mup work checkin

codes/models/audio/music/transformer_diffusion8_mup.py

@@ -0,0 +1,388 @@
+import itertools
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils.data import DataLoader
+
+from models.arch_util import ResBlock
+from models.audio.music.music_quantizer2 import MusicQuantizer2
+from models.diffusion.nn import timestep_embedding, normalization, zero_module, conv_nd, linear
+from models.diffusion.unet_diffusion import TimestepBlock
+from models.lucidrains.x_transformers import Encoder, Attention, FeedForward, RMSScaleShiftNorm, RotaryEmbedding
+from trainer.networks import register_model
+from utils.util import checkpoint, print_network
+
+
+def is_latent(t):
+    return t.dtype == torch.float
+
+def is_sequence(t):
+    return t.dtype == torch.long
+
+
+class MultiGroupEmbedding(nn.Module):
+    def __init__(self, tokens, groups, dim):
+        super().__init__()
+        self.m = nn.ModuleList([nn.Embedding(tokens, dim // groups) for _ in range(groups)])
+
+    def forward(self, x):
+        h = [embedding(x[:, :, i]) for i, embedding in enumerate(self.m)]
+        return torch.cat(h, dim=-1)
+
+
+class TimestepRotaryEmbedSequential(nn.Sequential, TimestepBlock):
+    def forward(self, x, emb, rotary_emb):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb, rotary_emb)
+            else:
+                x = layer(x, rotary_emb)
+        return x
+
+
+class DietAttentionBlock(TimestepBlock):
+    def __init__(self, in_dim, dim, heads, dropout):
+        super().__init__()
+        self.rms_scale_norm = RMSScaleShiftNorm(in_dim)
+        self.proj = nn.Linear(in_dim, dim)
+        self.attn = Attention(dim, heads=heads, causal=False, dropout=dropout)
+        self.ff = FeedForward(dim, in_dim, mult=1, dropout=dropout, zero_init_output=True)
+
+    def forward(self, x, timestep_emb, rotary_emb):
+        h = self.rms_scale_norm(x, norm_scale_shift_inp=timestep_emb)
+        h = self.proj(h)
+        h, _, _, _ = checkpoint(self.attn, h, None, None, None, None, None, rotary_emb)
+        h = checkpoint(self.ff, h)
+        return h + x
+
+
+class TransformerDiffusion(nn.Module):
+    """
+    A diffusion model composed entirely of stacks of transformer layers. Why would you do it any other way?
+    """
+    def __init__(
+            self,
+            prenet_channels=256,
+            prenet_layers=3,
+            model_channels=512,
+            block_channels=256,
+            num_layers=8,
+            in_channels=256,
+            rotary_emb_dim=32,
+            input_vec_dim=512,
+            out_channels=512,  # mean and variance
+            dropout=0,
+            use_fp16=False,
+            ar_prior=False,
+            # Parameters for regularization.
+            unconditioned_percentage=.1,  # This implements a mechanism similar to what is used in classifier-free training.
+            # mUp base shapes.
+            mup_base_shapes=None,
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.prenet_channels = prenet_channels
+        self.out_channels = out_channels
+        self.dropout = dropout
+        self.unconditioned_percentage = unconditioned_percentage
+        self.enable_fp16 = use_fp16
+
+        self.inp_block = conv_nd(1, in_channels, prenet_channels, 3, 1, 1)
+
+        self.time_embed = nn.Sequential(
+            linear(prenet_channels, prenet_channels),
+            nn.SiLU(),
+            linear(prenet_channels, model_channels),
+        )
+        prenet_heads = min(16, prenet_channels//64)
+
+        self.ar_prior = ar_prior
+        if ar_prior:
+            self.ar_input = nn.Linear(input_vec_dim, prenet_channels)
+            self.ar_prior_intg = Encoder(
+                    dim=prenet_channels,
+                    depth=prenet_layers,
+                    heads=prenet_heads,
+                    ff_dropout=dropout,
+                    attn_dropout=dropout,
+                    use_rmsnorm=True,
+                    ff_glu=True,
+                    rotary_pos_emb=True,
+                    zero_init_branch_output=True,
+                    ff_mult=1,
+                )
+        else:
+            self.input_converter = nn.Linear(input_vec_dim, prenet_channels)
+            self.code_converter = Encoder(
+                        dim=prenet_channels,
+                        depth=prenet_layers,
+                        heads=prenet_heads,
