time temperature and bias; rename midi-duet to generate.scd

Author: victor-shepardson

examples/notepredictor/autopitch.scd

@@ -67,6 +67,7 @@ s.waitForBoot{
 };
 )
 
+OSCdef.trace(false)
 // ~linn_reset.()
 
 (

examples/notepredictor/midi-duet.scd examples/notepredictor/generate.scdexamples/notepredictor/midi-duet.scd renamed to examples/notepredictor/generate.scd

@@ -42,8 +42,11 @@ b.sendMsg("/predictor/predict", \pitch, 60+12.rand, \time, 0, \vel, 0);
 // duet with the model
 // feeds the model's predictions back to it as well as player input
 (
+~step = 0;
 ~gate = 1;
-t = Process.elapsedTime;
+t = nil;
+~player_t = t;
+~machine_t = t;
 b.sendMsg("/predictor/reset");
 
 // footswitch
@@ -58,6 +61,7 @@ MIDIdef.program(\switch, {
         b.sendMsg("/predictor/reset");
         y.release;
         SystemClock.clear;
+        ~step = 0;
         };
     ~gate.postln;
 });
@@ -73,7 +77,12 @@ MIDIdef.noteOn(\input, {
     SystemClock.clear;
 
     //get a new prediction in light of current note
-    b.sendMsg("/predictor/predict", \pitch, num, \time, dt, \vel, val);
+    b.sendMsg("/predictor/predict",
+        \pitch, num, \time, dt, \vel, val,
+        \allow_start, false, \allow_end, false,
+        \time_temp, 0, \min_time, 0.1, \max_time, 5
+        // \fix_time, 9
+    );
 
     // release the previous note
     y.release(0.1);
@@ -86,6 +95,9 @@ MIDIdef.noteOn(\input, {
 
     // mark time of current note
     t = t2;
+    ~player_t = t;
+
+    ~step = ~step + 1;
 });
 
 
@@ -98,8 +110,8 @@ OSCdef(\return, {
 
     // time-to-next note gets 'censored' by the model
     // when over a threshold, in this case 10 seconds,
-    // meaning it just predicts 10s rather than a any longer time
-    var censor = dt==10.0;
+    // meaning it just predicts 10s rather than any longer time
+    var censor = dt>10.0;
 
     censor.if{
         // if the predicted time is > 10 seconds, don't schedule it, just stop.
@@ -108,6 +120,8 @@ OSCdef(\return, {
         // schedule the predicted note
         SystemClock.sched(dt-~delay, {
             (~gate>0).if{
+                var t2 = Process.elapsedTime;
+                var dt_actual = t2 - t;
                 (num==129).if{
                     // 129 is the 'stop token',  meaning 'end-of-performance'
                     // in this case don't schedule a note, and reset the model
@@ -124,7 +138,13 @@ OSCdef(\return, {
                     SystemClock.clear;
                     // feed model its own prediction as input
                     b.sendMsg("/predictor/predict",
-                        \pitch, num, \time, dt, \vel, val);
+                        \pitch, num, \time, dt_actual, \vel, val,
+                        \allow_start, false, \allow_end, false,
+                        \time_temp, 0.1, \min_time, 0.1, \max_time, 5
+                        // \fix_time, (~step%4==0).if{0.6}{0}  // tetrachords
+                        // \fix_time, (~step%8)*0.1 // specific rhythm
+
+                    );
                     // release the previous note
                     (dt<3e-2).if{
                         // if the time delay is very small, slow release for chord
@@ -138,14 +158,17 @@ OSCdef(\return, {
                         \freq, num.midicps, \vel, val/127]);//.release(1);
                     // post the current note
                     [\model, dt, num, val].postln;
-                    // mark the time of current note
-                    t = Process.elapsedTime;
+                    // mark the actual time of current note
+                    t = t2;
+                    ~machine_t = t;
                     // crudely draw note on piano GUI
                     ~gui.if{
                         AppClock.sched(0,{k.keyDown(num)});
                         AppClock.sched(0.2,{k.keyUp(num)});
                     }
-                }
+                };
+                ~step = ~step+1;
+                [\late, dt_actual-dt].postln;
             }
     })};
 

notepredictor/notepredictor/distributions.py

@@ -115,11 +115,17 @@ def cdf_components(self, loc, s, x):
         x_ = (x[...,None] - loc) * s
         return x_.sigmoid()
 
