def experiment(exp, arch, loss, double_softmax, confidence_thresh, rampup,
teacher_alpha, fix_ema, unsup_weight, cls_bal_scale,
cls_bal_scale_range, cls_balance, cls_balance_loss,
combine_batches, learning_rate, standardise_samples,
src_affine_std, src_xlat_range, src_hflip, src_intens_flip,
src_intens_scale_range, src_intens_offset_range,
src_gaussian_noise_std, tgt_affine_std, tgt_xlat_range,
tgt_hflip, tgt_intens_flip, tgt_intens_scale_range,
tgt_intens_offset_range, tgt_gaussian_noise_std, num_epochs,
batch_size, epoch_size, seed, log_file, model_file, device):
settings = locals().copy()
import os
import sys
import pickle
import cmdline_helpers
if log_file == '':
log_file = 'output_aug_log_{}.txt'.format(exp)
elif log_file == 'none':
log_file = None
if log_file is not None:
if os.path.exists(log_file):
print('Output log file {} already exists'.format(log_file))
return
use_rampup = rampup > 0
src_intens_scale_range_lower, src_intens_scale_range_upper, src_intens_offset_range_lower, src_intens_offset_range_upper = \
cmdline_helpers.intens_aug_options(src_intens_scale_range, src_intens_offset_range)
tgt_intens_scale_range_lower, tgt_intens_scale_range_upper, tgt_intens_offset_range_lower, tgt_intens_offset_range_upper = \
cmdline_helpers.intens_aug_options(tgt_intens_scale_range, tgt_intens_offset_range)
import time
import math
import numpy as np
from batchup import data_source, work_pool
import data_loaders
import standardisation
import network_architectures
import augmentation
import torch, torch.cuda
from torch import nn
from torch.nn import functional as F
import optim_weight_ema
torch_device = torch.device(device)
pool = work_pool.WorkerThreadPool(2)
n_chn = 0
if exp == 'svhn_mnist':
d_source = data_loaders.load_svhn(zero_centre=False, greyscale=True)
d_target = data_loaders.load_mnist(invert=False,
zero_centre=False,
pad32=True,
val=False)
elif exp == 'mnist_svhn':
d_source = data_loaders.load_mnist(invert=False,
zero_centre=False,
pad32=True)
d_target = data_loaders.load_svhn(zero_centre=False,
greyscale=True,
val=False)
elif exp == 'svhn_mnist_rgb':
d_source = data_loaders.load_svhn(zero_centre=False, greyscale=False)
d_target = data_loaders.load_mnist(invert=False,
zero_centre=False,
pad32=True,
val=False,
rgb=True)
elif exp == 'mnist_svhn_rgb':
d_source = data_loaders.load_mnist(invert=False,
zero_centre=False,
pad32=True,
rgb=True)
d_target = data_loaders.load_svhn(zero_centre=False,
greyscale=False,
val=False)
elif exp == 'cifar_stl':
d_source = data_loaders.load_cifar10(range_01=False)
d_target = data_loaders.load_stl(zero_centre=False, val=False)
elif exp == 'stl_cifar':
d_source = data_loaders.load_stl(zero_centre=False)
d_target = data_loaders.load_cifar10(range_01=False, val=False)
elif exp == 'mnist_usps':
d_source = data_loaders.load_mnist(zero_centre=False)
d_target = data_loaders.load_usps(zero_centre=False,
scale28=True,
val=False)
elif exp == 'usps_mnist':
d_source = data_loaders.load_usps(zero_centre=False, scale28=True)
d_target = data_loaders.load_mnist(zero_centre=False, val=False)
elif exp == 'syndigits_svhn':
d_source = data_loaders.load_syn_digits(zero_centre=False)
d_target = data_loaders.load_svhn(zero_centre=False, val=False)
elif exp == 'synsigns_gtsrb':
d_source = data_loaders.load_syn_signs(zero_centre=False)
d_target = data_loaders.load_gtsrb(zero_centre=False, val=False)
else:
print('Unknown experiment type \'{}\''.format(exp))
return
# Delete the training ground truths as we should not be using them
del d_target.train_y
if standardise_samples:
standardisation.standardise_dataset(d_source)
standardisation.standardise_dataset(d_target)
n_classes = d_source.n_classes
print('Loaded data')
if arch == '':
if exp in {'mnist_usps', 'usps_mnist'}:
arch = 'mnist-bn-32-64-256'
if exp in {'svhn_mnist', 'mnist_svhn'}:
arch = 'grey-32-64-128-gp'
if exp in {
'cifar_stl', 'stl_cifar', 'syndigits_svhn', 'svhn_mnist_rgb',
'mnist_svhn_rgb'
}:
arch = 'rgb-128-256-down-gp'
if exp in {'synsigns_gtsrb'}:
arch = 'rgb40-96-192-384-gp'
net_class, expected_shape = network_architectures.get_net_and_shape_for_architecture(
arch)
if expected_shape != d_source.train_X.shape[1:]:
print(
'Architecture {} not compatible with experiment {}; it needs samples of shape {}, '
'data has samples of shape {}'.format(arch, exp, expected_shape,
d_source.train_X.shape[1:]))
return
student_net = net_class(n_classes).to(torch_device)
teacher_net = net_class(n_classes).to(torch_device)
student_params = list(student_net.parameters())
teacher_params = list(teacher_net.parameters())
for param in teacher_params:
param.requires_grad = False
student_optimizer = torch.optim.Adam(student_params, lr=learning_rate)
if fix_ema:
teacher_optimizer = optim_weight_ema.EMAWeightOptimizer(
teacher_net, student_net, alpha=teacher_alpha)
else:
teacher_optimizer = optim_weight_ema.OldWeightEMA(teacher_net,
student_net,
alpha=teacher_alpha)
classification_criterion = nn.CrossEntropyLoss()
print('Built network')
src_aug = augmentation.ImageAugmentation(
src_hflip,
src_xlat_range,
src_affine_std,
intens_flip=src_intens_flip,
intens_scale_range_lower=src_intens_scale_range_lower,
intens_scale_range_upper=src_intens_scale_range_upper,
intens_offset_range_lower=src_intens_offset_range_lower,
intens_offset_range_upper=src_intens_offset_range_upper,
gaussian_noise_std=src_gaussian_noise_std)
tgt_aug = augmentation.ImageAugmentation(
tgt_hflip,
tgt_xlat_range,
tgt_affine_std,
intens_flip=tgt_intens_flip,
intens_scale_range_lower=tgt_intens_scale_range_lower,
intens_scale_range_upper=tgt_intens_scale_range_upper,
intens_offset_range_lower=tgt_intens_offset_range_lower,
intens_offset_range_upper=tgt_intens_offset_range_upper,
gaussian_noise_std=tgt_gaussian_noise_std)
if combine_batches:
def augment(X_sup, y_src, X_tgt):
X_src_stu, X_src_tea = src_aug.augment_pair(X_sup)
X_tgt_stu, X_tgt_tea = tgt_aug.augment_pair(X_tgt)
return X_src_stu, X_src_tea, y_src, X_tgt_stu, X_tgt_tea
else:
def augment(X_src, y_src, X_tgt):
X_src = src_aug.augment(X_src)
X_tgt_stu, X_tgt_tea = tgt_aug.augment_pair(X_tgt)
return X_src, y_src, X_tgt_stu, X_tgt_tea
rampup_weight_in_list = [0]
cls_bal_fn = network_architectures.get_cls_bal_function(cls_balance_loss)
def compute_aug_loss(stu_out, tea_out):
# Augmentation loss
if use_rampup:
unsup_mask = None
conf_mask_count = None
unsup_mask_count = None
else:
conf_tea = torch.max(tea_out, 1)[0]
unsup_mask = conf_mask = (conf_tea > confidence_thresh).float()
unsup_mask_count = conf_mask_count = conf_mask.sum()
if loss == 'bce':
aug_loss = network_architectures.robust_binary_crossentropy(
stu_out, tea_out)
else:
d_aug_loss = stu_out - tea_out
aug_loss = d_aug_loss * d_aug_loss
# Class balance scaling
if cls_bal_scale:
if use_rampup:
n_samples = float(aug_loss.shape[0])
else:
n_samples = unsup_mask.sum()
avg_pred = n_samples / float(n_classes)
bal_scale = avg_pred / torch.clamp(tea_out.sum(dim=0), min=1.0)
if cls_bal_scale_range != 0.0:
bal_scale = torch.clamp(bal_scale,
min=1.0 / cls_bal_scale_range,
max=cls_bal_scale_range)
bal_scale = bal_scale.detach()
aug_loss = aug_loss * bal_scale[None, :]
aug_loss = aug_loss.mean(dim=1)
if use_rampup:
unsup_loss = aug_loss.mean() * rampup_weight_in_list[0]
else:
unsup_loss = (aug_loss * unsup_mask).mean()
# Class balance loss
if cls_balance > 0.0:
