1326 lines
61 KiB
Python
1326 lines
61 KiB
Python
# coding=utf-8
|
|
# Copyright 2023 The Pop2Piano Authors and The HuggingFace Inc. team. All rights reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# 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.
|
|
"""PyTorch Pop2Piano model."""
|
|
|
|
import copy
|
|
import math
|
|
from typing import Optional, Union
|
|
|
|
import torch
|
|
from torch import nn
|
|
from torch.nn import CrossEntropyLoss
|
|
|
|
from transformers.generation import GenerationConfig
|
|
|
|
from ...activations import ACT2FN
|
|
from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache
|
|
from ...generation import GenerationMixin
|
|
from ...modeling_attn_mask_utils import AttentionMaskConverter
|
|
from ...modeling_layers import GradientCheckpointingLayer
|
|
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPastAndCrossAttentions, Seq2SeqLMOutput
|
|
from ...modeling_utils import PreTrainedModel
|
|
from ...pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer
|
|
from ...utils import auto_docstring, is_torch_flex_attn_available, is_torch_fx_proxy, is_torchdynamo_compiling, logging
|
|
from .configuration_pop2piano import Pop2PianoConfig
|
|
|
|
|
|
if is_torch_flex_attn_available():
|
|
from torch.nn.attention.flex_attention import BlockMask
|
|
|
|
from ...integrations.flex_attention import make_flex_block_causal_mask
|
|
|
|
|
|
logger = logging.get_logger(__name__)
|
|
|
|
_load_pop2piano_layer_norm = True
|
|
|
|
try:
|
|
from apex.normalization import FusedRMSNorm
|
|
|
|
_load_pop2piano_layer_norm = False
|
|
|
|
logger.info("Discovered apex.normalization.FusedRMSNorm - will use it instead of Pop2PianoLayerNorm")
|
|
except ImportError:
|
|
# using the normal Pop2PianoLayerNorm
|
|
pass
|
|
except Exception:
|
|
logger.warning("Discovered apex but it failed to load, falling back to Pop2PianoLayerNorm")
|
|
pass
|
|
|
|
|
|
# Copied from transformers.models.t5.modeling_t5.T5LayerNorm with T5->Pop2Piano
|
|
class Pop2PianoLayerNorm(nn.Module):
|
|
def __init__(self, hidden_size, eps=1e-6):
|
|
"""
|
|
Construct a layernorm module in the Pop2Piano style. No bias and no subtraction of mean.
|
|
"""
|
|
super().__init__()
|
|
self.weight = nn.Parameter(torch.ones(hidden_size))
|
|
self.variance_epsilon = eps
|
|
|
|
def forward(self, hidden_states):
|
|
# Pop2Piano uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean
|
|
# Square Layer Normalization https://huggingface.co/papers/1910.07467 thus variance is calculated
|
|
# w/o mean and there is no bias. Additionally we want to make sure that the accumulation for
|
|
# half-precision inputs is done in fp32
|
|
|
|
variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
|
|
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
|
|
|
# convert into half-precision if necessary
|
|
if self.weight.dtype in [torch.float16, torch.bfloat16]:
|
|
hidden_states = hidden_states.to(self.weight.dtype)
|
|
|
|
return self.weight * hidden_states
|
|
|
|
|
|
if not _load_pop2piano_layer_norm:
|
|
Pop2PianoLayerNorm = FusedRMSNorm # noqa
|
|
|
|
|
|
# Copied from transformers.models.t5.modeling_t5.T5DenseActDense with T5->Pop2Piano,t5->pop2piano
|
|
class Pop2PianoDenseActDense(nn.Module):
|
|
def __init__(self, config: Pop2PianoConfig):
|
|
super().__init__()
|
|
self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)
|
|
self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
|
|
self.dropout = nn.Dropout(config.dropout_rate)
|
|
self.act = ACT2FN[config.dense_act_fn]
|
|
|
|
def forward(self, hidden_states):
|
|
hidden_states = self.wi(hidden_states)
|
|
hidden_states = self.act(hidden_states)
|
|
hidden_states = self.dropout(hidden_states)
|
|
if (
|
|
isinstance(self.wo.weight, torch.Tensor)
|
|
and hidden_states.dtype != self.wo.weight.dtype
|
|
and self.wo.weight.dtype != torch.int8
|
|
):
|
|
hidden_states = hidden_states.to(self.wo.weight.dtype)
|
|
hidden_states = self.wo(hidden_states)
|
|
return hidden_states
|
|
|
|
|
|
# Copied from transformers.models.t5.modeling_t5.T5DenseGatedActDense with T5->Pop2Piano
|
|
class Pop2PianoDenseGatedActDense(nn.Module):
|
|
def __init__(self, config: Pop2PianoConfig):
|
|
super().__init__()
|
|
self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)
|
|
self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)
|
|
self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
|
|
self.dropout = nn.Dropout(config.dropout_rate)
|
|
self.act = ACT2FN[config.dense_act_fn]
|
|
|
|
def forward(self, hidden_states):
|
|
hidden_gelu = self.act(self.wi_0(hidden_states))
|
|
hidden_linear = self.wi_1(hidden_states)
|
|
hidden_states = hidden_gelu * hidden_linear
|
|
hidden_states = self.dropout(hidden_states)
|
|
|
|
# To make 8bit quantization work for google/flan-t5-xxl, self.wo is kept in float32.
|
|
# See https://github.com/huggingface/transformers/issues/20287
|
|
# we also make sure the weights are not in `int8` in case users will force `_keep_in_fp32_modules` to be `None``
|
|
if (
|
|
isinstance(self.wo.weight, torch.Tensor)
|
|
and hidden_states.dtype != self.wo.weight.dtype
|
|
and self.wo.weight.dtype != torch.int8
|
|
):
|
|
hidden_states = hidden_states.to(self.wo.weight.dtype)
|
|
|
|
hidden_states = self.wo(hidden_states)
|
|
return hidden_states
|
|
|
|
|
|
# Copied from transformers.models.t5.modeling_t5.T5LayerFF with T5->Pop2Piano
|
|
class Pop2PianoLayerFF(nn.Module):
|
|
def __init__(self, config: Pop2PianoConfig):
|
|
super().__init__()
|
|
if config.is_gated_act:
|
|
self.DenseReluDense = Pop2PianoDenseGatedActDense(config)
|
|
else:
|
|
self.DenseReluDense = Pop2PianoDenseActDense(config)
|
|
|
|
self.layer_norm = Pop2PianoLayerNorm(config.d_model, eps=config.layer_norm_epsilon)
|
|
self.dropout = nn.Dropout(config.dropout_rate)
|
|
|
|
def forward(self, hidden_states):
|
|
forwarded_states = self.layer_norm(hidden_states)
|
|
forwarded_states = self.DenseReluDense(forwarded_states)
|
|
hidden_states = hidden_states + self.dropout(forwarded_states)
|
|
return hidden_states
|
|
|
|
|
|
# Copied from transformers.models.t5.modeling_t5.T5Attention with T5->Pop2Piano,t5->pop2piano
|
|
class Pop2PianoAttention(nn.Module):
|
|
def __init__(
|
|
self,
|
|
config: Pop2PianoConfig,
|
|
has_relative_attention_bias=False,
|
|
layer_idx: Optional[int] = None,
|
|
):
|
|
super().__init__()
|
|
self.is_decoder = config.is_decoder
|
|
self.has_relative_attention_bias = has_relative_attention_bias
|
|
self.relative_attention_num_buckets = config.relative_attention_num_buckets
|
|
self.relative_attention_max_distance = config.relative_attention_max_distance
|
|
self.d_model = config.d_model
|
|
self.key_value_proj_dim = config.d_kv
|
|
self.n_heads = config.num_heads
|
|
self.dropout = config.dropout_rate
|
|
self.inner_dim = self.n_heads * self.key_value_proj_dim
|
|
self.layer_idx = layer_idx
|
|
if layer_idx is None and self.is_decoder:
|
|
logger.warning_once(
|
|
f"Instantiating a decoder {self.__class__.__name__} without passing `layer_idx` is not recommended and "
|
|
"will to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
|
|
"when creating this class."
