# @title Model helper # import spaces # for zero-GPU import os from collections import Counter import argparse import torch import torchaudio import numpy as np from model.init_train import initialize_trainer, update_config from utils.task_manager import TaskManager from config.vocabulary import drum_vocab_presets from utils.utils import str2bool from utils.utils import Timer from utils.audio import slice_padded_array from utils.note2event import mix_notes from utils.event2note import merge_zipped_note_events_and_ties_to_notes from utils.utils import write_model_output_as_midi, write_err_cnt_as_json from model.ymt3 import YourMT3 def debug_model_task_config(model): """Debug function to inspect what task configurations are available in the model""" print("=== Model Task Configuration Debug ===") if hasattr(model, 'task_manager'): print(f"✓ Model has task_manager") print(f" Task name: {getattr(model.task_manager, 'task_name', 'Unknown')}") if hasattr(model.task_manager, 'task'): task_config = model.task_manager.task print(f" Task config keys: {list(task_config.keys())}") if 'eval_subtask_prefix' in task_config: print(f" Available subtask prefixes: {list(task_config['eval_subtask_prefix'].keys())}") for key, value in task_config['eval_subtask_prefix'].items(): print(f" {key}: {value}") else: print(" No eval_subtask_prefix found") if 'subtask_tokens' in task_config: print(f" Subtask tokens: {task_config['subtask_tokens']}") else: print(" No task config found") if hasattr(model.task_manager, 'tokenizer'): tokenizer = model.task_manager.tokenizer print(f" Tokenizer available: {type(tokenizer)}") # Try to inspect available events in the codec if hasattr(tokenizer, 'codec'): codec = tokenizer.codec print(f" Codec type: {type(codec)}") if hasattr(codec, '_event_ranges'): print(f" Event ranges: {codec._event_ranges}") else: print(" No tokenizer found") else: print("✗ Model doesn't have task_manager") print("=" * 40) def create_instrument_task_tokens(model, instrument_hint, n_segments): """Create task tokens for instrument-specific transcription conditioning. Args: model: YourMT3 model instance instrument_hint: String indicating desired instrument ('vocals', 'guitar', 'piano', etc.) n_segments: Number of audio segments Returns: torch.LongTensor: Task tokens for conditioning the model """ device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # Check what task configuration is available in the model if not hasattr(model, 'task_manager') or not hasattr(model.task_manager, 'task'): print(f"Warning: Model doesn't have task configuration, skipping task tokens for {instrument_hint}") return None task_config = model.task_manager.task # Check if this model supports subtask prefixes if 'eval_subtask_prefix' in task_config: print(f"Model supports subtask prefixes: {list(task_config['eval_subtask_prefix'].keys())}") # Map instrument hints to available subtask prefixes if instrument_hint.lower() in ['vocals', 'singing', 'voice']: if 'singing-only' in task_config['eval_subtask_prefix']: prefix_tokens = task_config['eval_subtask_prefix']['singing-only'] print(f"Using singing-only task tokens: {prefix_tokens}") else: prefix_tokens = task_config['eval_subtask_prefix'].get('default', []) print(f"Singing task not available, using default: {prefix_tokens}") elif instrument_hint.lower() in ['drums', 'drum', 'percussion']: if 'drum-only' in task_config['eval_subtask_prefix']: prefix_tokens = task_config['eval_subtask_prefix']['drum-only'] print(f"Using drum-only task tokens: {prefix_tokens}") else: prefix_tokens = task_config['eval_subtask_prefix'].get('default', []) print(f"Drum task not available, using default: {prefix_tokens}") else: # For other instruments, use default transcribe_all prefix_tokens = task_config['eval_subtask_prefix'].get('default', []) print(f"Using default task tokens for {instrument_hint}: {prefix_tokens}") else: print(f"Model doesn't support subtask prefixes, using general transcription for {instrument_hint}") # For models without subtask support, return None to use regular transcription return None # Convert to token IDs if we have prefix tokens if prefix_tokens: try: tokenizer = model.task_manager.tokenizer task_token_ids = [] for event in prefix_tokens: try: token_id = tokenizer.codec.encode_event(event) task_token_ids.append(token_id) print(f"Encoded