Evaluate models gpu

.gitignore

@@ -8,3 +8,7 @@
 build/
 *.DS_Store
 .vscode/*
+.idea
+logs
+wandb
+.python-version

magenta/models/coconet/.gitignore

@@ -0,0 +1 @@
+coconet_checkpoint

magenta/models/coconet/coconet_evaluate.py

@@ -19,10 +19,10 @@
 
 import os
 
-from magenta.models.coconet import lib_data
-from magenta.models.coconet import lib_evaluation
-from magenta.models.coconet import lib_graph
-from magenta.models.coconet import lib_util
+import lib_data
+import lib_evaluation
+import lib_graph
+import lib_util
 import numpy as np
 import tensorflow as tf
 
@@ -38,16 +38,16 @@
 flags.DEFINE_string('fold_index', None,
                     'Optionally, index of particular data point in fold to '
                     'evaluate.')
-flags.DEFINE_string('unit', None, 'Note or frame or example.')
+flags.DEFINE_string('unit', 'note', 'Note or frame or example.')
 flags.DEFINE_integer('ensemble_size', 5,
                      'Number of ensemble members to average.')
-flags.DEFINE_bool('chronological', False,
+flags.DEFINE_bool('chronological', True,
                   'Indicates evaluation should proceed in chronological order.')
 flags.DEFINE_string('checkpoint', None, 'Path to checkpoint directory.')
 flags.DEFINE_string('sample_npy_path', None, 'Path to samples to be evaluated.')
 
 
-EVAL_SUBDIR = 'eval_stats'
+EVAL_SUBDIR = 'eval'
 
 
 def main(unused_argv):
@@ -132,6 +132,7 @@ def evaluate_paths(paths, evaluator, unused_hparams, eval_logdir):
 def get_fold_pianorolls(fold, hparams):
   dataset = lib_data.get_dataset(FLAGS.data_dir, hparams, fold)
   pianorolls = dataset.get_pianorolls()
+  pianorolls[0:2]
   tf.logging.info('Retrieving pianorolls from %s set of %s dataset.',
                   fold, hparams.dataset)
   print_statistics(pianorolls)

magenta/models/coconet/coconet_sample.py

@@ -23,13 +23,13 @@
 import re
 import time
 
-from magenta.models.coconet import lib_graph
-from magenta.models.coconet import lib_logging
-from magenta.models.coconet import lib_mask
-from magenta.models.coconet import lib_pianoroll
-from magenta.models.coconet import lib_sampling
-from magenta.models.coconet import lib_tfsampling
-from magenta.models.coconet import lib_util
+import lib_graph
+import lib_logging
+import lib_mask
+import lib_pianoroll
+import lib_sampling
+import lib_tfsampling
+import lib_util
 import numpy as np
 import pretty_midi
 import tensorflow as tf

magenta/models/coconet/coconet_train.py

@@ -109,7 +109,7 @@
                   'Optimize masked predictions only.')
 flags.DEFINE_bool('rescale_loss', True, 'Rescale loss based on context size.')
 flags.DEFINE_integer(
-    'patience', 10,
+    'patience', 5,
     'Number of epochs to wait for improvement before decaying learning rate.')
 
 flags.DEFINE_float('corrupt_ratio', 0.5, 'Fraction of variables to mask out.')

magenta/models/coconet/evalmodel_bazel.sh

@@ -28,22 +28,21 @@ checkpoint=$1
 data_dir="testdata"
 
 # Change this to where evaluation results are stored.
-eval_logdir="eval_logdir"
+eval_logdir=$1
 
 # Evaluation settings.
 fold=valid
-fold_index=1  # Optionally can specify index of specific piece to be evaluated.
-unit=frame
-chronological=false
-ensemble_size=5  # Number of different orderings to average.
+#fold_index=1  # Optionally can specify index of specific piece to be evaluated.
+unit=note
+chronological=True
+ensemble_size=1  # Number of different orderings to average.
 
 # Run command.
-python coconet_evaluate.py \
+python3 coconet_evaluate.py \
 --data_dir=$data_dir \
 --eval_logdir=$eval_logdir \
 --checkpoint=$checkpoint \
 --fold=$fold \
---unit=$unit \
 --chronological=$chronological \
---ensemble_size=5 \
-#--fold_index=$fold_index
+--ensemble_size=$ensemble_size \
+--unit=$unit

magenta/models/coconet/lib_evaluation.py

@@ -19,10 +19,10 @@
 
 import time
 
-from magenta.models.coconet import lib_tfutil
-from magenta.models.coconet import lib_util
+import lib_tfutil
+import lib_util
 import numpy as np
-from scipy.misc import logsumexp
+from scipy.special import logsumexp
 import tensorflow as tf
 
 
@@ -287,7 +287,7 @@ def __call__(self, pianoroll):
     self._update_lls(lls, xs, pxhats, ts, ds)
     return lls
 
-  def _draw_ordering(self, tt, dd):
+  def draw_ordering(self, tt, dd):
     o = np.arange(tt * dd, dtype=np.int32)
     if not self.chronological:
       np.random.shuffle(o)

magenta/models/coconet/lib_logging.py

@@ -20,7 +20,7 @@
 import contextlib
 import os
 
-from magenta.models.coconet import lib_util
+import lib_util
 import numpy as np
 
 

magenta/models/coconet/lib_sampling.py

@@ -18,11 +18,11 @@
 from __future__ import division
 from __future__ import print_function
 
-from magenta.models.coconet import lib_data
-from magenta.models.coconet import lib_logging
-from magenta.models.coconet import lib_mask
-from magenta.models.coconet import lib_tfutil
-from magenta.models.coconet import lib_util
+import lib_data
+import lib_logging
+import lib_mask
+import lib_tfutil
+import lib_util
 import numpy as np
 
 ################

magenta/models/coconet/sample_bazel.sh

@@ -35,7 +35,7 @@ strategy=igibbs
 tfsample=true
 
 # Run command.
-python coconet_sample.py \
+python3 coconet_sample.py \
 --checkpoint="$checkpoint" \
 --gen_batch_size=$gen_batch_size \
 --piece_length=$piece_length \
@@ -44,5 +44,3 @@ python coconet_sample.py \
 --tfsample=$tfsample \
 --generation_output_dir=$generation_output_dir \
 --logtostderr
-
-

magenta/models/coconet/train_bazel.sh

@@ -33,7 +33,7 @@ quantization_level=0.125  # 16th notes
 maskout_method=orderless
 num_layers=32
 num_filters=64
-batch_size=20
+batch_size=10
 use_sep_conv=True
 architecture='dilated'
 num_dilation_blocks=1

magenta/models/coconet/transfer.txt

@@ -1,3 +1,5 @@
+coconet_evaluate.py
+evalmodel_bazel.sh
 coconet_train.py
 train_bazel.sh
 lib_util.py
@@ -7,3 +9,4 @@ lib_data.py
 lib_mask.py
 lib_pianoroll.py
 lib_tfutil.py
+lib_evaluation.py