class CNN_skipCo_trainer(object):
def __init__(self):
self.image_type = 'US'
self.dataset = ProcessData(train_ratio=0.9,
process_raw_data=False,
do_augment=False,
add_augment=False,
do_flip=True,
do_blur=True,
do_deform=True,
do_crop=True,
image_type=self.image_type,
get_scale_center=False,
single_sample=True)
self.model = deep_model.deep_model(criterion=nn.MSELoss(),
optimizer=torch.optim.Adam,
learning_rate=0.001,
weight_decay=0)
if torch.cuda.is_available():
torch.cuda.current_device()
self.model.cuda()
self.logger = Logger(model=self.model,
project_root_dir=self.dataset.project_root_dir,
image_type=self.image_type)
self.epochs = 250
def fit(self, learning_rate, use_one_cycle=False):
# get scale and center parameters
scale_params_low, scale_params_high = self.dataset.load_params(
param_type="scale_params")
mean_image_low, mean_image_high = self.dataset.load_params(
param_type="mean_images")
# currently for one image:
'''
self.dataset.batch_names(batch_size=5)
X, Y = self.dataset.create_train_batches(self.dataset.train_batch_chunks[1])
print(X.shape)
X = X[0,:,:]
Y = Y[0,:,:]
'''
# load validation set, normalize and parse into tensor
input_tensor_val, target_tensor_val = self.dataset.scale_and_parse_to_tensor(
batch_files=self.dataset.val_file_names,
scale_params_low=scale_params_low,
scale_params_high=scale_params_high,
mean_image_low=mean_image_low,
mean_image_high=mean_image_high)
if torch.cuda.is_available():
input_tensor_val = input_tensor_val.cuda()
target_tensor_val = target_tensor_val.cuda()
if use_one_cycle:
higher_rate = learning_rate
lower_rate = 1 / 10 * higher_rate
num_epochs = self.epochs
up_num = int((num_epochs - 50) / 2)
down_num = up_num
ann_num = num_epochs - up_num - down_num
lr_up = np.linspace(lower_rate, higher_rate, num=up_num)
lr_down = np.linspace(higher_rate, lower_rate, num=down_num)
lr_anihilating = np.linspace(lower_rate, 0, num=ann_num)
learning_rates = np.append(np.append(lr_up, lr_down),
lr_anihilating)
else:
self.model.set_learning_rate(learning_rate)
for e in range(0, self.epochs):
if use_one_cycle:
lr = learning_rates[e]
self.model.set_learning_rate(lr)
# separate names into random batches and shuffle every epoch
self.dataset.batch_names(batch_size=32)
# in self.batch_number is the number of batches in the training set
for i in range(self.dataset.batch_number):
input_tensor, target_tensor = self.dataset.scale_and_parse_to_tensor(
batch_files=self.dataset.train_batch_chunks[i],
scale_params_low=scale_params_low,
scale_params_high=scale_params_high,
mean_image_low=mean_image_low,
mean_image_high=mean_image_high)
if torch.cuda.is_available():
input_tensor = input_tensor.cuda()
target_tensor = target_tensor.cuda()
self.model.train_model(input_tensor,
target_tensor,
current_epoch=e)
# calculate the validation loss and add to validation history
self.logger.get_val_loss(val_in=input_tensor_val,
val_target=target_tensor_val)
# save model every x epochs
if e % 25 == 0 or e == self.epochs - 1:
self.logger.log(
save_appendix='_epoch_' + str(e),
current_epoch=e,
epochs=self.epochs,
mean_images=[mean_image_low, mean_image_high],
scale_params=[scale_params_low, scale_params_high])
# how to undo the scaling:
# unscaled_X = utils.scale_and_center_reverse(scale_center_X,
# scale_params_low, mean_image_low, image_type = self.dataset.image_type)
# unscaled_Y = utils.scale_and_center_reverse(scale_center_Y, scale_params_high,
# mean_image_high, image_type=self.dataset.image_type)
def predict(self):
# self.model.predict()
# see self.dataset.X_val and self.dataset.Y_val
pass
def find_lr(self, init_value=1e-8, final_value=10., beta=0.98):
"""
learning rate finder. goes through multiple learning rates and does one forward pass with each and tracks the loss.
it returns the learning rates and losses so one can make an image of the loss from which one can find a suitable learning rate. Pick the highest one on which the loss is still decreasing (so not the minimum)
:param train_in: training data input
:param target: training data target
:param init_value: inital value which the learning rate start with. init_value < final_value. Default: 1e-8
:param final_value: last value of the learning rate. Default: 10.