|
|
)
|
|
|
|
# Mesh TensorFlow initialization to avoid scaling before softmax
|
|
self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
|
|
self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
|
|
self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
|
|
self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)
|
|
|
|
if self.has_relative_attention_bias:
|
|
self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
|
|
self.pruned_heads = set()
|
|
self.gradient_checkpointing = False
|
|
|
|
def prune_heads(self, heads):
|
|
if len(heads) == 0:
|
|
return
|
|
heads, index = find_pruneable_heads_and_indices(
|
|
heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads
|
|
)
|
|
# Prune linear layers
|
|
self.q = prune_linear_layer(self.q, index)
|
|
self.k = prune_linear_layer(self.k, index)
|
|
self.v = prune_linear_layer(self.v, index)
|
|
self.o = prune_linear_layer(self.o, index, dim=1)
|
|
# Update hyper params
|
|
self.n_heads = self.n_heads - len(heads)
|
|
self.inner_dim = self.key_value_proj_dim * self.n_heads
|
|
self.pruned_heads = self.pruned_heads.union(heads)
|
|
|
|
@staticmethod
|
|
def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
|
|
"""
|
|
Adapted from Mesh Tensorflow:
|
|
https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
|
|
|
|
Translate relative position to a bucket number for relative attention. The relative position is defined as
|
|
memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to
|
|
position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for
|
|
small absolute relative_position and larger buckets for larger absolute relative_positions. All relative
|
|
positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.
|
|
This should allow for more graceful generalization to longer sequences than the model has been trained on
|
|
|
|
Args:
|
|
relative_position: an int32 Tensor
|
|
bidirectional: a boolean - whether the attention is bidirectional
|
|
num_buckets: an integer
|
|
max_distance: an integer
|
|
|
|
Returns:
|
|
a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)
|
|
"""
|
|
relative_buckets = 0
|
|
if bidirectional:
|
|
num_buckets //= 2
|
|
relative_buckets += (relative_position > 0).to(torch.long) * num_buckets
|
|
relative_position = torch.abs(relative_position)
|
|
else:
|
|
relative_position = -torch.min(relative_position, torch.zeros_like(relative_position))
|
|
# now relative_position is in the range [0, inf)
|
|
|
|
# half of the buckets are for exact increments in positions
|
|
max_exact = num_buckets // 2
|
|
is_small = relative_position < max_exact
|
|
|
|
# The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
|
|
relative_position_if_large = max_exact + (
|
|
torch.log(relative_position.float() / max_exact)
|
|
/ math.log(max_distance / max_exact)
|
|
* (num_buckets - max_exact)
|
|
).to(torch.long)
|
|
relative_position_if_large = torch.min(
|
|
relative_position_if_large, torch.full_like(relative_position_if_large, num_buckets - 1)
|
|
)
|
|
|
|
relative_buckets += torch.where(is_small, relative_position, relative_position_if_large)
|
|
return relative_buckets
|
|
|
|
def compute_bias(self, query_length, key_length, device=None, cache_position=None):
|
|
"""Compute binned relative position bias"""
|
|
if device is None:
|
|
device = self.relative_attention_bias.weight.device
|
|
if cache_position is None:
|
|
context_position = torch.arange(query_length, dtype=torch.long, device=device)[:, None]
|
|
else:
|
|
context_position = cache_position[:, None].to(device)
|
|
memory_position = torch.arange(key_length, dtype=torch.long, device=device)[None, :]
|
|
relative_position = memory_position - context_position # shape (query_length, key_length)
|
|
relative_position_bucket = self._relative_position_bucket(
|
|
relative_position, # shape (query_length, key_length)
|
|
bidirectional=(not self.is_decoder),
|
|
num_buckets=self.relative_attention_num_buckets,
|
|
max_distance=self.relative_attention_max_distance,
|
|
)
|
|
values = self.relative_attention_bias(relative_position_bucket) # shape (query_length, key_length, num_heads)
|
|
values = values.permute([2, 0, 1]).unsqueeze(0) # shape (1, num_heads, query_length, key_length)
|
|
return values
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states,
|
|
mask=None,
|
|
key_value_states=None,
|
|
position_bias=None,
|
|
past_key_value=None,
|
|
layer_head_mask=None,
|
|
query_length=None,
|
|
use_cache=False,
|
|
output_attentions=False,
|
|
cache_position=None,
|
|
):
|
|
"""
|
|
Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states).
