def run():
# Test code goes here
from src.model.model import RubiksModel
import random
import numpy as np
import torch
ret = True
msg = ""
# Ensure model initialization
try:
model = RubiksModel()
except:
msg += "ERR: Model unable to be initialized.\n"
return ret, msg
# Ensure forward pass
state = np.array([[random.randint(0, 5) for i in range(6)] * 3,\
[random.randint(0, 5) for i in range(6)] * 3,\
[random.randint(0, 5) for i in range(6)] * 3], dtype=int)
state = torch.unsqueeze(torch.tensor(state), 0)
try:
ret = model(state)
expected = torch.random([1, 21])
flag = 1 / torch.all(ret.shape == expected.shape)
except:
msg += "ERR: Forward pass failed.\n"
return ret, msg
def get_random_range(self, board, seed):
seed = seed or torch.random()
gen = torch.Generator()
torch.manualSeed(gen, seed)
out = torch.rand(gen,
game_settings.card_count).typeAs(arguments.Tensor())
out.cmul(self.get_possible_hand_indexes(board))
out.div(out.sum())
return out
def __init__(self, n, k):
'''
n -> gaussian size
k -> number of gaussians
'''
super().__init__()
self.n = n
self.k = k
self.mu = net.Parameter(torch.rand(k, n))
self.sigma = net.Parameter(torch.rand(k, n, n))
self.weights = net.Parameter(torch.random(k))
def getbatch():
batch = torch.Tensor(128,3,1,64,64)
for i = 1,128 do
seed = torch.random(1, 100000) -- fix seed
gen = torch.Generator()
torch.manualSeed(gen, i*seed)
r1 = torch.random(gen,1,cn)
r2 = torch.random(gen,1,cn)
r3 = torch.random(gen,1,mn[r1])
path1 = cloth_table[r1]
path2 = cloth_table[r2]
path3 = models_table[r1][r3]
img1 = loadImage(path1)
img2 = loadImage(path2)
img3 = loadImage(path3)
batch[i][1] = img1
batch[i][2] = img2
batch[i][3] = img3
def _generate_recursion(self, cards, mass):
batch_size = cards.size(0)
assert mass.size(0) == batch_size
# we terminate recursion at size of 1
card_count = cards.size(1)
if card_count == 1:
cards.copy_(mass)
else:
rand = torch.rand(batch_size)
if arguments.gpu:
rand = rand.cuda()
mass1 = mass.clone().cmul(rand)
mass2 = mass - mass1
halfSize = card_count / 2
# if the tensor contains an odd number of cards, randomize which way the
# middle card goes
if halfSize % 1 != 0:
halfSize = halfSize - 0.5
halfSize = halfSize + torch.random(0, 1)
self._generate_recursion(cards[:, 0:halfSize, :], mass1)
self._generate_recursion(cards[:, halfSize + 1, -1, :], mass2)
def first_hidden_state(self):
initial_hidden = torch.random(self.batch_size, self.seq_len_size,
self.d_model)
return initial_hidden
def reparameterize(self,mu,logvar):
epsi = Variable(torch.random(mu.size(0), mu.size(1))).cuda()
z = mu + epsi*torch.exp(logvar/2)
return z
def initHidden(self):
return (torch.random(1, self.hidden_size, self.hidden_size),
torch.random(1, self.hidden_size, self.hidden_size))
def initHidden(self):
return (torch.random((1, hidden_size, hidden_size)),
torch.random((1, hidden_size, hidden_size)))
def main():
# Prepare Dataset
train_dataset = MyDataset(config.train_info, train=True)
val_dataset = MyDataset(config.train_info, train=False)
# for k, v in config.train_info.items():
# pass
# train_dataset = Mscoco(v, train=True)
# val_dataset = Mscoco(v, train=False)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=config.train_batch,
shuffle=True,
num_workers=config.train_mum_worker,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=config.val_batch,
shuffle=True,
num_workers=config.val_num_worker,
pin_memory=True)
# for k, v in config.train_info.items():
# train_dataset = Mscoco([v[0], v[1]], train=True, val_img_num=v[2])
# val_dataset = Mscoco([v[0], v[1]], train=False, val_img_num=v[2])
#
# train_loaders[k] = torch.utils.data.DataLoader(
# train_dataset, batch_size=config.train_batch, shuffle=True, num_workers=config.train_mum_worker,
# pin_memory=True)
#
# val_loaders[k] = torch.utils.data.DataLoader(
# val_dataset, batch_size=config.val_batch, shuffle=False, num_workers=config.val_num_worker, pin_memory=True)
#
# train_loader = torch.utils.data.DataLoader(
# train_dataset, batch_size=config.train_batch, shuffle=True, num_workers=config.train_mum_worker,
# pin_memory=True)
# val_loader = torch.utils.data.DataLoader(
# val_dataset, batch_size=config.val_batch, shuffle=False, num_workers=config.val_num_worker, pin_memory=True)
# assert train_loaders != {}, "Your training data has not been specific! "
# Model Initialize
if device != "cpu":
m = createModel(cfg=model_cfg).cuda()
else:
m = createModel(cfg=model_cfg).cpu()
begin_epoch = 0
pre_train_model = config.loadModel
flops = print_model_param_flops(m)
print("FLOPs of current model is {}".format(flops))
params = print_model_param_nums(m)
print("Parameters of current model is {}".format(params))
if pre_train_model:
print('Loading Model from {}'.format(pre_train_model))
m.load_state_dict(torch.load(pre_train_model))
opt.trainIters = config.train_batch * (begin_epoch - 1)
opt.valIters = config.val_batch * (begin_epoch - 1)
begin_epoch = int(pre_train_model.split("_")[-1][:-4]) + 1
os.makedirs("exp/{}/{}".format(dataset, save_folder), exist_ok=True)
else:
print('Create new model')
with open("log/{}.txt".format(save_folder), "a+") as f:
f.write("FLOPs of current model is {}\n".format(flops))
f.write("Parameters of current model is {}\n".format(params))
if not os.path.exists("exp/{}/{}".format(dataset, save_folder)):
try:
os.mkdir("exp/{}/{}".format(dataset, save_folder))
except FileNotFoundError:
os.mkdir("exp/{}".format(dataset))
os.mkdir("exp/{}/{}".format(dataset, save_folder))
if optimize == 'rmsprop':
optimizer = torch.optim.RMSprop(m.parameters(),
lr=config.lr,
momentum=config.momentum,
weight_decay=config.weightDecay)
elif optimize == 'adam':
optimizer = torch.optim.Adam(m.parameters(),
lr=config.lr,
weight_decay=config.weightDecay)
else:
raise Exception
if mix_precision:
m, optimizer = amp.initialize(m, optimizer, opt_level="O1")
writer = SummaryWriter('tensorboard/{}/{}'.format(dataset, save_folder))
# Model Transfer
if device != "cpu":
m = torch.nn.DataParallel(m).cuda()
criterion = torch.nn.MSELoss().cuda()
else:
m = torch.nn.DataParallel(m)
criterion = torch.nn.MSELoss()
rnd_inps = torch.random([2, 3, 224, 224])
writer.add_graph(m, rnd_inps)
# Start Training
for i in range(config.epochs)[begin_epoch:]:
os.makedirs("log/{}".format(dataset), exist_ok=True)
log = open("log/{}/{}.txt".format(dataset, save_folder), "a+")
print('############# Starting Epoch {} #############'.format(i))
log.write('############# Starting Epoch {} #############\n'.format(i))
for name, param in m.named_parameters():
writer.add_histogram(name, param.clone().data.to("cpu").numpy(), i)
loss, acc = train(train_loader, m, criterion, optimizer, writer)
print('Train-{idx:d} epoch | loss:{loss:.8f} | acc:{acc:.4f}'.format(
idx=i, loss=loss, acc=acc))
log.write(
'Train-{idx:d} epoch | loss:{loss:.8f} | acc:{acc:.4f}\n'.format(
