def predict(model, filename, weights_path):
"""
Predicts the label for a single file with MFCC features.
Args:
model: Keras model
weights_path: filepath to the weights of the model
filename: filepath to the file which should be processed
Returns:
The label.
"""
if os.path.isfile(weights_path):
model.load_weights(weights_path, by_name=False)
sample = data_wrangler.single_data_processing(filename)
pred = model.predict(sample)
print(pred)
print(data_read.label_from_oh(pred))
return data_read.label_from_oh(pred)
def predict_preprocessed(weights_path, file_paths, output_filename):
"""
Predict list processed files and saves it to CSV file.
Args:
weights_path: weights for model
file_paths: files to predict a label for
output_filename: name of CSV file
Returns:
-
"""
if os.path.isfile(weights_path):
print("Loading weights")
model.load_weights(weights_path, by_name=False)
with open(output_filename, "w") as csv_file:
csv_writer = csv.writer(csv_file, delimiter=",", dialect="unix", quoting = csv.QUOTE_NONE)
iteration = 0
max_iterations = math.ceil(len(file_paths)/256)
csv_writer.writerow(["fname","label"])
while len(file_paths) > 0:
batch = file_paths[:256]
batch = [np.reshape(np.loadtxt(filename), (101, 128)) for filename in batch]
batch = np.asarray(batch)
pred = model.predict(batch, batch_size=256)
for filename, prediction in zip(file_paths, pred):
clean_filename = filename.split("/")[-1].split(".")[0]+".wav"
csv_writer.writerow([clean_filename, data_read.label_from_oh(prediction)])
file_paths=file_paths[len(batch):] # TODO
iteration +=1
print("%s/%s done." % (iteration, max_iterations))
def predict_single_unpreprocessed(weights_path, file, model):
"""
Predict list of raw '.wav' files and saves it to CSV file.
Args:
weights_path: weights for model
file_paths: files to predict a label for
output_filename: name of CSV file
Returns:
-
"""
if os.path.isfile(weights_path):
#print("Loading weights")
model.load_weights(weights_path, by_name=False)
batch = [data_wrangler.single_data_processing_raw(file)]
batch = np.reshape(np.array(batch), (len(batch),101,13))
pred = model.predict(batch, batch_size=256)
print(data_read.label_from_oh(pred))
def predict_unpreprocessed(weights_path, file_paths, output_filename):
"""
Predict list of raw '.wav' files and saves it to CSV file.
Args:
weights_path: weights for model
file_paths: files to predict a label for
output_filename: name of CSV file
Returns:
-
"""
if os.path.isfile(weights_path):
print("Loading weights")
model.load_weights(weights_path, by_name=False)
with open(output_filename, "w") as csv_file:
csv_writer = csv.writer(csv_file, delimiter=",", dialect="unix", quoting = csv.QUOTE_NONE)
iteration = 0
max_iterations = math.ceil(len(file_paths)/256)
while len(file_paths) >0:
batch_filenames = file_paths[:256]
batch = [data_wrangler.single_data_processing(filename) for filename in batch_filenames]
batch = np.reshape(np.array(batch), (len(batch),101,13))
pred = model.predict(batch, batch_size=256)
for filename, prediction in zip(batch_filenames, pred):
clean_filename = filename.split("/")[-1]
csv_writer.writerow([clean_filename, data_read.label_from_oh(prediction)]) # TODO
file_paths = file_paths[min(len(batch),256):]
iteration = iteration+1
sys.stdout.write('\r>> Predicting classes for batch %d of %d.' % (iteration, max_iterations))
sys.stdout.flush()
sample = data_wrangler.single_data_processing_with_mel_spectrogram(
filename)
pred = model.predict(sample)
print(pred)
print(data_read.label_from_oh(pred))
if __name__ == "__main__":
model = my_model()
x_test, y_test, x_train, y_train = load_dataset()
d = {}
for l in y_train:
d[data_read.label_from_oh(l)] = d.get(data_read.label_from_oh(l),
0) + 1
print(d)
log_path = "models/v54"
weights_path = log_path + "/weights.best.hdf5"
learning_rate = 0.0001
train_model(model, 1000, x_test, y_test, x_train, y_train, weights_path,
log_path, learning_rate)
if os.path.isfile(weights_path):
model.load_weights(weights_path, by_name=False)
pred = model.predict(x_test, batch_size=128)
x = 0
for p_i, l_i in zip(pred, [data_read.label_from_oh(l) for l in y_test]):
