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()
from meta import DMMetaManager
meta_man = DMMetaManager(img_folder='preprocessedData/png_288x224/',
img_extension='png')
exam_list = meta_man.get_flatten_exam_list()
img_list = meta_man.get_flatten_img_list()
from dm_image import DMImageDataGenerator
img_gen = DMImageDataGenerator(featurewise_center=True,
featurewise_std_normalization=True)
img_gen.mean = 7772.
img_gen.std = 12187.
datgen_exam = img_gen.flow_from_exam_list(exam_list,
target_size=(288, 224),
batch_size=8,
shuffle=False,
seed=123)
datgen_image = img_gen.flow_from_img_list(img_list[0],
img_list[1],
target_size=(288, 224),
batch_size=32,
shuffle=False,
seed=123)
import numpy as np
def run(train_dir,
val_dir,
test_dir,
patch_model_state=None,
resume_from=None,
img_size=[1152, 896],
img_scale=None,
rescale_factor=None,
featurewise_center=True,
featurewise_mean=52.16,
equalize_hist=False,
augmentation=True,
class_list=['neg', 'pos'],
patch_net='resnet50',
block_type='resnet',
top_depths=[512, 512],
top_repetitions=[3, 3],
bottleneck_enlarge_factor=4,
add_heatmap=False,
avg_pool_size=[7, 7],
add_conv=True,
add_shortcut=False,
hm_strides=(1, 1),
hm_pool_size=(5, 5),
fc_init_units=64,
fc_layers=2,
top_layer_nb=None,
batch_size=64,
train_bs_multiplier=.5,
nb_epoch=5,
all_layer_epochs=20,
load_val_ram=False,
load_train_ram=False,
weight_decay=.0001,
hidden_dropout=.0,
weight_decay2=.0001,
hidden_dropout2=.0,
optim='sgd',
init_lr=.01,
lr_patience=10,
es_patience=25,
auto_batch_balance=False,
pos_cls_weight=1.0,
neg_cls_weight=1.0,
all_layer_multiplier=.1,
best_model='./modelState/image_clf.h5',
final_model="NOSAVE"):
'''Train a deep learning model for image classifications
'''
# ======= Environmental 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))
# ========= Image generator ============== #
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()
# Add augmentation options.
if augmentation:
train_imgen.horizontal_flip = True
train_imgen.vertical_flip = True
train_imgen.rotation_range = 25. # in degree.
train_imgen.shear_range = .2 # in radians.
train_imgen.zoom_range = [.8, 1.2] # in proportion.
train_imgen.channel_shift_range = 20. # in pixel intensity values.
# ================= Model creation ============== #
if resume_from is not None:
image_model = load_model(resume_from, compile=False)
else:
patch_model = load_model(patch_model_state, compile=False)
image_model, top_layer_nb = add_top_layers(
patch_model,
img_size,
patch_net,
block_type,
top_depths,
top_repetitions,
bottleneck_org,
nb_class=len(class_list),
shortcut_with_bn=True,
bottleneck_enlarge_factor=bottleneck_enlarge_factor,
dropout=hidden_dropout,
weight_decay=weight_decay,
add_heatmap=add_heatmap,
avg_pool_size=avg_pool_size,
add_conv=add_conv,
add_shortcut=add_shortcut,
hm_strides=hm_strides,
hm_pool_size=hm_pool_size,
fc_init_units=fc_init_units,
fc_layers=fc_layers)
if gpu_count > 1:
image_model, org_model = make_parallel(image_model, gpu_count)
else:
org_model = image_model
# ============ Train & validation set =============== #
train_bs = int(batch_size * train_bs_multiplier)
dup_3_channels = True
if load_train_ram:
raw_imgen = DMImageDataGenerator()
print "Create generator for raw train set"
raw_generator = raw_imgen.flow_from_directory(
train_dir,
target_size=img_size,
target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
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=auto_batch_balance,
shuffle=True,
seed=random_seed)
else:
print "Create generator for train set"
train_generator = train_imgen.flow_from_directory(
train_dir,
target_size=img_size,
target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
class_mode='categorical',
auto_batch_balance=auto_batch_balance,
batch_size=train_bs,
shuffle=True,
seed=random_seed)
print "Create generator for val set"
validation_set = val_imgen.flow_from_directory(
val_dir,
target_size=img_size,
target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
class_mode='categorical',
batch_size=batch_size,
shuffle=False)
sys.stdout.flush()
if load_val_ram:
print "Loading validation set into RAM.",
sys.stdout.flush()
validation_set = load_dat_ram(validation_set, validation_set.nb_sample)
print "Done."
sys.stdout.flush()
# ==================== Model training ==================== #
# Do 2-stage training.
train_batches = int(train_generator.nb_sample / train_bs) + 1
if isinstance(validation_set, tuple):
val_samples = len(validation_set[0])
else:
val_samples = validation_set.nb_sample
validation_steps = int(val_samples / batch_size)
#### DEBUG ####
# train_batches = 1
# val_samples = batch_size*5
# validation_steps = 5
#### DEBUG ####
if load_val_ram:
auc_checkpointer = DMAucModelCheckpoint(best_model,
validation_set,
batch_size=batch_size)
else:
auc_checkpointer = DMAucModelCheckpoint(best_model,
validation_set,
test_samples=val_samples)
# import pdb; pdb.set_trace()
image_model, loss_hist, acc_hist = do_2stage_training(
image_model,
org_model,
train_generator,
validation_set,
validation_steps,
best_model,
train_batches,
top_layer_nb,
nb_epoch=nb_epoch,
all_layer_epochs=all_layer_epochs,
optim=optim,
init_lr=init_lr,
all_layer_multiplier=all_layer_multiplier,
es_patience=es_patience,
lr_patience=lr_patience,
auto_batch_balance=auto_batch_balance,
pos_cls_weight=pos_cls_weight,
neg_cls_weight=neg_cls_weight,
nb_worker=nb_worker,
auc_checkpointer=auc_checkpointer,
weight_decay=weight_decay,
hidden_dropout=hidden_dropout,
weight_decay2=weight_decay2,
hidden_dropout2=hidden_dropout2,
)
# Training report.
if len(loss_hist) > 0:
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
if final_model != "NOSAVE":
image_model.save(final_model)
# ==== Predict on test set ==== #
print "\n==== Predicting on test set ===="
test_generator = test_imgen.flow_from_directory(
test_dir,
target_size=img_size,
target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
class_mode='categorical',
batch_size=batch_size,
shuffle=False)
test_samples = test_generator.nb_sample
#### DEBUG ####
# test_samples = 5
#### DEBUG ####
print "Test samples =", test_samples
print "Load saved best model:", best_model + '.',
sys.stdout.flush()
org_model.load_weights(best_model)
print "Done."
