def predict(datagen: SoundDatagen, model_name: str) -> np.array:
""" Predicts results on test, using the best model. """
clips_per_sample = datagen.get_clips_per_sample()
print("predicting results")
model = keras.models.load_model(get_best_model_path(model_name))
y_test = model.predict_generator(datagen, verbose=1)
print("y_test.shape after predict", y_test.shape)
pos = 0
y_merged = []
for count in clips_per_sample:
if count != 0:
y_merged.append(
merge_predictions(y_test[pos:pos + count], "max", 0))
else:
y_merged.append(np.zeros_like(y_merged[0]))
pos += count
y_test = np.array(y_merged)
print("y_test.shape after merge", y_test.shape)
return y_test
def train_model(train_datagen: SoundDatagen, val_datagen: SoundDatagen,
name: str) -> None:
""" Creates model, trains it and saves it to the file. """
shape = train_datagen.shape()
x = inp = keras.layers.Input(shape=shape[1:])
x = keras.layers.Convolution2D(32, (4,10), padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation("relu")(x)
x = keras.layers.MaxPool2D()(x)
x = keras.layers.Convolution2D(32, (4,10), padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation("relu")(x)
x = keras.layers.MaxPool2D()(x)
x = keras.layers.Convolution2D(32, (4,10), padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation("relu")(x)
x = keras.layers.MaxPool2D()(x)
x = keras.layers.Convolution2D(32, (4,10), padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation("relu")(x)
x = keras.layers.MaxPool2D()(x)
x = keras.layers.Convolution2D(32, (4,10), padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation("relu")(x)
x = keras.layers.MaxPool2D()(x)
x = keras.layers.Convolution2D(32, (4,10), padding="same")(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation("relu")(x)
x = keras.layers.MaxPool2D()(x)
x = keras.layers.Flatten()(x)
x = keras.layers.Dense(64)(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation("relu")(x)
out = keras.layers.Dense(NUM_CLASSES, activation="softmax")(x)
model = keras.models.Model(inputs=inp, outputs=out)
model.summary()
map3 = Map3Metric(val_datagen, name)
model.compile(loss="categorical_crossentropy", optimizer="adam",
metrics=["accuracy"])
model.fit_generator(train_datagen, epochs=NUM_EPOCHS,
verbose=1, shuffle=False,
# use_multiprocessing=False,
use_multiprocessing=True, workers=12,
validation_data=val_datagen, callbacks=[map3])
print("best MAP@3 value: %.04f at epoch %d" % (map3.best_map3, map3.best_epoch))
assert (len(train_files) == 9473)
train_labels = train_df["label"]
test_files = np.array(find_files("../data/audio_test/"))
test_idx = [os.path.basename(f) for f in test_files]
build_caches(np.concatenate([train_files, test_files]))
fit_labels(train_labels)
if USE_HYPEROPT:
x_train, x_val, y_train, y_val = train_test_split(train_files,
train_labels,
test_size=TEST_SIZE,
shuffle=False)
train_datagen = SoundDatagen(x_train, y_train)
val_datagen = SoundDatagen(x_val, y_val)
'''
cnn_dropout_coeff = float(params["cnn_dropout_coeff"])
cnn_kern_width = int(params["cnn_kern_width"])
cnn_dim_decay = int(params["cnn_dim_decay"])
cnn_depth_growth = int(params["cnn_depth_growth"])
conv2d_depth = int(params["conv2d_depth"])
conv2d_len = int(params["conv2d_len"])
pooling = params["pooling"]
'''
hyperopt_space = {
"num_cnn_layers":
hp.quniform("num_cnn_layers", 4, 7, 1),
"cnn_depth":
train_labels = train_df["label"]
test_files = np.array(find_files("../data/audio_test/"))
test_idx = [os.path.basename(f) for f in test_files]
build_caches(np.concatenate([train_files, test_files]))
fit_labels(train_labels)
if not ENABLE_KFOLD:
x_train, x_val, y_train, y_val = train_test_split(train_files,
train_labels,
test_size=TEST_SIZE,
shuffle=False)
if not PREDICT_ONLY:
train_model(SoundDatagen(x_train, y_train),
SoundDatagen(x_val, y_val), "nofolds")
pred = predict(SoundDatagen(test_files, None), "nofolds")
pred = encode_predictions(pred)
else:
kf = KFold(n_splits=KFOLDS, shuffle=False)
pred = np.zeros((len(test_idx), KFOLDS, NUM_CLASSES))
for k, (train_idx, val_idx) in enumerate(kf.split(train_files)):
print("fold %d ==============================================" % k)
name = "fold_%d" % k
x_train, y_train = train_files[train_idx], train_labels[train_idx]
x_val, y_val = train_files[val_idx], train_labels[val_idx]
