def main(CONST, data):
config = get_config("BiRNN")
#if not CONST.TRAIN:
# config.num_steps = 1
config.vocab_size = len(data["embeddings"])
config.max_max_epoch = CONST.MAX_EPOCH
from util.evaluations import print_config
print_config(config)
tf.reset_default_graph()
# start graph and session
# config=tf.ConfigProto(log_device_placement=True) pass to tf.Session
# to see which devices all operations can run on
with tf.Graph().as_default(), tf.Session(config=session_conf) as session:
with tf.variable_scope(
"model",
reuse=None,
initializer=tf.contrib.layers.xavier_initializer()):
training_model = RNNModel(is_training=CONST.TRAIN,
config=config) # model class
if CONST.TRAIN:
tf.initialize_all_variables().run()
session.run(training_model.embedding.assign(
data["embeddings"])) # train
# Reload save epoch training time
if CONST.RELOAD_TRAIN:
saver = tf.train.Saver()
saver.restore(sess=session,
save_path=CONST.SLOT3_TARGET_MODEL_PATH +
"slot3_target")
all_losses = []
train_writer = run_metadata = run_options = False
if config.get_summary:
# session = tf.InteractiveSession()
train_writer = tf.train.SummaryWriter(
CONST.SLOT3_TARGET_MODEL_PATH +
"attention_graph/slot3_target", session.graph)
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
for i in range(config.max_max_epoch): # setup epoch
train_perplexity, loss, losses = run_epoch(
session,
training_model,
data["x_train"],
data["p_train"],
train_writer,
run_options,
run_metadata,
targets=data["l_train"],
verbose=True)
all_losses = all_losses + [np.mean(losses)]
print("Epoch: %d Avg. Total Mean Loss: %.6f" %
(i + 1, np.mean(all_losses)))
from util.evaluations import print_config
print_config(config)
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
# plt.plot([np.mean(all_losses[i-50:i]) for i in range(len(all_losses))])
figure_name = CONST.OUT_DIR + "loss/" + "losses_slot3" + config.__class__.__name__ + ".png"
x = [i for i in range(len(all_losses))]
plt.plot(np.array(x), np.array(all_losses))
plt.savefig(figure_name)
save_pickle(
CONST.DATA_DIR + config.__class__.__name__ + "_slot3_target",
all_losses)
print("saved slot3 losses.png and losses data", figure_name)
print("loss: ", figure_name)
print(
"loss data: ", CONST.DATA_DIR + config.__class__.__name__ +
"_slot3_target" + ".pickle")
saver = tf.train.Saver(tf.all_variables())
path = saver.save(sess=session,
save_path=CONST.SLOT3_TARGET_MODEL_PATH +
"slot3_target")
print("model saved: " + path)
if config.get_summary:
train_writer.close()
session.close() # doesn't seem to close under scope??
if not CONST.TRAIN:
with tf.variable_scope(
"model", reuse=True): # reuse scope to evaluate model :-)
#config.batch_size = len(data["x_test"])
# tf.initialize_all_variables().run()
# set embeddings again
session.run(training_model.embedding.assign(
data["embeddings"]))
# Load Data Back Trained Model
saver = tf.train.Saver()
saver.restore(sess=session,
save_path=CONST.SLOT3_TARGET_MODEL_PATH +
"slot3_target")
# Initialize Model Graph
validation_model = RNNModel(is_training=False, config=config)
predictions = []
alphas = []
# Get Predictions
for step, (x, y, t) in enumerate(
semeval_itterator(data["x_test"],
data["p_test"],
validation_model.batch_size,
validation_model.num_steps,
target=data["l_test"],
shuffle_examples=False)):
# do not pass in test targets
if CONST.HEATMAP:
alpha = session.run(validation_model.alpha,
{validation_model.input_data: x
}) #, training_model.targets: y})
alphas = alphas + alpha.tolist()
else:
pred = session.run(validation_model.predictions,
{validation_model.input_data: x
}) #, training_model.targets: y})
predictions = predictions + pred.tolist()
if not CONST.HEATMAP:
even_batch = len(data["x_test"]) % config.batch_size
remove_added_batch = config.batch_size - even_batch
del predictions[-remove_added_batch:]
predictions = np.asarray(predictions)
from util.evaluations import print_config
print_config(config)
from util.evaluations import evaluate_multiclass
y = [np.asarray(e) for e in data["p_test"]]
save_pickle(
CONST.DATA_DIR + config.__class__.__name__ +
"slot3_predictions", {
"predictions": predictions,
"y": y
})
evaluate_multiclass(predictions, y, True)
print("predictions saved")
