def test(model, device, test_loader, images, texts, lengths, converter,
prediction_dir):
model.to(device)
images = images.to(device)
model.eval()
pred_json = {}
pred_list = []
make_folder(prediction_dir)
for i, datas in enumerate(test_loader):
datas, targets = datas
batch_size = datas.size(0)
dataloader.loadData(images, datas)
t, l = converter.encode(targets)
dataloader.loadData(texts, t)
dataloader.loadData(lengths, l)
preds = model(images)
preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size))
_, preds = preds.max(2)
preds = preds.transpose(1, 0).contiguous().view(-1)
pred_string = converter.decode(preds.data, preds_size.data, raw=False)
pred_dict = {
'image_path': test_loader.dataset.get_img_path(i),
'prediction': pred_string
}
pred_list.append(pred_dict)
pred_json = {'predict': pred_list}
with open(os.path.join(prediction_dir, 'predict.json'), 'w') as save_json:
json.dump(pred_json, save_json, indent=2, ensure_ascii=False)
def get_model_results_for_folder(model: torch.nn.Module,
folder: str,
device: Union[torch.device, str] = None,
pbar_desc: str = None) -> pd.DataFrame:
"""Collect model float prediction for all image files in folder.
The model must return a 2D tensor of size (batch_size, binary predictions).
All non-directory files ending with '.png' in folder are assumed to be valid image files
loadable by PIL.Image.open.
:param model: the model to use
:param device: if given, the device to move the model onto before evaluation
:param folder: the folder to search for image files in
:param pbar_desc: description for the progress bar
:return: pd.DataFrame with columns 'img' (the file name of the image relative to the folder),
and 'pred' (the float sigmoid of the prediction of the model).
"""
with torch.no_grad():
model.eval()
if device is not None:
model.to(device)
img_fns = [
fn for fn in os.listdir(folder)
if os.path.isfile(os.path.join(folder, fn)) and fn.endswith('.png')
]
row_list = []
for img_fn in tqdm(img_fns, desc=pbar_desc): # TODO: batch-processing
img = PIL.Image.open(os.path.join(folder, img_fn))
img_t = to_tens(img).to(device)
pred_t = torch.sigmoid(model(img_t.unsqueeze(0)).squeeze(0))
row_list.append({'img': img_fn, 'pred': float(pred_t)})
return pd.DataFrame(row_list)
def __init__(self,loader_test,loader_train,dataprep,file_R,cuda,device,model,crop=True):
self.loader_test = loader_test
self.loader_train=loader_train
self.root_M = file_R
self.dataprep=dataprep
self.crop = crop
self.fileC = torch.load(self.root_M,map_location='cpu')
model.load_state_dict(self.fileC['model_state_dict'])
self.losses_epoc = self.fileC['m_los']
self.psnr = self.fileC['psnr']
self.ssim = self.fileC['ssim']
self.best_psnr =0
self.best_t = []
self.best_d = []
self.best_s = []
if cuda:
model.to(device)
# from torchsummary import summary
# summary(model,input_size=(1,256,64,64))
self.model = model
self.device = device
self.loss = nn.MSELoss()
self.cuda = cuda
self.test_all()
if self.crop:
self.reconstruct()
def test(device, testset, testloader, model):
model.to(device)
model.eval()
acc = []
with tqdm(testloader, total=config.num_batches) as pbar:
for batch_idx, batch in enumerate(pbar):
train_inputs, train_targets = batch['train']
train_inputs = train_inputs.to(device=device)
train_targets = train_targets.to(device=device)
train_embeddings = model(train_inputs)
test_inputs, test_targets = batch['test']
test_inputs = test_inputs.to(device=device)
test_targets = test_targets.to(device=device)
test_embeddings = model(test_inputs)
prototypes = get_prototypes(train_embeddings, train_targets,
testset.num_classes_per_task)
with torch.no_grad():
accuracy = get_accuracy(prototypes, test_embeddings,
test_targets)
pbar.set_postfix(accuracy='{0:.4f}'.format(accuracy.item()))
acc.append(accuracy)
if batch_idx >= config.num_batches:
break
return acc
def run(args, trainset, testset, action):
if not torch.cuda.is_available():
args.device = 'cpu'
args.device = torch.device(args.device)
model = action.create_model()
if args.store and os.path.isfile(args.store):
model.load_state_dict(torch.load(args.store, map_location='cpu'))
if args.pretrained:
assert os.path.isfile(args.pretrained)
model.load_state_dict(torch.load(args.pretrained, map_location='cpu'))
model.to(args.device)
checkpoint = None
if args.resume:
assert os.path.isfile(args.resume)
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['model'])
# dataloader
testloader = torch.utils.data.DataLoader(testset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers)
# optimizer
min_loss = float('inf')
learnable_params = filter(lambda p: p.requires_grad, model.parameters())
if args.optimizer == 'Adam':
optimizer = torch.optim.Adam(learnable_params)
else:
optimizer = torch.optim.SGD(learnable_params, lr=0.1)
if checkpoint is not None:
min_loss = checkpoint['min_loss']
optimizer.load_state_dict(checkpoint['optimizer'])
# training
LOGGER.debug('train, begin')
for epoch in range(args.start_epoch, args.epochs):
running_loss = action.train(model, trainloader, optimizer, args.device)
val_loss = action.validate(model, testloader, args.device)
is_best = val_loss < min_loss
min_loss = min(val_loss, min_loss)
LOGGER.info('epoch, %04d, %f, %f', epoch + 1, running_loss, val_loss)
print('epoch, %04d, floss_train=%f, floss_val=%f' %
(epoch + 1, running_loss, val_loss))
if is_best:
save_checkpoint(model.state_dict(), args.outfile, 'model')
LOGGER.debug('train, end')
def evaluate(data_source, ngramProb=None):
# Turn on evaluation mode which disables dropout.
model.to(device)
model.eval()
model.set_mode('eval')
total_loss = 0.