# Compute per-sample average predicated probability
# Average over samples to get average class prediction
avg_cls_prob = stu_out.mean(dim=0)
# Compute loss
equalise_cls_loss = cls_bal_fn(avg_cls_prob,
float(1.0 / n_classes))
equalise_cls_loss = equalise_cls_loss.mean() * n_classes
if use_rampup:
equalise_cls_loss = equalise_cls_loss * rampup_weight_in_list[0]
else:
if rampup == 0:
equalise_cls_loss = equalise_cls_loss * unsup_mask.mean(
dim=0)
unsup_loss += equalise_cls_loss * cls_balance
return unsup_loss, conf_mask_count, unsup_mask_count
if combine_batches:
def f_train(X_src0, X_src1, y_src, X_tgt0, X_tgt1):
X_src0 = torch.tensor(X_src0,
dtype=torch.float,
device=torch_device)
X_src1 = torch.tensor(X_src1,
dtype=torch.float,
device=torch_device)
y_src = torch.tensor(y_src, dtype=torch.long, device=torch_device)
X_tgt0 = torch.tensor(X_tgt0,
dtype=torch.float,
device=torch_device)
X_tgt1 = torch.tensor(X_tgt1,
dtype=torch.float,
device=torch_device)
n_samples = X_src0.size()[0]
n_total = n_samples + X_tgt0.size()[0]
student_optimizer.zero_grad()
student_net.train()
teacher_net.train()
# Concatenate source and target mini-batches
X0 = torch.cat([X_src0, X_tgt0], 0)
X1 = torch.cat([X_src1, X_tgt1], 0)
student_logits_out = student_net(X0)
student_prob_out = F.softmax(student_logits_out, dim=1)
src_logits_out = student_logits_out[:n_samples]
src_prob_out = student_prob_out[:n_samples]
teacher_logits_out = teacher_net(X1)
teacher_prob_out = F.softmax(teacher_logits_out, dim=1)
# Supervised classification loss
if double_softmax:
clf_loss = classification_criterion(src_prob_out, y_src)
else:
clf_loss = classification_criterion(src_logits_out, y_src)
unsup_loss, conf_mask_count, unsup_mask_count = compute_aug_loss(
student_prob_out, teacher_prob_out)
loss_expr = clf_loss + unsup_loss * unsup_weight
loss_expr.backward()
student_optimizer.step()
teacher_optimizer.step()
outputs = [
float(clf_loss) * n_samples,
float(unsup_loss) * n_total
]
if not use_rampup:
mask_count = float(conf_mask_count) * 0.5
unsup_count = float(unsup_mask_count) * 0.5
outputs.append(mask_count)
outputs.append(unsup_count)
return tuple(outputs)
else:
def f_train(X_src, y_src, X_tgt0, X_tgt1):
X_src = torch.tensor(X_src, dtype=torch.float, device=torch_device)
y_src = torch.tensor(y_src, dtype=torch.long, device=torch_device)
X_tgt0 = torch.tensor(X_tgt0,
dtype=torch.float,
device=torch_device)
X_tgt1 = torch.tensor(X_tgt1,
dtype=torch.float,
device=torch_device)
student_optimizer.zero_grad()
student_net.train()
teacher_net.train()
src_logits_out = student_net(X_src)
student_tgt_logits_out = student_net(X_tgt0)
student_tgt_prob_out = F.softmax(student_tgt_logits_out, dim=1)
teacher_tgt_logits_out = teacher_net(X_tgt1)
teacher_tgt_prob_out = F.softmax(teacher_tgt_logits_out, dim=1)
# Supervised classification loss
if double_softmax:
clf_loss = classification_criterion(
F.softmax(src_logits_out, dim=1), y_src)
else:
clf_loss = classification_criterion(src_logits_out, y_src)
unsup_loss, conf_mask_count, unsup_mask_count = compute_aug_loss(
student_tgt_prob_out, teacher_tgt_prob_out)
loss_expr = clf_loss + unsup_loss * unsup_weight
loss_expr.backward()
student_optimizer.step()
teacher_optimizer.step()
n_samples = X_src.size()[0]
outputs = [
float(clf_loss) * n_samples,
float(unsup_loss) * n_samples
]
if not use_rampup:
mask_count = float(conf_mask_count)
unsup_count = float(unsup_mask_count)
outputs.append(mask_count)
outputs.append(unsup_count)
return tuple(outputs)
print('Compiled training function')
def f_pred_src(X_sup):
X_var = torch.tensor(X_sup, dtype=torch.float, device=torch_device)
student_net.eval()
teacher_net.eval()
return (F.softmax(student_net(X_var), dim=1).detach().cpu().numpy(),
F.softmax(teacher_net(X_var), dim=1).detach().cpu().numpy())
def f_pred_tgt(X_sup):
X_var = torch.tensor(X_sup, dtype=torch.float, device=torch_device)
student_net.eval()
teacher_net.eval()
return (F.softmax(student_net(X_var), dim=1).detach().cpu().numpy(),
F.softmax(teacher_net(X_var), dim=1).detach().cpu().numpy())
def f_eval_src(X_sup, y_sup):
y_pred_prob_stu, y_pred_prob_tea = f_pred_src(X_sup)
y_pred_stu = np.argmax(y_pred_prob_stu, axis=1)
y_pred_tea = np.argmax(y_pred_prob_tea, axis=1)
return (float(
(y_pred_stu != y_sup).sum()), float((y_pred_tea != y_sup).sum()))
def f_eval_tgt(X_sup, y_sup):
y_pred_prob_stu, y_pred_prob_tea = f_pred_tgt(X_sup)
y_pred_stu = np.argmax(y_pred_prob_stu, axis=1)
y_pred_tea = np.argmax(y_pred_prob_tea, axis=1)
return (float(
(y_pred_stu != y_sup).sum()), float((y_pred_tea != y_sup).sum()))
print('Compiled evaluation function')
# Setup output
def log(text):
print(text)
if log_file is not None:
with open(log_file, 'a') as f:
f.write(text + '\n')
f.flush()
f.close()
cmdline_helpers.ensure_containing_dir_exists(log_file)
# Report setttings
log('Settings: {}'.format(', '.join([
'{}={}'.format(key, settings[key])
for key in sorted(list(settings.keys()))
])))
# Report dataset size
log('Dataset:')
log('SOURCE Train: X.shape={}, y.shape={}'.format(d_source.train_X.shape,
d_source.train_y.shape))
log('SOURCE Test: X.shape={}, y.shape={}'.format(d_source.test_X.shape,
d_source.test_y.shape))
log('TARGET Train: X.shape={}'.format(d_target.train_X.shape))
log('TARGET Test: X.shape={}, y.shape={}'.format(d_target.test_X.shape,
d_target.test_y.shape))
print('Training...')
sup_ds = data_source.ArrayDataSource([d_source.train_X, d_source.train_y],
repeats=-1)
tgt_train_ds = data_source.ArrayDataSource([d_target.train_X], repeats=-1)
train_ds = data_source.CompositeDataSource([sup_ds,
tgt_train_ds]).map(augment)
train_ds = pool.parallel_data_source(train_ds)
if epoch_size == 'large':
n_samples = max(d_source.train_X.shape[0], d_target.train_X.shape[0])
elif epoch_size == 'small':
n_samples = min(d_source.train_X.shape[0], d_target.train_X.shape[0])
elif epoch_size == 'target':
n_samples = d_target.train_X.shape[0]
n_train_batches = n_samples // batch_size
source_test_ds = data_source.ArrayDataSource(
[d_source.test_X, d_source.test_y])
target_test_ds = data_source.ArrayDataSource(
[d_target.test_X, d_target.test_y])
if seed != 0:
shuffle_rng = np.random.RandomState(seed)
else:
shuffle_rng = np.random
train_batch_iter = train_ds.batch_iterator(batch_size=batch_size,
shuffle=shuffle_rng)
best_teacher_model_state = {
k: v.cpu().numpy()
for k, v in teacher_net.state_dict().items()
}
best_conf_mask_rate = 0.0
best_src_test_err = 1.0
for epoch in range(num_epochs):
t1 = time.time()
if use_rampup:
if epoch < rampup:
p = max(0.0, float(epoch)) / float(rampup)
p = 1.0 - p
rampup_value = math.exp(-p * p * 5.0)
else:
rampup_value = 1.0
rampup_weight_in_list[0] = rampup_value
train_res = data_source.batch_map_mean(f_train,
train_batch_iter,
n_batches=n_train_batches)
train_clf_loss = train_res[0]
if combine_batches:
unsup_loss_string = 'unsup (both) loss={:.6f}'.format(train_res[1])
else:
unsup_loss_string = 'unsup (tgt) loss={:.6f}'.format(train_res[1])
src_test_err_stu, src_test_err_tea = source_test_ds.batch_map_mean(
f_eval_src, batch_size=batch_size * 2)
tgt_test_err_stu, tgt_test_err_tea = target_test_ds.batch_map_mean(
f_eval_tgt, batch_size=batch_size * 2)
if use_rampup:
unsup_loss_string = '{}, rampup={:.3%}'.format(
unsup_loss_string, rampup_value)
if src_test_err_stu < best_src_test_err:
best_src_test_err = src_test_err_stu
best_teacher_model_state = {
k: v.cpu().numpy()
for k, v in teacher_net.state_dict().items()
}
improve = '*** '
else:
improve = ''
else:
conf_mask_rate = train_res[-2]
unsup_mask_rate = train_res[-1]
if conf_mask_rate > best_conf_mask_rate:
best_conf_mask_rate = conf_mask_rate
improve = '*** '
best_teacher_model_state = {
k: v.cpu().numpy()
for k, v in teacher_net.state_dict().items()
}
else:
improve = ''
unsup_loss_string = '{}, conf mask={:.3%}, unsup mask={:.3%}'.format(
unsup_loss_string, conf_mask_rate, unsup_mask_rate)
t2 = time.time()
log('{}Epoch {} took {:.2f}s: TRAIN clf loss={:.6f}, {}; '
'SRC TEST ERR={:.3%}, TGT TEST student err={:.3%}, TGT TEST teacher err={:.3%}'
.format(improve, epoch, t2 - t1, train_clf_loss, unsup_loss_string,
src_test_err_stu, tgt_test_err_stu, tgt_test_err_tea))
# Save network
if model_file != '':
cmdline_helpers.ensure_containing_dir_exists(model_file)
with open(model_file, 'wb') as f:
torch.save(best_teacher_model_state, f)
def experiment(exp, arch, rnd_init, img_size, confidence_thresh, teacher_alpha,
unsup_weight, cls_balance, cls_balance_loss, learning_rate,
pretrained_lr_factor, fix_layers, double_softmax, use_dropout,
src_scale_u_range, src_scale_x_range, src_scale_y_range,
src_affine_std, src_xlat_range, src_rot_std, src_hflip,
src_intens_scale_range, src_colour_rot_std, src_colour_off_std,
src_greyscale, src_cutout_prob, src_cutout_size,
tgt_scale_u_range, tgt_scale_x_range, tgt_scale_y_range,
tgt_affine_std, tgt_xlat_range, tgt_rot_std, tgt_hflip,
tgt_intens_scale_range, tgt_colour_rot_std, tgt_colour_off_std,
tgt_greyscale, tgt_cutout_prob, tgt_cutout_size, constrain_crop,
img_pad_width, num_epochs, batch_size, epoch_size, seed,
log_file, skip_epoch_eval, result_file, record_history,
model_file, hide_progress_bar, subsetsize, subsetseed, device,
num_threads):
settings = locals().copy()
if rnd_init:
if fix_layers != '':
print('`rnd_init` and `fix_layers` are mutually exclusive')
return
if epoch_size not in {'source', 'target'}:
try:
epoch_size = int(epoch_size)
except ValueError:
print(
'epoch_size should be an integer, \'source\', or \'target\', not {}'
.format(epoch_size))
return
import os
import sys
import pickle
import cmdline_helpers
fix_layers = [lyr.strip() for lyr in fix_layers.split(',')]
if log_file == '':
log_file = 'output_aug_log_{}.txt'.format(exp)
elif log_file == 'none':
log_file = None
if log_file is not None:
if os.path.exists(log_file):
print('Output log file {} already exists'.format(log_file))
return
src_intens_scale_range_lower, src_intens_scale_range_upper = cmdline_helpers.colon_separated_range(
src_intens_scale_range)
tgt_intens_scale_range_lower, tgt_intens_scale_range_upper = cmdline_helpers.colon_separated_range(
tgt_intens_scale_range)
src_scale_u_range = cmdline_helpers.colon_separated_range(
src_scale_u_range)
tgt_scale_u_range = cmdline_helpers.colon_separated_range(
tgt_scale_u_range)
src_scale_x_range = cmdline_helpers.colon_separated_range(
src_scale_x_range)
tgt_scale_x_range = cmdline_helpers.colon_separated_range(
tgt_scale_x_range)
src_scale_y_range = cmdline_helpers.colon_separated_range(
src_scale_y_range)
tgt_scale_y_range = cmdline_helpers.colon_separated_range(
tgt_scale_y_range)
import time
import tqdm
import math
import tables
import numpy as np
from batchup import data_source, work_pool
import image_dataset, visda17_dataset, office_dataset
import network_architectures
import augmentation
import image_transforms
from sklearn.model_selection import StratifiedShuffleSplit, ShuffleSplit
import torch, torch.cuda
from torch import nn
from torch.nn import functional as F
import optim_weight_ema
if hide_progress_bar:
progress_bar = None
else:
progress_bar = tqdm.tqdm
with torch.cuda.device(device):
pool = work_pool.WorkerThreadPool(num_threads)
n_chn = 0
half_batch_size = batch_size // 2
if arch == '':
if exp in {'train_val', 'train_test'}:
arch = 'resnet50'
if rnd_init:
mean_value = np.array([0.5, 0.5, 0.5])
std_value = np.array([0.5, 0.5, 0.5])
else:
mean_value = np.array([0.485, 0.456, 0.406])
std_value = np.array([0.229, 0.224, 0.225])
img_shape = (img_size, img_size)
img_padding = (img_pad_width, img_pad_width)
if exp == 'visda_train_val':
d_source = visda17_dataset.TrainDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = visda17_dataset.ValidationDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'visda_train_test':
d_source = visda17_dataset.TrainDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = visda17_dataset.TestDataset(img_size=img_shape,
range01=True,
rgb_order=True)
if not skip_epoch_eval:
print('WARNING: setting skip_epoch_eval to True')
skip_epoch_eval = True
elif exp == 'office_amazon_dslr':
d_source = office_dataset.OfficeAmazonDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeDSLRDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_amazon_webcam':
d_source = office_dataset.OfficeAmazonDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeWebcamDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_dslr_amazon':
d_source = office_dataset.OfficeDSLRDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeAmazonDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_dslr_webcam':
d_source = office_dataset.OfficeDSLRDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeWebcamDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_webcam_amazon':
d_source = office_dataset.OfficeWebcamDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeAmazonDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_webcam_dslr':
d_source = office_dataset.OfficeWebcamDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeDSLRDataset(img_size=img_shape,
range01=True,
rgb_order=True)
else:
print('Unknown experiment type \'{}\''.format(exp))
return
# Tensorboard log
# Subset
source_indices, target_indices, n_src, n_tgt = image_dataset.subset_indices(
d_source, d_target, subsetsize, subsetseed)
#
# Result file
#
if result_file != '':
cmdline_helpers.ensure_containing_dir_exists(result_file)
h5_filters = tables.Filters(complevel=9, complib='blosc')
f_target_pred = tables.open_file(result_file, mode='w')
g_tgt_pred = f_target_pred.create_group(f_target_pred.root,
'target_pred_y',
'Target prediction')
if record_history:
arr_tgt_pred_history = f_target_pred.create_earray(
g_tgt_pred,
'y_prob_history',
tables.Float32Atom(), (0, n_tgt, d_target.n_classes),
filters=h5_filters)
else:
arr_tgt_pred_history = None
else:
arr_tgt_pred_history = None
f_target_pred = None
g_tgt_pred = None
n_classes = d_source.n_classes
print('Loaded data')
net_class = network_architectures.get_build_fn_for_architecture(arch)
student_net = net_class(n_classes, img_size, use_dropout,
not rnd_init).cuda()
teacher_net = net_class(n_classes, img_size, use_dropout,
not rnd_init).cuda()
student_params = list(student_net.parameters())
teacher_params = list(teacher_net.parameters())
for param in teacher_params:
param.requires_grad = False
if rnd_init:
new_student_optimizer = torch.optim.Adam(student_params,
lr=learning_rate)
pretrained_student_optimizer = None
else:
named_params = list(student_net.named_parameters())
new_params = []
pretrained_params = []
for name, param in named_params:
if name.startswith('new_'):
new_params.append(param)
else:
fix = False
