def run(img_folder,
img_height=1024,
img_scale=4095,
roi_per_img=32,
roi_size=(256, 256),
low_int_threshold=.05,
blob_min_area=3,
blob_min_int=.5,
blob_max_int=.85,
blob_th_step=10,
roi_state=None,
roi_bs=32,
do_featurewise_norm=True,
featurewise_mean=884.7,
featurewise_std=745.3,
img_tsv='./metadata/images_crosswalk_prediction.tsv',
exam_tsv=None,
dl_state=None,
dl_bs=32,
nb_top_avg=1,
validation_mode=False,
val_size=None,
img_voting=False,
out_pred='./output/predictions.tsv'):
'''Run SC1 inference using the candidate ROI approach
Notes:
"mean=884.7, std=745.3" are estimated from 20 subjects on the
training data.
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
# nb_worker = int(os.getenv('NUM_CPU_CORES', 4))
gpu_count = int(os.getenv('NUM_GPU_DEVICES', 1))
# Setup data generator for inference.
meta_man = DMMetaManager(img_tsv=img_tsv,
exam_tsv=exam_tsv,
img_folder=img_folder,
img_extension='dcm')
if val_size is not None: # Use a subset for validation.
subj_list, subj_labs = meta_man.get_subj_labs()
_, subj_test = train_test_split(subj_list,
test_size=val_size,
random_state=random_seed,
stratify=subj_labs)
else:
subj_test = None
if validation_mode:
exam_list = meta_man.get_flatten_exam_list(subj_list=subj_test,
flatten_img_list=True)
else:
exam_list = meta_man.get_last_exam_list(subj_list=subj_test,
flatten_img_list=True)
if do_featurewise_norm:
img_gen = DMImageDataGenerator(featurewise_center=True,
featurewise_std_normalization=True)
img_gen.mean = featurewise_mean
img_gen.std = featurewise_std
else:
img_gen = DMImageDataGenerator(samplewise_center=True,
samplewise_std_normalization=True)
if validation_mode:
class_mode = 'categorical'
else:
class_mode = None
# Load ROI classifier.
if roi_state is not None:
roi_clf = load_model(roi_state,
custom_objects={
'sensitivity': DMMetrics.sensitivity,
'specificity': DMMetrics.specificity
})
if gpu_count > 1:
roi_clf = make_parallel(roi_clf, gpu_count)
else:
roi_clf = None
# Load model.
if dl_state is not None:
model = load_model(dl_state)
else:
raise Exception('At least one model state must be specified.')
if gpu_count > 1:
model = make_parallel(model, gpu_count)
# A function to make predictions on image patches from an image list.
def pred_img_list(img_list):
roi_generator = img_gen.flow_from_candid_roi(
img_list,
target_height=img_height,
target_scale=img_scale,
class_mode=class_mode,
validation_mode=True,
img_per_batch=len(img_list),
roi_per_img=roi_per_img,
roi_size=roi_size,
low_int_threshold=low_int_threshold,
blob_min_area=blob_min_area,
blob_min_int=blob_min_int,
blob_max_int=blob_max_int,
blob_th_step=blob_th_step,
roi_clf=roi_clf,
clf_bs=roi_bs,
return_sample_weight=True,
seed=random_seed)
roi_dat, roi_w = roi_generator.next()
# import pdb; pdb.set_trace()
pred = model.predict(roi_dat, batch_size=dl_bs)
pred = pred[:, 1] # cancer class predictions.
if roi_clf is not None:
# return np.average(pred, weights=roi_w)
# import pdb; pdb.set_trace()
return pred[np.argsort(roi_w)[-nb_top_avg:]].mean()
elif img_voting:
pred = pred.reshape((-1, roi_per_img))
img_preds = [np.sort(row)[-nb_top_avg:].mean() for row in pred]
return np.mean(img_preds)
else:
return np.sort(pred)[-nb_top_avg:].mean()
# Print header.
fout = open(out_pred, 'w')
if validation_mode:
fout.write(dminfer.INFER_HEADER_VAL)
else:
fout.write(dminfer.INFER_HEADER)
for subj, exidx, exam in exam_list:
try:
predL = pred_img_list(exam['L']['img'])
except KeyError:
predL = .0
try:
predR = pred_img_list(exam['R']['img'])
except KeyError:
predR = .0
try:
cancerL = int(exam['L']['cancer'])
except ValueError:
cancerL = 0
try:
cancerR = int(exam['R']['cancer'])
except ValueError:
cancerR = 0
if validation_mode:
fout.write("%s\t%s\tL\t%f\t%d\n" % \
(str(subj), str(exidx), predL, cancerL))
fout.write("%s\t%s\tR\t%f\t%d\n" % \
(str(subj), str(exidx), predR, cancerR))
else:
fout.write("%s\tL\t%f\n" % (str(subj), predL))
fout.write("%s\tR\t%f\n" % (str(subj), predR))
fout.close()
def run(img_folder, img_extension='dcm',
img_height=1024, img_scale=4095,
do_featurewise_norm=True, norm_fit_size=10,
img_per_batch=2, roi_per_img=32, roi_size=(256, 256),
one_patch_mode=False,
low_int_threshold=.05, blob_min_area=3,
blob_min_int=.5, blob_max_int=.85, blob_th_step=10,
data_augmentation=False, roi_state=None, clf_bs=32, cutpoint=.5,
amp_factor=1., return_sample_weight=True, auto_batch_balance=True,
patches_per_epoch=12800, nb_epoch=20,
neg_vs_pos_ratio=None, all_neg_skip=0.,
nb_init_filter=32, init_filter_size=5, init_conv_stride=2,
pool_size=2, pool_stride=2,
weight_decay=.0001, alpha=.0001, l1_ratio=.0,
inp_dropout=.0, hidden_dropout=.0, init_lr=.01,
test_size=.2, val_size=.0,
lr_patience=3, es_patience=10,
resume_from=None, net='resnet50', load_val_ram=False,
load_train_ram=False, no_pos_skip=0., balance_classes=0.,
pred_img_per_batch=1, pred_roi_per_img=32,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
best_model='./modelState/dm_candidROI_best_model.h5',
final_model="NOSAVE",
pred_trainval=False, pred_out="dl_pred_out.pkl"):
'''Run ResNet training on candidate ROIs from mammograms
Args:
norm_fit_size ([int]): the number of patients used to calculate
feature-wise mean and std.