# session.close() # doesn't seem to close under scope??
elif CONST.HEATMAP:
even_batch = len(data["x_test"]) % config.batch_size
remove_added_batch = config.batch_size - even_batch
del alphas[-remove_added_batch:]
distance = 0
sentences = data["test_sentences"]
print("computing heatmaps and avg distance...")
from util.heatmap import avg_distance_and_heatmaps
avg_distance_and_heatmaps(alphas, sentences,
CONST.SENTENCE_PATH + "hursh/")
def train_task(CONST, data):
config = multitask()
config.vocab_size = len(data["embeddings"])
tf.reset_default_graph()
with tf.Graph().as_default(), tf.Session() as session:
with tf.variable_scope(
"model",
reuse=None,
initializer=tf.contrib.layers.xavier_initializer()):
training_model = slot1_and_slot3_model(
is_training=True, config=config) # model class
if CONST.TRAIN:
tf.initialize_all_variables().run()
# Check if Model is Saved and then Load
if check_saved_file(CONST.MULTITASK_CHECKPOINT_PATH +
"checkpoint"):
saver = tf.train.Saver()
saver.restore(sess=session,
save_path=CONST.MULTITASK_CHECKPOINT_PATH)
session.run(
training_model.embeddings.embeddings.assign(
data["embeddings"])) # train
slot1_losses = []
slot3_losses = []
train_writer = run_metadata = run_options = False
if config.get_summary:
# session = tf.InteractiveSession()
train_writer = tf.train.SummaryWriter(
CONST.SLOT1_MODEL_PATH + "attention_graph/" +
config.__class__.__name__, session.graph)
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
for i in range(config.max_max_epoch): # setup epoch
train_perplexity, slot1_loss, slot3_loss = run_epoch(
session,
training_model,
data["x_train"],
data["y_train"],
data["p_train"],
train_writer,
run_options,
run_metadata,
category=data["a_train"],
verbose=True,
config=config)
slot1_losses = slot1_losses + [np.mean(slot1_loss)]
slot3_losses = slot3_losses + [np.mean(slot3_loss)]
print(
"Epoch: %d Avg. Total Mean Loss slot1: %.6f slot3: %.6f" %
(i + 1, np.mean(slot1_losses), np.mean(slot3_losses)))
# Output Config/Losses
from util.evaluations import print_config
print_config(config)
# Save CheckPoint
saver = tf.train.Saver(tf.all_variables())
path = saver.save(sess=session,
save_path=CONST.MULTITASK_CHECKPOINT_PATH)
print("model saved: " + path)
# Save Losses for Later
loss = {"slot1": slot1_losses, "slot3": slot3_losses}
save_pickle(
CONST.DATA_DIR + config.__class__.__name__ + "multi_task",
loss)
# Try and plot the losses
import matplotlib.pyplot as plt
x = [i for i in range(len(slot1_losses))]
plt.plot(np.array(x), np.array(slot1_losses))
plt.plot(np.array(x), np.array(slot3_losses))
plt.savefig("losses_" + config.__class__.__name__ + ".png")
if not CONST.TRAIN:
with tf.variable_scope(
"model", reuse=True): # reuse scope to evaluate model
validation_model = slot1_and_slot3_model(
is_training=True, config=config) # model class
session.run(
validation_model.embeddings.embeddings.assign(
data["embeddings"])) # load embeddings
saver = tf.train.Saver()
saver.restore(sess=session,
save_path=CONST.MULTITASK_CHECKPOINT_PATH)
slot1_predictions = []
slot3_predictions = []
# Get Predictions
for step, (x, y_1, y_3, e, a) in enumerate(
semeval_itterator(data["x_test"],
data["y_test"],
config.batch_size,
config.num_steps,
category=data["aspects"],
polarity=data["p_test"])):
slot1_prediction, slot3_prediction = session.run(
[
validation_model.slot1_model.predictions,
validation_model.slot3_model.predicitions