for lyr in fix_layers:
if name.startswith(lyr + '.'):
fix = True
break
if not fix:
pretrained_params.append(param)
else:
print('Fixing param {}'.format(name))
param.requires_grad = False
new_student_optimizer = torch.optim.Adam(new_params,
lr=learning_rate)
if len(pretrained_params) > 0:
pretrained_student_optimizer = torch.optim.Adam(
pretrained_params, lr=learning_rate * pretrained_lr_factor)
else:
pretrained_student_optimizer = None
teacher_optimizer = optim_weight_ema.WeightEMA(teacher_params,
student_params,
alpha=teacher_alpha)
classification_criterion = nn.CrossEntropyLoss()
print('Built network')
# Image augmentation
src_aug = augmentation.ImageAugmentation(
src_hflip,
src_xlat_range,
src_affine_std,
rot_std=src_rot_std,
intens_scale_range_lower=src_intens_scale_range_lower,
intens_scale_range_upper=src_intens_scale_range_upper,
colour_rot_std=src_colour_rot_std,
colour_off_std=src_colour_off_std,
greyscale=src_greyscale,
scale_u_range=src_scale_u_range,
scale_x_range=src_scale_x_range,
scale_y_range=src_scale_y_range,
cutout_probability=src_cutout_prob,
cutout_size=src_cutout_size)
tgt_aug = augmentation.ImageAugmentation(
tgt_hflip,
tgt_xlat_range,
tgt_affine_std,
rot_std=tgt_rot_std,
intens_scale_range_lower=tgt_intens_scale_range_lower,
intens_scale_range_upper=tgt_intens_scale_range_upper,
colour_rot_std=tgt_colour_rot_std,
colour_off_std=tgt_colour_off_std,
greyscale=tgt_greyscale,
scale_u_range=tgt_scale_u_range,
scale_x_range=tgt_scale_x_range,
scale_y_range=tgt_scale_y_range,
cutout_probability=tgt_cutout_prob,
cutout_size=tgt_cutout_size)
test_aug = augmentation.ImageAugmentation(
tgt_hflip,
tgt_xlat_range,
0.0,
rot_std=0.0,
scale_u_range=tgt_scale_u_range,
scale_x_range=tgt_scale_x_range,
scale_y_range=tgt_scale_y_range)
border_value = int(np.mean(mean_value) * 255 + 0.5)
sup_xf = image_transforms.Compose(
image_transforms.ScaleCropAndAugmentAffine(img_shape, img_padding,
True, src_aug,
border_value,
mean_value, std_value),
image_transforms.ToTensor(),
)
if constrain_crop >= 0:
unsup_xf = image_transforms.Compose(
image_transforms.ScaleCropAndAugmentAffinePair(
img_shape, img_padding, constrain_crop, True, tgt_aug,
border_value, mean_value, std_value),
image_transforms.ToTensor(),
)
else:
unsup_xf = image_transforms.Compose(
image_transforms.ScaleCropAndAugmentAffine(
img_shape, img_padding, True, tgt_aug, border_value,
mean_value, std_value),
image_transforms.ToTensor(),
)
test_xf = image_transforms.Compose(
image_transforms.ScaleAndCrop(img_shape, img_padding, False),
image_transforms.ToTensor(),
image_transforms.Standardise(mean_value, std_value),
)
test_xf_aug_mult = image_transforms.Compose(
image_transforms.ScaleCropAndAugmentAffineMultiple(
16, img_shape, img_padding, True, test_aug, border_value,
mean_value, std_value),
image_transforms.ToTensorMultiple(),
)
if constrain_crop >= 0:
def augment(X_sup, y_sup, X_tgt):
X_sup = sup_xf(X_sup)[0]
X_unsup_both = unsup_xf(X_tgt)[0]
X_unsup_stu = X_unsup_both[:len(X_tgt)]
X_unsup_tea = X_unsup_both[len(X_tgt):]
return X_sup, y_sup, X_unsup_stu, X_unsup_tea
else:
def augment(X_sup, y_sup, X_tgt):
X_sup = sup_xf(X_sup)[0]
X_unsup_stu = unsup_xf(X_tgt)[0]
X_unsup_tea = unsup_xf(X_tgt)[0]
return X_sup, y_sup, X_unsup_stu, X_unsup_tea
cls_bal_fn = network_architectures.get_cls_bal_function(
cls_balance_loss)
def compute_aug_loss(stu_out, tea_out):
# Augmentation loss
conf_tea = torch.max(tea_out, 1)[0]
conf_mask = torch.gt(conf_tea, confidence_thresh).float()
d_aug_loss = stu_out - tea_out
aug_loss = d_aug_loss * d_aug_loss
aug_loss = torch.mean(aug_loss, 1) * conf_mask
# Class balance loss
if cls_balance > 0.0:
# Average over samples to get average class prediction
avg_cls_prob = torch.mean(stu_out, 0)
# Compute loss
equalise_cls_loss = cls_bal_fn(avg_cls_prob,
float(1.0 / n_classes))
equalise_cls_loss = torch.mean(equalise_cls_loss) * n_classes
equalise_cls_loss = equalise_cls_loss * torch.mean(
conf_mask, 0)
else:
equalise_cls_loss = None
return aug_loss, conf_mask, equalise_cls_loss
_one = torch.autograd.Variable(
torch.from_numpy(np.array([1.0]).astype(np.float32)).cuda())
def f_train(X_sup, y_sup, X_unsup0, X_unsup1):
X_sup = torch.autograd.Variable(torch.from_numpy(X_sup).cuda())
y_sup = torch.autograd.Variable(
torch.from_numpy(y_sup).long().cuda())
X_unsup0 = torch.autograd.Variable(
torch.from_numpy(X_unsup0).cuda())
X_unsup1 = torch.autograd.Variable(
torch.from_numpy(X_unsup1).cuda())
if pretrained_student_optimizer is not None:
pretrained_student_optimizer.zero_grad()
new_student_optimizer.zero_grad()
student_net.train(mode=True)
teacher_net.train(mode=True)
sup_logits_out = student_net(X_sup)
student_unsup_logits_out = student_net(X_unsup0)
student_unsup_prob_out = F.softmax(student_unsup_logits_out)
teacher_unsup_logits_out = teacher_net(X_unsup1)
teacher_unsup_prob_out = F.softmax(teacher_unsup_logits_out)
# Supervised classification loss
if double_softmax:
clf_loss = classification_criterion(F.softmax(sup_logits_out),
y_sup)
else:
clf_loss = classification_criterion(sup_logits_out, y_sup)
aug_loss, conf_mask, cls_bal_loss = compute_aug_loss(
student_unsup_prob_out, teacher_unsup_prob_out)
conf_mask_count = torch.sum(conf_mask)
unsup_loss = torch.mean(aug_loss)
loss_expr = clf_loss + unsup_loss * unsup_weight
if cls_bal_loss is not None:
loss_expr = loss_expr + cls_bal_loss * cls_balance * unsup_weight
loss_expr.backward()
if pretrained_student_optimizer is not None:
pretrained_student_optimizer.step()
new_student_optimizer.step()
teacher_optimizer.step()
n_samples = X_sup.size()[0]
mask_count = conf_mask_count.data.cpu()[0]
outputs = [
float(clf_loss.data.cpu()[0]) * n_samples,
float(unsup_loss.data.cpu()[0]) * n_samples, mask_count
]
return tuple(outputs)
print('Compiled training function')
def f_pred_src(X_sup):
X_var = torch.autograd.Variable(torch.from_numpy(X_sup).cuda())
teacher_net.train(mode=False)
return (F.softmax(teacher_net(X_var)).data.cpu().numpy(), )
def f_pred_tgt(X_sup):
X_var = torch.autograd.Variable(torch.from_numpy(X_sup).cuda())
teacher_net.train(mode=False)
return (F.softmax(teacher_net(X_var)).data.cpu().numpy(), )
def f_pred_tgt_mult(X_sup):
teacher_net.train(mode=False)
y_pred_aug = []
for aug_i in range(len(X_sup)):
X_var = torch.autograd.Variable(
torch.from_numpy(X_sup[aug_i, ...]).cuda())
y_pred = F.softmax(teacher_net(X_var)).data.cpu().numpy()
y_pred_aug.append(y_pred[None, ...])
y_pred_aug = np.concatenate(y_pred_aug, axis=0)
return (y_pred_aug.mean(axis=0), )
print('Compiled evaluation function')
# Setup output
cmdline_helpers.ensure_containing_dir_exists(log_file)
def log(text):
print(text)
if log_file is not None:
with open(log_file, 'a') as f:
f.write(text + '\n')
f.flush()
f.close()
# Report setttings
log('Program = {}'.format(sys.argv[0]))
log('Settings: {}'.format(', '.join([
'{}={}'.format(key, settings[key])
for key in sorted(list(settings.keys()))
])))
# Report dataset size
log('Dataset:')
log('SOURCE len(X)={}, y.shape={}'.format(len(d_source.images),
d_source.y.shape))
log('TARGET len(X)={}'.format(len(d_target.images)))
if epoch_size == 'source':
n_samples = n_src
elif epoch_size == 'target':
n_samples = n_tgt
else:
n_samples = epoch_size
n_train_batches = n_samples // batch_size
n_test_batches = n_tgt // (batch_size * 2) + 1
print('Training...')