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
# Use of multiple CPU cores is not working!
# When nb_worker>1 and pickle_safe=True, this error is encountered:
# "failed to enqueue async memcpy from host to device: CUDA_ERROR_NOT_INITIALIZED"
# To avoid the error, only this combination worked:
# nb_worker=1 and pickle_safe=False.
nb_worker = int(os.getenv('NUM_CPU_CORES', 4))
gpu_count = int(os.getenv('NUM_GPU_DEVICES', 1))
# Setup training and validation data.
# Load image or exam lists and split them into train and val sets.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
# Split data based on subjects.
subj_list, subj_labs = meta_man.get_subj_labs()
subj_train, subj_test, slab_train, slab_test = train_test_split(
subj_list, subj_labs, test_size=test_size, random_state=random_seed,
stratify=subj_labs)
if val_size > 0: # train/val split.
subj_train, subj_val, slab_train, slab_val = train_test_split(
subj_train, slab_train, test_size=val_size,
random_state=random_seed, stratify=slab_train)
else: # use test as val. make a copy of the test list.
subj_val = list(subj_test)
slab_val = list(slab_test)
# import pdb; pdb.set_trace()
# Subset subject lists to desired ratio.
if neg_vs_pos_ratio is not None:
subj_train, slab_train = DMMetaManager.subset_subj_list(
subj_train, slab_train, neg_vs_pos_ratio, random_seed)
subj_val, slab_val = DMMetaManager.subset_subj_list(
subj_val, slab_val, neg_vs_pos_ratio, random_seed)
print "After sampling, Nb of subjects for train=%d, val=%d, test=%d" \
% (len(subj_train), len(subj_val), len(subj_test))
# Get image and label lists.
img_train, lab_train = meta_man.get_flatten_img_list(subj_train)
img_val, lab_val = meta_man.get_flatten_img_list(subj_val)
# Create image generators for train, fit and val.
imgen_trainval = DMImageDataGenerator()
if data_augmentation:
imgen_trainval.horizontal_flip=True
imgen_trainval.vertical_flip=True
imgen_trainval.rotation_range = 45.
imgen_trainval.shear_range = np.pi/8.
# imgen_trainval.width_shift_range = .05
# imgen_trainval.height_shift_range = .05
# imgen_trainval.zoom_range = [.95, 1.05]
if do_featurewise_norm:
imgen_trainval.featurewise_center = True
imgen_trainval.featurewise_std_normalization = True
# Fit feature-wise mean and std.
img_fit,_ = meta_man.get_flatten_img_list(
subj_train[:norm_fit_size]) # fit on a subset.
print ">>> Fit image generator <<<"; sys.stdout.flush()
fit_generator = imgen_trainval.flow_from_candid_roi(
img_fit,
target_height=img_height, target_scale=img_scale,
class_mode=None, validation_mode=True,
img_per_batch=len(img_fit), roi_per_img=roi_per_img,
roi_size=roi_size,
low_int_threshold=low_int_threshold, blob_min_area=blob_min_area,
blob_min_int=blob_min_int, blob_max_int=blob_max_int,
blob_th_step=blob_th_step,
roi_clf=None, return_sample_weight=False, seed=random_seed)
imgen_trainval.fit(fit_generator.next())
print "Estimates from %d images: mean=%.1f, std=%.1f." % \
(len(img_fit), imgen_trainval.mean, imgen_trainval.std)
sys.stdout.flush()
else:
imgen_trainval.samplewise_center = True
imgen_trainval.samplewise_std_normalization = True
# Load ROI classifier.
if roi_state is not None:
roi_clf = load_model(
roi_state,
custom_objects={
'sensitivity': DMMetrics.sensitivity,
'specificity': DMMetrics.specificity
}
)
graph = tf.get_default_graph()
else:
roi_clf = None
graph = None
# Set some DL training related parameters.
if one_patch_mode:
class_mode = 'binary'
loss = 'binary_crossentropy'
metrics = [DMMetrics.sensitivity, DMMetrics.specificity]
else:
class_mode = 'categorical'
loss = 'categorical_crossentropy'
metrics = ['accuracy', 'precision', 'recall']
if load_train_ram:
validation_mode = True
return_raw_img = True
else:
validation_mode = False
return_raw_img = False
# Create train and val generators.
print ">>> Train image generator <<<"; sys.stdout.flush()
train_generator = imgen_trainval.flow_from_candid_roi(
img_train, lab_train,
target_height=img_height, target_scale=img_scale,
class_mode=class_mode, validation_mode=validation_mode,
img_per_batch=img_per_batch, roi_per_img=roi_per_img,
roi_size=roi_size, one_patch_mode=one_patch_mode,
low_int_threshold=low_int_threshold, blob_min_area=blob_min_area,
blob_min_int=blob_min_int, blob_max_int=blob_max_int,
blob_th_step=blob_th_step,
tf_graph=graph, roi_clf=roi_clf, clf_bs=clf_bs, cutpoint=cutpoint,
amp_factor=amp_factor, return_sample_weight=return_sample_weight,
auto_batch_balance=auto_batch_balance,
all_neg_skip=all_neg_skip, shuffle=True, seed=random_seed,
return_raw_img=return_raw_img)
print ">>> Validation image generator <<<"; sys.stdout.flush()
val_generator = imgen_trainval.flow_from_candid_roi(
img_val, lab_val,
target_height=img_height, target_scale=img_scale,
class_mode=class_mode, validation_mode=True,
img_per_batch=img_per_batch, roi_per_img=roi_per_img,
roi_size=roi_size, one_patch_mode=one_patch_mode,
low_int_threshold=low_int_threshold, blob_min_area=blob_min_area,
blob_min_int=blob_min_int, blob_max_int=blob_max_int,
blob_th_step=blob_th_step,
tf_graph=graph, roi_clf=roi_clf, clf_bs=clf_bs, cutpoint=cutpoint,
amp_factor=amp_factor, return_sample_weight=False,
auto_batch_balance=False,
seed=random_seed)
# Load train and validation set into RAM.
if one_patch_mode:
nb_train_samples = len(img_train)
nb_val_samples = len(img_val)
else:
nb_train_samples = len(img_train)*roi_per_img
nb_val_samples = len(img_val)*roi_per_img
if load_val_ram:
print "Loading validation data into RAM.",
sys.stdout.flush()
validation_set = load_dat_ram(val_generator, nb_val_samples)
print "Done."; sys.stdout.flush()
sparse_y = to_sparse(validation_set[1])
for uy in np.unique(sparse_y):
print "Nb of samples for class:%d = %d" % \
(uy, (sparse_y==uy).sum())
sys.stdout.flush()
if load_train_ram:
print "Loading train data into RAM.",
sys.stdout.flush()
train_set = load_dat_ram(train_generator, nb_train_samples)
print "Done."; sys.stdout.flush()
sparse_y = to_sparse(train_set[1])
for uy in np.unique(sparse_y):
print "Nb of samples for class:%d = %d" % \
(uy, (sparse_y==uy).sum())
sys.stdout.flush()
train_generator = imgen_trainval.flow(
train_set[0], train_set[1], batch_size=clf_bs,
auto_batch_balance=auto_batch_balance, no_pos_skip=no_pos_skip,
balance_classes=balance_classes, shuffle=True, seed=random_seed)
