def main():
input_lang, output_lang, pairs = prepare_data('ques',
'ans',
'../debug.json',
reverse=False)
encoder = EncoderRNN(input_lang.n_words, hidden_size).to(device)
attn_decoder = AttnDecoderRNN(hidden_size,
output_lang.n_words,
dropout_p=0.1,
max_length=1000).to(device)
rate = 0.9
pairs_train, pairs_test = pairs[0:int(len(pairs) *
rate)], pairs[int(len(pairs) *
rate):]
encoder.load_state_dict(torch.load('model/encoder-0.model'))
encoder.eval()
attn_decoder.load_state_dict(torch.load('model/decoder-0.model'))
attn_decoder.eval()
evaluate_all(encoder,
attn_decoder,
pairs_test,
max_length=1000,
input_lang=input_lang,
output_lang=output_lang,
n=len(pairs_test))
# show_plot(loss_history)
print('done test')
def main():
input_lang, output_lang, pairs = prepare_data('ques',
'ans',
'../test.json',
reverse=False)
model = Transformer(
src_vocab_size=input_lang.n_words,
src_max_len=MAX_LENGTH,
tgt_vocab_size=output_lang.n_words,
tgt_max_len=MAX_LENGTH,
).to(device)
rate = 0.9
pairs_train, pairs_test = pairs[0:int(len(pairs) *
rate)], pairs[int(len(pairs) *
rate):]
model.load_state_dict(torch.load('model/transformer-0.model'))
model.eval()
evaluate_all(model,
pairs_train,
max_length=100,
input_lang=input_lang,
output_lang=output_lang,
n=len(pairs_train))
# show_plot(loss_history)
print('done test')
def main():
print(HELLO_MSG)
args = parse_args()
print(args)
if args.action == "run":
launch_on_all_nodes()
elif args.action == "eval":
evaluate_all()
def human_performance(args, params):
x_gt, ground_truths, processor, dl = load.load_test(
params, fit_processor=True)
ground_truths, probs = human_gt_and_probs(
params, x_gt, ground_truths, processor, review_indiv=True)
evaluate.evaluate_all(
ground_truths, probs, processor.classes,
model_title='Human Performance Average', plot_flag=args.plot)
def train(model, data, words, params):
start_time = time.time()
counter = 0
try:
for eidx in xrange(params.epochs):
kf = utils.get_minibatches_idx(len(data), params.batchsize, shuffle=True)
uidx = 0
for _, train_index in kf:
uidx += 1
batch = [data[t] for t in train_index]
for i in batch:
i[0].populate_embeddings(words)
i[1].populate_embeddings(words)
(g1x, g1mask, g2x, g2mask, p1x, p1mask, p2x, p2mask) = getpairs(model, batch, params)
cost = model.train_function(g1x, g2x, p1x, p2x, g1mask, g2mask, p1mask, p2mask)
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected'
if (utils.checkIfQuarter(uidx, len(kf))):
if (params.save):
counter += 1
utils.saveParams(model, params.outfile + str(counter) + '.pickle')
if (params.evaluate):
evaluate_all(model, words)
sys.stdout.flush()
#undo batch to save RAM
for i in batch:
i[0].representation = None
i[1].representation = None
i[0].unpopulate_embeddings()
i[1].unpopulate_embeddings()
#print 'Epoch ', (eidx+1), 'Update ', (uidx+1), 'Cost ', cost
if (params.save):
counter += 1
utils.saveParams(model, params.outfile + str(counter) + '.pickle')
if (params.evaluate):
evaluate_all(model, words)
print 'Epoch ', (eidx + 1), 'Cost ', cost
except KeyboardInterrupt:
