Masond / jukebox /prior /prior.py
MasonCrinr's picture
Upload 331 files
814c8cf
import numpy as np
import torch as t
import torch.nn as nn
import jukebox.utils.dist_adapter as dist
from jukebox.transformer.ops import LayerNorm
from jukebox.prior.autoregressive import ConditionalAutoregressive2D
from jukebox.prior.conditioners import Conditioner, LabelConditioner
from jukebox.data.labels import EmptyLabeller, Labeller
from jukebox.utils.torch_utils import assert_shape
from jukebox.utils.dist_utils import print_once
from jukebox.vqvae.vqvae import calculate_strides
"""
Model the prior on vq codes conditioned on timing, artist, genre, lyrics and codes from levels above.
To condition on the timing, genre and artist, we use the LabelConditioner class
To condition on the codes from the level above, we use the Conditioner class
To condition on lyrics, we allow two types of priors:
- Separate Encoder Decoder: This is the usual encoder-decoder style transformer. The encoder transformer autoregressively
models the lyrics, and we use its last layer to produce keys/values that are attened to by the decoder transformer
- Single Encoder Decoder: This is a simplification where we combine them into a single model. We merge the text vocab
and VQ vocab into a single large vocab, and the lyric tokens and VQ tokens into a single longer sequence of tokens which
we autoregressively model together.
"""
class SimplePrior(nn.Module):
def __init__(self, z_shapes, l_bins, encoder, decoder, level,
downs_t, strides_t, labels, prior_kwargs, x_cond_kwargs, y_cond_kwargs,
prime_kwargs, copy_input, labels_v3=False,
merged_decoder=False, single_enc_dec=False):
super().__init__()
self.use_tokens = prime_kwargs.pop('use_tokens')
self.n_tokens = prime_kwargs.pop('n_tokens')
self.prime_loss_fraction = prime_kwargs.pop('prime_loss_fraction')
self.copy_input = copy_input
if self.copy_input:
prime_kwargs['bins'] = l_bins
self.z_shapes = z_shapes
self.levels = len(self.z_shapes)
self.z_shape = self.z_shapes[level]
self.level = level
assert level < self.levels, f"Total levels {self.levels}, got level {level}"
self.l_bins = l_bins
# Passing functions instead of the vqvae module to avoid getting params
self.encoder = encoder
self.decoder = decoder
# X conditioning
self.x_cond = (level != (self.levels - 1))
self.cond_level = level + 1
# Y conditioning
self.y_cond = labels
self.single_enc_dec = single_enc_dec
# X conditioning
if self.x_cond:
self.conditioner_blocks = nn.ModuleList()
conditioner_block = lambda _level: Conditioner(input_shape=z_shapes[_level],
bins=l_bins,
down_t=downs_t[_level],
stride_t=strides_t[_level],
**x_cond_kwargs)
if dist.get_rank() == 0: print(f"Conditioning on 1 above level(s)")
self.conditioner_blocks.append(conditioner_block(self.cond_level))
# Y conditioning
if self.y_cond:
self.n_time = self.z_shape[0] # Assuming STFT=TF order and raw=T1 order, so T is first dim
self.y_emb = LabelConditioner(n_time=self.n_time,include_time_signal=not self.x_cond,**y_cond_kwargs)
# Lyric conditioning
if single_enc_dec:
# Single encoder-decoder transformer
self.prior_shapes = [(self.n_tokens,), prior_kwargs.pop('input_shape')]
self.prior_bins = [prime_kwargs['bins'], prior_kwargs.pop('bins')]
self.prior_dims = [np.prod(shape) for shape in self.prior_shapes]
self.prior_bins_shift = np.cumsum([0, *self.prior_bins])[:-1]
self.prior_width = prior_kwargs['width']
print_once(f'Creating cond. autoregress with prior bins {self.prior_bins}, ')
print_once(f'dims {self.prior_dims}, ')
print_once(f'shift {self.prior_bins_shift}')
