None,
static_k=None,
static_v=None,
average_attn_weights=True,
is_causal=False,
cache=None,
):
r"""
@@ -156,9 +155,7 @@ def multi_head_attention_forward_patched(
cache=cache,
)
is_batched = _mha_shape_check(
query, key, value, key_padding_mask, attn_mask, num_heads
)
is_batched = _mha_shape_check(query, key, value, key_padding_mask, attn_mask, num_heads)
# For unbatched input, we unsqueeze at the expected batch-dim to pretend that the input
# is batched, run the computation and before returning squeeze the
@@ -211,45 +208,33 @@ def multi_head_attention_forward_patched(
# longer causal.
is_causal = False
assert (
embed_dim == embed_dim_to_check
), f"was expecting embedding dimension of {embed_dim_to_check}, but got {embed_dim}"
assert embed_dim == embed_dim_to_check, (
f"was expecting embedding dimension of {embed_dim_to_check}, but got {embed_dim}"
)
if isinstance(embed_dim, torch.Tensor):
# embed_dim can be a tensor when JIT tracing
head_dim = embed_dim.div(num_heads, rounding_mode="trunc")
else:
head_dim = embed_dim // num_heads
assert (
head_dim * num_heads == embed_dim
), f"embed_dim {embed_dim} not divisible by num_heads {num_heads}"
assert head_dim * num_heads == embed_dim, f"embed_dim {embed_dim} not divisible by num_heads {num_heads}"
if use_separate_proj_weight:
# allow MHA to have different embedding dimensions when separate projection weights are used
assert (
key.shape[:2] == value.shape[:2]
), f"key's sequence and batch dims {key.shape[:2]} do not match value's {value.shape[:2]}"
assert key.shape[:2] == value.shape[:2], (
f"key's sequence and batch dims {key.shape[:2]} do not match value's {value.shape[:2]}"
)
else:
assert (
key.shape == value.shape
), f"key shape {key.shape} does not match value shape {value.shape}"
assert key.shape == value.shape, f"key shape {key.shape} does not match value shape {value.shape}"
#
# compute in-projection
#
if not use_separate_proj_weight:
assert (
in_proj_weight is not None
), "use_separate_proj_weight is False but in_proj_weight is None"
assert in_proj_weight is not None, "use_separate_proj_weight is False but in_proj_weight is None"
q, k, v = _in_projection_packed(query, key, value, in_proj_weight, in_proj_bias)
else:
assert (
q_proj_weight is not None
), "use_separate_proj_weight is True but q_proj_weight is None"
assert (
k_proj_weight is not None
), "use_separate_proj_weight is True but k_proj_weight is None"
assert (
v_proj_weight is not None
), "use_separate_proj_weight is True but v_proj_weight is None"
assert q_proj_weight is not None, "use_separate_proj_weight is True but q_proj_weight is None"
assert k_proj_weight is not None, "use_separate_proj_weight is True but k_proj_weight is None"
assert v_proj_weight is not None, "use_separate_proj_weight is True but v_proj_weight is None"
if in_proj_bias is None:
b_q = b_k = b_v = None
else:
@@ -312,9 +297,7 @@ def multi_head_attention_forward_patched(
f"The shape of the 3D attn_mask is {attn_mask.shape}, but should be {correct_3d_size}."
)
else:
raise RuntimeError(
f"attn_mask's dimension {attn_mask.dim()} is not supported"
)
raise RuntimeError(f"attn_mask's dimension {attn_mask.dim()} is not supported")
# add bias along batch dimension (currently second)
if bias_k is not None and bias_v is not None:
@@ -338,34 +321,26 @@ def multi_head_attention_forward_patched(
k = k.view(k.shape[0], bsz * num_heads, head_dim).transpose(0, 1)
else:
# TODO finish disentangling control flow so we don't do in-projections when statics are passed
assert (
static_k.size(0) == bsz * num_heads
), f"expecting static_k.size(0) of {bsz * num_heads}, but got {static_k.size(0)}"
assert (
static_k.size(2) == head_dim
