def get_model(self):
if self.encoder_type == "HC":
self.model = HarmonicCNN()
elif self.encoder_type == "SC":
self.model = resemul(block_type=self.block)
elif self.encoder_type == "MS":
self.model = resemul(ms=True, block_type="rese")
def get_model(self):
if self.encoder_type == "HC":
self.wave_encoder = WaveEncoder()
elif self.encoder_type == "SC":
self.wave_encoder = resemul(block_type=self.block)
elif self.encoder_type == "MS":
self.wave_encoder = resemul(ms=True, block_type="rese")
self.wave_projector = WaveProjector(128)
self.classifier = SupervisedClassifier()
self.tag_encoder = TagEncoder()
self.tag_decoder = TagDecoder(50)
def get_model(weights):
"""
Define the model arch + load weights
"""
# network arch
print("Creating model..")
input_tensor = Input(shape=(16000, ))
output_tensor = resemul(input_tensor,
block_type="rese",
init_features=128,
amplifying_ratio=16,
drop_rate=0.5,
weight_decay=0.0,
num_classes=31)
model = Model(input_tensor, output_tensor)
sgd = optimizers.SGD(lr=0.01, momentum=0.9, nesterov=True)
model.compile(sgd, "categorical_crossentropy", metrics=["accuracy"])
print("Model compiled.")
model.load_weights(weights)
print("weights loaded..")
return model
from utils import data_generator
from kaggle_utils import score_generator
import sys
import os
from fgsm_callback import fgsm_callback
try:
import cPickle as pickle
except:
import pickle
# network arch
input_tensor = Input(shape=(16000, ))
output_tensor = resemul(input_tensor,
block_type="rese",
init_features=128,
amplifying_ratio=16,
drop_rate=0.5,
weight_decay=0.0,
num_classes=31)
model = Model(input_tensor, output_tensor)
print(model.summary())
sgd = optimizers.SGD(lr=0.01, momentum=0.9, nesterov=True)
model.compile(sgd, "categorical_crossentropy", metrics=["accuracy"])
print("Model compiled.")
# scoring
if sys.argv[1] == "score":
model.load_weights(
"/home/tejaswin.p/Can-You-Hear-Me/checkpoints/2018-06-22_08:40:45.299850/weights.04-0.25.hdf5"
)
print("Weights loaded.")
def train(initial_lr, stage_train_dir, ckpt_path=None, initial_epoch=0):
# Prepare TFRecord batches.
x_train, y_train = batch_inputs(
file_pattern=make_path(args.data_dir, 'tfrecord',
'train-????-of-????.tfrecord'),
batch_size=args.batch_size,
is_training=True,
num_read_threads=args.num_read_threads,
examples_per_shard=args.num_audios_per_shard *
args.num_segments_per_audio,
shard_queue_name='train_filename_queue',
example_queue_name='train_input_queue')
x_val, y_val = batch_inputs(file_pattern=make_path(
args.data_dir, 'tfrecord', 'val-????-of-????.tfrecord'),
batch_size=args.batch_size,
is_training=False,
num_read_threads=1,
examples_per_shard=args.num_audios_per_shard *
args.num_segments_per_audio,
shard_queue_name='val_filename_queue',
example_queue_name='val_input_queue')
# Create a model.
out = resemul(x_train,
block_type=args.block,
multi=not args.no_multi,
amplifying_ratio=args.alpha,
drop_rate=args.dropout,
weight_decay=args.weight_decay)
model = TFRecordModel(inputs=x_train, val_inputs=x_val, outputs=out)
# Load weights from a checkpoint if exists.
if ckpt_path:
print('=> Load weights from "{}".'.format(ckpt_path))
model.load_weights(ckpt_path)
# Setup an optimizer and compile the models.
optimizer = SGD(lr=initial_lr,
momentum=args.momentum,
decay=1e-6,
nesterov=True)
model.compile_tfrecord(optimizer,
loss='binary_crossentropy',
y=y_train,
y_val=y_val)
# --- Setup callbacks.
# Setup a TensorBoard callback.
tensor_board = TensorBoard(log_dir=stage_train_dir)
# Use early stopping mechanism.
early_stopping = EarlyStopping(monitor='val_loss', patience=args.patience)
# Setup a checkpointer.
checkpointer = ModelCheckpoint(make_path(
stage_train_dir, 'ckpt-e{epoch:03d}-l{val_loss:.4f}.hdf5'),
monitor='val_loss',
save_best_only=True)
# Setup a CSV logger.
csv_logger = CSVLogger(make_path(stage_train_dir, 'training.csv'),
append=True)
# Kick-off the training!
num_train, num_val, num_test = 15250, 1529, 4332
train_steps = calculate_steps(num_train, args.num_segments_per_audio,
args.batch_size)
val_steps = calculate_steps(num_val, args.num_segments_per_audio,
args.batch_size)
model.fit_tfrecord(
epochs=100,
initial_epoch=initial_epoch,
steps_per_epoch=train_steps,
validation_steps=val_steps,
callbacks=[tensor_board, early_stopping, checkpointer, csv_logger])
# The end of the stage. Evaluate on test set.
best_ckpt_path, *_ = find_best_checkpoint(stage_train_dir)
print(
'=> The end of the stage. Start evaluation on test set using checkpoint "{}".'
.format(best_ckpt_path))
evaluate(model, args.data_dir, num_test, args.num_audios_per_shard)