Spaces:
Running
on
Zero
Running
on
Zero
File size: 4,903 Bytes
9a6dac6 c7710a2 9a6dac6 c7710a2 9a6dac6 c7710a2 9a6dac6 c7710a2 9a6dac6 c7710a2 9a6dac6 92fe890 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 |
import io
import copy
import torch.nn.functional as F
from torchvision.transforms.functional import pil_to_tensor, to_pil_image
import matplotlib.pyplot as plt
from PIL import Image
from src.audio_morphix import AudioMorphix
from src.utils.factory import plot_spectrogram, get_edit_mask
from src.utils.audio_processing import maybe_add_dimension
DESPLAY_RES = (1600, 900)
SPEC_RES = (1024, 64)
N_SAMPLE_PER_SEC = 100 # 1024frames / 10.24s
def func_clear(*args):
result = []
for arg in args:
if isinstance(arg, list):
result.append([])
else:
result.append(None)
return tuple(result)
def create_model(model_type):
model = AudioMorphix(pretrained_model_path=model_type, device="cpu")
return model
def process_audio(model, audio, config):
fbank, log_stft, wav = model.editor.get_fbank(
audio,
config.audio_processor,
return_intermediate=True,
)
fbank = maybe_add_dimension(fbank, 4)
# Generate spectrogram plot
spec_plot = plot_spectrogram(
fbank.permute(0, 1, 3, 2)[:,:,:,:10*N_SAMPLE_PER_SEC], auto_amp=True)
return fbank, spec_plot
def get_spec_pil(model, audio, config):
try:
fbank, spec_plot = process_audio(model, audio, config)
buf = io.BytesIO()
spec_plot.figure.savefig(buf, format='png')
buf.seek(0)
pil_spec = Image.open(buf)
plt.close()
except:
print("Warning: the streaming is not ready. Please repeate uploading again.")
fbank, pil_spec = None, None
return fbank, pil_spec
def get_spec_pil_with_original(model, audio, config):
fbank, pil_spec = get_spec_pil(model, audio, config)
pil_spec_ori = copy.deepcopy(pil_spec)
return fbank, pil_spec, pil_spec_ori
def get_spec_pils_for_moving(model, audio, config):
src_fbank, src_pil_spec = get_spec_pil(model, audio, config)
ref_fbank, ref_pil_spec = copy.deepcopy(src_fbank), copy.deepcopy(src_pil_spec)
ref_pil_spec_ori = copy.deepcopy(ref_pil_spec)
return src_fbank, src_pil_spec, ref_fbank, ref_pil_spec, ref_pil_spec_ori
def get_mask_region(img):
layers = img['layers']
if len(layers) > 0:
print("Warning: Multiple layers exist while only the first layer is considered as the mask.")
# Use the channel of opacity as mask
mask = pil_to_tensor(layers[0])[-1,:,:] # RGBA
mask = mask.permute(1, 0) # (F, T) -> (T, F)
# Flip the freq axis to ensure the orignal point on the top left
mask = mask.flip(1)
mask = (mask > 0).float()
# Rescale mask to spectrum size
mask = F.interpolate(mask.unsqueeze(0).unsqueeze(0), SPEC_RES).squeeze()
return mask
def get_mask_regions(img):
def _prepare_mask(m):
m = m.permute(1, 0)
# Flip the freq axis to ensure the orignal point on the top left
m = m.flip(1)
m = (m > 0).float()
m = F.interpolate(m.unsqueeze(0).unsqueeze(0), SPEC_RES).squeeze()
return m
layers = img['layers']
if len(layers) > 0:
print("Warning: Multiple layers exist while the first layer is considered as the mask to edit and the second is the mask to keep.")
if len(layers) > 1:
mask_src = pil_to_tensor(layers[0])[-1,:,:] # RGBA
mask_keep = pil_to_tensor(layers[1])[-1,:,:]
mask_src, mask_keep = _prepare_mask(mask_src), _prepare_mask(mask_keep)
elif len(layers) == 1:
mask_src = pil_to_tensor(layers[0])[-1,:,:]
mask_src = _prepare_mask(mask_src)
mask_keep = None
else:
mask_src, mask_keep = None, None
return mask_src, mask_keep
def update_reference_spec(ref_spec_pil_ori, mask_src, dt, df, resize_scale_t, resize_scale_f):
if mask_src is not None:
mask_ref = get_edit_mask(
mask_src, dx=df, dy=dt,
resize_scale_x=resize_scale_f,
resize_scale_y=resize_scale_t,
)
mask_ref = mask_ref.float() # match the PIL format, channel last
mask_ref_pil = F.interpolate(mask_ref.unsqueeze(0).unsqueeze(0), DESPLAY_RES).squeeze()
# Match the shape to the PIL format (H, W, C)
if mask_ref_pil.ndim > 2:
mask_ref_pil = mask_ref_pil.squeeze()
mask_ref_pil = mask_ref_pil.permute(1, 0)
# De-flip freq exis to match pil imshow style
mask_ref_pil = mask_ref_pil.flip(0)
mask_ref_pil = mask_ref_pil * 0.5 # for transparency
# Convert to PIL
mask_ref_pil = to_pil_image(mask_ref_pil).convert("L")
# mask_ref_pil = mask_ref_pil.resize(ref_spec_pil_ori.size)
overlay = Image.new("RGBA", mask_ref_pil.size, (128, 255, 255, 50)) # create overlay
ref_spec_pil = Image.composite(overlay, ref_spec_pil_ori, mask_ref_pil)
else:
ref_spec_pil = ref_spec_pil_ori
mask_ref = None
return mask_ref, ref_spec_pil |