+                        ff_dropout=dropout,
+                        attn_dropout=dropout,
+                        use_rmsnorm=True,
+                        ff_glu=True,
+                        rotary_pos_emb=True,
+                        zero_init_branch_output=True,
+                        ff_mult=1,
+                    )
+
+        self.unconditioned_embedding = nn.Parameter(torch.randn(1,1,prenet_channels))
+        self.rotary_embeddings = RotaryEmbedding(rotary_emb_dim)
+        self.intg = nn.Linear(prenet_channels*2, model_channels)
+        self.layers = TimestepRotaryEmbedSequential(*[DietAttentionBlock(model_channels, block_channels,
+                                                                         min(16, block_channels//64), dropout) for _ in range(num_layers)])
+
+        self.out = nn.Sequential(
+            normalization(model_channels),
+            nn.SiLU(),
+            zero_module(conv_nd(1, model_channels, out_channels, 3, padding=1)),
+        )
+
+        if mup_base_shapes is not None:
+            from mup import set_base_shapes
+            set_base_shapes(self, mup_base_shapes, rescale_params=False)
+
+        self.debug_codes = {}
+
+    def get_grad_norm_parameter_groups(self):
+        groups = {
+            'layers': list(self.layers.parameters()) + list(self.inp_block.parameters()),
+            'code_converters': list(self.input_converter.parameters()) + list(self.code_converter.parameters()),
+            'time_embed': list(self.time_embed.parameters()),
+        }
+        return groups
+
+    def timestep_independent(self, prior, expected_seq_len):
+        code_emb = self.ar_input(prior) if self.ar_prior else self.input_converter(prior)
+
+        # Mask out the conditioning branch for whole batch elements, implementing something similar to classifier-free guidance.
+        if self.training and self.unconditioned_percentage > 0:
+            unconditioned_batches = torch.rand((code_emb.shape[0], 1, 1),
+                                               device=code_emb.device) < self.unconditioned_percentage
+            code_emb = torch.where(unconditioned_batches, self.unconditioned_embedding.repeat(prior.shape[0], 1, 1),
+                                   code_emb)
+        code_emb = self.ar_prior_intg(code_emb) if self.ar_prior else self.code_converter(code_emb)
+
+        expanded_code_emb = F.interpolate(code_emb.permute(0,2,1), size=expected_seq_len, mode='nearest').permute(0,2,1)
+        return expanded_code_emb
+
+    def forward(self, x, timesteps, codes=None, conditioning_input=None, precomputed_code_embeddings=None, conditioning_free=False):
+        if precomputed_code_embeddings is not None:
+            assert codes is None and conditioning_input is None, "Do not provide precomputed embeddings and the other parameters. It is unclear what you want me to do here."
+
+        unused_params = []
+        if conditioning_free:
+            code_emb = self.unconditioned_embedding.repeat(x.shape[0], x.shape[-1], 1)
+        else:
+            if precomputed_code_embeddings is not None:
+                code_emb = precomputed_code_embeddings
+            else:
+                code_emb = self.timestep_independent(codes, x.shape[-1])
+            unused_params.append(self.unconditioned_embedding)
+
+        blk_emb = self.time_embed(timestep_embedding(timesteps, self.prenet_channels))
+        x = self.inp_block(x).permute(0,2,1)
+
+        rotary_pos_emb = self.rotary_embeddings(x.shape[1], x.device)
+        x = self.intg(torch.cat([x, code_emb], dim=-1))
+        for layer in self.layers:
+            x = checkpoint(layer, x, blk_emb, rotary_pos_emb)
+
+        x = x.float().permute(0,2,1)
+        out = self.out(x)
+
+        # Involve probabilistic or possibly unused parameters in loss so we don't get DDP errors.
+        extraneous_addition = 0
+        for p in unused_params:
+            extraneous_addition = extraneous_addition + p.mean()
+        out = out + extraneous_addition * 0
+
+        return out
+
+
+class TransformerDiffusionWithQuantizer(nn.Module):
+    def __init__(self, freeze_quantizer_until=20000, **kwargs):
+        super().__init__()
+
+        self.internal_step = 0
+        self.freeze_quantizer_until = freeze_quantizer_until
+        self.diff = TransformerDiffusion(**kwargs)