), f"expecting static_k.size(2) of {head_dim}, but got {static_k.size(2)}"
assert static_k.size(0) == bsz * num_heads, (
f"expecting static_k.size(0) of {bsz * num_heads}, but got {static_k.size(0)}"
)
assert static_k.size(2) == head_dim, f"expecting static_k.size(2) of {head_dim}, but got {static_k.size(2)}"
k = static_k
if static_v is None:
v = v.view(v.shape[0], bsz * num_heads, head_dim).transpose(0, 1)
else:
# TODO finish disentangling control flow so we don't do in-projections when statics are passed
assert (
static_v.size(0) == bsz * num_heads
), f"expecting static_v.size(0) of {bsz * num_heads}, but got {static_v.size(0)}"
assert (
static_v.size(2) == head_dim
), f"expecting static_v.size(2) of {head_dim}, but got {static_v.size(2)}"
assert static_v.size(0) == bsz * num_heads, (
f"expecting static_v.size(0) of {bsz * num_heads}, but got {static_v.size(0)}"
)
assert static_v.size(2) == head_dim, f"expecting static_v.size(2) of {head_dim}, but got {static_v.size(2)}"
v = static_v
# add zero attention along batch dimension (now first)
if add_zero_attn:
zero_attn_shape = (bsz * num_heads, 1, head_dim)
k = torch.cat(
[k, torch.zeros(zero_attn_shape, dtype=k.dtype, device=k.device)], dim=1
)
v = torch.cat(
[v, torch.zeros(zero_attn_shape, dtype=v.dtype, device=v.device)], dim=1
)
k = torch.cat([k, torch.zeros(zero_attn_shape, dtype=k.dtype, device=k.device)], dim=1)
v = torch.cat([v, torch.zeros(zero_attn_shape, dtype=v.dtype, device=v.device)], dim=1)
if attn_mask is not None:
attn_mask = pad(attn_mask, (0, 1))
if key_padding_mask is not None:
@@ -381,9 +356,7 @@ def multi_head_attention_forward_patched(
src_len,
), f"expecting key_padding_mask shape of {(bsz, src_len)}, but got {key_padding_mask.shape}"
key_padding_mask = (
key_padding_mask.view(bsz, 1, 1, src_len)
.expand(-1, num_heads, -1, -1)
.reshape(bsz * num_heads, 1, src_len)
key_padding_mask.view(bsz, 1, 1, src_len).expand(-1, num_heads, -1, -1).reshape(bsz * num_heads, 1, src_len)
)
if attn_mask is None:
attn_mask = key_padding_mask
@@ -402,14 +375,10 @@ def multi_head_attention_forward_patched(
B, Nt, E = q.shape
q_scaled = q / math.sqrt(E)
assert not (
is_causal and attn_mask is None
), "FIXME: is_causal not implemented for need_weights"
assert not (is_causal and attn_mask is None), "FIXME: is_causal not implemented for need_weights"
if attn_mask is not None:
attn_output_weights = torch.baddbmm(
attn_mask, q_scaled, k.transpose(-2, -1)
)
attn_output_weights = torch.baddbmm(attn_mask, q_scaled, k.transpose(-2, -1))
else:
attn_output_weights = torch.bmm(q_scaled, k.transpose(-2, -1))
attn_output_weights = softmax(attn_output_weights, dim=-1)
@@ -418,9 +387,7 @@ def multi_head_attention_forward_patched(
attn_output = torch.bmm(attn_output_weights, v)
attn_output = (
attn_output.transpose(0, 1).contiguous().view(tgt_len * bsz, embed_dim)
)
attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len * bsz, embed_dim)
attn_output = linear(attn_output, out_proj_weight, out_proj_bias)
attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
@@ -449,13 +416,9 @@ def multi_head_attention_forward_patched(
v = v.view(bsz, num_heads, src_len, head_dim)
# with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=True):
attn_output = scaled_dot_product_attention(
q, k, v, attn_mask, dropout_p, is_causal
)
attn_output = scaled_dot_product_attention(q, k, v, attn_mask, dropout_p, is_causal)
attn_output = (
attn_output.permute(2, 0, 1, 3).contiguous().view(bsz * tgt_len, embed_dim)
)
attn_output = attn_output.permute(2, 0, 1, 3).contiguous().view(bsz * tgt_len, embed_dim)
attn_output = linear(attn_output, out_proj_weight, out_proj_bias)
attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))