print_once(f'input shape {sum(self.prior_dims)}')
print_once(f'input bins {sum(self.prior_bins)}')
print_once(f'Self copy is {self.copy_input}')
self.prime_loss_dims, self.gen_loss_dims = self.prior_dims[0], self.prior_dims[1]
self.total_loss_dims = self.prime_loss_dims + self.gen_loss_dims
self.prior = ConditionalAutoregressive2D(input_shape=(sum(self.prior_dims),),
bins=sum(self.prior_bins),
x_cond=(self.x_cond or self.y_cond), y_cond=True,
prime_len=self.prime_loss_dims,
**prior_kwargs)
else:
# Separate encoder-decoder transformer
if self.n_tokens != 0 and self.use_tokens:
from jukebox.transformer.ops import Conv1D
prime_input_shape = (self.n_tokens,)
self.prime_loss_dims = np.prod(prime_input_shape)
self.prime_acts_width, self.prime_state_width = prime_kwargs['width'], prior_kwargs['width']
self.prime_prior = ConditionalAutoregressive2D(input_shape=prime_input_shape, x_cond=False, y_cond=False,
only_encode=True,
**prime_kwargs)
self.prime_state_proj = Conv1D(self.prime_acts_width, self.prime_state_width, init_scale=prime_kwargs['init_scale'])
self.prime_state_ln = LayerNorm(self.prime_state_width)
self.prime_bins = prime_kwargs['bins']
self.prime_x_out = nn.Linear(self.prime_state_width, self.prime_bins, bias=False)
nn.init.normal_(self.prime_x_out.weight, std=0.02 * prior_kwargs['init_scale'])
else:
self.prime_loss_dims = 0
self.gen_loss_dims = np.prod(self.z_shape)
self.total_loss_dims = self.prime_loss_dims + self.gen_loss_dims
self.prior = ConditionalAutoregressive2D(x_cond=(self.x_cond or self.y_cond), y_cond=self.y_cond,
encoder_dims = self.prime_loss_dims, merged_decoder=merged_decoder,
**prior_kwargs)
self.n_ctx = self.gen_loss_dims
self.downsamples = calculate_strides(strides_t, downs_t)
self.cond_downsample = self.downsamples[level+1] if level != self.levels - 1 else None
self.raw_to_tokens = np.prod(self.downsamples[:level+1])
self.sample_length = self.n_ctx*self.raw_to_tokens
if labels:
self.labels_v3 = labels_v3
self.labeller = Labeller(self.y_emb.max_bow_genre_size, self.n_tokens, self.sample_length, v3=self.labels_v3)
else:
self.labeller = EmptyLabeller()
print(f"Level:{level}, Cond downsample:{self.cond_downsample}, Raw to tokens:{self.raw_to_tokens}, Sample length:{self.sample_length}")
def get_y(self, labels, start, get_indices=False):
if isinstance(self.labeller, EmptyLabeller):
return None
y = labels['y'].clone()
# Set sample_length to match this level
y[:, 2] = int(self.sample_length)
# Set offset
y[:, 1:2] = y[:, 1:2] + int(start * self.raw_to_tokens)
# Set lyric tokens
indices = self.labeller.set_y_lyric_tokens(y, labels)
if get_indices:
return y, indices
else:
return y
def get_z_conds(self, zs, start, end):
if self.level != self.levels - 1:
assert start % self.cond_downsample == end % self.cond_downsample == 0
z_cond = zs[self.level + 1][:,start//self.cond_downsample:end//self.cond_downsample]
assert z_cond.shape[1] == self.n_ctx//self.cond_downsample
z_conds = [z_cond]
else:
z_conds = None
return z_conds
def prior_preprocess(self, xs, conds):
N = xs[0].shape[0]
for i in range(len(xs)):
x, shape, dims = xs[i], self.prior_shapes[i], self.prior_dims[i]
bins, bins_shift = int(self.prior_bins[i]), int(self.prior_bins_shift[i])
assert isinstance(x, t.cuda.LongTensor), x
assert (0 <= x).all() and (x < bins).all()
#assert_shape(x, (N, *shape))
xs[i] = (xs[i] + bins_shift).view(N, -1)
for i in range(len(conds)):
cond, shape, dims = conds[i], self.prior_shapes[i], self.prior_dims[i]
if cond is not None:
assert_shape(cond, (N, dims, self.prior_width))
else:
conds[i] = t.zeros((N, dims, self.prior_width), dtype=t.float, device='cuda')
return t.cat(xs, dim=1), t.cat(conds, dim=1)
def prior_postprocess(self, z):
N = z.shape[0]
dims = (self.prior_dims[0], z.shape[1] - self.prior_dims[0])
# xs = list(t.split(z, self.prior_dims, dim=1))
xs = list(t.split(z, dims, dim=1))
for i in range(len(xs)):
# x, shape, dims, bins, bins_shift = xs[i], self.prior_shapes[i], self.prior_dims[i], self.prior_bins[i], self.prior_bins_shift[i]
# assert_shape(x, (N, dims))
shape = self.prior_shapes[i]
bins, bins_shift = int(self.prior_bins[i]), int(self.prior_bins_shift[i])
# xs[i] = (xs[i] - bins_shift).view(N, *shape) #view(N, -1, *shape[1:])
xs[i] = (xs[i] - bins_shift).view(N, -1, *shape[1:])
xs[i] = t.clamp(xs[i], min=0) # If not masking loss, model may have generated lyric/midi tokens which are now shifted <0 by bin_shift
assert (xs[i] < bins).all(), f'rank: {dist.get_rank()}, bins: {bins}, dims {dims}, shape {shape}, prior_shape {self.prior_shapes}, bins_shift {bins_shift}, xs[i]: {xs[i]}'
return xs[-1]
def x_emb(self, z_conds):
z_conds = z_conds[:self.cond_level - self.level]
assert len(z_conds) == len(self.conditioner_blocks) == self.cond_level - self.level, f"Expected {len(z_conds)} == {len(self.conditioner_blocks)} == {self.cond_level} - {self.level}"
x_cond = None
for z_cond, conditioner_block in reversed(list(zip(z_conds, self.conditioner_blocks))):
x_cond = conditioner_block(z_cond, x_cond)
return x_cond
def encode(self, x, start_level=None, end_level=None, bs_chunks=1):
if start_level == None:
start_level = self.level
if end_level == None:
end_level = self.levels
# Get latents
with t.no_grad():
zs = self.encoder(x, start_level=start_level, end_level=end_level, bs_chunks=bs_chunks)
return zs
def decode(self, zs, start_level=None, end_level=None, bs_chunks=1):
if start_level == None:
start_level = self.level
if end_level == None:
end_level = self.levels
assert len(zs) == end_level - start_level
with t.no_grad():
x_out = self.decoder(zs, start_level=start_level, end_level=end_level, bs_chunks=bs_chunks)
return x_out
def get_cond(self, z_conds, y):
if y is not None:
assert y.shape[1] == 4 + self.y_emb.max_bow_genre_size + self.n_tokens, f"Expected {4} + {self.y_emb.max_bow_genre_size} + {self.n_tokens}, got {y.shape[1]}"
n_labels = y.shape[1] - self.n_tokens
y, prime = y[:,:n_labels], y[:,n_labels:]
else:
y, prime = None, None
y_cond, y_pos = self.y_emb(y) if self.y_cond else (None, None)
x_cond = self.x_emb(z_conds) if self.x_cond else y_pos
return x_cond, y_cond, prime
def sample(self, n_samples, z=None, z_conds=None, y=None, fp16=False, temp=1.0, top_k=0, top_p=0.0,
chunk_size=None, sample_tokens=None):
N = n_samples
if z is not None: assert z.shape[0] == N, f"Expected shape ({N},**), got shape {z.shape}"
if y is not None: assert y.shape[0] == N, f"Expected shape ({N},**), got shape {y.shape}"
if z_conds is not None:
for z_cond in z_conds:
assert z_cond.shape[0] == N, f"Expected shape ({N},**), got shape {z_cond.shape}"
no_past_context = (z is None or z.shape[1] == 0)
if dist.get_rank() == 0:
name = {True: 'Ancestral', False: 'Primed'}[no_past_context]
print(f"{name} sampling {n_samples} samples with temp={temp}, top_k={top_k}, top_p={top_p}")
with t.no_grad():
# Currently x_cond only uses immediately above layer
x_cond, y_cond, prime = self.get_cond(z_conds, y)
if self.single_enc_dec:
# assert chunk_size % self.prime_loss_dims == 0. TODO: Check if needed
if no_past_context:
z, x_cond = self.prior_preprocess([prime], [None, x_cond])
else:
z, x_cond = self.prior_preprocess([prime, z], [None, x_cond])
if sample_tokens is not None:
sample_tokens += self.n_tokens
z = self.prior.primed_sample(n_samples, z, x_cond, y_cond, fp16=fp16, temp=temp,
top_k=top_k, top_p=top_p, chunk_size=chunk_size, sample_tokens=sample_tokens)
z = self.prior_postprocess(z)
else:
encoder_kv = self.get_encoder_kv(prime, fp16=fp16, sample=True)
if no_past_context:
z = self.prior.sample(n_samples, x_cond, y_cond, encoder_kv, fp16=fp16, temp=temp, top_k=top_k,
top_p=top_p, sample_tokens=sample_tokens)
else:
z = self.prior.primed_sample(n_samples, z, x_cond, y_cond, encoder_kv, fp16=fp16, temp=temp,
top_k=top_k, top_p=top_p, chunk_size=chunk_size, sample_tokens=sample_tokens)
if sample_tokens is None:
assert_shape(z, (N, *self.z_shape))
return z
def get_encoder_kv(self, prime, fp16=False, sample=False):
if self.n_tokens != 0 and self.use_tokens:
if sample:
self.prime_prior.cuda()
N = prime.shape[0]
prime_acts = self.prime_prior(prime, None, None, None, fp16=fp16)
assert_shape(prime_acts, (N, self.prime_loss_dims, self.prime_acts_width))
assert prime_acts.dtype == t.float, f'Expected t.float, got {prime_acts.dtype}'
encoder_kv = self.prime_state_ln(self.prime_state_proj(prime_acts))
assert encoder_kv.dtype == t.float, f'Expected t.float, got {encoder_kv.dtype}'
if sample:
self.prime_prior.cpu()
if fp16:
encoder_kv = encoder_kv.half()
else:
encoder_kv = None
return encoder_kv
def get_prime_loss(self, encoder_kv, prime_t):
if self.use_tokens:
encoder_kv = encoder_kv.float()
encoder_kv = self.prime_x_out(encoder_kv)
prime_loss = nn.functional.cross_entropy(encoder_kv.view(-1, self.prime_bins), prime_t.view(-1)) / np.log(2.)
else:
prime_loss = t.tensor(0.0, device='cuda')
return prime_loss
def z_forward(self, z, z_conds=[], y=None, fp16=False, get_preds=False, get_attn_weights=False):
"""
Arguments:
get_attn_weights (bool or set): Makes forward prop dump
self-attention softmaxes to self.prior.transformer.ws. Either a
set of layer indices indicating which layers to store, or a
boolean value indicating whether to dump all.
"""
assert isinstance(get_attn_weights, (bool, set))
if get_attn_weights:
self.prior.transformer.set_record_attn(get_attn_weights)
x_cond, y_cond, prime = self.get_cond(z_conds, y)
if self.copy_input:
prime = z[:,:self.n_tokens]
if self.single_enc_dec:
z, x_cond = self.prior_preprocess([prime, z], [None, x_cond])
(prime_loss, gen_loss), preds = self.prior(z, x_cond, y_cond, fp16=fp16, get_sep_loss=True, get_preds=get_preds)
else:
encoder_kv = self.get_encoder_kv(prime, fp16=fp16)
prime_loss = self.get_prime_loss(encoder_kv, prime)
gen_loss, preds = self.prior(z, x_cond, y_cond, encoder_kv, fp16=fp16, get_preds=get_preds)
loss = (self.prime_loss_fraction*prime_loss*self.prime_loss_dims/self.total_loss_dims) + \
(gen_loss*self.gen_loss_dims/self.total_loss_dims)
metrics=dict(bpd=gen_loss.clone().detach(), prime_loss=prime_loss.clone().detach(),
gen_loss=gen_loss.clone().detach())
if get_preds:
metrics["preds"] = preds.clone().detach()
if get_attn_weights:
ws = self.prior.transformer.ws
self.prior.transformer.set_record_attn(False)
return ws
else:
return loss, metrics
def forward(self, x, y=None, fp16=False, decode=False, get_preds=False):
bs = x.shape[0]
z, *z_conds = self.encode(x, bs_chunks=bs)
loss, metrics = self.z_forward(z=z, z_conds=z_conds, y=y, fp16=fp16, get_preds=get_preds)
if decode:
x_out = self.decode([z, *z_conds])
else:
x_out = None
return x_out, loss, metrics