+        self.quantizer = MusicQuantizer2(inp_channels=256, inner_dim=[1024], codevector_dim=1024, codebook_size=256,
+                                        codebook_groups=2, max_gumbel_temperature=4, min_gumbel_temperature=.5)
+        self.quantizer.quantizer.temperature = self.quantizer.min_gumbel_temperature
+        del self.quantizer.up
+
+    def update_for_step(self, step, *args):
+        self.internal_step = step
+        qstep = max(0, self.internal_step - self.freeze_quantizer_until)
+        self.quantizer.quantizer.temperature = max(
+            self.quantizer.max_gumbel_temperature * self.quantizer.gumbel_temperature_decay ** qstep,
+                    self.quantizer.min_gumbel_temperature,
+                )
+
+    def forward(self, x, timesteps, truth_mel, conditioning_input, disable_diversity=False, conditioning_free=False):
+        quant_grad_enabled = self.internal_step > self.freeze_quantizer_until
+        with torch.set_grad_enabled(quant_grad_enabled):
+            proj, diversity_loss = self.quantizer(truth_mel, return_decoder_latent=True)
+            proj = proj.permute(0,2,1)
+
+        # Make sure this does not cause issues in DDP by explicitly using the parameters for nothing.
+        if not quant_grad_enabled:
+            unused = 0
+            for p in self.quantizer.parameters():
+                unused = unused + p.mean() * 0
+            proj = proj + unused
+            diversity_loss = diversity_loss * 0
+
+        diff = self.diff(x, timesteps, codes=proj, conditioning_input=conditioning_input, conditioning_free=conditioning_free)
+        if disable_diversity:
+            return diff
+        return diff, diversity_loss
+
+    def get_debug_values(self, step, __):
+        if self.quantizer.total_codes > 0:
+            return {'histogram_codes': self.quantizer.codes[:self.quantizer.total_codes]}
+        else:
+            return {}
+
+    def get_grad_norm_parameter_groups(self):
+        groups = {
+            'attention_layers': list(itertools.chain.from_iterable([lyr.attn.parameters() for lyr in self.diff.layers])),
+            'ff_layers': list(itertools.chain.from_iterable([lyr.ff.parameters() for lyr in self.diff.layers])),
+            'quantizer_encoder': list(self.quantizer.encoder.parameters()),
+            'quant_codebook': [self.quantizer.quantizer.codevectors],
+            'rotary_embeddings': list(self.diff.rotary_embeddings.parameters()),
+            'out': list(self.diff.out.parameters()),
+            'x_proj': list(self.diff.inp_block.parameters()),
+            'layers': list(self.diff.layers.parameters()),
+            'code_converters': list(self.diff.input_converter.parameters()) + list(self.diff.code_converter.parameters()),
+            'time_embed': list(self.diff.time_embed.parameters()),
+        }
+        return groups
+
+
+class TransformerDiffusionWithARPrior(nn.Module):
+    def __init__(self, freeze_diff=False, **kwargs):
+        super().__init__()
+
+        self.internal_step = 0
+        from models.audio.music.gpt_music import GptMusicLower
+        self.ar = GptMusicLower(dim=512, layers=12)
+        for p in self.ar.parameters():
+            p.DO_NOT_TRAIN = True
+            p.requires_grad = False
+
+        self.diff = TransformerDiffusion(ar_prior=True, **kwargs)
+        if freeze_diff:
+            for p in self.diff.parameters():
+                p.DO_NOT_TRAIN = True
+                p.requires_grad = False
+            for p in list(self.diff.ar_prior_intg.parameters()) + list(self.diff.ar_input.parameters()):
+                del p.DO_NOT_TRAIN
+                p.requires_grad = True
+
+    def get_grad_norm_parameter_groups(self):
+        groups = {
+            'attention_layers': list(itertools.chain.from_iterable([lyr.attn.parameters() for lyr in self.diff.layers])),
+            'ff_layers': list(itertools.chain.from_iterable([lyr.ff.parameters() for lyr in self.diff.layers])),
+            'rotary_embeddings': list(self.diff.rotary_embeddings.parameters()),
+            'out': list(self.diff.out.parameters()),
+            'x_proj': list(self.diff.inp_block.parameters()),
+            'layers': list(self.diff.layers.parameters()),
+            'ar_prior_intg': list(self.diff.ar_prior_intg.parameters()),
+            'time_embed': list(self.diff.time_embed.parameters()),