# Load or create model.
if resume_from is not None:
model = load_model(
resume_from,
custom_objects={
'sensitivity': DMMetrics.sensitivity,
'specificity': DMMetrics.specificity
}
)
else:
builder = ResNetBuilder
if net == 'resnet18':
model = builder.build_resnet_18(
(1, roi_size[0], roi_size[1]), 3, nb_init_filter, init_filter_size,
init_conv_stride, pool_size, pool_stride, weight_decay, alpha, l1_ratio,
inp_dropout, hidden_dropout)
elif net == 'resnet34':
model = builder.build_resnet_34(
(1, roi_size[0], roi_size[1]), 3, nb_init_filter, init_filter_size,
init_conv_stride, pool_size, pool_stride, weight_decay, alpha, l1_ratio,
inp_dropout, hidden_dropout)
elif net == 'resnet50':
model = builder.build_resnet_50(
(1, roi_size[0], roi_size[1]), 3, nb_init_filter, init_filter_size,
init_conv_stride, pool_size, pool_stride, weight_decay, alpha, l1_ratio,
inp_dropout, hidden_dropout)
elif net == 'resnet101':
model = builder.build_resnet_101(
(1, roi_size[0], roi_size[1]), 3, nb_init_filter, init_filter_size,
init_conv_stride, pool_size, pool_stride, weight_decay, alpha, l1_ratio,
inp_dropout, hidden_dropout)
elif net == 'resnet152':
model = builder.build_resnet_152(
(1, roi_size[0], roi_size[1]), 3, nb_init_filter, init_filter_size,
init_conv_stride, pool_size, pool_stride, weight_decay, alpha, l1_ratio,
inp_dropout, hidden_dropout)
if gpu_count > 1:
model = make_parallel(model, gpu_count)
# Model training.
sgd = SGD(lr=init_lr, momentum=0.9, decay=0.0, nesterov=True)
model.compile(optimizer=sgd, loss=loss, metrics=metrics)
reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5,
patience=lr_patience, verbose=1)
early_stopping = EarlyStopping(monitor='val_loss', patience=es_patience,
verbose=1)
if load_val_ram:
auc_checkpointer = DMAucModelCheckpoint(
best_model, validation_set, batch_size=clf_bs)
else:
auc_checkpointer = DMAucModelCheckpoint(
best_model, val_generator, nb_test_samples=nb_val_samples)
hist = model.fit_generator(
train_generator,
samples_per_epoch=patches_per_epoch,
nb_epoch=nb_epoch,
validation_data=validation_set if load_val_ram else val_generator,
nb_val_samples=nb_val_samples,
callbacks=[reduce_lr, early_stopping, auc_checkpointer],
# nb_worker=1, pickle_safe=False,
nb_worker=nb_worker if load_train_ram else 1,
pickle_safe=True if load_train_ram else False,
verbose=2)
if final_model != "NOSAVE":
print "Saving final model to:", final_model; sys.stdout.flush()
model.save(final_model)
# Training report.
min_loss_locs, = np.where(hist.history['val_loss'] == min(hist.history['val_loss']))
best_val_loss = hist.history['val_loss'][min_loss_locs[0]]
if one_patch_mode:
best_val_sensitivity = hist.history['val_sensitivity'][min_loss_locs[0]]
best_val_specificity = hist.history['val_specificity'][min_loss_locs[0]]
else:
best_val_precision = hist.history['val_precision'][min_loss_locs[0]]
best_val_recall = hist.history['val_recall'][min_loss_locs[0]]
best_val_accuracy = hist.history['val_acc'][min_loss_locs[0]]
print "\n==== Training summary ===="
print "Minimum val loss achieved at epoch:", min_loss_locs[0] + 1
print "Best val loss:", best_val_loss
if one_patch_mode:
print "Best val sensitivity:", best_val_sensitivity
print "Best val specificity:", best_val_specificity
else:
print "Best val precision:", best_val_precision
print "Best val recall:", best_val_recall
print "Best val accuracy:", best_val_accuracy
# Make predictions on train, val, test exam lists.
if best_model != 'NOSAVE':
print "\n==== Making predictions ===="
print "Load best model for prediction:", best_model
sys.stdout.flush()
pred_model = load_model(best_model)
if gpu_count > 1:
pred_model = make_parallel(pred_model, gpu_count)
if pred_trainval:
print "Load exam lists for train, val sets"; sys.stdout.flush()
exam_train = meta_man.get_flatten_exam_list(
subj_train, flatten_img_list=True)
print "Train exam list length=", len(exam_train); sys.stdout.flush()
exam_val = meta_man.get_flatten_exam_list(
subj_val, flatten_img_list=True)
print "Val exam list length=", len(exam_val); sys.stdout.flush()
print "Load exam list for test set"; sys.stdout.flush()
exam_test = meta_man.get_flatten_exam_list(
subj_test, flatten_img_list=True)
print "Test exam list length=", len(exam_test); sys.stdout.flush()
if do_featurewise_norm:
imgen_pred = DMImageDataGenerator()
imgen_pred.featurewise_center = True
imgen_pred.featurewise_std_normalization = True
imgen_pred.mean = imgen_trainval.mean
imgen_pred.std = imgen_trainval.std
else:
imgen_pred.samplewise_center = True
imgen_pred.samplewise_std_normalization = True
if pred_trainval:
print "Make predictions on train exam list"; sys.stdout.flush()
meta_prob_train = get_exam_pred(
exam_train, pred_roi_per_img, imgen_pred,
target_height=img_height, target_scale=img_scale,
img_per_batch=pred_img_per_batch, roi_size=roi_size,
low_int_threshold=low_int_threshold, blob_min_area=blob_min_area,
blob_min_int=blob_min_int, blob_max_int=blob_max_int,
blob_th_step=blob_th_step, seed=random_seed,
dl_model=pred_model)
print "Train prediction list length=", len(meta_prob_train)
print "Make predictions on val exam list"; sys.stdout.flush()
meta_prob_val = get_exam_pred(
exam_val, pred_roi_per_img, imgen_pred,
target_height=img_height, target_scale=img_scale,
img_per_batch=pred_img_per_batch, roi_size=roi_size,
low_int_threshold=low_int_threshold, blob_min_area=blob_min_area,
blob_min_int=blob_min_int, blob_max_int=blob_max_int,
blob_th_step=blob_th_step, seed=random_seed,
dl_model=pred_model)
print "Val prediction list length=", len(meta_prob_val)
print "Make predictions on test exam list"; sys.stdout.flush()
meta_prob_test = get_exam_pred(
exam_test, pred_roi_per_img, imgen_pred,
target_height=img_height, target_scale=img_scale,
img_per_batch=pred_img_per_batch, roi_size=roi_size,
low_int_threshold=low_int_threshold, blob_min_area=blob_min_area,
blob_min_int=blob_min_int, blob_max_int=blob_max_int,
blob_th_step=blob_th_step, seed=random_seed,
dl_model=pred_model)
print "Test prediction list length=", len(meta_prob_test)
if pred_trainval:
pickle.dump((meta_prob_train, meta_prob_val, meta_prob_test),
open(pred_out, 'w'))
else:
pickle.dump(meta_prob_test, open(pred_out, 'w'))