# test_steps = int(test_generator.nb_sample/batch_size)
# test_res = image_model.evaluate_generator(
# test_generator, test_steps, nb_worker=nb_worker,
# pickle_safe=True if nb_worker > 1 else False)
test_auc = DMAucModelCheckpoint.calc_test_auc(test_generator,
image_model,
test_samples=test_samples)
print "AUROC on test set:", test_auc
def run(img_folder, dl_state, fprop_mode=False,
img_size=(1152, 896), img_height=None, img_scale=None,
rescale_factor=None,
equalize_hist=False, featurewise_center=False, featurewise_mean=71.8,
net='vgg19', batch_size=128, patch_size=256, stride=8,
avg_pool_size=(7, 7), hm_strides=(1, 1),
pat_csv='./full_img/pat.csv', pat_list=None,
out='./output/prob_heatmap.pkl'):
'''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))
# Create image generator.
imgen = DMImageDataGenerator(featurewise_center=featurewise_center)
imgen.mean = featurewise_mean
# Get image and label lists.
df = pd.read_csv(pat_csv, header=0)
df = df.set_index(['patient_id', 'side'])
df.sort_index(inplace=True)
if pat_list is not None:
pat_ids = pd.read_csv(pat_list, header=0).values.ravel()
pat_ids = pat_ids.tolist()
print "Read %d patient IDs" % (len(pat_ids))
df = df.loc[pat_ids]
# Load DL model, preprocess.
print "Load patch classifier:", dl_state; sys.stdout.flush()
dl_model, preprocess_input, _ = get_dl_model(net, resume_from=dl_state)
if fprop_mode:
dl_model = add_top_layers(dl_model, img_size, patch_net=net,
avg_pool_size=avg_pool_size,
return_heatmap=True, hm_strides=hm_strides)
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
if featurewise_center:
preprocess_input = None
# Sweep the whole images and classify patches.
def const_filename(pat, side, view):
basename = '_'.join([pat, side, view]) + '.png'
return os.path.join(img_folder, basename)
print "Generate prob heatmaps"; sys.stdout.flush()
heatmaps = []
cases_seen = 0
nb_cases = len(df.index.unique())
for i, (pat,side) in enumerate(df.index.unique()):
## DEBUG ##
#if i >= 10:
# break
## DEBUG ##
cancer = df.loc[pat].loc[side]['cancer']
cc_fn = const_filename(pat, side, 'CC')
if os.path.isfile(cc_fn):
if fprop_mode:
cc_x = read_img_for_pred(
cc_fn, equalize_hist=equalize_hist, data_format=data_format,
dup_3_channels=True,
transformer=imgen.random_transform,
standardizer=imgen.standardize,
target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor)
cc_x = cc_x.reshape((1,) + cc_x.shape)
cc_hm = dl_model.predict_on_batch(cc_x)[0]
# import pdb; pdb.set_trace()
else:
cc_hm = get_prob_heatmap(
cc_fn, 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)
else:
cc_hm = None
mlo_fn = const_filename(pat, side, 'MLO')
if os.path.isfile(mlo_fn):
if fprop_mode:
mlo_x = read_img_for_pred(
mlo_fn, equalize_hist=equalize_hist, data_format=data_format,
dup_3_channels=True,
transformer=imgen.random_transform,
standardizer=imgen.standardize,
target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor)
mlo_x = mlo_x.reshape((1,) + mlo_x.shape)
mlo_hm = dl_model.predict_on_batch(mlo_x)[0]
else:
mlo_hm = get_prob_heatmap(
mlo_fn, 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)
else:
mlo_hm = None
heatmaps.append({'patient_id':pat, 'side':side, 'cancer':cancer,
'cc':cc_hm, 'mlo':mlo_hm})
print "scored %d/%d cases" % (i + 1, nb_cases)
sys.stdout.flush()
print "Done."
# Save the result.
print "Saving result to external files.",
sys.stdout.flush()
pickle.dump(heatmaps, open(out, 'w'))
print "Done."
def run(img_folder,
img_size=[288, 224],
do_featurewise_norm=True,
featurewise_mean=485.9,
featurewise_std=765.2,
batch_size=16,
img_tsv='./metadata/images_crosswalk_prediction.tsv',
exam_tsv=None,
dl_state=None,
enet_state=None,
validation_mode=False,
use_mean=False,
out_pred='./output/predictions.tsv'):
'''Run SC1 inference
Args:
featurewise_mean, featurewise_std ([float]): they are estimated from
1152 x 896 images. Using different sized images give very close
results. For png, mean=7772, std=12187.
'''
# Setup data generator for inference.
meta_man = DMMetaManager(img_tsv=img_tsv,
exam_tsv=exam_tsv,
img_folder=img_folder,
img_extension='dcm')
if validation_mode:
exam_list = meta_man.get_flatten_exam_list()
else:
exam_list = meta_man.get_last_exam_list()
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 = 'binary'
else:
class_mode = None
datgen_exam = img_gen.flow_from_exam_list(exam_list,
target_size=(img_size[0],
img_size[1]),
class_mode=class_mode,
prediction_mode=True,
batch_size=batch_size)
if enet_state is not None:
model = MultiViewDLElasticNet(*enet_state)
elif dl_state is not None:
model = load_model(dl_state)
else:
raise Exception('At least one model state must be specified.')
exams_seen = 0
fout = open(out_pred, 'w')
# Print header.
if validation_mode:
fout.write(dminfer.INFER_HEADER_VAL)
else:
fout.write(dminfer.INFER_HEADER)
while exams_seen < len(exam_list):
ebat = next(datgen_exam)
if class_mode is not None:
bat_x = ebat[0]
bat_y = ebat[1]
else:
bat_x = ebat
subj_batch = bat_x[0]
exam_batch = bat_x[1]
cc_batch = bat_x[2]
mlo_batch = bat_x[3]
for i, subj in enumerate(subj_batch):
exam = exam_batch[i]
li = i * 2 # left breast index.
ri = i * 2 + 1 # right breast index.
left_pred = dminfer.pred_2view_img_list(cc_batch[li],
mlo_batch[li], model,
use_mean)
right_pred = dminfer.pred_2view_img_list(cc_batch[ri],
mlo_batch[ri], model,
use_mean)
if validation_mode:
fout.write("%s\t%s\tL\t%f\t%f\n" % \
(str(subj), str(exam), left_pred, bat_y[li]))
fout.write("%s\t%s\tR\t%f\t%f\n" % \
(str(subj), str(exam), right_pred, bat_y[ri]))
else:
fout.write("%s\tL\t%f\n" % (str(subj), left_pred))
fout.write("%s\tR\t%f\n" % (str(subj), right_pred))
exams_seen += len(subj_batch)
fout.close()
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."
def run(train_dir, val_dir, test_dir,
img_size=[256, 256], img_scale=None, rescale_factor=None,
featurewise_center=True, featurewise_mean=59.6,
equalize_hist=True, augmentation=False,
class_list=['background', 'malignant', 'benign'],
batch_size=64, train_bs_multiplier=.5, nb_epoch=5,
top_layer_epochs=10, all_layer_epochs=20,
load_val_ram=False, load_train_ram=False,
net='resnet50', use_pretrained=True,
nb_init_filter=32, init_filter_size=5, init_conv_stride=2,
pool_size=2, pool_stride=2,
weight_decay=.0001, weight_decay2=.0001,
alpha=.0001, l1_ratio=.0,
inp_dropout=.0, hidden_dropout=.0, hidden_dropout2=.0,
optim='sgd', init_lr=.01, lr_patience=10, es_patience=25,
resume_from=None, auto_batch_balance=False,
pos_cls_weight=1.0, neg_cls_weight=1.0,
top_layer_nb=None, top_layer_multiplier=.1, all_layer_multiplier=.01,
best_model='./modelState/patch_clf.h5',
final_model="NOSAVE"):
'''Train a deep learning model for patch classifications
'''
#给块分类训练一个深度学习模型
# ======= Environmental 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))
# ========= Image generator ============== #图片生成
if featurewise_center:#数据集去中心化
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()
# Add augmentation options.
#图像增强
if augmentation:
train_imgen.horizontal_flip = True #进行随机水平翻转
train_imgen.vertical_flip = True#进行随机垂直翻转
train_imgen.rotation_range = 25. # in degree.#整数，数据提升时图片随机转动的角度
train_imgen.shear_range = .2 # in radians.浮点数，剪切强度（逆时针方向的剪切变换角度）
train_imgen.zoom_range = [.8, 1.2] # in proportion.