|
|
"""
|
|
# Input is (batch_size, seq_length, dim)
|
|
# Mask is (batch_size, 1, 1, key_length) (non-causal encoder) or (batch_size, 1, seq_length, key_length) (causal decoder)
|
|
batch_size, seq_length = hidden_states.shape[:2]
|
|
|
|
# if key_value_states are provided this layer is used as a cross-attention layer for the decoder
|
|
is_cross_attention = key_value_states is not None
|
|
|
|
query_states = self.q(hidden_states)
|
|
query_states = query_states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2)
|
|
|
|
# Check is encoder-decoder model is being used. Otherwise we'll get `DynamicCache`
|
|
if past_key_value is not None and isinstance(past_key_value, EncoderDecoderCache):
|
|
is_updated = past_key_value.is_updated.get(self.layer_idx)
|
|
if is_cross_attention:
|
|
# after the first generated id, we can subsequently re-use all key/value_states from cache
|
|
curr_past_key_value = past_key_value.cross_attention_cache
|
|
else:
|
|
curr_past_key_value = past_key_value.self_attention_cache
|
|
else:
|
|
curr_past_key_value = past_key_value
|
|
|
|
current_states = key_value_states if is_cross_attention else hidden_states
|
|
if is_cross_attention and past_key_value is not None and is_updated:
|
|
# reuse k,v, cross_attentions
|
|
key_states = curr_past_key_value.layers[self.layer_idx].keys
|
|
value_states = curr_past_key_value.layers[self.layer_idx].values
|
|
else:
|
|
key_states = self.k(current_states)
|
|
value_states = self.v(current_states)
|
|
key_states = key_states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2)
|
|
value_states = value_states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2)
|
|
|
|
if past_key_value is not None:
|
|
# save all key/value_states to cache to be re-used for fast auto-regressive generation
|
|
cache_position = cache_position if not is_cross_attention else None
|
|
key_states, value_states = curr_past_key_value.update(
|
|
key_states, value_states, self.layer_idx, {"cache_position": cache_position}
|
|
)
|
|
# set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls
|
|
if is_cross_attention:
|
|
past_key_value.is_updated[self.layer_idx] = True
|
|
|
|
# compute scores, equivalent of torch.einsum("bnqd,bnkd->bnqk", query_states, key_states), compatible with onnx op>9
|
|
scores = torch.matmul(query_states, key_states.transpose(3, 2))
|
|
|
|
if position_bias is None:
|
|
key_length = key_states.shape[-2]
|
|
# cache position is 0-indexed so we add 1 to get the real length of queries (aka with past)
|
|
real_seq_length = query_length if query_length is not None else cache_position[-1] + 1
|
|
if not self.has_relative_attention_bias:
|
|
position_bias = torch.zeros(
|
|
(1, self.n_heads, seq_length, key_length), device=scores.device, dtype=scores.dtype
|
|
)
|
|
if self.gradient_checkpointing and self.training:
|
|
position_bias.requires_grad = True
|
|
else:
|
|
position_bias = self.compute_bias(
|
|
real_seq_length, key_length, device=scores.device, cache_position=cache_position
|
|
)
|
|
position_bias = position_bias[:, :, -seq_length:, :]
|
|
|
|
if mask is not None:
|
|
causal_mask = mask[:, :, :, : key_states.shape[-2]]
|
|
position_bias = position_bias + causal_mask
|
|
|
|
if self.pruned_heads:
|
|
mask = torch.ones(position_bias.shape[1])
|
|
mask[list(self.pruned_heads)] = 0
|
|
position_bias_masked = position_bias[:, mask.bool()]
|
|
else:
|
|
position_bias_masked = position_bias
|
|
|
|
scores += position_bias_masked
|
|
|
|
# (batch_size, n_heads, seq_length, key_length)
|
|
attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(scores)
|
|
attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
|
|
|
|
# Mask heads if we want to
|
|
if layer_head_mask is not None:
|
|
attn_weights = attn_weights * layer_head_mask
|
|
|
|
attn_output = torch.matmul(attn_weights, value_states)
|
|
|
|
attn_output = attn_output.transpose(1, 2).contiguous()
|
|
attn_output = attn_output.view(batch_size, -1, self.inner_dim)
|
|
attn_output = self.o(attn_output)
|
|
|
|
outputs = (attn_output, position_bias)
|
|
|
|
if output_attentions:
|
|
outputs = outputs + (attn_weights,)
|
|
return outputs
|
|
|
|
|
|
# Copied from transformers.models.t5.modeling_t5.T5LayerSelfAttention with T5->Pop2Piano,t5->pop2piano
|
|
class Pop2PianoLayerSelfAttention(nn.Module):
|
|
def __init__(self, config, has_relative_attention_bias=False, layer_idx: Optional[int] = None):
|
|
super().__init__()
|
|
self.SelfAttention = Pop2PianoAttention(
|
|
config, has_relative_attention_bias=has_relative_attention_bias, layer_idx=layer_idx
|
|
)
|
|
self.layer_norm = Pop2PianoLayerNorm(config.d_model, eps=config.layer_norm_epsilon)
|
|
self.dropout = nn.Dropout(config.dropout_rate)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states,
|
|
attention_mask=None,
|
|
position_bias=None,
|
|
layer_head_mask=None,
|
|
past_key_value=None,
|
|
use_cache=False,
|
|
output_attentions=False,
|
|
cache_position=None,
|
|
):
|
|
normed_hidden_states = self.layer_norm(hidden_states)
|
|
attention_output = self.SelfAttention(
|
|
normed_hidden_states,
|
|
mask=attention_mask,
|
|
position_bias=position_bias,
|
|
layer_head_mask=layer_head_mask,
|
|
past_key_value=past_key_value,
|
|
use_cache=use_cache,
|
|
output_attentions=output_attentions,
|
|
cache_position=cache_position,
|
|
)
|
|
hidden_states = hidden_states + self.dropout(attention_output[0])
|
|
outputs = (hidden_states,) + attention_output[1:] # add attentions if we output them
|
|
return outputs
|
|
|
|
|
|
# Copied from transformers.models.t5.modeling_t5.T5LayerCrossAttention with T5->Pop2Piano,t5->pop2piano
|
|
class Pop2PianoLayerCrossAttention(nn.Module):
|
|
def __init__(self, config, layer_idx: Optional[int] = None):
|
|
super().__init__()
|
|
self.EncDecAttention = Pop2PianoAttention(config, has_relative_attention_bias=False, layer_idx=layer_idx)
|
|
self.layer_norm = Pop2PianoLayerNorm(config.d_model, eps=config.layer_norm_epsilon)
|
|
self.dropout = nn.Dropout(config.dropout_rate)
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states,
|
|
key_value_states,
|
|
attention_mask=None,
|
|
position_bias=None,
|
|
layer_head_mask=None,
|
|
past_key_value=None,
|
|
use_cache=False,
|
|
query_length=None,
|
|
output_attentions=False,
|
|
cache_position=None,
|
|
):
|
|
normed_hidden_states = self.layer_norm(hidden_states)
|
|
attention_output = self.EncDecAttention(
|
|
normed_hidden_states,
|
|
mask=attention_mask,
|
|
key_value_states=key_value_states,
|
|
position_bias=position_bias,
|
|
layer_head_mask=layer_head_mask,
|
|
past_key_value=past_key_value,
|
|
use_cache=use_cache,
|
|
query_length=query_length,
|
|
output_attentions=output_attentions,
|
|
cache_position=cache_position,
|
|
)
|
|
layer_output = hidden_states + self.dropout(attention_output[0])
|
|
outputs = (layer_output,) + attention_output[1:] # add attentions if we output them
|
|
return outputs
|
|
|
|
|
|
# Copied from transformers.models.t5.modeling_t5.T5Block with T5->Pop2Piano,t5->pop2piano
|
|
class Pop2PianoBlock(GradientCheckpointingLayer):
|
|
def __init__(self, config, has_relative_attention_bias=False, layer_idx: Optional[int] = None):
