def main(config):
with SummaryWriter(
comment='_{}_{}'.format(config.arch, config.dataset)) as writer:
dataset_config = datasets.cifar10(
) if config.dataset == 'cifar10' else datasets.cifar100()
num_classes = dataset_config.pop('num_classes')
train_loader, eval_loader = create_data_loaders(**dataset_config,
config=config)
dummy_input = (torch.randn(10, 3, 32, 32), )
net = arch[config.arch](num_classes)
writer.add_graph(net, dummy_input)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
criterion = create_loss_fn(config)
if config.is_parallel:
net = torch.nn.DataParallel(net).to(device)
else:
device = 'cuda:{}'.format(
config.gpu) if torch.cuda.is_available() else 'cpu'
net = net.to(device)
optimizer = create_optim(net.parameters(), config)
scheduler = create_lr_scheduler(optimizer, config)
trainer = Trainer.PseudoLabel(net, optimizer, criterion, device,
config, writer)
trainer.loop(config.epochs,
train_loader,
eval_loader,
scheduler=scheduler,
print_freq=config.print_freq)
def train_mdn_with_proposal(save=True):
"""Use the prior proposal learnt by bootstrapping to train a brand new mdn."""
# load prior proposal and observations
_, obs_stats = helper.load(datadir + 'observed_data.pkl')
net, _, prior_proposal, _ = helper.load(netsdir + 'mdn_svi_proposal_prior_{0}.pkl'.format(n_bootstrap_iter-1))
n_inputs = n_percentiles
n_outputs = 3
n_samples = 5000
# generate data
ps = np.empty([n_samples, n_outputs])
stats = np.empty([n_samples, n_inputs])
for i in xrange(n_samples):
prior = 0.0
while prior < 0.5:
ps[i] = prior_proposal.gen()[0]
prior = eval_prior(*ps[i])
_, _, _, idts, _ = sim_likelihood(*ps[i])
stats[i] = calc_summary_stats(idts)
# train an mdn to give the posterior
minibatch = 100
maxiter = int(10000 * n_samples / minibatch)
monitor_every = 1000
net = mdn.replicate_gaussian_mdn(net, 8)
regularizer = lf.regularizerSvi(net.mps, net.sps, 0.1)
trainer = Trainer.Trainer(
model=net,
trn_data=[stats, ps],
trn_loss=net.mlprob + regularizer / n_samples,
trn_target=net.y
)
trainer.train(
maxiter=maxiter,
minibatch=minibatch,
show_progress=True,
monitor_every=monitor_every
)
# calculate the approximate posterior
mdn_mog = net.get_mog(obs_stats)
mdn_mog.prune_negligible_components(1.0e-6)
approx_posterior = mdn_mog / prior_proposal
# save the net
if save:
filename = netsdir + 'mdn_svi_proposal_hiddens_50_tanh_comps_8_sims_5k.pkl'
helper.save((net, approx_posterior), filename)
def run(self):
best_err1 = 100.
best_epoch = 0
logger.info('==> creating model "{}"'.format(args.model_name))
model = Util.getModel(**vars(args))
model = model.to(DEVICE)
# 大部分情况下，设置这个flag可以让内置的cuDNN的auto - tuner自动寻找最适合当前配置的高效算法，来达到优化运行效率的问题。
cudnn.benchmark = True
# define loss function (criterion) and pptimizer
# criterion = nn.CrossEntropyLoss().to(DEVICE)
# 标签平滑
criterion = LabelSmoothingLoss(classes=self.args.num_classes, smoothing=0.2)
# define optimizer
optimizer = Util.getOptimizer(model=model, args=self.args)
trainer = Trainer(dataset=self.dataset, criterion=criterion, optimizer=optimizer, args=self.args, logger=logger)
logger.info('train: {} test: {}'.format(self.dataset.get_train_length(), self.dataset.get_validation_length()))
for epoch in range(0, self.args.EPOCHS):
# train for one epoch
model = trainer.train(model=model, epoch=epoch)
# evaluate on validation set
model, val_loss, val_err1 = trainer.test(model=model, epoch=epoch)
# remember best err@1 and save checkpoint
is_best = val_err1 < best_err1
if is_best:
best_err1 = val_err1
best_epoch = epoch
logger.info('Best var_err1 {}'.format(best_err1))
Util.save_checkpoint(model.state_dict(), is_best, args.output_models_dir)
if not is_best and epoch - best_epoch >= args.patience > 0:
break
logger.info('Best val_err1: {:.4f} at epoch {}'.format(best_err1, best_epoch))
def train_mdn_with_proposal(save=True):
"""Use the prior proposal learnt by bootstrapping to train an mdn."""
# load prior proposal and observations
_, x, obs_data = helper.load(datadir + 'observed_data.pkl')
net, _, prior_proposal, _ = helper.load(
netsdir +
'mdn_svi_proposal_prior_{0}.pkl'.format(n_bootstrap_iter - 1))
n_inputs = n_data
n_outputs = n_dim
n_samples = 2000
# generate data
ws = np.empty([n_samples, n_outputs])
data = np.empty([n_samples, n_inputs])
for i in xrange(n_samples):
ws[i] = prior_proposal.gen()[0]
data[i] = gen_y_data(ws[i], x)
# train an mdn to give the posterior
minibatch = 100
maxiter = int(5000 * n_samples / minibatch)
monitor_every = 1000
regularizer = lf.regularizerSvi(net.mps, net.sps, 0.01)
trainer = Trainer.Trainer(model=net,
trn_data=[data, ws],
trn_loss=net.mlprob + regularizer / n_samples,
trn_target=net.y)
trainer.train(maxiter=maxiter,
minibatch=minibatch,
show_progress=True,
monitor_every=monitor_every)
# calculate the approximate posterior
mdn_mog = net.get_mog(obs_data)
approx_posterior = (mdn_mog * get_prior()) / prior_proposal
# save the net
if save:
filename = netsdir + 'mdn_svi_proposal_hiddens_50_tanh.pkl'
helper.save((net, approx_posterior), filename)
def train_mdn_on_sims_from_prior(save=True):
"""
Loads simulations done on parameters sampled from the prior, and trains an mdn on them.
"""
# read data
params, stats, _ = load_sims_from_prior()
n_data = 10 ** 5
params, stats = params[:n_data], stats[:n_data]
# split data into train and validation sets
trn_perc = 0.95
n_trn_data = int(trn_perc * n_data)
params_trn, stats_trn = params[:n_trn_data], stats[:n_trn_data]
params_val, stats_val = params[n_trn_data:], stats[n_trn_data:]
# train an mdn to give the posterior
n_components = 1
minibatch = 100
maxiter = int(1000 * n_data / minibatch)
monitor_every = 1000
net = mdn.MDN(n_inputs=9, n_hiddens=[50, 50], act_fun='tanh', n_outputs=4, n_components=n_components)
trainer = Trainer.Trainer(
model=net,
trn_data=[stats_trn, np.log(params_trn)],
trn_loss=net.mlprob,
trn_target=net.y,
val_data=[stats_val, np.log(params_val)],
val_loss=net.mlprob,
val_target=net.y
)
trainer.train(
maxiter=maxiter,
minibatch=minibatch,
show_progress=True,
monitor_every=monitor_every
)
# save the net
if save:
filename = netsdir + 'mdn_prior_hiddens_50_50_tanh_comps_1_sims_100k.pkl'
helper.save(net, filename)
def train_mdn_on_sims_from_prior(save=True):
"""
Loads simulations done on parameters sampled from the prior, and trains an mdn on them.
"""
# read data
ws, data, _ = load_sims_from_prior(n_files=1)
n_sims = 10**5
ws, data = ws[:n_sims], data[:n_sims]
# split data into train and validation sets
trn_perc = 0.95
n_trn_data = int(trn_perc * n_sims)
ws_trn, data_trn = ws[:n_trn_data], data[:n_trn_data]
ws_val, data_val = ws[n_trn_data:], data[n_trn_data:]
# train an mdn to give the posterior
minibatch = 100
maxiter = int(1000 * n_trn_data / minibatch)
monitor_every = 1000
net = mdn.MDN(n_inputs=data.shape[1],
n_hiddens=[50],
act_fun='tanh',
n_outputs=n_dim,
n_components=1)
trainer = Trainer.Trainer(model=net,
trn_data=[data_trn, ws_trn],
trn_loss=net.mlprob,
trn_target=net.y,
val_data=[data_val, ws_val],
val_loss=net.mlprob,
val_target=net.y)
trainer.train(maxiter=maxiter,
minibatch=minibatch,
show_progress=True,
monitor_every=monitor_every)
# save the net
if save:
filename = netsdir + 'mdn_prior_hiddens_50_tanh_sims_100k.pkl'
helper.save(net, filename)
model = DDPG(
'MlpPolicy',
env,
action_noise=action_noise,
verbose=1,
tensorboard_log="./h={}/".format(horizons[rank]),
gamma=0.99,
learning_rate=0.0003,
)
# model = DDPG.load("Model_DDPG_FS_30.zip")
# model.learning_rate = 0.0003
# model.gamma = 0.99
# action_noise = OrnsteinUhlenbeckActionNoise(mean=np.zeros(n_actions), sigma=0.05*np.ones(n_actions))
# action_noise = NormalActionNoise(mean=np.zeros(n_actions), sigma=0.075 * np.ones(n_actions))
# model.action_noise = action_noise
trainer = Trainer(env)
trainer.retrain_rl(model,
episodes=20000,
path="./h={}/".format(horizons[rank]))
# ## Training on horizon observations
# env = HorizonObservationWrapper(gym.make("reference_environment:reference-environment-v0"),
# horizon_length=horizons[rank],
# transform_name="Standard")
# trainer = Trainer(env)
# trainer.train_rl(models_to_train=1, episodes_per_model=20000, path='./h={}/'.format(horizons[rank]))
# ## Testing random action wrapper
# env = JoesActionWrapper(gym.make("reference_environment:reference-environment-v0"))
# trainer = Trainer(env)
# trainer.train_rl(models_to_train=1, episodes_per_model=20000)
def train_mdn_proposal_prior(save=True):
"""Trains an svi mdn to return the proposal prior with boostrapping."""
n_iterations = n_bootstrap_iter
n_samples = 200
true_w, x, y = helper.load(datadir + 'observed_data.pkl')
obs_data = y
# create an mdn
n_inputs = obs_data.size
net = mdn.MDN_SVI(n_inputs=n_inputs,
n_hiddens=[50],
act_fun='tanh',
n_outputs=n_dim,
n_components=1)
regularizer = lf.regularizerSvi(net.mps, net.sps, 0.01)
prior = get_prior()
prior_proposal = prior
for iter in xrange(n_iterations):
# generate new data
ws = np.empty([n_samples, n_dim])
data = np.empty([n_samples, n_inputs])
dist = np.empty(n_samples)
for i in xrange(n_samples):
w = prior_proposal.gen()[0]
y = gen_y_data(w, x)
this_data = y
ws[i] = w
data[i] = this_data
dist[i] = calc_dist(this_data, obs_data)
print 'simulation {0}, distance = {1}'.format(i, dist[i])
# plot distance histogram
fig = plt.figure()
ax = fig.add_subplot(111)
ax.hist(dist, bins=int(np.sqrt(n_samples)))
ax.set_title('iteration = {0}'.format(iter + 1))
ax.set_xlim([0.0, 20.0])
plt.show(block=False)
# train an mdn to give the posterior
minibatch = 50
maxiter = int(1000 * n_samples / minibatch)
monitor_every = 10
trainer = Trainer.Trainer(model=net,
trn_data=[data, ws],
trn_loss=net.mlprob +
regularizer / n_samples,
trn_target=net.y)
trainer.train(maxiter=maxiter,
minibatch=minibatch,
show_progress=True,
monitor_every=monitor_every)
# calculate the approximate posterior
mdn_mog = net.get_mog(obs_data, n_samples=None)
approx_posterior = (mdn_mog * prior) / prior_proposal
prior_proposal = approx_posterior.project_to_gaussian()
# save the net and the approximate posterior
if save:
helper.save(
(net, approx_posterior, prior_proposal, dist),
netsdir + 'mdn_svi_proposal_prior_{0}.pkl'.format(iter))
from models import ConvModel
dataset = KPIDataset(
'../data/train_preprocessed.csv',
seq_length=1001,
step_width=1
)
model = ConvModel(1001)
args = {
"lr": 0.5e-4,
"betas": (0.9, 0.999),
"eps": 1e-8,
"weight_decay": 0.0
}
trainer = Trainer(
model,
dataset,
batch_size=512,
epochs=100,
log_nth=800,
validation_size=0.2,
optim_args=args,
loss_func=CrossEntropyLoss()
)
trainer.train()
from util import Trainer
from torchvision import models
from torch.utils.data import random_split, DataLoader
device = "cuda:1"
dataset = CervixDataset(root='./',
csv_path="table_label_v2.csv",
transform=preprocess)
# print(len(dataset))
trainset, valset = random_split(dataset, [75, 23])
trainloader = DataLoader(trainset, shuffle=True)
valloader = DataLoader(valset, shuffle=True)
model = models.resnet50(pretrained=False)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters())
data_loader_dict = {'train': trainloader, 'val': valloader}
model.fc = torch.nn.Linear(in_features=2048, out_features=2, bias=True)
model = model.to(device)
trainer = Trainer(model=model,
dataloaders=data_loader_dict,
criterion=criterion,
optimizer=optimizer,
device=device,
filename="RESNET50_ADAM_Cervix.pth",
num_epochs=100)
best_model, val_acc_history = trainer.train_model()
def train_prior_proposal_with_bootstrapping(save=True):
"""Trains an svi mdn to return the posterior with boostrapping."""
n_samples = 400
true_ps, obs_stats = helper.load(datadir + 'observed_data.pkl')
# create an mdn
n_inputs = len(obs_stats)
n_outputs = len(true_ps)
net = mdn.MDN_SVI(n_inputs=n_inputs, n_hiddens=[50], act_fun='tanh', n_outputs=n_outputs, n_components=1)
regularizer = lf.regularizerSvi(net.mps, net.sps, 0.01)
prior_proposal = None
for iter in xrange(n_bootstrap_iter):
# generate new data
ps = np.empty([n_samples, n_outputs])
stats = np.empty([n_samples, n_inputs])
dist = np.empty(n_samples)
for i in xrange(n_samples):
prior = 0.0
while prior < 0.5:
ps[i] = sim_prior() if iter == 0 else prior_proposal.gen()[0]
prior = eval_prior(*ps[i])
_, _, _, idts, _ = sim_likelihood(*ps[i])
stats[i] = calc_summary_stats(idts)
dist[i] = calc_dist(stats[i], obs_stats)
print 'simulation {0}, distance = {1}'.format(i, dist[i])
# plot distance histogram
fig = plt.figure()
ax = fig.add_subplot(111)
ax.hist(dist, bins=int(np.sqrt(n_samples)))
ax.set_title('iteration = {0}'.format(iter + 1))
ax.set_xlim([0.0, 1.0])
plt.show(block=False)
# train an mdn to give the posterior
minibatch = 50
maxiter = int(1500 * n_samples / minibatch)
monitor_every = 10
trainer = Trainer.Trainer(
model=net,
trn_data=[stats, ps],
trn_loss=net.mlprob + regularizer / n_samples,
trn_target=net.y
)
trainer.train(
maxiter=maxiter,
minibatch=minibatch,
show_progress=True,
monitor_every=monitor_every
)
# calculate the approximate posterior
mdn_mog = net.get_mog(obs_stats, n_samples=None)
approx_posterior = mdn_mog if iter == 0 else mdn_mog / prior_proposal
prior_proposal = approx_posterior.project_to_gaussian()
# save the net and the approximate posterior
if save:
helper.save((net, approx_posterior, prior_proposal, dist), netsdir + 'mdn_svi_proposal_prior_{0}.pkl'.format(iter))
def train_mdn_proposal_prior(save=True):
"""
Train a proposal prior using bootstrapping.