], {
validation_model.input_data: x,
validation_model.targets: y,
validation_model.input_e: e,
validation_model.input_a: a
})
slot1_predictions = slot1_predictions + slot1_prediction.tolist(
)
slot3_predictions = slot3_predictions + slot3_prediction.tolist(
)
even_batch = len(data["x_test"]) % config.batch_size
remove_added_batch = config.batch_size - even_batch
del slot1_predictions[-remove_added_batch:]
del slot3_predictions[-remove_added_batch:]
slot1_predictions = np.asarray(slot1_predictions)
slot3_predictions = np.asarray(slot3_predictions)
# print congiuration for test predictions
from util.evaluations import print_config
print_config(config)
# save predictions
predictions = {
"slot1": slot1_predictions,
"slot3": slot3_predictions,
"slot1_y": data["y_test"],
"slot3_y": data["p_test"]
}
save_pickle(
CONST.DATA_DIR + config.__class__.__name__ +
"_predictions", predictions)
print(
"predictions saved to file ", CONST.DATA_DIR +
config.__class__.__name__ + "_predictions")
from util.evaluations import evaluate_multilabel
from util.evaluations import evaluate_multiclass
evaluate_multilabel(predictions.slot1, predictions.slot1_y,
CONST.THRESHOLD)
evaluate_multiclass(predictions.slot3, predictions.slot3_y,
True)
def train_task(CONST, data):
config = multitask()
config.vocab_size = len(data["embeddings"])
config.max_max_epoch = CONST.MAX_EPOCH
tf.reset_default_graph()
with tf.Graph().as_default(), tf.Session(config=session_conf) as session:
with tf.variable_scope(
"model",
reuse=None,
initializer=tf.contrib.layers.xavier_initializer()):
training_model = slot1_and_slot3_model(
is_training=CONST.TRAIN, config=config) # model class
if CONST.TRAIN:
tf.initialize_all_variables().run()
# Check if Model is Saved and then Load
if CONST.RELOAD_TRAIN:
saver = tf.train.Saver()
print("CHECKPATH",
CONST.MULTITASK_CHECKPOINT_PATH_HURSH + "task")
saver.restore(sess=session,
save_path=CONST.MULTITASK_CHECKPOINT_PATH_HURSH +
"task")
session.run(
training_model.embeddings.embeddings.assign(
data["embeddings"])) # train
slot1_losses = []
slot3_losses = []
slot4_losses = []
train_writer = run_metadata = run_options = False
if config.get_summary:
# session = tf.InteractiveSession()
train_writer = tf.train.SummaryWriter(
CONST.MULTITASK_CHECKPOINT_PATH_HURSH +
"attention_graph/" + config.__class__.__name__,
session.graph)
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
for i in range(config.max_max_epoch): # setup epoch
train_perplexity, slot1_loss, slot3_loss = run_epoch(
session,
training_model,
data["x_train"],
data["y_train"],
data["p_train"],
train_writer,
run_options,
run_metadata,
target=data["l_train"],
verbose=True,
config=config)
slot1_losses = slot1_losses + [np.mean(slot1_loss)]
slot3_losses = slot3_losses + [np.mean(slot3_loss)]
print(
"Epoch: %d Avg. Total Mean Loss slot1: %.6f slot3: %.6f" %
(i + 1, np.mean(slot1_losses), np.mean(slot3_losses)))
# Output Config/Losses
from util.evaluations import print_config
print_config(config)
# Save Losses for Later
loss = {"slot1": slot1_losses, "slot3": slot3_losses}
save_pickle(
CONST.DATA_DIR + config.__class__.__name__ + "multi_task",
loss)
# Save CheckPoint
saver = tf.train.Saver(tf.all_variables())
path = saver.save(sess=session,
save_path=CONST.MULTITASK_CHECKPOINT_PATH_HURSH +
"task")
print("model saved: " + path)
if config.get_summary:
train_writer.close()
session.close()
