audio-model / mae.py
Tharya's picture
Upload 3 files
9442c34 verified
from functools import partial
import torch
import torch.nn as nn
import numpy as np
import torch.utils.checkpoint
from timm.models.swin_transformer import SwinTransformerBlock
from timm.models.vision_transformer import Block
from timm.models.layers import to_2tuple
class PatchEmbed(nn.Module):
""" Image to Patch Embedding
"""
def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
self.patch_hw = (img_size[1] // patch_size[1], img_size[0] // patch_size[0])
self.img_size = img_size
self.patch_size = patch_size
self.num_patches = num_patches
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
def forward(self, x):
B, C, H, W = x.shape
x = self.proj(x).flatten(2).transpose(1, 2)
return x
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
"""
embed_dim: output dimension for each position
pos: a list of positions to be encoded: size (M,)
out: (M, D)
"""
assert embed_dim % 2 == 0
omega = np.arange(embed_dim // 2, dtype=float)
omega /= embed_dim / 2.
omega = 1. / 10000 ** omega # (D/2,)
pos = pos.reshape(-1) # (M,)
out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product
emb_sin = np.sin(out) # (M, D/2)
emb_cos = np.cos(out) # (M, D/2)
emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
return emb
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
assert embed_dim % 2 == 0
# use half of dimensions to encode grid_h
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
return emb
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
"""
grid_size: int of the grid height and width
return:
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
"""
grid_h = np.arange(grid_size, dtype=np.float32)
grid_w = np.arange(grid_size, dtype=np.float32)
grid = np.meshgrid(grid_w, grid_h) # here w goes first
grid = np.stack(grid, axis=0)
grid = grid.reshape([2, 1, grid_size, grid_size])
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
if cls_token:
pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
return pos_embed
def get_2d_sincos_pos_embed_flexible(embed_dim, grid_size, cls_token=False):
"""
grid_size: int of the grid height and width
return:
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
"""
grid_h = np.arange(grid_size[0], dtype=np.float32)
grid_w = np.arange(grid_size[1], dtype=np.float32)
grid = np.meshgrid(grid_w, grid_h) # here w goes first
grid = np.stack(grid, axis=0)
grid = grid.reshape([2, 1, grid_size[0], grid_size[1]])
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
if cls_token:
pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
return pos_embed
class SwinTransformerBlockWrapper(torch.nn.Module):
"""
Wrap SwinTransformerBlock to fit the input shape of [B, N, C] like TransformerBlock.
The SwinTransformerBlock takes the input shape of [B, H, W, C], and TransformerBlock
takes the input shape of [B, N, C].
"""
def __init__(self, block: SwinTransformerBlock):
super().__init__()
self.block = block
self.input_resolution = block.input_resolution
def forward(self, x):
"""
:param x: [B, N, C]
:return: [B, N, C]
"""
B, N, C = x.shape
x = x.reshape(B, *self.input_resolution, C)
x = self.block(x)
x = x.reshape(B, N, C)
return x
class MaskedAutoencoderViT(nn.Module):
""" Masked Autoencoder with VisionTransformer backbone
"""
def __init__(
self,
img_size=224,
patch_size=16,
in_chans=3, # input channels. 1 for audio, 3 for image
embed_dim=1024,
depth=24, # transformer depth
num_heads=16,
decoder_mode=0, # 0: transformer (global attn), 1: swin-transformer (swined local attn)
no_shift=False, # invalid when decoder_mode=0. swin-transformer. shift patch or not
decoder_embed_dim=512,
decoder_depth=8, # invalid when decoder_mode=1. It will be fixed to 16 when decoder_mode=1.
decoder_num_heads=16, # invalid when decoder_mode=1. It will be fixed to 16 when decoder_mode=1.
