def one_epoch_training(model, loss, optimiser, scheduler, device, train_gen, test_gen,
timesteps, batch_size):
t0 = time.time()
scheduler = scheduler
optimiser = optimiser
loss_fn = loss
# Track losses:
loss_vals_train = np.zeros((len(train_gen)))
loss_vals_test = np.zeros((len(test_gen)))
total_time = 0
batch_nr = 0
for batch, labels in train_gen:
model.batch_size = batch_size
batch = batch.view(timesteps, batch_size, -1)
# print(batch.shape)
labels = labels.float()
# Transfer to GPU
batch, labels = batch.to(device), labels.to(device)
optimiser.zero_grad()
preds = model(batch)
loss = loss_fn(preds, labels)
loss_vals_train[batch_nr] = loss.item()
loss.backward()
optimiser.step()
if scheduler is not None:
scheduler.step()
batch_nr += 1
batch_nr = 0
for batch, labels in test_gen:
model.batch_size = 1
batch = batch.view(timesteps, 1, -1)
labels = labels.float()
# Transfer to GPU
batch, labels = batch.to(device), labels.to(device)
optimiser.zero_grad()
y_pred_test = model(batch)
loss = loss_fn(y_pred_test, labels)
loss_vals_test[batch_nr] = loss.item()
batch_nr += 1
error = np.mean(loss_vals_test)
print('The epoch took {} seconds'.format(time.time() - t0))
total_time += time.time() - t0
return error, model
def _predict_n_forward(self, dataset, model_key, window_normalisation=True):
"""
Idea:
Do a dataset with the longest tiemesteps
feed it to run_predictions
inside run predictions split it up to accommodate shorter timestep sequences
"""
timesteps = self.model_config[model_key]["timesteps"]
model = self.models_dict[model_key]
n_forward = self.model_config[model_key]['num_forward']
dt = dataset.get_test_data(timesteps, window_normalisation, n_forward)[0]
predicted = []
predicted_denormalized = []
skip = self.most_timesteps - timesteps
for i in range(int(len(dt) / self.prediction_length)):
if skip + i * self.prediction_length < dt.shape[0]:
curr_frame = dt[skip + i * self.prediction_length]
curr_frame = torch.from_numpy(curr_frame).type(torch.Tensor).detach().view(timesteps, 1, -1)
prediction = model(curr_frame)[0, 0].detach()
predicted.append(prediction.cpu().numpy())
predicted_denormalized.append(
denormalise("window", dataset.w_normalisation_p0_test[i * self.prediction_length][0],
prediction.cpu().numpy(), None)
)
return predicted, predicted_denormalized
def test(model, dataset):
avg_loss = tf.keras.metrics.Mean('loss', dtype=tf.float32)
avg_accuracy = tf.keras.metrics.Mean('accuracy', dtype=tf.float32)
for images, labels in dataset:
logits = model(images, training=False)
avg_loss(loss_fn(labels, logits))
avg_accuracy(accuracy_fn(labels, logits))
print('# test_loss :', avg_loss.result().numpy(), ', test accurary :', avg_accuracy.result().numpy())
def train_step(images, labels):
with tf.GradientTape() as tape:
logits = model(images, training=True)
loss = loss_fn(labels, logits)
accuracy = accuracy_fn(labels, logits)
grads = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
return loss, accuracy
def evaluate(model, val_loader, criterion):
model.eval()
epoch_loss = 0