# doesn't seem to close under scope??
# Try and plot the losses
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
x = [i for i in range(len(slot1_losses))]
plt.plot(np.array(x), np.array(slot1_losses))
plt.plot(np.array(x), np.array(slot3_losses))
plt.legend(['slot1', 'slot3'], loc='upper right')
plt.savefig("losses_" + config.__class__.__name__ + ".png")
if not CONST.TRAIN:
with tf.variable_scope(
"model", reuse=True): # reuse scope to evaluate model
#validation_model = slot1_and_slot3_model(is_training=False, config=config) # model class
session.run(
training_model.embeddings.embeddings.assign(
data["embeddings"])) # load embeddings
saver = tf.train.Saver()
saver.restore(sess=session,
save_path=CONST.MULTITASK_CHECKPOINT_PATH_HURSH +
"task")
slot1_predictions = []
slot3_predictions = []
alphas = []
# Get Predictions
for step, (x, y_1, y_3, t) in enumerate(
semeval_itterator(data["x_test"],
data["y_test"],
config.batch_size,
config.num_steps,
target=data["l_test"],
polarity=data["p_test"],
shuffle_examples=False)):
if CONST.HEATMAP:
alpha = session.run(
[training_model.slot3_model.alpha], {
training_model.embeddings.source_data: x,
training_model.slot1_model.targets: y_1,
training_model.slot3_model.targets: y_3
})
alphas = alphas + alpha[0].tolist()
else:
slot1_prediction, slot3_prediction = session.run(
[
training_model.slot1_model.predictions,
training_model.slot3_model.predictions
], {
training_model.embeddings.source_data: x,
training_model.slot1_model.targets: y_1,
training_model.slot3_model.targets: y_3
})
slot1_predictions = slot1_predictions + slot1_prediction.tolist(
)
slot3_predictions = slot3_predictions + slot3_prediction.tolist(
)
if not CONST.HEATMAP:
even_batch = len(data["x_test"]) % config.batch_size
remove_added_batch = config.batch_size - even_batch
del slot1_predictions[-remove_added_batch:]
del slot3_predictions[-remove_added_batch:]
slot1_predictions = [
np.asarray(x) for x in slot1_predictions
]
slot3_predictions = [
np.asarray(x) for x in slot3_predictions
]
# print congiuration for test predictions
from util.evaluations import print_config
print_config(config)
slot3_y = [np.asarray(x) for x in data["p_test"]]
# save predictions
predictions = {
"slot1": slot1_predictions,
"slot3": slot3_predictions,
"slot1_y": data["y_test"],
"slot3_y": slot3_y
}
save_pickle(
CONST.DATA_DIR + config.__class__.__name__ +
"_predictions", predictions)
print(
"predictions saved to file ", CONST.DATA_DIR +
config.__class__.__name__ + "_predictions")
from util.evaluations import evaluate_multilabel
from util.evaluations import evaluate_multiclass
from util.evaluations import find_best_slot1
# evaluate_multilabel(predictions["slot1"], predictions["slot1_y"], CONST.THRESHOLD)
print("\nslot3 sentiment ...\n")
evaluate_multiclass(np.asarray(predictions["slot3"]),
predictions["slot3_y"], True)
print("\nfinding best threshold slot1 E\A pairs...\n")
find_best_slot1("multitask_hursh",
np.asarray(predictions["slot1"]),
predictions["slot1_y"])
elif CONST.HEATMAP:
even_batch = len(data["x_test"]) % config.batch_size
remove_added_batch = config.batch_size - even_batch
del alphas[-remove_added_batch:]
distance = 0
sentences = data["test_sentences"]
print("computing heatmaps and avg distance...")
