fix transcribe

piano_transcription/models/cnn_1.py

@@ -47,7 +47,7 @@ def get_train_batch_iterator():
     return batch_iterator
 
 
-def predict(net, X, max_seq_len, out_len):
+def predict(pfun, X, max_seq_len, out_len):
     seq_len, feat_len = X.shape
 
     pad_width = (SPEC_CONTEXT - 1) / 2
@@ -68,9 +68,9 @@ def predict(net, X, max_seq_len, out_len):
             i0 = batch * MAX_PRED_SIZE
             i1 = i0 + MAX_PRED_SIZE
             i1o = min(i1, seq_len)
-            p_b[i0:i1o] = net.predict_proba(X_pred[i0:i1])[0:(i1o-i0)]
+            p_b[i0:i1o] = pfun(X_pred[i0:i1])[0:(i1o-i0)]
     else:
-        p_b = net.predict_proba(X_pred)[:seq_len]
+        p_b = pfun(X_pred)[:seq_len]
 
     return p_b
 

piano_transcription/models/crnn_1.py

@@ -52,7 +52,7 @@ def get_train_batch_iterator():
     return batch_iterator
 
 
-def predict(net, X, max_seq_len, out_len):
+def predict(pfun, X, max_seq_len, out_len):
     seq_len, feat_len = X.shape
 
     n_batches = int(np.ceil(seq_len / float(MAX_PRED_SIZE)))
@@ -82,9 +82,9 @@ def predict(net, X, max_seq_len, out_len):
             i0 = batch * MAX_PRED_SIZE
             i1 = i0 + MAX_PRED_SIZE
             i1o = min(i1, seq_len)
-            p_b[i0:i1o] = net.predict_proba([X_pred[:, i0:i1], m_b[:, i0:i1]])[0, 0:(i1o-i0)]
+            p_b[i0:i1o] = pfun([X_pred[:, i0:i1], m_b[:, i0:i1]])[0, 0:(i1o-i0)]
     else:
-        p_b = net.predict_proba([X_pred, m_b])[0, :seq_len]
+        p_b = pfun([X_pred, m_b])[0, :seq_len]
 
     return p_b
 

piano_transcription/models/rnn_1.py

@@ -26,6 +26,16 @@ RNN_GRAD_CLIP = 50
 FF_GRAD_CLIP = 50
 
 
+def predict(pfun, X, max_seq_len, out_len):
+    seq_len, feat_len = X.shape
+    # pad and create mask
+    m_b = np.ones((1, seq_len), dtype=theano.config.floatX)
+    X_pred = np.pad(X, ((0, max_seq_len - seq_len), (0, 0)), 'constant').astype(theano.config.floatX)
+    m_b = np.pad(m_b, ((0, 0), (0, max_seq_len - seq_len)), 'constant').astype(theano.config.floatX)
+
+    return pfun([X_pred, m_b])[0, :seq_len]
+
+
 def get_batch_iterator_train():
     def batch_iterator(batch_size, k_samples, shuffle):
         return BatchIterator(batch_size=batch_size, prepare=prepare_train, k_samples=k_samples, shuffle=shuffle)

piano_transcription/transcribe.py

@@ -6,29 +6,45 @@ import numpy as np
 from madmom.features.notes import NotePeakPickingProcessor
 
 from piano_transcription import BEST_MODEL_FILE_NAME, SETTINGS_FILE_NAME, LOSSES_FILE
-from piano_transcription.utils import select_model
+from piano_transcription.utils import select_model, collect_inputs
 from piano_transcription.data import FEAT_SIZE, OUT_SIZE
 from piano_transcription.data.annotations import write_txt_annotation
 from piano_transcription.data.features import extract_features
 
+import theano
+import lasagne
+
+
+def compile_prediction_function(net):
+    """
+    Compile theano prediction function
+    """
+    input_vars = collect_inputs(net)
+    net_output = lasagne.layers.get_output(net, deterministic=True)
+    return theano.function(inputs=input_vars, outputs=net_output, on_unused_input='warn')
+
 
 def run(model_path, input_file):
 
-    settings = np.load(os.path.join(model_path, SETTINGS_FILE_NAME))
+    settings = np.load(os.path.join(model_path, SETTINGS_FILE_NAME)).item(0)
 
-    model = select_model(settings['model'])
-    model_seq_len = model.MAX_PRED_SIZE
+    model, model_name = select_model(settings['model'])
 
     features = extract_features(input_file)
 
-    network = model.build_eval_model(model_seq_len, FEAT_SIZE, OUT_SIZE)
+    seq_len, feat_len = features.shape
+
+    network = model.build_eval_model(seq_len, feat_len, OUT_SIZE)
 
     with np.load(os.path.join(model_path, BEST_MODEL_FILE_NAME)) as f:
         param_values = [f['arr_%d' % i] for i in range(len(f.files))]
     lasagne.layers.set_all_param_values(network, param_values)
 
+    pfun = compile_prediction_function(network)
+
     # transcribe using model
-    pred = model.predict(network, features, model_seq_len, OUT_SIZE)
+
+    pred = model.predict(pfun, features, seq_len, OUT_SIZE)
 
     peak_picker = NotePeakPickingProcessor(pitch_offset=0)
     notes = peak_picker.process(pred)
@@ -51,4 +67,5 @@ def main():
 
 
 if __name__ == '__main__':
+    #run('/Users/rich/python-workspace/piano_trans/piano_transcription/output/2018-06-17-16-57-22', '/Users/rich/Desktop/piano_test/CH/MAPS_ISOL_CH0.3_F_AkPnBsdf.flac')
     main()