'''
浮点数或形如[lower,upper]的列表，随机缩放的幅度，若为浮点数，则相当于[lower,upper] = [1 - zoom_range, 1+zoom_range]
'''
train_imgen.channel_shift_range = 20. # in pixel intensity values.
#.浮点数，随机通道偏移的幅度
#通过对颜色通道的数值偏移，改变图片的整体的颜色
# ================= Model creation ============== #模型创建
'''
一、weight decay（权值衰减）使用的目的是防止过拟合。
在损失函数中，weight decay是放在正则项（regularization）前面的一个系数，正则项一般指示模型的复杂度，
所以weight decay的作用是调节模型复杂度对损失函数的影响，若weight decay很大，则复杂的模型损失函数的值也就大。
hidden_dropout 防止过拟合
init_conv_stride 卷积核步幅大小
pool_size 池化层大小，pool_stride 池化层步幅（一般是最大值池化，和平均值）
alpha 给图像添加透明度
l1_ratio 交叉验证选择l1和l2惩罚之间的折中,类可以通过交叉验证来设置 alpha(α) 和 l1_ratio(ρ) **参数 :l1_ratio 参数来控制L1和L2的凸组合
inp_dropout 输入权重随机抛弃
'''
model, preprocess_input, top_layer_nb = get_dl_model(
net, nb_class=len(class_list), use_pretrained=use_pretrained,
resume_from=resume_from, img_size=img_size, top_layer_nb=top_layer_nb,
weight_decay=weight_decay, hidden_dropout=hidden_dropout,
nb_init_filter=nb_init_filter, init_filter_size=init_filter_size,
init_conv_stride=init_conv_stride, pool_size=pool_size,
pool_stride=pool_stride, alpha=alpha, l1_ratio=l1_ratio,
inp_dropout=inp_dropout)
if featurewise_center:
preprocess_input = None
if gpu_count > 1:
model, org_model = make_parallel(model, gpu_count)#并行计算
else:
org_model = model
# ============ Train & validation set =============== #
#训练和验证集
train_bs = int(batch_size*train_bs_multiplier)#每批数据量的大小*乘数
if net != 'yaroslav':#dm_keras_ext.py
dup_3_channels = True
else:
dup_3_channels = False
if load_train_ram:
raw_imgen = DMImageDataGenerator()#t图片数据生成器
#创建行训练集数据生成器
print ("Create generator for raw train set")
#以文件夹路径为参数,生成经过数据提升/归一化后的数据,在一个无限循环中无限产生batch数据
'''
equalize_hist 直方图均衡，
shuffle 随机打乱数据
'''
raw_generator = raw_imgen.flow_from_directory(
train_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, class_mode='categorical',
batch_size=train_bs, shuffle=False)
#加载行训练数据集到内存
print ("Loading raw train set into RAM.",sys.stdout.flush())
#行数据集
raw_set = load_dat_ram(raw_generator, raw_generator.nb_sample)
print ("Done."); sys.stdout.flush()
#为训练集创建生成器
print ("Create generator for train set")
#接收numpy数组和标签为参数,生成经过数据提升或标准化后的batch数据,并在一个无限循环中不断的返回batch数据
train_generator = train_imgen.flow(
raw_set[0], raw_set[1], batch_size=train_bs,
auto_batch_balance=auto_batch_balance, preprocess=preprocess_input,
shuffle=True, seed=random_seed)
else:
print ("Create generator for train set")
#以文件夹路径为参数,生成经过数据提升/归一化后的数据,在一个无限循环中无限产生batch数据
train_generator = train_imgen.flow_from_directory(
train_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, class_mode='categorical',
auto_batch_balance=auto_batch_balance, batch_size=train_bs,
preprocess=preprocess_input, shuffle=True, seed=random_seed)
#创建验证集生成器
print ("Create generator for val set")
# 以文件夹路径为参数,生成经过数据提升/归一化后的数据,在一个无限循环中无限产生batch数据
validation_set = val_imgen.flow_from_directory(
val_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, class_mode='categorical',
batch_size=batch_size, preprocess=preprocess_input, shuffle=False)
sys.stdout.flush()
#是否加载验证集到内存中
if load_val_ram:
print ("Loading validation set into RAM.",
sys.stdout.flush())
validation_set = load_dat_ram(validation_set, validation_set.nb_sample)
print ("Done."); sys.stdout.flush()
# ==================== Model training ==================== #模型训练
# Do 3-stage training.三个阶段训练
train_batches = int(train_generator.nb_sample/train_bs) + 1
#判断验证集是否三元组
if isinstance(validation_set, tuple):
val_samples = len(validation_set[0])
else:
val_samples = validation_set.nb_sample
validation_steps = int(val_samples/batch_size)
#### DEBUG ####
# val_samples = 100
#### DEBUG ####
# import pdb; pdb.set_trace()
#通过三阶段训练得到模型，损失率，准确率
model, loss_hist, acc_hist = do_3stage_training(
model, org_model, train_generator, validation_set, validation_steps,
best_model, train_batches, top_layer_nb, net, nb_epoch=nb_epoch,
top_layer_epochs=top_layer_epochs, all_layer_epochs=all_layer_epochs,
use_pretrained=use_pretrained, 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=auto_batch_balance, nb_class=len(class_list),
pos_cls_weight=pos_cls_weight, neg_cls_weight=neg_cls_weight,
nb_worker=nb_worker, weight_decay2=weight_decay2,
hidden_dropout2=hidden_dropout2)
# Training report.
#训练报告
if len(loss_hist) > 0:
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)
#保存模型
if final_model != "NOSAVE":
model.save(final_model)
# ==== Predict on test set ==== #
#基于测试集的预测
print ("\n==== Predicting on test set ====")
# 以文件夹路径为参数,生成经过数据提升/归一化后的数据,在一个无限循环中无限产生batch数据
test_generator = test_imgen.flow_from_directory(
test_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, class_mode='categorical', batch_size=batch_size,
preprocess=preprocess_input, shuffle=False)
print ("Test samples =", test_generator.nb_sample)
#加载最好的模型
print ("Load saved best model:", best_model + '.',
sys.stdout.flush())
#原始模型加载最好模型的权重
org_model.load_weights(best_model)
print ("Done.")