mlp_ratio=4., # hidden dimension / embed dimension in feedforward layer of transformer
norm_layer=nn.LayerNorm,
norm_pix_loss=False, # use (per-patch) normalized pixels as targets for computing loss
pos_trainable=False,
):
super().__init__()
self.img_size = to_2tuple(img_size)
self.embed_dim = embed_dim
self.decoder_embed_dim = decoder_embed_dim
# MAE encoder specifics
self.patch_embed = PatchEmbed(img_size, patch_size, in_chans, embed_dim)
num_patches = self.patch_embed.num_patches
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim),
requires_grad=pos_trainable) # fixed sin-cos embedding
self.encoder_depth = depth
self.blocks = nn.ModuleList([
Block(embed_dim, num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer) for _ in range(depth)])
self.norm = norm_layer(embed_dim)
# --------------------------------------------------------------------------
# MAE decoder specifics
self.decoder_embed = nn.Linear(embed_dim, decoder_embed_dim, bias=True)
self.mask_token = nn.Parameter(torch.zeros(1, 1, decoder_embed_dim))
self.decoder_pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, decoder_embed_dim),
requires_grad=pos_trainable) # fixed sin-cos embedding
self.no_shift = no_shift
self.decoder_mode = decoder_mode
window_size = (4, 4)
feat_size = (self.img_size[0] // patch_size, 8)
if self.decoder_mode == 1:
decoder_modules = []
for index in range(16):
if self.no_shift:
shift_size = (0, 0)
else:
if (index % 2) == 0:
shift_size = (0, 0)
else:
shift_size = (2, 0)
# shift_size = tuple([0 if ((index % 2) == 0) else w // 2 for w in window_size])
decoder_modules.append(
SwinTransformerBlockWrapper(
SwinTransformerBlock(
dim=decoder_embed_dim,
input_resolution=feat_size,
num_heads=16,
window_size=window_size,
shift_size=shift_size,
mlp_ratio=mlp_ratio,
proj_drop=0.0,
attn_drop=0.0,
drop_path=0.0,
norm_layer=norm_layer,
)
)
)
self.decoder_blocks = nn.ModuleList(decoder_modules)
else:
# Transformer
self.decoder_blocks = nn.ModuleList([
Block(decoder_embed_dim, decoder_num_heads, mlp_ratio, qkv_bias=True, norm_layer=norm_layer)
for _ in range(decoder_depth)])
self.decoder_norm = norm_layer(decoder_embed_dim)
self.decoder_pred = nn.Linear(decoder_embed_dim, patch_size ** 2 * in_chans, bias=True) # decoder to patch
self.norm_pix_loss = norm_pix_loss
self.patch_size = patch_size
self.initialize_weights()
def initialize_weights(self):
# initialize (and freeze) pos_embed by sin-cos embedding
pos_embed = get_2d_sincos_pos_embed_flexible(self.pos_embed.shape[-1], self.patch_embed.patch_hw,
cls_token=True)
self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
decoder_pos_embed = get_2d_sincos_pos_embed_flexible(self.decoder_pos_embed.shape[-1],
self.patch_embed.patch_hw, cls_token=True)
self.decoder_pos_embed.data.copy_(torch.from_numpy(decoder_pos_embed).float().unsqueeze(0))
# initialize patch_embed like nn.Linear (instead of nn.Conv2d)
w = self.patch_embed.proj.weight.data
torch.nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
# timm's trunc_normal_(std=.02) is effectively normal_(std=0.02) as cutoff is too big (2.)
torch.nn.init.normal_(self.cls_token, std=.02)
torch.nn.init.normal_(self.mask_token, std=.02)
# initialize nn.Linear and nn.LayerNorm
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
# we use xavier_uniform following official JAX ViT:
torch.nn.init.xavier_uniform_(m.weight)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def patchify(self, imgs):
"""
imgs: (N, 3, H, W)
x: (N, L, patch_size**2 *3)
L = (H/p)*(W/p)
"""
p = self.patch_embed.patch_size[0]
h = imgs.shape[2] // p
w = imgs.shape[3] // p
# h,w = self.patch_embed.patch_hw
x = imgs.reshape(shape=(imgs.shape[0], 1, h, p, w, p))
x = torch.einsum('nchpwq->nhwpqc', x)
x = x.reshape(imgs.shape[0], h * w, p ** 2 * 1)
return x
def unpatchify(self, x):
"""
x: (N, L, patch_size**2 *3)
specs: (N, 1, H, W)
"""
p = self.patch_embed.patch_size[0]
h = self.img_size[0] // p
w = 128 // p
x = x.reshape(shape=(x.shape[0], h, w, p, p, 1))
x = torch.einsum('nhwpqc->nchpwq', x)
specs = x.reshape(x.shape[0], 1, h * p, w * p)
return specs
def random_masking(self, x, mask_ratio):
"""
Perform per-sample random masking by per-sample shuffling.
Per-sample shuffling is done by argsort random noise.
