midisimx

midisimx

Greatly improved, enhanced, and streamlined fork of midisim for calculating, searching, and analyzing MIDI-to-MIDI similarity at scale

---

What's new

🌟 midisimx vs midisim — comparison table

Main features

• Ultra-fast and flexible GPU/CPU MIDI-to-MIDI similarity calculation, search and analysis
• Quality pre-trained model and pre-computed embeddings sets
• Stand-alone, versatile, and extensive codebase for general or custom MIDI-to-MIDI similarity tasks
• Full cross-platform compatibility and support

---

Pre-trained model

• midisimx_trained_model_14391_steps_0.255_loss_0.9036_acc.pth - Unified and fast large model for a nuanced embeddings generation. Download checkpoint from Hugging Face

This model was trained on full Discover Piano dataset for 2 complete epochs

---

Pre-computed embeddings sets

Weighted Mean Pool Embeddings (1-2-1-2)

• These embeddings put more emphasis on pitches and chords (weights == 2) with start-times and durations left as is (weights == 1)

discover_midi_dataset_3267574_clean_midis_embeddings_1_2_1_2_weighted_cc_by_nc_sa.npy - 3267574 clean MIDIs weighted embeddings from Discover MIDI Dataset for large scale similarity search and analysis tasks

lakh_midi_dataset_17203_clean_midis_embeddings_1_2_1_2_weighted_cc_by_nc_sa.npy - 17203 LAKH clean_midi subset weighted embeddings tailored primarily for artist/song identification tasks

Source MIDI datasets: Discover MIDI Dataset and LAKH MIDI Dataset

---

Similarity search output samples

midisimx-similarity-search-output-samples-1-2-1-2-weighted-CC-BY-NC-SA.zip - ~182k+ MIDIs filtered by weighted midisimx music discovery pipeline

Source MIDI dataset: Discover MIDI Dataset

---

Installation

midisimx PyPI package (for general use)

!pip install -U midisimx

x-transformers 2.3.1 (for raw/custom tasks)

!pip install x-transformers==2.3.1

---

Basic use guide

General use example

# ================================================================================================
# Initalize midisimx
# ================================================================================================

# Import main midisimx module
import midisimx

# ================================================================================================
# Prepare midisimx embeddings
# ================================================================================================

# Option 1: Download sample pre-computed embeddings corpus from Hugging Face
emb_path = midisimx.download_embeddings()

# Option 2: use custom pre-computed embeddings corpus
# See custom embeddings generation section of this README for details
# emb_path = './custom_midis_embeddings_corpus.npy'

# Load downloaded embeddings corpus
corpus_midi_names, corpus_emb = midisimx.load_embeddings(emb_path)

# ================================================================================================
# Prepare midisimx model
# ================================================================================================

# Option 1: Download main pre-trained midisimx model from Hugging Face
model_path = midisimx.download_model()

# Option 2: Use main pre-trained midisimx model included in midisimx PyPI package
# model_path = midisimx.get_package_models()[0]['path']

# Load midisimx model
model, ctx, dtype = midisimx.load_model(model_path)

# ================================================================================================
# Prepare source MIDI
# ================================================================================================

# Load source MIDI
input_toks_seqs = midisimx.midi_to_tokens('Come To My Window.mid')

# ================================================================================================
# Calculate and analyze embeddings
# ================================================================================================

# Compute source/query embeddings
query_emb = midisimx.get_embeddings_bf16(model,
										 input_toks_seqs,
                                         device=torch.device('cuda'),
                                         pooling='weighted_mean',
  										 # The following arg is optional but recommended if
										 # you want to make an emphasis on music
										 # Remove it for overall/general similarity searches
										 # PLEAE NOTE: You must enable it if you are using
										 # included pre-computed weighted embeddings
										 token_type_weights={(128, 256): 2, # Pitches weight
                                                             (384, 718): 2  # Chords weight
                                                            },
										)