-    def sample(self, h, truncate=None, shape=None):
+    def sample(self, h, truncate=None, shape=None, temp=None, bias=None):
         """
         Args:
             h: Tensor[...,n_params]
-            shape: additional sample shape to be prepended to dims or None
+            truncate: Optional[Tuple[2]]. lower and upper bound for truncation.
+            shape: Optional[int]. additional sample shape to be prepended to dims.
+            temp: Optional[float]. pseudo-temperature (temperature of each mixture 
+                component). default is 1. 0 would sample component location only,
+                ignoring sharpness.
+            bias: applied outside of truncation but inside of clamping,
+                useful e.g. for latency correction when sampling delta-time
         Returns:
             Tensor[*shape,...] (h without last dimension, prepended with `shape`)
         """
@@ -133,6 +139,12 @@ def sample(self, h, truncate=None, shape=None):
             truncate = (-np.inf, np.inf)
         truncate = torch.tensor(truncate)
 
+        if temp is None:
+            temp = 1
+
+        if bias is None:
+            bias = 0
+
         log_pi, loc, s = self.get_params(h)
         scale = 1/s
 
@@ -154,7 +166,7 @@ def sample(self, h, truncate=None, shape=None):
         u = u * (upper-lower) + lower
 
         # x = loc + scale * (u.log() - (1 - u).log())
-        x = loc - scale * (1/u - 1).log()
+        x = loc + bias - scale * temp * (1/u - 1).log()
         x = x.clamp(self.lo, self.hi)
         return x[0] if unwrap else x
 

notepredictor/notepredictor/model.py

@@ -200,7 +200,11 @@ def cell_state(self):
 
     def get_samplers(self, 
             pitch_topk=None, index_pitch=None, allow_start=False, allow_end=False, 
-            sweep_time=False, trunc_time=None):
+            sweep_time=False, min_time=None, max_time=None, bias_time=None, time_temp=None):
+        """
+        this method converts the many arguments to `predict` into functions for
+        sampling each note modality (e.g. pitch, time, velocity)
+        """
 
         def sample_pitch(x):
             if not allow_start:
@@ -217,7 +221,7 @@ def sample_pitch(x):
         def sample_time(x):
             # TODO: respect trunc_time when sweep_time is True
             if sweep_time:
-                if trunc_time is not None:
+                if min_time is not None or max_time is not None:
                     raise NotImplementedError("""
                     trunc_time with sweep_time needs implementation
                     """)
@@ -228,7 +232,11 @@ def sample_time(x):
                 # print(loc.shape)
                 return loc
             else:
-                return self.time_dist.sample(x, truncate=trunc_time)
+                trunc = (
+                    -np.inf if min_time is None else min_time,
+                    np.inf if max_time is None else max_time)
+                return self.time_dist.sample(x, 
+                    truncate=trunc, temp=time_temp, bias=bias_time)
 
         return (
             sample_pitch, 
@@ -316,7 +324,7 @@ def predict(self,
             pitch, time, vel, 
             fix_pitch=None, fix_time=None, fix_vel=None, 
             pitch_topk=None, index_pitch=None, allow_start=False, allow_end=False,
-            sweep_time=False, trunc_time=None):
+            sweep_time=False, min_time=None, max_time=None, bias_time=None, time_temp=None):
         """
         consume the most recent note and return a prediction for the next note.
 
@@ -335,7 +343,8 @@ def predict(self,
             allow_end: if False, zero probaility for sampling the end token
             sweep_time: if True, instead of sampling time, choose a diverse set of
                 times and stack along the batch dimension
-            trunc_time: if not None, truncate the time distribution to (lo, hi)
+            min_time, max_time: if not None, truncate the time distribution
+            time_temp: if not None, apply pseudo-temperature to the time distribution.
 
         Returns: dict of
             'pitch': int. predicted MIDI number of next note.
@@ -366,7 +375,7 @@ def predict(self,
                 self.projections,
                 self.get_samplers(
                     pitch_topk, index_pitch, allow_start, allow_end, 
-                    sweep_time, trunc_time),
+                    sweep_time, min_time, max_time, bias_time, time_temp),
                 self.embeddings,
                 ))
 
@@ -388,7 +397,9 @@ def predict(self,
             for i,(item, embed) in enumerate(zip(fix, self.embeddings)):
                 if item is None:
                     if (
-                        i==1 and (sweep_time or (trunc_time is not None)) or
+                        i==1 and (sweep_time 
+                            or (min_time is not None) or (max_time is not None)
+                            or (time_temp is not None)) or
                         i==0 and pitch_topk
                         ):
                         cons_idx.append(i)
@@ -403,6 +414,9 @@ def predict(self,
             perm = det_idx + undet_idx # permutation from the canonical order
             iperm = np.argsort(perm) # inverse permutation back to canonical order
 
+            md = ['pitch', 'time', 'vel']
+            print([md[i] for i in perm])
+
             # for each undetermined modality, 
             # sample a new value conditioned on alteady determined ones