epoch_acc = 0
with torch.no_grad():
for X, Y in val_loader:
X = X.to(c.device)
Y = Y.to(c.device)
outputs = model(X)
epoch_acc += torch.sum((torch.round(outputs) == Y)).float() / Y.size(0)
loss = criterion(outputs, Y)
epoch_loss += loss
model.train()
return epoch_loss / len(val_loader), epoch_acc / len(val_loader)
def train(train_loader, model, criterion_encoder, optimizer_encoder, classifier, criterion_fc, optimizer_fc, epoch, opt):
model.train()
losses = AverageMeter()
fc_losses = AverageMeter()
encoder_losses = AverageMeter()
top1 = AverageMeter()
_print("--" * 50)
for i, (images, labels) in enumerate(train_loader):
images = torch.cat(images, dim=0).cuda()
labels = torch.cat([labels, labels], dim=0).cuda()
BSZ = labels.shape[0]
warmup_learning_rate(opt, epoch, i, len(
train_loader), optimizer_encoder)
warmup_learning_rate(opt, epoch, i, len(train_loader), optimizer_fc)
features = model(images, labels)
encoder_loss = criterion_encoder(features, labels)
encoder_losses.update(encoder_loss.item(), BSZ)
logits = classifier(features)
fc_loss = criterion_fc(logits, labels)
fc_losses.update(fc_loss.item(), BSZ)
acc1, acc5 = accuracy(logits, labels, topk=(1, 5))
top1.update(acc1[0], BSZ)
loss = encoder_loss + fc_loss
# loss = fc_loss
losses.update(loss.item(), BSZ)
optimizer_encoder.zero_grad()
optimizer_fc.zero_grad()
loss.backward()
optimizer_encoder.step()
optimizer_fc.step()
progress_bar(i, len(train_loader), "train")
_print("epoch:{}".format(epoch))
_print(
"loss: {loss.avg:.4f} "
"encoder_loss: {encoder_loss.avg:.4f} "
"fc_loss: {fc_loss.avg:.4f} "
"train acc@1 {top1.avg:.4f} ".format(
loss=losses, encoder_loss=encoder_losses, fc_loss=fc_losses, top1=top1))
return losses.avg, top1.avg
def test(model, dataset):
for images, labels in dataset:
cams = model(images, training=False, using_cam=True)
cams = tf.image.resize(cams, (image_size, image_size),
method='bilinear')
cams = normalize(cams) * 255
for image, cam in zip(images, cams.numpy()):
for class_index in range(classes):
cv2.imshow('cam - {}'.format(class_index),
cam[..., class_index].astype(np.uint8))
cv2.imshow('show', image.astype(np.uint8))
cv2.waitKey(0)
def ocr(absolute_path, img_type, tune_angle=0):
print("OCR Starting!")
img = Image.open(absolute_path).convert("RGB")
t = time.time()
if tune_angle == 1:
global_tune = True
fine_tune = True
else:
global_tune = False
fine_tune = False
params = select_config(img_type)
img, res_origin, res_sorted, angle = core.model(img,
global_tune=global_tune, # 图片的整体大方向调整，逆时针旋转 镜像. 大约0.5s
fine_tune=fine_tune,
# 微调倾斜角（如果能保证图像水平，或者global_tune之后为水平，则不需要微调）. 大约2s
config=params[0],
if_im=params[1],
left_adjust=params[2], # 对检测的文本行进行向左延伸
right_adjust=params[3], # 对检测的文本行进行向右延伸
alpha=params[4], # 对检测的文本行进行向右、左延伸的倍数
result_typeset_opotion=0, # 结果排版方案 TODO: other two type
)
print("检测识别1 总耗时:{}s\n".format(time.time() - t))
json_str = ''
# for index, _ in enumerate(re_origin):
# print(res_origin[index]["text"])
if len(res_sorted) < 1:
# TODO: return
return 'we are sorry, no text detected!'