|
|
super().__init__()
|
|
self.is_decoder = config.is_decoder
|
|
self.layer = nn.ModuleList()
|
|
self.layer.append(
|
|
Pop2PianoLayerSelfAttention(
|
|
config, has_relative_attention_bias=has_relative_attention_bias, layer_idx=layer_idx
|
|
)
|
|
)
|
|
if self.is_decoder:
|
|
self.layer.append(Pop2PianoLayerCrossAttention(config, layer_idx=layer_idx))
|
|
|
|
self.layer.append(Pop2PianoLayerFF(config))
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states,
|
|
attention_mask=None,
|
|
position_bias=None,
|
|
encoder_hidden_states=None,
|
|
encoder_attention_mask=None,
|
|
encoder_decoder_position_bias=None,
|
|
layer_head_mask=None,
|
|
cross_attn_layer_head_mask=None,
|
|
past_key_value=None,
|
|
use_cache=False,
|
|
output_attentions=False,
|
|
return_dict=True,
|
|
cache_position=None,
|
|
):
|
|
self_attention_outputs = self.layer[0](
|
|
hidden_states,
|
|
attention_mask=attention_mask,
|
|
position_bias=position_bias,
|
|
layer_head_mask=layer_head_mask,
|
|
past_key_value=past_key_value,
|
|
use_cache=use_cache,
|
|
output_attentions=output_attentions,
|
|
cache_position=cache_position,
|
|
)
|
|
hidden_states = self_attention_outputs[0]
|
|
attention_outputs = self_attention_outputs[1:] # Keep self-attention outputs and relative position weights
|
|
|
|
# clamp inf values to enable fp16 training
|
|
if hidden_states.dtype == torch.float16:
|
|
clamp_value = torch.where(
|
|
torch.isinf(hidden_states).any(),
|
|
torch.finfo(hidden_states.dtype).max - 1000,
|
|
torch.finfo(hidden_states.dtype).max,
|
|
)
|
|
hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
|
|
|
|
do_cross_attention = self.is_decoder and encoder_hidden_states is not None
|
|
if do_cross_attention:
|
|
cross_attention_outputs = self.layer[1](
|
|
hidden_states,
|
|
key_value_states=encoder_hidden_states,
|
|
attention_mask=encoder_attention_mask,
|
|
position_bias=encoder_decoder_position_bias,
|
|
layer_head_mask=cross_attn_layer_head_mask,
|
|
past_key_value=past_key_value,
|
|
query_length=cache_position[-1] + 1,
|
|
use_cache=use_cache,
|
|
output_attentions=output_attentions,
|
|
)
|
|
hidden_states = cross_attention_outputs[0]
|
|
|
|
# clamp inf values to enable fp16 training
|
|
if hidden_states.dtype == torch.float16:
|
|
clamp_value = torch.where(
|
|
torch.isinf(hidden_states).any(),
|
|
torch.finfo(hidden_states.dtype).max - 1000,
|
|
torch.finfo(hidden_states.dtype).max,
|
|
)
|
|
hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
|
|
|
|
# Keep cross-attention outputs and relative position weights
|
|
attention_outputs = attention_outputs + cross_attention_outputs[1:]
|
|
|
|
# Apply Feed Forward layer
|
|
hidden_states = self.layer[-1](hidden_states)
|
|
|
|
# clamp inf values to enable fp16 training
|
|
if hidden_states.dtype == torch.float16:
|
|
clamp_value = torch.where(
|
|
torch.isinf(hidden_states).any(),
|
|
torch.finfo(hidden_states.dtype).max - 1000,
|
|
torch.finfo(hidden_states.dtype).max,
|
|
)
|
|
hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
|
|
|
|
outputs = (hidden_states,)
|
|
|
|
return (
|
|
outputs + attention_outputs
|
|
) # hidden-states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)
|
|
|
|
|
|
@auto_docstring
|
|
class Pop2PianoPreTrainedModel(PreTrainedModel):
|
|
config: Pop2PianoConfig
|
|
base_model_prefix = "transformer"
|
|
is_parallelizable = False
|
|
supports_gradient_checkpointing = True
|
|
|
|
_can_compile_fullgraph = False
|
|
_no_split_modules = ["Pop2PianoBlock"]
|
|
_keep_in_fp32_modules = ["wo"]
|
|
|
|
def _init_weights(self, module):
|
|
"""Initialize the weights"""
|
|
factor = self.config.initializer_factor # Used for testing weights initialization
|
|
if isinstance(module, Pop2PianoLayerNorm):
|
|
module.weight.data.fill_(factor * 1.0)
|
|
elif isinstance(module, Pop2PianoConcatEmbeddingToMel):
|
|
module.embedding.weight.data.normal_(mean=0.0, std=factor * 1.0)
|
|
elif isinstance(module, Pop2PianoForConditionalGeneration):
|
|
# Mesh TensorFlow embeddings initialization
|
|
# See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624
|
|
module.shared.weight.data.normal_(mean=0.0, std=factor * 1.0)
|
|
if hasattr(module, "lm_head") and not self.config.tie_word_embeddings:
|
|
module.lm_head.weight.data.normal_(mean=0.0, std=factor * 1.0)
|
|
elif isinstance(module, Pop2PianoDenseActDense):
|
|
# Mesh TensorFlow FF initialization
|
|
# See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56
|
|
# and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89
|
|
module.wi.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
|
|
if hasattr(module.wi, "bias") and module.wi.bias is not None:
|
|
module.wi.bias.data.zero_()
|
|
module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5))
|
|
if hasattr(module.wo, "bias") and module.wo.bias is not None:
|
|
module.wo.bias.data.zero_()
|
|
elif isinstance(module, Pop2PianoDenseGatedActDense):
|
|
module.wi_0.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
|
|
if hasattr(module.wi_0, "bias") and module.wi_0.bias is not None:
|
|
module.wi_0.bias.data.zero_()
|
|
module.wi_1.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
|
|
if hasattr(module.wi_1, "bias") and module.wi_1.bias is not None:
|
|
module.wi_1.bias.data.zero_()
|
|
module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5))
|
|
if hasattr(module.wo, "bias") and module.wo.bias is not None:
|
|
module.wo.bias.data.zero_()
|
|
elif isinstance(module, Pop2PianoAttention):
|
|
# Mesh TensorFlow attention initialization to avoid scaling before softmax
|
|
# See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136
|
|
d_model = self.config.d_model
|
|
key_value_proj_dim = self.config.d_kv
|
|
n_heads = self.config.num_heads
|
|
module.q.weight.data.normal_(mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5))
|
|
module.k.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))
|
|
module.v.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))
|
|
module.o.weight.data.normal_(mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5))
|
|
if module.has_relative_attention_bias:
|
|
module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor * ((d_model) ** -0.5))
|
|
|
|
def _shift_right(self, input_ids):
|
|
decoder_start_token_id = self.config.decoder_start_token_id
|
|
pad_token_id = self.config.pad_token_id
|
|
|
|
if decoder_start_token_id is None:
|
|
raise ValueError(
|
|
"self.model.config.decoder_start_token_id has to be defined. In Pop2Piano it is usually set to the pad_token_id."
|
|
)
|
|
|
|
# shift inputs to the right
|
|
if is_torch_fx_proxy(input_ids):
|
|
# Item assignment is not supported natively for proxies.
|
|
shifted_input_ids = torch.full(input_ids.shape[:-1] + (1,), decoder_start_token_id)
|
|
shifted_input_ids = torch.cat([shifted_input_ids, input_ids[..., :-1]], dim=-1)
|
|
else:
|
|
shifted_input_ids = input_ids.new_zeros(input_ids.shape)
|
|
shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
|
|
shifted_input_ids[..., 0] = decoder_start_token_id
|
|
|
|
if pad_token_id is None:
|
|
raise ValueError("self.model.config.pad_token_id has to be defined.")
