Spaces:
Running
Running
Commit
·
4059958
1
Parent(s):
8a83a8e
add preset selection dropdown, fix incorrect x2z and json not updated
Browse files
app.py
CHANGED
@@ -47,6 +47,8 @@ CONFIG_PATH = "presets/rt_config.yaml"
|
|
47 |
PCA_PARAM_FILE = "presets/internal/gaussian.npz"
|
48 |
INFO_PATH = "presets/internal/info.json"
|
49 |
MASK_PATH = "presets/internal/feature_mask.npy"
|
|
|
|
|
50 |
|
51 |
|
52 |
with open(CONFIG_PATH) as fp:
|
@@ -56,16 +58,21 @@ with open(CONFIG_PATH) as fp:
|
|
56 |
global_fx = instantiate(fx_config)
|
57 |
global_fx.eval()
|
58 |
|
|
|
|
|
|
|
|
|
|
|
59 |
pca_params = np.load(PCA_PARAM_FILE)
|
60 |
mean = pca_params["mean"]
|
61 |
cov = pca_params["cov"]
|
62 |
eigvals, eigvecs = np.linalg.eigh(cov)
|
63 |
eigvals = np.flip(eigvals, axis=0)
|
64 |
eigvecs = np.flip(eigvecs, axis=1)
|
65 |
-
|
66 |
-
U = torch.from_numpy(
|
67 |
mean = torch.from_numpy(mean).float()
|
68 |
-
|
69 |
# Global latent variable
|
70 |
# z = torch.zeros_like(mean)
|
71 |
|
@@ -104,7 +111,7 @@ meter = pyln.Meter(44100)
|
|
104 |
def z2x(z):
|
105 |
# close all figures to avoid too many open figures
|
106 |
plt.close("all")
|
107 |
-
x = U @ z + mean
|
108 |
# # print(z)
|
109 |
# fx.load_state_dict(vec2dict(x), strict=False)
|
110 |
# fx.apply(partial(clip_delay_eq_Q, Q=0.707))
|
@@ -123,7 +130,7 @@ def fx2x(fx):
|
|
123 |
@torch.no_grad()
|
124 |
def x2z(x):
|
125 |
z = U.T @ (x - mean)
|
126 |
-
return z
|
127 |
|
128 |
|
129 |
@torch.no_grad()
|
@@ -393,13 +400,23 @@ with gr.Blocks() as demo:
|
|
393 |
|
394 |
sliders = [s1, s2, s3, s4]
|
395 |
|
396 |
-
|
397 |
-
|
398 |
-
|
399 |
-
|
400 |
-
|
401 |
-
|
402 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
403 |
|
404 |
with gr.Column():
|
405 |
audio_output = default_audio_block(label="Output Audio", interactive=False)
|
@@ -853,8 +870,15 @@ with gr.Blocks() as demo:
|
|
853 |
t60_plot,
|
854 |
]
|
855 |
|
856 |
-
update_all =
|
857 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
858 |
|
859 |
random_button.click(
|
860 |
chain_functions(
|
@@ -912,4 +936,14 @@ with gr.Blocks() as demo:
|
|
912 |
outputs=extra_slider,
|
913 |
)
|
914 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
915 |
demo.launch()
|
|
|
47 |
PCA_PARAM_FILE = "presets/internal/gaussian.npz"
|
48 |
INFO_PATH = "presets/internal/info.json"
|
49 |
MASK_PATH = "presets/internal/feature_mask.npy"
|
50 |
+
PRESET_PATH = "presets/internal/raw_params.npy"
|
51 |
+
TRAIN_INDEX_PATH = "presets/internal/train_index.npy"
|
52 |
|
53 |
|
54 |
with open(CONFIG_PATH) as fp:
|
|
|
58 |
global_fx = instantiate(fx_config)
|
59 |
global_fx.eval()
|
60 |
|
61 |
+
raw_params = torch.from_numpy(np.load(PRESET_PATH))
|
62 |
+
train_index = torch.from_numpy(np.load(TRAIN_INDEX_PATH))
