From a659cd865c1878dff661a96b72886371e689aadc Mon Sep 17 00:00:00 2001
From: James Betker <jbetker@gmail.com>
Date: Sun, 19 Jun 2022 23:12:52 -0600
Subject: [PATCH] All the stuff needed for cheater latent generation

---
 codes/data/__init__.py                        |   2 +
 codes/data/audio/preprocessed_mel_dataset.py  |  71 +++++
 ...ormer_diffusion_with_point_conditioning.py | 272 ++++++++++++++++++
 codes/train.py                                |   2 +-
 codes/trainer/injectors/audio_injectors.py    |  15 +
 5 files changed, 361 insertions(+), 1 deletion(-)
 create mode 100644 codes/data/audio/preprocessed_mel_dataset.py
 create mode 100644 codes/models/audio/music/transformer_diffusion_with_point_conditioning.py

diff --git a/codes/data/__init__.py b/codes/data/__init__.py
index f296a53c..ac426ba1 100644
--- a/codes/data/__init__.py
+++ b/codes/data/__init__.py
@@ -95,6 +95,8 @@ def create_dataset(dataset_opt, return_collate=False):
         from data.audio.unsupervised_audio_dataset import UnsupervisedAudioDataset as D
     elif mode == 'unsupervised_audio_with_noise':
         from data.audio.audio_with_noise_dataset import AudioWithNoiseDataset as D
+    elif mode == 'preprocessed_mel':
+        from data.audio.preprocessed_mel_dataset import PreprocessedMelDataset as D
     elif mode == 'grand_conjoined_voice':
         from data.audio.grand_conjoined_dataset import GrandConjoinedDataset as D
         from data.zero_pad_dict_collate import ZeroPadDictCollate as C
diff --git a/codes/data/audio/preprocessed_mel_dataset.py b/codes/data/audio/preprocessed_mel_dataset.py
new file mode 100644
index 00000000..3c8e813d
--- /dev/null
+++ b/codes/data/audio/preprocessed_mel_dataset.py
@@ -0,0 +1,71 @@
+import os
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+import torch.utils.data
+import torchaudio
+import torchvision
+from tqdm import tqdm
+
+from utils.util import opt_get
+
+
+class PreprocessedMelDataset(torch.utils.data.Dataset):
+
+    def __init__(self, opt):
+        path = opt['path']
+        cache_path = opt['cache_path']  # Will fail when multiple paths specified, must be specified in this case.
+        if os.path.exists(cache_path):
+            self.paths = torch.load(cache_path)
+        else:
+            path = Path(path)
+            self.paths = [str(p) for p in path.rglob("*.npz")]
+            torch.save(self.paths, cache_path)
+        self.pad_to = opt_get(opt, ['pad_to_samples'], 10336)
+
+    def __getitem__(self, index):
+        with np.load(self.paths[index]) as npz_file:
+            mel = torch.tensor(npz_file['arr_0'])
+        assert mel.shape[-1] <= self.pad_to
+        padding_needed = self.pad_to - mel.shape[-1]
+        mask = torch.zeros_like(mel)
+        if padding_needed > 0:
+            mel = F.pad(mel, (0,padding_needed))
+            mask = F.pad(mask, (0,padding_needed), value=1)
+
+        output = {
+            'mel': mel,
+            'mel_lengths': torch.tensor(mel.shape[-1]),
+            'mask': mask,
+            'mask_lengths': torch.tensor(mask.shape[-1]),
+            'path': self.paths[index],
+        }
+        return output
+
+    def __len__(self):
+        return len(self.paths)
+
+
+if __name__ == '__main__':
+    params = {
+        'mode': 'preprocessed_mel',
+        'path': 'Y:\\separated\\large_mels',
+        'cache_path': 'Y:\\separated\\large_mels.pth',
+        'pad_to_samples': 10336,
+        'phase': 'train',