else:
basename = os.path.basename(absolute_path)
print(basename)
out_path = './users/out_result/' + basename[:-4] + '.txt'
print(out_path)
out = open(out_path, 'w')
for row in range(res_sorted.shape[0]):
for col in range(res_sorted.shape[1]):
json_str += res_sorted[row][col]
out.write(res_sorted[row][col])
json_str += '\n'
out.write('\n')
out.close()
print(json_str)
return json_str
def train(train_loader, model, optimizer, criterion):
model.train()
epoch_loss = 0
bar = tqdm(total=len(train_loader))
b_ix = 1
for X, Y in train_loader:
X = X.to(c.device)
Y = Y.to(c.device)
optimizer.zero_grad()
outputs = model(X)
loss = criterion(outputs, Y)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
if b_ix % 10 == 0:
bar.update(10)
bar.set_description('current loss:{:.4f}'.format(epoch_loss / b_ix))
b_ix += 1
bar.update((b_ix - 1) % 10)
bar.close()
return epoch_loss / len(train_loader)
def validate(val_loader, model, criterion_encoder, classifier, criterion_fc, epoch, opt):
model.eval()
losses = AverageMeter()
fc_losses = AverageMeter()
encoder_losses = AverageMeter()
top1 = AverageMeter()
with torch.no_grad():
for i, (images, labels) in enumerate(val_loader):
images = images.cuda()
labels = labels.cuda()
BSZ = labels.shape[0]
# caculate class loss
features = model(images, labels)
# encoder_loss = criterion_encoder(features, labels)
# encoder_losses.update(encoder_loss.item(), BSZ)
logits = classifier(features)
fc_loss = criterion_fc(logits, labels)
fc_losses.update(fc_loss.item(), BSZ)
loss = fc_loss
losses.update(loss.item(), BSZ)
acc1, acc5 = accuracy(logits, labels, topk=(1, 5))
top1.update(acc1[0], BSZ)
progress_bar(i, len(val_loader), "eval test set")
_print(
"loss: {loss.avg:.4f} "
"fc_loss :{fc_loss.avg:.4f} "
"acc@1 {top1.avg:.4f}".format(loss=losses, fc_loss=fc_losses, top1=top1))
return losses.avg, top1.avg
SOFTWARE.
"""
__author__ = "Marcus T. Andersson"
__copyright__ = "Copyright 2020, Marcus T. Andersson"
__credits__ = ["Marcus T. Andersson"]
__license__ = "MIT"
__version__ = "2"
__maintainer__ = "Marcus T. Andersson"
import core
statements = []
# Entities have Identity
statements += core.model("EntityE -- LabelEScn1")
statements += core.model("EntityE -- IdentityEScn1")
# Files have Path
statements += core.model("FileE -- PathEScn1")
# Mutables can change content but Constants remain the same
statements += core.model("MutableE -- ContentEEcnn -- ConstantE")
statements += core.model("ConstantE -- ContentTypeEScn1")
statements += core.model("ConstantE -- CreationTimeEScn1")
statements += core.model("ConstantE -- ContentEBcn1")
statements += core.model("ConstantE -- ShaEScn1")
# Embedded things are located in Containers
statements += core.model("ContainerE -- EmbeddedEEcnn -- EmbeddedE")
def rolling_window(model_config,
datasetName,
start_from,
restart,
restart_window,
window_length,
timesteps,
max_epochs,
data_folder,
save_folder,
col='log_ret',
returns=True):
### Model:
model, loss_fn, optimiser = give_me_model(model_config)
### CUDA for PyTorch:
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if use_cuda else "cpu")
cudnn.benchmark = True
model = model
if torch.cuda.is_available():
print("We're running on GPU")
model.cuda()
### Model end
path = data_folder + datasetName + '.csv'
dataset = miniDataLoader(path, start_from, cols=col, returns=returns)
train_dt = dataset.get_data(timesteps, False, 1)
predictions = []
true_labels = []
end_of_window = 253 * window_length