# Calculate cosine similarity between source/query MIDI embeddings and embeddings corpus
idxs, sims = midisimx.cosine_similarity_topk(query_emb, corpus_emb)

# ================================================================================================
# Processs, print and save results
# ================================================================================================

# Convert the results to sorted list with transpose values
idxs_sims_tvs_list = midisimx.idxs_sims_to_sorted_list(idxs, sims)

# Print corpus matches (and optionally) convert the final result to a handy list for further processing
corpus_matches_list = midisimx.print_sorted_idxs_sims_list(idxs_sims_tvs_list, corpus_midi_names, return_as_list=True)

# ================================================================================================
# Copy matched MIDIs from the MIDI corpus for listening and further evaluation and analysis
# ================================================================================================

# Copy matched corpus MIDI to a desired directory for easy evaluation and analysis
out_dir_path = midisimx.copy_corpus_files(corpus_matches_list)

# ================================================================================================

Raw/custom use example

import torch
from x_transformers import TransformerWrapper, Encoder

# Original model hyperparameters
SEQ_LEN = 3072

MASK_IDX     = 718 # Use this value for masked modelling
PAD_IDX      = 719 # Model pad index
VOCAB_SIZE   = 720 # Total vocab size

MASK_PROB    = 0.15 # Original training mask probability value (use for masked modelling)

DEVICE = 'cuda' # You can use any compatible device or CPU
DTYPE  = torch.bfloat16 # Original training dtype

# Official main midisimx model checkpoint name
MODEL_CKPT = 'midisimx_trained_model_14391_steps_0.255_loss_0.9036_acc.pth'

# Model architecture using x-transformers
model = TransformerWrapper(
    num_tokens = VOCAB_SIZE,
    max_seq_len = SEQ_LEN,
    attn_layers = Encoder(
        dim   = 768,
        depth = 16,
        heads = 12,
        rotary_pos_emb = True,
        attn_flash = True,
    ),
)

model.load_state_dict(torch.load(MODEL_CKPT, map_location=DEVICE))

model.to(DEVICE)
model.eval()

# Original training autoxast setup
autocast_ctx = torch.amp.autocast(device_type=DEVICE, dtype=DTYPE)

---

Creating custom MIDI corpus embeddings

# ================================================================================================

# Load main midisimx module
import midisimx

# Import helper modules
import os
import tqdm

# ================================================================================================

# Call included TMIDIX module through midisimx to create MIDI files list
custom_midi_corpus_file_names = midisimx.TMIDIX.create_files_list(['./custom_midi_corpus_dir/'])

# ================================================================================================

# Create two lists: one with MIDI corpus file names 
# and another with MIDI corpus tokens representations suitable for embeddings generation
midi_corpus_file_names = []
midi_corpus_tokens = []

for midi_file in tqdm.tqdm(custom_midi_corpus_file_names):
    midi_corpus_file_names.append(os.path.splitext(os.path.basename(midi_file))[0])
    
    midi_tokens = midisimx.midi_to_tokens(midi_file, transpose_factor=0, verbose=False)[0]
    midi_corpus_tokens.append(midi_tokens)

# It is highly recommended to sort the resulting corpus by tokens sequence length
# This greatly speeds up embeddings calculations
sorted_midi_corpus = sorted(zip(midi_corpus_file_names, midi_corpus_tokens), key=lambda x: len(x[1]))
midi_corpus_file_names, midi_corpus_tokens = map(list, zip(*sorted_midi_corpus))

# ================================================================================================
# Now you are ready to generate embeddings as follows:
# ================================================================================================

# Load main midisimx model
model, ctx, dtype = midisimx.load_model(verbose=False)

# Generate MIDI corpus embeddings
midi_corpus_embeddings = midisimx.get_embeddings_bf16(model, midi_corpus_tokens, verbose=False)

