Rob Jaret
Merge branch 'main' of https://huggingface.co/spaces/rjaret/rma2
198141f
import huggingface_hub
#
# paths to various models
model_path_configs = {
"Humpback Whales": ("Intelligent-Instruments-Lab/rave-models", "humpbacks_pondbrain_b2048_r48000_z20.ts"),
"Magnets": ("Intelligent-Instruments-Lab/rave-models", "magnets_b2048_r48000_z8.ts"),
"Big Ensemble": ("Intelligent-Instruments-Lab/rave-models", "crozzoli_bigensemblesmusic_18d.ts"),
"Bird Dawn Chorus": ("Intelligent-Instruments-Lab/rave-models", "birds_dawnchorus_b2048_r48000_z8.ts"),
"Speaking & Singing": ("Intelligent-Instruments-Lab/rave-models", "voice-multi-b2048-r48000-z11.ts"),
"Resonator Piano": ("Intelligent-Instruments-Lab/rave-models", "mrp_strengjavera_b2048_r44100_z16.ts"),
"Multimbral Guitar": ("Intelligent-Instruments-Lab/rave-models", "guitar_iil_b2048_r48000_z16.ts"),
"Organ Archive": ("Intelligent-Instruments-Lab/rave-models", "organ_archive_b2048_r48000_z16.ts"),
"Water": ("Intelligent-Instruments-Lab/rave-models", "water_pondbrain_b2048_r48000_z16.ts"),
"Brass Sax": ("shuoyang-zheng/jaspers-rave-models", "aam_brass_sax_b2048_r44100_z8_noncausal.ts"),
"Speech": ("shuoyang-zheng/jaspers-rave-models", "librispeech100_b2048_r44100_z8_noncausal.ts"),
"String": ("shuoyang-zheng/jaspers-rave-models" ,"aam_string_b2048_r44100_z16_noncausal.ts"),
"Singer": ("shuoyang-zheng/jaspers-rave-models","gtsinger_b2048_r44100_z16_noncausal.ts"),
"Bass": ("shuoyang-zheng/jaspers-rave-models","aam_bass_b2048_r44100_z16_noncausal.ts"),
"Drum": ("shuoyang-zheng/jaspers-rave-models","aam_drum_b2048_r44100_z16_noncausal.ts"),
"Gtr Picking": ("shuoyang-zheng/jaspers-rave-models","guitar_picking_dm_b2048_r44100_z8_causal.ts"),
}
available_audio_files=[
"SilverCaneAbbey-Voices.wav",
"Chimes.wav",
"FrenchChildren.wav",
"Organ-ND.wav",
"SpigotsOfChateauLEtoge.wav",
"GesturesPercStrings.wav",
"SingingBowl-OmniMic.wav",
"BirdCalls.mp3",
]
model_path_config_keys = sorted(model_path_configs)
model_paths_cache = {}
def GetModelPath(model_path_name):
model_path = ()
if model_path_name in model_paths_cache.keys():
model_path = model_paths_cache[model_path_name]
else:
repo_id, filename = model_path_configs[model_path_name]
model_path = huggingface_hub.hf_hub_download(
repo_id =repo_id,
filename = filename,
cache_dir="../huggingface_hub_cache",
force_download=False,
)
print(f"Generated Model Path for {filename}.")
model_paths_cache[model_path_name] = model_path
return model_path
def saveAudio(file_path, audio):
with open(file_path + '.wav', 'wb') as f:
f.write(audio.data)
import torch
import pandas as pd
import copy
import librosa
import ast
import os
def AverageRaveModels(rave_a, rave_b, bias = 0):
r1_ratio = .5
r2_ratio = .5
messages = {}
# bias between -1 and 1
if abs(bias) <= 1:
if bias > 0:
r1_ratio = .5 + bias/2
r2_ratio = 1.0 - r1_ratio
rave_temp = rave_a
elif bias < 0:
r2_ratio = .5 + abs(bias)/2
r1_ratio = 1.0 - r2_ratio
else:
print(f"Unable to apply bias {bias} - bias must be between -1 and 1.")