print "Training interupted"
end_time = time.time()
print "total time:", (end_time - start_time)
def train(model, data, words, params):
start_time = time.time()
counter = 0
try:
for eidx in xrange(params.epochs):
kf = utils.get_minibatches_idx(len(data), params.batchsize, shuffle=True)
uidx = 0
for _, train_index in kf:
uidx += 1
batch = [data[t] for t in train_index]
for i in batch:
i[0].populate_embeddings(words)
i[1].populate_embeddings(words)
(g1x, g1mask, g2x, g2mask, p1x, p1mask, p2x, p2mask) = getpairs(model, batch, params)
cost = model.train_function(g1x, g2x, p1x, p2x, g1mask, g2mask, p1mask, p2mask)
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected'
if (utils.checkIfQuarter(uidx, len(kf))):
if (params.save):
counter += 1
utils.saveParams(model, params.outfile + str(counter) + '.pickle')
if (params.evaluate):
evaluate_all(model, words)
sys.stdout.flush()
# undo batch to save RAM
for i in batch:
i[0].representation = None
i[1].representation = None
i[0].unpopulate_embeddings()
i[1].unpopulate_embeddings()
# print 'Epoch ', (eidx+1), 'Update ', (uidx+1), 'Cost ', cost
if (params.save):
counter += 1
utils.saveParams(model, params.outfile + str(counter) + '.pickle')
if (params.evaluate):
evaluate_all(model, words)
print 'Epoch ', (eidx + 1), 'Cost ', cost
except KeyboardInterrupt:
print "Training interupted"
end_time = time.time()
print "total time:", (end_time - start_time)
def agreement(args, params):
_, ground_truths, classes = load.load_test(params)
NUM_REPETITIONS = 10
gt_all = []
probs_all = []
for i in range(NUM_REPETITIONS):
gt, probs = get_ground_truths_and_human_probs(ground_truths,
params["num_reviewers"])
gt_all.append(gt)
probs_all.append(probs)
ground_truths = np.concatenate(tuple(gt_all), axis=1)
probs = np.concatenate(tuple(probs_all), axis=0)
evaluate.evaluate_all(ground_truths,
probs,
classes,
model_title='Human Agreement')
def run_and_evaluate(**suite_params):
suite = EntailmentSuite(**suite_params)
try:
type = eval(suite.cfgparser.get('DEFAULT', 'type'))
except:
logging.info(
"Warning: type of experiment not specified. Assuming cross-validation."
)
type = "cv"
if type == "heldout":
suite = EntailmentSuiteHeldOut(**suite_params)
elif type == "traintest":
suite = EntailmentSuiteTrainTest(**suite_params)
elif type == "heldoutstrict":
suite = EntailmentSuiteHeldOutStrict(**suite_params)
suite.start()
experiments = suite.cfgparser.sections()
for experiment in experiments:
logging.info("Running experiment: %s", experiment)
experiment_path = os.path.join(
eval(suite.cfgparser.get('DEFAULT', 'path')), experiment)
params = suite.get_params(experiment_path)
path = os.path.join(params['path'], params['name'])
rows = evaluate.evaluate_all(path)
evaluate.write_summary(rows, os.path.join(path, 'analysis.csv'))
def run_and_evaluate(**suite_params):
suite = EntailmentSuite(**suite_params)
try:
type = eval(suite.cfgparser.get('DEFAULT', 'type'))
except:
logging.info("Warning: type of experiment not specified. Assuming cross-validation.")
