hparam tweaks for final output

Xefyr0 · Nov 26, 2024 · 7a94545 · 7a94545
1 parent bffcf38
commit 7a94545
Show file tree

Hide file tree

Showing 4 changed files with 5 additions and 5 deletions.
diff --git a/magenta/models/coconet/coconet_train.py b/magenta/models/coconet/coconet_train.py
@@ -113,7 +113,7 @@
 
 flags.DEFINE_float('corrupt_ratio', 0.5, 'Fraction of variables to mask out.')
 # Run parameters.
-flags.DEFINE_integer('num_epochs', 50,
+flags.DEFINE_integer('num_epochs', 100,
                      'The number of epochs to train the model. Default '
                      'is 0, which means to run until terminated '
                      'manually.')

diff --git a/magenta/models/coconet/lib_graph.py b/magenta/models/coconet/lib_graph.py
@@ -112,8 +112,8 @@ def build(self):
         logits=self.logits, labels=self.pianorolls)
 
     self.compute_loss(
-        self.cpep_calculator.calculate_voice_range_penalty(self.predictions) + 
-        self.cpep_calculator.calculate_kernel_penalty(self.predictions) + 
+        self.cpep_calculator.calculate_voice_range_penalty(self.predictions) +
+        self.cpep_calculator.calculate_kernel_penalty(self.predictions) +
         self.cpep_calculator.calculate_parallel_perfect_penalty(self.predictions) +
         self.cross_entropy)
     self.setup_optimizer()

diff --git a/magenta/models/coconet/sample_bazel.sh b/magenta/models/coconet/sample_bazel.sh
@@ -29,7 +29,7 @@ generation_output_dir=$HOME/samples
 
 # Generation parameters.
 # Number of samples to generate in a batch.
-gen_batch_size=10
+gen_batch_size=20
 piece_length=64
 strategy=igibbs
 tfsample=true

diff --git a/magenta/models/coconet/train_bazel.sh b/magenta/models/coconet/train_bazel.sh
@@ -25,7 +25,7 @@ data_dir=$HOME/JSB-Chorales-dataset/
 dataset=Jsb16thSeparated
 
 # Data preprocessing.
-crop_piece_len=48
+crop_piece_len=64
 separate_instruments=True
 quantization_level=0.125  # 16th notes