def visualize(model, test_file, train_desc_file):
""" Get the prediction using the model, and visualize softmax outputs
Params:
model (keras.models.Model): Trained speech model
test_file (str): Path to an audio clip
train_desc_file(str): Path to the training file used to train this
model
"""
datagen = DataGenerator()
datagen.load_train_data(train_desc_file)
datagen.fit_train(100)
print("Compiling test function...")
test_fn = compile_output_fn(model)
inputs = [datagen.featurize(test_file)]
prediction = np.squeeze(test_fn([inputs, True]))
softmax_file = "softmax.npy".format(test_file)
softmax_img_file = "softmax.png".format(test_file)
print("Prediction: {}".format(argmax_decode(prediction)))
print("Saving network output to: {}".format(softmax_file))
print("As image: {}".format(softmax_img_file))
np.save(softmax_file, prediction)
sm = softmax(prediction.T)
sm = np.vstack((sm[0], sm[2], sm[3:][::-1]))
fig, ax = plt.subplots()
ax.pcolor(sm, cmap=plt.cm.Greys_r)
column_labels = [chr(i) for i in range(97, 97 + 26)] + ['space', 'blank']
ax.set_yticks(np.arange(sm.shape[0]) + 0.5, minor=False)
ax.set_yticklabels(column_labels[::-1], minor=False)
plt.savefig(softmax_img_file)
def interactive_vis(model_dir, train_desc_file, weights_file=None):
""" Get the prediction using the model, and visualize softmax outputs, able
to predict multiple inputs.
Params:
model_dir (str): Trained speech model or None. If None given will ask
code to make model.
train_desc_file(str): Path to the training file used to train this
model
weights_file(str): Path to stored weights file for model being made
"""
datagen = DataGenerator()
datagen.load_train_data(train_desc_file)
datagen.fit_train(100)
if model_dir is None:
assert weights_file is not None
print("""Make and store new model into model, e.g.
>>> model_wrp = HalfPhonemeModelWrapper()
>>> model = model_wrp.compile(nodes=1000, recur_layers=5,
conv_context=5)
""")
model = prompt_loop('[model=]> ', locals())['model']
model.load_weights(weights_file)
else:
model = load_model(model_dir, weights_file)
print("""Make and store test function to test_fn, e.g.
>>> test_fn = model_wrp.compile_output_fn()
""")
test_fn = prompt_loop('[test_fn=]> ', locals())['test_fn']
while True:
try:
test_file = raw_input('Input file: ')
except EOFError:
comm_mode = True
while comm_mode:
try:
comm = raw_input("[w: load wieghts\t s: shell ] > ")
if comm.strip() == 'w':
w_path = raw_input("weights file path: ").strip()
model.load_weights(w_path)
if comm.strip() == 's':
prompt_loop('> ', locals())
except EOFError:
comm_mode = False
except Exception as exc:
print(exc)
continue
if test_file.strip() == '':
break
try:
inputs = [datagen.featurize(test_file)]
except Exception as exc:
print(exc)
continue
prediction = np.squeeze(test_fn([inputs, True]))
softmax_file = "softmax.npy".format(test_file)
softmax_img_file = "softmax.png".format(test_file)
print("Prediction: {}".format(argmax_decode(prediction)))
print("Saving network output to: {}".format(softmax_file))
print("As image: {}".format(softmax_img_file))
np.save(softmax_file, prediction)
sm = softmax(prediction.T)
sm = np.vstack((sm[0], sm[2], sm[3:][::-1]))
fig, ax = plt.subplots()
ax.pcolor(sm, cmap=plt.cm.Greys_r)
column_labels = [chr(i)
for i in range(97, 97 + 26)] + ['space', 'blank']
ax.set_yticks(np.arange(sm.shape[0]) + 0.5, minor=False)
ax.set_yticklabels(column_labels[::-1], minor=False)
plt.savefig(softmax_img_file)