stout = []
ntokens = len(dictionary)
hidden = model.init_hidden(eval_batch_size)
with torch.no_grad():
for i in range(0, data_source.size(0) - 1, args.bptt):
data, targets = get_batch(data_source, i)
# import pdb; pdb.set_trace()
if ngramProb != None:
_, batch_ngramProb = get_batch(ngramProb, i)
# gs534 add sentence resetting
eosidx = dictionary.get_eos()
output, hidden = model(data, hidden, separate=args.reset, eosidx=eosidx)
output_flat = output.view(-1, ntokens)
logProb = interpCrit(output.view(-1, ntokens), targets)
rnnProbs = torch.exp(-logProb)
if args.interp and args.evalmode:
final_prob = args.factor * rnnProbs + (1 - args.factor) * batch_ngramProb
else:
final_prob = rnnProbs
if args.stream_out:
stout += final_prob.tolist()
total_loss += (-torch.log(final_prob).sum()) / data.size(1)
hidden = repackage_hidden(hidden)
return total_loss / len(data_source), stout
def train(num_epochs, model, device, train_loader, val_loader, images, texts, lengths, converter, optimizer, lr_scheduler, prediction_dir, print_iter, opt) :
# criterion = CTCLoss()
# criterion.to(device)
criterion = torch.nn.CrossEntropyLoss(ignore_index = 0).to(device)
images = images.to(device)
model.to(device)
for epoch in range(num_epochs) :
count = 0
model.train()
for i, datas in enumerate(train_loader) :
datas, targets = datas
batch_size = datas.size(0)
count += batch_size
dataloader.loadData(images, datas)
t, l = converter.encode(targets, opt.batch_max_length)
dataloader.loadData(texts, t)
dataloader.loadData(lengths, l)
preds = model(images, t[:, :-1])
# preds_size = Variable(torch.IntTensor([preds.size(0)] * batch_size))
cost = criterion(preds.view(-1, preds.shape[-1]), t[:, 1:].contiguous().view(-1))
model.zero_grad()
cost.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), opt.grad_clip)
optimizer.step()
if count % print_iter < train_loader.batch_size :
print('epoch {} [{}/{}]loss : {}'.format(epoch, count, len(train_loader.dataset), cost))
if (epoch %3==0) &(epoch !=0 ):
res = validation(model, device, val_loader, images, texts, lengths, converter, prediction_dir, opt)
save_model(opt.save_dir, f'{epoch}_{round(float(res),3)}', model, optimizer, lr_scheduler, opt)
def test(model, data_loader_test, device, predict_path) :
class_name = {1 :'sidewalk_blocks' , 2 : 'alley_damaged', 3 :'sidewalk_damaged', 4 : 'caution_zone_manhole', 5 : 'braille_guide_blocks_damaged',
6 : 'alley_speed_bump', 7 : 'roadway_crosswalk', 8 : 'sidewalk_urethane', 9 : 'caution_zone_repair_zone', 10 : 'sidewalk_asphalt',
11 : 'sidewalk_other', 12 : 'alley_crosswalk', 13 : 'caution_zone_tree_zone', 14 : 'caution_zone_grating', 15 : 'roadway_normal',
16 : 'bike_lane', 17 : 'caution_zone_stairs', 18 : 'alley_normal', 19 : 'sidewalk_cement', 20 : 'braille_guide_blocks_normal',
21 : 'sidewalk_soil_stone'}
model.to(device)
model.eval()
try :
os.mkdir(os.path.join('./predictions'))
except :
pass
# 이미지 전체 반복
pred_xml = elemTree.Element('predictions')
pred_xml.text = '\n '
for idx, data in enumerate(data_loader_test) :
print('{} / {}'.format(idx+1, len(data_loader_test)))
images, target = data
images = list(image.to(device) for image in images)
outputs = model(images)
output = outputs[0]
masks, labels, scores = output['masks'], output['labels'], output['scores']
texts = []
# 이미지 한장에 대하여
xml_image = elemTree.SubElement(pred_xml, 'image')
xml_image.attrib['name'] = target[0]['image_id'].split('.')[0]
xml_image.text = '\n '
for index in range(len(masks)) :
mask, label, score = masks[index], int(labels[index]), scores[index]
# class, score, x1, y1, x2, y2
mask_arr = mask[0].cpu().detach().numpy()
mask_bin = np.where(mask_arr > 0.3, True, False)
polygons = Mask(mask_bin).polygons()
points = polygons.points
point = ''
for p in points[0]:
point += str(p[0]) + ',' + str(p[1]) +';'
xml_predict = elemTree.SubElement(xml_image, 'predict')
xml_predict.tail = '\n '
xml_predict.attrib['class_name'] = class_name[label]
xml_predict.attrib['score'] = str(float(score))
xml_predict.attrib['polygon'] = point
if index == len(masks) - 1 :
xml_predict.tail = '\n '
xml_image.tail = '\n '
if idx == len(data_loader_test) - 1:
xml_image.tail = '\n'
pred_xml = elemTree.ElementTree(pred_xml)
pred_xml.write('./predictions/'+ predict_path + '.xml')
def evaluate(model, testset, device):
# eval on valid set
loss_val_, acc_val_ = 0., 0.