return hist
def run(img_folder,
dl_state,
best_model,
img_extension='dcm',
img_height=1024,
img_scale=255.,
equalize_hist=False,
featurewise_center=False,
featurewise_mean=91.6,
neg_vs_pos_ratio=1.,
val_size=.1,
test_size=.15,
net='vgg19',
batch_size=128,
train_bs_multiplier=.5,
patch_size=256,
stride=8,
roi_cutoff=.9,
bkg_cutoff=[.5, 1.],
sample_bkg=True,
train_out='./scratch/train',
val_out='./scratch/val',
test_out='./scratch/test',
out_img_ext='png',
neg_name='benign',
pos_name='malignant',
bkg_name='background',
augmentation=True,
load_train_ram=False,
load_val_ram=False,
top_layer_nb=None,
nb_epoch=10,
top_layer_epochs=0,
all_layer_epochs=0,
optim='sgd',
init_lr=.01,
top_layer_multiplier=.01,
all_layer_multiplier=.0001,
es_patience=5,
lr_patience=2,
weight_decay2=.01,
bias_multiplier=.1,
hidden_dropout2=.0,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
out='./modelState/subj_lists.pkl'):
'''Finetune a trained DL model on a different dataset
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
rng = RandomState(random_seed) # an rng used across board.
nb_worker = int(os.getenv('NUM_CPU_CORES', 4))
gpu_count = int(os.getenv('NUM_GPU_DEVICES', 1))
# Load and split image and label lists.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
subj_list, subj_labs = meta_man.get_subj_labs()
subj_labs = np.array(subj_labs)
print "Found %d subjests" % (len(subj_list))
print "cancer patients=%d, normal patients=%d" \
% ((subj_labs==1).sum(), (subj_labs==0).sum())
if neg_vs_pos_ratio is not None:
subj_list, subj_labs = DMMetaManager.subset_subj_list(
subj_list, subj_labs, neg_vs_pos_ratio, random_seed)
subj_labs = np.array(subj_labs)
print "After subsetting, there are %d subjects" % (len(subj_list))
print "cancer patients=%d, normal patients=%d" \
% ((subj_labs==1).sum(), (subj_labs==0).sum())
subj_train, subj_test, labs_train, labs_test = train_test_split(
subj_list,
subj_labs,
test_size=test_size,
stratify=subj_labs,
random_state=random_seed)
subj_train, subj_val, labs_train, labs_val = train_test_split(
subj_train,
labs_train,
test_size=val_size,
stratify=labs_train,
random_state=random_seed)
# Get image lists.
# >>>> Debug <<<< #
# subj_train = subj_train[:5]
# subj_val = subj_val[:5]
# subj_test = subj_test[:5]
# >>>> Debug <<<< #
print "Get flattened image lists"
img_train, ilab_train = meta_man.get_flatten_img_list(subj_train)
img_val, ilab_val = meta_man.get_flatten_img_list(subj_val)
img_test, ilab_test = meta_man.get_flatten_img_list(subj_test)
ilab_train = np.array(ilab_train)
ilab_val = np.array(ilab_val)
ilab_test = np.array(ilab_test)
print "On train set, positive img=%d, negative img=%d" \
% ((ilab_train==1).sum(), (ilab_train==0).sum())
print "On val set, positive img=%d, negative img=%d" \
% ((ilab_val==1).sum(), (ilab_val==0).sum())
print "On test set, positive img=%d, negative img=%d" \
% ((ilab_test==1).sum(), (ilab_test==0).sum())
sys.stdout.flush()
# Save the subj lists.
print "Saving subject lists to external files.",
sys.stdout.flush()
pickle.dump((subj_train, subj_val, subj_test), open(out, 'w'))
print "Done."
# Load DL model, preprocess function.
print "Load patch classifier:", dl_state
sys.stdout.flush()
dl_model, preprocess_input, top_layer_nb = get_dl_model(
net,
use_pretrained=True,
resume_from=dl_state,
top_layer_nb=top_layer_nb)
if featurewise_center:
preprocess_input = None
if gpu_count > 1:
print "Make the model parallel on %d GPUs" % (gpu_count)
sys.stdout.flush()
dl_model, org_model = make_parallel(dl_model, gpu_count)
parallelized = True
else:
org_model = dl_model
parallelized = False
# Sweep the whole images and classify patches.
print "Score image patches and write them to:", train_out
sys.stdout.flush()
nb_roi_train, nb_bkg_train = score_write_patches(
img_train,
ilab_train,
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
neg_out=os.path.join(train_out, neg_name),
pos_out=os.path.join(train_out, pos_name),
bkg_out=os.path.join(train_out, bkg_name),
preprocess=preprocess_input,
equalize_hist=equalize_hist,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
roi_cutoff=roi_cutoff,
bkg_cutoff=bkg_cutoff,
sample_bkg=sample_bkg,
img_ext=out_img_ext,
random_seed=random_seed,
parallelized=parallelized)
print "Wrote %d ROI and %d bkg patches" % (nb_roi_train, nb_bkg_train)
####
print "Score image patches and write them to:", val_out
sys.stdout.flush()
nb_roi_val, nb_bkg_val = score_write_patches(
img_val,
ilab_val,
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
neg_out=os.path.join(val_out, neg_name),
pos_out=os.path.join(val_out, pos_name),
bkg_out=os.path.join(val_out, bkg_name),
preprocess=preprocess_input,
equalize_hist=equalize_hist,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
roi_cutoff=roi_cutoff,
bkg_cutoff=bkg_cutoff,
sample_bkg=sample_bkg,
img_ext=out_img_ext,
random_seed=random_seed,
parallelized=parallelized)
print "Wrote %d ROI and %d bkg patches" % (nb_roi_val, nb_bkg_val)
####
print "Score image patches and write them to:", test_out
sys.stdout.flush()
nb_roi_test, nb_bkg_test = score_write_patches(
img_test,
ilab_test,
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
neg_out=os.path.join(test_out, neg_name),
pos_out=os.path.join(test_out, pos_name),
bkg_out=os.path.join(test_out, bkg_name),
preprocess=preprocess_input,
equalize_hist=equalize_hist,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
roi_cutoff=roi_cutoff,
bkg_cutoff=bkg_cutoff,
sample_bkg=sample_bkg,
img_ext=out_img_ext,
random_seed=random_seed,
parallelized=parallelized)
print "Wrote %d ROI and %d bkg patches" % (nb_roi_test, nb_bkg_test)
sys.stdout.flush()
# ==== Image generators ==== #
if featurewise_center:
train_imgen = DMImageDataGenerator(featurewise_center=True)
val_imgen = DMImageDataGenerator(featurewise_center=True)
test_imgen = DMImageDataGenerator(featurewise_center=True)
train_imgen.mean = featurewise_mean
val_imgen.mean = featurewise_mean
test_imgen.mean = featurewise_mean
else:
train_imgen = DMImageDataGenerator()
val_imgen = DMImageDataGenerator()
test_imgen = DMImageDataGenerator()
if augmentation:
train_imgen.horizontal_flip = True
train_imgen.vertical_flip = True
train_imgen.rotation_range = 45.
train_imgen.shear_range = np.pi / 8.
# ==== Train & val set ==== #
# Note: the images are histogram equalized before they were written to
# external folders.
train_bs = int(batch_size * train_bs_multiplier)
if load_train_ram:
raw_imgen = DMImageDataGenerator()
print "Create generator for raw train set"
raw_generator = raw_imgen.flow_from_directory(
train_out,
target_size=(patch_size, patch_size),
target_scale=img_scale,
equalize_hist=False,
dup_3_channels=True,
classes=[bkg_name, pos_name, neg_name],
class_mode='categorical',
batch_size=train_bs,
shuffle=False)
print "Loading raw train set into RAM.",
sys.stdout.flush()
raw_set = load_dat_ram(raw_generator, raw_generator.nb_sample)
print "Done."
sys.stdout.flush()
print "Create generator for train set"
train_generator = train_imgen.flow(raw_set[0],
raw_set[1],
batch_size=train_bs,
auto_batch_balance=True,
preprocess=preprocess_input,
shuffle=True,
seed=random_seed)
else:
print "Create generator for train set"
train_generator = train_imgen.flow_from_directory(
train_out,
target_size=(patch_size, patch_size),
target_scale=img_scale,
equalize_hist=False,
dup_3_channels=True,
classes=[bkg_name, pos_name, neg_name],
class_mode='categorical',
auto_batch_balance=True,
batch_size=train_bs,
preprocess=preprocess_input,
shuffle=True,
seed=random_seed)
print "Create generator for val set"
sys.stdout.flush()
validation_set = val_imgen.flow_from_directory(
val_out,
target_size=(patch_size, patch_size),
target_scale=img_scale,
equalize_hist=False,
dup_3_channels=True,
classes=[bkg_name, pos_name, neg_name],
class_mode='categorical',
batch_size=batch_size,
preprocess=preprocess_input,
shuffle=False)
val_samples = validation_set.nb_sample
if parallelized and val_samples % batch_size != 0:
val_samples -= val_samples % batch_size
print "Validation samples =", val_samples
sys.stdout.flush()
if load_val_ram:
print "Loading validation set into RAM.",
sys.stdout.flush()
validation_set = load_dat_ram(validation_set, val_samples)
print "Done."
print "Loaded %d val samples" % (len(validation_set[0]))
sys.stdout.flush()
# ==== Model finetuning ==== #
train_batches = int(train_generator.nb_sample / train_bs) + 1
samples_per_epoch = train_bs * train_batches
# import pdb; pdb.set_trace()
dl_model, loss_hist, acc_hist = do_3stage_training(
dl_model,
org_model,
train_generator,
validation_set,
val_samples,
best_model,
samples_per_epoch,
top_layer_nb,
net,
nb_epoch=nb_epoch,
top_layer_epochs=top_layer_epochs,
all_layer_epochs=all_layer_epochs,
use_pretrained=True,
optim=optim,
init_lr=init_lr,
top_layer_multiplier=top_layer_multiplier,
all_layer_multiplier=all_layer_multiplier,
es_patience=es_patience,
lr_patience=lr_patience,
auto_batch_balance=True,
nb_worker=nb_worker,
weight_decay2=weight_decay2,
bias_multiplier=bias_multiplier,
hidden_dropout2=hidden_dropout2)
# Training report.
min_loss_locs, = np.where(loss_hist == min(loss_hist))
best_val_loss = loss_hist[min_loss_locs[0]]
best_val_accuracy = acc_hist[min_loss_locs[0]]
print "\n==== Training summary ===="
print "Minimum val loss achieved at epoch:", min_loss_locs[0] + 1
print "Best val loss:", best_val_loss
print "Best val accuracy:", best_val_accuracy
# ==== Predict on test set ==== #
print "\n==== Predicting on test set ===="
print "Create generator for test set"
test_generator = test_imgen.flow_from_directory(
test_out,
target_size=(patch_size, patch_size),
target_scale=img_scale,
equalize_hist=False,
dup_3_channels=True,
classes=[bkg_name, pos_name, neg_name],
class_mode='categorical',
batch_size=batch_size,
preprocess=preprocess_input,
shuffle=False)
test_samples = test_generator.nb_sample
if parallelized and test_samples % batch_size != 0:
test_samples -= test_samples % batch_size
print "Test samples =", test_samples
print "Load saved best model:", best_model + '.',
sys.stdout.flush()
org_model.load_weights(best_model)
print "Done."
test_res = dl_model.evaluate_generator(
test_generator,
test_samples,
nb_worker=nb_worker,
pickle_safe=True if nb_worker > 1 else False)
print "Evaluation result on test set:", test_res
def run(img_folder, dl_state, img_extension='dcm',
img_height=1024, img_scale=4095, val_size=.2, neg_vs_pos_ratio=10.,
do_featurewise_norm=True, featurewise_mean=873.6, featurewise_std=739.3,
img_per_batch=2, roi_per_img=32, roi_size=(256, 256),
low_int_threshold=.05, blob_min_area=3,
blob_min_int=.5, blob_max_int=.85, blob_th_step=10,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
train_out='./modelState/meta_prob_train.pkl',
test_out='./modelState/meta_prob_test.pkl'):
'''Calculate bag of deep visual words count matrix for all breasts
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
rng = RandomState(random_seed) # an rng used across board.