#测试的步数
test_steps = int(test_generator.nb_sample/batch_size)
#### DEBUG ####
# test_samples = 10
#### DEBUG ####
test_res = model.evaluate_generator(
test_generator, test_steps, nb_worker=nb_worker,
pickle_safe=True if nb_worker > 1 else False)
print ("Evaluation result on test set:", test_res)
def run(train_dir, val_dir, test_dir,
img_size=[256, 256], img_scale=None, rescale_factor=None,
featurewise_center=True, featurewise_mean=59.6,
equalize_hist=True, augmentation=False,
class_list=['background', 'malignant', 'benign'],
batch_size=64, train_bs_multiplier=.5, nb_epoch=5,
top_layer_epochs=10, all_layer_epochs=20,
load_val_ram=False, load_train_ram=False,
net='resnet50', use_pretrained=True,
nb_init_filter=32, init_filter_size=5, init_conv_stride=2,
pool_size=2, pool_stride=2,
weight_decay=.0001, weight_decay2=.0001,
alpha=.0001, l1_ratio=.0,
inp_dropout=.0, hidden_dropout=.0, hidden_dropout2=.0,
optim='sgd', init_lr=.01, lr_patience=10, es_patience=25,
resume_from=None, auto_batch_balance=False,
pos_cls_weight=1.0, neg_cls_weight=1.0,
top_layer_nb=None, top_layer_multiplier=.1, all_layer_multiplier=.01,
best_model='./modelState/patch_clf.h5',
final_model="NOSAVE"):
'''Train a deep learning model for patch classifications
'''
# ======= Environmental 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))
# ========= Image generator ============== #
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()
# Add augmentation options.
if augmentation:
train_imgen.horizontal_flip = True
train_imgen.vertical_flip = True
train_imgen.rotation_range = 25. # in degree.
train_imgen.shear_range = .2 # in radians.
train_imgen.zoom_range = [.8, 1.2] # in proportion.
train_imgen.channel_shift_range = 20. # in pixel intensity values.
# ================= Model creation ============== #
model, preprocess_input, top_layer_nb = get_dl_model(
net, nb_class=len(class_list), use_pretrained=use_pretrained,
resume_from=resume_from, img_size=img_size, top_layer_nb=top_layer_nb,
weight_decay=weight_decay, hidden_dropout=hidden_dropout,
nb_init_filter=nb_init_filter, init_filter_size=init_filter_size,
init_conv_stride=init_conv_stride, pool_size=pool_size,
pool_stride=pool_stride, alpha=alpha, l1_ratio=l1_ratio,
inp_dropout=inp_dropout)
if featurewise_center:
preprocess_input = None
if gpu_count > 1:
model, org_model = make_parallel(model, gpu_count)
else:
org_model = model
# ============ Train & validation set =============== #
train_bs = int(batch_size*train_bs_multiplier)
if net != 'yaroslav':
dup_3_channels = True
else:
dup_3_channels = False
if load_train_ram:
raw_imgen = DMImageDataGenerator()
print "Create generator for raw train set"
raw_generator = raw_imgen.flow_from_directory(
train_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, 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=auto_batch_balance, preprocess=preprocess_input,
shuffle=True, seed=random_seed)
else:
print "Create generator for train set"
train_generator = train_imgen.flow_from_directory(
train_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, class_mode='categorical',
auto_batch_balance=auto_batch_balance, batch_size=train_bs,
preprocess=preprocess_input, shuffle=True, seed=random_seed)
print "Create generator for val set"
validation_set = val_imgen.flow_from_directory(
val_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, class_mode='categorical',
batch_size=batch_size, preprocess=preprocess_input, shuffle=False)
sys.stdout.flush()
if load_val_ram:
print "Loading validation set into RAM.",
sys.stdout.flush()
validation_set = load_dat_ram(validation_set, validation_set.nb_sample)
print "Done."; sys.stdout.flush()
# ==================== Model training ==================== #
# Do 3-stage training.
train_batches = int(train_generator.nb_sample/train_bs) + 1
if isinstance(validation_set, tuple):
val_samples = len(validation_set[0])
else:
val_samples = validation_set.nb_sample
validation_steps = int(val_samples/batch_size)
#### DEBUG ####
# val_samples = 100
#### DEBUG ####
# import pdb; pdb.set_trace()
model, loss_hist, acc_hist = do_3stage_training(
model, org_model, train_generator, validation_set, validation_steps,
best_model, train_batches, top_layer_nb, net, nb_epoch=nb_epoch,
top_layer_epochs=top_layer_epochs, all_layer_epochs=all_layer_epochs,
use_pretrained=use_pretrained, 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=auto_batch_balance, nb_class=len(class_list),
pos_cls_weight=pos_cls_weight, neg_cls_weight=neg_cls_weight,
nb_worker=nb_worker, weight_decay2=weight_decay2,
hidden_dropout2=hidden_dropout2)
# Training report.
if len(loss_hist) > 0:
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
if final_model != "NOSAVE":
model.save(final_model)
# ==== Predict on test set ==== #
print "\n==== Predicting on test set ===="
test_generator = test_imgen.flow_from_directory(
test_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, class_mode='categorical', batch_size=batch_size,
preprocess=preprocess_input, shuffle=False)
print "Test samples =", test_generator.nb_sample
print "Load saved best model:", best_model + '.',
sys.stdout.flush()
org_model.load_weights(best_model)
print "Done."
test_steps = int(test_generator.nb_sample/batch_size)
#### DEBUG ####
# test_samples = 10
#### DEBUG ####
test_res = model.evaluate_generator(
test_generator, test_steps, 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, 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,
img_size=[288, 224],
do_featurewise_norm=True,
featurewise_mean=485.9,
featurewise_std=765.2,
img_tsv='./metadata/images_crosswalk.tsv',
exam_tsv='./metadata/exams_metadata.tsv',
dl_state=None,
enet_state=None,
xgb_state=None,
validation_mode=False,
use_mean=False,
out_pred='./output/predictions.tsv'):
'''Run SC2 inference
Args:
featurewise_mean, featurewise_std ([float]): they are estimated from
1152 x 896 images. Using different sized images give very close
results. For png, mean=7772, std=12187.
'''
# Setup data generator for inference.
meta_man = DMMetaManager(img_tsv=img_tsv,
exam_tsv=exam_tsv,
img_folder=img_folder,
img_extension='dcm')
last2_exgen = meta_man.last_2_exam_generator()
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 = 'binary'
else:
class_mode = None
# Image prediction model.
if enet_state is not None:
model = MultiViewDLElasticNet(*enet_state)
elif dl_state is not None:
model = load_model(dl_state)
else:
raise Exception('At least one image model state must be specified.')
# XGB model.
xgb_clf = pickle.load(open(xgb_state))
# Print header.
fout = open(out_pred, 'w')
if validation_mode:
fout.write(dminfer.INFER_HEADER_VAL)
else:
fout.write(dminfer.INFER_HEADER)
# Loop through all last 2 exam pairs.
for subj_id, curr_idx, curr_dat, prior_idx, prior_dat in last2_exgen:
# Get meta info for both breasts.
left_record, right_record = meta_man.get_info_exam_pair(
curr_dat, prior_dat)
nb_days = left_record['daysSincePreviousExam']