model.eval()
model.to(device)
for j, sample in enumerate(testset):
img = sample['img'].to(device)
score = sample['score'].to(device)
score_pred = model(img)
loss_val_ += bce_loss(score_pred, score).item()
acc_val_ += accuracy(score_pred, score)
avg_loss, avg_acc = loss_val_/(j+1), acc_val_/(j+1)
return avg_loss, avg_acc
def train(model,
trainloader,
testloader,
criterion,
optimizer,
epochs=10,
print_every=40,
print_accuracy=False,
device='cpu'):
steps = 0
# turn on drop out in the network (default)
model.train()
# change to device
model.to(device)
for e in range(epochs):
running_loss = 0
for ii, (inputs, labels) in enumerate(trainloader):
steps += 1
inputs, labels = inputs.to(device), labels.to(device)
optimizer.zero_grad()
# Forward and backward passes
outputs = model.forward(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
running_loss += loss.item()
if steps % print_every == 0:
#print("Epoch: {}/{}... ".format(e+1, epochs),
# "Loss: {:.4f}".format(running_loss/print_every))
#running_loss = 0
if print_accuracy:
train_accuracy = accuracy_score(model,
trainloader,
device=device,
print_score=False)
test_accuracy = accuracy_score(model,
testloader,
device=device,
print_score=False)
print(
"Epoch: {}/{}... ".format(e + 1, epochs),
"Loss: {:.4f}".format(running_loss / print_every),
"Train accuracy: {} %".format(f'{train_accuracy:.1f}'),
"Test accuracy: {} %".format(f'{test_accuracy:.1f}'))
else:
print("Epoch: {}/{}... ".format(e + 1, epochs),
"Loss: {:.4f}".format(running_loss / print_every))
running_loss = 0
def load_checkpoint(opts, model, label):
""" Load a model from a file. """
cur_dir = os.getcwd()
model_dir = os.path.join(cur_dir, opts.results_dir, opts.experiment_name,
'checkpoints')
model_file = os.path.join(model_dir, '{}_net.pth.tar'.format(label))
print("Loading model from {}".format(model_file))
model_dict = torch.load(model_file, map_location=opts.device)
device = torch.device(opts.device)
model.to(opts.device)
model.load_state_dict(model_dict['state_dict'])
return model
def __init__(self, dataset, model, model_state_file, save_dir, device, shuffle,
num_epochs, batch_size, learning_rate, early_stopping_criteria):
self.dataset = dataset
self.class_weights = dataset.class_weights.to(device)
self.model = model.to(device)
self.save_dir = save_dir
self.device = device
self.shuffle = shuffle
self.num_epochs = num_epochs
self.batch_size = batch_size
self.loss_func = nn.CrossEntropyLoss(self.class_weights)
self.optimizer = optim.Adam(self.model.parameters(), lr=learning_rate)
self.scheduler = optim.lr_scheduler.ReduceLROnPlateau(
optimizer=self.optimizer, mode='min', factor=0.5, patience=1)
self.train_state = {
'done_training': False,
'stop_early': False,
'early_stopping_step': 0,
'early_stopping_best_val': 1e8,
'early_stopping_criteria': early_stopping_criteria,
'learning_rate': learning_rate,
'epoch_index': 0,
'train_loss': [],
'train_acc': [],
'val_loss': [],
'val_acc': [],
'test_loss': -1,
'test_acc': -1,
'model_filename': model_state_file}
def load_model(model, checkpoint, device):
ckpt = torch.load(os.path.join('checkpoint', checkpoint))
if 'args' in ckpt:
args = ckpt['args']
if model == 'vqvae':
model = VQVAE()
elif model == 'pixelsnail_bottom':
model = PixelSNAIL(
[64, 64],
512,
args.channel,
5,
4,
args.n_res_block,
args.n_res_channel,
attention=False,
dropout=args.dropout,
n_cond_res_block=args.n_cond_res_block,
cond_res_channel=args.n_res_channel,
)
if 'model' in ckpt:
ckpt = ckpt['model']
model.load_state_dict(ckpt)
model = model.to(device)
model.eval()
return model
def train(model, optimizer, epochs, print_every=1000, plot_every=20):
scheduler = ReduceLROnPlateau(optimizer=optimizer,
mode='max',
factor=0.1,
patience=10,
verbose=True)
model = model.to(device=device)
for e in range(epochs):
for t, (x, y) in enumerate(loader_train):
model.train()
x = x.to(device=device, dtype=dtype)
y = y.to(device=device, dtype=torch.long)
scores = model(x)
_, preds = scores.max(1)
loss = F.cross_entropy(scores, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if t % print_every == 0:
print('Epoch %d, Iteration %d, loss = %.4f' % (e, t, loss))
val_acc = check_accuracy(model)
scheduler.step(val_acc)
print()
if e % plot_every == 0 and e > 0:
plt.plot(np.arange(len(a)), a, 'b', np.arange(len(b)), b, 'y')
plt.ylim(0, 1)
# plt.show()
plt.savefig("./logs/Epoch %d.png" % e)
def train(device, dataset, dataloader, model):
print("in train")
model = model.to(device)
model.train()
optimizer = torch.optim.Adam(model.parameters(), lr=config.learning_rate)
# Training loop
images_per_batch = {}
batch_count, images_per_batch['train'], images_per_batch[
'test'] = 0, [], []
with tqdm(dataloader, total=config.num_batches) as pbar:
for batch_idx, batch in enumerate(pbar):
model.zero_grad()
train_inputs, train_targets = batch['train']
train_inputs = train_inputs.to(device=device)
train_targets = train_targets.to(device=device)
train_embeddings = model(train_inputs)
test_inputs, test_targets = batch['test']
test_inputs = test_inputs.to(device=device)
test_targets = test_targets.to(device=device)
test_embeddings = model(test_inputs)
prototypes = get_prototypes(train_embeddings, train_targets,
dataset.num_classes_per_task)
loss = prototypical_loss(prototypes, test_embeddings, test_targets)
loss.backward()
optimizer.step()
#Just keeping the count here
batch_count += 1
images_per_batch['train'].append(train_inputs.shape[1])
images_per_batch['test'].append(test_inputs.shape[1])
with torch.no_grad():
accuracy = get_accuracy(prototypes, test_embeddings,
test_targets)
pbar.set_postfix(accuracy='{0:.4f}'.format(accuracy.item()))
if batch_idx >= config.num_batches:
break
print("Number of batches in the dataloader: ", batch_count)
# Save model
if check_dir() is not None:
filename = os.path.join(
'saved_models',
'protonet_cifar_fs_{0}shot_{1}way.pt'.format(config.k, config.n))
with open(filename, 'wb') as f:
state_dict = model.state_dict()
torch.save(state_dict, f)
print("Model saved")
return batch_count, images_per_batch
def __init__(self, model, optimizer):
self.model = model.to(DEVICE)
self.optimizer = optimizer
self.epoch = 0
self.iterations = 0
self.accuracies = np.array([])
self.losses = np.array([])
self.eval_accuracies = np.array([])
self.log_softmax = nn.LogSoftmax(dim=1).to(DEVICE)
def inference(model, data_loader, **kwargs):
dev = kwargs['device']