gpu_count = int(os.getenv('NUM_GPU_DEVICES', 1))
# Load and split image and label lists.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
subj_list, subj_labs = meta_man.get_subj_labs()
subj_train, subj_test, labs_train, labs_test = train_test_split(
subj_list, subj_labs, test_size=val_size, stratify=subj_labs,
random_state=random_seed)
if neg_vs_pos_ratio is not None:
def subset_subj(subj, labs):
subj = np.array(subj)
labs = np.array(labs)
pos_idx = np.where(labs==1)[0]
neg_idx = np.where(labs==0)[0]
nb_neg_desired = int(len(pos_idx)*neg_vs_pos_ratio)
if nb_neg_desired >= len(neg_idx):
return subj.tolist()
else:
neg_chosen = rng.choice(neg_idx, nb_neg_desired, replace=False)
subset_idx = np.concatenate([pos_idx, neg_chosen])
return subj[subset_idx].tolist()
subj_train = subset_subj(subj_train, labs_train)
subj_test = subset_subj(subj_test, labs_test)
# Create image generator for ROIs for representation extraction.
print "Create an image generator for ROIs"; sys.stdout.flush()
if do_featurewise_norm:
imgen = DMImageDataGenerator(
featurewise_center=True,
featurewise_std_normalization=True)
imgen.mean = featurewise_mean
imgen.std = featurewise_std
else:
imgen = DMImageDataGenerator(
samplewise_center=True,
samplewise_std_normalization=True)
# Load DL model.
print "Load DL classification model:", dl_state; sys.stdout.flush()
dl_model = load_model(
dl_state,
custom_objects={
'sensitivity': dmm.sensitivity,
'specificity': dmm.specificity
}
)
if gpu_count > 1:
print "Make the model parallel on %d GPUs" % (gpu_count)
sys.stdout.flush()
dl_model = make_parallel(dl_model, gpu_count)
# Read exam lists.
exam_train = meta_man.get_flatten_exam_list(
subj_train, flatten_img_list=True)
exam_test = meta_man.get_flatten_exam_list(
subj_test, flatten_img_list=True)
exam_labs_train = np.array(meta_man.exam_labs(exam_train))
exam_labs_test = np.array(meta_man.exam_labs(exam_test))
nb_pos_exams_train = (exam_labs_train==1).sum()
nb_neg_exams_train = (exam_labs_train==0).sum()
nb_pos_exams_test = (exam_labs_test==1).sum()
nb_neg_exams_test = (exam_labs_test==0).sum()
print "Train set - Nb of pos exams: %d, Nb of neg exams: %d" % \
(nb_pos_exams_train, nb_neg_exams_train)
print "Test set - Nb of pos exams: %d, Nb of neg exams: %d" % \
(nb_pos_exams_test, nb_neg_exams_test)
# Make predictions for exam lists.
print "Predicting for train exam list"; sys.stdout.flush()
meta_prob_train = get_exam_pred(
exam_train, roi_per_img, imgen,
target_height=img_height, target_scale=img_scale,
img_per_batch=img_per_batch, roi_size=roi_size,
low_int_threshold=low_int_threshold, blob_min_area=blob_min_area,
blob_min_int=blob_min_int, blob_max_int=blob_max_int,
blob_th_step=blob_th_step, seed=random_seed,
dl_model=dl_model)
print "Length of train prediction list:", len(meta_prob_train)
sys.stdout.flush()
print "Predicting for test exam list"; sys.stdout.flush()
meta_prob_test = get_exam_pred(
exam_test, roi_per_img, imgen,
target_height=img_height, target_scale=img_scale,
img_per_batch=img_per_batch, roi_size=roi_size,
low_int_threshold=low_int_threshold, blob_min_area=blob_min_area,
blob_min_int=blob_min_int, blob_max_int=blob_max_int,
blob_th_step=blob_th_step, seed=random_seed,
dl_model=dl_model)
print "Length of test prediction list:", len(meta_prob_test)
sys.stdout.flush()
pickle.dump(meta_prob_train, open(train_out, 'w'))
pickle.dump(meta_prob_test, open(test_out, 'w'))
print "Done."
def run(img_folder,
dl_state,
img_extension='dcm',
img_height=1024,
img_scale=255.,
equalize_hist=False,
featurewise_center=False,
featurewise_mean=91.6,
neg_vs_pos_ratio=1.,
net='vgg19',
batch_size=128,
patch_size=256,
stride=8,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
out='./modelState/prob_heatmap.pkl',
predicted_subj_file=None,
add_subjs=500):
'''Sweep mammograms with trained DL model to create prob heatmaps
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
rng = RandomState(random_seed) # an rng used across board.
gpu_count = int(os.getenv('NUM_GPU_DEVICES', 1))
# Load and split image and label lists.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
subj_list, subj_labs = meta_man.get_subj_labs()
subj_labs = np.array(subj_labs)
print "Found %d subjests" % (len(subj_list))
print "cancer patients=%d, normal patients=%d" \
% ((subj_labs==1).sum(), (subj_labs==0).sum())
if predicted_subj_file is not None:
predicted_subjs = np.load(predicted_subj_file)
subj_list = np.setdiff1d(subj_list, predicted_subjs)
subj_list = subj_list[:add_subjs]
print "Will predict additional %d subjects" % (len(subj_list))
elif neg_vs_pos_ratio is not None:
subj_list, subj_labs = DMMetaManager.subset_subj_list(
subj_list, subj_labs, neg_vs_pos_ratio, random_seed)
subj_labs = np.array(subj_labs)
print "After subsetting, there are %d subjects" % (len(subj_list))
print "cancer patients=%d, normal patients=%d" \
% ((subj_labs==1).sum(), (subj_labs==0).sum())
