Big UI overhaul

app.py

@@ -24,11 +24,46 @@ def generate(*args):
 
 
 def main():
-    gr.Interface(inputs=[
-        (gr.inputs.Number(label=k, default=v0) if t in (float, int) else
-         gr.inputs.Checkbox(label=k, default=v0) if t == bool else gr.inputs.Textbox(label=k, default=v0) if t == str
-         else gr.inputs.Dropdown(label=k, default=v0, choices=t) if isinstance(t, (tuple, list)) else 1/0)
-        for k, v0, t in CONFIG_SPEC], outputs=gr.outputs.Video(), fn=generate).launch()
+    with gr.Blocks() as ui:
+        gr.Markdown("# Pulsar+CLIP")
+        gr.Markdown("Generate 3D point clouds from text!")
+        
+        with gr.Group():
+            gr.Markdown("## Settings")
+            inputs = []
+            with gr.Tabs():
+                for name, section in CONFIG_SPEC:
+                    with gr.TabItem(name):
+                        for k, v0, t in section:
+                            if t in (float, int):
+                                element = gr.Number(label=k, value=v0)
+                            elif t == str:
+                                element = gr.Textbox(label=k, value=v0)
+                            elif t == bool:
+                                element = gr.Checkbox(label=k, value=v0)
+                            elif isinstance(t, tuple):
+                                element = gr.Slider(*t, label=k, value=v0)
+                            elif isinstance(t, list):
+                                element = gr.Dropdown(label=k, value=v0, choices=t)
+                            else:
+                                raise TypeError(f"Input format {t} should be one of str, int, bool, tuple, list")
+                                element = 1/0
+                            inputs.append(element)
+        button = gr.Button("Run")
+        gr.Markdown("## Result")
+        output = gr.Video()
+        
+        button.click(fn=generate,
+                     inputs=inputs,
+                     outputs=[output])
+    
+    ui.launch()
+    
+    #gr.Interface(inputs=[
+        #(gr.inputs.Number(label=k, default=v0) if t in (float, int) else
+         #gr.inputs.Checkbox(label=k, default=v0) if t == bool else gr.inputs.Textbox(label=k, default=v0) if t == str
+         #else gr.inputs.Dropdown(label=k, default=v0, choices=t) if isinstance(t, (tuple, list)) else 1/0)
+        #for k, v0, t in CONFIG_SPEC], outputs=gr.outputs.Video(), fn=generate).launch()
 
 
 if __name__ == '__main__':