type="cv"
if type=="heldout":
suite = EntailmentSuiteHeldOut(**suite_params)
elif type=="traintest":
suite = EntailmentSuiteTrainTest(**suite_params)
elif type=="heldoutstrict":
suite = EntailmentSuiteHeldOutStrict(**suite_params)
suite.start()
experiments = suite.cfgparser.sections()
for experiment in experiments:
logging.info("Running experiment: %s", experiment)
experiment_path = os.path.join(eval(suite.cfgparser.get('DEFAULT', 'path')),
experiment)
params = suite.get_params(experiment_path)
path = os.path.join(params['path'],
params['name'])
rows = evaluate.evaluate_all(path)
evaluate.write_summary(rows, os.path.join(path, 'analysis.csv'))
def test(self, dataloader):
epoch = self.best_valid_epoch
if self.data_parallel:
model = nn.DataParallel(self.custom_net(
**self.custom_net_args)).to(self.device).eval()
else:
model = self.custom_net(**self.custom_net_args).to(
self.device).eval()
params_path = os.path.join(self.params_dir, self.descriptor)
print(
'For test set predictions, loading model params from params_path=',
params_path)
check_point = torch.load(params_path)
model.load_state_dict(check_point['params'])
with torch.no_grad():
epoch_loss, pred_epoch, gr_truth_epoch, volume_accs_epoch = self.iterate_through_batches(
model, dataloader, epoch, training=False)
self.eval_results_test = evaluate.evaluate_all(self.eval_results_test,
epoch,
self.label_meanings,
gr_truth_epoch,
pred_epoch)
self.plot_roc_and_pr_curves('test', epoch, pred_epoch, gr_truth_epoch)
self.save_all_pred_probs('test', epoch, pred_epoch, gr_truth_epoch,
volume_accs_epoch)
print("{:5s} {:<3d} {:11s} {:.3f}".format('Epoch', epoch, 'Test Loss',
epoch_loss))
def human_gt_and_probs(params, x_gt, ground_truths, processor, review_indiv=False):
gt_all = []
probs_all = []
for i in TEST_REVIEWS:
params["epi_ext"] = "_rev" + str(i) + ".episodes.json"
x_rev, probs, dl = load.load_x_y_with_processor(params, processor)
gt_i, rev_i = get_matching_indices(x_gt, x_rev)
gt = ground_truths[:, gt_i]
probs = probs[rev_i]
if review_indiv is True:
evaluate.evaluate_all(
gt, probs, processor.classes,
model_title='Human Performance with review ' + str(i))
gt_all.append(gt)
probs_all.append(probs)
ground_truths = np.concatenate(tuple(gt_all), axis=1)
probs = np.concatenate(tuple(probs_all), axis=0)
return ground_truths, probs
def main():
input_lang, output_lang, pairs = prepare_data('ques', 'ans', '../data.json',reverse=False)
encoder = Encoder(input_lang.n_words, MAX_LENGTH).to(device)
attn_decoder = AttnDecoderRNN(hidden_size, output_lang.n_words, dropout_p=0.1, max_length=MAX_LENGTH).to(device)
rate = 0.9
epoch = 10
pairs_train,pairs_test = pairs[0:int(len(pairs)*rate)], pairs[int(len(pairs)*rate):]
for i in range(epoch):
encoder.train()
attn_decoder.train()
train(encoder, attn_decoder, len(pairs_train), pairs=pairs_train, input_lang=input_lang,output_lang=output_lang, print_every=10)
encoder.eval()
attn_decoder.eval()
evaluate_all(encoder, attn_decoder, pairs_test, max_length=MAX_LENGTH, input_lang=input_lang, output_lang=output_lang,
n=len(pairs_test))
torch.save(encoder.state_dict(), 'model/encoder-' + str(i) + '.model')
torch.save(attn_decoder.state_dict(), 'model/decoder-' + str(i) + '.model')
#show_plot(loss_history)
print('done training')
def valid(self, dataloader, epoch):
model = self.model.eval()
with torch.no_grad():
epoch_loss, pred_epoch, gr_truth_epoch, volume_accs_epoch = self.iterate_through_batches(
model, dataloader, epoch, training=False)
self.valid_loss[epoch] = epoch_loss
self.eval_results_valid = evaluate.evaluate_all(