# Get exam lists.
# >>>> Debug <<<< #
# subj_list = subj_list[:2]
# >>>> Debug <<<< #
print "Get flattened exam list"
exam_list = meta_man.get_flatten_exam_list(subj_list, cc_mlo_only=True)
exam_labs = meta_man.exam_labs(exam_list)
exam_labs = np.array(exam_labs)
print "positive exams=%d, negative exams=%d" \
% ((exam_labs==1).sum(), (exam_labs==0).sum())
sys.stdout.flush()
# Load DL model.
print "Load patch classifier:", dl_state
sys.stdout.flush()
dl_model = load_model(dl_state,
custom_objects={
'sensitivity': dmm.sensitivity,
'specificity': dmm.specificity
})
if gpu_count > 1:
print "Make the model parallel on %d GPUs" % (gpu_count)
sys.stdout.flush()
dl_model, _ = make_parallel(dl_model, gpu_count)
parallelized = True
else:
parallelized = False
# Load preprocess function.
if featurewise_center:
preprocess_input = None
else:
print "Load preprocess function for net:", net
if net == 'resnet50':
from keras.applications.resnet50 import preprocess_input
elif net == 'vgg16':
from keras.applications.vgg16 import preprocess_input
elif net == 'vgg19':
from keras.applications.vgg19 import preprocess_input
elif net == 'xception':
from keras.applications.xception import preprocess_input
elif net == 'inception':
from keras.applications.inception_v3 import preprocess_input
else:
raise Exception("Pretrained model is not available: " + net)
# Sweep the whole images and classify patches.
print "Generate prob heatmaps for exam list"
sys.stdout.flush()
heatmap_dat_list = []
for i, e in enumerate(exam_list):
dat = (e[0], e[1], {
'L': {
'cancer': e[2]['L']['cancer']
},
'R': {
'cancer': e[2]['R']['cancer']
}
})
dat[2]['L']['CC'] = get_prob_heatmap(
e[2]['L']['CC'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
dat[2]['L']['MLO'] = get_prob_heatmap(
e[2]['L']['MLO'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
dat[2]['R']['CC'] = get_prob_heatmap(
e[2]['R']['CC'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
dat[2]['R']['MLO'] = get_prob_heatmap(
e[2]['R']['MLO'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
heatmap_dat_list.append(dat)
print "processed %d/%d exams" % (i + 1, len(exam_list))
sys.stdout.flush()
### DEBUG ###
# if i >= 1:
# break
### DEBUG ###
print "Done."
# Save the result.
print "Saving result to external files.",
sys.stdout.flush()
pickle.dump(heatmap_dat_list, open(out, 'w'))
print "Done."
def run(img_folder,
dl_state,
img_extension='dcm',
img_height=1024,
img_scale=4095,
val_size=.2,
neg_vs_pos_ratio=10.,
do_featurewise_norm=True,
featurewise_mean=873.6,
featurewise_std=739.3,
img_per_batch=2,
roi_per_img=32,
roi_size=(256, 256),
low_int_threshold=.05,
blob_min_area=3,
blob_min_int=.5,
blob_max_int=.85,
blob_th_step=10,
layer_name=['flatten_1', 'dense_1'],
layer_index=None,
roi_state=None,
roi_clf_bs=32,
pc_components=.95,
pc_whiten=True,
nb_words=[512],
km_max_iter=100,
km_bs=1000,
km_patience=20,
km_init=10,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
pca_km_states='./modelState/dlrepr_pca_km_models.pkl',
bow_train_out='./modelState/bow_dat_train.pkl',
bow_test_out='./modelState/bow_dat_test.pkl'):
'''Calculate bag of deep visual words count matrix for all breasts
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
rng = RandomState(random_seed) # an rng used across board.
# Load and split image and label lists.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
subj_list, subj_labs = meta_man.get_subj_labs()
subj_train, subj_test, labs_train, labs_test = train_test_split(
subj_list,
subj_labs,
test_size=val_size,
stratify=subj_labs,
random_state=random_seed)
if neg_vs_pos_ratio is not None:
def subset_subj(subj, labs):
subj = np.array(subj)
labs = np.array(labs)
pos_idx = np.where(labs == 1)[0]
neg_idx = np.where(labs == 0)[0]
nb_neg_desired = int(len(pos_idx) * neg_vs_pos_ratio)
if nb_neg_desired >= len(neg_idx):
return subj.tolist()
else:
neg_chosen = rng.choice(neg_idx, nb_neg_desired, replace=False)
subset_idx = np.concatenate([pos_idx, neg_chosen])
return subj[subset_idx].tolist()
subj_train = subset_subj(subj_train, labs_train)
subj_test = subset_subj(subj_test, labs_test)
img_list, lab_list = meta_man.get_flatten_img_list(subj_train)
lab_list = np.array(lab_list)
print "Train set - Nb of positive images: %d, Nb of negative images: %d" \
% ( (lab_list==1).sum(), (lab_list==0).sum())
sys.stdout.flush()
# Create image generator for ROIs for representation extraction.
print "Create an image generator for ROIs"
sys.stdout.flush()
if do_featurewise_norm:
imgen = DMImageDataGenerator(featurewise_center=True,
featurewise_std_normalization=True)
imgen.mean = featurewise_mean
imgen.std = featurewise_std
else:
imgen = DMImageDataGenerator(samplewise_center=True,
samplewise_std_normalization=True)
# Load ROI classifier.
if roi_state is not None:
print "Load ROI classifier"
sys.stdout.flush()
roi_clf = load_model(roi_state,
custom_objects={
'sensitivity': dmm.sensitivity,
'specificity': dmm.specificity
})
graph = tf.get_default_graph()
else:
roi_clf = None
graph = None
# Create ROI generators for pos and neg images separately.
print "Create ROI generators for pos and neg images"
sys.stdout.flush()
roi_generator = imgen.flow_from_candid_roi(
img_list,
target_height=img_height,
target_scale=img_scale,
class_mode=None,
validation_mode=True,
img_per_batch=img_per_batch,
roi_per_img=roi_per_img,
roi_size=roi_size,
low_int_threshold=low_int_threshold,
blob_min_area=blob_min_area,
blob_min_int=blob_min_int,
blob_max_int=blob_max_int,
blob_th_step=blob_th_step,
tf_graph=graph,
roi_clf=roi_clf,
clf_bs=roi_clf_bs,
return_sample_weight=False,
seed=random_seed)
# Generate image patches and extract their DL representations.
print "Load DL representation model"
sys.stdout.flush()
dlrepr_model = DLRepr(dl_state,
custom_objects={
'sensitivity': dmm.sensitivity,
'specificity': dmm.specificity
},
layer_name=layer_name,
layer_index=layer_index)
last_output_size = dlrepr_model.get_output_shape()[-1][-1]
if last_output_size != 3 and last_output_size != 1:
raise Exception("The last output must be prob outputs (size=3 or 1)")
nb_tot_samples = len(img_list) * roi_per_img
print "Extract ROIs from pos and neg images"
sys.stdout.flush()
pred = dlrepr_model.predict_generator(roi_generator,
val_samples=nb_tot_samples)
for i, d in enumerate(pred):
print "Shape of representation/output data %d:" % (i), d.shape
sys.stdout.flush()
# Flatten feature maps, e.g. an 8x8 feature map will become a 64-d vector.
pred = [d.reshape((-1, d.shape[-1])) for d in pred]
for i, d in enumerate(pred):
print "Shape of flattened data %d:" % (i), d.shape
sys.stdout.flush()
# Split representations and prob outputs.
dl_repr = pred[0]
prob_out = pred[1]
if prob_out.shape[1] == 3:
prob_out = prob_out[:, 1] # pos class.