|
|
# replace possible -100 values in labels by `pad_token_id`
|
|
shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
|
|
|
|
return shifted_input_ids
|
|
|
|
|
|
class Pop2PianoStack(Pop2PianoPreTrainedModel):
|
|
# Copied from transformers.models.t5.modeling_t5.T5Stack.__init__ with T5->Pop2Piano,t5->pop2piano
|
|
def __init__(self, config, embed_tokens=None):
|
|
super().__init__(config)
|
|
|
|
self.embed_tokens = embed_tokens
|
|
self.is_decoder = config.is_decoder
|
|
|
|
self.block = nn.ModuleList(
|
|
[
|
|
Pop2PianoBlock(config, has_relative_attention_bias=bool(i == 0), layer_idx=i)
|
|
for i in range(config.num_layers)
|
|
]
|
|
)
|
|
self.final_layer_norm = Pop2PianoLayerNorm(config.d_model, eps=config.layer_norm_epsilon)
|
|
self.dropout = nn.Dropout(config.dropout_rate)
|
|
|
|
# Initialize weights and apply final processing
|
|
self.post_init()
|
|
# Model parallel
|
|
self.model_parallel = False
|
|
self.device_map = None
|
|
self.gradient_checkpointing = False
|
|
|
|
# Copied from transformers.models.t5.modeling_t5.T5Stack.set_input_embeddings
|
|
def set_input_embeddings(self, new_embeddings):
|
|
self.embed_tokens = new_embeddings
|
|
|
|
def forward(
|
|
self,
|
|
input_ids=None,
|
|
attention_mask=None,
|
|
encoder_hidden_states=None,
|
|
encoder_attention_mask=None,
|
|
inputs_embeds=None,
|
|
head_mask=None,
|
|
cross_attn_head_mask=None,
|
|
past_key_values=None,
|
|
use_cache=None,
|
|
output_attentions=None,
|
|
output_hidden_states=None,
|
|
return_dict=None,
|
|
cache_position=None,
|
|
):
|
|
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
|
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
|
output_hidden_states = (
|
|
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
|
)
|
|
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
|
if input_ids is not None and inputs_embeds is not None:
|
|
err_msg_prefix = "decoder_" if self.is_decoder else ""
|
|
raise ValueError(
|
|
f"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}inputs_embeds at the same time"
|
|
)
|
|
elif input_ids is not None:
|
|
input_shape = input_ids.size()
|
|
input_ids = input_ids.view(-1, input_shape[-1])
|
|
elif inputs_embeds is not None:
|
|
input_shape = inputs_embeds.size()[:-1]
|
|
else:
|
|
err_msg_prefix = "decoder_" if self.is_decoder else ""
|
|
raise ValueError(f"You have to specify either {err_msg_prefix}input_ids or {err_msg_prefix}inputs_embeds")
|
|
|
|
if self.gradient_checkpointing and self.training:
|
|
if use_cache:
|
|
logger.warning_once(
|
|
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
|
)
|
|
use_cache = False
|
|
|
|
if inputs_embeds is None:
|
|
if self.embed_tokens is None:
|
|
raise ValueError("You have to initialize the model with valid token embeddings")
|
|
inputs_embeds = self.embed_tokens(input_ids)
|
|
|
|
batch_size, seq_length = input_shape
|
|
|
|
if use_cache is True:
|
|
if not self.is_decoder:
|
|
raise ValueError(f"`use_cache` can only be set to `True` if {self} is used as a decoder")
|
|
|
|
if self.is_decoder:
|
|
if use_cache and past_key_values is None:
|
|
if self.config.is_encoder_decoder:
|
|
past_key_values = EncoderDecoderCache(DynamicCache(), DynamicCache())
|
|
else:
|
|
past_key_values = DynamicCache()
|
|
elif not self.is_decoder:
|
|
# do not pass cache object down the line for encoder stack
|
|
# it messes indexing later in decoder-stack because cache object is modified in-place
|
|
past_key_values = None
|
|
|
|
past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0
|
|
if cache_position is None:
|
|
cache_position = torch.arange(
|
|
past_key_values_length, past_key_values_length + seq_length, device=inputs_embeds.device
|
|
)
|
|
|
|
if attention_mask is None and not is_torchdynamo_compiling():
|
|
# required mask seq length can be calculated via length of past cache
|
|
mask_seq_length = past_key_values_length + seq_length
|
|
attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)
|
|
|
|
if self.config.is_decoder:
|
|
causal_mask = self._update_causal_mask(
|
|
attention_mask,
|
|
inputs_embeds,
|
|
cache_position,
|
|
past_key_values.self_attention_cache
|
|
if isinstance(past_key_values, EncoderDecoderCache)
|
|
else past_key_values,
|
|
output_attentions,
|
|
)
|
|
else:
|
|
causal_mask = attention_mask[:, None, None, :]
|
|
causal_mask = causal_mask.to(dtype=inputs_embeds.dtype)
|
|
causal_mask = (1.0 - causal_mask) * torch.finfo(inputs_embeds.dtype).min
|
|
|
|
# If a 2D or 3D attention mask is provided for the cross-attention
|
|
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
|
|
if self.is_decoder and encoder_hidden_states is not None:
|
|
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
|
|
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
|
|
if encoder_attention_mask is None:
|
|
encoder_attention_mask = torch.ones(encoder_hidden_shape, device=inputs_embeds.device)
|
|
encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
|
|
else:
|
|
encoder_extended_attention_mask = None
|
|
|
|
# Prepare head mask if needed
|
|
head_mask = self.get_head_mask(head_mask, self.config.num_layers)
|
|
cross_attn_head_mask = self.get_head_mask(cross_attn_head_mask, self.config.num_layers)
|
|
all_hidden_states = () if output_hidden_states else None
|
|
all_attentions = () if output_attentions else None
|
|
all_cross_attentions = () if (output_attentions and self.is_decoder) else None
|
|
position_bias = None
|
|
encoder_decoder_position_bias = None
|
|
|
|
hidden_states = self.dropout(inputs_embeds)
|
|
|
|
for i, layer_module in enumerate(self.block):
|
|
layer_head_mask = head_mask[i]
|
|
cross_attn_layer_head_mask = cross_attn_head_mask[i]
|
|
if output_hidden_states:
|
|
all_hidden_states = all_hidden_states + (hidden_states,)
|
|
|
|
layer_outputs = layer_module(
|
|
hidden_states,
|
|
causal_mask,
|
|
position_bias,
|
|
encoder_hidden_states,
|
|
encoder_extended_attention_mask,
|
|
encoder_decoder_position_bias, # as a positional argument for gradient checkpointing
|
|
layer_head_mask=layer_head_mask,
|
|
cross_attn_layer_head_mask=cross_attn_layer_head_mask,
|
|
past_key_value=past_key_values,
|
|
use_cache=use_cache,
|
|
output_attentions=output_attentions,
|
|
cache_position=cache_position,
|
|
)
|
|
|
|
hidden_states = layer_outputs[0]
|
|
|
|
# We share the position biases between the layers - the first layer store them
|
|
# layer_outputs = hidden-states, key-value-states (self-attention position bias), (self-attention weights),
|