sup_ds = data_source.ArrayDataSource([d_source.images, d_source.y],
repeats=-1,
indices=source_indices)
tgt_train_ds = data_source.ArrayDataSource([d_target.images],
repeats=-1,
indices=target_indices)
train_ds = data_source.CompositeDataSource([sup_ds,
tgt_train_ds]).map(augment)
train_ds = pool.parallel_data_source(train_ds,
batch_buffer_size=min(
20, n_train_batches))
target_ds_for_test = data_source.ArrayDataSource(
[d_target.images], indices=target_indices)
target_test_ds = target_ds_for_test.map(test_xf)
target_test_ds = pool.parallel_data_source(target_test_ds,
batch_buffer_size=min(
20, n_test_batches))
target_mult_test_ds = target_ds_for_test.map(test_xf_aug_mult)
target_mult_test_ds = pool.parallel_data_source(target_mult_test_ds,
batch_buffer_size=min(
20,
n_test_batches))
if seed != 0:
shuffle_rng = np.random.RandomState(seed)
else:
shuffle_rng = np.random
if d_target.has_ground_truth:
evaluator = d_target.prediction_evaluator(target_indices)
else:
evaluator = None
best_mask_rate = 0.0
best_teacher_model_state = {
k: v.cpu().numpy()
for k, v in teacher_net.state_dict().items()
}
train_batch_iter = train_ds.batch_iterator(batch_size=batch_size,
shuffle=shuffle_rng)
for epoch in range(num_epochs):
t1 = time.time()
if not skip_epoch_eval:
test_batch_iter = target_test_ds.batch_iterator(
batch_size=batch_size * 2)
else:
test_batch_iter = None
train_clf_loss, train_unsup_loss, mask_rate = data_source.batch_map_mean(
f_train,
train_batch_iter,
progress_iter_func=progress_bar,
n_batches=n_train_batches)
# train_clf_loss, train_unsup_loss, mask_rate = train_ds.batch_map_mean(
# f_train, batch_size=batch_size, shuffle=shuffle_rng, n_batches=n_train_batches,
# progress_iter_func=progress_bar)
if mask_rate > best_mask_rate:
best_mask_rate = mask_rate
improve = True
improve_str = '*** '
best_teacher_model_state = {
k: v.cpu().numpy()
for k, v in teacher_net.state_dict().items()
}
else:
improve = False
improve_str = ''
if not skip_epoch_eval:
tgt_pred_prob_y, = data_source.batch_map_concat(
f_pred_tgt,
test_batch_iter,
progress_iter_func=progress_bar)
mean_class_acc, cls_acc_str = evaluator.evaluate(
tgt_pred_prob_y)
t2 = time.time()
log('{}Epoch {} took {:.2f}s: TRAIN clf loss={:.6f}, unsup loss={:.6f}, mask={:.3%}; '
'TGT mean class acc={:.3%}'.format(improve_str, epoch,
t2 - t1, train_clf_loss,
train_unsup_loss,
mask_rate,
mean_class_acc))
log(' per class: {}'.format(cls_acc_str))
# Save results
if arr_tgt_pred_history is not None:
arr_tgt_pred_history.append(
tgt_pred_prob_y[None, ...].astype(np.float32))
else:
t2 = time.time()
log('{}Epoch {} took {:.2f}s: TRAIN clf loss={:.6f}, unsup loss={:.6f}, mask={:.3%}'
.format(improve_str, epoch, t2 - t1, train_clf_loss,
train_unsup_loss, mask_rate))
# Save network
if model_file != '':
cmdline_helpers.ensure_containing_dir_exists(model_file)
with open(model_file, 'wb') as f:
pickle.dump(best_teacher_model_state, f)
# Restore network to best state
teacher_net.load_state_dict({
k: torch.from_numpy(v)
for k, v in best_teacher_model_state.items()
})
# Predict on test set, without augmentation
tgt_pred_prob_y, = target_test_ds.batch_map_concat(
f_pred_tgt, batch_size=batch_size, progress_iter_func=progress_bar)
if d_target.has_ground_truth:
mean_class_acc, cls_acc_str = evaluator.evaluate(tgt_pred_prob_y)
log('FINAL: TGT mean class acc={:.3%}'.format(mean_class_acc))
log(' per class: {}'.format(cls_acc_str))
# Predict on test set, using augmentation
tgt_aug_pred_prob_y, = target_mult_test_ds.batch_map_concat(
f_pred_tgt_mult,
batch_size=batch_size,
progress_iter_func=progress_bar)
if d_target.has_ground_truth:
aug_mean_class_acc, aug_cls_acc_str = evaluator.evaluate(
tgt_aug_pred_prob_y)
log('FINAL: TGT AUG mean class acc={:.3%}'.format(
aug_mean_class_acc))
log(' per class: {}'.format(aug_cls_acc_str))
if f_target_pred is not None:
f_target_pred.create_array(g_tgt_pred, 'y_prob', tgt_pred_prob_y)
f_target_pred.create_array(g_tgt_pred, 'y_prob_aug',
tgt_aug_pred_prob_y)
f_target_pred.close()
def experiment(args):
exp = args.exp
# learning_rate = args.learning_rate
# num_epochs = args.num_epochs
batch_size = args.batch_size
epoch_size = args.epoch_size = 'dataplots'
seed = args.seed
log_file = args.log_file
workers = args.workers
image_size = args.image_size
nc = args.nc
# nz = args.nz
use_gpu = True
gpus = args.gpus
img_dir = args.img_dir
logs_dir = args.logs_dir
# ckpt_dir = args.ckpt_dir
# plot_interval = args.plot_interval
experiment = args.exp
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
# Some variable hardcoding - Only for development
# seed = 1126
seed = args.seed
simul_train_src_tgt = True
# use_scheduler = args.use_scheduler # True/False
use_sampler = True # True/False
machinename = socket.gethostname()
# hostname = timestamp = datetime.datetime.now().strftime("%d_%b_%Y_%H_%M_%S")
# hostname = timestamp = datetime.datetime.now().strftime("%Y_%b_%d_%H_%M_%S")
hostname = timestamp = datetime.datetime.now().strftime(
"%y_%m_%d_%H_%M_%S")
absolute_pyfile_path = os.path.abspath(sys.argv[0])
args.absolute_pyfile_path = os.path.abspath(sys.argv[0])
absolute_base_path = os.path.dirname(absolute_pyfile_path)
args.absolute_base_path = os.path.dirname(absolute_pyfile_path)
args.dataroot = os.path.expanduser(args.dataroot)
dataroot = os.path.join(absolute_base_path, args.dataroot)
dataset_path = os.path.join(absolute_base_path, args.dataroot,
args.dataset)
args.dataroot = os.path.join(absolute_base_path, args.dataroot)
args.dataset_path = os.path.join(absolute_base_path, args.dataroot,
args.dataset)
args.logs_dir = os.path.expanduser(args.logs_dir)
# logs_dir = args.logs_dir
logs_dir = os.path.join(absolute_base_path, args.logs_dir)
args.logs_dir = logs_dir
log_num = 0
log_file = logs_dir + '/' + hostname + '_' + machinename + '_' + args.exp + '_' + str(
args.seed) + '.txt'
# Setup logfile to store output logs
if log_file is not None:
while os.path.exists(log_file):
log_num += 1
log_file = '{0}/{3}_{1}_{2}_{4}.txt'.format(
args.logs_dir, 'datasplots', log_num, hostname,
args.epoch_size)
# return
args.log_file = log_file
args.img_dir = os.path.expanduser(args.img_dir)
img_dir = os.path.join(absolute_base_path, args.img_dir,
args.exp) + '_' + str(log_num)
args.img_dir = img_dir
# Create folder to save log files
os.system(
'mkdir -p {0}'.format(logs_dir)) # create folder to store log files
settings = locals().copy()
# net.writer = SummaryWriter(comment='_' + args.exp + '_' + args.epoch_size)
#
# ckpt_dir = os.path.join(absolute_base_path, net.writer.log_dir)
# args.ckpt_dir = ckpt_dir
# Setup output
def log(text):
# timestamp = datetime.datetime.now().strftime("%d_%b_%Y_%H_%M_%S")
# timestamp = datetime.datetime.now().strftime("%Y_%b_%d_%H_%M_%S")
timestamp = datetime.datetime.now().strftime("%y_%m_%d_%H_%M_%S")
# print('[' + timestamp + '] ' + text)
print(text)
if log_file is not None:
with open(log_file, 'a') as f:
# f.write('[' + timestamp + '] ' + text + '\n')
str = text + '\n'
f.write(text + '\n')
f.flush()
f.close()
return str
cmdline_helpers.ensure_containing_dir_exists(log_file)
n_iter = 0
log_str = ''
log_str += log('\n')
log_str += log('Output log file {0} created'.format(log_file))
log_str += log(
'File used to run the experiment : {0}'.format(absolute_pyfile_path))
log_str += log(
'Output image files are stored in {0} directory'.format(img_dir))
# Report setttings
log_str += log('Settings: {}'.format(', '.join([
'{}={}'.format(key, settings[key])
for key in sorted(list(settings.keys()))
])))
num_gpu = len(args.gpus.split(','))
log_str += log('num_gpu: {0}, GPU-ID: {1}'.format(num_gpu, args.gpus))
random.seed(seed)
np.random.seed(seed=seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
cudnn.benchmark = True
# if num_gpu < 4:
# os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
dataset_list = [
'usps',
'mnist',
'svhn',
'syndigits',
'cifar9',
'stl9',
'gtsrb',
'synsigns',
'amazon',
'dslr',
'webcam',
# 'cifar10', 'cifar100',
# 'fmnist', 'emnist',
# 'lsun', 'imagenet',
# 'folder', 'lfw',
]
transform = None
# Get datasets
if exp == 'usps_mnist':
args.nc = 1
args.image_size = 28
args.dataset = 'usps'
d_src_train, d_src_test = get_data(args,
transform=transform,
train_flag=True)
args.dataset = 'mnist'