# Get state dictionaries of both models
rave_a_params = rave_a.state_dict()
rave_b_params = rave_b.state_dict()
# intialize the averaged rave with model_a
rave_avg = copy.deepcopy(rave_a)
avg = rave_avg.state_dict()
# for reporting
keys_averaged={}
keys_not_averaged={}
for key in rave_a_params:
if key in rave_b_params:
try:
avg[key] = ((rave_a_params[key] * r1_ratio) + (rave_b_params[key] * r2_ratio))
keys_averaged[key]=(key, rave_a_params[key].shape, rave_b_params[key].shape, "")
except Exception as e:
print(f"Error averaging key {key}: {e}")
keys_not_averaged[key]=(key, rave_a_params[key].shape, rave_b_params[key].shape, e)
else:
print(f"Key {key} not found in rave_b parameters, skipping.")
# keys_not_averaged(key)
keys_not_averaged[key]=(key, rave_a_params[key].shape, "n/a", "Key not found in rave_b parameters.")
messages["keys_averaged"] = keys_averaged
messages["keys_not_averaged"] = keys_not_averaged
messages["stats"] = f'Numb Params Averaged: {len(keys_averaged)}\nNumb Params Unable to Average: {len(keys_not_averaged)}\nPercent Averaged: {len(keys_averaged) * 100/(len(keys_not_averaged) + len(keys_averaged)):5.2f}%'
# Commit the changes
rave_avg.load_state_dict(avg)
return rave_avg, messages
def ProcessRequest(model_name_a, model_name_b, audio_file_name, audio_file, sr_multiple=1, bias=0, function="AverageModels"):
if function == "AverageModels":
return GenerateRaveEncDecAudio(model_name_a, model_name_b, audio_file_name, audio_file, sr_multiple, bias)
elif function == "StyleTransfer":
print("Style Transfer not implemented yet.")
return None
def GenerateStyleTransfer(model_name_a, model_name_b, audio_file_name, audio_file, sr_multiple=1):
model_path_a = GetModelPath(model_name_a)
model_path_b = GetModelPath(model_name_b)
# Choose Audio File to encode/decode
if audio_file is None:
audio_file = os.path.join('assets', audio_file_name)
# print("Audio File Name:", audio_file_name)
# Generate Audio Files
# Audio files are created in the assets folder
generate_audio_files = False
rave_a = torch.jit.load(model_path_a)
rave_b = torch.jit.load(model_path_b)
# Let's load a sample audio file
y, sr = librosa.load(audio_file)
sr_multiplied = sr * sr_multiple # Adjust sample rate if needed
print(f"Audio File Loaded: {audio_file}, sample_rate = {sr}")
# Convert audio to a PyTorch tensor and reshape it to the
# required shape: (batch_size, n_channels, n_samples)
audio = torch.from_numpy(y).float()
audio = audio.reshape(1, 1, -1)
messages={}
audio_outputs={}
# perform style transfer
with torch.no_grad():
# encode the audio with the new averaged models
try:
latent = rave_a.encode(audio)
# decode individual and averaged models
decoded = rave_a.decode(latent)
style_transfer_decoded = rave_b.decode(latent)
audio_outputs['decoded'] = decoded
audio_outputs['style_transfer'] = style_transfer_decoded
except Exception as e:
print(f'Encoding process generated error: ', e)
model_a_file=model_path_a.rsplit("/")[-1]
model_b_file=model_path_b.rsplit("/")[-1]
# Original Audio
original_audio = (sr, y)
# Decoded Audio
print("Encoded and Decoded using original models")
decoded_audio = (sr, decoded.detach().numpy().squeeze())
style_transfer_audio = (sr, style_transfer_decoded.detach().numpy().squeeze())
# saveAudio('assets/' + model_a_file[: 7] + '_enc-dec.wav', a)
# saveAudio('assets/' model_a_file[: 7] + '-' model_b_file[: 7] + '_style_transfer.wav', a)
messages["stats"] = f"Model A: {model_name_a}\nModel B: {model_name_b}\nAudio file: {os.path.basename(audio_file)}\nSample Rate Multiple for Style Transfer Version: {sr_multiple}\n\n" + messages["stats"]
return original_audio, decoded_audio, style_transfer_audio, '', messages["stats"], '', ''
def GenerateRaveEncDecAudio(model_name_a, model_name_b, audio_file_name, audio_file, sr_multiple=1, bias=0):