if dev == 'cuda':
args.cpu = False
args.ngraph = ''
# from trainer import Trainer
model = model.to(torch.device('cpu' if args.cpu else 'cuda'))
t = Trainer(args, data_loader, model, loss, checkpoint)
metric = t.test()
if dev == 'cpu':
args.cpu = True
# args.ngraph = ''
args.ngraph = './edsr.model'
# from trainer import Trainer
model = model.to(torch.device('cpu' if args.cpu else 'cuda'))
t = Trainer(args, data_loader, model, loss, checkpoint)
metric = t.test()
if metric == 'nan':
metric = 0
return metric
def run(args, testset, action):
if not torch.cuda.is_available():
args.device = 'cpu'
args.device = torch.device(args.device)
model = action.create_model()
if args.pretrained:
assert os.path.isfile(args.pretrained)
model.load_state_dict(torch.load(args.pretrained, map_location='cpu'))
model.to(args.device)
# dataloader
testloader = torch.utils.data.DataLoader(testset,
batch_size=1,
shuffle=False,
num_workers=args.workers)
# testing
LOGGER.debug('tests, begin')
action.evaluate(model, testloader, args.device)
LOGGER.debug('tests, end')
def print_examples(model, device, dataset):
transform = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((299, 299)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
model.to(device)
model.eval()
test_img1 = transform(
cv2.imread("../Data/flickr8k/test_examples/dog.jpg")).unsqueeze(0)
print("Example 1 CORRECT: Dog on a beach by the ocean")
print("Example 1 OUTPUT: " +
" ".join(model.caption_image(test_img1.to(device), dataset.vocab)))
print('=' * 50)
test_img2 = transform(
cv2.imread("../Data/flickr8k/test_examples/child.jpg")).unsqueeze(0)
print("Example 2 CORRECT: Child holding red frisbee outdoors")
print("Example 2 OUTPUT: " +
" ".join(model.caption_image(test_img2.to(device), dataset.vocab)))
print('=' * 50)
test_img3 = transform(
cv2.imread("../Data/flickr8k/test_examples/bus.png")).unsqueeze(0)
print("Example 3 CORRECT: Bus driving by parked cars")
print("Example 3 OUTPUT: " +
" ".join(model.caption_image(test_img3.to(device), dataset.vocab)))
print('=' * 50)
test_img4 = transform(
cv2.imread("../Data/flickr8k/test_examples/boat.png")).unsqueeze(0)
print("Example 4 CORRECT: A small boat in the ocean")
print("Example 4 OUTPUT: " +
" ".join(model.caption_image(test_img4.to(device), dataset.vocab)))
print('=' * 50)
test_img5 = transform(
cv2.imread("../Data/flickr8k/test_examples/horse.png")).unsqueeze(0)
print("Example 5 CORRECT: A cowboy riding a horse in the desert")
print("Example 5 OUTPUT: " +
" ".join(model.caption_image(test_img5.to(device), dataset.vocab)))
print('=' * 50)
model.train()
def train(num_epochs, model, device, train_loader, val_loader, images, texts,
lengths, converter, optimizer, lr_scheduler, prediction_dir,
print_iter):
criterion = CTCLoss()
criterion.to(device)
images = images.to(device)
model.to(device)
for epoch in range(num_epochs):
print(epoch)
count = 0
model.train()
for i, datas in enumerate(train_loader):
datas, targets = datas
batch_size = datas.size(0)
count += batch_size
dataloader.loadData(images, datas)
t, l = converter.encode(targets)
dataloader.loadData(texts, t)
dataloader.loadData(lengths, l)
preds = model(images)
preds_size = Variable(torch.IntTensor([preds.size(0)] *
batch_size))
cost = criterion(preds, texts, preds_size, lengths) / batch_size
model.zero_grad()
cost.backward()
optimizer.step()
if count % print_iter < train_loader.batch_size:
print('epoch {} [{}/{}]loss : {}'.format(
epoch, count, len(train_loader.dataset), cost))
validation(model, device, val_loader, images, texts, lengths,
converter, prediction_dir)
save_model('{}'.format(epoch), model, optimizer, lr_scheduler)
lr_scheduler.step()
ntokens = len(corpus.dictionary)
if args.load is None:
model = model.RNNModel(args.model,
ntokens,
args.emsize,
args.nhid,
args.nlayers,
args.dropout,
args.tied,
corpus=corpus,
embeddings=embeddings)
else:
with open(args.load, 'rb') as f:
model = torch.load(f)
model = model.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()),
lr=args.lr)
###############################################################################
# Training code
###############################################################################
def repackage_hidden(h):
"""Wraps hidden states in new Variables, to detach them from their history."""
if type(h) == Variable:
return Variable(h.data)
else:
# This Preprocess internally Triage class
# This will split data and encode using passing tokenizer
# Creating instance of the class
Preprocess = prepare_data.Preprocess(dataframe=df_new_reduced,
tokenizer=bert_tokenizer,
max_len=config.MAX_LEN,
train_batch_size=config.TRAIN_BATCH_SIZE,
valid_batch_size=config.VALID_BATCH_SIZE,
test_batch_size=config.TEST_BATCH_SIZE)
# Accessing the process_data_for_model method of Preprocess class
testing_loader = Preprocess.process_data_for_test()
#################################################################################
model = model.DistillBERTClass() # Creating the model shape
model.to(device)
# Loading back the model from checkpoint
# checkpoint = torch.load(config.checkpoint_path, map_location=device) # Loading the model from check point
# model.load_state_dict(checkpoint['model_state_dict'])
model = torch.load(config.checkpoint_path, map_location=device)
model.eval()
model.to(device) # Loading model to GPU
# Validation on test data
# Creating the loss function
# Optimizer is not needed since its for prediction
loss_function = torch.nn.CrossEntropyLoss()
test_loss, test_accu, y_test_actual, y_test_predicted, y_test_predicted_prob_list = valid(
model=model, testing_loader=testing_loader, loss_fn=loss_function)
optimizer.param_groups[0]['lr'] = args.lr
model.dropouti, model.dropouth, model.dropout, args.dropoute = (
args.dropouti, args.dropouth, args.dropout, args.dropoute)
if args.wdrop:
from weight_drop import WeightDrop
for rnn in model.rnns:
if type(rnn) == WeightDrop:
rnn.dropout = args.wdrop
elif rnn.zoneout > 0:
rnn.zoneout = args.wdrop
###
if not criterion:
criterion = torch.nn.CrossEntropyLoss()
###
if args.cuda:
model = model.to('cuda')
###
params = list(model.parameters()) + list(criterion.parameters())
total_params = sum(x.size()[0] *
x.size()[1] if len(x.size()) > 1 else x.size()[0]
for x in params if x.size())
print('Args:', args)
print('Model total parameters:', total_params)
if args.torchscript:
print("Scripting the module...")