:param beta: used for smoothing the loss. Default: 0.98
:return: log_learning_rates, losses
"""
# get scale and center parameters
scale_params_low, scale_params_high = self.dataset.load_params(
param_type="scale_params")
mean_image_low, mean_image_high = self.dataset.load_params(
param_type="mean_images")
import math
print(len(self.dataset.train_file_names))
num = len(self.dataset.train_file_names) - 1
mult = (final_value / init_value)**(1 / num)
lr = init_value
self.model.optimizer.param_groups[0]['lr'] = lr
avg_loss = 0.
best_loss = 0.
batch_num = 0
losses = []
log_lrs = []
#for i in range(1, train_in.shape[0]):
self.dataset.batch_names(batch_size=2)
# in self.batch_number is the number of batches in the training set
print(self.dataset.batch_number)
for i in range(self.dataset.batch_number):
sys.stdout.write('\r current iteration : ' + str(i))
input_tensor, target_tensor = self.dataset.scale_and_parse_to_tensor(
batch_files=self.dataset.val_file_names,
scale_params_low=scale_params_low,
scale_params_high=scale_params_high,
mean_image_low=mean_image_low,
mean_image_high=mean_image_high)
if torch.cuda.is_available():
input_tensor = input_tensor.cuda()
target_tensor = target_tensor.cuda()
batch_num += 1
# As before, get the loss for this mini-batch of inputs/outputs
self.model.optimizer.zero_grad()
outputs = self.model.forward(input_tensor)
loss = self.model.criterion(outputs, target_tensor)
# Compute the smoothed loss
avg_loss = beta * avg_loss + (1 - beta) * loss.item()
smoothed_loss = avg_loss / (1 - beta**batch_num)
# Stop if the loss is exploding
if batch_num > 1 and smoothed_loss > 4 * best_loss:
print('loss exploding')
return log_lrs, losses
# Record the best loss
if smoothed_loss < best_loss or batch_num == 1:
best_loss = smoothed_loss
# Store the values
losses.append(smoothed_loss)
log_lrs.append(math.log10(lr))
# Do the SGD step
loss.backward()
self.model.optimizer.step()
# Update the lr for the next step
lr *= mult
self.model.optimizer.param_groups[0]['lr'] = lr
# Plot the results
'''
lrs = 10 ** np.array(log_lrs)
fig, ax = plt.subplots(1)
ax.plot(lrs, losses)
ax.set_xscale('log')
ax.set_xlim((1e-8, 1))
ax.figure.show()
ax.figure.savefig('learning_rate_finder.png')
'''
return log_lrs, losses
class CNN_skipCo_trainer(object):
def __init__(self, image_type, batch_size, log_period, epochs, data_type, train_ratio,
process_raw_data, pro_and_augm_only_image_type, do_heavy_augment,do_augment,
add_augment, do_rchannels,do_flip, do_blur, do_deform, do_crop,do_speckle_noise,
trunc_points, get_scale_center, single_sample,do_scale_center,
height_channel_oa, use_regressed_oa, include_regression_error, add_f_test,
only_f_test_in_target, channel_slice_oa, process_all_raw_folders,
conv_channels,kernels, model_name, input_size,output_channels, drop_probs,
di_conv_channels, dilations, learning_rates, hetero_mask_to_mask,hyper_no, input_ds_mask,
input_ss_mask, ds_mask_channels, add_skip):
self.image_type = image_type
self.batch_size = batch_size
self.log_period = log_period
self.epochs = epochs
self.dataset = ProcessData(data_type=data_type, train_ratio=train_ratio, process_raw_data=process_raw_data,
pro_and_augm_only_image_type=pro_and_augm_only_image_type,
do_heavy_augment=do_heavy_augment,
do_augment=do_augment, add_augment=add_augment, do_rchannels=do_rchannels,
do_flip=do_flip, do_blur=do_blur, do_deform=do_deform, do_crop=do_crop,
do_speckle_noise=do_speckle_noise,trunc_points=trunc_points,
image_type=image_type, get_scale_center=get_scale_center,
single_sample=single_sample,
do_scale_center=do_scale_center,
height_channel_oa=height_channel_oa, use_regressed_oa=use_regressed_oa,
include_regression_error=include_regression_error,
