def main():
# Get label encoder
lb = LabelBinarizer()
lbenc = lb.fit(utils.get_classes())
# Get train data
X_train, y_train, train_filenames = utils.get_train(
'../input/train', list(lbenc.classes_), img_width, img_height)
# Create and train model
model = train(X_train, y_train, epochs=100, batch_size=32)
print("+++++++++++++++++++++++++++++++++++++++++++")
# Load model ...
#model = load_model('../models/'+ 'model2_f0.86/'+ 'model2-64-0.341.h5')
# Get test data
X_test, X_test_id = utils.get_test('../input/test', img_width, img_height)
# Predict on test data
preds = model.predict(X_test, verbose=1)
# Create submission
utils.create_submission(lbenc.inverse_transform(preds),
X_test_id,
output_path="../submissions/",
filename=modelname,
isSubmission=True)
utils.to_csv_ens(lbenc.inverse_transform(preds),
preds,
X_test_id,
utils.get_classes(),
output_path="../submissions/",
filename=modelname)
print('Finished.')
def eval():
# Get classes
lb = LabelBinarizer()
lbenc = lb.fit(utils.get_classes())
# Load model
model = load_model('../models/' + 'model.h5')
# Get test data
X_test, X_test_id = utils.get_test('../input/test', img_width, img_height)
# Predict on test data
preds = model.predict(X_test, verbose=1)
# Create ensembling file
df_csv = utils.to_csv_ens(lbenc.inverse_transform(preds),
preds,
X_test_id,
utils.get_classes(),
output_path="../submissions/",
filename=modelname)
# Create submission file
subm = utils.create_submission(lbenc.inverse_transform(preds),
X_test_id,
output_path="../submissions/",
filename=modelname,
isSubmission=True)
def read_vals(objective, graph_files):
tool_files = ["/".join(path) for path in graph_files if utils.get_tool(utils.get_fn(path)) == tool]
# Extract times
name_to_n = utils.key_functions[objective]
time_pairs = [(name_to_n(utils.get_test(utils.get_fn(path.split("/")))),
utils.read_times(in_dir + "/" + path))
for path in tool_files]
# Sort values
times_sorted = sorted(time_pairs, key=lambda pair: pair[0])
n_vals = list(map(lambda pair: pair[0], times_sorted))
t_objective_vals = list(map(lambda pair: pair[1][0], times_sorted))
t_jacobian_vals = list(map(lambda pair: pair[1][1], times_sorted))
return (n_vals, t_objective_vals, t_jacobian_vals)
def get_test_instance(self):
if os.path.exists('../DataSets/mydata/test_txtid.pkl') and os.path.exists('../DataSets/mydata/test_instances.pkl'):
test_txtid_pkl = open('../DataSets/mydata/test_txtid.pkl', 'rb')
test_txtid = pickle.load(test_txtid_pkl)
test_instances_pkl = open('../DataSets/mydata/test_instances.pkl', 'rb')
test_instances = pickle.load(test_instances_pkl)
else:
test_txtid, test_instances = get_test()
test_paper_list = self.load_data(test_instances)
test_tag_list, test_id_list, test_p1_list, test_p2_list, test_y_list = self.test_padding(test_paper_list,
self.word2id_dic,
self.max_sequence_length,
self.position_max)
self.testb_dataset = (test_txtid, test_tag_list, test_id_list, test_p1_list, test_p2_list, test_y_list)
return self.testb_dataset
def read_vals(objective, graph_files, tool, in_dir):
'''Extracts data for files of the specified tool.'''
def get_violations(file_name):
'''Extracts jacobian calculation correctness.'''
