def valid(self, epoch):
self.dataset.mode = 'valid'
dataloader = DataLoad(self.dataset, batch_size = self.batch_size, shuffle = False)
self.model.eval()
total_acc = 0
total_loss = 0
cnt = 0
for batch_idx, data in enumerate(dataloader):
contexts, questions, answers = data
batch_size = len(answers)
output = self.model(contexts, questions, self.dataset.train_tree_dict)
predict = self.model.predict(output, questions, self.dataset.train_tree_dict)
acc = self.metrics(predict, answers)
m = nn.LogSoftmax(dim = 1)
loss = self.criteria(m(output), answers)
print('|', batch_size, 'acc:', acc, 'loss:', loss,'|')
total_loss += loss.item()
total_acc += acc * batch_size
cnt += batch_size
total_acc = total_acc / cnt
if total_acc > self.best_acc:
self.best_acc = total_acc
self.best_state = self.model.state_dict()
self.early_stopping_cnt = 0
else:
self.early_stopping_cnt += 1
if self.early_stopping_cnt >= 256:
self.early_stopping_flag = True
with open(self.LOG_PATH, 'a+') as fp:
fp.write(f'[Run {self.run}, Task {self.task_id}, Epoch {epoch}] [Validate] Accuracy : {total_acc: {5}.{4}} Loss : {total_loss / cnt: {5}.{4}}' + '\n')
return total_acc
def train(self, epoch):
self.dataset.mode = 'train'
# 每一次取出 batch 对 batch 再按照contexts长度进行分组
dataloader = DataLoad(self.dataset, batch_size = self.batch_size)
self.model.train()
total_acc = 0
total_loss = 0
cnt = 0
# 0-6 7-12 13-18... 有多少组循环多少次 每个data里的context长度相似
for batch_idx, data in enumerate(dataloader):
self.optimizer.zero_grad()
contexts, questions, answers = data
batch_size = len(answers)
'''
batch_idx 0
contexts ([0, 1, 3, 4, 6, 7, 9, 10, 12, 13], [0, 1, 3, 4, 6, 7], [15, 16, 18, 19, 21, 22], [15, 16, 18, 19, 21, 22, 24, 25], [0, 1, 3, 4, 6, 7, 9, 10])
questions tensor([[ 10, 11, 14],
[ 10, 11, 12],
[ 10, 11, 14],
[ 10, 11, 7],
[ 10, 11, 12]])
answers tensor([ 5, 9, 5, 16, 16])
'''
output = self.model(contexts, questions, self.dataset.train_tree_dict)
predict = self.model.predict(output, questions, self.dataset.train_tree_dict)
m = nn.LogSoftmax(dim = 1)
loss = self.criteria(m(output), answers)
total_loss += loss.item()
acc = self.metrics(predict, answers)
total_acc += acc * batch_size
cnt += batch_size
loss.backward()
self.optimizer.step()
with open(self.LOG_PATH, 'a+') as fp:
fp.write(f'[Run {self.run}, Task {self.task_id}, Epoch {epoch}] [Training] loss : {total_loss / cnt: {10}.{8}}, acc : {total_acc / cnt: {5}.{4}}, batch_idx : {batch_idx}' + '\n')
def test(self):
self.dataset.mode = 'test'
dataloader = DataLoad(self.dataset,
batch_size=self.batch_size,
shuffle=False)
self.model.load_state_dict(self.best_state)
self.model.eval()
total_acc = 0
cnt = 0
for batch_idx, data in enumerate(dataloader):
contexts, questions, answers = data
batch_size = len(answers)
output = self.model(contexts, questions,
self.dataset.test_tree_dict)
predict = self.model.predict(output, questions,
self.dataset.test_tree_dict)
acc = self.metrics(predict, answers)
total_acc += acc * batch_size
cnt += batch_size
with open(self.LOG_PATH, 'a+') as fp:
fp.write(
f'[Run {self.run}, Task {self.task_id}, Epoch {epoch}] [Test] Accuracy : {total_acc / cnt: {5}.{4}}'
+ '\n')
os.makedirs('models', exist_ok=True)
with open(
f'models/task{self.task_id}_epoch{epoch}_run{self.run}_acc{total_acc/cnt}.pth',