event {event} -> token {token_id}") except Exception as e: print(f"Warning: Could not encode event {event}: {e}") continue if task_token_ids: # Create task token array: (n_segments, 1, task_len) for single channel task_len = len(task_token_ids) task_tokens = torch.zeros((n_segments, 1, task_len), dtype=torch.long, device=device) for i in range(n_segments): task_tokens[i, 0, :] = torch.tensor(task_token_ids, dtype=torch.long) print(f"Created task tokens with shape: {task_tokens.shape}") return task_tokens else: print("No valid task tokens could be created") return None except Exception as e: print(f"Warning: Could not create task tokens for {instrument_hint}: {e}") return None def filter_instrument_consistency(pred_notes, primary_instrument=None, confidence_threshold=0.7, instrument_hint=None): """Post-process transcribed notes to maintain instrument consistency. Args: pred_notes: List of Note objects from transcription primary_instrument: Target instrument program number (if known) confidence_threshold: Threshold for maintaining instrument consistency instrument_hint: Original instrument hint to help with mapping Returns: List of filtered Note objects """ if not pred_notes: return pred_notes # Count instrument occurrences to find dominant instrument instrument_counts = {} total_notes = len(pred_notes) for note in pred_notes: program = getattr(note, 'program', 0) instrument_counts[program] = instrument_counts.get(program, 0) + 1 print(f"Found instruments in transcription: {instrument_counts}") # Determine primary instrument if primary_instrument is None: primary_instrument = max(instrument_counts, key=instrument_counts.get) primary_count = instrument_counts.get(primary_instrument, 0) primary_ratio = primary_count / total_notes if total_notes > 0 else 0 print(f"Primary instrument: {primary_instrument} ({primary_ratio:.2%} of notes)") # Map instrument hints to preferred MIDI programs instrument_program_map = { 'vocals': 100, # Singing voice in YourMT3 'singing': 100, 'voice': 100, 'piano': 0, # Acoustic Grand Piano 'guitar': 24, # Acoustic Guitar (nylon) 'violin': 40, # Violin 'drums': 128, # Drum kit 'bass': 32, # Acoustic Bass 'saxophone': 64, # Soprano Sax 'flute': 73, # Flute } # If we have an instrument hint, try to use the appropriate program if instrument_hint and instrument_hint.lower() in instrument_program_map: target_program = instrument_program_map[instrument_hint.lower()] print(f"Target program for {instrument_hint}: {target_program}") # Check if the target program exists in the transcription if target_program in instrument_counts: primary_instrument = target_program primary_ratio = instrument_counts[target_program] / total_notes print(f"Found target instrument in transcription: {primary_ratio:.2%} of notes") # If primary instrument is dominant enough, filter out other instruments if primary_ratio >= confidence_threshold: print(f"Applying consistency filter (threshold: {confidence_threshold:.2%})") filtered_notes = [] converted_count = 0 for note in pred_notes: note_program = getattr(note, 'program', 0) if note_program == primary_instrument: filtered_notes.append(note) else: # Convert note to primary instrument try: note_copy = note._replace(program=primary_instrument) filtered_notes.append(note_copy) converted_count += 1 except AttributeError: # Handle different note types note_copy = note.__class__( start=note.start, end=note.end, pitch=note.pitch, velocity=note.velocity, program=primary_instrument ) filtered_notes.append(note_copy) converted_count += 1 print(f"Converted {converted_count} notes to primary instrument {primary_instrument}") return filtered_notes else: print(f"Primary instrument ratio ({primary_ratio:.2%}) below threshold ({confidence_threshold:.2%}), keeping all instruments") return pred_notes def load_model_checkpoint(args=None, device='cpu'): parser = argparse.ArgumentParser(description="YourMT3") # General parser.add_argument('exp_id', type=str, help='A unique identifier for the experiment is used to resume training. The "@" symbol can be used to load a specific checkpoint.') parser.add_argument('-p', '--project', type=str, default='ymt3', help='project name') parser.add_argument('-ac', '--audio-codec', type=str, default=None, help='audio codec (default=None). {"spec", "melspec"}. If None, default value defined in config.py