update demo

README.md

@@ -1,6 +1,6 @@
 ---
 title: Musika
-emoji: 🚀
+emoji: 🎵
 colorFrom: purple
 colorTo: blue
 sdk: gradio

utils.py

@@ -34,11 +34,7 @@ class Utils_functions:
         )
         mel_f = tf.convert_to_tensor(librosa.mel_frequencies(n_mels=args.mel_bins + 2, fmin=0.0, fmax=args.sr // 2))
         enorm = tf.cast(
-            tf.expand_dims(
-                tf.constant(2.0 / (mel_f[2 : args.mel_bins + 2] - mel_f[: args.mel_bins])),
-                0,
-            ),
-            tf.float32,
+            tf.expand_dims(tf.constant(2.0 / (mel_f[2 : args.mel_bins + 2] - mel_f[: args.mel_bins])), 0,), tf.float32,
         )
         melmat = tf.multiply(melmat, enorm)
         melmat = tf.divide(melmat, tf.reduce_sum(melmat, axis=0))
@@ -165,8 +161,9 @@ class Utils_functions:
                 outls.append(model(x[i * bs : i * bs + bs], training=False))
 
         if dual_out:
-            return np.concatenate([outls[k][0] for k in range(len(outls))], 0), np.concatenate(
-                [outls[k][1] for k in range(len(outls))], 0
+            return (
+                np.concatenate([outls[k][0] for k in range(len(outls))], 0),
+                np.concatenate([outls[k][1] for k in range(len(outls))], 0),
             )
         else:
             return np.concatenate(outls, 0)
@@ -199,8 +196,9 @@ class Utils_functions:
             inp = tf.concat(inpls, 0)
             res = model(inp, training=False)
             outls.append(res)
-        return np.concatenate([outls[k][0] for k in range(len(outls))], 0), np.concatenate(
-            [outls[k][1] for k in range(len(outls))], 0
+        return (
+            np.concatenate([outls[k][0] for k in range(len(outls))], 0),
+            np.concatenate([outls[k][1] for k in range(len(outls))], 0),
         )
 
     def distribute_dec2(self, x, model, bs=64):
@@ -228,17 +226,7 @@ class Utils_functions:
         return tf.image.random_crop(noisetot, [1, self.args.latlen, 64 + 64])
 
     def generate_example_stereo(self, models_ls):
-        (
-            critic,
-            gen,
-            enc,
-            dec,
-            enc2,
-            dec2,
-            critic_rec,
-            gen_ema,
-            [opt_dec, opt_disc],
-        ) = models_ls
+        (critic, gen, enc, dec, enc2, dec2, critic_rec, gen_ema, [opt_dec, opt_disc],) = models_ls
         abb = gen_ema(self.get_noise_interp(), training=False)
         abbls = tf.split(abb, abb.shape[-2] // 16, -2)
         abb = tf.concat(abbls, 0)
@@ -247,13 +235,7 @@ class Utils_functions:
         for channel in range(2):
 
             ab = self.distribute_dec2(
-                abb[
-                    :,
-                    :,
-                    :,
-                    channel * self.args.latdepth : channel * self.args.latdepth + self.args.latdepth,
-                ],
-                dec2,
+                abb[:, :, :, channel * self.args.latdepth : channel * self.args.latdepth + self.args.latdepth,], dec2,
             )
             abls = tf.split(ab, ab.shape[-2] // self.args.shape, -2)
             ab = tf.concat(abls, 0)
@@ -291,28 +273,14 @@ class Utils_functions:
 