'wb') as fp:
torch.save(self.model.state_dict(), fp)
D['X_val'] = np.load(datadirec + 'X_val.npy')
D['y_val'] = np.load(datadirec + 'y_val.npy')
print('Read data from pickles')
except:
print('One or more files are missing from ' + datadirec)
else:
os.makedirs(datadirec)
names = 3
filenames = []
for i in range(names):
filenames.append(
('/home/rob/Dropbox/ml_projects/deepclust/data/male' + str(i + 1) +
'.wav', '/home/rob/Dropbox/ml_projects/deepclust/data/female' +
str(i + 1) + '.wav'))
dl = DataLoad(seq_len, Nn, low_freq=6000)
dl.read_data(filenames)
dl.strip_zero()
D = dl.return_data()
for key, value in D.iteritems():
np.save(datadirec + key + '.npy', value)
print('Finished saving data')
#Obtain some sizes
N = D['X_train'].shape[0]
Nval = D['X_val'].shape[0]
"""Visualize some data"""
row = 10
ind = np.random.choice(D['X_train'].shape[0], row)
f, axarr = plt.subplots(row, 1)
def test_file_exists(self):
# A dataframe must be returned when the path exists
print('testing valid file...')
DL = DataLoad()
df = DL.load_csv('data/ross_train.csv', show_details=False)
self.assertIsInstance(df, pd.DataFrame)
def test_invalid_format(self):
print('testing invalid file format - not CSV...')
DL = DataLoad()
df = DL.load_csv('data/transportation.json')
self.assertEqual(df, None)
def test_dir(self):
print('testing dir instead of file...')
DL = DataLoad()
df = DL.load_csv('data')
self.assertEqual(df, None)
def test_corrupted_file(self):
print('testing corrupted file...')
DL = DataLoad()
df = DL.load_csv('data/corrupted.csv')
self.assertEqual(df, None)
def test_file_does_not_exist(self):
print('testing inexistent file...')
DL = DataLoad()
df = DL.load_csv('data/ross_trai.csv')
self.assertEqual(df, None)
from dataload import DataLoad
from evaluation import EvaluateModel
from preprocessing import DataSplit
from models import Exponential_Smoothing, AutoArima, EnsembleModels, FBProphet
import matplotlib.pyplot as plt
import pandas as pd
import warnings
warnings.simplefilter("ignore")
from datapreparation import DataVisualization
print('Time Series model composition')
# Loading Data
DL = DataLoad()
df = DL.load_csv('data/ross_train.csv')
df = df[df['Store'] == 1]
df = df[['Date', 'Sales']]
df['Date'] = pd.to_datetime(df['Date'], format='%Y-%m-%d')
df.set_index('Date', inplace=True)
df.sort_index(inplace=True)
print(df)
# Analyzing using different frequencies
x = []
y = []
title = []
# Daily
x.append(df.index)
y.append(df.Sales)
title.append('Daily')
from models import Exponential_Smoothing, Arima, EnsembleModels, FBProphet, RandomForest, XGBoost, SupportVectorRegressor
import warnings
from dataload import DataLoad
from analytics import DataVisualization
from evaluation import EvaluateModel
from preprocessing import FeatureEngineering
warnings.simplefilter("ignore")
print('Time Series model composition')
# Loading Data
DL = DataLoad()
df = DL.load_csv(file_path='data/ross_train.csv',
date='Date',
target='Sales',
show_details=True)
print(df)
# Analyzing using different frequencies
x = []
y = []
title = []
# Daily
x.append(df.index)
y.append(df.Sales)
title.append('Daily')
freq = {'W': 'Weekly', 'M': 'Monthly', 'Y': 'Yearly'}
for f in freq.keys():
sampled_df = df.resample(f).sum()
x.append(sampled_df.index)