end_of_loop = len(train_dt[0])
print(end_of_loop)
i = 0
while end_of_window + i + 1 < end_of_loop:
print(i)
t0 = time.time()
train_windows = (train_dt[0][i:end_of_window + i],
train_dt[1][i:end_of_window + i])
test_window = (train_dt[0][end_of_window + i + 1],
train_dt[1][end_of_window + i + 1])
train_windows = Dataset(train_windows)
dataloader_params = {
'batch_size': 1,
'shuffle': True,
'drop_last': True,
'num_workers': 0
}
training_windows_generator = data.DataLoader(train_windows,
**dataloader_params)
avg_loss_epoch = np.zeros((max_epochs, 1))
best_accuracy = torch.FloatTensor([1000])
if i >= restart_window:
restart = False
if restart is False:
for epoch in range(max_epochs):
# print("Epoch nr: " + str(epoch))
loss_vals = np.zeros(
(max_epochs, len(training_windows_generator)))
batch_nr = 0
for batch, labels in training_windows_generator:
model.batch_size = 1
batch = batch.view(timesteps, 1, -1)
labels = labels.float()
# Transfer to GPU
batch, labels = batch.to(device), labels.to(device)
optimiser.zero_grad()
preds = model(batch)
loss = loss_fn(preds.view(-1), labels)
loss_vals[epoch, batch_nr] = loss.item()
loss.backward()
optimiser.step()
batch_nr += 1
avg_loss_epoch[epoch] = np.mean(loss_vals[epoch, :])
is_best = bool(avg_loss_epoch[epoch] < best_accuracy.numpy())
best_accuracy = torch.FloatTensor(
min(avg_loss_epoch[epoch], best_accuracy.numpy()))
save_checkpoint(
{
'epoch': 0 + epoch + 1,
'state_dict': model.state_dict(),
'best_accuracy': best_accuracy
}, i, is_best, save_folder + datasetName)
print(save_folder)
path_to_checkpoint = save_folder + datasetName + '/' + 'checkpoint' + str(
i) + '.pth.tar'
checkpoint = torch.load(path_to_checkpoint)
model.load_state_dict(checkpoint['state_dict'])
test_seq = torch.from_numpy(test_window[0]).float().view(
timesteps, 1, -1).to(device)
predictions.append(model(test_seq))
true_labels.append(test_window[1])
output = (predictions, true_labels, dataset.dates)
file_output = open(save_folder + datasetName + '/last_pickle.obj',
'wb')
pickle.dump(output, file_output)
print('The window took {} seconds'.format(time.time() - t0))
i += 1
output = (predictions, true_labels)
return output
if __name__ == '__main__':
print("OCR Starting!")
img = Image.open(paths[0]).convert("RGB")
t = time.time()
# TODO if res_sorted is null
img, res_origin, res_sorted, angle = core.model(
img,
global_tune=False, # 图片的整体大方向调整，逆时针旋转 镜像. 大约0.5s
fine_tune=False,
# 微调倾斜角（如果能保证图像水平，或者global_tune之后为水平，则不需要微调）. 大约2s
config=dict(
MAX_HORIZONTAL_GAP=50,
# (1.5~2.5较佳)字符之间的最大间隔，用于文本行的合并 TODO:最好是自动计算
MIN_V_OVERLAPS=0.6, # 小 ==》斜 TODO
MIN_SIZE_SIM=0.6,
TEXT_PROPOSALS_MIN_SCORE=0.1,
# 值越小,候选框越多（一些模棱两可的文字）
TEXT_PROPOSALS_NMS_THRESH=0.3 # 候选框非极大值抑制
),
if_im=False,
left_adjust=False, # 对检测的文本行进行向左延伸
right_adjust=False, # 对检测的文本行进行向右延伸
alpha=0.2, # 对检测的文本行进行向右、左延伸的倍数
result_typeset_opotion=0, # 结果排版方案 TODO: other two type
)
print("检测识别1 总耗时:{}s\n".format(time.time() - t))
# for index, _ in enumerate(re_origin):
# print(res_origin[index]["text"])
if len(res_sorted) < 1:
print("we are sorry, no text are detected!")
else:
print("OCR Starting!")
with open('testA_part_result_test.csv', 'w') as csv_file:
fieldheader = ['filename', 'content']
writer = csv.DictWriter(csv_file, fieldheader)
writer.writeheader()
for path in paths:
img = Image.open(path).convert("RGB")
_, result, angle = core.model(
img,
global_tune=cfg.global_tune, # 图片的整体大方向调整，逆时针旋转 镜像. 大约0.5s
fine_tune=cfg.