+        }
+        return groups
+
+    def forward(self, x, timesteps, truth_mel, disable_diversity=False, conditioning_input=None, conditioning_free=False):
+        with torch.no_grad():
+            prior = self.ar(truth_mel, conditioning_input, return_latent=True)
+
+        diff = self.diff(x, timesteps, prior, conditioning_free=conditioning_free)
+        return diff
+
+
+@register_model
+def register_transformer_diffusion8_mup(opt_net, opt):
+    return TransformerDiffusion(**opt_net['kwargs'])
+
+
+@register_model
+def register_transformer_diffusion8_with_quantizer_mup(opt_net, opt):
+    return TransformerDiffusionWithQuantizer(**opt_net['kwargs'])
+
+
+@register_model
+def register_transformer_diffusion8_with_ar_prior_mup(opt_net, opt):
+    return TransformerDiffusionWithARPrior(**opt_net['kwargs'])
+
+
+def test_quant_model():
+    clip = torch.randn(2, 256, 400)
+    cond = torch.randn(2, 256, 400)
+    ts = torch.LongTensor([600, 600])
+    model = TransformerDiffusionWithQuantizer(model_channels=2048, block_channels=1024, prenet_channels=1024,
+                                              input_vec_dim=1024, num_layers=16, prenet_layers=6)
+    model.get_grad_norm_parameter_groups()
+
+    quant_weights = torch.load('D:\\dlas\\experiments\\train_music_quant_r4\\models\\5000_generator.pth')
+    #diff_weights = torch.load('X:\\dlas\\experiments\\train_music_diffusion_tfd5\\models\\48000_generator_ema.pth')
+    model.quantizer.load_state_dict(quant_weights, strict=False)
+    #model.diff.load_state_dict(diff_weights)
+
+    torch.save(model.state_dict(), 'sample.pth')
+    print_network(model)
+    o = model(clip, ts, clip, cond)
+
+
+def test_ar_model():
+    clip = torch.randn(2, 256, 400)
+    cond = torch.randn(2, 256, 400)
+    ts = torch.LongTensor([600, 600])
+    model = TransformerDiffusionWithARPrior(model_channels=2048, block_channels=1024, prenet_channels=1024,
+                                              input_vec_dim=512, num_layers=16, prenet_layers=6, freeze_diff=True)
+    model.get_grad_norm_parameter_groups()
+
+    ar_weights = torch.load('D:\\dlas\\experiments\\train_music_gpt\\models\\44500_generator_ema.pth')
+    model.ar.load_state_dict(ar_weights, strict=True)
+    diff_weights = torch.load('X:\\dlas\\experiments\\train_music_diffusion_tfd8\\models\\47500_generator_ema.pth')
+    pruned_diff_weights = {}
+    for k,v in diff_weights.items():
+        if k.startswith('diff.'):
+            pruned_diff_weights[k.replace('diff.', '')] = v
+    model.diff.load_state_dict(pruned_diff_weights, strict=False)
+    torch.save(model.state_dict(), 'sample.pth')
+
+    model(clip, ts, cond, conditioning_input=cond)
+
+
+def init_mup():
+    base_model = TransformerDiffusion(model_channels=768, block_channels=768, prenet_channels=768,
+                                                   input_vec_dim=1024, num_layers=16, prenet_layers=4)
+    delta_model = TransformerDiffusion(model_channels=2048, block_channels=1024, prenet_channels=1024,
+                                                   input_vec_dim=1024, num_layers=16, prenet_layers=4)
+    target_model = TransformerDiffusion(model_channels=3072, block_channels=1536, prenet_channels=1536,
+                                                   input_vec_dim=1024, num_layers=16, prenet_layers=4)
+    from mup import set_base_shapes, save_base_shapes
+    set_base_shapes(target_model, base_model, delta=delta_model)
+    save_base_shapes(target_model, 'mup_base_shapes.bsh')
+
+    """
+    # Ah to have a simple loss..
+    def lazy_model(width):
+        return lambda: set_base_shapes(TransformerDiffusion(model_channels=width*2, block_channels=width,
+                                                            prenet_channels=width, num_layers=16, prenet_layers=4,
+                                                            input_vec_dim=1024),
+                                       'mup_base_shapes.bsh')
+    from mup.coord_check import get_coord_data, plot_coord_data
+    models = {256: lazy_model(256), 512: lazy_model(512), 1024: lazy_model(1024), 1536: lazy_model(1536)}
+    dataloader = DataLoader(MupSampleDataset())
+    df = get_coord_data(models, dataloader, dict_in_out=True)
+    plot_coord_data(df, 'coord_check')
+    """
+
+if __name__ == '__main__':
+    init_mup()