self.eval_results_valid, epoch, self.label_meanings,
gr_truth_epoch, pred_epoch)
self.early_stopping_check(epoch, pred_epoch, gr_truth_epoch,
volume_accs_epoch)
print("{:5s} {:<3d} {:11s} {:.3f}".format('Epoch', epoch, 'Valid Loss',
epoch_loss))
def run_and_evaluate(**suite_params):
suite = EntailmentSuite(**suite_params)
suite.start()
experiments = suite.cfgparser.sections()
for experiment in experiments:
logging.info("Running experiment: %s", experiment)
experiment_path = os.path.join(eval(suite.cfgparser.get('DEFAULT', 'path')),
experiment)
params = suite.get_params(experiment_path)
path = os.path.join(params['path'],
params['name'])
rows = evaluate.evaluate_all(path)
evaluate.write_summary(rows, os.path.join(path, 'analysis.csv'))
def predict():
"""
Uses the models to make the predictions and displays results on the page
"""
article = get_article(request.form['article_url'])
if article is None:
return "The article URL is invalid/not supported. " + \
"Try one of the following news sources: {}".format(TRUSTED_SOURCES)
predictions = evaluate_all(article.text, article.title, BASE_PATH, True)
sentiment = predictions['sentiment']
fake = predictions['fake']
category = predictions['category']
emotion = predictions['emotion']
sentiment = '{:.1f}% positive'.format(float(sentiment) * 100 / 4)
return render_template('index.html',
sentiment=sentiment,
fake=fake,
emotion=emotion,
category=category)
def plot_bleu(model_filter=None, token_filter=None, opt_filter=None):
# Calculate all scores first
bleu_scores = evaluate_all(model_filter, token_filter, opt_filter)
print('Finished calculating BLEU scores: %s' %
"\n".join(bleu_scores.keys()))
file_out = "plots/bleu"
if model_filter is not None:
file_out += '_' + model_filter
if token_filter is not None:
file_out += '_' + token_filter
if opt_filter is not None:
file_out += '_' + opt_filter
labels = []
xs = []
ys = []
plt.figure() # reset the plot
with open(file_out + '.csv',
mode='w') as csv_file: # overwrite any existing file
csv_writer = csv.writer(csv_file,
delimiter=',',
quotechar='"',
quoting=csv.QUOTE_MINIMAL)
# For each, lookup the val_loss and plot them
for model_name, model_class in models.items():
if model_filter is None or model_filter == model_name:
for token_id, tokenizer in tokenizers.items():
if token_filter is None or token_filter == token_id:
for opt_id, optimizer in optimizer_opts.items():
if opt_filter is None or opt_filter == opt_id:
# save each one
label = model_name + '_' + token_id + '_' + opt_id
filename = label + '_' + version
try:
history = genfromtxt('checkpoints/' +
filename + '.csv',
delimiter=',',
skip_header=1)
bleu = bleu_scores[filename]
val_loss = history[:, 2][-1]
csv_writer.writerow(
[filename, bleu, val_loss])
print("bleu=%s, val_loss=%s" %
(bleu, val_loss))
labels.append(label)
xs.append(bleu)
ys.append(val_loss)
# if isinstance(bleu, numbers.Number):
# plt.plot(bleu, val_loss, label=filename, markersize=12)
print("Plotted: " + filename)
except UserWarning as uw:
# print(uw)
traceback.print_exc()
# No model trained yet
print('No val_los history: ' + filename)
except Exception as e:
# print(e)
traceback.print_exc()
# No model trained yet
print('No model logs for: ' + filename)
if not labels:
print("No matching data for filter %s, %s, %s" %
(model_filter, token_filter, opt_filter))
else:
plt.scatter(xs, ys, marker='o')
for label, x, y in zip(labels, xs, ys):
plt.annotate(label,
xy=(x, y),
xytext=(-20, 20),
textcoords='offset points',
ha='right',
va='bottom',
bbox=dict(boxstyle='round,pad=0.5',
fc='yellow',