folder, fn = os.path.split(file_name)
correctness_file_name = os.path.join(
in_dir, folder, fn.replace(TIMES_SUBSTRING, CORRECTNESS_SUBSTRING))
if not os.path.isfile(correctness_file_name):
print(f"WARNING: correctness file {correctness_file_name} "
"doesn't exist\n")
return False
try:
with open(correctness_file_name, "r", encoding="ascii") as cf:
correctness_data = json.load(cf)
return correctness_data["ViolationsHappened"]
except Exception as e:
print(f"WARNING: correctness file {correctness_file_name} parsing "
f"failed.\nError message:{e.args}\n")
return False
tool_files = [
os.path.join(*path) for path in graph_files
if utils.get_tool_from_path(path) == tool
]
if has_manual(tool):
violation_info = [False for file in tool_files]
else:
violation_info = [get_violations(file) for file in tool_files]
# Extract times
name_to_n = utils.key_functions[objective]
info = [(name_to_n(utils.get_test(utils.get_fn(path.split("/")))),
utils.read_times(os.path.join(in_dir, path)), violation)
for (path, violation) in zip(tool_files, violation_info)]
# Sort values
info_sorted = sorted(info, key=lambda t: t[0])
n_vals = list(map(lambda t: t[0], info_sorted))
t_objective_vals = list(map(lambda t: t[1][0], info_sorted))
t_jacobian_vals = list(map(lambda t: t[1][1], info_sorted))
violation_vals = list(map(lambda t: t[2], info_sorted))
return (n_vals, t_objective_vals, t_jacobian_vals, violation_vals)
def get_vocab(dset):
return set(w for sent in dset for w in sent['ws'])
def get_contexts(sent, c):
ws = (['<s>']*c) + sent['ws'] + (['</s>']*c)
contexts = []
for i, w in enumerate(sent['ws']):
wi = i + c
if sent['ii'][i]:
contexts.append(' '.join([w for w in ws[wi-c:wi] + ['___'] + ws[wi+1:wi+c+1]]))
return contexts
if __name__ == '__main__':
trn = get_dset()
tst = get_test()
print map(len, map(get_tagged_vocab, [trn,tst]))
print 'tagged vocab size trn {} tst {}'.format(*map(len, map(get_tagged_vocab, [trn,tst])))
print 'all vocab size trn {} tst {}'.format(*map(len, map(get_vocab, [trn,tst])))
vtrn, vtst = map(get_tagged_vocab, [trn,tst])
print 'tagged vtst diff: {:.2f}'.format( len(vtst.difference(vtrn)) / len(vtst) )
vtrn, vtst = map(get_vocab, [trn,tst])
print 'all vtst diff: {:.2f}'.format( len(vtst.difference(vtrn)) / len(vtst) )
precnt = Counter(w[:j] for sent in trn for w, lbl in zip(sent['ws'],sent['ls']) for j in range(3,5) if lbl==1 and len(w)>j)
sufcnt = Counter(w[-j:] for sent in trn for w, lbl in zip(sent['ws'],sent['ls']) for j in range(3,5) if lbl==1 and len(w)>j)
print 'most common prefixes:', precnt.most_common(100)
print 'most common suffixes:', sufcnt.most_common(100)
def submission(model,
sampling_method,
data_dir,
results_dir,
device='cpu',
verbose=True):
if verbose:
print("Using device: {}".format(device))
print("Reading train data in...")
if model == 'lgbm':
X_train, Y_train, feature_labels = get_train(data_dir, one_hot=False)
else:
X_train, Y_train, feature_labels = get_train(data_dir)
X_test = get_test(data_dir)
train_ids, test_ids = get_ids(data_dir)
country_names = get_country_names(data_dir)
if verbose:
print("Successfully loaded data")
lgbm_params = {
'task': 'train',
'objective': 'multiclass',
'num_class': 12,
'num_leaves': 31,
'learning_rate': 0.3,
'lambda_l2': 1.0,
'feature_fraction': 0.9,
'min_child_weight': 1.0,
'device': device,
'gpu_device_id': 0,
'gpu_platform_id': 0,
'max_bin': 63,
'verbose': 0
}
if device == 'cpu':
xgb_params = {
"objective": "multi:softprob",
"num_class": 12,
"tree_method": "hist",
"colsample_bytree": 0.9,
"n_jobs": 2,
"silent": 1
}
else:
xgb_params = {
"objective": "multi:softprob",
"num_class": 12,
"tree_method": "gpu_hist",
"colsample_bytree": 0.9,
"gpu_id": 0,
"max_bin": 16,
"silent": 1
}
if verbose:
print("{} sampling process started...".format(sampling_method))
curr_time = time.time()
if sampling_method == "adasyn":
X_train_resampled, Y_train_resampled = ADASYN().fit_sample(
X_train, Y_train)
elif sampling_method == "smote":
X_train_resampled, Y_train_resampled = SMOTE().fit_sample(
X_train, Y_train)