###############################################
# Choose models from filenames dictionary created in previous cell
# Note: model_path_a is always used to initialize the averaged model.
# Switching them gets different results if the parameters are not all matched.
###############################################
# Examples - this matches only 21 params, but it sounds like maybe sosme of both are in the result.
model_path_a = GetModelPath(model_name_a)
model_path_b = GetModelPath(model_name_b)
#####################################
# Set biases between -1 and 1 to bias the result towards one of the models
# 0 = no bias; >0 biased towards model_a; <0 = biased towards model_b
#####################################
# Note: multiple biases not implemented for gradio version
biases=[bias]
####################################
# Choose Audio File to encode/decode
#####################################
# audio_file_name = "RJM1240-Gestures.wav"
if audio_file is None:
audio_file = os.path.join('assets', audio_file_name)
# print("Audio File Name:", audio_file_name)
####################################
# Generate Audio Files
# Audio files are created in the assets folder
generate_audio_files = False
rave_a = torch.jit.load(model_path_a)
rave_b = torch.jit.load(model_path_b)
# Let's load a sample audio file
y, sr = librosa.load(audio_file)
sr_multiplied = sr * sr_multiple # Adjust sample rate if needed
print(f"Audio File Loaded: {audio_file}, sample_rate = {sr}")
# Convert audio to a PyTorch tensor and reshape it to the
# required shape: (batch_size, n_channels, n_samples)
audio = torch.from_numpy(y).float()
audio = audio.reshape(1, 1, -1)
messages={}
audio_outputs={}
for bias in biases:
# Average the rave models
# rave_avg, numb_params_mod, numb_params_unable_to_mod = AverageRaveModels(rave_a, rave_b, bias=bias)
rave_avg, new_msgs = AverageRaveModels(rave_a, rave_b, (-1 * bias))
messages |= new_msgs
# no decode the results back to audio
with torch.no_grad():
# encode the audio with the new averaged models
try:
latent_a = rave_a.encode(audio)
latent_b = rave_b.encode(audio)
latent_avg = rave_avg.encode(audio)
# decode individual and averaged models
decoded_a = rave_a.decode(latent_a)
decoded_b = rave_b.decode(latent_b)
decoded_avg = rave_avg.decode(latent_avg)
audio_outputs[bias] = decoded_avg[0]
except:
print(f'Bias {bias} generated an error. Removing it from list of biases.')
biases.remove(bias)
# print(biases)
model_a_file=model_path_a.rsplit("/")[-1]
model_b_file=model_path_b.rsplit("/")[-1]
# Original Audio
original_audio = (sr, y)
# Decoded Audio
print("Encoded and Decoded using original models")
model_a_audio = (sr, decoded_a[0].detach().numpy().squeeze())
# saveAudio('assets/' + model_a_file[: 7] + '_only.wav', a)
model_b_audio = (sr, decoded_b[0].detach().numpy().squeeze())
# # saveAudio('assets/' + model_b_file[: 7] + '_only.wav', a)
print("Encoded and Decoded using Averaged Models")
print("with Biases: ", biases)
print("\nNumber of params able to average:", len(messages["keys_averaged"]))
print("Number of params unable to average:", len(messages["keys_not_averaged"]))
output_file_prefix = f'assets/{model_a_file[: 7]}-{model_b_file[: 7]}_'
bias = biases[0]
averaged_audio = (sr_multiplied, audio_outputs[bias].detach().numpy().squeeze())
df_averaged = pd.DataFrame(messages['keys_averaged']).transpose() #reset_index(names='Param Key')
df_averaged.columns=['Param Name', 'Model A Shape', 'Model B Shape', 'Notes']
df_not_averaged = pd.DataFrame(messages["keys_not_averaged"]).transpose()
# case when all params are averaged
if len(df_not_averaged.columns) == 0:
data = {'Param Name': [], 'Modeal A Shape': [], 'Model B Shape': [], 'Notes': []}
df_not_averaged = pd.DataFrame(data)
