def try_list(datapath, modelargs, outfile="ensemble_list.txt"):
"""
Score a list of models, to find the best of a fixed number of combinations for ensembling.
Expects `modelargs` to be a list of model descriptions, see predict().
"""
df = pd.read_csv(datapath)
df = preproc.preproc(df, lower=True)
lines = []
already_tried = load_tried_set(outfile)
with open(datapath) as f:
gt = f.readlines()[1:]
while len(modelargs) > 0:
lines.clear()
ensemble = modelargs.pop()
if len(set(ensemble)) != len(ensemble):
logger.warning("Truncating ensemble: {}".format(str(ensemble)))
ensemble = list(set(ensemble))
results = predict(list(ensemble), list(df['text']), None)
lines = []
for i, result in enumerate(results):
line = "{},{},{},{}".format(df['text'][i], df['sex'][i],
df['age'][i], results[i])
lines.append(line)
en_score = score(gt, lines)
print("{}: {}".format(en_score, str(ensemble)))
with open(outfile, "a") as f:
f.write("{}: {}\n".format(en_score, str(ensemble)))
def main(datapath: ("The CSV file to use as input", "option", "d",
str) = "valid5000.csv",
permute: ("Permute the order of words in sentences randomly", "flag",
"p") = False,
reverse: ("Reverse the order of words in sentences", "flag",
"r") = False,
search:
("Do an exhaustive search to find the best ensemble of SEARCH models",
"option", "s", int) = 0,
models: ("Path to a file with a list of models", "option", "m") = ""):
global prediction_cache_file
global prediction_cache
prediction_cache_file = datapath + ".prediction_cache.pickle"
if os.path.exists(prediction_cache_file):
with open(prediction_cache_file, "rb") as f:
prediction_cache = pickle.load(f)
logger.warning("Loaded prediction cache from {}".format(
prediction_cache_file))
if search > 0:
try_permutations(datapath, num_ensembles=search, outfile="tmp.txt")
sys.exit()
elif models != "":
already_tried = load_tried_set("tmp.txt")
already_tried = truncate_ensembles(already_tried)
to_try = load_tried_set(models)
to_try = truncate_ensembles(to_try)
ensembles = to_try - already_tried
logger.info("List of {} ensembles to try: {}".format(
len(ensembles), ensembles))
try_list(datapath, ensembles, outfile="tmp.txt")
sys.exit()
ensemble = build_ensemble()
df = pd.read_csv(datapath)
df_nonums = pd.read_csv(datapath)
df = preproc.preproc(df, lower=True) #, spelling=True)
df_nonums = preproc.preproc(df_nonums, lower=True, nonumbers=True)
if permute:
results = permute_predict(ensemble, df)
elif reverse:
results = reverse_predict(ensemble, df)
else:
results = predict(ensemble, list(df['text']), list(df_nonums['text']))
lines = []
for i, result in enumerate(results):
line = "{},{},{},{}".format(df['text'][i], df['sex'][i], df['age'][i],
results[i])
lines.append(line)
write_solution(lines)
with open(datapath) as f:
gt = f.readlines()
print("Score: {}".format(score(gt[1:], lines)))
def fit_score(word, pos, batch_pmass):
"""
Convenience function wrapper for the score function found
in main.py. It also happens to hide the ugliness of passing
a ton of arguments to it.