d_tgt_train, d_tgt_test = get_data(args,
transform=transform,
train_flag=False)
elif exp == 'mnist_usps':
args.nc = 1
args.image_size = 28
args.dataset = 'mnist'
d_src_train, d_src_test = get_data(args,
transform=transform,
train_flag=True)
args.dataset = 'usps'
d_tgt_train, d_tgt_test = get_data(args,
transform=transform,
train_flag=False)
elif exp == 'svhn_mnist':
# args.nc = 1
# args.image_size = 28
if args.network_type == 'dade':
args.nc = 1
args.image_size = 28
else:
args.nc = 3
args.image_size = 32
args.dataset = 'svhn'
d_src_train, d_src_test = get_data(args,
transform=transform,
train_flag=True)
args.dataset = 'mnist'
d_tgt_train, d_tgt_test = get_data(args,
transform=transform,
train_flag=False)
elif exp == 'mnist_svhn':
# args.nc = 1
# args.image_size = 32
# args.image_size = 28
if args.network_type == 'dade':
args.nc = 1
args.image_size = 28
else:
args.nc = 3
args.image_size = 32
args.dataset = 'mnist'
d_src_train_notinverted, _ = get_data(args,
transform=transform,
train_flag=False)
d_src_train_inverted, d_src_test = get_data(args,
transform=transform,
train_flag=True)
d_src_train = torch.utils.data.ConcatDataset(
[d_src_train_inverted, d_src_train_notinverted])
d_src_train.dataset_name = 'mnist'
d_src_train.n_classes = 10
d_src_train.transform = d_src_train_notinverted.transform
args.dataset = 'svhn'
d_tgt_train, d_tgt_test = get_data(args,
transform=transform,
train_flag=False)
elif exp == 'cifar_stl':
args.nc = 3
args.image_size = 32
args.dataset = 'cifar9'
d_src_train, d_src_test = get_data(args,
transform=transform,
train_flag=True)
args.dataset = 'stl9'
d_tgt_train, d_tgt_test = get_data(args,
transform=transform,
train_flag=False)
elif exp == 'stl_cifar':
args.nc = 3
args.image_size = 32
args.dataset = 'stl9'
d_src_train, d_src_test = get_data(args,
transform=transform,
train_flag=True)
args.dataset = 'cifar9'
d_tgt_train, d_tgt_test = get_data(args,
transform=transform,
train_flag=False)
elif exp == 'syndigits_svhn':
# args.nc = 1
# # args.image_size = 28
# args.image_size = 32
args.nc = 3
args.image_size = 32
args.dataset = 'syndigits'
d_src_train, d_src_test = get_data(args,
transform=transform,
train_flag=True)
args.dataset = 'svhn'
d_tgt_train, d_tgt_test = get_data(args,
transform=transform,
train_flag=False)
elif exp == 'svhn_syndigits':
# args.nc = 1
# args.image_size = 28
args.nc = 3
args.image_size = 32
args.dataset = 'svhn'
d_src_train, d_src_test = get_data(args,
transform=transform,
train_flag=True)
args.dataset = 'syndigits'
d_tgt_train, d_tgt_test = get_data(args,
transform=transform,
train_flag=False)
elif exp == 'synsigns_gtsrb':
args.nc = 3
args.image_size = 40
# args.nc = 3
# args.image_size = 32
args.dataset = 'synsigns'
d_src_train, d_src_test = get_data(args,
transform=transform,
train_flag=True)
args.dataset = 'gtsrb'
d_tgt_train, d_tgt_test = get_data(args,
transform=transform,
train_flag=False)
elif exp in {
'amazon_dslr', 'amazon_webcam', 'dslr_amazon', 'dslr_webcam',
'webcam_amazon', 'webcam_dslr'
}:
args.nc = 3
args.image_size = 224
args.dataset = exp.split('_')[0]
d_src_train, d_src_test = get_data(args,
transform=transform,
train_flag=True)
args.dataset = exp.split('_')[1]
d_tgt_train, d_tgt_test = get_data(args,
transform=transform,
train_flag=False)
else:
log_str += log('Error : Unknown experiment type \'{}\''.format(exp))
raise ValueError('Unknown experiment type \'{}\''.format(exp))
log_str += log(
'\nSRC : {}: train: count={}, X.shape={} test: count={}, X.shape={}'.
format(d_src_train.dataset_name.upper(),
d_src_train.__len__(), d_src_train[0][0].shape,
d_src_test.__len__(), d_src_test[0][0].shape))
log_str += log(
'TGT : {}: train: count={}, X.shape={} test: count={}, X.shape={}'.
format(d_tgt_train.dataset_name.upper(),
d_tgt_train.__len__(), d_tgt_train[0][0].shape,
d_tgt_test.__len__(), d_tgt_test[0][0].shape))
n_classes = d_src_train.n_classes
log_str += log('\nTransformations for SRC and TGT datasets ...')
log_str += log('SRC : {0} - transformation : {1}'.format(
d_src_train.dataset_name.upper(), d_src_train.transform))
log_str += log('TGT : {0} - transformation : {1}'.format(
d_tgt_train.dataset_name.upper(), d_tgt_train.transform))
log_str += log('\nNumber of classes : {}'.format(n_classes))
log_str += log('\nLoaded Source and Target data respectively')
if exp == 'synsigns_gtsrb':
image_grid_n_row = 10
elif exp in {
'amazon_dslr', 'amazon_webcam', 'dslr_amazon', 'dslr_webcam',
'webcam_amazon', 'webcam_dslr'
}:
image_grid_n_row = 8
else:
image_grid_n_row = 6
image_grid = torch.zeros(
(n_classes * 2, args.nc, args.image_size, args.image_size),
requires_grad=False)
try:
d_src_train_labels = d_src_train.train_labels.numpy()
except:
d_src_train = d_src_train_notinverted
d_src_train_labels = d_src_train.train_labels.numpy()
labels, src_indices = np.unique(d_src_train_labels, return_index=True)
d_tgt_train_labels = d_tgt_train.train_labels.numpy()
labels, tgt_indices = np.unique(d_tgt_train_labels, return_index=True)
for i, (src_index, tgt_index) in enumerate(
zip(np.nditer(src_indices), np.nditer(tgt_indices))):
image_grid[2 * i, :, :, :] = d_src_train[src_index][0]
image_grid[2 * i + 1, :, :, :] = d_tgt_train[tgt_index][0]
image_grid.data.mul_(0.5).add_(0.5)
img_grid = vutils.make_grid(
image_grid,
nrow=image_grid_n_row) # make an square grid of images and plot
vutils.save_image(img_grid, '{}.png'.format(log_file[:-4]))
def experiment(exp, arch, learning_rate, standardise_samples, affine_std,
xlat_range, hflip, intens_flip, intens_scale_range,
intens_offset_range, gaussian_noise_std, num_epochs, batch_size,
seed, log_file, device):
import os
import sys
import cmdline_helpers
if log_file == '':
log_file = 'output_aug_log_{}.txt'.format(exp)
elif log_file == 'none':
log_file = None
if log_file is not None:
if os.path.exists(log_file):
print('Output log file {} already exists'.format(log_file))
return
intens_scale_range_lower, intens_scale_range_upper, intens_offset_range_lower, intens_offset_range_upper = \
cmdline_helpers.intens_aug_options(intens_scale_range, intens_offset_range)
import time
import math
import numpy as np
from batchup import data_source, work_pool
import data_loaders
import standardisation
import network_architectures
import augmentation
import torch, torch.cuda
from torch import nn
from torch.nn import functional as F
with torch.cuda.device(device):
pool = work_pool.WorkerThreadPool(2)
n_chn = 0
if exp == 'svhn_mnist':
d_source = data_loaders.load_svhn(zero_centre=False,
greyscale=True)
d_target = data_loaders.load_mnist(invert=False,
zero_centre=False,
pad32=True,
val=False)
elif exp == 'mnist_svhn':
d_source = data_loaders.load_mnist(invert=False,
zero_centre=False,
pad32=True)
d_target = data_loaders.load_svhn(zero_centre=False,
greyscale=True,
val=False)
elif exp == 'svhn_mnist_rgb':
d_source = data_loaders.load_svhn(zero_centre=False,
greyscale=False)
d_target = data_loaders.load_mnist(invert=False,
zero_centre=False,
pad32=True,
val=False,
rgb=True)
elif exp == 'mnist_svhn_rgb':
d_source = data_loaders.load_mnist(invert=False,
zero_centre=False,
pad32=True,
rgb=True)
d_target = data_loaders.load_svhn(zero_centre=False,
greyscale=False,
val=False)
elif exp == 'cifar_stl':
d_source = data_loaders.load_cifar10(range_01=False)
d_target = data_loaders.load_stl(zero_centre=False, val=False)
elif exp == 'stl_cifar':
d_source = data_loaders.load_stl(zero_centre=False)
d_target = data_loaders.load_cifar10(range_01=False, val=False)
elif exp == 'mnist_usps':
d_source = data_loaders.load_mnist(zero_centre=False)
d_target = data_loaders.load_usps(zero_centre=False,
scale28=True,
val=False)
elif exp == 'usps_mnist':
d_source = data_loaders.load_usps(zero_centre=False, scale28=True)
d_target = data_loaders.load_mnist(zero_centre=False, val=False)
elif exp == 'syndigits_svhn':
d_source = data_loaders.load_syn_digits(zero_centre=False)
d_target = data_loaders.load_svhn(zero_centre=False, val=False)
elif exp == 'svhn_syndigits':
d_source = data_loaders.load_svhn(zero_centre=False, val=False)
d_target = data_loaders.load_syn_digits(zero_centre=False)
elif exp == 'synsigns_gtsrb':
d_source = data_loaders.load_syn_signs(zero_centre=False)
d_target = data_loaders.load_gtsrb(zero_centre=False, val=False)