will be used.') parser.add_argument('-hop', '--hop-length', type=int, default=None, help='hop length in frames (default=None). {128, 300} 128 for MT3, 300 for PerceiverTFIf None, default value defined in config.py will be used.') parser.add_argument('-nmel', '--n-mels', type=int, default=None, help='number of mel bins (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-if', '--input-frames', type=int, default=None, help='number of audio frames for input segment (default=None). If None, default value defined in config.py will be used.') # Model configurations parser.add_argument('-sqr', '--sca-use-query-residual', type=str2bool, default=None, help='sca use query residual flag. Default follows config.py') parser.add_argument('-enc', '--encoder-type', type=str, default=None, help="Encoder type. 't5' or 'perceiver-tf' or 'conformer'. Default is 't5', following config.py.") parser.add_argument('-dec', '--decoder-type', type=str, default=None, help="Decoder type. 't5' or 'multi-t5'. Default is 't5', following config.py.") parser.add_argument('-preenc', '--pre-encoder-type', type=str, default='default', help="Pre-encoder type. None or 'conv' or 'default'. By default, t5_enc:None, perceiver_tf_enc:conv, conformer:None") parser.add_argument('-predec', '--pre-decoder-type', type=str, default='default', help="Pre-decoder type. {None, 'linear', 'conv1', 'mlp', 'group_linear'} or 'default'. Default is {'t5': None, 'perceiver-tf': 'linear', 'conformer': None}.") parser.add_argument('-cout', '--conv-out-channels', type=int, default=None, help='Number of filters for pre-encoder conv layer. Default follows "model_cfg" of config.py.') parser.add_argument('-tenc', '--task-cond-encoder', type=str2bool, default=True, help='task conditional encoder (default=True). True or False') parser.add_argument('-tdec', '--task-cond-decoder', type=str2bool, default=True, help='task conditional decoder (default=True). True or False') parser.add_argument('-df', '--d-feat', type=int, default=None, help='Audio feature will be projected to this dimension for Q,K,V of T5 or K,V of Perceiver (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-pt', '--pretrained', type=str2bool, default=False, help='pretrained T5(default=False). True or False') parser.add_argument('-b', '--base-name', type=str, default="google/t5-v1_1-small", help='base model name (default="google/t5-v1_1-small")') parser.add_argument('-epe', '--encoder-position-encoding-type', type=str, default='default', help="Positional encoding type of encoder. By default, pre-defined PE for T5 or Perceiver-TF encoder in config.py. For T5: {'sinusoidal', 'trainable'}, conformer: {'rotary', 'trainable'}, Perceiver-TF: {'trainable', 'rope', 'alibi', 'alibit', 'None', '0', 'none', 'tkd', 'td', 'tk', 'kdt'}.") parser.add_argument('-dpe', '--decoder-position-encoding-type', type=str, default='default', help="Positional encoding type of decoder. By default, pre-defined PE for T5 in config.py. {'sinusoidal', 'trainable'}.") parser.add_argument('-twe', '--tie-word-embedding', type=str2bool, default=None, help='tie word embedding (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-el', '--event-length', type=int, default=None, help='event length (default=None). If None, default value defined in model cfg of config.py will be used.') # Perceiver-TF configurations parser.add_argument('-dl', '--d-latent', type=int, default=None, help='Latent dimension of Perceiver. On T5, this will be ignored (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-nl', '--num-latents', type=int, default=None, help='Number of latents of Perceiver. On T5, this will be ignored (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-dpm', '--perceiver-tf-d-model', type=int, default=None, help='Perceiver-TF d_model (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-npb', '--num-perceiver-tf-blocks', type=int, default=None, help='Number of blocks of Perceiver-TF. On T5, this will be ignored (default=None). If None, default value defined in config.py.') parser.add_argument('-npl', '--num-perceiver-tf-local-transformers-per-block', type=int, default=None, help='Number of local layers per block of Perceiver-TF. On T5, this will be ignored (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-npt', '--num-perceiver-tf-temporal-transformers-per-block', type=int, default=None, help='Number