+        'n_workers': 0,
+        'batch_size': 16,
+    }
+    from data import create_dataset, create_dataloader
+
+    ds = create_dataset(params)
+    dl = create_dataloader(ds, params)
+    i = 0
+    for b in tqdm(dl):
+        #pass
+        torchvision.utils.save_image((b['mel'].unsqueeze(1)+1)/2, f'{i}.png')
+        i += 1
+        if i > 20:
+            break
diff --git a/codes/models/audio/music/transformer_diffusion_with_point_conditioning.py b/codes/models/audio/music/transformer_diffusion_with_point_conditioning.py
new file mode 100644
index 00000000..68b09b4a
--- /dev/null
+++ b/codes/models/audio/music/transformer_diffusion_with_point_conditioning.py
@@ -0,0 +1,272 @@
+import itertools
+from time import time
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from models.arch_util import ResBlock, AttentionBlock
+from models.audio.music.gpt_music2 import UpperEncoder, GptMusicLower
+from models.audio.music.music_quantizer2 import MusicQuantizer2
+from models.audio.tts.lucidrains_dvae import DiscreteVAE
+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, RMSScaleShiftNorm, RotaryEmbedding, \
+    FeedForward
+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 SubBlock(nn.Module):
+    def __init__(self, inp_dim, contraction_dim, heads, dropout):
+        super().__init__()
+        self.attn = Attention(inp_dim, out_dim=contraction_dim, heads=heads, dim_head=contraction_dim//heads, causal=False, dropout=dropout)
+        self.attnorm = nn.LayerNorm(contraction_dim)
+        self.ff = FeedForward(inp_dim+contraction_dim, dim_out=contraction_dim, mult=2, dropout=dropout)
+        self.ffnorm = nn.LayerNorm(contraction_dim)
+
+    def forward(self, x, rotary_emb):
+        ah, _, _, _ = checkpoint(self.attn, x, None, None, None, None, None, rotary_emb)
+        ah = F.gelu(self.attnorm(ah))
+        h = torch.cat([ah, x], dim=-1)
+        hf = checkpoint(self.ff, h)
+        hf = F.gelu(self.ffnorm(hf))
+        h = torch.cat([h, hf], dim=-1)
+        return h
+
+
+class ConcatAttentionBlock(TimestepBlock):
+    def __init__(self, trunk_dim, contraction_dim, time_embed_dim, heads, dropout):
+        super().__init__()
+        self.prenorm = RMSScaleShiftNorm(trunk_dim, embed_dim=time_embed_dim, bias=False)
+        self.block1 = SubBlock(trunk_dim, contraction_dim, heads, dropout)
+        self.block2 = SubBlock(trunk_dim+contraction_dim*2, contraction_dim, heads, dropout)
+        self.out = nn.Linear(contraction_dim*4, trunk_dim, bias=False)
+        self.out.weight.data.zero_()
+
+    def forward(self, x, conditioning, timestep_emb, rotary_emb):
+        h = self.prenorm(x, norm_scale_shift_inp=timestep_emb)
+        h = torch.cat([conditioning, h], dim=1)
+        h = self.block1(h, rotary_emb)
+        h = self.block2(h, rotary_emb)
+        h = self.out(h[:,:,x.shape[-1]:])
+        return h[:, 1:] + 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,
+            in_channels=256,
+            out_channels=512,  # mean and variance
+            model_channels=1024,
+            contraction_dim=256,
+            time_embed_dim=256,
+            num_layers=8,
+            rotary_emb_dim=32,
+            input_cond_dim=1024,
+            num_heads=8,
+            dropout=0,
+            use_fp16=False,
+            # Parameters for regularization.
+            unconditioned_percentage=.1,  # This implements a mechanism similar to what is used in classifier-free training.
