[fix] fix infer_params file extension bug, rename test audio

README.md

@@ -1,18 +1,19 @@
-# Syntheon
+![syntheon_logo](docs/syntheon-logo.png)
 
-Syntheon - the Pantheon (temple of gods) for music synthesizers.
+# Syntheon
 
-DX7 patches: https://yamahablackboxes.com/collection/yamaha-dx7-synthesizer/patches/
+Syntheon - [Pantheon]() for music synthesizers. 
 
-## Testing
+Syntheon aims to provide **parameter inference** for major music synthesizers using *deep learning models* to simplify sound design process. For now [Vital](https://vital.audio/) is supported, and [Dexed](https://asb2m10.github.io/dexed/) is on its way.
 
-```
-python3 -m inferencer.vital.vital_inferencer
+**Check out [this presentation](https://docs.google.com/presentation/d/1PA4fom6QvCW_YG8L0MMVumrAluljcymndNlaK2HW5t0/edit?usp=sharing) on the recent advances of synth parameter inference.
 
-python3 -m inferencer.dexed.dexed_inferencer
+## Installation
 
-python3 -m pytest
+Syntheon needs `python 3.9`. Clone this repo and install the dependencies using:
 
+```
+python3 -m pip install -r requirements.txt
 ```
 
 ## Usage
@@ -25,4 +26,32 @@ output_params_file, eval_dict = infer_params(
     "vital", 
     enable_eval=True
 )
-```
+```
+
+## Testing
+
+```
+python3 -m pytest
+```
+
+## Structure
+
+For each synthesizer, we need to define:
+
+- **converter** for preset format conversion: 
+    - `serializeToDict`: convert preset file to a Python dictionary to be handled by inferencer
+    - `parseToPluginFile`: convert Python dictionary back to preset file, to be loaded by the synthesizer
+
+- **inferencer** for model inference:
+    - `convert`: define the workflow of `load_model` -> `inference` -> `convert_to_preset`
+
+## Contribution
+
+Syntheon is actively under development, and contributions are welcomed. Some TODOs we have in mind include:
+
+- Replicating state-of-the-art approaches
+- Improving current model performance
+- Incorporating new synthesizers 
+- Code refactoring 😅
+
+This repo will only host the serving code, training code shall be released on a separate repo.

syntheon/converter/dexed/dexed_constants.py

@@ -3,6 +3,7 @@
 
 Credits to source: https://github.com/Nintorac/NeuralDX7/blob/master/scratch/dx7_constants.py
 DX7 specification: https://github.com/asb2m10/dexed/blob/master/Documentation/sysex-format.txt
+DX7 patches: DX7 patches: https://yamahablackboxes.com/collection/yamaha-dx7-synthesizer/patches/
 """
 import bitstruct
 

syntheon/inferencer/dexed/models/ddx7/loss_functions.py

@@ -1,7 +1,6 @@
 import ddx7.core as core
 import torch
 import torch.nn as nn
-import torchcrepe
 from functools import partial
 
 

syntheon/inferencer/inferencer.py

@@ -31,4 +31,7 @@ def inference(self, model, audio_fname):
         return NotImplementedError
 
     def convert_to_preset(self, inference_output):
+        """
+        Output a Python dictionary to be handled by the converter.
+        """
         return NotImplementedError

syntheon/inferencer/vital/models/adsr_envelope.py

@@ -1,5 +1,5 @@
 """
-ADSR envelope shaper
+Differentiable ADSR envelope shaper.
 """
 import numpy as np
 import torch
@@ -177,7 +177,7 @@ def forward(self,
 
         attack_ratio = attack_secs / total_secs
         decay_ratio = decay_secs / total_secs
-        # TODO: parameteize release_ratio
+        # TODO: parameterize release_ratio
         release_ratio = torch.tensor([0.]).repeat(attack_secs.size(0), 1, 1)
         if device == "cuda":
             release_ratio = release_ratio.cuda()
@@ -200,47 +200,6 @@ def get_amp_shaper(
                 sustain_level,
                 offsets=None):
     """
-    implement case with no offset first.
-    """
-    if offsets is None:
-        # if offset not specified, take next onset as offset
-        offsets = onsets[1:]
-        onsets = onsets[:len(onsets) - 1]
-
-    start_offset = int(onsets[0] * 100)        # TODO: 100 is block size
-    onsets, offsets = torch.tensor(onsets), torch.tensor(offsets)
-    if device == "cuda":
-        onsets, offsets = onsets.cuda(), offsets.cuda()
-    dur_vec = offsets - onsets
-    lst = []
-
-    # append zeros first before first onset
-    if device == "cuda":
-        lst.append(torch.zeros(start_offset).cuda())
-    else:
-        lst.append(torch.zeros(start_offset))
-
-    for dur in dur_vec:
-        dur = round(dur.item(), 2)
-        adsr = shaper(
-            attack_secs=torch.tensor([0.2, 0.1]), 
-            decay_secs=torch.tensor([0.1, 0.2]),
-            sustain_level=torch.tensor([0.9, 0.2]),
-            total_secs=dur)
-        lst.append(adsr[0].squeeze())   # TODO: fix the batch size case
-
-    final_signal = torch.cat(lst, dim=0)
-    return final_signal
-
-
-def get_amp_shaper_v2(
-                shaper, 
-                onsets, 
-                attack_secs,
-                decay_secs,
-                sustain_level,
-                offsets=None):
-    """
     implement case with no offset first. enable batches
     """
     if offsets is None: