def main(load=False, lr=1e-3, savestr="", reset=True, palette=False):
total_epochs = 10
iter_per_epoch = 100000
lr = lr
optim = None
starting_epoch = 0
starting_iteration = 0
logfile = "log.txt"
num_workers = 3
ig = InputGenD()
trainds, validds = train_valid_split(ig, split_fold=10)
traindl = DataLoader(dataset=trainds,
batch_size=1,
num_workers=num_workers)
validdl = DataLoader(dataset=validds, batch_size=1)
print("Using", num_workers, "workers for training set")
computer = NotMySam(input_size=47764,
hidden_size=128,
last_output_size=3620,
rnn_type='lstm',
num_layers=4,
nr_cells=100,
cell_size=32,
read_heads=4,
sparse_reads=4,
batch_first=True,
gpu_id=0)
# load model:
if load:
print("loading model")
computer, optim, starting_epoch, starting_iteration = load_model(
computer, optim, starting_epoch, starting_iteration, savestr)
computer = computer.cuda()
if optim is None:
print("Using Adam with lr", lr)
optimizer = torch.optim.Adam(
[i for i in computer.parameters() if i.requires_grad], lr=lr)
else:
# print('use Adadelta optimizer with learning rate ', lr)
# optimizer = torch.optim.Adadelta(computer.parameters(), lr=lr)
optimizer = optim
real_criterion = nn.SmoothL1Loss()
binary_criterion = nn.BCEWithLogitsLoss(size_average=False)
# starting with the epoch after the loaded one
train(computer, optimizer, real_criterion, binary_criterion, traindl,
iter(validdl), int(starting_epoch), total_epochs,
int(starting_iteration), iter_per_epoch, savestr, logfile)
def main(load=False, lr=1e-4, savestr="6"):
total_epochs = 10
iter_per_epoch = 10000
lr = lr
optim = None
starting_epoch = 0
starting_iteration = 0
logfile = "smalltacolog.txt"
num_workers = 32
ig = InputGenD()
# multiprocessing disabled, because socket request seems unstable.
# performance should not be too bad?
trainds, validds = train_valid_split(ig, split_fold=10)
traindl = DataLoader(dataset=trainds,
batch_size=32,
num_workers=num_workers,
collate_fn=pad_collate)
validdl = DataLoader(dataset=validds,
batch_size=8,
num_workers=4,
collate_fn=pad_collate)
print("Using", num_workers, "workers for training set")
computer = Tacotron()
# load model:
if load:
print("loading model")
computer, optim, starting_epoch, starting_iteration = load_model(
computer, optim, starting_epoch, starting_iteration, savestr)
computer = computer.cuda()
if optim is None:
print("Using Adam with lr", lr)
optimizer = torch.optim.Adam(
[i for i in computer.parameters() if i.requires_grad], lr=lr)
else:
# print('use Adadelta optimizer with learning rate ', lr)
# optimizer = torch.optim.Adadelta(computer.parameters(), lr=lr)
optimizer = optim
real_criterion = nn.SmoothL1Loss()
# time-wise sum, label-wise average.
binary_criterion = nn.BCEWithLogitsLoss()
# starting with the epoch after the loaded one
train(computer, optimizer, real_criterion, binary_criterion,
traindl, validdl, int(starting_epoch), total_epochs,
int(starting_iteration), iter_per_epoch, savestr, logfile)
def main():
total_epochs = 10
iter_per_epoch = 100000
lr = 1e-5
target_dim = 3656
logfile = "log.txt"
num_workers = 3
ig = InputGenD()