elif exp == 'gtsrb_synsigns':
d_source = data_loaders.load_gtsrb(zero_centre=False, val=False)
d_target = data_loaders.load_syn_signs(zero_centre=False)
else:
print('Unknown experiment type \'{}\''.format(exp))
return
# Delete the training ground truths as we should not be using them
del d_target.train_y
if standardise_samples:
standardisation.standardise_dataset(d_source)
standardisation.standardise_dataset(d_target)
n_classes = d_source.n_classes
print('Loaded data')
if arch == '':
if exp in {'mnist_usps', 'usps_mnist'}:
arch = 'mnist-bn-32-64-256'
if exp in {'svhn_mnist', 'mnist_svhn'}:
arch = 'grey-32-64-128-gp'
if exp in {
'cifar_stl', 'stl_cifar', 'syndigits_svhn',
'svhn_syndigits', 'svhn_mnist_rgb', 'mnist_svhn_rgb'
}:
arch = 'rgb-48-96-192-gp'
if exp in {'synsigns_gtsrb', 'gtsrb_synsigns'}:
arch = 'rgb40-48-96-192-384-gp'
net_class, expected_shape = network_architectures.get_net_and_shape_for_architecture(
arch)
if expected_shape != d_source.train_X.shape[1:]:
print(
'Architecture {} not compatible with experiment {}; it needs samples of shape {}, '
'data has samples of shape {}'.format(
arch, exp, expected_shape, d_source.train_X.shape[1:]))
return
net = net_class(n_classes).cuda()
params = list(net.parameters())
optimizer = torch.optim.Adam(params, lr=learning_rate)
classification_criterion = nn.CrossEntropyLoss()
print('Built network')
aug = augmentation.ImageAugmentation(
hflip,
xlat_range,
affine_std,
intens_scale_range_lower=intens_scale_range_lower,
intens_scale_range_upper=intens_scale_range_upper,
intens_offset_range_lower=intens_offset_range_lower,
intens_offset_range_upper=intens_offset_range_upper,
intens_flip=intens_flip,
gaussian_noise_std=gaussian_noise_std)
def augment(X_sup, y_sup):
X_sup = aug.augment(X_sup)
return [X_sup, y_sup]
def f_train(X_sup, y_sup):
X_sup = torch.autograd.Variable(torch.from_numpy(X_sup).cuda())
y_sup = torch.autograd.Variable(
torch.from_numpy(y_sup).long().cuda())
optimizer.zero_grad()
net.train(mode=True)
sup_logits_out = net(X_sup)
# Supervised classification loss
clf_loss = classification_criterion(sup_logits_out, y_sup)
loss_expr = clf_loss
loss_expr.backward()
optimizer.step()
n_samples = X_sup.size()[0]
return float(clf_loss.data.cpu().numpy()) * n_samples
print('Compiled training function')
def f_pred_src(X_sup):
X_var = torch.autograd.Variable(torch.from_numpy(X_sup).cuda())
net.train(mode=False)
return F.softmax(net(X_var)).data.cpu().numpy()
def f_pred_tgt(X_sup):
X_var = torch.autograd.Variable(torch.from_numpy(X_sup).cuda())
net.train(mode=False)
return F.softmax(net(X_var)).data.cpu().numpy()
def f_eval_src(X_sup, y_sup):
y_pred_prob = f_pred_src(X_sup)
y_pred = np.argmax(y_pred_prob, axis=1)
return float((y_pred != y_sup).sum())
def f_eval_tgt(X_sup, y_sup):
y_pred_prob = f_pred_tgt(X_sup)
y_pred = np.argmax(y_pred_prob, axis=1)
return float((y_pred != y_sup).sum())
print('Compiled evaluation function')
# Setup output
def log(text):
print(text)
if log_file is not None:
with open(log_file, 'a') as f:
f.write(text + '\n')
f.flush()
f.close()
cmdline_helpers.ensure_containing_dir_exists(log_file)
# Report setttings
log('sys.argv={}'.format(sys.argv))
# Report dataset size
log('Dataset:')
log('SOURCE Train: X.shape={}, y.shape={}'.format(
d_source.train_X.shape, d_source.train_y.shape))
log('SOURCE Test: X.shape={}, y.shape={}'.format(
d_source.test_X.shape, d_source.test_y.shape))
log('TARGET Train: X.shape={}'.format(d_target.train_X.shape))
log('TARGET Test: X.shape={}, y.shape={}'.format(
d_target.test_X.shape, d_target.test_y.shape))
print('Training...')
train_ds = data_source.ArrayDataSource(
[d_source.train_X, d_source.train_y]).map(augment)
source_test_ds = data_source.ArrayDataSource(
[d_source.test_X, d_source.test_y])
target_test_ds = data_source.ArrayDataSource(
[d_target.test_X, d_target.test_y])
if seed != 0:
shuffle_rng = np.random.RandomState(seed)
else:
shuffle_rng = np.random
best_src_test_err = 1.0
for epoch in range(num_epochs):
t1 = time.time()
train_res = train_ds.batch_map_mean(f_train,
batch_size=batch_size,
shuffle=shuffle_rng)
train_clf_loss = train_res[0]
src_test_err, = source_test_ds.batch_map_mean(
f_eval_src, batch_size=batch_size * 4)
tgt_test_err, = target_test_ds.batch_map_mean(
f_eval_tgt, batch_size=batch_size * 4)
t2 = time.time()
if src_test_err < best_src_test_err:
log('*** Epoch {} took {:.2f}s: TRAIN clf loss={:.6f}; '
'SRC TEST ERR={:.3%}, TGT TEST err={:.3%}'.format(
epoch, t2 - t1, train_clf_loss, src_test_err,
tgt_test_err))
best_src_test_err = src_test_err
else:
log('Epoch {} took {:.2f}s: TRAIN clf loss={:.6f}; '
'SRC TEST ERR={:.3%}, TGT TEST err={:.3%}'.format(
epoch, t2 - t1, train_clf_loss, src_test_err,
tgt_test_err))
def experiment(exp, arch, rnd_init, img_size, standardise_samples,
learning_rate, pretrained_lr_factor, fix_layers, double_softmax,
use_dropout, scale_u_range, scale_x_range, scale_y_range,
affine_std, xlat_range, rot_std, hflip, intens_scale_range,
colour_rot_std, colour_off_std, greyscale, img_pad_width,
num_epochs, batch_size, seed, log_file, result_file,
hide_progress_bar, subsetsize, subsetseed, device):
settings = locals().copy()
if rnd_init:
if fix_layers != '':
print('`rnd_init` and `fix_layers` are mutually exclusive')
return
import os
import sys
import cmdline_helpers
fix_layers = [lyr.strip() for lyr in fix_layers.split(',')]
if log_file == '':
log_file = 'output_aug_log_{}.txt'.format(exp)
elif log_file == 'none':
log_file = None
if log_file is not None:
if os.path.exists(log_file):
print('Output log file {} already exists'.format(log_file))
return
intens_scale_range_lower, intens_scale_range_upper = cmdline_helpers.colon_separated_range(
intens_scale_range)
scale_u_range = cmdline_helpers.colon_separated_range(scale_u_range)
scale_x_range = cmdline_helpers.colon_separated_range(scale_x_range)
scale_y_range = cmdline_helpers.colon_separated_range(scale_y_range)
import time
import tqdm
import math
import tables
import numpy as np
from batchup import data_source, work_pool
import image_dataset, visda17_dataset, office_dataset
import network_architectures
import augmentation
import image_transforms
from sklearn.model_selection import StratifiedShuffleSplit
import torch, torch.cuda
from torch import nn
from torch.nn import functional as F
if hide_progress_bar:
progress_bar = None
else:
progress_bar = tqdm.tqdm
with torch.cuda.device(device):
pool = work_pool.WorkerThreadPool(4)
n_chn = 0
half_batch_size = batch_size // 2
RESNET_ARCHS = {'resnet50', 'resnet101', 'resnet152'}
RNDINIT_ARCHS = {'vgg13_48_gp'}
if arch == '':
if exp in {'train_val', 'train_test'}:
arch = 'resnet50'
if arch in RESNET_ARCHS and not rnd_init:
mean_value = np.array([0.485, 0.456, 0.406])
std_value = np.array([0.229, 0.224, 0.225])
elif arch in RNDINIT_ARCHS:
mean_value = np.array([0.5, 0.5, 0.5])
std_value = np.array([0.5, 0.5, 0.5])
rnd_init = True
else:
mean_value = std_value = None
img_shape = (img_size, img_size)
img_padding = (img_pad_width, img_pad_width)
if exp == 'visda_train_val':
d_source = visda17_dataset.TrainDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = visda17_dataset.ValidationDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'visda_train_test':
d_source = visda17_dataset.TrainDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = visda17_dataset.TestDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_amazon_dslr':
d_source = office_dataset.OfficeAmazonDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeDSLRDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_amazon_webcam':
d_source = office_dataset.OfficeAmazonDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeWebcamDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_dslr_amazon':
d_source = office_dataset.OfficeDSLRDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeAmazonDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_dslr_webcam':
d_source = office_dataset.OfficeDSLRDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeWebcamDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_webcam_amazon':
d_source = office_dataset.OfficeWebcamDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeAmazonDataset(img_size=img_shape,
range01=True,
rgb_order=True)