model = torch.jit.script(model)
if args.trace:
example_input = torch.randn(args.bptt, args.batch_size).to('cuda')
print('data.size():{}\n'.format(example_input.size()))
def main(model, path):
print(path)
t1 = time.time()
checkpoint_folder = "Model_Checkpoints"
project_path = os.getcwd()
save_path = os.path.join(project_path, checkpoint_folder)
if not os.path.exists(checkpoint_folder):
os.makedirs(checkpoint_folder)
else:
shutil.rmtree(save_path)
os.makedirs(checkpoint_folder)
in_features = 300
hidden_size = 256
layer_num = 2
print("\n")
print(" Loading Data ... ")
print("="*30)
print("\n")
train_dl, valid_dl, trn, vld = dataloader.train_val_loader(path)
print(" Got train_dataloader and validation_dataloader ")
print("="*30)
print("\n")
print(" Loading LSTM Model ...")
print("="*30)
print("\n")
model = model.Rnn_Lstm(in_features, hidden_size, layer_num, 391)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-2)
criterion = nn.BCEWithLogitsLoss()
epochs = 10
print(" Training started ... ")
print("="*30)
print("\n")
for epoch in range(1, epochs + 1):
checkpoint_name = "checkpoint_"+ str(epoch) +".pth"
checkpoint_save_path = os.path.join(save_path, checkpoint_name)
running_loss = 0.0
model.train() # turn on training mode
for x, y in tqdm.tqdm(train_dl):
x, y = x.to(device), y.to(device)
optimizer.zero_grad()
preds = model(x)
loss = criterion(preds, y)
loss.backward()
optimizer.step()
running_loss += loss.item() * x.size(0)
epoch_loss = running_loss / len(trn)
# calculate the validation loss for this epoch
val_loss = 0.0
model.eval() # turn on evaluation mode
for x, y in valid_dl:
x, y = x.to(device), y.to(device)
preds = model(x)
loss = criterion(preds, y)
val_loss += loss.item() * x.size(0)
val_loss /= len(vld)
print('Epoch: {}, Training Loss: {:.4f}, Validation Loss: {:.4f} \n'.format(epoch, epoch_loss, val_loss))
print("Checkpoint saved after {} epoch\n".format(epoch))
torch.save(model.state_dict(), checkpoint_save_path)
print("Training completed -> Finished -- {} \n".format(time.time()-t1))
print("="*30)
print("\n")
def train_model(model, trainds, testds, config, device, writer=None):
batch_size = config['data']['batch_size']
status = config['training']['status']
epochs = config['training']['epochs']
balanced_loss = config['loss']['balanced']
# nval = config['nval']
nval_tests = config['nval_tests']
nsave = config['nsave']
model_save = config['model_save']
rank = config['rank']
nranks = config['nranks']
hvd = config['hvd']
num_classes = config['data']['num_classes']
## create samplers for these datasets
train_sampler = torch.utils.data.distributed.DistributedSampler(
trainds, nranks, rank, shuffle=True, drop_last=True)
test_sampler = torch.utils.data.distributed.DistributedSampler(
testds, nranks, rank, shuffle=True, drop_last=True)
## create data loaders
train_loader = torch.utils.data.DataLoader(
trainds,
shuffle=False,
sampler=train_sampler,
num_workers=config['data']['num_parallel_readers'],
batch_size=batch_size,
persistent_workers=True)
test_loader = torch.utils.data.DataLoader(
testds,
shuffle=False,
sampler=test_sampler,
num_workers=config['data']['num_parallel_readers'],
batch_size=batch_size,
persistent_workers=True)
loss_func = loss.get_loss(config)
ave_loss = CalcMean.CalcMean()
acc_func = accuracy.get_accuracy(config)
ave_acc = CalcMean.CalcMean()
opt_func = optimizer.get_optimizer(config)
opt = opt_func(model.parameters(), **config['optimizer']['args'])
lrsched_func = optimizer.get_learning_rate_scheduler(config)
lrsched = lrsched_func(opt, **config['lr_schedule']['args'])
# Add Horovod Distributed Optimizer
if hvd:
opt = hvd.DistributedOptimizer(
opt, named_parameters=model.named_parameters())
# Broadcast parameters from rank 0 to all other processes.
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
model.to(device)
for epoch in range(epochs):
logger.info(' epoch %s of %s', epoch, epochs)
train_sampler.set_epoch(epoch)
test_sampler.set_epoch(epoch)
model.to(device)
for batch_counter, (inputs, targets, class_weights,
nonzero_mask) in enumerate(train_loader):
# move data to device
inputs = inputs.to(device)
targets = targets.to(device)
class_weights = class_weights.to(device)
nonzero_mask = nonzero_mask.to(device)
# zero grads
opt.zero_grad()
outputs, endpoints = model(inputs)
# set the weights
if balanced_loss:
weights = class_weights
nonzero_to_class_scaler = torch.sum(
nonzero_mask.type(torch.float32)) / torch.sum(
class_weights.type(torch.float32))
else:
weights = nonzero_mask
nonzero_to_class_scaler = torch.ones(1, device=device)
loss_value = loss_func(outputs, targets.long())
loss_value = torch.mean(
loss_value * weights) * nonzero_to_class_scaler
# backward calc grads
loss_value.backward()
# apply grads
opt.step()
ave_loss.add_value(float(loss_value.to('cpu')))
# calc acc
ave_acc.add_value(
float(acc_func(outputs, targets, weights).to('cpu')))
# print statistics
if batch_counter % status == 0:
logger.info(
'<[%3d of %3d, %5d of %5d]> train loss: %6.4f acc: %6.4f',
epoch + 1, epochs, batch_counter,
len(trainds) / nranks / batch_size, ave_loss.mean(),
ave_acc.mean())
if writer and rank == 0:
global_batch = epoch * len(
trainds) / nranks / batch_size + batch_counter
writer.add_scalars('loss', {'train': ave_loss.mean()},
global_batch)
writer.add_scalars('accuracy', {'train': ave_acc.mean()},
global_batch)
#writer.add_histogram('input_trans',endpoints['input_trans'].view(-1),global_batch)
ave_loss = CalcMean.CalcMean()
ave_acc = CalcMean.CalcMean()
# release tensors for memory
del inputs, targets, weights, endpoints, loss_value
if config['batch_limiter'] and batch_counter > config[
'batch_limiter']:
logger.info('batch limiter enabled, stop training early')
break
# save at end of epoch
torch.save(model.state_dict(),
model_save + '_%05d.torch_model_state_dict' % epoch)
if nval_tests == -1:
nval_tests = len(testds) / nranks / batch_size
logger.info('epoch %s complete, running validation on %s batches',
epoch, nval_tests)
model.to(device)
# every epoch, evaluate validation data set
with torch.no_grad():
vloss = CalcMean.CalcMean()
vacc = CalcMean.CalcMean()
vious = [CalcMean.CalcMean() for i in range(num_classes)]
for valid_batch_counter, (inputs, targets, class_weights,
nonzero_mask) in enumerate(test_loader):
inputs = inputs.to(device)
targets = targets.to(device)
class_weights = class_weights.to(device)
nonzero_mask = nonzero_mask.to(device)
# set the weights
if balanced_loss:
weights = class_weights
nonzero_to_class_scaler = torch.sum(
nonzero_mask.type(torch.float32)) / torch.sum(
class_weights.type(torch.float32))
else:
weights = nonzero_mask
nonzero_to_class_scaler = torch.ones(1, device=device)
outputs, endpoints = model(inputs)
loss_value = loss_func(outputs, targets.long())
loss_value = torch.mean(
loss_value * weights) * nonzero_to_class_scaler
vloss.add_value(float(loss_value.to('cpu')))
# calc acc
vacc.add_value(
float(acc_func(outputs, targets, weights).to('cpu')))
# calc ious
ious = get_ious(outputs, targets, weights, num_classes)
for i in range(num_classes):
vious[i].add_value(float(ious[i]))
if valid_batch_counter > nval_tests:
break
mean_acc = vacc.mean()
mean_loss = vloss.mean()
# if config['hvd'] is not None:
# mean_acc = config['hvd'].allreduce(torch.tensor([mean_acc]))
# mean_loss = config['hvd'].allreduce(torch.tensor([mean_loss]))
mious = float(
torch.sum(torch.FloatTensor([x.mean()
for x in vious]))) / num_classes
ious_out = {
'jet': vious[0].mean(),
'electron': vious[1].mean(),
'bkgd': vious[2].mean(),
'all': mious
}
# add validation to tensorboard
if writer and rank == 0:
global_batch = epoch * len(
trainds) / nranks / batch_size + batch_counter
writer.add_scalars('loss', {'valid': mean_loss}, global_batch)
writer.add_scalars('accuracy', {'valid': mean_acc},
global_batch)
writer.add_scalars('IoU', ious_out, global_batch)
logger.warning(
'>[%3d of %3d, %5d of %5d]<<< ave valid loss: %6.4f ave valid acc: %6.4f on %s batches >>>',
epoch + 1, epochs, batch_counter,
len(trainds) / nranks / batch_size, mean_loss, mean_acc,
valid_batch_counter + 1)
logger.warning(' >> ious: %s', ious_out)
# update learning rate
lrsched.step()
model = model.MCCNNNet()
#判断是否启用GPU
device = torch.device(args.gpu_device if args.cuda else "cpu")
#加载训练好的model
if args.loadmodel is not None:
pretrain_dict = torch.load(args.loadmodel)
model.load_state_dict(pretrain_dict)
train_datatset = MyDataLoader(args.datapath)
train_loader = torch.utils.data.DataLoader(dataset=train_datatset,
batch_size=args.batch_size,
shuffle=True)
model.to(device)
criterion = nn.BCELoss()
optimizer = optim.SGD(model.parameters(), lr=0.003, momentum=0.9)
def main():
model.train()
for epoch in range(1, args.epochs):
i = 0
for left_image, right_image, label in train_loader:
left_image = left_image.to(device)
right_image = right_image.to(device)
label = label.to(device)
label = label.float()
optimizer.zero_grad()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--epochs',
'-e',
default=300,
type=int,
help='number of epochs to learn')
parser.add_argument('--d_hidden',
'-d',
default=650,
type=int,
help='number of units in hidden layers')
parser.add_argument('--seq_len',
'-b',
type=int,
default=20,
help='learning minibatch size')
parser.add_argument('--betapoint',
'-c',
type=int,
default=10,
help='betapoint for decrease beta')
parser.add_argument('--seed', type=int, default=1111)
parser.add_argument('--lr', type=float, default=0.0001)
parser.add_argument('--batch_size', type=int, default=160)
parser.add_argument('--d_emb', type=int, default=512)
parser.add_argument('--n', type=int, default=2)
parser.add_argument('--h', type=int, default=8)
parser.add_argument('--num', type=int, default=3)
parser.add_argument('--beta', type=float, default=1.0)
parser.add_argument('--training_capacity', type=int, default=3e5)
global device
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
args = parser.parse_args()
np.random.seed(args.seed)
torch.manual_seed(args.seed)
lr = args.lr
base_path = os.path.dirname(os.path.realpath(__file__))
text_data_dir = os.path.join(base_path, '../misspelling/data/')
output_dir = os.path.join(base_path, 'output/')
######################################################
########### DATA PREPARE ###########
######################################################
train_file = 'train_correct.txt'
val_file = 'dev_correct.txt'
test_file = 'test_correct.txt'
train_noise_filename = text_data_dir + 'train.txt'
train_filename = text_data_dir + train_file
valid_noise_filename = text_data_dir + 'dev.txt'
valid_filename = text_data_dir + val_file
test_noise_filename = text_data_dir + 'test.txt'
test_filename = text_data_dir + test_file
vocab, id2vocab = {'<eos>': 0}, {0: '<eos>'}
vocab, id2vocab, train_len = update_vocab(text_data_dir + train_file,
vocab, id2vocab)
#train_noise_tokens = open(train_noise_filename).read().replace('\n', ' <eos> ').strip().split()
#train_tokens = open(train_filename).read().replace('\n', ' <eos> ').strip().split()
vocab, id2vocab, _ = update_vocab(text_data_dir + val_file, vocab,
id2vocab)
valid_noise_tokens = open(valid_noise_filename).read().replace(
'\n', ' <eos> ').strip().split()