from util.heatmap import avg_distance_and_heatmaps
avg_distance_and_heatmaps(
alphas, sentences,
CONST.SENTENCE_PATH + "/hursh_multitask")
def main(CONST, data):
config = get_config("CNN")
# if not CONST.TRAIN:
# config.num_steps = 1
config.vocab_size = len(data["embeddings"])
config.max_max_epoch = CONST.MAX_EPOCH
#from util.evaluations import print_config
#print_config(config)
tf.reset_default_graph()
# start graph and session
# config=tf.ConfigProto(log_device_placement=True) pass to tf.Session
# to see which devices all operations can run on
with tf.Graph().as_default(), tf.Session(config=session_conf) as session:
with tf.variable_scope(
"model",
reuse=None,
initializer=tf.contrib.layers.xavier_initializer()):
training_model = CNNModel(is_training=True,
config=config) # model class
if CONST.TRAIN:
tf.initialize_all_variables().run()
session.run(training_model.embedding.assign(
data["embeddings"])) # train
# Reload save epoch training time
if CONST.RELOAD_TRAIN:
print("Reloading previous run...")
saver = tf.train.Saver()
saver.restore(sess=session,
save_path=CONST.CNN_MODEL_PATH + "cnn")
all_losses = []
train_writer = run_metadata = run_options = False
if config.get_summary:
# session = tf.InteractiveSession()
train_writer = tf.train.SummaryWriter(
CONST.SLOT1_MODEL_PATH + "attention_graph/" +
config.__class__.__name__, session.graph)
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
for i in range(config.max_max_epoch): # setup epoch
train_perplexity, loss, losses = run_epoch(
session,
training_model,
data["x_train"],
data["p_train"],
train_writer,
run_options,
run_metadata,
category=data["a_train"],
verbose=True)
all_losses = all_losses + [np.mean(losses)]
print("Epoch: %d Avg. Cost: %.6f" %
(i + 1, np.mean(all_losses)))
from util.evaluations import print_config
print_config(config)
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
# plt.plot([np.mean(all_losses[i-50:i]) for i in range(len(all_losses))])
figure_name = CONST.OUT_DIR + "loss/" + "losses_slot3" + config.__class__.__name__ + ".png"
x = [i for i in range(len(all_losses))]
plt.plot(np.array(x), np.array(all_losses))
plt.savefig(figure_name)
save_pickle(CONST.DATA_DIR + config.__class__.__name__ + "_slot3",
all_losses)
print("saved slot3 losses.png and losses data", figure_name)
print("loss: ", figure_name)
print(
"loss data: ", CONST.DATA_DIR + config.__class__.__name__ +
"_slot3" + ".pickle")
#saver = tf.train.Saver(tf.all_variables())
#path = saver.save(sess=session, save_path=CONST.CNN_MODEL_PATH + "cnn")
#print("model saved: " + path)
if config.get_summary:
train_writer.close()
# session.close() # doesn't seem to close under scope??
if not False:
with tf.variable_scope(
"model", reuse=True): # reuse scope to evaluate model :-)
# Initialize Model Graph
validation_model = training_model
predictions = []
correct_predictions = []
# Get Predictions
for step, (x, y, l, e, a) in enumerate(
semeval_itterator(data["x_test"],
data["p_test"],
validation_model.batch_size,
validation_model.num_steps,
category=data["a_train"],
shuffle_examples=False)):
pred, scores = session.run(
[
validation_model.predictions,
validation_model.scores
], {
validation_model.input_data: x,
validation_model.targets: y
})
predictions = predictions + pred.tolist()
correct_predictions = correct_predictions + scores.tolist()
even_batch = len(data["x_test"]) % config.batch_size
remove_added_batch = config.batch_size - even_batch
predictions = correct_predictions
del predictions[-remove_added_batch:]
predictions = np.asarray(predictions)
from util.evaluations import print_config
from util.evaluations import evaluate_multiclass
y = [np.asarray(e) for e in data["p_test"]]
save_pickle(
CONST.DATA_DIR + config.__class__.__name__ +
"slot3_predictions", {
"predictions": predictions,
"y": y
})
evaluate_multiclass(predictions, y, True)
print("predictions saved")