def train_model(model, X_train, X_test, y_train, y_test):
cp = ModelCheckpoint('model-{epoch:03d}.h5', monitor = 'val_loss', verbose = 0, save_best_only=True, mode='auto')
model.compile(loss='mean_squared_error', optimizer=Nadam(lr=1.0e-4))
model.fit_generator(batch_generator('data', X_train, y_train, 40, True),
20000,
30,
max_q_size=1,
validation_data=batch_generator('data', X_test, y_test, 40, False),
nb_val_samples=len(X_test),
callbacks=[cp],
verbose=1)
anomalous_test_data_list_artificial_hot_towers = os.listdir(
data_folder.replace("good_2016",
"bad_2016/hot_towers")) #artificially created
anomalous_test_data_list_artificial_missing_modules = os.listdir(
data_folder.replace("good_2016",
"bad_2016/missing_modules")) #artificially created
#print(anomalous_test_data_list_natural)
# print(anomalous_test_data_list_artificial_hot_towers)
# print(anomalous_test_data_list_artificial_missing_modules)
print("Current training set is made from " + str(len(train_data_list)) +
" files and has " +
str(helper_functions.get_num_samples(train_data_list)) + " examples")
my_training_data_generator = helper_functions.batch_generator(
4, train_data_list, group=image_type, prep_level=args.prep_level)
training_losses = []
for batch in my_training_data_generator:
loss = trained_model.evaluate(batch, batch, batch_size=4, verbose=0)
training_losses.append(loss)
print("Current test set is made from " + str(len(test_data_list)) +
" files and has " +
str(helper_functions.get_num_samples(test_data_list)) + " examples")
my_test_data_generator = helper_functions.batch_generator(
4, test_data_list, group=image_type, prep_level=args.prep_level)
test_losses = []
for batch in my_test_data_generator:
loss = trained_model.evaluate(batch, batch, batch_size=4, verbose=0)
+ str(args.regularize)
# create a secondary validation model to run our similarity checks during training
similarity = dot([word_embedding, context_embedding], axes=1, normalize=True)
validation_model = Model(inputs=[word_index, context_index], outputs=[similarity])
sim_cb = helpers.SimilarityCallback(validation_model=validation_model)
loss_plot = TensorBoard(log_dir=train_name + '_logs', write_graph=False)
earlystopping = EarlyStopping(monitor='loss', min_delta=0.0001, patience=1, verbose=1,
mode='auto')
# How many times per epoch we will ask the batch generator to yield a batch?
steps = no_train_pairs / batch_size
# Let's start training!
start = time.time()
history = keras_model.fit_generator(
helpers.batch_generator(wordpairs, vocab_dict, vocab_size, negative, batch_size,
args.use_neighbors, neighbors_count),
callbacks=[sim_cb, loss_plot, earlystopping], steps_per_epoch=steps, epochs=args.epochs,
workers=cores, verbose=2)
end = time.time()
print('Training took:', int(end - start), 'seconds', file=sys.stderr)
# Saving the resulting vectors:
filename = train_name + '_' + run_name + '.vec.gz'
helpers.save_word2vec_format(filename, vocab_dict, word_embedding_layer.get_weights()[0])
backend.clear_session()
# sim_cb = helpers.SimilarityCallback(validation_model=validation_model)
loss_plot = TensorBoard(log_dir=train_name + '_logs', write_graph=False)
earlystopping = EarlyStopping(monitor='loss',
min_delta=0.0001,
patience=1,
verbose=1,
mode='auto')
# How many times per epoch we will ask the batch generator to yield a batch?
steps = no_train_pairs / batch_size
# Let's start training!
start = time.time()
history = keras_model.fit_generator(helpers.batch_generator(
wordpairs, vocab_dict, vocab_size, negative, batch_size,
args.use_neighbors, neighbors_count),
callbacks=[loss_plot, earlystopping],
steps_per_epoch=steps,
epochs=args.epochs,
workers=1,
verbose=2)
end = time.time()
print('Training took:', int(end - start), 'seconds', file=sys.stderr)
# Saving the resulting vectors:
filename = train_name + '_' + run_name + '.vec.gz'
helpers.save_word2vec_format(filename, vocab_dict,
word_embedding_layer.get_weights()[0])
X_val = { 'ims': np.array( validation_ims ),
'bxs': all_validation_bxs,
'detectors': val_detectors_mask,
'true_bxs': val_matching_true_boxes }
data = {'train': X_train, 'val': X_val
### TRAIN
num_epochs = 10
batch_size = 32
steps_per_epoch_fit = np.ceil(len(data[ 'train' ][ 'ims' ]) / batch_size)
steps_per_epoch_val = np.ceil(len(data[ 'val' ][ 'ims' ]) / batch_size)
fit_gen = batch_generator( 'train', batch_size, shuffle = True)
val_gen = batch_generator( 'val' , batch_size)
model_body, model, global_step, metrics = create_model(anchors, class_names, reset_weights=False)
model.compile( optimizer='adam',
loss = {'yolo_loss': lambda y_true, y_pred: y_pred},
metrics = metrics)
annealing_step = (steps_per_epoch_fit*num_epochs) / 4
hist = model.fit_generator(generator = fit_gen,
steps_per_epoch = steps_per_epoch_fit,
validation_data = val_gen,
validation_steps = steps_per_epoch_val,
epochs = num_epochs,
data_folder=os.environ["DATA"]
except KeyError:
"Please cd into the module's base folder and run set_env from there."
file_list=os.listdir(data_folder)
np.random.seed(1)
np.random.shuffle(file_list)
train_data_list=file_list[0:160] #choosing 63 here keeps ~80% of data for testing, rest for training and val, need to automatize this
test_data_list=file_list[160:]
print("Current training set is made from "+str(len(train_data_list))+" files and has "+str(helper_functions.get_num_samples(train_data_list))+" examples: ", train_data_list)
print("Current test set is made from "+str(len(test_data_list))+" files and has "+str(helper_functions.get_num_samples(test_data_list))+" examples: ", test_data_list)
my_test_data_generator=helper_functions.batch_generator(22900,test_data_list,image_type)
y_true_test=[]
y_predicted_test=[]
for batch in my_test_data_generator:
test_x,test_y=batch
predicted=trained_model.predict(test_x)
y_true_test=test_y[:,0]
y_predicted_test=predicted[:,0]
my_training_data_generator=helper_functions.batch_generator(44600,train_data_list,image_type)
y_true_train=[]
emoji_index, _, emoji_sorted = build_emoji_index(vocab_f, emoji_64)
# build vocab
word2index, index2word = build_vocab(vocab_f)
start_i, end_i = word2index['<s>'], word2index['</s>']
vocab_size = len(word2index)
classifier = EmojiClassifier(batch_size, vocab_size, emoji_num, embed_size,
num_unit, num_gpu)
# build data
train_data = build_data(train_ori_f, train_rep_f, word2index)
test_data = build_data(test_ori_f, test_rep_f, word2index)
test_batches = batch_generator(test_data,
start_i,
end_i,
batch_size,
permutate=False)
print_out("*** CLASSIFIER DATA READY ***")
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
saver = tf.train.Saver()
with tf.Session(config=config) as sess:
classifier.set_sess(sess)
classifier.set_emoji_index(emoji_index)
global_step = best_step = 1
start_epoch = best_epoch = 1
best_loss = 1000.
sess.run(tf.global_variables_initializer())