x: [N, L, D], sequence
"""
N, L, D = x.shape # batch, length, dim
len_keep = int(L * (1 - mask_ratio))
noise = torch.rand(N, L, device=x.device) # noise in [0, 1]
# sort noise for each sample
ids_shuffle = torch.argsort(noise, dim=1) # ascend: small is keep, large is remove
ids_restore = torch.argsort(ids_shuffle, dim=1)
# keep the first subset
ids_keep = ids_shuffle[:, :len_keep]
x_masked = torch.gather(x, dim=1, index=ids_keep.unsqueeze(-1).repeat(1, 1, D))
# generate the binary mask: 0 is keep, 1 is remove
mask = torch.ones([N, L], device=x.device)
mask[:, :len_keep] = 0
# unshuffle to get the binary mask
mask = torch.gather(mask, dim=1, index=ids_restore)
return x_masked, mask, ids_restore
def forward_encoder(self, x, mask_ratio):
"""
:param x: [N, C, H, W]
:param mask_ratio: float. ratio of masked patches
:return: tuple. x: [N, L', D], mask: [N, L], ids_restore: [N, L], None
"""
# embed patches
x = self.patch_embed(x)
B, L, D = x.shape
# add pos embed w/o cls token
x = x + self.pos_embed[:, 1:L + 1, :]
# masking: length -> length * mask_ratio
x, mask, ids_restore = self.random_masking(x, mask_ratio)
# append cls token
cls_token = self.cls_token + self.pos_embed[:, :1, :]
cls_tokens = cls_token.expand(x.shape[0], -1, -1)
x = torch.cat((cls_tokens, x), dim=1)
# apply Transformer blocks
for blk in self.blocks:
x = blk(x)
x = self.norm(x)
return x, mask, ids_restore
def forward_encoder_no_mask(
self,
x,
header='mean'
):
"""
:param x: [N, C, H, W]
:param header: str. 'cls' or 'mean'
:param key_padding_mask: [N, L], 0 is keep, 1 is remove
:return: contextual_emb: [N, L, D]
"""
# embed patches
x = self.patch_embed(x)
B, L, D = x.shape
# add pos embed w/o cls token
x = x + self.pos_embed[:, 1:L + 1, :]
# append cls token
cls_token = self.cls_token + self.pos_embed[:, :1, :]
cls_tokens = cls_token.expand(x.shape[0], -1, -1)
x = torch.cat((cls_tokens, x), dim=1)
# apply Transformer blocks
for n, blk in enumerate(self.blocks):
x = blk(x)
x = self.norm(x)
if header == 'cls':
emb = x[:, 0, :]
elif header == 'mean':
emb = x[:, 1:, :].mean(dim=1)
else:
raise NotImplementedError
return emb
def forward_decoder(self, x, ids_restore):
"""
:param x: [N, L, D]
:param ids_restore: [N, L]
:return: pred: [N, L, p*p*3], None, None
"""
# embed tokens
x = self.decoder_embed(x) # [N, L, D] -> [N, L, D']
# append mask tokens to sequence
mask_tokens = self.mask_token.repeat(x.shape[0], ids_restore.shape[1] + 1 - x.shape[1], 1)
x_ = torch.cat([x[:, 1:, :], mask_tokens], dim=1) # no cls token
x_ = torch.gather(x_, dim=1, index=ids_restore.unsqueeze(-1).repeat(1, 1, x.shape[2])) # unshuffle
x = torch.cat([x[:, :1, :], x_], dim=1) # append cls token
B, L, D = x.shape
# add pos embed
x = x + self.decoder_pos_embed[:, :L, :]
if self.decoder_mode != 0:
B, L, D = x.shape
x = x[:, 1:, :]
if self.decoder_mode > 3: # mvit
x = self.decoder_blocks(x)
else:
# apply Transformer blocks
for blk in self.decoder_blocks:
x = blk(x)
x = self.decoder_norm(x)
# predictor projection
pred = self.decoder_pred(x)
# remove cls token
if self.decoder_mode == 0:
pred = pred[:, 1:, :]
return pred
def forward_loss(self, imgs, pred, mask, norm_pix_loss=False):
"""
imgs: [N, 3, H, W]
pred: [N, L, p*p*3]
mask: [N, L], 0 is keep, 1 is remove,
"""
target = self.patchify(imgs)
if norm_pix_loss:
mean = target.mean(dim=-1, keepdim=True)
var = target.var(dim=-1, keepdim=True)
target = (target - mean) / (var + 1.e-6) ** .5
loss = (pred - target) ** 2
loss = loss.mean(dim=-1) # [N, L], mean loss per patch
loss = (loss * mask).sum() / mask.sum() # mean loss on removed patches
return loss
def forward(self, imgs, mask_ratio=0.8):
"""
:param imgs: [N, C, H, W]
:param mask_ratio: float. ratio of masked patches
:return: tuple. loss_recon: float, pred: [N, L, p*p*3], mask: [N, L], None
"""
emb_enc, mask, ids_restore = self.forward_encoder(imgs, mask_ratio)
pred = self.forward_decoder(emb_enc, ids_restore) # [N, L, p*p*3]
loss_recon = self.forward_loss(imgs, pred, mask, norm_pix_loss=self.norm_pix_loss)
return loss_recon, pred, mask
if __name__ == '__main__':
device = 'cpu'
# device = 'cuda'
# Model
audio_mae = MaskedAutoencoderViT(
img_size=(2048, 128),
patch_size=16,
in_chans=1,
embed_dim=768,
depth=12,
num_heads=12,
decoder_mode=1,
no_shift=False,
decoder_embed_dim=512,
norm_layer=partial(nn.LayerNorm, eps=1e-6),
norm_pix_loss=False,
pos_trainable=False,
)
# Load pre-trained weights
ckpt_path = 'music-mae-32kHz.pth'
audio_mae.load_state_dict(torch.load(ckpt_path, map_location='cpu'))
audio_mae.to(device)
# Generate a batch of random inputs: (N, C, H, W), N=4 (batch size), C=1 (channel), H=2048, W=128
# Each input is a mel spectrogram with shape (2048, 128)
x = torch.randn(4, 1, 2048, 128).to(device)
# Compute the representation of the input batch
emb = audio_mae.forward_encoder_no_mask(x, header='mean') # torch.Size([4, 768])