idx=i, loss=loss, acc=acc))
opt.acc = acc
opt.loss = loss
m_dev = m.module
loss, acc = valid(val_loader, m, criterion, optimizer, writer)
print('Valid:-{idx:d} epoch | loss:{loss:.8f} | acc:{acc:.4f}'.format(
idx=i, loss=loss, acc=acc))
log.write(
'Valid:-{idx:d} epoch | loss:{loss:.8f} | acc:{acc:.4f}\n'.format(
idx=i, loss=loss, acc=acc))
log.close()
if i % config.save_interval == 0:
torch.save(
m_dev.state_dict(),
'exp/{}/{}/model_{}.pkl'.format(dataset, save_folder, i))
torch.save(opt,
'exp/{}/{}/option.pkl'.format(dataset, save_folder, i))
torch.save(optimizer,
'exp/{}/{}/optimizer.pkl'.format(dataset, save_folder))
writer.close()
import torch.nn as nn from torch.nn import Parameter import torch.nn.functional as F from torch import exp, lgamma from hydroDL.new.model import flowPath import importlib importlib.reload(flowPath) nt = 1000 rho = 365 nb = 100 nx = 1 nq = 40 x = torch.randn(nt, nb, nx).cuda() y = torch.random(nt, nb, 1).cuda() hiddenSize = 64 inputSize = nx convSize = nq batchSize = x.shape[1] # model = flowPath(nx, 256, nq).cuda() # yp = model(x, rho) rnn = nn.RNN(inputSize, hiddenSize) linear = torch.nn.Linear(hiddenSize, convSize) aT = exp(Parameter(torch.Tensor(nq))) bT = exp(Parameter(torch.Tensor(nq))) out1, hn = rnn(x)
full_loss = torch.where(tru_mask > 0.5, 1.,
0.) # truth_table[:, :, (3,)] > 0.5
m_only_loss = torch.where(
torch.where(tru_mask > 0.5, 0., 1.) * out_mask > 0.5, 1., 0.
) # truth_table[:, :, (3,)] < 0.5 and output_table[:, :, (3,)] > 0.5
no_loss = torch.where(
torch.where(tru_mask > 0.5, 0., 1.) * out_mask <= 0.5, 1., 0.
) # truth_table[:, :, (3,)] < 0.5 and output_table[:, :, (3,)] < 0.5
total_loss = (dirjovis_loss * m_only_loss + mask_loss) * no_loss
'''
out m | truth m
1 1 -> apply loss to all
1 0 -> apply loss to all
0 1 -> apply loss only to m
0 0 -> apply no loss
'''
return torch.sum(total_loss)
if __name__ == "__main__":
output_table = torch.random((9, 16, 6)) * 0.5
truth_table = torch.random((9, 16, 6)) * 0.5
loss = CustomLoss()
out = loss(output_table, truth_table)
print(out)
def sample(self, sample_shape):
return torch.random(self._batch_shape)
def train(self):
print("Training Started")
bce = torch.nn.BCELoss()
mse = torch.nn.MSELoss()
l1 = torch.nn.L1Loss()
for epoch in range(self.epochs):
iter = 0
# print(type(self.dataset))
# print(self.dataset)
# i = 0
for point in enumerate(self.dataLoader):
# print(len(point))
correct_image = point['correct_image']
incorrect_image = point['incorrect_image']
correct_embed = point['correct_embed']
#---------------------------------------------------------------------
#For discriminator
correct_image = Variable(correct_image.float()).cuda()
incorrect_image = Variable(incorrect_image.float()).cuda()
correct_embed = Variable(correct_embed.float()).cuda()
incorrect_labels = Variable(np.zeroes(self.batch_size)).cuda()
# One Sided Label Smoothing
correct_labels = torch.FloatTensor(
np.ones(self.batch_size) + -1)
correct_labels = Variable(correct_labels).cuda()
self.disc.zero_grad()
# Right images and right caption
output, activations = self.disc(correct_image, correct_labels)
correct_loss = bce(output, correct_labels)
# Wrong image and right caption
output, activations = self.disc(incorrect_image,
correct_labels)
incorrect_loss = bce(output, incorrect_labels)
#Generated image and right captions
noise = Variable(torch.random(self.batch_size, 100)).cuda()
noise = noise.view(self.batch_size, 100, 1, 1)
# Feeding it to the discriminator
generated_images = Variable(self.gen(noise,
correct_labels)).cuda()
output, activations = self.disc(generated_images,
correct_labels)
generated_loss = torch.mean(output)
# Calculating the net loss
net_loss = generated_loss + correct_loss + incorrect_loss
net_loss.backward()
# Taking one more step towards convergence
self.optimD.step()
# ----------------------------------------------------------------------------
#For generator
self.gen.zero_grad()
noise = Variable(torch.random(self.batch_size, 100)).cuda()
noise = noise.view(self.batch_size, 100, 1, 1)
generated_images = Variable(self.gen(noise,
correct_labels)).cuda()
output, generated = self.disc(generated_images, correct_labels)
output, real = self.disc(correct_image, correct_labels)
generated = torch.mean(generated, 0)
real = torch.mean(real, 0)
net_loss = bce(
output,
correct_labels) + mse(generated, real) * 100 + 50 * l1(
generated_images, correct_image)
net_loss.backward()
self.optimG.step()
def initHidden(self):
return torch.random(self.batch_size * self.input_size, self.h1), torch.random(self.h1, self.h2)
import torch
import xitorch as xt
@xt.module(shape=(25, 25))
def A(x, diag):
return x * diag
@A.set_precond
def precond(y, diag, biases=None):
return y / diag
@xt.module_like(A)
def AA(x, diag2):
return x * diag2 * diag2
class Aclass(xt.Module):
def forward(self, x, diag):
return x * diag
def precond(self, y, diag):
return y / diag
eigvals, eigvecs = xt.lsymeig(A, (diag, ), 3)
B = torch.random((nbatch, A.shape[1], 3))
c = xt.solve(A, (diag, ), B)
# # Mask a token that we will try to predict back with `BertForMaskedLM`
# masked_index = 8
# tokenized_text[masked_index] = '[MASK]'
# assert tokenized_text == ['[CLS]', 'who', 'was', 'jim', 'henson', '?', '[SEP]', 'jim', '[MASK]', 'was', 'a', 'puppet', '##eer', '[SEP]']
#
# # Convert token to vocabulary indices
# indexed_tokens = tokenizer.convert_tokens_to_ids(tokenized_text)
# # Define sentence A and B indices associated to 1st and 2nd sentences (see paper)
segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
input_mask = [1] * len(segments_ids)
#
# # Convert inputs to PyTorch tensors
tokens_tensor = torch.tensor([indexed_tokens])
segments_tensors = torch.tensor([segments_ids])
masks_tensors = torch.tensor([input_mask])
img = torch.random(1, 5, 768)
#
# model1 = BertModel.from_pretrained('bert-base-uncased')
# model2 = BertForPreTraining.from_pretrained('bert-base-uncased')
# model3 = BertForMaskedLM.from_pretrained('bert-base-uncased')
# model4 = BertForMultipleChoice.from_pretrained('bert-base-uncased', num_choices=10)
# model5 = BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=10)
# model6 = BertForTokenClassification.from_pretrained('bert-base-uncased', num_labels=10)
# model7 = BertForQuestionAnswering.from_pretrained('bert-base-uncased')
#
# tokens_tensor = tokens_tensor.to('cuda')
# segments_tensors = segments_tensors.to('cuda')
# masks_tensors = masks_tensors.to('cuda')
#
# # Predict hidden states features for each layer
# print(1)
from itertools import product
from math import prod
from multiprocessing import process
from multiprocessing.spawn import prepare
from torch import random
print("sdfgsdfg")
print()
prepare(product())
random
prepare
random()
print("asdf")
asdfa
asdf