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.time_embed_dim = time_embed_dim
+        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, model_channels, 3, 1, 1)
+
+        self.time_embed = nn.Sequential(
+            linear(time_embed_dim, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        self.conditioner = nn.Linear(input_cond_dim, model_channels) if input_cond_dim != model_channels else nn.Identity()
+        self.unconditioned_embedding = nn.Parameter(torch.randn(1,1,model_channels))
+        self.rotary_embeddings = RotaryEmbedding(rotary_emb_dim)
+        self.layers = TimestepRotaryEmbedSequential(*[ConcatAttentionBlock(model_channels, contraction_dim, time_embed_dim, num_heads, 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)),
+        )
+
+        self.debug_codes = {}
+
+    def get_grad_norm_parameter_groups(self):
+        attn1 = list(itertools.chain.from_iterable([lyr.block1.attn.parameters() for lyr in self.layers]))
+        attn2 = list(itertools.chain.from_iterable([lyr.block2.attn.parameters() for lyr in self.layers]))
+        ff1 = list(itertools.chain.from_iterable([lyr.block1.ff.parameters() for lyr in self.layers]))
+        ff2 = list(itertools.chain.from_iterable([lyr.block2.ff.parameters() for lyr in self.layers]))
+        blkout_layers = list(itertools.chain.from_iterable([lyr.out.parameters() for lyr in self.layers]))
+        groups = {
+            'prenorms': list(itertools.chain.from_iterable([lyr.prenorm.parameters() for lyr in self.layers])),
+            'blk1_attention_layers': attn1,
+            'blk2_attention_layers': attn2,
+            'attention_layers': attn1 + attn2,
+            'blk1_ff_layers': ff1,
+            'blk2_ff_layers': ff2,
+            'ff_layers': ff1 + ff2,
+            'block_out_layers': blkout_layers,
+            'rotary_embeddings': list(self.rotary_embeddings.parameters()),
+            'out': list(self.out.parameters()),
+            'x_proj': list(self.inp_block.parameters()),
+            'layers': list(self.layers.parameters()),
+            'time_embed': list(self.time_embed.parameters()),
+        }
+        return groups
+
+    def forward(self, x, timesteps, conditioning_input, conditioning_free=False):
+        unused_params = []
+        if conditioning_free:
+            cond = self.unconditioned_embedding.repeat(x.shape[0], x.shape[-1], 1)
+        else:
+            cond = self.conditioner(conditioning_input)
+            # 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((cond.shape[0], 1, 1),
+                                                   device=cond.device) < self.unconditioned_percentage
+                cond = torch.where(unconditioned_batches, self.unconditioned_embedding.repeat(cond.shape[0], 1, 1),
+                                       cond)
+            unused_params.append(self.unconditioned_embedding)
+
+        with torch.autocast(x.device.type, enabled=self.enable_fp16):
+            blk_emb = self.time_embed(timestep_embedding(timesteps, self.time_embed_dim))
+            x = self.inp_block(x).permute(0,2,1)
+
+            rotary_pos_emb = self.rotary_embeddings(x.shape[1]+1, x.device)
+            for layer in self.layers:
+                x = checkpoint(layer, x, cond, 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 ConditioningEncoder(nn.Module):
+    def __init__(self,
+                 cond_dim,
+                 embedding_dim,
+                 attn_blocks=6,
+                 num_attn_heads=8,
+                 do_checkpointing=False):
+        super().__init__()
+        attn = []
+        self.init = nn.Conv1d(cond_dim, embedding_dim, kernel_size=1)
+        for a in range(attn_blocks):
+            attn.append(AttentionBlock(embedding_dim, num_attn_heads, do_checkpoint=do_checkpointing))
+        self.attn = nn.Sequential(*attn)
+        self.dim = embedding_dim
+        self.do_checkpointing = do_checkpointing
