sync ms

app.py

@@ -172,20 +172,23 @@ def generate_music(args, region: str):
         audio = xml2(xml, "wav")
         pdf, jpg = xml2img(xml)
         mxl = xml2(xml, "mxl")
-        return tunes, midi, pdf, xml, mxl, jpg, audio
+        return audio, midi, pdf, xml, mxl, tunes, jpg
 
     except Exception as e:
         print(f"Invalid abc generated: {e}, retrying...")
         return generate_music(args, region)
 
 
-def infer(region: str):
+def infer(p, k, t, region: str):
     if os.path.exists(TEMP_DIR):
         shutil.rmtree(TEMP_DIR)
 
     os.makedirs(TEMP_DIR, exist_ok=True)
     parser = argparse.ArgumentParser()
     args = get_args(parser)
+    args.top_p = p
+    args.top_k = k
+    args.temperature = t
     if region == "Natlan":
         region = "Teyvat"
 
@@ -194,49 +197,30 @@ def infer(region: str):
 
 if __name__ == "__main__":
     warnings.filterwarnings("ignore")
-    with gr.Blocks() as demo:
-        with gr.Row():
-            with gr.Column():
-                region_opt = gr.Dropdown(
-                    choices=TEYVAT,
-                    value="Mondstadt",
-                    label="Region",
-                )
-                gen_btn = gr.Button("Generate")
-                gr.Markdown(
-                    """
-    Welcome to this space based on the Tunesformer open source project, which is totally free!
-    
-    The current model is still in debugging, the plan is in the Genshin Impact after the main line is killed, all countries and regions after all the characters are open, the second creation of the concert will be complete and the sample is balanced, at that time to re-fine-tune the model and add the reality of the style of screening to assist in the game of each country's output to strengthen the learning in order to enhance the output differentiation and quality.
-
-    Data source: <a href="https://musescore.org">MuseScore</a><br>
-    Tags source: <a href="https://genshin-impact.fandom.com/wiki/Genshin_Impact_Wiki">Genshin Impact Wiki | Fandom</a><br>
-    Model base: <a href="https://github.com/sander-wood/tunesformer">Tunesformer</a>
-
-    Note: Data engineering on the Star Rail is in operation, and will hopefully be baselined in the future as well with the mainline kill."""
-                )
-
-            with gr.Column():
-                wav_output = gr.Audio(label="Audio", type="filepath")
-                dld_midi = gr.File(label="Download MIDI")
-                pdf_score = gr.File(label="Download PDF")
-                dld_xml = gr.File(label="Download MusicXML")
-                dld_mxl = gr.File(label="Download MXL")
-                abc_output = gr.Textbox(label="ABC notation", show_copy_button=True)
-                img_score = gr.Image(label="Staff", type="filepath")
-
-        gen_btn.click(
-            infer,
-            inputs=region_opt,
-            outputs=[
-                abc_output,
-                dld_midi,
-                pdf_score,
-                dld_xml,
-                dld_mxl,
-                img_score,
-                wav_output,
-            ],
-        )
-
-    demo.launch()
+    gr.Interface(
+        fn=infer,
+        inputs=[
+            gr.Slider(0, 1.0, 0.8, label="Top-P sample"),
+            gr.Slider(0, 80, 8, label="Top-K sample (0=closed)"),
+            gr.Slider(0, 2.0, 1.2, label="Temperature"),
+            gr.Dropdown(
+                choices=TEYVAT,
+                value="Mondstadt",
+                label="Region",
+            ),
+        ],
+        outputs=[
+            gr.Audio(label="Audio", type="filepath"),
+            gr.File(label="Download MIDI"),
+            gr.File(label="Download PDF"),
+            gr.File(label="Download MusicXML"),
+            gr.File(label="Download MXL"),
+            gr.Textbox(label="ABC notation", show_copy_button=True),
+            gr.Image(label="Staff", type="filepath"),
+        ],
+        flagging_mode="never",
+        description="""
+    Welcome to this space based on the Tunesformer open source project, which is totally free! The current model is still in debugging, the plan is in the Genshin Impact after the main line is killed, all countries and regions after all the characters are open, the second creation of the concert will be complete and the sample is balanced, at that time to re-fine-tune the model and add the reality of the style of screening to assist in the game of each country's output to strengthen the learning in order to enhance the output differentiation and quality. Note: Data engineering on the Star Rail is in operation, and will hopefully be baselined in the future as well with the mainline kill.<br>
+    Data source: <a href="https://musescore.org">MuseScore</a> Tags source: <a href="https://genshin-impact.fandom.com/wiki/Genshin_Impact_Wiki">Genshin Impact Wiki | Fandom</a> Model base: <a href="https://github.com/sander-wood/tunesformer">Tunesformer</a>
+    """,
+    ).launch()

requirements.txt

@@ -3,8 +3,8 @@ music21
 pymupdf
 autopep8
 unidecode
-modelscope[framework]==1.18
-transformers==4.18.0
-samplings==0.1.7
 pillow==9.4.0
-numpy==1.26.4
+numpy==1.26.4
+samplings==0.1.7
+transformers==4.18.0
+modelscope[framework]==1.18

utils.py

@@ -13,15 +13,13 @@ WEIGHTS_PATH = (
     + "/weights.pth"
 )
 DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-TEMP_DIR = "./flagged"
+TEMP_DIR = "./__pycache__/tmp"
 PATCH_LENGTH = 128  # Patch Length
 PATCH_SIZE = 32  # Patch Size
 PATCH_NUM_LAYERS = 9  # Number of layers in the encoder
 CHAR_NUM_LAYERS = 3  # Number of layers in the decoder
-# Batch size for patch during training, 0 for full context
-PATCH_SAMPLING_BATCH_SIZE = 0
-# Whether to share weights between the encoder and decoder
-SHARE_WEIGHTS = False
+PATCH_SAMPLING_BATCH_SIZE = 0  # Batchsize for patch during training, 0=full context
+SHARE_WEIGHTS = False  # Whether to share weights between the encoder and decoder
 
 
 def download(filename: str, url: str):