codes/models/lucidrains/x_transformers.py

@@ -509,9 +509,10 @@ class Attention(nn.Module):
             rel_pos_bias=False,
             rel_pos_num_buckets=32,
             rel_pos_max_distance=128,
+            mup_scale=False
     ):
         super().__init__()
-        self.scale = dim_head ** -0.5
+        self.scale = 8/dim_head if mup_scale else dim_head ** -0.5
 
         self.heads = heads
         self.causal = causal

codes/requirements.txt

@@ -14,6 +14,7 @@ tensorboard
 orjson
 einops
 lambda-networks
+mup
 
 # For image generation stuff
 opencv-python

codes/trainer/steps.py

@@ -127,6 +127,16 @@ class ConfigurableStep(Module):
                                        weight_decay=opt_get(opt_config, ['weight_decay'], 1e-2),
                                        betas=(opt_get(opt_config, ['beta1'], .9), opt_get(opt_config, ['beta2'], .999)))
                 opt._group_names = [params_names_weights, params_names_notweights]
+            elif self.step_opt['optimizer'] == 'mu_adamw':
+                groups = [
+                    { 'params': params_weights, 'weight_decay': opt_get(opt_config, ['weight_decay'], 0) },
+                    { 'params': params_notweights, 'weight_decay': 0 }
+                ]
+                from mup.optim import MuAdamW
+                opt = MuAdamW(groups, lr=opt_config['lr'],
+                             weight_decay=opt_get(opt_config, ['weight_decay'], 1e-2),
+                             betas=(opt_get(opt_config, ['beta1'], .9), opt_get(opt_config, ['beta2'], .999)))
+                opt._group_names = [params_names_weights, params_names_notweights]
             elif self.step_opt['optimizer'] == 'adamw_zero':
                 # The torch ZeRO implementation does not seem to support parameter groups, so do not shard the non-weighted
                 # parameters and just use a normal AdamW implementation. In a large network, these weights will normally