+
+    def forward(self, x):
+        h = self.init(x)
+        h = self.attn(h)
+        return h.mean(dim=2).unsqueeze(1)
+
+
+class TransformerDiffusionWithConditioningEncoder(nn.Module):
+    def __init__(self, **kwargs):
+        super().__init__()
+        self.internal_step = 0
+        self.diff = TransformerDiffusion(**kwargs)
+        self.conditioning_encoder = ConditioningEncoder(256, kwargs['model_channels'])
+        self.encoder = UpperEncoder(256, 1024, 256).eval()
+        for p in self.encoder.parameters():
+            p.DO_NOT_TRAIN = True
+            p.requires_grad = False
+
+    def forward(self, x, timesteps, true_cheater, conditioning_input=None, disable_diversity=False, conditioning_free=False):
+        cond = self.conditioning_encoder(true_cheater)
+        diff = self.diff(x, timesteps, conditioning_input=cond, conditioning_free=conditioning_free)
+        return diff
+
+    def get_debug_values(self, step, __):
+        self.internal_step = step
+        return {}
+
+    def get_grad_norm_parameter_groups(self):
+        groups = self.diff.get_grad_norm_parameter_groups()
+        groups['conditioning_encoder'] = list(self.conditioning_encoder.parameters())
+        return
+
+    def before_step(self, step):
+        scaled_grad_parameters = list(itertools.chain.from_iterable([lyr.out.parameters() for lyr in self.diff.layers])) + \
+                                 list(itertools.chain.from_iterable([lyr.prenorm.parameters() for lyr in self.diff.layers]))
+        # Scale back the gradients of the blkout and prenorm layers by a constant factor. These get two orders of magnitudes
+        # higher gradients. Ideally we would use parameter groups, but ZeroRedundancyOptimizer makes this trickier than
+        # directly fiddling with the gradients.
+        for p in scaled_grad_parameters:
+            if hasattr(p, 'grad') and p.grad is not None:
+                p.grad *= .2
+
+
+@register_model
+def register_transformer_diffusion_with_point_conditioning(opt_net, opt):
+    return TransformerDiffusion(**opt_net['kwargs'])
+
+
+@register_model
+def register_tfdpc_with_conditioning_encoder(opt_net, opt):
+    return TransformerDiffusionWithConditioningEncoder(**opt_net['kwargs'])
+
+
+def test_cheater_model():
+    clip = torch.randn(2, 256, 400)
+    cl = torch.randn(2, 1, 400)
+    ts = torch.LongTensor([600, 600])
+
+    # For music:
+    model = TransformerDiffusionWithConditioningEncoder(model_channels=1024)
+    print_network(model)
+    o = model(clip, ts, cl)
+    pg = model.get_grad_norm_parameter_groups()
+
+
+if __name__ == '__main__':
+    test_cheater_model()
diff --git a/codes/train.py b/codes/train.py
index 033bb894..42027511 100644
--- a/codes/train.py
+++ b/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_cheater_gen.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)
diff --git a/codes/trainer/injectors/audio_injectors.py b/codes/trainer/injectors/audio_injectors.py
index e0cb97d8..7634effb 100644
--- a/codes/trainer/injectors/audio_injectors.py
+++ b/codes/trainer/injectors/audio_injectors.py
@@ -378,5 +378,20 @@ class ChannelClipInjector(Injector):
         return {self.output: inp[:,self.lo:self.hi]}
 
 
+class MusicCheaterLatentInjector(Injector):
+    def __init__(self, opt, env):
+        super().__init__(opt, env)
+        from models.audio.music.gpt_music2 import UpperEncoder
+        self.encoder = UpperEncoder(256, 1024, 256).eval()
+        self.encoder.load_state_dict(torch.load('../experiments/music_cheater_encoder_256.pth', map_location=torch.device('cpu')))
+
+    def forward(self, state):
+        with torch.no_grad():
+            mel = state[self.input]
+            self.encoder = self.encoder.to(mel.device)
+            proj = self.encoder(mel)
+            return {self.output: proj}
+
+
 if __name__ == '__main__':
     print('hi')
\ No newline at end of file