elif sampling_method == "random":
X_train_resampled, Y_train_resampled = RandomOverSampler().fit_sample(
X_train, Y_train)
elif sampling_method == "smoteenn":
X_train_resampled, Y_train_resampled = SMOTEENN().fit_sample(
X_train, Y_train)
else:
X_train_resampled, Y_train_resampled = X_train, Y_train
if verbose:
print("Oversampling completed")
print("Time Taken: {:.2f}".format(time.time() - curr_time))
print("Size of Oversampled data: {}".format(X_train_resampled.shape))
print("{} selected for classification".format(model))
curr_time = time.time()
if model == 'lgbm':
categorical_feature = [
'age_bucket', 'gender', 'signup_method', 'signup_flow', 'language',
'affiliate_channel', 'affiliate_provider',
'first_affiliate_tracked', 'signup_app', 'first_device_type',
'first_browser'
]
lgb_train = lgb.Dataset(data=X_train_resampled,
label=Y_train_resampled,
feature_name=feature_labels,
categorical_feature=categorical_feature)
clf = lgb.train(lgbm_params, lgb_train, num_boost_round=30)
print("Time taken: {:.2f}".format(time.time() - curr_time))
Y_probs = clf.predict(X_test)
order = np.argsort(-Y_probs[:, :5], axis=1)
else:
X_train_xgb = xgb.DMatrix(X_train_resampled,
Y_train_resampled,
feature_names=feature_labels)
X_test_xgb = xgb.DMatrix(X_test, feature_names=feature_labels)
clf = xgb.train(xgb_params, X_train_xgb, 30)
print("Time taken: {:.2f}".format(time.time() - curr_time))
Y_probs = clf.predict(X_test_xgb)
order = np.argsort(-Y_probs[:, :5], axis=1)
print("Generating submission csv...")
with open(os.path.join(results_dir, 'submission_{}.csv'.format(model)),
'w') as f:
writer = csv.writer(f, delimiter=',', quoting=csv.QUOTE_MINIMAL)
writer.writerow(['id', 'country'])
for i in range(len(test_ids)):
for k in range(5):
writer.writerow([test_ids[i], country_names[order[i, k]]])
print("Finished.")
pred_labels = []
for sent, pred in zip(dset,preds):
gold_labels.extend([sent['ls'][ii] for ii, interested in enumerate(sent['ii']) if interested])
# pred_labels.extend([pred[ii] for ii, interested in enumerate(sent['ii']) if interested])
pred_labels.extend(pred)
logging.debug(tabulate(confusion_matrix(np.array(gold_labels), np.array(pred_labels)), headers=[0,1]))
p, r, f = evaluate_system.evaluateIdentifier(gold_labels, pred_labels)
return p,r,f
if __name__ == '__main__':
parser = get_arg_parser()
args = vars(parser.parse_args())
setup_logger(args)
logging.debug(tabulate([OrderedDict((k,v) for k,v in sorted(args.iteritems()))], headers='keys'))
if args['testf']:
trn = utils.get_dset(args['data'])
tst = utils.get_test()
ytrn, ytst = fit_predict(trn, tst, args, Emb(trn+tst))
with open(args['testf'], 'w') as out:
out.write('\n'.join([str(y) for y in ytst]))
else:
dset = utils.get_dset(args['data'])
if args['sample'] > 0:
random.seed(0)
dset = random.sample(dset, args['sample'])
xvalidate(dset, args, Emb(dset))
# create initializations for the three different test problems
# for every problem: create 100*5 random X values, distributed on the given slices
# for every problem: create corresponding responses
# save the data to be used in the torch and GPy implementation
from utils import get_test
import numpy as np
import torch
import pandas as pd
torch.manual_seed(42)
np.random.seed(42)
problems = ['branin', 'eggholder', 'camel']
for problem in problems:
testfunction, slices, scalers, hyperparameters = get_test(problem)
# create 500 points on [0, 1]
X1 = np.random.rand(500, 1)
X2 = np.random.choice(slices.numpy(),
(500, 1)) # slices are already scaled
X = scalers[0].inverse_transform(
torch.tensor(np.concatenate((X1, X2),
axis=1))) # scale back for evaluation
y = testfunction(torch.tensor(X))
# write to csv
df = pd.DataFrame(np.concatenate((X, y.reshape(-1, 1)), axis=1),
columns=['X1', 'X2', 'y'])
df.to_csv('initialization_' + problem + '.csv', index=False)
def min_max_optimization():
np.random.seed(42)
torch.manual_seed(32)
iterations_list = [20, 20, 100]
for problem_idx, problem in enumerate(['branin', 'camel', 'eggholder']):
testfunction, slices, scalers, hyperparameters = get_test(problem)