prob_out = prob_out.reshape((len(img_list), -1))
print "Reshape prob output to:", prob_out.shape
sys.stdout.flush()
# Use PCA to reduce dimension of the representation data.
if pc_components is not None:
print "Start PCA dimension reduction on DL representation"
sys.stdout.flush()
pca = PCA(n_components=pc_components, whiten=pc_whiten)
pca.fit(dl_repr)
print "Nb of PCA components:", pca.n_components_
print "Total explained variance ratio: %.4f" % \
(pca.explained_variance_ratio_.sum())
dl_repr_pca = pca.transform(dl_repr)
print "Shape of transformed representation data:", dl_repr_pca.shape
sys.stdout.flush()
else:
pca = None
# Use K-means to create a codebook for deep visual words.
print "Start K-means training on DL representation"
sys.stdout.flush()
clf_list = []
clust_list = []
# Shuffling indices for mini-batches learning.
perm_idx = rng.permutation(len(dl_repr))
for n in nb_words:
print "Train K-means with %d cluster centers" % (n)
sys.stdout.flush()
clf = MiniBatchKMeans(n_clusters=n,
init='k-means++',
max_iter=km_max_iter,
batch_size=km_bs,
compute_labels=True,
random_state=random_seed,
tol=0.0,
max_no_improvement=km_patience,
init_size=None,
n_init=km_init,
reassignment_ratio=0.01,
verbose=0)
clf.fit(dl_repr[perm_idx])
clf_list.append(clf)
clust = np.zeros_like(clf.labels_)
clust[perm_idx] = clf.labels_
clust = clust.reshape((len(img_list), -1))
clust_list.append(clust)
if pca is not None:
print "Start K-means training on transformed representation"
sys.stdout.flush()
clf_list_pca = []
clust_list_pca = []
# Shuffling indices for mini-batches learning.
perm_idx = rng.permutation(len(dl_repr_pca))
for n in nb_words:
print "Train K-means with %d cluster centers" % (n)
sys.stdout.flush()
clf = MiniBatchKMeans(n_clusters=n,
init='k-means++',
max_iter=km_max_iter,
batch_size=km_bs,
compute_labels=True,
random_state=random_seed,
tol=0.0,
max_no_improvement=km_patience,
init_size=None,
n_init=km_init,
reassignment_ratio=0.01,
verbose=0)
clf.fit(dl_repr_pca[perm_idx])
clf_list_pca.append(clf)
clust = np.zeros_like(clf.labels_)
clust[perm_idx] = clf.labels_
clust = clust.reshape((len(img_list), -1))
clust_list_pca.append(clust)
# Read exam lists.
exam_train = meta_man.get_flatten_exam_list(subj_train,
flatten_img_list=True)
exam_test = meta_man.get_flatten_exam_list(subj_test,
flatten_img_list=True)
exam_labs_train = np.array(meta_man.exam_labs(exam_train))
exam_labs_test = np.array(meta_man.exam_labs(exam_test))
nb_pos_exams_train = (exam_labs_train == 1).sum()
nb_neg_exams_train = (exam_labs_train == 0).sum()
nb_pos_exams_test = (exam_labs_test == 1).sum()
nb_neg_exams_test = (exam_labs_test == 0).sum()
print "Train set - Nb of pos exams: %d, Nb of neg exams: %d" % \
(nb_pos_exams_train, nb_neg_exams_train)
print "Test set - Nb of pos exams: %d, Nb of neg exams: %d" % \
(nb_pos_exams_test, nb_neg_exams_test)
# Do BoW counts for each breast.
print "BoW counting for train exam list"
sys.stdout.flush()
bow_dat_train = get_exam_bow_dat(exam_train,
nb_words,
roi_per_img,
img_list=img_list,
prob_out=prob_out,
clust_list=clust_list)
for i, d in enumerate(bow_dat_train[1]):
print "Shape of train BoW matrix %d:" % (i), d.shape
sys.stdout.flush()
print "BoW counting for test exam list"
sys.stdout.flush()
bow_dat_test = get_exam_bow_dat(exam_test,
nb_words,
roi_per_img,
imgen=imgen,
clf_list=clf_list,
transformer=None,
target_height=img_height,
target_scale=img_scale,
img_per_batch=img_per_batch,
roi_size=roi_size,
low_int_threshold=low_int_threshold,
blob_min_area=blob_min_area,
blob_min_int=blob_min_int,
blob_max_int=blob_max_int,
blob_th_step=blob_th_step,
seed=random_seed,
dlrepr_model=dlrepr_model)
for i, d in enumerate(bow_dat_test[1]):
print "Shape of test BoW matrix %d:" % (i), d.shape
sys.stdout.flush()
if pca is not None:
print "== Do same BoW counting on PCA transformed data =="
print "BoW counting for train exam list"
sys.stdout.flush()
bow_dat_train_pca = get_exam_bow_dat(exam_train,
nb_words,
roi_per_img,
img_list=img_list,
prob_out=prob_out,
clust_list=clust_list_pca)
for i, d in enumerate(bow_dat_train_pca[1]):
print "Shape of train BoW matrix %d:" % (i), d.shape
sys.stdout.flush()
print "BoW counting for test exam list"
sys.stdout.flush()
bow_dat_test_pca = get_exam_bow_dat(
exam_test,
nb_words,
roi_per_img,
imgen=imgen,
clf_list=clf_list_pca,
transformer=pca,
target_height=img_height,
target_scale=img_scale,
img_per_batch=img_per_batch,
roi_size=roi_size,
low_int_threshold=low_int_threshold,
blob_min_area=blob_min_area,
blob_min_int=blob_min_int,
blob_max_int=blob_max_int,
blob_th_step=blob_th_step,
seed=random_seed,
dlrepr_model=dlrepr_model)
for i, d in enumerate(bow_dat_test_pca[1]):
print "Shape of test BoW matrix %d:" % (i), d.shape
sys.stdout.flush()
# Save K-means model and BoW count data.
if pca is None:
pickle.dump(clf_list, open(pca_km_states, 'w'))
pickle.dump(bow_dat_train, open(bow_train_out, 'w'))
pickle.dump(bow_dat_test, open(bow_test_out, 'w'))
else:
pickle.dump((pca, clf_list), open(pca_km_states, 'w'))
pickle.dump((bow_dat_train, bow_dat_train_pca),
open(bow_train_out, 'w'))
pickle.dump((bow_dat_test, bow_dat_test_pca), open(bow_test_out, 'w'))
print "Done."