# ================================================================================================

# Save generated MIDI corpus embeddings and MIDI corpus file names in one handy NumPy file
midisimx.save_embeddings(midi_corpus_file_names,
                        midi_corpus_embeddings,
                        verbose=False
                       )

# ================================================================================================

# You now can use this saved custom MIDI corpus NumPy file with midisimx.load_embeddings()
# and the rest of the pipeline outlined in the general use section above

---

Music discovery pipeline

Here is a complete MIDI music discovery pipeline example using midisimx and Discover MIDI Dataset

Install midisimx and discovermidi PyPI packages

!pip install -U midisimx

!pip install -U discovermidi

Download and unzip Discover MIDI Dataset

import discovermidi
from discovermidi import fast_parallel_extract

discovermidi.download_dataset()

fast_parallel_extract.fast_parallel_extract()

Prepare midisimx model and desired corresponding embeddings set

model_ckpt = 'midisimx_trained_model_14391_steps_0.255_loss_0.9036_acc.pth'
model_depth = 16

embeddings_file = 'discover_midi_dataset_3267574_clean_midis_embeddings_1_2_1_2_weighted_cc_by_nc_sa.npy'

Create Master MIDI dataset directory and upload your source/master MIDIs in it

import os

os.makedirs('./Master-MIDI-Dataset/', exist_ok=True)

Initialize midisimx, download and load midisimx model and embeddings set

# Import main midisimx module
import midisimx

# Download embeddings from Hugging Face
emb_path = midisimx.download_embeddings(filename=embeddings_file)

# Load downloaded embeddings corpus
corpus_midi_names, corpus_emb = midisimx.load_embeddings(embeddings_path=emb_path)

# Download midisimx model from Hugging Face
model_path = midisimx.download_model(filename=model_ckpt)

# Load midisimx model
model, ctx, dtype = midisimx.load_model(model_path,
                                       depth=model_depth
                                      )

Create Master MIDI dataset files list

filez = midisimx.TMIDIX.create_files_list(['./Master-MIDI-Dataset/'])

Launch the search

import os
import tqdm

for fa in tqdm.tqdm(filez):
    
    # Load source MIDI
    input_toks_seqs = midisimx.midi_to_tokens(fa, verbose=False)

    if input_toks_seqs:
    
        # ================================================================================================
        # Calculate and analyze embeddings
        # ================================================================================================
        
        # Compute source/query embeddings
        query_emb = midisimx.get_embeddings_bf16(model,
                                                input_toks_seqs,

		                                        device=torch.device('cuda'),
		                                        pooling='weighted_mean',
	  										    # The following arg is optional but recommended if
											    # you want to make an emphasis on music
											    # Remove it for overall/general similarity searches
											    # PLEAE NOTE: You must enable it if you are using
											    # included pre-computed weighted embeddings
												token_type_weights={(128, 256): 2, # Pitches weight
												     				 (384, 718): 2  # Chords weight
																	},
                                                verbose=False,
                                                show_progress_bar=False
                                               )
    
        # Calculate cosine similarity between source/query MIDI embeddings and embeddings corpus
        idxs, sims = midisimx.cosine_similarity_topk(query_emb,
													corpus_emb,
													verbose=False
												   )
       
        # ================================================================================================
        # Processs, print and save results
        # ================================================================================================
         
        # Convert the results to sorted list with transpose values
        idxs_sims_tvs_list = midisimx.idxs_sims_to_sorted_list(idxs, sims)
       
        # Print corpus matches (and optionally) convert the final result to a handy list for further processing
        corpus_matches_list = midisimx.print_sorted_idxs_sims_list(idxs_sims_tvs_list,
                                                                  corpus_midi_names,
                                                                  return_as_list=True
                                                                 )
         
        # ================================================================================================
        # Copy matched MIDIs from the MIDI corpus for listening and further evaluation and analysis
        # ================================================================================================
        
        # Copy matched corpus MIDI to a desired directory for easy evaluation and analysis
        out_dir_path = midisimx.copy_corpus_files(corpus_matches_list,
                                                 corpus_midis_dirs=['./Discover-MIDI-Dataset/MIDIs/'],
                                                 main_output_dir='Output-MIDI-Dataset',
                                                 sub_output_dir=os.path.splitext(os.path.basename(fa))[0],
                                                 verbose=False
                                                )
        # ================================================================================================

---

MIDI Representation Encoding

midisimx uses a compact, event‑structured token format that lets the model understand timing, harmony, melody, and rhythm with minimal overhead.

