network updates

codes/models/arch_util.py

@@ -369,7 +369,7 @@ class ResBlock(nn.Module):
     def __init__(
         self,
         channels,
-        dropout,
+        dropout=0,
         out_channels=None,
         use_conv=False,
         dims=2,

codes/models/audio/music/gpt_music.py

@@ -3,12 +3,12 @@ from torch import nn
 import torch.nn.functional as F
 from transformers import GPT2Config, GPT2Model
 
-from models.arch_util import AttentionBlock
+from models.arch_util import AttentionBlock, ResBlock
 from models.audio.music.music_quantizer import MusicQuantizer
 from models.audio.music.music_quantizer2 import MusicQuantizer2
 from models.lucidrains.x_transformers import Encoder
 from trainer.networks import register_model
-from utils.util import opt_get
+from utils.util import opt_get, checkpoint
 
 
 class ConditioningEncoder(nn.Module):
@@ -25,6 +25,32 @@ class ConditioningEncoder(nn.Module):
         self.attn = nn.Sequential(*attn)
         self.dim = embedding_dim
 
+    def forward(self, x):
+        h = checkpoint(self.init, x)
+        h = checkpoint(self.attn, h)
+        return h.mean(dim=2)
+
+
+class UpperConditioningEncoder(nn.Module):
+    def __init__(self,
+                 spec_dim,
+                 embedding_dim,
+                 attn_blocks=6,
+                 num_attn_heads=4):
+        super().__init__()
+        attn = []
+        self.init = nn.Sequential(nn.Conv1d(spec_dim, min(spec_dim+128, embedding_dim), kernel_size=3, stride=2, padding=1),
+                                  nn.Conv1d(min(spec_dim+128, embedding_dim), min(spec_dim+256, embedding_dim), kernel_size=3, stride=2, padding=1),
+                                  nn.Conv1d(min(spec_dim+256, embedding_dim), min(spec_dim+384, embedding_dim), kernel_size=3, stride=2, padding=1),
+                                  nn.Conv1d(min(spec_dim+384, embedding_dim), min(spec_dim+512, embedding_dim), kernel_size=3, stride=2, padding=1),
+                                  ResBlock(min(spec_dim+512, embedding_dim), dims=1),
+                                  nn.Conv1d(min(spec_dim+512, embedding_dim), min(spec_dim+512, embedding_dim), kernel_size=3, stride=2, padding=1),
+                                  ResBlock(min(spec_dim+512, embedding_dim), dims=1))
+        for a in range(attn_blocks):
+            attn.append(AttentionBlock(embedding_dim, num_attn_heads, do_activation=True))
+        self.attn = nn.Sequential(*attn)
+        self.dim = embedding_dim
+
     def forward(self, x):
         h = self.init(x)
         h = self.attn(h)
@@ -135,12 +161,92 @@ class GptMusicLower(nn.Module):
                 )
 
 
+class GptMusicUpper(nn.Module):
+    def __init__(self, dim, layers, dropout=0, num_upper_vectors=64, num_upper_groups=4):
+        super().__init__()
+        self.internal_step = 0
+        self.num_groups = num_upper_groups
+        self.config = GPT2Config(vocab_size=1, n_positions=8192, n_embd=dim, n_layer=layers, n_head=dim//64,
+                                 n_inner=dim*2, attn_pdrop=dropout, resid_pdrop=dropout, gradient_checkpointing=True,
+                                 use_cache=False)
+        self.upper_quantizer = MusicQuantizer2(inp_channels=256, inner_dim=[dim,
+                                                                            max(512,dim-128),
+                                                                            max(512,dim-256),
+                                                                            max(512,dim-384),
+                                                                            max(512,dim-512),
+                                                                            max(512,dim-512)], codevector_dim=dim,
+                                               codebook_size=num_upper_vectors, codebook_groups=num_upper_groups,
+                                               expressive_downsamples=True)
+        # Following are unused quantizer constructs we delete to avoid DDP errors (and to be efficient.. of course..)
+        del self.upper_quantizer.up
+        # Freeze the quantizer.
+        for p in self.upper_quantizer.parameters():
+            p.DO_NOT_TRAIN = True
+            p.requires_grad = False
+
+        self.conditioning_encoder = UpperConditioningEncoder(256, dim, attn_blocks=4, num_attn_heads=dim//64)
+
+        self.gpt = GPT2Model(self.config)
+        del self.gpt.wte  # Unused, we'll do our own embeddings.
+
+        self.embeddings = nn.ModuleList([nn.Embedding(num_upper_vectors, dim // num_upper_groups) for _ in range(num_upper_groups)])
+        self.heads = nn.ModuleList([nn.Linear(dim, num_upper_vectors) for _ in range(num_upper_groups)])
+
+
+    def forward(self, mel, conditioning, return_latent=False):
+        with torch.no_grad():
+            self.upper_quantizer.eval()
+            codes = self.upper_quantizer.get_codes(mel)
+
+        inputs = codes[:, :-1]
+        targets = codes
+        h = [embedding(inputs[:, :, i]) for i, embedding in enumerate(self.embeddings)]
+        h = torch.cat(h, dim=-1)
+
+        with torch.autocast(mel.device.type):
+            # Stick the conditioning embedding on the front of the input sequence.
+            # The transformer will learn how to integrate it.
+            # This statement also serves to pre-pad the inputs by one token, which is the basis of the next-token-prediction task. IOW: this is the "START" token.
+            cond_emb = self.conditioning_encoder(conditioning).unsqueeze(1)
+            h = torch.cat([cond_emb, h], dim=1)
+
+            h = self.gpt(inputs_embeds=h, return_dict=True).last_hidden_state
+
+            if return_latent:
+                return h.float()
+
+            losses = 0
+            for i, head in enumerate(self.heads):
+                logits = head(h).permute(0,2,1)
+                loss = F.cross_entropy(logits, targets[:,:,i])
+                losses = losses + loss
+
+        return losses / self.num_groups
+
+    def get_grad_norm_parameter_groups(self):
+        groups = {
+            'gpt': list(self.gpt.parameters()),
+            'conditioning': list(self.conditioning_encoder.parameters()),
+        }
+        return groups
+
+    def get_debug_values(self, step, __):
+        if self.upper_quantizer.total_codes > 0:
+            return {'histogram_upper_codes': self.upper_quantizer.codes[:self.upper_quantizer.total_codes]}
+        else:
+            return {}
+
+
 @register_model
 def register_music_gpt_lower(opt_net, opt):
     return GptMusicLower(**opt_get(opt_net, ['kwargs'], {}))
 