"""
return score(pos, word,
english_pmass,
batch_pmass,
wts["english_freq"],
wts[ "batch_freq" ],
wts["starts_word" ],
wts["repetitions" ],
wts["avoid_vowels"],
wts["consecutives"] )
def main(modelpath, datapath):
import main
logger.info("Loading model")
model = Model.load_model(modelpath)
df = pd.read_csv(datapath)
logger.info("Predicting")
ret, _ = model.predict_df(df)
import pdb
pdb.set_trace()
solution = []
for i in tqdm.tqdm(range(len(ret))):
solution.append(ret[i] + 1)
with open(datapath) as f:
gt = f.readlines()
print("Score: {}".format(main.score(gt[1:], solution)))
def batch_fit():
X = []
Y = []
prediction_data = []
prediction_labels = []
X, Y, prediction_data, prediction_labels = get_Data()
print('Train data length ={0} , Test data length ={1}'.format(
len(X), len(prediction_data)))
#print(X)
X = np.array(X) / 255.0
Y = np.array(Y)
prediction_data = np.array(prediction_data) / 255.0
prediction_labels = np.array(prediction_labels)
N, D = X.shape
#print ('Dshape {0}'.format(D))
M = 100
K = 7
# randomly initialize weights
W1 = np.random.randn(D, M) / np.sqrt(D + M)
b1 = np.zeros(M)
W2 = np.random.randn(M, K) / np.sqrt(M + K)
b2 = np.zeros(K)
learning_rate = 5 * 10e-7
costs = []
best_validation_error = 1
batch_sz = 500
n_batches = int(N / batch_sz)
#Use of Momentum
mu = 0.9
dW2 = 0
db2 = 0
dW1 = 0
db1 = 0
reg = 0.01
#rms propagation
cache_W2 = 1
cache_b2 = 1
cache_W1 = 1
cache_b1 = 1
decay_rate = 0.999
eps = 1e-10
#print ('{0} {1} {2}'.format(N,n_batches,D))
for m in range(1000):
tmpX, tmpY = shuffle(X, Y)
for j in range(n_batches):
x = tmpX[j * batch_sz:(j * batch_sz + batch_sz), :]
y = tmpY[j * batch_sz:(j * batch_sz + batch_sz)]
n = len(x) # Getting new length
T = np.zeros((n, K))
for i in range(n):
T[i, y[i]] = 1
output, hidden = forward(x, W1, b1, W2, b2)
gW2 = derivative_w2(hidden, T, output) + reg * W2
gb1 = derivative_b1(T, output, W2, hidden) + reg * b1
gW1 = derivative_w1(x, W2, hidden, T, output) + reg * W1
gb2 = derivative_b2(T, output) + reg * b2
dW1 = dW1 * mu + learning_rate * gW1
dW2 = dW2 * mu + learning_rate * gW2
db1 = db1 * mu + learning_rate * gb1
db2 = db2 * mu + learning_rate * gb2
#cache calculation
cache_W2 = decay_rate * cache_W2 + (1 - decay_rate) * gW2 * gW2
cache_b2 = decay_rate * cache_b2 + (1 - decay_rate) * gb2 * gb2
cache_W1 = decay_rate * cache_W1 + (1 - decay_rate) * gW1 * gW1
cache_b1 = decay_rate * cache_b1 + (1 - decay_rate) * gb1 * gb1
W2 += (dW2 / (np.sqrt(cache_W2) + eps))
b1 += (db1 / (np.sqrt(cache_b1) + eps))
W1 += (dW1 / (np.sqrt(cache_W1) + eps))
b2 += (db2 / (np.sqrt(cache_b2) + eps))
#print('i value ={0}'.format(i))
if j % (n_batches / 2) == 0:
c = cost(T, output)
P = np.argmax(output, axis=1)
r = score(y, P)
e = error_rate(y, P)
if e < best_validation_error:
best_validation_error = e
print('Cost = {0} , Training Score = {1} ,Error ={2}'.format(
c, r, e))
costs.append(c)
print("best_validation_error:", best_validation_error)
plt.plot(costs)
plt.show()
# Test npar_train
output, hidden = forward(prediction_data, W1, b1, W2, b2)
P = np.argmax(output, axis=1)
print('Test Score {0}'.format(score(prediction_labels, P))) # Final output
parser = argparse.ArgumentParser()
parser.add_argument('file', type=str)
args = parser.parse_args()
n_books, n_libs, n_days, scores_of_books, libs = load_libraries(args.file)
scores_of_books_dict = dict(enumerate(scores_of_books))
# print(n_books)
# print(n_libs)
# print(n_days)
# print(book_scores)
# print(libs[:10])
# lib = libs[0]
# n = lib[0]
# signup_days = lib[1]
# bs_per_day = lib[2]
# bs = lib[3]
# score = score_library(bs, signup_days, bs_per_day, [], scores_of_books, n_days)
# print(score)
library_signup_times = [lib[1] for lib in libs]
library_ship_capacities = [lib[2] for lib in libs]
solution = order_libraries(libs, scores_of_books_dict, n_days)
print(solution)
print(scores_of_books)
scores_of_books = np.asarray(scores_of_books, dtype=np.uint)
s = score(solution, n_days, scores_of_books,
library_signup_times,
library_ship_capacities)
print(s)
def hello():
return score()
with open("responses.json") as file:
data = json.load(file)
d1 = dict(list(data.items())[:len(data) // 2])
d2 = dict(list(data.items())[len(data) // 2:])
propscrDict = {}
const1 = 0
const2 = 0
while len(d1) != 0 and len(d2) != 0:
scoreDict = {}
for men in d2:
scoreDict[men] = []
for women in d1:
value = score(d2[men], d1[women])
if value > 0:
scoreDict[men].append((women, value))
scoreDict[men].sort(key=lambda x: x[1], reverse=True)
for proposer in scoreDict:
if len(scoreDict[proposer]) == 0:
continue
if scoreDict[proposer][0][0] not in propscrDict.keys():
propscrDict[scoreDict[proposer][0][0]] = (
proposer, scoreDict[proposer][0][1])
else:
if scoreDict[proposer][0][1] > propscrDict[scoreDict[proposer][0]
[0]][1]:
propscrDict[scoreDict[proposer][0][0]] = (
def try_permutations(datapath, num_ensembles=8, outfile="tmp.txt"):
"""
Brute force random walk over the list of available checkpoints. Tries combinations at random
to find the highest scoring ensemble. Never returns.