Each event is encoded in a strict order, and notes and chords share the same structure—chords simply contain multiple pitch–duration pairs.

Event Structure

Notes

A note event always has four tokens:

[delta‑start, note-token, pitch, duration]

Chords

A chord event starts with the same two tokens, but then includes multiple (pitch, duration) pairs:

[delta‑start, chord-token, pitch, duration, pitch, duration, pitch, duration, ...]

This allows encoding triads, extended chords, clusters, or any multi‑note harmony.

Sample Encoded Sequence

Below is a real midisimx token sequence excerpt, formatted for readability.
Events are grouped to show how notes and chords appear:

[0, 643, 193, 321]  
[186, 321, 179, 325]  
[16, 391, 195, 265]  
[9, 391, 195, 298]  
[16, 387, 191, 272]  
[16, 387, 191, 266]  
[8, 689, 193, 323, 186, 321]        ← chord (two pitch–duration pairs)
[1, 386, 178, 323]  
[15, 391, 195, 265]  
[9, 391, 195, 298]  
[16, 387, 191, 283]  
[24, 711, 196, 321, 186, 321]        ← chord
[1, 384, 176, 320]  
[15, 391, 195, 296]  
[9, 389, 193, 297]  
[23, 387, 191, 273]  
[8, 391, 195, 315]  
[9, 707, 183, 323, 174, 324]         ← chord
[50, 391, 195, 274]  
...

You can clearly see:

• Delta‑times drive the rhythm
• Chord tokens (≥396) introduce multi‑pitch structures
• Single notes → 4 tokens
• Chords → 2 + (pitch, duration) × N tokens

---

Documentation

midisimx API Reference

midisimx API Functions Index

Legacy midisimx API Reference

---

Project Structure

midisimx/                                   # Project root
├── LICENSE                                 # Apache-2.0 license text
├── MANIFEST.in                             # Setuptools manifest — package-data inclusion rules for sdist/wheel
├── README.md                               # Main project README — features, usage guides, links, citations
├── midisimx/                               # The installable Python package
│   ├── API_REFERENCE.md                    # This document — complete public API reference
│   ├── MIDI.py                             # LEGACY — original parent of TMIDIX; unused, kept for reference/posterity
│   ├── README.md                           # Package README (PyPI landing page)
│   ├── TMIDIX.py                           # TMIDIX MIDI parsing/processing suite; re-exported as midisimx.TMIDIX
│   ├── artwork/                            # Project images
│   │   ├── Project-Los-Angeles.png         # Project Los Angeles logo
│   │   ├── README.md                       # Artwork notes and credits
│   │   ├── Tegridy-Code-2026.png           # Tegridy Code 2026 branding image
│   │   └── midisimx.png                    # Project banner (embedded in READMEs)
│   ├── crossmodal_mapper.py                # Non-ML closed-form bi-directional cross-modal embedding mapper (procrustes/ridge/cca)
│   ├── embeddings/                         # Bundled pre-computed embeddings
│   │   ├── README.md                       # Notes on bundled embeddings sets
│   │   └── lakh_midi_dataset_17209......   # Tiny 128-dim weighted (1-2-1-2) embeddings for 17 209 clean LAKH MIDIs — pairs with the bundled tiny model
│   ├── helpers.py                          # Utilities — bundled assets listing, MIDI normalization, file hashing, apt install
│   ├── instrumentation_similarity.py       # Deterministic timbre-aware GM instrumentation similarity scoring
│   ├── ldmb.py                             # LDMB — mmap-backed binary storage for large lists of dicts (lazy reads, byte-level merges)
│   ├── memmap.py                           # Single-file memmap storage for paired names + float32 embeddings
│   ├── midi_to_colab_audio.py              # AUX (optional) — renders MIDIs to audio via fluidsynth + SF2 soundfont banks
│   ├── midisimx.py                         # CORE — model/embeddings I/O, MIDI↔tokens, embedding computation, similarity search
│   ├── models/                             # Bundled model checkpoints
│   │   ├── README.md                       # Notes on bundled models
│   │   └── midisimx_tiny_trained_model...  # Tiny 6.49M-param Transformer checkpoint (14 401 steps · 0.5146 loss · 0.8202 acc)
│   ├── pca_reduce.py                       # Streaming, GPU-accelerated PCA reduction (PCAReductor, PCAReductionResult)
│   └── x_transformer_2_3_1.py              # CORE (models) — vendored, stand-alone x-transformers v2.3.1 by lucidrains
└── pyproject.toml                          # PEP 621 packaging metadata — version, dependencies, PyPI URLs, classifiers