of temporal layers per block of Perceiver-TF. On T5, this will be ignored (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-atc', '--attention-to-channel', type=str2bool, default=None, help='Attention to channel flag of Perceiver-TF. On T5, this will be ignored (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-ln', '--layer-norm-type', type=str, default=None, help='Layer normalization type (default=None). {"layer_norm", "rms_norm"}. If None, default value defined in config.py will be used.') parser.add_argument('-ff', '--ff-layer-type', type=str, default=None, help='Feed forward layer type (default=None). {"mlp", "moe", "gmlp"}. If None, default value defined in config.py will be used.') parser.add_argument('-wf', '--ff-widening-factor', type=int, default=None, help='Feed forward layer widening factor for MLP/MoE/gMLP (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-nmoe', '--moe-num-experts', type=int, default=None, help='Number of experts for MoE (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-kmoe', '--moe-topk', type=int, default=None, help='Top-k for MoE (default=None). If None, default value defined in config.py will be used.') parser.add_argument('-act', '--hidden-act', type=str, default=None, help='Hidden activation function (default=None). {"gelu", "silu", "relu", "tanh"}. If None, default value defined in config.py will be used.') parser.add_argument('-rt', '--rotary-type', type=str, default=None, help='Rotary embedding type expressed in three letters. e.g. ppl: "pixel" for SCA and latents, "lang" for temporal transformer. If None, use config.') parser.add_argument('-rk', '--rope-apply-to-keys', type=str2bool, default=None, help='Apply rope to keys (default=None). If None, use config.') parser.add_argument('-rp', '--rope-partial-pe', type=str2bool, default=None, help='Whether to apply RoPE to partial positions (default=None). If None, use config.') # Decoder configurations parser.add_argument('-dff', '--decoder-ff-layer-type', type=str, default=None, help='Feed forward layer type of decoder (default=None). {"mlp", "moe", "gmlp"}. If None, default value defined in config.py will be used.') parser.add_argument('-dwf', '--decoder-ff-widening-factor', type=int, default=None, help='Feed forward layer widening factor for decoder MLP/MoE/gMLP (default=None). If None, default value defined in config.py will be used.') # Task and Evaluation configurations parser.add_argument('-tk', '--task', type=str, default='mt3_full_plus', help='tokenizer type (default=mt3_full_plus). See config/task.py for more options.') parser.add_argument('-epv', '--eval-program-vocab', type=str, default=None, help='evaluation vocabulary (default=None). If None, default vocabulary of the data preset will be used.') parser.add_argument('-edv', '--eval-drum-vocab', type=str, default=None, help='evaluation vocabulary for drum (default=None). If None, default vocabulary of the data preset will be used.') parser.add_argument('-etk', '--eval-subtask-key', type=str, default='default', help='evaluation subtask key (default=default). See config/task.py for more options.') parser.add_argument('-t', '--onset-tolerance', type=float, default=0.05, help='onset tolerance (default=0.05).') parser.add_argument('-os', '--test-octave-shift', type=str2bool, default=False, help='test optimal octave shift (default=False). True or False') parser.add_argument('-w', '--write-model-output', type=str2bool, default=True, help='write model test output to file (default=False). True or False') # Trainer configurations parser.add_argument('-pr','--precision', type=str, default="bf16-mixed", help='precision (default="bf16-mixed") {32, 16, bf16, bf16-mixed}') parser.add_argument('-st', '--strategy', type=str, default='auto', help='strategy (default=auto). auto or deepspeed or ddp') parser.add_argument('-n', '--num-nodes', type=int, default=1, help='number of nodes (default=1)') parser.add_argument('-g', '--num-gpus', type=str, default='auto', help='number of gpus (default="auto")') parser.add_argument('-wb', '--wandb-mode', type=str, default="disabled", help='wandb mode for logging (default=None). "disabled" or "online" or "offline". If None, default value defined in config.py will be used.') # Debug parser.add_argument('-debug', '--debug-mode', type=str2bool, default=False, help='debug mode (default=False). True or False') parser.add_argument('-tps', '--test-pitch-shift', type=int, default=None, help='use pitch shift when testing. debug-purpose only. (default=None). semitone in int.') args = parser.parse_args(args) # yapf: enable if torch.