add_f_test=add_f_test, only_f_test_in_target=only_f_test_in_target,
channel_slice_oa=channel_slice_oa,
process_all_raw_folders=process_all_raw_folders,
hetero_mask_to_mask=hetero_mask_to_mask)
self.model_convdeconv = ConvDeconv(conv_channels=conv_channels,
kernels=kernels,
model_name=model_name, input_size=input_size,
output_channels=output_channels, drop_probs=drop_probs,
add_skip=add_skip)
self.model_dilated = DilatedTranslator(conv_channels=di_conv_channels, dilations=dilations)
self.model = ImageTranslator([self.model_dilated])
if torch.cuda.is_available():
torch.cuda.current_device()
self.model.cuda()
self.learning_rates = learning_rates
self.logger = Logger(model=self.model, project_root_dir=self.dataset.project_root_dir,
image_type=self.image_type, dataset=self.dataset, batch_size=self.batch_size,
epochs=self.epochs,learning_rates=self.learning_rates,hyper_no=hyper_no)
def fit(self, learning_rate, lr_method='standard'):
# get scale and center parameters
scale_params_low, scale_params_high = self.dataset.load_params(param_type="scale_params")
mean_image_low, mean_image_high = self.dataset.load_params(param_type="mean_images")
# load validation set, normalize and parse into tensor
input_tensor_val, target_tensor_val = self.dataset.scale_and_parse_to_tensor(
batch_files=self.dataset.val_file_names,
scale_params_low=scale_params_low,
scale_params_high=scale_params_high,
mean_image_low=mean_image_low,
mean_image_high=mean_image_high)
if torch.cuda.is_available():
input_tensor_val = input_tensor_val.cuda()
target_tensor_val = target_tensor_val.cuda()
# activate optimizer with the base learning rate
self.model.activate_optimizer(learning_rate)
# now calculate the learning rates list
self.learning_rates = self.get_learning_rate(learning_rate, self.epochs, lr_method)
for e in range(0, self.epochs):
# setting the learning rate each epoch
lr = self.learning_rates[e]
self.model.set_learning_rate(lr)
# separate names into random batches and shuffle every epoch
self.dataset.batch_names(batch_size=self.batch_size)
# in self.batch_number is the number of batches in the training set
# go through all the batches
for i in range(self.dataset.batch_number):
input_tensor, target_tensor = self.dataset.scale_and_parse_to_tensor(
batch_files=self.dataset.train_batch_chunks[i],
scale_params_low=scale_params_low,
scale_params_high=scale_params_high,
mean_image_low=mean_image_low,
mean_image_high=mean_image_high)
if torch.cuda.is_available():
input_tensor = input_tensor.cuda()
target_tensor = target_tensor.cuda()
self.model.train_model(input_tensor, target_tensor, current_epoch=e)
# calculate the validation loss and add to validation history
self.logger.get_val_loss(val_in=input_tensor_val, val_target=target_tensor_val)
# save model every x epochs
if e % self.log_period == 0 or e == self.epochs - 1:
self.logger.log(save_appendix='_epoch_' + str(e),
current_epoch=e,
epochs=self.epochs,
mean_images=[mean_image_low, mean_image_high],
scale_params=[scale_params_low, scale_params_high],
learning_rates=self.learning_rates)
def find_lr(self, init_value=1e-8, final_value=10., beta=0.98):
"""
learning rate finder. goes through multiple learning rates and does one forward
pass with each and tracks the loss.
it returns the learning rates and losses so one can make an image of the loss from which one can find
a suitable learning rate. Pick the highest one on which the loss is still decreasing (so not the minimum)
:param train_in: training data input
:param target: training data target
:param init_value: inital value which the learning rate start with. init_value < final_value. Default: 1e-8
:param final_value: last value of the learning rate. Default: 10.