elif exp == 'office_webcam_dslr':
d_source = office_dataset.OfficeWebcamDataset(img_size=img_shape,
range01=True,
rgb_order=True)
d_target = office_dataset.OfficeDSLRDataset(img_size=img_shape,
range01=True,
rgb_order=True)
else:
print('Unknown experiment type \'{}\''.format(exp))
return
#
# Result file
#
if result_file != '':
cmdline_helpers.ensure_containing_dir_exists(result_file)
h5_filters = tables.Filters(complevel=9, complib='blosc')
f_target_pred = tables.open_file(result_file, mode='w')
g_tgt_pred = f_target_pred.create_group(f_target_pred.root,
'target_pred_y',
'Target prediction')
arr_tgt_pred = f_target_pred.create_earray(
g_tgt_pred,
'y',
tables.Float32Atom(),
(0, len(d_target.images), d_target.n_classes),
filters=h5_filters)
else:
f_target_pred = None
g_tgt_pred = None
arr_tgt_pred = None
# Delete the training ground truths as we should not be using them
# del d_target.y
n_classes = d_source.n_classes
print('Loaded data')
net_class = network_architectures.get_build_fn_for_architecture(arch)
net = net_class(n_classes, img_size, use_dropout, not rnd_init).cuda()
if arch in RESNET_ARCHS and not rnd_init:
named_params = list(net.named_parameters())
new_params = []
pretrained_params = []
for name, param in named_params:
if name.startswith('new_'):
new_params.append(param)
else:
fix = False
for lyr in fix_layers:
if name.startswith(lyr + '.'):
fix = True
break
if not fix:
pretrained_params.append(param)
else:
print('Fixing param {}'.format(name))
param.requires_grad = False
new_optimizer = torch.optim.Adam(new_params, lr=learning_rate)
if len(pretrained_params) > 0:
pretrained_optimizer = torch.optim.Adam(pretrained_params,
lr=learning_rate *
pretrained_lr_factor)
else:
pretrained_optimizer = None
else:
new_optimizer = torch.optim.Adam(net.parameters(),
lr=learning_rate)
pretrained_optimizer = None
classification_criterion = nn.CrossEntropyLoss()
print('Built network')
# Image augmentation
aug = augmentation.ImageAugmentation(
hflip,
xlat_range,
affine_std,
rot_std=rot_std,
intens_scale_range_lower=intens_scale_range_lower,
intens_scale_range_upper=intens_scale_range_upper,
colour_rot_std=colour_rot_std,
colour_off_std=colour_off_std,
greyscale=greyscale,
scale_u_range=scale_u_range,
scale_x_range=scale_x_range,
scale_y_range=scale_y_range)
test_aug = augmentation.ImageAugmentation(hflip,
xlat_range,
0.0,
rot_std=0.0,
scale_u_range=scale_u_range,
scale_x_range=scale_x_range,
scale_y_range=scale_y_range)
border_value = int(np.mean(mean_value) * 255 + 0.5)
sup_xf = image_transforms.Compose(
image_transforms.ScaleCropAndAugmentAffine(img_shape, img_padding,
True, aug, border_value,
mean_value, std_value),
image_transforms.ToTensor(),
)
test_xf = image_transforms.Compose(
image_transforms.ScaleAndCrop(img_shape, img_padding, False),
image_transforms.ToTensor(),
image_transforms.Standardise(mean_value, std_value),
)
test_xf_aug_mult = image_transforms.Compose(
image_transforms.ScaleCropAndAugmentAffineMultiple(
16, img_shape, img_padding, True, test_aug, border_value,
mean_value, std_value),
image_transforms.ToTensorMultiple(),
)
def augment(X_sup, y_sup):
X_sup = sup_xf(X_sup)[0]
return X_sup, y_sup
_one = torch.autograd.Variable(
torch.from_numpy(np.array([1.0]).astype(np.float32)).cuda())
def f_train(X_sup, y_sup):
X_sup = torch.autograd.Variable(torch.from_numpy(X_sup).cuda())
y_sup = torch.autograd.Variable(
torch.from_numpy(y_sup).long().cuda())
if pretrained_optimizer is not None:
pretrained_optimizer.zero_grad()
new_optimizer.zero_grad()
net.train(mode=True)
sup_logits_out = net(X_sup)
# Supervised classification loss
if double_softmax:
clf_loss = classification_criterion(F.softmax(sup_logits_out),
y_sup)
else:
clf_loss = classification_criterion(sup_logits_out, y_sup)
loss_expr = clf_loss
loss_expr.backward()
if pretrained_optimizer is not None:
pretrained_optimizer.step()
new_optimizer.step()
n_samples = X_sup.size()[0]
return (float(clf_loss.data.cpu()[0]) * n_samples, )
print('Compiled training function')
def f_pred(X_sup):
X_var = torch.autograd.Variable(torch.from_numpy(X_sup).cuda())
net.train(mode=False)
return F.softmax(net(X_var)).data.cpu().numpy()
def f_pred_tgt_mult(X_sup):
net.train(mode=False)
y_pred_aug = []
for aug_i in range(len(X_sup)):
X_var = torch.autograd.Variable(
torch.from_numpy(X_sup[aug_i, ...]).cuda())
y_pred = F.softmax(net(X_var)).data.cpu().numpy()
y_pred_aug.append(y_pred[None, ...])
y_pred_aug = np.concatenate(y_pred_aug, axis=0)
return (y_pred_aug.mean(axis=0), )
print('Compiled evaluation function')
# Setup output
def log(text):
print(text)
if log_file is not None:
with open(log_file, 'a') as f:
f.write(text + '\n')
f.flush()
f.close()
cmdline_helpers.ensure_containing_dir_exists(log_file)
# Report setttings
log('Program = {}'.format(sys.argv[0]))
log('Settings: {}'.format(', '.join([
'{}={}'.format(key, settings[key])
for key in sorted(list(settings.keys()))
])))
# Report dataset size
log('Dataset:')
print('SOURCE len(X)={}, y.shape={}'.format(len(d_source.images),
d_source.y.shape))
print('TARGET len(X)={}'.format(len(d_target.images)))
# Subset
source_indices, target_indices, n_src, n_tgt = image_dataset.subset_indices(
d_source, d_target, subsetsize, subsetseed)
n_train_batches = n_src // batch_size + 1
n_test_batches = n_tgt // (batch_size * 2) + 1
print('Training...')
train_ds = data_source.ArrayDataSource([d_source.images, d_source.y],
indices=source_indices)
train_ds = train_ds.map(augment)
train_ds = pool.parallel_data_source(train_ds,
batch_buffer_size=min(
20, n_train_batches))
# source_test_ds = data_source.ArrayDataSource([d_source.images])
# source_test_ds = pool.parallel_data_source(source_test_ds)
target_ds_for_test = data_source.ArrayDataSource(
[d_target.images], indices=target_indices)
target_test_ds = target_ds_for_test.map(test_xf)
target_test_ds = pool.parallel_data_source(target_test_ds,
batch_buffer_size=min(
20, n_test_batches))
target_mult_test_ds = target_ds_for_test.map(test_xf_aug_mult)
target_mult_test_ds = pool.parallel_data_source(target_mult_test_ds,
batch_buffer_size=min(
20,
n_test_batches))
if seed != 0:
shuffle_rng = np.random.RandomState(seed)
else:
shuffle_rng = np.random
if d_target.has_ground_truth:
evaluator = d_target.prediction_evaluator(target_indices)
else:
evaluator = None
train_batch_iter = train_ds.batch_iterator(batch_size=batch_size,
shuffle=shuffle_rng)
for epoch in range(num_epochs):
t1 = time.time()
test_batch_iter = target_test_ds.batch_iterator(
batch_size=batch_size)
train_clf_loss, = data_source.batch_map_mean(
f_train,
train_batch_iter,
n_batches=n_train_batches,
progress_iter_func=progress_bar)
# train_clf_loss, train_unsup_loss, mask_rate, train_align_loss = train_ds.batch_map_mean(
# lambda *x: 1.0, batch_size=batch_size, shuffle=shuffle_rng, n_batches=n_train_batches,
# progress_iter_func=progress_bar)
if d_target.has_ground_truth or arr_tgt_pred is not None:
tgt_pred_prob_y, = data_source.batch_map_concat(
f_pred, test_batch_iter, progress_iter_func=progress_bar)
else:
tgt_pred_prob_y = None
train_batch_iter = train_ds.batch_iterator(batch_size=batch_size,
shuffle=shuffle_rng)
if d_target.has_ground_truth:
mean_class_acc, cls_acc_str = evaluator.evaluate(
tgt_pred_prob_y)
t2 = time.time()
log('Epoch {} took {:.2f}s: TRAIN clf loss={:.6f}; '
'TGT mean class acc={:.3%}'.format(epoch, t2 - t1,
train_clf_loss,
mean_class_acc))
log(' per class: {}'.format(cls_acc_str))
else:
t2 = time.time()
log('Epoch {} took {:.2f}s: TRAIN clf loss={:.6f}'.format(
epoch, t2 - t1, train_clf_loss))
# Save results
if arr_tgt_pred is not None:
arr_tgt_pred.append(tgt_pred_prob_y[None,
...].astype(np.float32))
# Predict on test set, using augmentation
tgt_aug_pred_prob_y, = target_mult_test_ds.batch_map_concat(
f_pred_tgt_mult,
batch_size=batch_size,
progress_iter_func=progress_bar)
if d_target.has_ground_truth:
aug_mean_class_acc, aug_cls_acc_str = evaluator.evaluate(
tgt_aug_pred_prob_y)
log('FINAL: TGT AUG mean class acc={:.3%}'.format(
aug_mean_class_acc))
log(' per class: {}'.format(aug_cls_acc_str))
if f_target_pred is not None:
f_target_pred.create_array(g_tgt_pred, 'y_prob', tgt_pred_prob_y)
f_target_pred.create_array(g_tgt_pred, 'y_prob_aug',
tgt_aug_pred_prob_y)
f_target_pred.close()