df_not_averaged.columns=['Param Name', 'Model A Shape', 'Model B Shape', 'Notes']
messages["stats"] = f"Model A: {model_name_a}\nModel B: {model_name_b}\nAudio file: {os.path.basename(audio_file)}\nSample Rate Multiple for Averaged Version: {sr_multiple}\n\n" + messages["stats"]
return original_audio, model_a_audio, model_b_audio, averaged_audio, messages["stats"], df_averaged, df_not_averaged
import gradio as gr
column_widths=['35%', '20%', '20%', '25%']
waveform_options = gr.WaveformOptions(waveform_color="#01C6FF",
waveform_progress_color="#0066B4",
skip_length=2,)
description = "<p style='line-height: 1'>This app attempts to average two RAVE models and then encode and decode an audio file through the original and averaged models.</p>" \
"<ul style='padding-bottom: 0px; padding-top: 5px;'>Instructions:<li style='line-height: .8; padding-top: 5px; padding-bottom: 0px;'>Select the two models from the lists of pre-trained RAVE models.</li>" \
"<li style='padding-top: 0px; line-height: .8'>Select an audio file from the ones available in the dropdown, or upload an audio file of up to 60 seconds. Click 'Submit' at the bottom of the page.</li></ul>"\
"<p style='line-height: 1.2; padding-top: 0px; margin-top: 10px;'>Note that in most cases not all parameters can be averaged. They may not exist in both models or the two values may not have the same shape. The data sets in the output list which ones were and weren't averaged with their shapes and any notes. (You can copy them into a spreadsheet by clicking the icon at the top right corner of each.)</p>"
"<!-- <li>Select a sample rate multiple for the averaged model. When there is a useful result, it sometimes sounds better at double the sample rate.</li>" \
"<li>Select a bias towards one of the models. A bias of 0 will average the two models equally. A positive bias will favor Model A, and a negative bias will favor Model B.</li></ul>" \
"-->"
AverageModels = gr.Interface(title="Process Audio Through the Average of Two Rave Models", description=description,
fn=GenerateRaveEncDecAudio,
inputs=[
gr.Radio(model_path_config_keys, label="Select Model A", value="Multimbral Guitar", container=True),
gr.Radio(model_path_config_keys, label="Select Model B", value="Water", container=True),
gr.Dropdown(available_audio_files, label="Select from these audio files or upload your own below:", value="SingingBowl-OmniMic.wav",container=True),
gr.Audio(label="Upload an audio file (wav)", type="filepath", sources=["upload", "microphone"], max_length=60,
waveform_options=waveform_options, format='wav'),],
additional_inputs=[
gr.Radio([.2, .5, .75, 1, 2, 4], label="Sample Rate Multiple (Averaged version only)", value=1, container=True),
gr.Slider(label="Bias towards Model A or B", minimum=-1, maximum=1, value=0, step=0.1, container=True),
],
# if no way to pass dictionary, pass separate keys and values and zip them.
outputs=[
gr.Audio(label="Original Audio", sources=None, waveform_options=waveform_options, interactive=False),
gr.Audio(label="Encoded/Decoded through Model A", sources=None, waveform_options=waveform_options,),
gr.Audio(label="Encoded/Decoded through Model B", sources=None, waveform_options=waveform_options,),
gr.Audio(label="Encoded/Decoded through averaged model", sources=None, waveform_options=waveform_options,),
gr.Textbox(label="Info:"),
gr.Dataframe(label="Params Averaged", show_copy_button="True", scale=100, column_widths=column_widths, headers=['Param Name', 'Model A Shape', 'Model B Shape', 'Notes']),
gr.Dataframe(label="Params Not Averaged", show_copy_button="True", scale=100, column_widths=column_widths, headers=['Param Name', 'Model A Shape', 'Model B Shape', 'Notes'])
]
,fill_width=True
)
AverageModels.launch(max_file_size=10 * gr.FileSize.MB, share=True)