:param beta: used for smoothing the loss. Default: 0.98
:return: log_learning_rates, losses
"""
# get scale and center parameters
scale_params_low, scale_params_high = self.dataset.load_params(param_type="scale_params")
mean_image_low, mean_image_high = self.dataset.load_params(param_type="mean_images")
import math
self.dataset.batch_names(batch_size=2)
print('number of files: ', len(self.dataset.train_file_names))
num = self.dataset.batch_number-1
mult = (final_value / init_value) ** (1 / num)
lr = init_value
self.model.optimizer.param_groups[0]['lr'] = lr
avg_loss = 0.; best_loss = 0.; batch_num = 0; losses = []; log_lrs = []
# in self.batch_number is the number of batches in the training set
print('batch numbers: ', self.dataset.batch_number)
for i in range(self.dataset.batch_number):
sys.stdout.write('\r' + 'current iteration : ' + str(i))
input_tensor, target_tensor = self.dataset.scale_and_parse_to_tensor(
batch_files=self.dataset.val_file_names,
scale_params_low=scale_params_low,
scale_params_high=scale_params_high,
mean_image_low=mean_image_low,
mean_image_high=mean_image_high)
if torch.cuda.is_available():
input_tensor = input_tensor.cuda()
target_tensor = target_tensor.cuda()
batch_num += 1
# As before, get the loss for this mini-batch of inputs/outputs
self.model.optimizer.zero_grad()
print('lr: ', lr)
outputs = self.model.forward(input_tensor)
loss = self.model.criterion(outputs, target_tensor)
# Compute the smoothed loss
avg_loss = beta * avg_loss + (1 - beta) * loss.item()
smoothed_loss = avg_loss / (1 - beta ** batch_num)
# Stop if the loss is exploding
if batch_num > 1 and smoothed_loss > 4 * best_loss:
print('loss exploding')
return log_lrs, losses
# Record the best loss
if smoothed_loss < best_loss or batch_num == 1:
best_loss = smoothed_loss
# Store the values
losses.append(smoothed_loss)
log_lrs.append(math.log10(lr))
# Do the SGD step
loss.backward()
self.model.optimizer.step()
# Update the lr for the next step
lr *= mult
self.model.optimizer.param_groups[0]['lr'] = lr
# Plot the results
'''
lrs = 10 ** np.array(log_lrs)
fig, ax = plt.subplots(1)
ax.plot(lrs, losses)
ax.set_xscale('log')
#ax.set_xlim((1e-8, 100))
ax.figure.show()
ax.figure.savefig('learning_rate_finder.png')
'''
return log_lrs, losses
def get_learning_rate(self, learning_rate, epochs, method):
"""
Method creating the learning rates corresponding to the corresponding adaptive-method.