Legend:

• CORE — required by the main similarity pipeline.
• AUX — optional convenience module; requires fluidsynth (installable via midisimx.helpers.install_apt_package('fluidsynth')) and SF2 banks; audio rendering only.
• LEGACY — not used anywhere in the project; provided for reference, convenience, and posterity.
• The vendored x_transformer_2_3_1.py makes the core pipeline independent of the PyPI x-transformers package (which is only needed for raw/custom tasks).

---

Limitations

• Current code and models support only MIDI music elements similarity (start-times, durations, pitches and chords)
• MIDI channels and velocities are not currently supported due to practicality considerations
• Current model is limited by 3k sequence length (~1000 MIDI music notes) so long-running MIDIs can only be analyzed in chunks

---

Citations

@misc{project_los_angeles_2026,
	author       = { Project Los Angeles and Tegridy Code },
	title        = { midisimx (Revision cfed861) },
	year         = 2026,
	url          = { https://huggingface.co/projectlosangeles/midisimx },
	doi          = { 10.57967/hf/10032 },
	publisher    = { Hugging Face }
}

@misc{project_los_angeles_2026,
	author       = { Project Los Angeles and Tegridy Code },
	title        = { midisimx-embeddings (Revision 0af7bbc) },
	year         = 2026,
	url          = { https://huggingface.co/datasets/projectlosangeles/midisimx-embeddings },
	doi          = { 10.57967/hf/10082 },
	publisher    = { Hugging Face }
}

@misc{project_los_angeles_2026,
	author       = { Project Los Angeles and Tegridy Code },
	title        = { midisimx-samples (Revision 3c28df7) },
	year         = 2026,
	url          = { https://huggingface.co/datasets/projectlosangeles/midisimx-samples },
	doi          = { 10.57967/hf/10085 },
	publisher    = { Hugging Face }
}

@misc{project_los_angeles_2025,
	author       = { Project Los Angeles },
	title        = { Discover-MIDI-Dataset (Revision 0eaecb5) },
	year         = 2025,
	url          = { https://huggingface.co/datasets/projectlosangeles/Discover-MIDI-Dataset },
	doi          = { 10.57967/hf/7361 },
	publisher    = { Hugging Face }
}

@phdthesis{raffel2016learning,
  author       = { Colin Raffel },
  title        = { Learning-Based Methods for Comparing Sequences, with Applications to Audio-to-{MIDI} Alignment and Matching },
  school       = { Columbia University },
  year         = { 2016 },
  url          = { https://colinraffel.com/projects/lmd/ }
}

---

Project Los Angeles

Tegridy Code 2026