def train(epoch, learningRate, batchSize):
r"""
Training and saving my model.
:param epoch:
:param learningRate:
:param batchSize:
:return: None
"""
train = load_train()
trainset = torch.utils.data.DataLoader(train, batch_size=batchSize)
test = load_test()
testset = torch.utils.data.DataLoader(test, batch_size=batchSize)
vectorNet = VectorNetWithPredicting(len=9, timeStampNumber=30)
# for para in vectorNet.named_parameters():
# print(para)
# exit(0)
vectorNet = vectorNet.to(device)
lossfunc = torch.nn.MSELoss()
lr = learningRate
optimizer = torch.optim.Adam(vectorNet.parameters(), lr=lr)
for iterator in range(epoch):
for data, target in trainset:
data = data.to(device)
target = target.to(device)
optimizer.zero_grad()
offset = data[:, -1, :] # [0, 0, 0, 0, 0, maxX, maxY, ..., 0]
data = data[:, 0:data.shape[1] - 1, :]
outputs = vectorNet(data) # [batch size, len*2]
loss = lossfunc(outputs, target)
loss.backward()
# print(iterator)
# print('loss=',loss.item())
# t = askADE(outputs, target)
# print('minDis=', t[0].item(), 'ADE=', t[1].item())
optimizer.step()
print(iterator)
if (iterator + 1) % 5 == 0:
lr *= 0.3
optimizer = torch.optim.Adam(vectorNet.parameters(), lr=lr)
minADE = torch.zeros(1).to(device)
minDis = torch.zeros(1).to(device)
minDis[0] = float('inf')
for data, target in testset:
data = data.to(device)
target = target.to(device)
offset = data[:, -1, :] # [0, 0, 0, 0, 0, maxX, maxY, ..., 0]
data = data[:, 0:data.shape[1] - 1, :]
outputs = vectorNet(data)
for i in range(0, outputs.shape[1], 2):
outputs[:, i] *= offset[:, 5]
target[:, i] *= offset[:, 5]
outputs[:, i + 1] *= offset[:, 6]
target[:, i + 1] *= offset[:, 6]
print(outputs)
print(target)
# loss = lossfunc(outputs, target)
# print('loss=', loss.item())
t = askADE(outputs, target)
# print('minDis=', t[0].item(), 'ADE=', t[1].item())
# minADE += t[1]
if minDis > t[0]:
minDis = t[0]
minADE = t[1]
print('minDis=', minDis.item(), 'minADE=', minADE.item())
torch.save(vectorNet, 'VectorNet-test.model')
listX.append(y[0, i])
listY.append(y[0, i + 1])
print(listX)
print(listY)
print('---------')
plt.plot(listX, listY, 'g', linewidth=3) # ground truth
x = x.to(device)
myPredict = vectorNet(x)
listX = [0]
listY = [0]
for i in range(0, y.shape[1], 2):
listX.append(myPredict[0, i])
listY.append(myPredict[0, i + 1])
print(listX)
print(listY)
print('---------')
plt.plot(listX, listY, 'yellow', linewidth=3) # predict
plt.show()
if __name__ == '__main__':
data = load_train()
dataset = torch.utils.data.DataLoader(data, batch_size=1)
for data, y in dataset:
offset = data[:, -1, :] # [0, 0, 0, 0, 0, maxX, maxY, ..., 0]
data = data[:, 0:data.shape[1] - 1, :]
viz(data, y)
# exit(0)
import dataloader
import transformer
import numpy as np
import matplotlib.pyplot as plt
'''
base analysis
'''
desc_map, price_map = dataloader.load_train()
plt.scatter(desc_map['YearBuilt'], price_map['price'], s=10)
plt.title('SalePrice_YearBuilt')
plt.xlabel('YearBuilt')
plt.ylabel('SalePrice')
plt.show()
print('SalePrice mean:', np.mean(price_map['price']), 'median:',
np.median(price_map['price']), 'std:', np.std(price_map['price']))
'''
corrcoefs analysis
'''
corrcoefs = []
keys = []
# corrcoefs_map ={}
desc_map, price_map = dataloader.load_train()
price_map['price'] = transformer.normlize(price_map['price'])
for key in desc_map.keys():
desc_map[key] = transformer.normlize(desc_map[key])
corrcoefs.append(np.corrcoef(desc_map[key], price_map['price'])[0][1])
keys.append(key)
# corrcoefs_map[key]
zipped = zip(keys, corrcoefs)
sort_zipped = sorted(zipped, key=lambda x: (x[1], x[0]), reverse=True)
#decay by time def act_score_by_time(act, time, decay_weight = 0.98): return act_score(act) * (decay_weight ** (time)) def cal_pop(train): pop = defaultdict(float) for record in train: uid, bid, act, time = record[0], record[1], record[2], record[3] pop[bid] += act_score_by_time(act, time) #pop[bid] += act_score(act) #normalize pop_sum = sum(pop) for bid in pop: pop[bid] /= pop_sum return pop def dump(dic, path): fo = open(path, 'w') for bid in dic: fo.write(str(bid) + '\t' + str(dic[bid]) + '\n') fo.close() if __name__ == "__main__": path = 'data/train.csv' train = load_train(path) pop = cal_pop(train) dump(pop, 'data/pop.csv')
def shuffle(f_submit): f = open(f_submit) submit = [] for line in f: submit.append(line) f.close() for i in range(0, len(submit)): idx = random.randint(0,len(submit)-1) submit[i], submit[idx] = submit[idx], submit[i] f = open(f_submit, 'w') for i in range(0, len(submit)): f.write(submit[i]) f.close() if __name__ == "__main__": f_train = 'data/train.csv' f_pref = 'data/long_pref.csv' f_pop = 'data/pop.csv' train = load_train(f_train) pref = load_pref(f_pref) short_pref = cal_short_pref(train) pop = load_pop(f_pop) ans = gen_ans(pref, pop, short_pref) topk = cal_topk(train) write_ans(ans, topk, 'data/pop_pref_0414.txt')