13
GPT_SoVITS/AR/modules/patched_mha_with_cache_onnx.py
View File

@@ -1,11 +1,9 @@
from torch.nn.functional import *
from torch.nn.functional import (
_mha_shape_check,
_canonical_mask,
_none_or_dtype,
_in_projection_packed,
)
def multi_head_attention_forward_patched(
query,
key,
@@ -34,7 +32,6 @@ def multi_head_attention_forward_patched(
is_causal: bool = False,
cache=None,
) -> Tuple[Tensor, Optional[Tensor]]:
# set up shape vars
_, _, embed_dim = query.shape
attn_mask = _canonical_mask(
@@ -80,12 +77,8 @@ def multi_head_attention_forward_patched(
q = q.view(num_heads, -1, head_dim).unsqueeze(0)
k = k.view(num_heads, -1, head_dim).unsqueeze(0)
v = v.view(num_heads, -1, head_dim).unsqueeze(0)
attn_output = scaled_dot_product_attention(
q, k, v, attn_mask, dropout_p, is_causal
)
attn_output = (
attn_output.permute(2, 0, 1, 3).contiguous().view(-1, embed_dim)
)
attn_output = scaled_dot_product_attention(q, k, v, attn_mask, dropout_p, is_causal)
attn_output = attn_output.permute(2, 0, 1, 3).contiguous().view(-1, embed_dim)
attn_output = linear(attn_output, out_proj_weight, out_proj_bias)
attn_output = attn_output.view(-1, 1, attn_output.size(1))

33
GPT_SoVITS/AR/modules/scaling.py
View File

@@ -13,12 +13,9 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
import math
import random
from typing import Optional
from typing import Tuple
from typing import Union
import torch
import torch.nn as nn
@@ -61,9 +58,7 @@ class DoubleSwishFunction(torch.autograd.Function):
# floors), should be expectation-preserving.
floor = -0.043637
ceil = 1.2
d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like(
deriv
)
d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like(deriv)
if __name__ == "__main__":
# for self-testing only.
assert d_scaled.min() >= 0.0
@@ -153,13 +148,9 @@ def _compute_scale_factor(
else:
# below_threshold is 0 if x_abs_mean > min_abs, can be at most max_factor if
# x_abs)_mean , min_abs.
below_threshold = ((min_abs - x_abs_mean) * (gain_factor / min_abs)).clamp(
min=0, max=max_factor
)
below_threshold = ((min_abs - x_abs_mean) * (gain_factor / min_abs)).clamp(min=0, max=max_factor)
above_threshold = ((x_abs_mean - max_abs) * (gain_factor / max_abs)).clamp(
min=0, max=max_factor
)
above_threshold = ((x_abs_mean - max_abs) * (gain_factor / max_abs)).clamp(min=0, max=max_factor)
return below_threshold - above_threshold
@@ -181,18 +172,16 @@ def _compute_sign_factor(
else:
# 0 if proportion_positive >= min_positive, else can be
# as large as max_factor.
factor1 = (
(min_positive - proportion_positive) * (gain_factor / min_positive)
).clamp_(min=0, max=max_factor)
factor1 = ((min_positive - proportion_positive) * (gain_factor / min_positive)).clamp_(min=0, max=max_factor)
if max_positive == 1.0:
factor2 = 0.0
else:
# 0 if self.proportion_positive <= max_positive, else can be
# as large as -max_factor.
factor2 = (
(proportion_positive - max_positive) * (gain_factor / (1.0 - max_positive))
).clamp_(min=0, max=max_factor)
factor2 = ((proportion_positive - max_positive) * (gain_factor / (1.0 - max_positive))).clamp_(
min=0, max=max_factor
)
sign_factor = factor1 - factor2
# require min_positive != 0 or max_positive != 1:
assert not isinstance(sign_factor, float)
@@ -320,15 +309,11 @@ class ActivationBalancer(torch.nn.Module):
return _no_op(x)
def BalancedDoubleSwish(
d_model, channel_dim=-1, max_abs=10.0, min_prob=0.25
) -> nn.Sequential:
def BalancedDoubleSwish(d_model, channel_dim=-1, max_abs=10.0, min_prob=0.25) -> nn.Sequential:
"""
ActivationBalancer -> DoubleSwish
"""
balancer = ActivationBalancer(
d_model, channel_dim=channel_dim, max_abs=max_abs, min_prob=min_prob
)
balancer = ActivationBalancer(d_model, channel_dim=channel_dim, max_abs=max_abs, min_prob=min_prob)
return nn.Sequential(
balancer,
DoubleSwish(),