|
# (cross-attention position bias), (cross-attention weights)
|
|
position_bias = layer_outputs[1]
|
|
if self.is_decoder and encoder_hidden_states is not None:
|
|
encoder_decoder_position_bias = layer_outputs[3 if output_attentions else 2]
|
|
|
|
if output_attentions:
|
|
all_attentions = all_attentions + (layer_outputs[2],)
|
|
if self.is_decoder:
|
|
all_cross_attentions = all_cross_attentions + (layer_outputs[4],)
|
|
|
|
hidden_states = self.final_layer_norm(hidden_states)
|
|
hidden_states = self.dropout(hidden_states)
|
|
|
|
# Add last layer
|
|
if output_hidden_states:
|
|
all_hidden_states = all_hidden_states + (hidden_states,)
|
|
|
|
if not return_dict:
|
|
return tuple(
|
|
v
|
|
for v in [
|
|
hidden_states,
|
|
past_key_values,
|
|
all_hidden_states,
|
|
all_attentions,
|
|
all_cross_attentions,
|
|
]
|
|
if v is not None
|
|
)
|
|
return BaseModelOutputWithPastAndCrossAttentions(
|
|
last_hidden_state=hidden_states,
|
|
past_key_values=past_key_values,
|
|
hidden_states=all_hidden_states,
|
|
attentions=all_attentions,
|
|
cross_attentions=all_cross_attentions,
|
|
)
|
|
|
|
# Copied from transformers.models.gptj.modeling_gptj.GPTJModel._update_causal_mask
|
|
def _update_causal_mask(
|
|
self,
|
|
attention_mask: Union[torch.Tensor, "BlockMask"],
|
|
input_tensor: torch.Tensor,
|
|
cache_position: torch.Tensor,
|
|
past_key_values: Cache,
|
|
output_attentions: bool = False,
|
|
):
|
|
if self.config._attn_implementation == "flash_attention_2":
|
|
if attention_mask is not None and (attention_mask == 0.0).any():
|
|
return attention_mask
|
|
return None
|
|
if self.config._attn_implementation == "flex_attention":
|
|
if isinstance(attention_mask, torch.Tensor):
|
|
attention_mask = make_flex_block_causal_mask(attention_mask)
|
|
return attention_mask
|
|
|
|
# For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
|
|
# order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
|
|
# to infer the attention mask.
|
|
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
|
using_compilable_cache = past_key_values.is_compileable if past_key_values is not None else False
|
|
|
|
# When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
|
|
if self.config._attn_implementation == "sdpa" and not using_compilable_cache and not output_attentions:
|
|
if AttentionMaskConverter._ignore_causal_mask_sdpa(
|
|
attention_mask,
|
|
inputs_embeds=input_tensor,
|
|
past_key_values_length=past_seen_tokens,
|
|
is_training=self.training,
|
|
):
|
|
return None
|
|
|
|
dtype = input_tensor.dtype
|
|
sequence_length = input_tensor.shape[1]
|
|
if using_compilable_cache:
|
|
target_length = past_key_values.get_max_cache_shape()
|
|
else:
|
|
target_length = (
|
|
attention_mask.shape[-1]
|
|
if isinstance(attention_mask, torch.Tensor)
|
|
else past_seen_tokens + sequence_length + 1
|
|
)
|
|
|
|
# In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
|
|
causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
|
|
attention_mask,
|
|
sequence_length=sequence_length,
|
|
target_length=target_length,
|
|
dtype=dtype,
|
|
cache_position=cache_position,
|
|
batch_size=input_tensor.shape[0],
|
|
)
|
|
|
|
if (
|
|
self.config._attn_implementation == "sdpa"
|
|
and attention_mask is not None
|
|
and attention_mask.device.type in ["cuda", "xpu", "npu"]
|
|
and not output_attentions
|
|
):
|
|
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
|
|
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
|
|
# Details: https://github.com/pytorch/pytorch/issues/110213
|
|
min_dtype = torch.finfo(dtype).min
|
|
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
|
|
|
|
return causal_mask
|
|
|
|
@staticmethod
|
|
# Copied from transformers.models.gptj.modeling_gptj.GPTJModel._prepare_4d_causal_attention_mask_with_cache_position
|
|
def _prepare_4d_causal_attention_mask_with_cache_position(
|
|
attention_mask: torch.Tensor,
|
|
sequence_length: int,
|
|
target_length: int,
|
|
dtype: torch.dtype,
|
|
cache_position: torch.Tensor,
|
|
batch_size: int,
|
|
**kwargs,
|
|
):
|
|
"""
|
|
Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
|
|
`(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
|
|
|
|
Args:
|
|
attention_mask (`torch.Tensor`):
|
|
A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
|
|
`(batch_size, 1, query_length, key_value_length)`.
|
|
sequence_length (`int`):
|
|
The sequence length being processed.
|
|
target_length (`int`):
|
|
The target length: when generating with static cache, the mask should be as long as the static cache,
|
|
to account for the 0 padding, the part of the cache that is not filled yet.
|
|
dtype (`torch.dtype`):
|
|
The dtype to use for the 4D attention mask.
|
|
cache_position (`torch.Tensor`):
|
|
Indices depicting the position of the input sequence tokens in the sequence.
|
|
batch_size (`torch.Tensor`):
|
|
Batch size.
|
|
"""
|
|
if attention_mask is not None and attention_mask.dim() == 4:
|
|
# In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
|
|
causal_mask = attention_mask
|
|
else:
|
|
min_dtype = torch.finfo(dtype).min
|
|
causal_mask = torch.full(
|
|
(sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device
|
|
)
|
|
if sequence_length != 1:
|
|
causal_mask = torch.triu(causal_mask, diagonal=1)
|
|
causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(-1, 1)
|
|
causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
|
|
if attention_mask is not None:
|
|
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
|
mask_length = attention_mask.shape[-1]
|
|
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
|
|
causal_mask.device
|
|
)
|
|
padding_mask = padding_mask == 0
|
|
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
|
|
padding_mask, min_dtype
|
|
)
|
|
|
|
return causal_mask
|
|
|
|
|
|
class Pop2PianoConcatEmbeddingToMel(nn.Module):
|
|
"""Embedding Matrix for `composer` tokens."""
|
|
|
|
def __init__(self, config):
|
|
super().__init__()
|
|
self.embedding = nn.Embedding(num_embeddings=config.composer_vocab_size, embedding_dim=config.d_model)
|
|
|
|
def forward(self, feature, index_value, embedding_offset):
|
|
index_shifted = index_value - embedding_offset
|
|
composer_embedding = self.embedding(index_shifted).unsqueeze(1)
|
|
inputs_embeds = torch.cat([composer_embedding, feature], dim=1)
|
|
return inputs_embeds
|
|
|
|
|
|
@auto_docstring(
|
|
custom_intro="""
|
|
Pop2Piano Model with a `language modeling` head on top.