__version__ >= "1.13": torch.set_float32_matmul_precision("high") args.epochs = None # Initialize and update config _, _, dir_info, shared_cfg = initialize_trainer(args, stage='test') shared_cfg, audio_cfg, model_cfg = update_config(args, shared_cfg, stage='test') if args.eval_drum_vocab != None: # override eval_drum_vocab eval_drum_vocab = drum_vocab_presets[args.eval_drum_vocab] # Initialize task manager tm = TaskManager(task_name=args.task, max_shift_steps=int(shared_cfg["TOKENIZER"]["max_shift_steps"]), debug_mode=args.debug_mode) print(f"Task: {tm.task_name}, Max Shift Steps: {tm.max_shift_steps}") # Use GPU if available device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # Model model = YourMT3( audio_cfg=audio_cfg, model_cfg=model_cfg, shared_cfg=shared_cfg, optimizer=None, task_manager=tm, # tokenizer is a member of task_manager eval_subtask_key=args.eval_subtask_key, write_output_dir=dir_info["lightning_dir"] if args.write_model_output or args.test_octave_shift else None ).to(device) checkpoint = torch.load(dir_info["last_ckpt_path"], map_location=device, weights_only=False) state_dict = checkpoint['state_dict'] new_state_dict = {k: v for k, v in state_dict.items() if 'pitchshift' not in k} model.load_state_dict(new_state_dict, strict=False) return model.eval() # load checkpoint on cpu first def transcribe(model, audio_info, instrument_hint=None): t = Timer() # Converting Audio t.start() audio, sr = torchaudio.load(uri=audio_info['filepath']) audio = torch.mean(audio, dim=0).unsqueeze(0) audio = torchaudio.functional.resample(audio, sr, model.audio_cfg['sample_rate']) audio_segments = slice_padded_array(audio, model.audio_cfg['input_frames'], model.audio_cfg['input_frames']) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") audio_segments = torch.from_numpy(audio_segments.astype('float32')).to(device).unsqueeze(1) # (n_seg, 1, seg_sz) t.stop(); t.print_elapsed_time("converting audio"); # Inference t.start() # Debug model configuration when using instrument hints if instrument_hint: print(f"Attempting to create task tokens for instrument: {instrument_hint}") debug_model_task_config(model) # Create task tokens for instrument-specific transcription task_tokens = None if instrument_hint: task_tokens = create_instrument_task_tokens(model, instrument_hint, audio_segments.shape[0]) pred_token_arr, _ = model.inference_file(bsz=8, audio_segments=audio_segments, task_token_array=task_tokens) t.stop(); t.print_elapsed_time("model inference"); # Post-processing t.start() num_channels = model.task_manager.num_decoding_channels n_items = audio_segments.shape[0] start_secs_file = [model.audio_cfg['input_frames'] * i / model.audio_cfg['sample_rate'] for i in range(n_items)] pred_notes_in_file = [] n_err_cnt = Counter() for ch in range(num_channels): pred_token_arr_ch = [arr[:, ch, :] for arr in pred_token_arr] # (B, L) zipped_note_events_and_tie, list_events, ne_err_cnt = model.task_manager.detokenize_list_batches( pred_token_arr_ch, start_secs_file, return_events=True) pred_notes_ch, n_err_cnt_ch = merge_zipped_note_events_and_ties_to_notes(zipped_note_events_and_tie) pred_notes_in_file.append(pred_notes_ch) n_err_cnt += n_err_cnt_ch pred_notes = mix_notes(pred_notes_in_file) # This is the mixed notes from all channels # Apply instrument consistency filter if instrument hint was provided if instrument_hint: print(f"Applying instrument consistency filter for: {instrument_hint}") # Use more aggressive filtering if task tokens weren't available confidence_threshold = 0.6 if task_tokens is not None else 0.4 print(f"Using confidence threshold: {confidence_threshold}") pred_notes = filter_instrument_consistency(pred_notes, confidence_threshold=confidence_threshold, instrument_hint=instrument_hint) # Write MIDI write_model_output_as_midi(pred_notes, './', audio_info['track_name'], model.midi_output_inverse_vocab) t.stop(); t.print_elapsed_time("post processing"); midifile = os.path.join('./model_output/', audio_info['track_name'] + '.mid') assert os.path.exists(midifile) return midifile