alpha=0.5),
arrowprops=dict(arrowstyle='->',
connectionstyle='arc3,rad=0'))
# summarize history for loss
plt.style.use('seaborn-whitegrid')
plt.title('BLEU-1 vs val_loss')
plt.ylabel('val_loss')
plt.xlabel('BLEU')
# plt.show()
plt.subplots_adjust(left=0.2, right=0.9, top=0.9, bottom=0.1)
plt.savefig(file_out + '.png')
print("Wrote plot to " + file_out)
for l in [0, 1, 10, 100]:
params['CLR_c k={} l={}'.format(
k, l)] = [CLRcRegressor(k, l, constr_id=constr_id), X, y]
params['CLR_c k={} l={} ens=10'.format(k, l)] = [
RegressorEnsemble(CLRcRegressor(k, l, constr_id=constr_id)), X,
y
]
params['kplane k={} l={} w=size'.format(
k, l)] = [KPlaneRegressor(k, l, weighted='size'), X, y]
params['kplane k={} l={} w=size ens=10'.format(k, l)] = [
RegressorEnsemble(KPlaneRegressor(k, l, weighted='size')), X, y
]
results = evaluate_all(
params,
file_name="results/{}-tmp1.csv".format(args.dataset),
n_jobs=args.n_jobs,
gl_parallel=args.global_parallel,
)
results = results.sort_values('test_mse_mean')
add_params = {}
# assuming ensembles are always best
algos = [CLRpRegressor, KPlaneRegressor]
algo_names = ['CLR_p', 'kplane']
for algo, algo_name in zip(algos, algo_names):
for idx in results.index:
if algo_name in idx:
k = int(idx.split()[1].split('=')[1])
l = int(idx.split()[2].split('=')[1])
w = idx.split()[3].split('=')[1] == 'True'
f = idx.split()[4].split('=')[1] == 'True'
import evaluate
if __name__ == '__main__':
evaluate.evaluate_all(models_dir='models/robotics_klcoeff')
def train(self, data, words, params):
start_time = time.time()
evaluate_all(self, words)
counter = 0
try:
for eidx in xrange(params.epochs):
kf = self.get_minibatches_idx(len(data),
params.batchsize,
shuffle=True)
uidx = 0
for _, train_index in kf:
uidx += 1
batch = [data[t] for t in train_index]
for i in batch:
i[0].populate_embeddings(words, True)
i[1].populate_embeddings(words, True)
(g1x, g1mask, g2x, g2mask, p1x, p1mask, p2x,
p2mask) = self.getpairs(batch, params)
cost = self.train_function(g1x, g2x, p1x, p2x, g1mask,
g2mask, p1mask, p2mask)
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected'
if utils.check_if_quarter(uidx, len(kf)):
if params.save:
counter += 1
self.save_params(
params.outfile + str(counter) + '.pickle',
words)
if params.evaluate:
evaluate_all(self, words)
for i in batch:
i[0].representation = None
i[1].representation = None
i[0].unpopulate_embeddings()
i[1].unpopulate_embeddings()
if params.save:
counter += 1
self.save_params(params.outfile + str(counter) + '.pickle',
words)
if params.evaluate:
evaluate_all(self, words)
print 'Epoch ', (eidx + 1), 'Cost ', cost
except KeyboardInterrupt:
print "Training interupted"
end_time = time.time()
print "total time:", (end_time - start_time)
def train(self, data, words, params):
start_time = time.time()
evaluate_all(self, words)
counter = 0
try:
for eidx in xrange(params.epochs):
kf = self.get_minibatches_idx(len(data),
params.batchsize,
shuffle=True)
uidx = 0
for _, train_index in kf:
uidx += 1
batch = [data[t] for t in train_index]
for i in batch:
if params.scramble:
n = np.random.binomial(1, params.scramble, 1)[0]
if n > 0:
self.scramble(i[0], words)
self.scramble(i[1], words)
else:
i[0].populate_embeddings(words)
i[1].populate_embeddings(words)
else:
i[0].populate_embeddings(words)
i[1].populate_embeddings(words)
(g1x, g1mask, g2x, g2mask, p1x, p1mask, p2x,
p2mask) = self.get_pairs(batch, params)
cost = self.train_function(g1x, g2x, p1x, p2x, g1mask,
g2mask, p1mask, p2mask)
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected. Exiting.'