"""
n_iterations = n_bootstrap_iter
n_data = 500
# read data
pilot_means, pilot_stds = helper.load(datadir + 'pilot_run_results.pkl')
obs_stats = helper.load(datadir + 'obs_stats.pkl')
obs_stats -= pilot_means
obs_stats /= pilot_stds
# create an mdn
net = mdn.MDN_SVI(n_inputs=9, n_hiddens=[50], act_fun='tanh', n_outputs=4, n_components=1)
regularizer = lf.regularizerSvi(net.mps, net.sps, 0.01)
prior_proposal = None
for iter in xrange(n_iterations):
# generate new data
params = []
stats = []
dist = []
i = 0
while i < n_data:
prop_params = sim_prior_params() if iter == 0 else np.exp(prior_proposal.gen())[0]
if np.any(np.log(prop_params) < log_prior_min) or np.any(np.log(prop_params) > log_prior_max):
continue
try:
lv = mjp.LotkaVolterra(init, prop_params)
states = lv.sim_time(dt, duration, max_n_steps=max_n_steps)
except mjp.SimTooLongException:
continue
sum_stats = calc_summary_stats(states)
sum_stats -= pilot_means
sum_stats /= pilot_stds
params.append(prop_params)
stats.append(sum_stats)
dist.append(calc_dist(sum_stats, obs_stats))
i += 1
print 'simulation {0}, distance = {1}'.format(i, dist[-1])
params = np.array(params)
stats = np.array(stats)
dist = np.array(dist)
# plot distance histogram
fig = plt.figure()
ax = fig.add_subplot(111)
ax.hist(dist, bins=int(np.sqrt(n_data)))
ax.set_title('iteration = {0}'.format(iter + 1))
ax.set_xlim([0.0, 12.0])
plt.show(block=False)
# train an mdn to give the posterior
minibatch = 100
maxiter = int(2000 * n_data / minibatch)
monitor_every = 100
trainer = Trainer.Trainer(
model=net,
trn_data=[stats, np.log(params)],
trn_loss=net.mlprob + regularizer / n_data,
trn_target=net.y
)
trainer.train(
maxiter=maxiter,
minibatch=minibatch,
show_progress=True,
monitor_every=monitor_every
)
# calculate the approximate posterior
mdn_mog = net.get_mog(obs_stats)
approx_posterior = mdn_mog if iter == 0 else mdn_mog / prior_proposal
prior_proposal = approx_posterior.project_to_gaussian()
# save the net and the approximate posterior
if save:
helper.save((net, approx_posterior, prior_proposal, dist), netsdir + 'mdn_svi_proposal_prior_{0}.pkl'.format(iter))
def train_mdn_with_proposal(save=True):
"""Use the prior proposal learnt by bootstrapping to train an mdn."""
# load prior proposal and observations
pilot_means, pilot_stds = helper.load(datadir + 'pilot_run_results.pkl')
obs_stats = helper.load(datadir + 'obs_stats.pkl')
obs_stats -= pilot_means
obs_stats /= pilot_stds
net, _, prior_proposal, _ = helper.load(netsdir + 'mdn_svi_proposal_prior_{0}.pkl'.format(n_bootstrap_iter-1))
n_samples = 2000
# generate data
params = []
stats = []
i = 0
while i < n_samples:
prop_params = np.exp(prior_proposal.gen())[0]
if np.any(np.log(prop_params) < log_prior_min) or np.any(np.log(prop_params) > log_prior_max):
continue
try:
lv = mjp.LotkaVolterra(init, prop_params)
states = lv.sim_time(dt, duration, max_n_steps=max_n_steps)
except mjp.SimTooLongException:
continue
sum_stats = calc_summary_stats(states)
sum_stats -= pilot_means
sum_stats /= pilot_stds
params.append(prop_params)
stats.append(sum_stats)
i += 1
params = np.array(params)
stats = np.array(stats)
# train an mdn to give the posterior
minibatch = 100
maxiter = int(5000 * n_samples / minibatch)
monitor_every = 1000
regularizer = lf.regularizerSvi(net.mps, net.sps, 0.01)
trainer = Trainer.Trainer(
model=net,
trn_data=[stats, np.log(params)],
trn_loss=net.mlprob + regularizer / n_samples,
trn_target=net.y
)
trainer.train(
maxiter=maxiter,
minibatch=minibatch,
show_progress=True,
monitor_every=monitor_every
)
# calculate the approximate posterior
mdn_mog = net.get_mog(obs_stats)
mdn_mog.prune_negligible_components(1.0e-3)
approx_posterior = mdn_mog / prior_proposal
# save the net
if save:
filename = netsdir + 'mdn_svi_proposal_hiddens_50_tanh_comps_1_sims_2k.pkl'
helper.save((net, approx_posterior), filename)
def main(config):
