def run_mlm():
global LAMBDA, b, MU, WEIGHTS, INCREMENT
timer = Timer('main')
timer.start()
try:
for j in range(0, int(1 / INCREMENT)):
k = 1
for i in range(0, int(1 / INCREMENT)):
run(k=k, b=b, lam=LAMBDA, mu=MU, weights=WEIGHTS)
baseline = eval.full_eval(QRELS_FILE, RANKING_FILE)
mlm = eval.full_eval(QRELS_FILE, MLM_FILE)
with open('data/results.txt', 'a') as rfile:
rfile.write('SCR P@10 (M)AP (M)RR using parameters '
'lambda={0:05.3f}, weights={1:05.3f}\n'.format(
LAMBDA, WEIGHTS))
rfile.write(
'BLN {0:05.3f} {1:05.3f} {2:05.3f}\n'.format(
baseline['p10'], baseline['ap'], baseline['rr']))
rfile.write(
'MLM {0:05.3f} {1:05.3f} {2:05.3f}\n'.format(
mlm['p10'], mlm['ap'], mlm['rr']))
LAMBDA += INCREMENT
for w in WEIGHTS:
w += INCREMENT
except Exception as e:
dump_args([LAMBDA, b])
print('There was an error \n {}'.format(e))
timer.stop()
print(timer.total_running_time_long)
for i in range(int(1/increment)):
SIM["similarity"]["default"]["k1"] = 1
for j in range(int(1/increment)):
loop_vars = get_vars(SIM)
print('Running loop no. {}...'.format(loop_no))
file = 'data/bmo/results_b_{0:04.2f}_k_{1:04.2f}_.txt'.format(loop_vars[1], loop_vars[0])
if exists(file):
print('Results for b={0:04.2f}, k={1:04.2f} already exist'.format(loop_vars[1], loop_vars[0]))
else:
run_query(file, es)
SIM["similarity"]["default"]["k1"] += increment
change_model(SIM, es)
loop_no += 1
SIM["similarity"]["default"]["b"] += increment
change_model(SIM, es)
for i in os.listdir('data/BMO/'):
with open('data/bmo/evaluations.txt', 'a') as f:
cur_file = 'data/bmo/{}'.format(i)
print('Evaluating {}...'.format(cur_file))
if cur_file.startswith('data/bmo/results'):
feval = eval.full_eval(QRELS_FILE, cur_file)
fvars = i.split('_')
f.write('\tP@10 (M)AP (M)RR using parameters b={0:04.2f}, k={1:04.2f}\n'.format(float(fvars[2]),
float(fvars[4])))
f.write('\t{0:05.3f} {1:05.3f} {2:05.3f}\n'.format(feval['p10'], feval['ap'], feval['rr']))
else:
print('Invalid file, skipping...')
change_model(DEFAULT_SIM, es) # Resets the search model
def singleprocess_run():
run(k=k, b=b, lam=LAMBDA, mu=MU)
baseline = eval.full_eval(QRELS_FILE, RANKING_FILE)
mlm = eval.full_eval(QRELS_FILE, MLM_FILE)
bm = eval.full_eval(QRELS_FILE, BM_FILE)
return [baseline, mlm, bm]
def main():
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
if flags.data_source == 'video_sub': flags.dim_mcb = 4096
with tf.Session(config=gpu_config) as sess:
_inputs = {
'query':
tf.placeholder(dtype=tf.float32,
shape=[None, flags.dim_text],
name='query_input'),
'answer':
tf.placeholder(dtype=tf.float32,
shape=[None, 5, flags.dim_text],
name='answer_input'),
'story':
tf.placeholder(dtype=tf.float32,
shape=[None, flags.dim_mcb],
name='story_input'),
'cor_idx':
tf.placeholder(dtype=tf.int64, shape=[None], name='cor_idx_input'),
'rgb':
tf.placeholder(dtype=tf.float32, shape=[None, flags.dim_rgb]),
'sub':
tf.placeholder(dtype=tf.float32, shape=[None, flags.dim_sub])
}
if FLAGS.video_features == True:
_inputs.pop("story", None)
else:
_inputs.pop("rgb", None)
_inputs.pop("sub", None)
if FLAGS.video_features == True:
model = ltm_video_model(flags=flags, inputs=_inputs)
else:
model = ltm_model(flags=flags, inputs=_inputs)
model.build_model()
model_vars = tf.contrib.framework.get_model_variables()
tf.contrib.slim.model_analyzer.analyze_vars(model_vars,
print_info=True)
saved_op = {}
for var in tf.trainable_variables():
print(var.name)
saved_op[var.name] = var
train_queue = LTM_Queue(story_filename=STORY_FILE,
qa_filelist=QA_TRAIN,
capacity=30,
batch_size=32,
num_threads=20)
val_queue = LTM_Queue(story_filename=STORY_FILE,
qa_filelist=QA_VAL,
capacity=30,
batch_size=32,
num_threads=10)
train_queue.start_threads()
val_queue.start_threads()
merge_op, train_writer, val_writer = build_summary_ops(
model, flags, sess)
checkpoint_op, checkpoint_dir = build_checkpoint_ops(flags)
tf.global_variables_initializer().run()
best_accuracy = 0.0
for step in xrange(FLAGS.max_steps):
ts = time()
queue_inputs = train_queue.get_inputs()
feed = {}
for key, val in _inputs.iteritems():
try:
feed[_inputs[key]] = queue_inputs[key + '_rep']
except:
feed[_inputs[key]] = queue_inputs[key]
loss, _ = sess.run([model.loss, model.train_op], feed_dict=feed)
loss_mean = np.mean(loss)
print('| Step %07d | Loss %.3f | Time spent %.3f |' %
(step, loss_mean, time() - ts),
end='\r')
if step == 0: continue
if step % SUMMARY_STEP == 0:
summary = sess.run(merge_op, feed_dict=feed)
train_writer.add_summary(summary, step)
if step % VAL_STEP == 0:
queue_inputs = val_queue.get_inputs()
feed = {}
for key, val in _inputs.iteritems():
try:
feed[_inputs[key]] = queue_inputs[key + '_rep']
except:
feed[_inputs[key]] = queue_inputs[key]
loss = sess.run(model.loss, feed_dict=feed)
loss_mean = np.mean(loss)
summary = sess.run(merge_op, feed_dict=feed)
val_writer.add_summary(summary, step)
if step % FULL_EVAL_STEP == 0 and step >= 900:
num_val_examples = {
'VideoSubInception': 886,
'VideoSubResnet': 886,
'video_sub': 886,
'sub': 1958,
"sub_part": 1958,
'dvs': 282,
'dvs_pe': 282,
'script': 976,
'plot': 1958,
'sub_fasttext': 1958
}
accuracy = full_eval(step, _inputs, model, sess, STORY_FILE,
QA_VAL,
num_val_examples[FLAGS.data_source])
best_accuracy = max([best_accuracy, accuracy])
if accuracy >= FLAGS.save_threshold:
print("Saving checkpoint| Step {} | Accuracy : {} ".format(
step, accuracy))
save_checkpoint(sess, checkpoint_dir, checkpoint_op, step)
print("Best Accuracy : {}".format(best_accuracy))