valid_tokens = open(valid_filename).read().replace(
'\n', ' <eos> ').strip().split()
vocab, id2vocab, _ = update_vocab(text_data_dir + test_file, vocab,
id2vocab)
test_noise_tokens = open(test_noise_filename).read().replace(
'\n', ' <eos> ').strip().split()
test_tokens = open(test_filename).read().replace(
'\n', ' <eos> ').strip().split()
alph = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz.,:;'*!?`$%&(){}[]-/\@_#"
seq_len = args.seq_len
#X_train, mask_train, Y_train = make_input_data(train_noise_tokens, train_tokens, seq_len, alph, vocab)
X_valid, mask_valid, Y_valid = make_input_data(valid_noise_tokens,
valid_tokens, seq_len, alph,
vocab)
X_test, mask_test, Y_test = make_input_data(test_noise_tokens, test_tokens,
seq_len, alph, vocab)
#X_train, mask_train, Y_train = X_train.to(device), mask_train.to(device), Y_train.to(device)
X_valid, mask_valid, Y_valid = X_valid.to(device), mask_valid.to(
device), Y_valid.to(device)
X_test, mask_test, Y_test = X_test.to(device), mask_test.to(
device), Y_test.to(device)
######################################################
########### MODEL AND TRAINING CONFIG ###########
######################################################
model_name = "beta_{}_emb_{}_h_{}_hidden_{}_n_{}_lr_{}_bs_{}_check_{}".format(
args.beta, args.d_emb, args.h, args.d_hidden, args.n, args.lr,
args.batch_size, args.betapoint)
global message_filename
message_filename = output_dir + 'r_' + model_name + '.txt'
model_filename = output_dir + 'm_' + model_name + '.pt'
with open(message_filename, 'w') as out:
out.write('start\n')
char_vocab_size = len(alph) + 5
global model
if args.num == 3:
model = model.MUDE(char_vocab_size,
d_emb=args.d_emb,
h=args.h,
n=args.n,
d_hidden=args.d_hidden,
vocab_size=len(vocab),
dropout=0.01)
#model = nn.DataParallel(model)
model.to(device)
global criterion
criterion = nn.NLLLoss()
global seq_criterion
seq_criterion = nn.NLLLoss(ignore_index=char_vocab_size - 1)
global optimizer
optimizer = getattr(optim, 'RMSprop')(model.parameters(), lr=lr)
######################################################
########### START TRAINING ###########
######################################################
print(args)
print(message_filename)
best_acc = 0
for epoch in range(1, args.epochs + 1):
with open(train_noise_filename) as f_input:
with open(train_filename) as f_label:
sentence_count = 0
epoch_start_time = time.time()
input_text = ""
label_text = ""
for input_line, label_line in zip(f_input, f_label):
input_text += input_line
label_text += label_line
sentence_count += 1
if sentence_count % args.training_capacity == 0: # or train_len <= args.training_capacity and sentence_count == train_len:
train_noise_tokens = input_text.replace(
'\n', ' <eos> ').strip().split()
train_tokens = label_text.replace(
'\n', ' <eos> ').strip().split()
input_text = ""
label_text = ""
X_train, mask_train, Y_train = make_input_data(
train_noise_tokens, train_tokens, seq_len, alph,
vocab)
train(epoch, X_train, mask_train, Y_train,
args.batch_size, args.seq_len, len(vocab),
char_vocab_size, args)
#val_acc = evaluate(X_valid, mask_valid, Y_valid, args.batch_size, args.seq_len, len(vocab), args)
#test_acc = evaluate(X_test, mask_test, Y_test, args.batch_size, args.seq_len, len(vocab), args)
val_precision, val_recall, val_acc, val_f05, val_real_precision, val_real_recall, val_real_acc, val_real_f05, val_non_precision, val_non_recall, val_non_acc, val_non_f05 = check_performance(
X_valid, mask_valid, Y_valid,
valid_noise_tokens, valid_tokens, id2vocab,
len(vocab), args.seq_len, args)
test_precision, test_recall, test_acc, test_f05, test_real_precision, test_real_recall, test_real_acc, test_real_f05, test_non_precision, test_non_recall, test_non_acc, test_non_f05 = check_performance(
X_test, mask_test,
Y_test, test_noise_tokens, test_tokens, id2vocab,
len(vocab), args.seq_len, args)
message = (
'-' * 89 +
'\n| end of epoch {:3d} | time: {:5.4f}s | valid precision {:5.4f} | valid recall {:5.4f} | valid accuracy {:5.4f} | valid F0.5 {:5.4f} | '
.format(epoch, (time.time() - epoch_start_time),
val_precision, val_recall, val_acc,
val_f05) +
'\n| end of epoch {:3d} | time: {:5.4f}s | valid real-word precision {:5.4f} | valid real-word recall {:5.4f} | valid real-word accuracy {:5.4f} | valid real-word F0.5 {:5.4f} | '
.format(epoch, (time.time() - epoch_start_time),
val_real_precision, val_real_recall,
val_real_acc, val_real_f05) +
'\n| end of epoch {:3d} | time: {:5.4f}s | valid non-word precision {:5.4f} | valid non-word recall {:5.4f} | valid non-word accuracy {:5.4f} | valid non-word F0.5 {:5.4f} | '
.format(epoch, (time.time() - epoch_start_time),
val_non_precision, val_non_recall,
val_non_acc, val_non_f05) +
'\n| end of epoch {:3d} | time: {:5.4f}s | test precision {:5.4f} | test recall {:5.4f} | test accuracy {:5.4f} | test F0.5 {:5.4f} | '
.format(epoch, (time.time() - epoch_start_time),
test_precision, test_recall, test_acc,
test_f05) +
'\n| end of epoch {:3d} | time: {:5.4f}s | test real-word precision {:5.4f} | test real-word recall {:5.4f} | test real-word accuracy {:5.4f} | test real-word F0.5 {:5.4f} | '
.format(epoch, (time.time() - epoch_start_time),
test_real_precision, test_real_recall,
test_real_acc, test_real_f05) +
'\n| end of epoch {:3d} | time: {:5.4f}s | test non-word precision {:5.4f} | test non-word recall {:5.4f} | test non-word accuracy {:5.4f} | test non-word F0.5 {:5.4f} | '
.format(epoch, (time.time() - epoch_start_time),
test_non_precision, test_non_recall,
test_non_acc, test_non_f05) + '-' * 89)
output_s(message, message_filename)