34
GPT_SoVITS/AR/modules/transformer.py
View File

@@ -42,12 +42,8 @@ class LayerNorm(nn.Module):
self.eps = eps
self.elementwise_affine = elementwise_affine
if self.elementwise_affine:
self.weight = nn.Parameter(
torch.empty(self.normalized_shape, **factory_kwargs)
)
self.bias = nn.Parameter(
torch.empty(self.normalized_shape, **factory_kwargs)
)
self.weight = nn.Parameter(torch.empty(self.normalized_shape, **factory_kwargs))
self.bias = nn.Parameter(torch.empty(self.normalized_shape, **factory_kwargs))
else:
self.register_parameter("weight", None)
self.register_parameter("bias", None)
@@ -74,15 +70,10 @@ class LayerNorm(nn.Module):
)
assert embedding is None
return F.layer_norm(
input, self.normalized_shape, self.weight, self.bias, self.eps
)
return F.layer_norm(input, self.normalized_shape, self.weight, self.bias, self.eps)
def extra_repr(self) -> str:
return (
"{normalized_shape}, eps={eps}, "
"elementwise_affine={elementwise_affine}".format(**self.__dict__)
)
return "{normalized_shape}, eps={eps}, elementwise_affine={elementwise_affine}".format(**self.__dict__)
class IdentityNorm(nn.Module):
@@ -121,6 +112,7 @@ class TransformerEncoder(nn.Module):
>>> src = torch.rand(10, 32, 512)
>>> out = transformer_encoder(src)
"""
__constants__ = ["norm"]
def __init__(self, encoder_layer, num_layers, norm=None):
@@ -218,13 +210,9 @@ class TransformerEncoderLayer(nn.Module):
)
# Implementation of Feedforward model
self.linear1 = linear1_feedforward_cls(
d_model, dim_feedforward, **factory_kwargs
)
self.linear1 = linear1_feedforward_cls(d_model, dim_feedforward, **factory_kwargs)
self.dropout = nn.Dropout(dropout)
self.linear2 = linear2_feedforward_cls(
dim_feedforward, d_model, **factory_kwargs
)
self.linear2 = linear2_feedforward_cls(dim_feedforward, d_model, **factory_kwargs)
self.norm_first = norm_first
self.dropout1 = nn.Dropout(dropout)
@@ -291,12 +279,8 @@ class TransformerEncoderLayer(nn.Module):
if src_key_padding_mask is not None:
_skpm_dtype = src_key_padding_mask.dtype
if _skpm_dtype != torch.bool and not torch.is_floating_point(
src_key_padding_mask
):
raise AssertionError(
"only bool and floating types of key_padding_mask are supported"
)
if _skpm_dtype != torch.bool and not torch.is_floating_point(src_key_padding_mask):
raise AssertionError("only bool and floating types of key_padding_mask are supported")
if self.norm_first:
x = x + self._sa_block(

27
GPT_SoVITS/AR/modules/transformer_onnx.py
View File

@@ -42,12 +42,8 @@ class LayerNorm(nn.Module):
self.eps = eps
self.elementwise_affine = elementwise_affine
if self.elementwise_affine:
self.weight = nn.Parameter(
torch.empty(self.normalized_shape, **factory_kwargs)
)
self.bias = nn.Parameter(
torch.empty(self.normalized_shape, **factory_kwargs)
)
self.weight = nn.Parameter(torch.empty(self.normalized_shape, **factory_kwargs))
self.bias = nn.Parameter(torch.empty(self.normalized_shape, **factory_kwargs))
else:
self.register_parameter("weight", None)
self.register_parameter("bias", None)
@@ -74,15 +70,10 @@ class LayerNorm(nn.Module):
)
assert embedding is None
return F.layer_norm(
input, self.normalized_shape, self.weight, self.bias, self.eps
)
return F.layer_norm(input, self.normalized_shape, self.weight, self.bias, self.eps)
def extra_repr(self) -> str:
return (
"{normalized_shape}, eps={eps}, "
"elementwise_affine={elementwise_affine}".format(**self.__dict__)
)
return "{normalized_shape}, eps={eps}, elementwise_affine={elementwise_affine}".format(**self.__dict__)
class IdentityNorm(nn.Module):
@@ -121,6 +112,7 @@ class TransformerEncoder(nn.Module):
>>> src = torch.rand(10, 32, 512)
>>> out = transformer_encoder(src)
"""
__constants__ = ["norm"]
def __init__(self, encoder_layer, num_layers, norm=None):
@@ -154,6 +146,7 @@ class TransformerEncoder(nn.Module):
class TransformerEncoderLayer(nn.Module):
__constants__ = ["batch_first", "norm_first"]
def __init__(
self,
d_model: int,
@@ -184,13 +177,9 @@ class TransformerEncoderLayer(nn.Module):
linear2_cls=linear2_self_attention_cls,
**factory_kwargs,
)
self.linear1 = linear1_feedforward_cls(
d_model, dim_feedforward, **factory_kwargs
)
self.linear1 = linear1_feedforward_cls(d_model, dim_feedforward, **factory_kwargs)
self.dropout = nn.Dropout(dropout)
self.linear2 = linear2_feedforward_cls(
dim_feedforward, d_model, **factory_kwargs
)
self.linear2 = linear2_feedforward_cls(dim_feedforward, d_model, **factory_kwargs)
self.norm_first = norm_first
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)