fine_tune, # 微调倾斜角（如果能保证图像水平，或者global_tune之后为水平，则不需要微调）. 大约2s
config=dict(
MAX_HORIZONTAL_GAP=80, # 字符之间的最大间隔，用于文本行的合并 TODO:最好是自动计算
MIN_V_OVERLAPS=0.6, # 小 ==》斜 TODO
MIN_SIZE_SIM=0.6,
TEXT_PROPOSALS_MIN_SCORE=0.1, # 值越小,候选框越多（一些模棱两可的文字）
TEXT_PROPOSALS_NMS_THRESH=0.4 # 候选框非极大值抑制
),
if_im=cfg.if_im,
left_adjust=True, # 对检测的文本行进行向左延伸
right_adjust=True, # 对检测的文本行进行向右延伸
alpha=0.2 # 对检测的文本行进行向右、左延伸的倍数
)
content = ''
row = {} # TODO 修改为多行一次写入
for index, _ in enumerate(result):
content += result[index]["text"]
row.update(filename=path[17:-4], content=content)
writer.writerow(row)
# writer.writerow({'mp': '7-DForc.郭美a刚g足扇僧:29370777707900273970ee0.o)110193721:2997311国警', 'mp1': '合号单币鄂昨eo电舒y营柱国)警翠女义职翠餐武署益罩发兴'})
def train(args):
"""Sets up the model to train"""
# Create a writer object to log events during training
writer = SummaryWriter(pjoin('runs', 'exp_1'))
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Load splits
x_train, y_train, x_val, y_val, seismic = train_val_split(args)
# Convert to torch tensors in the form (N, C, L)
x_train = torch.from_numpy(np.expand_dims(x_train, 1)).float().to(device)
y_train = torch.from_numpy(np.expand_dims(y_train, 1)).float().to(device)
x_val = torch.from_numpy(np.expand_dims(x_val, 1)).float().to(device)
y_val = torch.from_numpy(np.expand_dims(y_val, 1)).float().to(device)
seismic = torch.from_numpy(np.expand_dims(seismic, 1)).float().to(device)
# Set up the dataloader for training dataset
dataset = SeismicLoader(x_train, y_train)
train_loader = DataLoader(dataset=dataset,
batch_size=args.batch_size,
shuffle=False)
# import tcn
model = TCN(1, 1, args.tcn_layer_channels, args.kernel_size,
args.dropout).to(device)
#model = ANN().to(device)
#model = lstm().to(device)
# Set up loss
criterion = torch.nn.MSELoss()
# Define Optimizer
optimizer = torch.optim.Adam(model.parameters(),
weight_decay=args.weight_decay,
lr=args.lr)
# Set up list to store the losses
train_loss = [np.inf]
val_loss = [np.inf]
iter = 0
# Start training
for epoch in range(args.n_epoch):
for x, y in train_loader:
model.train()
optimizer.zero_grad()
y_pred = model(x)
loss = criterion(y_pred, y)
loss.backward()
optimizer.step()
train_loss.append(loss.item())
writer.add_scalar(tag='Training Loss',
scalar_value=loss.item(),
global_step=iter)
if epoch % 200 == 0:
with torch.no_grad():
model.eval()
y_pred = model(x_val)
loss = criterion(y_pred, y_val)
val_loss.append(loss.item())
writer.add_scalar(tag='Validation Loss',
scalar_value=loss.item(),
global_step=iter)
print(
'epoch:{} - Training loss: {:0.4f} | Validation loss: {:0.4f}'.
format(epoch, train_loss[-1], val_loss[-1]))
if epoch % 100 == 0:
with torch.no_grad():
model.eval()
AI_inv = model(seismic)
fig, ax = plt.subplots()
ax.imshow(AI_inv[:, 0].detach().cpu().numpy().squeeze().T,
cmap="rainbow")
ax.set_aspect(4)
writer.add_figure('Inverted Acoustic Impedance', fig, iter)
iter += 1
writer.close()
# Set up directory to save results
results_directory = 'results'
seismic_offsets = np.expand_dims(marmousi_seismic().squeeze()[:, 100:600],
1)
seismic_offsets = torch.from_numpy(
(seismic_offsets - seismic_offsets.mean()) /
seismic_offsets.std()).float()
with torch.no_grad():
model.cpu()
model.eval()
AI_inv = model(seismic_offsets)
if not os.path.exists(
results_directory
): # Make results directory if it doesn't already exist
os.mkdir(results_directory)
print('Saving results...')