|
63 |
+
feature_mask = torch.from_numpy(np.load(MASK_PATH))
|
64 |
+
presets = raw_params[train_index][:, feature_mask].contiguous()
|
65 |
+
|
66 |
pca_params = np.load(PCA_PARAM_FILE)
|
67 |
mean = pca_params["mean"]
|
68 |
cov = pca_params["cov"]
|
69 |
eigvals, eigvecs = np.linalg.eigh(cov)
|
70 |
eigvals = np.flip(eigvals, axis=0)
|
71 |
eigvecs = np.flip(eigvecs, axis=1)
|
72 |
+
eigsqrt = torch.from_numpy(eigvals.copy()).float().sqrt()
|
73 |
+
U = torch.from_numpy(eigvecs.copy()).float()
|
74 |
mean = torch.from_numpy(mean).float()
|
75 |
+
|
76 |
# Global latent variable
|
77 |
# z = torch.zeros_like(mean)
|
78 |
|
|
|
111 |
def z2x(z):
|
112 |
# close all figures to avoid too many open figures
|
113 |
plt.close("all")
|
114 |
+
x = U @ (z * eigsqrt) + mean
|
115 |
# # print(z)
|
116 |
# fx.load_state_dict(vec2dict(x), strict=False)
|
117 |
# fx.apply(partial(clip_delay_eq_Q, Q=0.707))
|
|
|
130 |
@torch.no_grad()
|
131 |
def x2z(x):
|
132 |
z = U.T @ (x - mean)
|
133 |
+
return z / eigsqrt
|
134 |
|
135 |
|
136 |
@torch.no_grad()
|
|
|
400 |
|
401 |
sliders = [s1, s2, s3, s4]
|
402 |
|
403 |
+
with gr.Row():
|
404 |
+
with gr.Column():
|
405 |
+
extra_pc_dropdown = gr.Dropdown(
|
406 |
+
list(range(NUMBER_OF_PCS + 1, mean.numel() + 1)),
|
407 |
+
label=f"PC > {NUMBER_OF_PCS}",
|
408 |
+
info="Select which extra PC to adjust",
|
409 |
+
interactive=True,
|
410 |
+
)
|
411 |
+
extra_slider = default_pc_slider(label="Extra PC")
|
412 |
+
|
413 |
+
preset_dropdown = gr.Dropdown(
|
414 |
+
["none"] + list(range(1, presets.shape[0] + 1)),
|
415 |
+
value="none",
|
416 |
+
label=f"Select Preset (1-{presets.shape[0]})",
|
417 |
+
info="Select a preset to load (this will override the current settings)",
|
418 |
+
interactive=True,
|
419 |
+
)
|
420 |
|
421 |
with gr.Column():
|
422 |
audio_output = default_audio_block(label="Output Audio", interactive=False)
|
|
|
870 |
t60_plot,
|
871 |
]
|
872 |
|
873 |
+
update_all = (
|
874 |
+
lambda z, fx, i: update_pc(z, i)
|
875 |
+
+ update_fx(fx)
|
876 |
+
+ update_plots(fx)
|
877 |
+
+ [model2json(fx)]
|
878 |
+
)
|
879 |
+
update_all_outputs = (
|
880 |
+
update_pc_outputs + update_fx_outputs + update_plots_outputs + [json_output]
|
881 |
+
)
|
882 |
|
883 |
random_button.click(
|
884 |
chain_functions(
|
|
|
936 |
outputs=extra_slider,
|
937 |
)
|
938 |
|
939 |
+
preset_dropdown.input(
|
940 |
+
chain_functions(
|
941 |
+
lambda i, _: (mean if i == "none" else presets[i - 1], _),
|
942 |
+
lambda x, i: (x2z(x), x, vec2fx(x), i),
|
943 |
+
lambda z, x, fx, i: [z, x] + update_all(z, fx, i),
|
944 |
+
),
|
945 |
+
inputs=[preset_dropdown, extra_pc_dropdown],
|
946 |
+
outputs=[z, fx_params] + update_all_outputs,
|
947 |
+
)
|
948 |
+
|
949 |
demo.launch()
|