+@register_model
+def register_music_gpt_upper(opt_net, opt):
+    return GptMusicUpper(**opt_get(opt_net, ['kwargs'], {}))
 
-if __name__ == '__main__':
+
+def test_lower():
     from models.audio.music.transformer_diffusion8 import TransformerDiffusionWithQuantizer
     base_diff = TransformerDiffusionWithQuantizer(in_channels=256, out_channels=512, model_channels=2048, block_channels=1024,
                                                   prenet_channels=1024, prenet_layers=6, num_layers=16, input_vec_dim=1024,
@@ -153,3 +259,18 @@ if __name__ == '__main__':
     mel = torch.randn(2,256,400)
     model(mel, mel)
     model.get_grad_norm_parameter_groups()
+
+
+def test_upper():
+    lower = GptMusicLower(512, 12)
+    lower.load_state_dict(torch.load('D:\\dlas\\experiments\\train_music_gpt\\models\\44500_generator_ema.pth'))
+    model = GptMusicUpper(512, 12)
+    model.upper_quantizer.load_state_dict(lower.upper_quantizer.state_dict())
+    torch.save(model.state_dict(), 'sample.pth')
+    mel = torch.randn(2,256,2500)
+    model(mel, mel)
+    model.get_grad_norm_parameter_groups()
+
+
+if __name__ == '__main__':
+    test_upper()

codes/models/audio/music/transformer_diffusion8.py

@@ -73,6 +73,7 @@ class TransformerDiffusion(nn.Module):
             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.
     ):
@@ -95,6 +96,22 @@ class TransformerDiffusion(nn.Module):
         )
         prenet_heads = 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,
@@ -130,16 +147,16 @@ class TransformerDiffusion(nn.Module):
         }
         return groups
 
-    def timestep_independent(self, codes, expected_seq_len):
+    def timestep_independent(self, prior, expected_seq_len):
-        code_emb = self.input_converter(codes)
+        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(codes.shape[0], 1, 1),
+            code_emb = torch.where(unconditioned_batches, self.unconditioned_embedding.repeat(prior.shape[0], 1, 1),
                                    code_emb)
-        code_emb = self.code_converter(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
@@ -151,7 +168,6 @@ class TransformerDiffusion(nn.Module):
         unused_params = []
         if conditioning_free:
             code_emb = self.unconditioned_embedding.repeat(x.shape[0], x.shape[-1], 1)
-            unused_params.extend(list(self.code_converter.parameters()))
         else:
             if precomputed_code_embeddings is not None:
                 code_emb = precomputed_code_embeddings
@@ -240,6 +256,47 @@ class TransformerDiffusionWithQuantizer(nn.Module):
         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(opt_net, opt):
     return TransformerDiffusion(**opt_net['kwargs'])
@@ -250,24 +307,17 @@ def register_transformer_diffusion8_with_quantizer(opt_net, opt):
     return TransformerDiffusionWithQuantizer(**opt_net['kwargs'])
 
 
-"""
+@register_model
-# For TFD5
+def register_transformer_diffusion8_with_ar_prior(opt_net, opt):
-if __name__ == '__main__':
+    return TransformerDiffusionWithARPrior(**opt_net['kwargs'])
-    clip = torch.randn(2, 256, 400)
-    aligned_sequence = torch.randn(2,100,512)
-    cond = torch.randn(2, 256, 400)
-    ts = torch.LongTensor([600, 600])
-    model = TransformerDiffusion(model_channels=3072, block_channels=1536, prenet_channels=1536)
-    torch.save(model, 'sample.pth')
-    print_network(model)
-    o = model(clip, ts, aligned_sequence, cond)
-"""
 
-if __name__ == '__main__':
+
+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 = 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')
@@ -279,3 +329,28 @@ if __name__ == '__main__':
     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)
+
+
+
+if __name__ == '__main__':
+    test_ar_model()

codes/models/audio/music/unet_diffusion_music_codes.py

@@ -780,6 +780,15 @@ class UNetMusicModelARPrior(nn.Module):
                 del p.DO_NOT_TRAIN
                 p.requires_grad = True
 
+    def get_grad_norm_parameter_groups(self):
+        groups = {
+            'input_blocks': list(self.diff.input_blocks.parameters()),
+            'output_blocks': list(self.diff.output_blocks.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)
@@ -805,6 +814,7 @@ if __name__ == '__main__':
                                   attention_resolutions=(2,4), channel_mult=(1,2,3), dims=1,
                                   use_scale_shift_norm=True, dropout=.1, num_heads=8, unconditioned_percentage=.4, freeze_unet=True)
     print_network(model)
+    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)

codes/train.py

@@ -339,7 +339,7 @@ class Trainer:
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_music_gpt.yml')
+    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_music_gpt_upper.yml')
     parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher')
     args = parser.parse_args()
     opt = option.parse(args.opt, is_train=True)