         fig, axs = plt.subplots(nrows=4, ncols=1, figsize=(20, 20))
         axs[0].imshow(
-            np.flip(
-                np.array(
-                    tf.transpose(
-                        self.wv2spec_hop((abwv[:, 0] + abwv[:, 1]) / 2.0, 80.0, 256),
-                        [1, 0],
-                    )
-                ),
-                -2,
-            ),
+            np.flip(np.array(tf.transpose(self.wv2spec_hop((abwv[:, 0] + abwv[:, 1]) / 2.0, 80.0, 256), [1, 0],)), -2,),
             cmap=None,
         )
         axs[0].axis("off")
         axs[0].set_title("Generated1")
         axs[1].imshow(
             np.flip(
-                np.array(
-                    tf.transpose(
-                        self.wv2spec_hop((abwv2[:, 0] + abwv2[:, 1]) / 2.0, 80.0, 256),
-                        [1, 0],
-                    )
-                ),
-                -2,
+                np.array(tf.transpose(self.wv2spec_hop((abwv2[:, 0] + abwv2[:, 1]) / 2.0, 80.0, 256), [1, 0],)), -2,
             ),
             cmap=None,
         )
@@ -320,13 +288,7 @@ class Utils_functions:
         axs[1].set_title("Generated2")
         axs[2].imshow(
             np.flip(
-                np.array(
-                    tf.transpose(
-                        self.wv2spec_hop((abwv3[:, 0] + abwv3[:, 1]) / 2.0, 80.0, 256),
-                        [1, 0],
-                    )
-                ),
-                -2,
+                np.array(tf.transpose(self.wv2spec_hop((abwv3[:, 0] + abwv3[:, 1]) / 2.0, 80.0, 256), [1, 0],)), -2,
             ),
             cmap=None,
         )
@@ -334,13 +296,7 @@ class Utils_functions:
         axs[2].set_title("Generated3")
         axs[3].imshow(
             np.flip(
-                np.array(
-                    tf.transpose(
-                        self.wv2spec_hop((abwv4[:, 0] + abwv4[:, 1]) / 2.0, 80.0, 256),
-                        [1, 0],
-                    )
-                ),
-                -2,
+                np.array(tf.transpose(self.wv2spec_hop((abwv4[:, 0] + abwv4[:, 1]) / 2.0, 80.0, 256), [1, 0],)), -2,
             ),
             cmap=None,
         )
@@ -351,26 +307,9 @@ class Utils_functions:
 
     # Save in training loop
     def save_end(
-        self,
-        epoch,
-        gloss,
-        closs,
-        mloss,
-        models_ls=None,
-        n_save=3,
-        save_path="checkpoints",
+        self, epoch, gloss, closs, mloss, models_ls=None, n_save=3, save_path="checkpoints",
     ):
-        (
-            critic,
-            gen,
-            enc,
-            dec,
-            enc2,
-            dec2,
-            critic_rec,
-            gen_ema,
-            [opt_dec, opt_disc],
-        ) = models_ls
+        (critic, gen, enc, dec, enc2, dec2, critic_rec, gen_ema, [opt_dec, opt_disc],) = models_ls
         if epoch % n_save == 0:
             print("Saving...")
             path = f"{save_path}/MUSIKA!_-{str(gloss)[:9]}-{str(closs)[:9]}-{str(mloss)[:9]}"
@@ -502,7 +441,7 @@ class Utils_functions:
         )
 
     def render_gradio(self, models_ls_techno, models_ls_classical, train=True):
-        article_text = "Original work by Marco Pasini ([Twitter](https://twitter.com/marco_ppasini)) at Johannes Kepler Universität Linz. Supervised by Jan Schlüter."
+        article_text = "Original work by Marco Pasini ([Twitter](https://twitter.com/marco_ppasini)) and Jan Schlüter at Johannes Kepler Universität Linz."
 
         def gradio_func(x, y, z):
             return self.stfunc(x, y, z, models_ls_techno, models_ls_classical)
@@ -514,13 +453,10 @@ class Utils_functions:
                     choices=["Techno/Experimental", "Classical"],
                     type="index",
                     default="Classical",
-                    label="Music Genre to Generate (Brace yourself for very high levels of weirdness!)",
+                    label="Music Genre to Generate",
                 ),
                 gr.inputs.Radio(
-                    choices=["23 s", "115 s", "230 s"],
-                    type="index",
-                    default="115 s",
-                    label="Generated Music Length",
+                    choices=["23s", "1m 58s", "3m 57s"], type="index", default="1m 58s", label="Generated Music Length",
                 ),
                 gr.inputs.Slider(
                     minimum=0,
@@ -536,7 +472,7 @@ class Utils_functions:
             ],
             allow_screenshot=False,
             title="musika!",
-            description="Blazingly Fast Stereo Waveform Music Generation of Arbitrary Length",
+            description="Blazingly Fast Stereo Waveform Music Generation of Arbitrary Length. Be patient and enjoy the weirdness!",
             article=article_text,
             layout="vertical",
             theme="huggingface",
@@ -551,7 +487,7 @@ class Utils_functions:
         if train:
             iface.launch(prevent_thread_lock=True)
         else:
-            iface.launch()
+            iface.launch(enable_queue=True)
         # iface.launch(share=True, enable_queue=True)
         print("--------------------------------")
         print("--------------------------------")