def main(args):
exp_path = args["exp_path"]
isTraining = args["isTraining"]
model_name = args["model_name"]
if isTraining:
data_path = "data/Train"
rel_path = "../Corpus/TDT2/Train/QDRelevanceTDT2_forHMMOutSideTrain"
else:
data_path = "data/Test"
rel_path = "../Corpus/TDT2/AssessmentTrainSet/AssessmentTrainSet.txt"
# Read data
with open(data_path + "/qry_IDs.pkl", "rb") as f:
qry_IDs = pickle.load(f)
with open(data_path + "/doc_IDs.pkl", "rb") as f:
doc_IDs = pickle.load(f)
qry_tf = np.load(data_path + "/x_qry_tf_mdl.npy")
doc = np.load(data_path + "/doc_mdl.npy")
#mean = np.load(exp_path + "/mean.npy")
#stdv = np.load(exp_path + "/stdv.npy")
#valid_idx = np.nonzero(stdv)
# Load model
model = load_model(exp_path + "/" + model_name)
# Evaluation
evaluate_model = EvaluateModel(rel_path, isTraining)
with tf.device('/device:GPU:0'):
# Train
#qry_tf[:, valid_idx] = (qry_tf[:, valid_idx] - mean[valid_idx]) / stdv[valid_idx]
rel_mdl = model.predict(qry_tf)
#rel_mdl = rel_mdl * stdv[-1] + mean[-1]
qry_docs_ranking = cosineFast(rel_mdl, qry_IDs, doc, doc_IDs)
mAP = evaluate_model.mAP(qry_docs_ranking)
print(mAP)
for eqe_list in EQE2:
query_model, query_model_dict = eqe_list
qry_mdl, qry_IDs = query_model
# pseudo relevance feedback (swm)
for step in range(2):
# kl divergence
query_result = np.dot(qry_mdl, np.log(doc_mdl.T))
result = np.argsort(-query_result, axis=1)
query_docs_point_dict = defaultdict(dict)
for q_idx in xrange(len(qry_IDs)):
docs_ranking = []
for doc_idx in result[q_idx]:
docs_ranking.append(doc_IDs[doc_idx])
query_docs_point_dict[qry_IDs[q_idx]] = docs_ranking
mAP = evaluate_model.mAP(query_docs_point_dict)
mAP_list.append(mAP)
print "mAP"
print mAP
if step < 1:
# save one-shot result
Pickle.dump(query_model_dict, open("model/query_model.pkl", "wb"),
True)
Pickle.dump(query_docs_point_dict,
open("model/query_docs_point_dict.pkl", "wb"), True)
# load one-shot result
query_docs_point_fb = Pickle.load(
open("model/query_docs_point_dict.pkl", "rb"))
query_model_fb = Pickle.load(open("model/query_model.pkl", "rb"))
# swm