# multiprocessing disabled, because socket request seems unstable.
# performance should not be too bad?
trainds, validds = train_valid_split(ig, split_fold=10)
traindl = DataLoader(dataset=trainds,
batch_size=1,
num_workers=num_workers)
validdl = DataLoader(dataset=validds, batch_size=1)
print("Using", num_workers, "workers for training set")
computer = DNC()
computer.train()
# load model:
if True:
print("loading model")
computer, optim, starting_epoch, starting_iteration = load_model(
computer)
computer = computer.cuda()
if optim is None:
optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad,
computer.parameters()),
lr=lr)
else:
# print('use Adadelta optimizer with learning rate ', lr)
# optimizer = torch.optim.Adadelta(computer.parameters(), lr=lr)
optimizer = optim
real_criterion = nn.SmoothL1Loss()
binary_criterion = nn.BCEWithLogitsLoss(size_average=False)
# starting with the epoch after the loaded one
train(computer, optimizer, real_criterion, binary_criterion, traindl,
iter(validdl), int(starting_epoch), total_epochs,
int(starting_iteration), iter_per_epoch, target_dim, logfile)
def main(load=False, lr=1e-3):
total_epochs = 10
iter_per_epoch = 100000
lr = lr
optim = None
starting_epoch = 0
starting_iteration = 0
logfile = "log.txt"
num_workers = 3
ig = InputGenD()
trainds, validds = train_valid_split(ig, split_fold=10)
traindl = DataLoader(dataset=trainds,
batch_size=1,
num_workers=num_workers)
validdl = DataLoader(dataset=validds, batch_size=1)
print("Using", num_workers, "workers for training set")
computer = DNC()
# load model:
if load:
print("loading model")
computer, optim, starting_epoch, starting_iteration = load_model(
computer, optim, starting_epoch, starting_iteration)
computer = computer.cuda()
if optim is None:
print("Using Adam with lr", lr)
optimizer = torch.optim.Adam(
[i for i in computer.parameters() if i.requires_grad], lr=lr)
else:
# print('use Adadelta optimizer with learning rate ', lr)
# optimizer = torch.optim.Adadelta(computer.parameters(), lr=lr)
optimizer = optim
real_criterion = nn.SmoothL1Loss()
binary_criterion = nn.BCEWithLogitsLoss(size_average=False)
# starting with the epoch after the loaded one
train(computer, optimizer, real_criterion, binary_criterion, traindl,
iter(validdl), int(starting_epoch), total_epochs,
int(starting_iteration), iter_per_epoch, logfile)
# this file contains all the statistical models for baselines.
# I chose to do all of these in Python, not R, because the data is finalized in Python, not R. We can import from
# python to R, but it seems unreasonable.
from death.post.inputgen_planD import InputGenD, train_valid_split
from sklearn.datasets import load_iris
from sklearn.linear_model import LogisticRegression
# Logistic regression.
# The goal is bunch of binary values.
ig = InputGenD()
trainds, validds = train_valid_split(ig, split_fold=10)
lr = LogisticRegression(random_state=0, multi_class="ovr")
# X is a nd array with (150,4)
# y is a nd array with (150)
X, y = load_iris(return_X_y=True)
# lrfit=lr.fit(X,y)
print("end")
### To see the pandas data frames # Note that DFManager allows you to load from pickle file or raw csv files # pickle loading is much faster. loading raw rebuilds pickle # from death.post.dfmanager import DFManager dfs = DFManager() dfs.load_pickle(verbose=True) # from here, you can see all dataframes as dfs' properties # for example, this is demographics csv: print(dfs.demo) # if you want to load_raw and rebuild pickle files, run: dfs.load_raw(save=True) # making dictionary necessary for one-hot encodings dfs.make_dictionary(verbose=True, save=True, skip=False) ### To see the inputs and outputs used by the deep learning model from death.post.inputgen_planD import InputGenD, train_valid_split ig = InputGenD(verbose=False) # split to training set and validation set if you want # it's fine if you don't do this step train, valid = train_valid_split(ig) # __getitem__() method is how you should access this dataset print(train[123]) ### Loading into PyTorch is trickier, because sequences don't have even lengths # I have two solutions, one with ChannelManager and one with padded sequences # you should see the script for BatchDNC and Tacotron respectively.