|
|
"""
|
|
)
|
|
class Pop2PianoForConditionalGeneration(Pop2PianoPreTrainedModel, GenerationMixin):
|
|
_tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
|
|
|
|
def __init__(self, config: Pop2PianoConfig):
|
|
super().__init__(config)
|
|
self.config = config
|
|
self.model_dim = config.d_model
|
|
|
|
self.shared = nn.Embedding(config.vocab_size, config.d_model)
|
|
|
|
self.mel_conditioner = Pop2PianoConcatEmbeddingToMel(config)
|
|
|
|
encoder_config = copy.deepcopy(config)
|
|
encoder_config.is_decoder = False
|
|
encoder_config.use_cache = False
|
|
encoder_config.tie_encoder_decoder = False
|
|
|
|
self.encoder = Pop2PianoStack(encoder_config, self.shared)
|
|
|
|
decoder_config = copy.deepcopy(config)
|
|
decoder_config.is_decoder = True
|
|
decoder_config.tie_encoder_decoder = False
|
|
decoder_config.num_layers = config.num_decoder_layers
|
|
self.decoder = Pop2PianoStack(decoder_config, self.shared)
|
|
|
|
self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
|
|
|
|
# Initialize weights and apply final processing
|
|
self.post_init()
|
|
|
|
def get_input_embeddings(self):
|
|
return self.shared
|
|
|
|
def set_input_embeddings(self, new_embeddings):
|
|
self.shared = new_embeddings
|
|
self.encoder.set_input_embeddings(new_embeddings)
|
|
self.decoder.set_input_embeddings(new_embeddings)
|
|
|
|
def get_encoder(self):
|
|
return self.encoder
|
|
|
|
def get_decoder(self):
|
|
return self.decoder
|
|
|
|
def get_mel_conditioner_outputs(
|
|
self,
|
|
input_features: torch.FloatTensor,
|
|
composer: str,
|
|
generation_config: GenerationConfig,
|
|
attention_mask: Optional[torch.FloatTensor] = None,
|
|
):
|
|
"""
|
|
This method is used to concatenate mel conditioner tokens at the front of the input_features in order to
|
|
control the type of MIDI token generated by the model.
|
|
|
|
Args:
|
|
input_features (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
|
input features extracted from the feature extractor.
|
|
composer (`str`):
|
|
composer token which determines the type of MIDI tokens to be generated.
|
|
generation_config (`~generation.GenerationConfig`):
|
|
The generation is used to get the composer-feature_token pair.
|
|
attention_mask (``, *optional*):
|
|
For batched generation `input_features` are padded to have the same shape across all examples.
|
|
`attention_mask` helps to determine which areas were padded and which were not.
|
|
- 1 for tokens that are **not padded**,
|
|
- 0 for tokens that are **padded**.
|
|
"""
|
|
composer_to_feature_token = generation_config.composer_to_feature_token
|
|
if composer not in composer_to_feature_token.keys():
|
|
raise ValueError(
|
|
f"Please choose a composer from {list(composer_to_feature_token.keys())}. Composer received - {composer}"
|
|
)
|
|
composer_value = composer_to_feature_token[composer]
|
|
composer_value = torch.tensor(composer_value, device=self.device)
|
|
composer_value = composer_value.repeat(input_features.shape[0])
|
|
|
|
embedding_offset = min(composer_to_feature_token.values())
|
|
|
|
input_features = self.mel_conditioner(
|
|
feature=input_features,
|
|
index_value=composer_value,
|
|
embedding_offset=embedding_offset,
|
|
)
|
|
if attention_mask is not None:
|
|
input_features[~attention_mask[:, 0].bool()] = 0.0
|
|
|
|
# since self.mel_conditioner adds a new array at the front of inputs_embeds we need to do the same for attention_mask to keep the shapes same
|
|
attention_mask = torch.concatenate([attention_mask[:, 0].view(-1, 1), attention_mask], axis=1)
|
|
return input_features, attention_mask
|
|
|
|
return input_features, None
|
|
|
|
@auto_docstring
|
|
def forward(
|
|
self,
|
|
input_ids: Optional[torch.LongTensor] = None,
|
|
attention_mask: Optional[torch.FloatTensor] = None,
|
|
decoder_input_ids: Optional[torch.LongTensor] = None,
|
|
decoder_attention_mask: Optional[torch.BoolTensor] = None,
|
|
head_mask: Optional[torch.FloatTensor] = None,
|
|
decoder_head_mask: Optional[torch.FloatTensor] = None,
|
|
cross_attn_head_mask: Optional[torch.Tensor] = None,
|
|
encoder_outputs: Optional[tuple[tuple[torch.Tensor]]] = None,
|
|
past_key_values: Optional[Cache] = None,
|
|
inputs_embeds: Optional[torch.FloatTensor] = None,
|
|
input_features: Optional[torch.FloatTensor] = None,
|
|
decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
|
|
labels: Optional[torch.LongTensor] = None,
|
|
use_cache: Optional[bool] = None,
|
|
output_attentions: Optional[bool] = None,
|
|
output_hidden_states: Optional[bool] = None,
|
|
return_dict: Optional[bool] = None,
|
|
cache_position: Optional[torch.LongTensor] = None,
|
|
) -> Union[tuple[torch.FloatTensor], Seq2SeqLMOutput]:
|
|
r"""
|
|
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
|
Indices of input sequence tokens in the vocabulary. Pop2Piano is a model with relative position embeddings
|
|
so you should be able to pad the inputs on both the right and the left. Indices can be obtained using
|
|
[`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for detail.
|
|
[What are input IDs?](../glossary#input-ids) To know more on how to prepare `input_ids` for pretraining
|
|
take a look a [Pop2Piano Training](./Pop2Piano#training).
|
|
decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
|
|
Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using
|
|
[`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details.
|
|
[What are decoder input IDs?](../glossary#decoder-input-ids) Pop2Piano uses the `pad_token_id` as the
|
|
starting token for `decoder_input_ids` generation. If `past_key_values` is used, optionally only the last
|
|
`decoder_input_ids` have to be input (see `past_key_values`). To know more on how to prepare
|
|
decoder_attention_mask (`torch.BoolTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
|
|
Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
|
|
be used by default.
|
|
decoder_head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
|
|
Mask to nullify selected heads of the self-attention modules in the decoder. Mask values selected in `[0,
|
|
1]`:
|
|
- 1 indicates the head is **not masked**,
|
|
- 0 indicates the head is **masked**.
|
|
cross_attn_head_mask (`torch.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
|
|
Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in
|
|
`[0, 1]`:
|
|
- 1 indicates the head is **not masked**,
|
|
- 0 indicates the head is **masked**.