pulsar_clip.py

@@ -8,58 +8,67 @@ import torch
 
 
 CONFIG_SPEC = [
-    ("text", "A cloud at dawn", str),
-    ("iterations", 5000, int),
-    ("turns", 4, int),
-    ("showoff", 5000, int),
-    ("seed", 12, int),
-    ("focal_length", 0.1, float),
-    ("plane_width", 0.1, float),
-    ("shade_strength", 0.25, float),
-    ("gamma", 0.5, float),
-    ("max_depth", 7, float),
-    ("lr", 0.5, float),
-    ("offset", 5, float),
-    ("offset_random", 0.75, float),
-    ("xyz_random", 0.25, float),
-    ("altitude_range", 0.3, float),
-    ("augments", 4, int),
-    ("show_every", 50, int),
-    ("epochs", 1, int),
-    ("w", 224, int),
-    ("h", 224, int),
-    ("num_objects", 256, int),
-    #@markdown CLIP loss type, might improve the results
-    ("loss_type", "spherical", ("spherical", "cosine")),
-    #@markdown CLIP loss weight
-    ("clip_weight", 1.0, float),        #@param {type: "number"}
-    #@markdown Number of dimensions. 0 is for point clouds (default), 1 will make
-    #@markdown strokes, 2 will make planes, 3 produces little cubes
-    ("ndim", 0, (0, 1, 2, 3)),  #@param {type: "integer"}
-
-    #@markdown Opacity scale:
-    ("min_opacity", 1e-4, float),       #@param {type: "number"}
-    ("max_opacity", 1.0, float),        #@param {type: "number"}
-    ("log_opacity", False, bool),      #@param {type: "boolean"}
-
-    ("min_radius", 0.030, float),
-    ("max_radius", 0.070, float),
-    ("log_radius", False, bool),
-
-    # TODO dynamically decide bezier_res
-    #@markdown Bezier resolution: how many points a line/plane/cube will have. Not applicable to points
-    ("bezier_res", 8, int),  #@param {type: "integer"}
-    #@markdown Maximum scale of parameters: position, velocity, acceleration
-    ("pos_scale", 0.4, float),  #@param {type: "number"}
-    ("vel_scale", 0.15, float),  #@param {type: "number"}
-    ("acc_scale", 0.15, float),  #@param {type: "number"}
-
-    #@markdown Scale of each individual 3D object. Master control for velocity and acceleration scale.
-    ("scale", 1, float),  #@param {type: "number"}
+    ("General", [
+        ("text", "A cloud at dawn", str),
+        ("iterations", 5000, (0, 7500)),
+        ("seed", 12, int),
+    ]),
+    ("Rendering", [
+        ("w", 224, [224, 252]),
+        ("h", 224, [224, 252]),
+        # ("show_every", 50, int),
+        ("showoff", 5000, (0, 10000)),
+        ("turns", 4, int),
+        ("focal_length", 0.1, float),
+        ("plane_width", 0.1, float),
+        ("shade_strength", 0.25, float),
+        ("gamma", 0.5, float),
+        ("max_depth", 7, float),
+        ("offset", 5, float),
+        ("offset_random", 0.75, float),
+        ("xyz_random", 0.25, float),
+        ("altitude_range", 0.3, float),
+        ("augments", 4, int),
+    ]),
+    ("Optimization", [
+        ("epochs", 1, int),
+        ("lr", 0.5, float),
+        #@markdown CLIP loss type, might improve the results
+        ("loss_type", "spherical", ["spherical", "cosine"]),
+        #@markdown CLIP loss weight
+        ("clip_weight", 1.0, float),        #@param {type: "number"}
+    ]),
+    ("Elements", [
+        ("num_objects", 256, int),
+        #@markdown Number of dimensions. 0 is for point clouds (default), 1 will make
+        #@markdown strokes, 2 will make planes, 3 produces little cubes
+        ("ndim", 0, [0, 1, 2, 3]),  #@param {type: "integer"}
+
+        #@markdown Opacity scale:
+        ("min_opacity", 1e-4, float),       #@param {type: "number"}
+        ("max_opacity", 1.0, float),        #@param {type: "number"}
+        ("log_opacity", False, bool),      #@param {type: "boolean"}
+
+        ("min_radius", 0.030, float),
+        ("max_radius", 0.070, float),
+        ("log_radius", False, bool),
+
+        # TODO dynamically decide bezier_res
+        #@markdown Bezier resolution: how many points a line/plane/cube will have. Not applicable to points
+        ("bezier_res", 8, int),  #@param {type: "integer"}
+        #@markdown Maximum scale of parameters: position, velocity, acceleration
+        ("pos_scale", 0.4, float),  #@param {type: "number"}
+        ("vel_scale", 0.15, float),  #@param {type: "number"}
+        ("acc_scale", 0.15, float),  #@param {type: "number"}
+
+        #@markdown Scale of each individual 3D object. Master control for velocity and acceleration scale.
+        ("scale", 1, float),  #@param {type: "number"}
+    ]),
 ]
 
 
 # TODO: one day separate the config into multiple parts and split this megaobject into multiple objects
+# 2022/08/09: halfway done
 class PulsarCLIP(object):
     def __init__(self, args):
         args = DotDict(**args)
@@ -208,11 +217,11 @@ class PulsarCLIP(object):
                         self.scheduler.step()
                     except AttributeError:
                         pass
-                if i % self.args.show_every == 0:
-                    cam_params = self.camera(i / self.args.iterations * np.pi * 2 * self.args.turns, use_random=False)
-                    img_show, _ = self.render(cam_params)
-                    img = Image.fromarray((img_show.cpu().detach().numpy() * 255).astype(np.uint8))
-                    yield img
+                #if i % self.args.show_every == 0:
+                    #cam_params = self.camera(i / self.args.iterations * np.pi * 2 * self.args.turns, use_random=False)
+                    #img_show, _ = self.render(cam_params)
+                    #img = Image.fromarray((img_show.cpu().detach().numpy() * 255).astype(np.uint8))
+                    #yield img
         except KeyboardInterrupt:
             pass