# Get image data and make predictions.
exam_list = []
exam_list.append(
(subj_id, curr_idx, meta_man.get_info_per_exam(curr_dat)))
if prior_idx is not None:
exam_list.append(
(subj_id, prior_idx, meta_man.get_info_per_exam(prior_dat)))
datgen_exam = img_gen.flow_from_exam_list(exam_list,
target_size=(img_size[0],
img_size[1]),
class_mode=class_mode,
prediction_mode=True,
batch_size=len(exam_list),
verbose=False)
ebat = next(datgen_exam)
if class_mode is not None:
bat_x = ebat[0]
bat_y = ebat[1]
else:
bat_x = ebat
cc_batch = bat_x[2]
mlo_batch = bat_x[3]
curr_left_score = dminfer.pred_2view_img_list(cc_batch[0],
mlo_batch[0], model,
use_mean)
curr_right_score = dminfer.pred_2view_img_list(cc_batch[1],
mlo_batch[1], model,
use_mean)
if prior_idx is not None:
prior_left_score = dminfer.pred_2view_img_list(
cc_batch[2], mlo_batch[2], model, use_mean)
prior_right_score = dminfer.pred_2view_img_list(
cc_batch[3], mlo_batch[3], model, use_mean)
diff_left_score = (curr_left_score -
prior_left_score) / nb_days * 365
diff_right_score = (curr_right_score -
prior_right_score) / nb_days * 365
else:
prior_left_score = np.nan
prior_right_score = np.nan
diff_left_score = np.nan
diff_right_score = np.nan
# Merge image scores into meta info.
left_record = left_record\
.assign(curr_score=curr_left_score)\
.assign(prior_score=prior_left_score)\
.assign(diff_score=diff_left_score)
right_record = right_record\
.assign(curr_score=curr_right_score)\
.assign(prior_score=prior_right_score)\
.assign(diff_score=diff_right_score)
dsubj = xgb.DMatrix(
pd.concat([left_record, right_record], ignore_index=True))
# Predict using XGB.
pred = xgb_clf.predict(dsubj, ntree_limit=xgb_clf.best_ntree_limit)
# Output.
if validation_mode:
fout.write("%s\t%s\tL\t%f\t%f\n" % \
(str(subj_id), str(curr_idx), pred[0], bat_y[0]))
fout.write("%s\t%s\tR\t%f\t%f\n" % \
(str(subj_id), str(curr_idx), pred[1], bat_y[1]))
else:
fout.write("%s\tL\t%f\n" % (str(subj_id), pred[0]))
fout.write("%s\tR\t%f\n" % (str(subj_id), pred[1]))
fout.close()
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,
img_extension='dcm',
img_size=[288, 224],
img_scale=4095,
multi_view=False,
do_featurewise_norm=True,
featurewise_mean=398.5,
featurewise_std=627.8,
batch_size=16,
samples_per_epoch=160,
nb_epoch=20,
balance_classes=.0,
all_neg_skip=0.,
pos_cls_weight=1.0,
nb_init_filter=64,
init_filter_size=7,
init_conv_stride=2,
pool_size=3,
pool_stride=2,
weight_decay=.0001,
alpha=1.,
l1_ratio=.5,
inp_dropout=.0,
hidden_dropout=.0,
init_lr=.01,
val_size=.2,
lr_patience=5,
es_patience=10,
resume_from=None,
net='resnet50',
load_val_ram=False,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
best_model='./modelState/dm_resnet_best_model.h5',
final_model="NOSAVE"):
'''Run ResNet training on mammograms using an exam or image list
Args:
featurewise_mean, featurewise_std ([float]): they are estimated from
1152 x 896 images. Using different sized images give very close
results. For png, mean=7772, std=12187.
'''
# 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 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)
if multi_view:
exam_list = meta_man.get_flatten_exam_list()
exam_train, exam_val = train_test_split(
exam_list,
test_size=val_size,
random_state=random_seed,
stratify=meta_man.exam_labs(exam_list))
val_size_ = len(exam_val) * 2 # L and R.
else:
img_list, lab_list = meta_man.get_flatten_img_list()
img_train, img_val, lab_train, lab_val = train_test_split(
img_list,
lab_list,
test_size=val_size,
random_state=random_seed,
stratify=lab_list)
val_size_ = len(img_val)
# Create image generator.
img_gen = DMImageDataGenerator(horizontal_flip=True, vertical_flip=True)
if do_featurewise_norm:
img_gen.featurewise_center = True
img_gen.featurewise_std_normalization = True
img_gen.mean = featurewise_mean
img_gen.std = featurewise_std
else:
img_gen.samplewise_center = True
img_gen.samplewise_std_normalization = True
if multi_view:
train_generator = img_gen.flow_from_exam_list(
exam_train,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=batch_size,
balance_classes=balance_classes,
all_neg_skip=all_neg_skip,
shuffle=True,
seed=random_seed,
class_mode='binary')
if load_val_ram:
val_generator = img_gen.flow_from_exam_list(
exam_val,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=val_size_,
validation_mode=True,
class_mode='binary')
else:
val_generator = img_gen.flow_from_exam_list(
exam_val,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=batch_size,
validation_mode=True,
class_mode='binary')
else:
train_generator = img_gen.flow_from_img_list(
img_train,
lab_train,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=batch_size,
balance_classes=balance_classes,
all_neg_skip=all_neg_skip,
shuffle=True,
seed=random_seed,
class_mode='binary')
if load_val_ram:
val_generator = img_gen.flow_from_img_list(
img_val,
lab_val,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=val_size_,
validation_mode=True,
class_mode='binary')
else:
val_generator = img_gen.flow_from_img_list(
img_val,
lab_val,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=batch_size,
validation_mode=True,
class_mode='binary')
# Load validation set into RAM.
if load_val_ram:
validation_set = next(val_generator)
if not multi_view and len(validation_set[0]) != val_size_:
raise Exception
elif len(validation_set[0][0]) != val_size_ \
or len(validation_set[0][1]) != val_size_:
raise Exception
# Create model.
if resume_from is not None:
model = load_model(resume_from,
custom_objects={
'sensitivity': DMMetrics.sensitivity,
'specificity': DMMetrics.specificity
})
else:
if multi_view:
builder = MultiViewResNetBuilder
else:
builder = ResNetBuilder
if net == 'resnet18':
model = builder.build_resnet_18(
(1, img_size[0], img_size[1]), 1, 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, img_size[0], img_size[1]), 1, 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, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'dmresnet14':
model = builder.build_dm_resnet_14(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'dmresnet47rb5':
model = builder.build_dm_resnet_47rb5(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'dmresnet56rb6':
model = builder.build_dm_resnet_56rb6(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'dmresnet65rb7':
model = builder.build_dm_resnet_65rb7(
(1, img_size[0], img_size[1]), 1, 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, img_size[0], img_size[1]), 1, 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, img_size[0], img_size[1]), 1, 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='binary_crossentropy',
metrics=[DMMetrics.sensitivity, DMMetrics.specificity])
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.1,
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=batch_size)
else:
auc_checkpointer = DMAucModelCheckpoint(best_model,
val_generator,
nb_test_samples=val_size_)
# checkpointer = ModelCheckpoint(
# best_model, monitor='val_loss', verbose=1, save_best_only=True)
hist = model.fit_generator(
train_generator,
samples_per_epoch=samples_per_epoch,
nb_epoch=nb_epoch,
class_weight={
0: 1.0,
1: pos_cls_weight
},
validation_data=validation_set if load_val_ram else val_generator,
nb_val_samples=val_size_,
callbacks=[reduce_lr, early_stopping, auc_checkpointer],
nb_worker=nb_worker,
pickle_safe=True, # turn on pickle_safe to avoid a strange error.