# Save the model if the validation loss is the best we've seen so far.
if val_f05 > best_acc:
save(model, model_filename)
best_acc = val_f05
OUTPUT_DIM = INPUT_DIM
HIDDEN_DIM = 15
BATCH_SIZE = 1
NUM_LAYERS = 1
NUM_EPOCHS = 1
SEQ_SIZE = 60
SKIP_SIZE = 1
model = model.LSTMModel(
input_dim=INPUT_DIM,
output_dim=OUTPUT_DIM,
hidden_dim=HIDDEN_DIM,
batch_size=BATCH_SIZE,
num_layers=NUM_LAYERS,
)
model.to(device=DEVICE)
loss_fn = torch.nn.MSELoss(reduction="sum")
optimiser = torch.optim.Adam(model.parameters(), lr=1e-1)
training_data = make_training_data()
print(training_data[0][0].shape[0])
real = csv_data.iloc[:, -1]
"""
loss = 0
for epoch in range(NUM_EPOCHS):
print("Epoch", epoch, "Loss", loss)
loss = train(model, training_data, loss_fn, optimiser, batch_size=BATCH_SIZE)
predictions = predict()
def main():
dataframe = pd.read_csv('../input/imdb.csv') # load dataframe
dataframe.sentiment = dataframe.sentiment.apply(lambda x: 1
if x == 'positive' else 0)
# sentiment is category target variable so we have to label encode it, we can do it like this by hand, or simply with sklearn.model_selection.LabelEncoder
# now split data into validation and training
df_train, df_valid = model_selection.train_test_split(
dataframe,
test_size=0.1, # 10 percent of dataframe will be for validation
random_state=
42, # if we are going to run multiple time this script, random state enables that everytime we get same split with same random state
shuffle=True, # shuffle indices
stratify=dataframe.sentiment.
values # same distribution in train and valid
)
df_train = df_train.reset_index(
drop=True) # we reset indices from 0 to len(df_train)
df_valid = df_valid.reset_index(
drop=True) # we reset indices from 0 to len(df_valid)
# make datasets with our class in order to make data loaders
training_dataset = dataset.BERTdataset(review=df_train.review.values,
sentiment=df_train.sentiment.values)
# from dataset to dataloader
training_data_loader = torch.utils.data.DataLoader(
dataset=training_dataset,
batch_size=config.TRAINING_BATCH_SIZE,
shuffle=True,
num_workers=4)
validation_dataset = dataset.BERTdataset(
review=df_valid.review.values,
sentiment=df_valid.sentiment.values,
)
# from dataset to dataloader
validation_data_loader = torch.utils.data.DataLoader(
dataset=validation_dataset,
batch_size=config.VALIDATION_BATCH_SIZE,
shuffle=False,
num_workers=4)
device = torch.device('cuda')
model = model.BERTsentiment()
model.to(device) # move model to cuda device
# params to optimize
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.001
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.00
}]
number_of_training_steps = int(
len(df_train) / config.TRAINING_BATCH_SIZE * config.EPOCHS)
#AdamW focuses on regularization and model does better on generalization
optimizer = AdamW(params=param_optimizer, lr=3e-5)
scheduler = get_linear_schedule_with_warmup(
optimizer=optimizer,
num_warmup_steps=0,
num_training_steps=number_of_training_steps,
)
best_accuracy = []
for epoch in range(config.EPOCHS):
print('EPOCH:', epoch + 1)
engine.training_loop(training_data_loader, model, optimizer, scheduler,
device)
outputs, sentiments = engine.validation_loop(validation_data_loader,
model, device)
# distribution is 50 50 so we can use acc score
outputs = np.array(outputs) >= 0.5 # positive class
accuracy = metrics.accuracy_score(sentiments, outputs)
print('ACCURACY SCORE', {accuracy})
if accuracy > best_accuracy:
torch.save(model.state_dict(),
config.MODEL_PATH) # save model in working dir
best_accuracy = accuracy
def main():
"""
Training and validation.
"""
global epochs_since_improvement, start_epoch, label_map, best_loss, epoch, checkpoint
# Initialize model or load checkpoint
if checkpoint is None:
model = SSD300(n_classes=n_classes)
# Initialize the optimizer, with twice the default learning rate for biases, as in the original Caffe repo
biases = list()
not_biases = list()
for param_name, param in model.named_parameters():
if param.requires_grad:
if param_name.endswith('.bias'):
biases.append(param)
else:
not_biases.append(param)
optimizer = torch.optim.SGD(params=[{
'params': biases,
'lr': 2 * lr
}, {
'params': not_biases
}],
lr=lr,
momentum=momentum,
weight_decay=weight_decay)
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
epochs_since_improvement = checkpoint['epochs_since_improvement']
best_loss = checkpoint['best_loss']
print(
'\nLoaded checkpoint from epoch %d. Best loss so far is %.3f.\n' %
(start_epoch, best_loss))
model = checkpoint['model']
optimizer = checkpoint['optimizer']
# Move to default device
model = model.to(device)
criterion = MultiBoxLoss(priors_cxcy=model.priors_cxcy).to(device)
# Custom dataloaders
train_dataset = PascalVOCDataset(data_folder, split='train')
val_dataset = PascalVOCDataset(data_folder, split='test')
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
collate_fn=train_dataset.collate_fn,
num_workers=workers,
pin_memory=True) # note that we're passing the collate function here
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=batch_size,
shuffle=True,
collate_fn=val_dataset.collate_fn,
num_workers=workers,
pin_memory=True)
# Epochs
for epoch in range(start_epoch, epochs):
# Paper describes decaying the learning rate at the 80000th, 100000th, 120000th 'iteration', i.e. model update or batch
# The paper uses a batch size of 32, which means there were about 517 iterations in an epoch
# Therefore, to find the epochs to decay at, you could do,
# if epoch in {80000 // 517, 100000 // 517, 120000 // 517}:
# adjust_learning_rate(optimizer, 0.1)
# In practice, I just decayed the learning rate when loss stopped improving for long periods,
# and I would resume from the last best checkpoint with the new learning rate,
# since there's no point in resuming at the most recent and significantly worse checkpoint.
# So, when you're ready to decay the learning rate, just set checkpoint = 'BEST_checkpoint_ssd300.pth.tar' above
# and have adjust_learning_rate(optimizer, 0.1) BEFORE this 'for' loop
# One epoch's training
train(train_loader=train_loader,
model=model,
criterion=criterion,
optimizer=optimizer,
epoch=epoch)
# One epoch's validation
val_loss = validate(val_loader=val_loader,
model=model,
criterion=criterion)
# Did validation loss improve?
is_best = val_loss < best_loss
best_loss = min(val_loss, best_loss)
if not is_best:
epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" %
(epochs_since_improvement, ))
else:
epochs_since_improvement = 0
# Save checkpoint
save_checkpoint(epoch, epochs_since_improvement, model, optimizer,
val_loss, best_loss, is_best)