"""
DIRS = [
"cp/bert/*/pytorch_model.bin",
"cp/roberta/*/pytorch_model.bin",
"cp/xlnet/*/pytorch_model.bin",
"cp/xlnet/09/checkpoint-21000",
"cp/xlnet/07/checkpoint-42500",
"cp/distilbert/*/pytorch_model.bin",
]
score_hash = {}
tried = load_tried_set(outfile)
for models in tried:
score_hash[tuple(set(models))] = 0.0
setlist = [set(x) for x in tried]
buf = []
full_list = []
df = pd.read_csv(datapath)
df = preproc.preproc(df, lower=True)
with open(datapath) as f:
gt = f.readlines()[1:]
gt = [int(x.split(',')[-1][:-1]) for x in gt]
for d in DIRS:
buf = glob.glob(d)
for i, tmp in enumerate(buf):
tmp = tmp.replace("/pytorch_model.bin", "")
full_list.append(tmp)
logger.warning("Models to permute: {}".format(str(full_list)))
combo = itertools.combinations(full_list, num_ensembles)
combo_list = list(combo)
random.shuffle(combo_list)
df_text_list = list(df['text'])
progbar = tqdm.tqdm(
combo_list,
total=len(combo_list) -
len(list(filter(lambda x: len(x) == num_ensembles, tried))),
desc="HiScore: 0.0")
hiscore = 0.0
for paths in progbar:
buf.clear()
for path in paths:
path = path.strip()
if 'roberta' in path:
buf.append(("roberta", path, ""))
elif "distilbert" in path:
buf.append(("distilbert", path, ""))
elif 'bert' in path:
buf.append(("bert", path, ""))
elif 'xlnet' in path:
buf.append(('xlnet', path, ''))
else:
logger.error("Epic fail, programmer.")
sys.exit()
buf_set = set(buf)
buf_tuple = tuple(buf_set)
if score_hash.get(buf_tuple, None) is not None:
continue
# if buf_set in setlist:
# continue
# setlist.append(buf_set)
# if buf_tuple in tried:
# continue
# tried.add(buf_tuple)
results = predict(buf_tuple, df_text_list, None)
# lines = []
# for i, result in enumerate(results):
# line = ("%s,%d,%d,%d" % (df['text'][i],
# df['sex'][i],
# df['age'][i],
# results[i]))
# lines.append(line)
en_score = score(gt, results)
score_hash[buf_tuple] = en_score
if en_score > hiscore:
hiscore = en_score
progbar.set_description("HiScore: {}".format(hiscore))
#print("{}: {}".format(en_score, str(buf)))
# if en_score > 0.83:
with open(outfile, "a") as f:
f.write("{}: {}\n".format(en_score, str(buf)))
def test2(self):
self.assertEqual(main.score([4, 4, 4, 3, 3]), 400)
def test1(self):
self.assertEqual(main.score([2, 3, 4, 6, 2]), 0)
def test3(self):
self.assertEqual(main.score([2, 4, 4, 5, 4]), 450)
nb_train = 2000000
#%%
x0 = data_array[:nb_train, 0:60] # - data_array[:nb_train, 30:60]
train_y = data_array[:nb_train, 60]
x1 = data_array[nb_train:, 0:60] # - data_array[nb_train:, 30:60]
test_y = data_array[nb_train:, 60]
reg = LinearRegression(normalize=True).fit(x0, train_y)
#%%
train_pred = reg.predict(x0)
test_pred = reg.predict(x1)
print('train score: {:.5f}'.format(score(train_y, train_pred)))
print('test score: {:.5f}'.format(score(test_y, test_pred)))
#%% [markdown]
# ### linear model with normalized data
#%%
with open('normed_data.pickle', 'rb') as f:
normed_data = pickle.load(f)
#%%
nb_train = 2000000
# train data
train_x = normed_data[:nb_train, :60]
train_y = normed_data[:nb_train, 60]
# test data