n_init = 5
jj = 0
# load data
df_read = pd.read_csv('initialization_' + problem + '.csv')
print("read data")
for initialization in range(int(df_read.shape[0] / n_init)):
X = np.array(df_read.iloc[(initialization *
n_init):(initialization * n_init +
n_init), 0:2])
z = np.array(df_read.iloc[(initialization *
n_init):(initialization * n_init +
n_init), 2]).reshape(-1, 1)
# scale
X_scaled = torch.tensor(scalers[0].transform(X),
dtype=torch.float32)
z_scaled = torch.tensor(scalers[1].transform(z),
dtype=torch.float32)
# run the optimization
iterations = iterations_list[problem_idx]
model = build_models(X_scaled, z_scaled, hyperparameters)
model = model.eval()
results = torch.zeros((1, 7))
print("started optimization")
for i in range(iterations):
new_candidate, min_max_location, current_min_max = thompson_sampling_acquisition(
model, slices)
current_min_max_unscaled = torch.tensor(
scalers[1].inverse_transform(
current_min_max.detach().numpy().reshape(1, 1)))
min_max_location_unscaled = torch.tensor(
scalers[0].inverse_transform(
min_max_location.detach().numpy()))
new_candidate_unscaled = torch.tensor(
scalers[0].inverse_transform(
new_candidate.detach().numpy()))
new_function_value = testfunction(
new_candidate_unscaled.reshape(1, -1))
# update the model
model = model.condition_on_observations(
new_candidate,
torch.tensor(scalers[1].transform(
new_function_value.numpy().reshape(-1, 1))))
print('new candidate:', new_candidate_unscaled)
print('min max location:', min_max_location_unscaled)
print('current min max:', current_min_max_unscaled)
print('iteration ', i)
results[0, 0] = i
results[0, 1:3] = new_candidate_unscaled
results[0, 3:5] = min_max_location_unscaled
results[0, 5] = current_min_max_unscaled
results[0, 6] = initialization
df = pd.DataFrame(results.detach().numpy(),
columns=[
'i', 'x_cand0', 'x_cand1', 'min_max0',
'min_max1', 'min_max_val', 'init'
])
df['problem'] = problem
if jj == 0:
df.to_csv(problem + '_results_thompson.csv', index=False)
else:
df.to_csv(problem + '_results_thompson.csv',
mode='a',
header=False,
index=False)
jj += 1
print('finished the optimization')
best_valid_mae = model.mae_valid
best_epoch_trial = model.early_stop_epoch
print('trial valid rmse of best epoch:', model.best_rmse)
print('trial valid r2 of best epoch:', model.r2_valid)
print('trial valid mae of best epoch:', model.mae_valid)
with open('params_dict.txt', 'w', encoding="utf8") as outfile:
json.dump(trials_dict, outfile)
return best_valid_rmse, best_valid_r_2, best_valid_mae, best_epoch_trial
if __name__ == '__main__':
train, validation = get_data(True, 1, 1)
test = get_test()
# SGD
best_valid_rmse_sgd, best_valid_r_2_sgd, best_valid_mae_sgd, best_epoch = \
hyper_param_tuning('SGD', SGD_HYPER_PARAMS)
print('best SGD model rmse:', best_valid_rmse_sgd)
print('best SGD model r2:', best_valid_r_2_sgd)
print('best SGD model mae:', best_valid_mae_sgd)
final_model = SGD()
final_model.fit_early_stop(train, validation, best_epoch)
test['pred'] = test.apply(lambda row:
final_model.predict(row[USER_COL],
row[ITEM_COL]), axis=1)
test[[USER_COL, ITEM_COL]] = test[[USER_COL, ITEM_COL]].apply(lambda col: col + 1)
print("Ensemble : ", mean_square_error(target_val, mean_pred_data))
"""
Create a file of the ensemble prediction
in the corresponding format for a submission on kaggle
"""
all_preds_han = []
for i in range(len(target_idx)):
idx = target_idx[i]
nb_model = 1
list_pred = []
# each model begin to predict on the documents test
for i in range(len(type_loss)):
name_doc = list_name_documents[i]
docs_val = get_test(name_doc=name_doc, idx_target=idx, config=config)
if type_loss[i] == "mse":
custom_loss = "mean_squared_error"
elif type_loss[i] == "higher":
custom_loss = mse_asymetric_higher
elif type_loss[i] == "lower":
custom_loss = mse_asymetric_lower
if type_model[i] == 1:
model = get_model_1(docs_train=docs_val,
config=config,
name_embeddings=list_embeddings[i],
custom_loss=custom_loss)
elif type_model[i] == 2:
model = get_model_2(docs_train=docs_val,