# SummaryWriter画图用的
with SummaryWriter(
comment='_{}_{}'.format(config.arch, config.dataset)) as writer:
# 选择datasets中的cifar10
dataset_config = datasets.FPN(
config) if config.dataset == 'FPN' else datasets.cifar10()
# dataset_config = datasets.cifar10() if config.dataset == 'cifar10' else datasets.cifar100()
# num_classes为类别数
# num_classes = dataset_config.pop('num_classes')
train_loader, eval_loader = data_loaders(**dataset_config,
config=config)
# torch.set_default_dtype(torch.float64)
dummy_input = torch.randn(1, 1, 200, 200) # 添加一个模型的图
#net = dw(FPN_ResNet18())
net = FPN_ResNet18()
# net = Fpn_n()
writer.add_graph(net, dummy_input)
###
# checkpoint = torch.load(config.PATH)
# net.load_state_dict(checkpoint['weight'])
###
device1 = 'cuda' if torch.cuda.is_available() else 'cpu'
criterion_l = create_loss_fn(config)
if config.is_parallel:
net = torch.nn.DataParallel(net).to(device1)
else:
device1 = 'cuda:{}'.format(
config.gpu) if torch.cuda.is_available() else 'cpu'
net = net.to(device1)
optimizer = create_optim(net.parameters(), config)
if config.train:
trainer = Trainer.PseudoLabel(net,
optimizer,
criterion_l,
device1,
config,
writer,
save_dir='./model')
scheduler = create_lr_scheduler(optimizer, config)
trainer.loop(config.epochs,
train_loader,
eval_loader,
scheduler=scheduler,
print_freq=config.print_freq)
else:
checkpoint = torch.load(config.PATH)
net.load_state_dict(checkpoint['weight'])
trainer = Trainer.PseudoLabel(net,
optimizer,
criterion_l,
device1,
config,
writer,
save_dir='./model')
trainer.testonce(eval_loader, print_freq=config.print_freq)
# class dw(torch.nn.Module):
# '''
# '''
#
# def __init__(self, model):
# '''
# '''
#
# # initialize the module using super() constructor
# super(dw, self).__init__()
# # assign the architectures
# self.model = model
# # assign the weights for each task
# self.weights = torch.nn.Parameter(torch.ones(2).float())
#
# def forward(self, x):
# out = self.model(x)
# return out
from stable_baselines3.common.callbacks import EvalCallback
env_action = RelativeActionWrapper(gym.make("reference_environment:reference-environment-v0"))
env_horizon = HorizonObservationWrapper(env_action,
horizon_length=4,
transform_name="Deltas")
env = PhaseRewardWrapper(env_horizon, phase="Full") # Set Phase to Full
eval_callback = EvalCallback(env, best_model_save_path='./logs/',
log_path='./logs/', eval_freq=500,
deterministic=True, render=False)
### DDPG Noise
### Try increasing the noise when retraining.
### Try less noise based on the policy plot.
n_actions = env.action_space.shape[-1]
action_noise = OrnsteinUhlenbeckActionNoise(mean=np.zeros(n_actions), sigma=1 * np.ones(n_actions))
# action_noise = NormalActionNoise(mean=np.zeros(n_actions), sigma=0.1 * np.ones(n_actions))
model = DDPG('MlpPolicy', env, action_noise=action_noise, verbose=1, tensorboard_log="./logs",
gamma=0.99,
learning_rate=0.0003,
)
# model = DDPG.load("Model_DDPG_FS_30.zip")
# model.learning_rate = 0.0003
# model.gamma = 0.99
# action_noise = OrnsteinUhlenbeckActionNoise(mean=np.zeros(n_actions), sigma=0.05*np.ones(n_actions))
# action_noise = NormalActionNoise(mean=np.zeros(n_actions), sigma=0.075 * np.ones(n_actions))
# model.action_noise = action_noise
trainer = Trainer(env)
trainer.retrain_rl(model, episodes=20000)
# Import resnet model
import torch
resnet18 = models.resnet18()
resnet18.fc = torch.nn.Linear(512, 10)
resnet18.eval()
# To see the structure of network, uncomment this line
# print(resnet18)
import time
import copy
device = "cuda:1"
from custom_model.alexresnet import alexresnet
model = alexresnet()
model.to(device)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
data_loader_dict = {'train': trainloader, 'val': valloader}
from util import Trainer
trainer = Trainer(model=model,
dataloaders=data_loader_dict,
criterion=criterion,
optimizer=optimizer)
best_model, val_acc_history = trainer.train_model()