|
|
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
|
Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ...,
|
|
config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for
|
|
labels in `[0, ..., config.vocab_size]`
|
|
"""
|
|
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
|
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
|
if inputs_embeds is not None and input_features is not None:
|
|
raise ValueError("Both `inputs_embeds` and `input_features` received! Please provide only one of them")
|
|
elif input_features is not None and inputs_embeds is None:
|
|
inputs_embeds = input_features
|
|
|
|
# Encode if needed (training, first prediction pass)
|
|
if encoder_outputs is None:
|
|
# Convert encoder inputs in embeddings if needed
|
|
encoder_outputs = self.encoder(
|
|
input_ids=input_ids,
|
|
attention_mask=attention_mask,
|
|
inputs_embeds=inputs_embeds,
|
|
head_mask=head_mask,
|
|
output_attentions=output_attentions,
|
|
output_hidden_states=output_hidden_states,
|
|
return_dict=return_dict,
|
|
)
|
|
elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
|
|
encoder_outputs = BaseModelOutput(
|
|
last_hidden_state=encoder_outputs[0],
|
|
hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
|
|
attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
|
|
)
|
|
|
|
hidden_states = encoder_outputs[0]
|
|
|
|
if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None:
|
|
# get decoder inputs from shifting lm labels to the right
|
|
decoder_input_ids = self._shift_right(labels)
|
|
|
|
# Decode
|
|
decoder_outputs = self.decoder(
|
|
input_ids=decoder_input_ids,
|
|
attention_mask=decoder_attention_mask,
|
|
inputs_embeds=decoder_inputs_embeds,
|
|
past_key_values=past_key_values,
|
|
encoder_hidden_states=hidden_states,
|
|
encoder_attention_mask=attention_mask,
|
|
head_mask=decoder_head_mask,
|
|
cross_attn_head_mask=cross_attn_head_mask,
|
|
use_cache=use_cache,
|
|
output_attentions=output_attentions,
|
|
output_hidden_states=output_hidden_states,
|
|
return_dict=return_dict,
|
|
cache_position=cache_position,
|
|
)
|
|
|
|
sequence_output = decoder_outputs[0]
|
|
|
|
if self.config.tie_word_embeddings:
|
|
# Rescale output before projecting on vocab
|
|
# See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
|
|
sequence_output = sequence_output * (self.model_dim**-0.5)
|
|
|
|
lm_logits = self.lm_head(sequence_output)
|
|
|
|
loss = None
|
|
if labels is not None:
|
|
loss_fct = CrossEntropyLoss(ignore_index=-100)
|
|
loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1))
|
|
|
|
if not return_dict:
|
|
output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs
|
|
return ((loss,) + output) if loss is not None else output
|
|
|
|
return Seq2SeqLMOutput(
|
|
loss=loss,
|
|
logits=lm_logits,
|
|
past_key_values=decoder_outputs.past_key_values,
|
|
decoder_hidden_states=decoder_outputs.hidden_states,
|
|
decoder_attentions=decoder_outputs.attentions,
|
|
cross_attentions=decoder_outputs.cross_attentions,
|
|
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
|
|
encoder_hidden_states=encoder_outputs.hidden_states,
|
|
encoder_attentions=encoder_outputs.attentions,
|
|
)
|
|
|
|
@torch.no_grad()
|
|
def generate(
|
|
self,
|
|
input_features,
|
|
attention_mask=None,
|
|
composer="composer1",
|
|
generation_config=None,
|
|
**kwargs,
|
|
):
|
|
"""
|
|
Generates token ids for midi outputs.
|
|
|
|
<Tip warning={true}>
|
|
|
|
Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
|
|
model's default generation configuration. You can override any `generation_config` by passing the corresponding
|
|
parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`. For an overview of generation
|
|
strategies and code examples, check out the [following guide](./generation_strategies).
|
|
|
|
</Tip>
|
|
|
|
Parameters:
|
|
input_features (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
|
This is the featurized version of audio generated by `Pop2PianoFeatureExtractor`.
|
|
attention_mask:
|
|
For batched generation `input_features` are padded to have the same shape across all examples.
|
|
`attention_mask` helps to determine which areas were padded and which were not.
|
|
- 1 for tokens that are **not padded**,
|
|
- 0 for tokens that are **padded**.
|
|
composer (`str`, *optional*, defaults to `"composer1"`):
|
|
This value is passed to `Pop2PianoConcatEmbeddingToMel` to generate different embeddings for each
|
|
`"composer"`. Please make sure that the composet value is present in `composer_to_feature_token` in
|
|
`generation_config`. For an example please see
|
|
https://huggingface.co/sweetcocoa/pop2piano/blob/main/generation_config.json .
|
|
generation_config (`~generation.GenerationConfig`, *optional*):
|
|
The generation configuration to be used as base parametrization for the generation call. `**kwargs`
|
|
passed to generate matching the attributes of `generation_config` will override them. If
|
|
`generation_config` is not provided, the default will be used, which had the following loading
|
|
priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
|
|
configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
|
|
default values, whose documentation should be checked to parameterize generation.
|
|
kwargs:
|
|
Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
|
|
forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
|
|
specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.
|
|
Return:
|
|
[`~utils.ModelOutput`] or `torch.LongTensor`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True`
|
|
or when `config.return_dict_in_generate=True`) or a `torch.FloatTensor`.
|
|
Since Pop2Piano is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible
|
|
[`~utils.ModelOutput`] types are:
|
|
- [`~generation.GenerateEncoderDecoderOutput`],
|
|
- [`~generation.GenerateBeamEncoderDecoderOutput`]
|
|
"""
|
|
|
|
if generation_config is None:
|
|
generation_config = self.generation_config
|
|
generation_config.update(**kwargs)
|
|
|
|
# check for composer_to_feature_token
|
|
if not hasattr(generation_config, "composer_to_feature_token"):
|
|
raise ValueError(
|
|
"`composer_to_feature_token` was not found! Please refer to "
|
|
"https://huggingface.co/sweetcocoa/pop2piano/blob/main/generation_config.json"
|
|
"and parse a dict like that."
|
|
)
|
|
|
|
if len(generation_config.composer_to_feature_token) != self.config.composer_vocab_size:
|
|
raise ValueError(
|
|
"config.composer_vocab_size must be same as the number of keys in "
|
|
f"generation_config.composer_to_feature_token! "
|
|
f"Found {self.config.composer_vocab_size} vs {len(generation_config.composer_to_feature_token)}."
|
|
)
|
|
|
|
# to control the variation of generated MIDI tokens we concatenate mel-conditioner tokens(which depends on composer_token)
|
|
# at the front of input_features.
|
|
input_features, attention_mask = self.get_mel_conditioner_outputs(
|
|
input_features=input_features,
|
|
attention_mask=attention_mask,
|
|
composer=composer,
|
|
generation_config=generation_config,
|
|
)
|
|
|
|
return super().generate(
|
|
inputs=None,
|
|
inputs_embeds=input_features,
|
|
attention_mask=attention_mask,
|
|
generation_config=generation_config,
|
|
**kwargs,
|
|
)
|
|
|
|
def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
|
|
return self._shift_right(labels)
|
|
|
|
|
|
__all__ = ["Pop2PianoForConditionalGeneration", "Pop2PianoPreTrainedModel"]
|