else:
print('Saving results...')
np.save(pjoin(results_directory, 'AI.npy'),
marmousi_model().T[452:2399, 400:2400])
np.save(pjoin(results_directory, 'AI_inv.npy'),
AI_inv.detach().numpy().squeeze()[452:2399])
print('Results successfully saved.')
def train_model(model, loss, optimiser, scheduler, max_epochs, train_gen, test_gen,
timesteps, batch_size, folder, resuming):
######## CUDA for PyTorch
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if use_cuda else "cpu")
cudnn.benchmark = True
model = model
if torch.cuda.is_available():
print("We're running on GPU")
model.cuda()
#######
scheduler = scheduler
optimiser = optimiser
loss_fn = loss
# Track losses:
loss_vals_train = np.zeros((max_epochs, len(train_gen)))
loss_vals_test = np.zeros((max_epochs, len(test_gen)))
avg_loss_epoch_train = np.zeros((max_epochs, 1))
avg_loss_epoch_test = np.zeros((max_epochs, 1))
print("Start Training")
total_time = 0
best_accuracy = torch.FloatTensor([1000])
start_epoch = 0
if resuming:
path_to_checkpoint = folder + '/checkpoint.pth.tar'
cuda = torch.cuda.is_available()
if cuda:
checkpoint = torch.load(path_to_checkpoint)
else:
# Load GPU model on CPU
checkpoint = torch.load(path_to_checkpoint,
map_location=lambda storage,
loc: storage)
start_epoch = checkpoint['epoch']
best_accuracy = checkpoint['best_accuracy']
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}' (trained for {} epochs)".format(path_to_checkpoint, checkpoint['epoch']))
for epoch in range(start_epoch, max_epochs, 1):
print("Epoch nr: " + str(epoch))
t0 = time.time()
batch_nr = 0
for batch, labels in train_gen:
model.batch_size = batch_size
batch = batch.view(timesteps, batch_size, -1)
# print(batch.shape)
labels = labels.float()
# Transfer to GPU
batch, labels = batch.to(device), labels.to(device)
optimiser.zero_grad()
preds = model(batch)
loss = loss_fn(preds, labels)
loss_vals_train[epoch, batch_nr] = loss.item()
loss.backward()
optimiser.step()
batch_nr += 1
if scheduler is not None:
scheduler.step()
batch_nr = 0
for batch, labels in test_gen:
model.batch_size = 1
batch = batch.view(timesteps, 1, -1)
labels = labels.float()
# Transfer to GPU
batch, labels = batch.to(device), labels.to(device)
optimiser.zero_grad()
y_pred_test = model(batch)
loss = loss_fn(y_pred_test, labels)
loss_vals_test[epoch, batch_nr] = loss.item()
batch_nr += 1
error = np.mean(loss_vals_test[epoch, :])
print('The epoch took {} seconds'.format(time.time() - t0))
total_time += time.time() - t0
avg_loss_epoch_train[epoch] = np.mean(loss_vals_train[epoch, :])
avg_loss_epoch_test[epoch] = np.mean(loss_vals_test[epoch, :])
is_best = bool(avg_loss_epoch_test[epoch] < best_accuracy.numpy())
best_accuracy = torch.FloatTensor(min(avg_loss_epoch_test[epoch], best_accuracy.numpy()))
print(best_accuracy)
save_checkpoint({
'epoch': start_epoch + epoch + 1,
'state_dict': model.state_dict(),
'best_accuracy': best_accuracy
}, is_best, folder)
print('Total training time is: ' + str(total_time) + ' seconds')
save_path = folder + '/' + 'last_model.pth.tar'
torch.save({
'epoch': start_epoch + epoch + 1,
'state_dict': model.state_dict(),
'best_accuracy': best_accuracy
}, save_path)
with open(folder + '/model_summary.txt', 'w+') as f:
f.write(str(model)) # Python 3.x
print("Last Model Saved")
return model, avg_loss_epoch_train, avg_loss_epoch_test, error