verbose=2)
# 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]]
best_val_sensitivity = hist.history['val_sensitivity'][min_loss_locs[0]]
best_val_specificity = hist.history['val_specificity'][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 sensitivity:", best_val_sensitivity
print "Best val specificity:", best_val_specificity
if final_model != "NOSAVE":
model.save(final_model)
return hist
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,
img_extension='png',
img_size=[288, 224],
multi_view=False,
do_featurewise_norm=True,
featurewise_mean=7772.,
featurewise_std=12187.,
batch_size=16,
samples_per_epoch=160,
nb_epoch=20,
val_size=.2,
balance_classes=0.,
all_neg_skip=False,
pos_cls_weight=1.0,
alpha=1.,
l1_ratio=.5,
init_lr=.01,
lr_patience=2,
es_patience=4,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
dl_state='./modelState/resnet50_288_best_model.h5',
best_model='./modelState/enet_288_best_model.h5',
final_model="NOSAVE"):
# 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 training and validation data.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
if multi_view:
exam_list = meta_man.get_flatten_exam_list()
exam_train, exam_val = train_test_split(
exam_list,
test_size=val_size,
random_state=random_seed,
stratify=meta_man.exam_labs(exam_list))
val_size_ = len(exam_val) * 2 # L and R.
else:
img_list, lab_list = meta_man.get_flatten_img_list()
img_train, img_val, lab_train, lab_val = train_test_split(
img_list,
lab_list,
test_size=val_size,
random_state=random_seed,
stratify=lab_list)
val_size_ = len(img_val)
img_gen = DMImageDataGenerator(horizontal_flip=True, vertical_flip=True)
if do_featurewise_norm:
img_gen.featurewise_center = True
img_gen.featurewise_std_normalization = True
img_gen.mean = featurewise_mean
img_gen.std = featurewise_std
else:
img_gen.samplewise_center = True
img_gen.samplewise_std_normalization = True
if multi_view:
train_generator = img_gen.flow_from_exam_list(
exam_train,
target_size=(img_size[0], img_size[1]),
batch_size=batch_size,
balance_classes=balance_classes,
all_neg_skip=all_neg_skip,
shuffle=True,
seed=random_seed,
class_mode='binary')
val_generator = img_gen.flow_from_exam_list(exam_val,
target_size=(img_size[0],
img_size[1]),
batch_size=batch_size,
validation_mode=True,
class_mode='binary')
else:
train_generator = img_gen.flow_from_img_list(
img_train,
lab_train,
target_size=(img_size[0], img_size[1]),
batch_size=batch_size,
balance_classes=balance_classes,
all_neg_skip=all_neg_skip,
shuffle=True,
seed=random_seed,
class_mode='binary')
val_generator = img_gen.flow_from_img_list(img_val,
lab_val,
target_size=(img_size[0],
img_size[1]),
batch_size=batch_size,
validation_mode=True,
class_mode='binary')
# Deep learning model.
dl_model = load_model(dl_state,
custom_objects={
'sensitivity': DMMetrics.sensitivity,
'specificity': DMMetrics.specificity
})
# Dummy compilation to turn off the "uncompiled" error when model was run on multi-GPUs.
# dl_model.compile(optimizer='sgd', loss='binary_crossentropy')
reprlayer_model = Model(input=dl_model.input,
output=dl_model.get_layer(index=-2).output)
if gpu_count > 1:
reprlayer_model = make_parallel(reprlayer_model, gpu_count)
# Setup test data in RAM.
X_test, y_test = dlrepr_generator(reprlayer_model, val_generator,
val_size_)
# import pdb; pdb.set_trace()
# Evaluat DL model on the test data.
val_generator.reset()
dl_test_pred = dl_model.predict_generator(val_generator,
val_samples=val_size_,
nb_worker=1,
pickle_safe=False)
# Set nb_worker to >1 can cause:
# either inconsistent result when pickle_safe is False,
# or broadcasting error when pickle_safe is True.
# This seems to be a Keras bug!!
# Further note: the broadcasting error may only happen when val_size_
# is not divisible by batch_size.
try:
dl_auc = roc_auc_score(y_test, dl_test_pred)
dl_loss = log_loss(y_test, dl_test_pred)
except ValueError:
dl_auc = 0.
dl_loss = np.inf
print "\nAUROC by the DL model: %.4f, loss: %.4f" % (dl_auc, dl_loss)
# import pdb; pdb.set_trace()
# Elastic net training.
target_classes = np.array([0, 1])
sgd_clf = SGDClassifier(loss='log',
penalty='elasticnet',
alpha=alpha,
l1_ratio=l1_ratio,
verbose=0,
n_jobs=nb_worker,
learning_rate='constant',
eta0=init_lr,
random_state=random_seed,
class_weight={
0: 1.0,
1: pos_cls_weight
})
curr_lr = init_lr
best_epoch = 0
best_auc = 0.
min_loss = np.inf
min_loss_epoch = 0
for epoch in xrange(nb_epoch):
samples_seen = 0
X_list = []
y_list = []
epoch_start = time.time()
while samples_seen < samples_per_epoch:
X, y = next(train_generator)
X_repr = reprlayer_model.predict_on_batch(X)
sgd_clf.partial_fit(X_repr, y, classes=target_classes)
samples_seen += len(y)
X_list.append(X_repr)
y_list.append(y)
# The training X, y are expected to change for each epoch due to
# image random sampling and class balancing.
X_train_epo = np.concatenate(X_list)
y_train_epo = np.concatenate(y_list)
# End of epoch summary.
pred_prob = sgd_clf.predict_proba(X_test)[:, 1]
train_prob = sgd_clf.predict_proba(X_train_epo)[:, 1]
try:
auc = roc_auc_score(y_test, pred_prob)
crossentropy_loss = log_loss(y_test, pred_prob)
except ValueError:
auc = 0.
crossentropy_loss = np.inf
try:
train_loss = log_loss(y_train_epo, train_prob)
except ValueError:
train_loss = np.inf
wei_sparseness = np.mean(sgd_clf.coef_ == 0)
epoch_span = time.time() - epoch_start
print ("%ds - Epoch=%d, auc=%.4f, train_loss=%.4f, test_loss=%.4f, "
"weight sparsity=%.4f") % \
(epoch_span, epoch + 1, auc, train_loss, crossentropy_loss,
wei_sparseness)
# Model checkpoint, reducing learning rate and early stopping.
if auc > best_auc:
best_epoch = epoch + 1
best_auc = auc
if best_model != "NOSAVE":
with open(best_model, 'w') as best_state:
pickle.dump(sgd_clf, best_state)
if crossentropy_loss < min_loss:
min_loss = crossentropy_loss
min_loss_epoch = epoch + 1
else:
if epoch + 1 - min_loss_epoch >= es_patience:
print 'Early stopping criterion has reached. Stop training.'
break
if epoch + 1 - min_loss_epoch >= lr_patience:
curr_lr *= .1
sgd_clf.set_params(eta0=curr_lr)
print "Reducing learning rate to: %s" % (curr_lr)
# End of training summary
print ">>> Found best AUROC: %.4f at epoch: %d, saved to: %s <<<" % \
(best_auc, best_epoch, best_model)
print ">>> Found best val loss: %.4f at epoch: %d. <<<" % \
(min_loss, min_loss_epoch)
#### Save elastic net model!! ####
if final_model != "NOSAVE":
with open(final_model, 'w') as final_state:
pickle.dump(sgd_clf, final_state)
def run(img_folder, dl_state, fprop_mode=False,
img_size=(1152, 896), img_height=None, img_scale=None,
rescale_factor=None,
equalize_hist=False, featurewise_center=False, featurewise_mean=71.8,
net='vgg19', batch_size=128, patch_size=256, stride=8,
avg_pool_size=(7, 7), hm_strides=(1, 1),
pat_csv='./full_img/pat.csv', pat_list=None,
out='./output/prob_heatmap.pkl'):
'''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))
# Create image generator.
imgen = DMImageDataGenerator(featurewise_center=featurewise_center)
imgen.mean = featurewise_mean
# Get image and label lists.
df = pd.read_csv(pat_csv, header=0)
df = df.set_index(['patient_id', 'side'])
df.sort_index(inplace=True)
if pat_list is not None:
pat_ids = pd.read_csv(pat_list, header=0).values.ravel()
pat_ids = pat_ids.tolist()
print ("Read %d patient IDs" % (len(pat_ids)))
df = df.loc[pat_ids]
# Load DL model, preprocess.
print ("Load patch classifier:", dl_state)
sys.stdout.flush()
dl_model, preprocess_input, _ = get_dl_model(net, resume_from=dl_state)
if fprop_mode:
dl_model = add_top_layers(dl_model, img_size, patch_net=net,
avg_pool_size=avg_pool_size,
return_heatmap=True, hm_strides=hm_strides)
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
if featurewise_center:
preprocess_input = None
# Sweep the whole images and classify patches.
def const_filename(pat, side, view):
basename = '_'.join([pat, side, view]) + '.png'
return os.path.join(img_folder, basename)
print ("Generate prob heatmaps")
sys.stdout.flush()
heatmaps = []
cases_seen = 0
nb_cases = len(df.index.unique())
for i, (pat,side) in enumerate(df.index.unique()):
## DEBUG ##
#if i >= 10:
# break
## DEBUG ##
cancer = df.loc[pat].loc[side]['cancer']
cc_fn = const_filename(pat, side, 'CC')
if os.path.isfile(cc_fn):
if fprop_mode:
cc_x = read_img_for_pred(
cc_fn, equalize_hist=equalize_hist, data_format=data_format,
dup_3_channels=True,
transformer=imgen.random_transform,
standardizer=imgen.standardize,
target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor)
cc_x = cc_x.reshape((1,) + cc_x.shape)
cc_hm = dl_model.predict_on_batch(cc_x)[0]
# import pdb; pdb.set_trace()
else:
cc_hm = get_prob_heatmap(
cc_fn, 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)
else:
cc_hm = None
mlo_fn = const_filename(pat, side, 'MLO')
if os.path.isfile(mlo_fn):
if fprop_mode:
mlo_x = read_img_for_pred(
mlo_fn, equalize_hist=equalize_hist, data_format=data_format,
dup_3_channels=True,
transformer=imgen.random_transform,
standardizer=imgen.standardize,
target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor)
mlo_x = mlo_x.reshape((1,) + mlo_x.shape)
mlo_hm = dl_model.predict_on_batch(mlo_x)[0]
else:
mlo_hm = get_prob_heatmap(
mlo_fn, 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)
else:
mlo_hm = None
heatmaps.append({'patient_id':pat, 'side':side, 'cancer':cancer,
'cc':cc_hm, 'mlo':mlo_hm})
print ("scored %d/%d cases" % (i + 1, nb_cases))
sys.stdout.flush()
print ("Done.")
# Save the result.
print ("Saving result to external files.",)
sys.stdout.flush()
pickle.dump(heatmaps, open(out, 'w'))
print ("Done.")
def run(train_dir, val_dir, test_dir, patch_model_state=None, resume_from=None,
img_size=[1152, 896], img_scale=None, rescale_factor=None,
featurewise_center=True, featurewise_mean=52.16,
equalize_hist=False, augmentation=True,
class_list=['neg', 'pos'], patch_net='resnet50',
block_type='resnet', top_depths=[512, 512], top_repetitions=[3, 3],
bottleneck_enlarge_factor=4,
add_heatmap=False, avg_pool_size=[7, 7],
add_conv=True, add_shortcut=False,
hm_strides=(1,1), hm_pool_size=(5,5),
fc_init_units=64, fc_layers=2,
top_layer_nb=None,
batch_size=64, train_bs_multiplier=.5,
nb_epoch=5, all_layer_epochs=20,
load_val_ram=False, load_train_ram=False,
weight_decay=.0001, hidden_dropout=.0,
weight_decay2=.0001, hidden_dropout2=.0,
optim='sgd', init_lr=.01, lr_patience=10, es_patience=25,
auto_batch_balance=False, pos_cls_weight=1.0, neg_cls_weight=1.0,
all_layer_multiplier=.1,
best_model='./modelState/image_clf.h5',
final_model="NOSAVE"):
'''Train a deep learning model for image classifications
'''
# ======= Environmental 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))
# ========= Image generator ============== #
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()
# Add augmentation options.
if augmentation:
train_imgen.horizontal_flip = True
train_imgen.vertical_flip = True
train_imgen.rotation_range = 25. # in degree.
train_imgen.shear_range = .2 # in radians.
train_imgen.zoom_range = [.8, 1.2] # in proportion.
train_imgen.channel_shift_range = 20. # in pixel intensity values.
# ================= Model creation ============== #
if resume_from is not None:
image_model = load_model(resume_from, compile=False)
else:
patch_model = load_model(patch_model_state, compile=False)
image_model, top_layer_nb = add_top_layers(
patch_model, img_size, patch_net, block_type,
top_depths, top_repetitions, bottleneck_org,
nb_class=len(class_list), shortcut_with_bn=True,
bottleneck_enlarge_factor=bottleneck_enlarge_factor,
dropout=hidden_dropout, weight_decay=weight_decay,
add_heatmap=add_heatmap, avg_pool_size=avg_pool_size,
add_conv=add_conv, add_shortcut=add_shortcut,
hm_strides=hm_strides, hm_pool_size=hm_pool_size,
fc_init_units=fc_init_units, fc_layers=fc_layers)
if gpu_count > 1:
image_model, org_model = make_parallel(image_model, gpu_count)
else:
org_model = image_model
# ============ Train & validation set =============== #
train_bs = int(batch_size*train_bs_multiplier)
if patch_net != 'yaroslav':
dup_3_channels = True
else:
dup_3_channels = False
if load_train_ram:
raw_imgen = DMImageDataGenerator()
print "Create generator for raw train set"
raw_generator = raw_imgen.flow_from_directory(
train_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, 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=auto_batch_balance,
shuffle=True, seed=random_seed)
else:
print "Create generator for train set"
train_generator = train_imgen.flow_from_directory(
train_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, class_mode='categorical',
auto_batch_balance=auto_batch_balance, batch_size=train_bs,
shuffle=True, seed=random_seed)
print "Create generator for val set"
validation_set = val_imgen.flow_from_directory(
val_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, class_mode='categorical',
batch_size=batch_size, shuffle=False)
sys.stdout.flush()
if load_val_ram:
print "Loading validation set into RAM.",
sys.stdout.flush()
validation_set = load_dat_ram(validation_set, validation_set.nb_sample)
print "Done."; sys.stdout.flush()