:param learning_rate: base learning rate. Used as max rate for one_cycle and cosine_annealing
:param epochs: number of epochs the model is trained
:param method: adaptive-method which should be used: standard, one_cycle, cosine_annealing
:return: learning_rates as a list
"""
lrs = []
if method == 'standard' or method is None:
lrs = [learning_rate for i in range(epochs)]
elif method == 'one_cycle':
higher_rate = learning_rate
lower_rate = 1 / 10 * higher_rate
ann_frac = min(50, int(epochs/3))
up_num = int((epochs - ann_frac) / 2)
down_num = up_num
ann_num = epochs - up_num - down_num
lr_up = np.linspace(lower_rate, higher_rate, num=up_num)
lr_down = np.linspace(higher_rate, lower_rate, num=down_num)
lr_anihilating = np.linspace(lower_rate, 0, num=ann_num)
lrs = np.append(np.append(lr_up, lr_down), lr_anihilating)
elif method=='cosine_annealing':
pass
return lrs
class CNN_skipCo_trainer(object):
def __init__(self):
self.dataset = ProcessData(train_ratio=0.9,
process_raw_data=False,
do_augment=False,
add_augment=True,
do_flip=True,
do_blur=True,
do_deform=True,
do_crop=True,
image_type='US',
get_scale_center=False,
single_sample=True)
self.model = awesomeImageTranslatorJunior.AwesomeImageTranslatorJunior(
criterion=nn.MSELoss(),
optimizer=torch.optim.Adam,
learning_rate=0.001,
weight_decay=0)
if torch.cuda.is_available():
torch.cuda.current_device()
self.model.cuda()
self.logger = Logger(model=self.model,
project_root_dir=self.dataset.project_root_dir)
self.epochs = 100
def fit(self, learning_rate, use_one_cycle=False):
# get scale and center parameters
scale_params_low, scale_params_high = self.dataset.load_params(
param_type="scale_params")
mean_image_low, mean_image_high = self.dataset.load_params(
param_type="mean_images")
# currently for one image:
'''
self.dataset.batch_names(batch_size=5)
X, Y = self.dataset.create_train_batches(self.dataset.train_batch_chunks[1])
print(X.shape)
X = X[0,:,:]
Y = Y[0,:,:]
'''
# load validation set, normalize and parse into tensor
input_tensor_val, target_tensor_val = self.dataset.scale_and_parse_to_tensor(
batch_files=self.dataset.val_file_names,
scale_params_low=scale_params_low,
scale_params_high=scale_params_high,
mean_image_low=mean_image_low,
mean_image_high=mean_image_high)
if torch.cuda.is_available():
input_tensor_val = input_tensor_val.cuda()
target_tensor_val = target_tensor_val.cuda()
if use_one_cycle:
higher_rate = learning_rate
lower_rate = 1 / 10 * higher_rate
num_epochs = self.epochs
up_num = int((num_epochs - 50) / 2)
down_num = up_num
ann_num = num_epochs - up_num - down_num
lr_up = np.linspace(lower_rate, higher_rate, num=up_num)
lr_down = np.linspace(higher_rate, lower_rate, num=down_num)
lr_anihilating = np.linspace(lower_rate, 0, num=ann_num)
learning_rates = np.append(np.append(lr_up, lr_down),
lr_anihilating)
for e in range(0, self.epochs):
if use_one_cycle:
lr = learning_rates[e]
self.model.set_learning_rate(lr)
# separate names into random batches and shuffle every epoch
self.dataset.batch_names(batch_size=32)
# in self.batch_number is the number of batches in the training set
for i in range(self.dataset.batch_number):
input_tensor, target_tensor = self.dataset.scale_and_parse_to_tensor(
batch_files=self.dataset.train_batch_chunks[i],
scale_params_low=scale_params_low,
scale_params_high=scale_params_high,
mean_image_low=mean_image_low,
mean_image_high=mean_image_high)
if torch.cuda.is_available():
input_tensor = input_tensor.cuda()
target_tensor = target_tensor.cuda()
self.model.train_model(input_tensor,
target_tensor,
current_epoch=e)
# calculate the validation loss and add to validation history
self.logger.get_val_loss(val_in=input_tensor_val,
val_target=target_tensor_val)
# save model every x epochs
if e % 25 == 0 or e == self.epochs - 1:
self.logger.log(
save_appendix='_epoch_' + str(e),
current_epoch=e,
epochs=self.epochs,
mean_images=[mean_image_low, mean_image_high],
scale_params=[scale_params_low, scale_params_high])
# how to undo the scaling:
# unscaled_X = utils.scale_and_center_reverse(scale_center_X,
# scale_params_low, mean_image_low, image_type = self.dataset.image_type)
# unscaled_Y = utils.scale_and_center_reverse(scale_center_Y, scale_params_high,
# mean_image_high, image_type=self.dataset.image_type)
def predict(self):
#self.model.predict()
# see self.dataset.X_val and self.dataset.Y_val
pass