sys.exit(0)
if (check_quarter(uidx, len(kf))):
if (params.save):
counter += 1
self.save_params(params.outfile + str(counter) +
'.pickle')
if (params.evaluate):
evaluate_all(self, words)
#undo batch to save RAM
for i in batch:
i[0].representation = None
i[1].representation = None
i[0].unpopulate_embeddings()
i[1].unpopulate_embeddings()
if (params.save):
counter += 1
self.save_params(params.outfile + str(counter) + '.pickle')
if (params.evaluate):
evaluate_all(self, words)
print 'Epoch ', (eidx + 1), 'Cost ', cost
except KeyboardInterrupt:
print "Training interupted"
end_time = time.time()
print "total time:", (end_time - start_time)
def train(self, data, words, params):
start_time = time.time()
evaluate_all(self, words, params)
old_v = 0
try:
for eidx in xrange(params.epochs):
kf = self.get_minibatches_idx(len(data),
params.batchsize,
shuffle=True)
lkf = len(kf)
uidx = 0
while (len(kf) > 0):
megabatch = []
idxs = []
idx = 0
for i in range(params.mb_batchsize):
if len(kf) > 0:
arr = [data[t] for t in kf[0][1]]
curr_idxs = [i + idx for i in range(len(kf[0][1]))]
kf.pop(0)
megabatch.extend(arr)
idxs.append(curr_idxs)
idx += len(curr_idxs)
uidx += len(idxs)
for i in megabatch:
if params.wordtype == "words":
if params.scramble > 0:
n = np.random.binomial(1, params.scramble,
1)[0]
if n > 0:
i[0].populate_embeddings_scramble(words)
i[1].populate_embeddings_scramble(words)
else:
i[0].populate_embeddings(words, True)
i[1].populate_embeddings(words, True)
else:
i[0].populate_embeddings(words, True)
i[1].populate_embeddings(words, True)
else:
i[0].populate_embeddings_ngrams(words, 3, True)
i[1].populate_embeddings_ngrams(words, 3, True)
(g1x, g1mask, g2x, g2mask, p1x, p1mask, p2x,
p2mask) = self.get_pairs(megabatch, params)
cost = 0
for i in idxs:
cost += self.train_function(g1x[i], g2x[i], p1x[i],
p2x[i], g1mask[i],
g2mask[i], p1mask[i],
p2mask[i])
cost = cost / len(idxs)
if np.isnan(cost) or np.isinf(cost):
print 'NaN detected'
if utils.check_if_quarter(uidx - len(idxs), uidx, lkf):
if params.evaluate:
v = evaluate_all(self, words, params)
if params.save:
if v > old_v:
old_v = v
self.save_params(params.outfile + '.pickle',
words)
for i in megabatch:
i[0].representation = None
i[1].representation = None
i[0].unpopulate_embeddings()
i[1].unpopulate_embeddings()
if params.evaluate:
v = evaluate_all(self, words, params)
if params.save:
if v > old_v:
old_v = v
self.save_params(params.outfile + '.pickle', words)
print 'Epoch ', (eidx + 1), 'Cost ', cost
except KeyboardInterrupt:
print "Training interupted"
end_time = time.time()
print "total time:", (end_time - start_time)
def train(self, data, words, params):
start_time = time.time()
evaluate_all(self, words, params)
old_v = 0
try:
for eidx in range(params.epochs):
kf = self.get_minibatches_idx(len(data), params.batchsize, shuffle=True)
lkf = len(kf)
uidx = 0
sentence_samples = []
while(len(kf) > 0):
megabatch = []
idxs = []
idx = 0
for i in range(params.mb_batchsize):
if len(kf) > 0:
arr = [data[t] for t in kf[0][1]]
curr_idxs = [i + idx for i in range(len(kf[0][1]))]
kf.pop(0)
megabatch.extend(arr)
idxs.append(curr_idxs)
idx += len(curr_idxs)
uidx += len(idxs)
for i in megabatch:
if params.wordtype == "words":
if params.scramble > 0:
n = np.random.binomial(1, params.scramble, 1)[0]
if n > 0:
i[0].populate_embeddings_scramble(words)
i[1].populate_embeddings_scramble(words)
else:
i[0].populate_embeddings(words, True)
i[1].populate_embeddings(words, True)
else:
i[0].populate_embeddings(words, True)
i[1].populate_embeddings(words, True)
else:
i[0].populate_embeddings_ngrams(words, 3, True)
i[1].populate_embeddings_ngrams(words, 3, True)
(g1x, g1mask, g2x, g2mask, p1x, p1mask, p2x, p2mask),(g1_s,g2_s,p1_s,p2_s) = self.get_pairs(megabatch, params)
cost = 0
for i in idxs:
# cc1,cc2 = self.cost_each_data(g1x[i], g2x[i], p1x[i], p2x[i], g1mask[i], g2mask[i], p1mask[i], p2mask[i])
cost += self.train_function(g1x[i], g2x[i], p1x[i], p2x[i], g1mask[i], g2mask[i], p1mask[i], p2mask[i])
# for j in range(len(i)):
# try:
# sentence_samples.append({'orig':g1_s[i[j]],'para':g2_s[i[j]], 'neg_orign':p1_s[i[j]], 'neg_para':p2_s[i[j]], 'orig_cost':str(cc1[j]),'para_cost':str(cc2[j])})
# except IndexError:
# print(j,i[j])
cost = cost / len(idxs)
if np.isnan(cost) or np.isinf(cost):
print('NaN detected')
if utils.check_if_quarter(uidx-len(idxs), uidx, lkf):
if params.evaluate:
v = evaluate_all(self, words, params)
if params.save:
if v > old_v:
old_v = v
self.save_params(params.outfile + '.pickle', words)
for i in megabatch:
i[0].representation = None
i[1].representation = None
i[0].unpopulate_embeddings()
i[1].unpopulate_embeddings()
if params.evaluate:
v = evaluate_all(self, words, params)
if params.save:
if v > old_v:
old_v = v
self.save_params(params.outfile + '.pickle', words)
print('Epoch ', (eidx + 1), 'Cost ', cost)
#with open("../data/sampled_samples/%s_epoch_%d.json" % (params.model, eidx),'w') as f:
# json_result = json.dumps(sentence_samples, indent=2)
# f.write(json_result)
except KeyboardInterrupt:
print("Training interupted")
end_time = time.time()
print("total time:", (end_time - start_time))
from evaluate import evaluate_all evaluate_all()
y = abalone_data.iloc[:, 8].as_matrix().astype(np.float)
constr_id = 10
elif args.dataset == 'auto-mpg':
data = pd.read_csv('data/auto-mpg.data', header=None, sep='\s+', na_values='?')
data = data.dropna()
X = pd.get_dummies(data.iloc[:,1:-1], columns=[7]).as_matrix().astype(np.float)
y = data[0].as_matrix().astype(np.float)
constr_id = 5
else:
print("Dataset is not supported")
sys.exit(0)
X, y = preprocess_data(X, y)
params = {}
for k in [2, 8]:
for ens in range(1, 21):
params['kplane k={} ens={}'.format(k, ens)] = [
RegressorEnsemble(KPlaneRegressor(k, 100), n_estimators=ens), X, y, 10, 20]
params['CLR_p k={} ens={}'.format(k, ens)] = [
RegressorEnsemble(CLRpRegressor(k, 10, weighted=True), n_estimators=ens), X, y, 10, 20]
params['CLR_c k={} ens={}'.format(k, ens)] = [
RegressorEnsemble(CLRcRegressor(k, 10, constr_id=constr_id), n_estimators=ens), X, y, 10, 20]
results = evaluate_all(
params,
file_name="results_ens/{}.csv".format(args.dataset),
n_jobs=args.n_jobs,
gl_parallel=args.global_parallel,
)
for min_samples_leaf in [1, 10, 30, 50]:
params['rf md={}, mf={}, mss={}, msl={}'.format(
max_depth, max_features, min_samples_split,
min_samples_leaf)] = [
RandomForestRegressor(
n_estimators=30,
max_depth=max_depth,
max_features=max_features,
min_samples_leaf=min_samples_leaf,
min_samples_split=min_samples_split,
n_jobs=n_jobs), X, y, 3, 1
]
results = evaluate_all(
params,
file_name="results/patient-claims-rf.csv",
n_jobs=args.n_jobs,
gl_parallel=args.global_parallel,
)
if args.run_clrs:
print("Run clrs")
if args.n_jobs == 1:
for k in [2, 4, 6, 8]:
for l in [0, 1, 10, 100, 1000, 10000]:
tm = time.time()
gen_clrs(k, l, X, y, max_iter=5, n_estimators=10)
print("k={}, l={}, time={}".format(k, l, time.time() - tm))
tm = time.time()
gen_clrs(k,
l,
X,
def predict(self, data, words, params):
start_time = time.time()
evaluate_all(self, words, params)
end_time = time.time()
print "total time:", (end_time - start_time)