imageWidth = trainSet.imgSize[1]
nChannels = trainSet.imgs.shape[3]
Xtrain = trainSet.imgs
Xvalid = validationSet.imgs
# import pdb; pdb.set_trace()
Ytrain = getLabelsForDataset(trainSet)
Yvalid = getLabelsForDataset(validationSet)
testIdxsTrainSet = range(len(Xvalid))
testIdxsValidSet = range(len(Xvalid))
meanImg = trainSet.meanImg
stdDevImg = trainSet.stdDevImg
initLandmarks = trainSet.initLandmarks[0].reshape((1, 136))
dan = DAN(initLandmarks)
STAGE = 2
with tf.Session() as sess:
Saver = tf.train.Saver()
Writer = tf.summary.FileWriter("logs/", sess.graph)
if STAGE < 2:
sess.run(tf.global_variables_initializer())
else:
Saver.restore(sess,'./Model/Model')
print('Pre-trained model has been loaded!')
# Landmark68Test(MeanShape,ImageMean,ImageStd,sess)
print("Starting training......")
for epoch in range(2):
if not os.path.exists(log_path):
os.mkdir(log_path)
model_path = 'model'
if not os.path.exists(model_path):
os.mkdir(model_path)
train_number, valid_number = samples_counter()
nb_epocs = 20
batch_size = 128
stage = 2
meanshape = init_meanshape()
model = DAN(meanshape)
train_image, train_landmark = read_and_decode('data/train_dataset.tfrecords', batch_size)
valid_image, valid_landmark = read_and_decode('data/valid_dataset.tfrecords', batch_size)
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.local_variables_initializer())
sess.run(tf.global_variables_initializer())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess = sess, coord = coord)
if stage >= 2:
saver.restore(sess,'model/model_step2_02.ckpt')
def main(args):
"""Experiment logic"""
# Get file separator and construct paths
sep = "\t" if args.file_type == "tsv" else ","
train_path = os.path.join(args.data_dir, "train.{}".format(args.file_type))
test_path = os.path.join(args.data_dir, "test.{}".format(args.file_type))
# Read column headings
headings = pd.read_csv(train_path, sep=sep, nrows=1).columns
text, label = "text", "gold_label_{}".format(args.task_type)
if args.elmo:
from elmo import TabularReader, ElmoLoader
# Pretrained urls
options_file = "https://s3-us-west-2.amazonaws.com/allennlp/models/elmo/2x4096_512_2048cnn_2xhighway/elmo_2x4096_512_2048cnn_2xhighway_options.json"
weight_file = "https://s3-us-west-2.amazonaws.com/allennlp/models/elmo/2x4096_512_2048cnn_2xhighway/elmo_2x4096_512_2048cnn_2xhighway_weights.hdf5"
# Read dataset
reader = TabularReader(text, label, sep)
loader = ElmoLoader(reader, train_path, test_path, args.batch_dims)
# Build model
label_map = loader.label_map
embedding_size = 1024
model = DAN(to_int(args.layers),
len(label_map),
embedding_size=embedding_size,
elmo_config=(options_file, weight_file))
else:
# Build data loader
loader = DataLoader(
args.data_dir,
args.file_type,
headings,
text,
label,
to_int(args.batch_dims),
(args.glove_type, args.glove_dim),
args.temp_dir,
)
# Build model
vocab, label_map = loader.vocab, loader.label_map
model = DAN(to_int(args.layers),
len(label_map),
vocab_size=len(vocab),
embedding_size=args.glove_dim,
pretrained_vecs=vocab.vectors)
# Define training functions
optimiser = optim.SGD(model.parameters(), lr=args.lr)
loss_fn = nn.CrossEntropyLoss()
# Train
logging.info("\n\nStarting training...\n\n")
if args.num_processes > 1:
model.share_memory()
processes = []
for pid in range(args.num_processes):
p = mp.Process(target=run.training_process,
args=(pid, loader, model, optimiser, loss_fn,
(args.num_steps // args.num_processes)))
p.start()
processes.append(p)
for p in processes:
p.join()
else:
report_every = int(args.num_steps * 0.01)
losses = run.train(loader, model, optimiser, loss_fn, label_map,
args.num_steps, report_every)
if args.plot:
logging.info("\n\nPlotting training schedule...\n\n")
plot_loss(losses, report_every, args.temp_dir)
# Save the trained model
logging.info("\n\nNow saving...\n\n")
torch.save(model, os.path.join(args.temp_dir, "saved_model.pt"))
# Test
model_acc = run.test(loader, label_map, args.temp_dir)
if args.baseline:
logging.info(
"\n\nComparing with multinomial naive bayes baseline...\n\n")
from bayes import multi_nb
train, test = pd.read_csv(train_path, sep=sep), pd.read_csv(test_path,
sep=sep)
train_txt, test_txt = (train[text], train[label]), (test[text],
test[label])
base_acc = multi_nb(train_txt, test_txt)
logging.info("Model accuracy: {:.6g}".format(model_acc))
logging.info("Baseline accuracy: {:.6g}".format(base_acc))
logging.info("{}".format(
"Model wins!" if model_acc > base_acc else "Baseline wins!"))