# ==================== Model training ==================== #
# Do 2-stage training.
train_batches = int(train_generator.nb_sample/train_bs) + 1
if isinstance(validation_set, tuple):
val_samples = len(validation_set[0])
else:
val_samples = validation_set.nb_sample
validation_steps = int(val_samples/batch_size)
#### DEBUG ####
# train_batches = 1
# val_samples = batch_size*5
# validation_steps = 5
#### DEBUG ####
if load_val_ram:
auc_checkpointer = DMAucModelCheckpoint(
best_model, validation_set, batch_size=batch_size)
else:
auc_checkpointer = DMAucModelCheckpoint(
best_model, validation_set, test_samples=val_samples)
# import pdb; pdb.set_trace()
image_model, loss_hist, acc_hist = do_2stage_training(
image_model, org_model, train_generator, validation_set, validation_steps,
best_model, train_batches, top_layer_nb, nb_epoch=nb_epoch,
all_layer_epochs=all_layer_epochs,
optim=optim, init_lr=init_lr,
all_layer_multiplier=all_layer_multiplier,
es_patience=es_patience, lr_patience=lr_patience,
auto_batch_balance=auto_batch_balance,
pos_cls_weight=pos_cls_weight, neg_cls_weight=neg_cls_weight,
nb_worker=nb_worker, auc_checkpointer=auc_checkpointer,
weight_decay=weight_decay, hidden_dropout=hidden_dropout,
weight_decay2=weight_decay2, hidden_dropout2=hidden_dropout2,)
# Training report.
if len(loss_hist) > 0:
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
if final_model != "NOSAVE":
image_model.save(final_model)
# ==== Predict on test set ==== #
print "\n==== Predicting on test set ===="
test_generator = test_imgen.flow_from_directory(
test_dir, target_size=img_size, target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist, dup_3_channels=dup_3_channels,
classes=class_list, class_mode='categorical', batch_size=batch_size,
shuffle=False)
test_samples = test_generator.nb_sample
#### DEBUG ####
# test_samples = 5
#### DEBUG ####
print "Test samples =", test_samples
print "Load saved best model:", best_model + '.',
sys.stdout.flush()
org_model.load_weights(best_model)
print "Done."
# test_steps = int(test_generator.nb_sample/batch_size)
# test_res = image_model.evaluate_generator(
# test_generator, test_steps, nb_worker=nb_worker,
# pickle_safe=True if nb_worker > 1 else False)
test_auc = DMAucModelCheckpoint.calc_test_auc(
test_generator, image_model, test_samples=test_samples)
print "AUROC on test set:", test_auc
def run(train_dir,
val_dir,
test_dir,
img_size=[256, 256],
img_scale=None,
rescale_factor=None,
featurewise_center=True,
featurewise_mean=59.6,
equalize_hist=True,
augmentation=False,
class_list=['background', 'malignant', 'benign'],
batch_size=64,
train_bs_multiplier=.5,
nb_epoch=5,
top_layer_epochs=10,
all_layer_epochs=20,
load_val_ram=False,
load_train_ram=False,
net='resnet50',
use_pretrained=True,
nb_init_filter=32,
init_filter_size=5,
init_conv_stride=2,
pool_size=2,
pool_stride=2,
weight_decay=.0001,
weight_decay2=.0001,
alpha=.0001,
l1_ratio=.0,
inp_dropout=.0,
hidden_dropout=.0,
hidden_dropout2=.0,
optim='sgd',
init_lr=.01,
lr_patience=10,
es_patience=25,
resume_from=None,
auto_batch_balance=False,
pos_cls_weight=1.0,
neg_cls_weight=1.0,
top_layer_nb=None,
top_layer_multiplier=.1,
all_layer_multiplier=.01,
best_model='./modelState/patch_clf.h5',
final_model="NOSAVE"):
'''Train a deep learning model for patch classifications
'''
best_model_dir = os.path.dirname(best_model)
if not os.path.exists(best_model_dir):
os.makedirs(best_model_dir)
if final_model != "NOSAVE":
final_model_dir = os.path.dirname(final_model)
if not os.path.exists(final_model_dir):
os.makedirs(final_model_dir)
# ======= Environmental 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))
# ========= Image generator ============== #
if featurewise_center:
print "Using feature-wise centering, mean:", featurewise_mean
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()
# Add augmentation options.
if augmentation:
train_imgen.horizontal_flip = True
train_imgen.vertical_flip = True
train_imgen.rotation_range = 25. # in degree.
train_imgen.shear_range = .2 # in radians.
train_imgen.zoom_range = [.8, 1.2] # in proportion.
train_imgen.channel_shift_range = 20. # in pixel intensity values.
# ================= Model creation ============== #
model, preprocess_input, top_layer_nb = get_dl_model(
net,
nb_class=len(class_list),
use_pretrained=use_pretrained,
resume_from=resume_from,
img_size=img_size,
top_layer_nb=top_layer_nb,
weight_decay=weight_decay,
hidden_dropout=hidden_dropout,
nb_init_filter=nb_init_filter,
init_filter_size=init_filter_size,
init_conv_stride=init_conv_stride,
pool_size=pool_size,
pool_stride=pool_stride,
alpha=alpha,
l1_ratio=l1_ratio,
inp_dropout=inp_dropout)
if featurewise_center:
preprocess_input = None
if gpu_count > 1:
model, org_model = make_parallel(model, gpu_count)
else:
org_model = model
# ============ Train & validation set =============== #
train_bs = int(batch_size * train_bs_multiplier)
if net != 'yaroslav':
dup_3_channels = True
else:
dup_3_channels = False
if load_train_ram:
raw_imgen = DMImageDataGenerator()
print "Create generator for raw train set"
raw_generator = raw_imgen.flow_from_directory(
train_dir,
target_size=img_size,
target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
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=auto_batch_balance,
preprocess=preprocess_input,
shuffle=True,
seed=random_seed)
else:
print "Create generator for train set"
train_generator = train_imgen.flow_from_directory(
train_dir,
target_size=img_size,
target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
class_mode='categorical',
auto_batch_balance=auto_batch_balance,
batch_size=train_bs,
preprocess=preprocess_input,
shuffle=True,
seed=random_seed)
print "Create generator for val set"
validation_set = val_imgen.flow_from_directory(
val_dir,
target_size=img_size,
target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
class_mode='categorical',
batch_size=batch_size,
preprocess=preprocess_input,
shuffle=False)
sys.stdout.flush()
if load_val_ram:
print "Loading validation set into RAM.",
sys.stdout.flush()
validation_set = load_dat_ram(validation_set, validation_set.nb_sample)
print "Done."
sys.stdout.flush()
# ==================== Model training ==================== #
# Do 3-stage training.
train_batches = int(train_generator.nb_sample / train_bs) + 1
if isinstance(validation_set, tuple):
val_samples = len(validation_set[0])
else:
val_samples = validation_set.nb_sample
validation_steps = int(val_samples / batch_size)
#### DEBUG ####
# val_samples = 100
#### DEBUG ####
# import pdb; pdb.set_trace()
model, loss_hist, acc_hist = do_3stage_training(
model,
org_model,
train_generator,
validation_set,
validation_steps,
best_model,
train_batches,
top_layer_nb,
net,
nb_epoch=nb_epoch,
top_layer_epochs=top_layer_epochs,
all_layer_epochs=all_layer_epochs,
use_pretrained=use_pretrained,
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=auto_batch_balance,
nb_class=len(class_list),
pos_cls_weight=pos_cls_weight,
neg_cls_weight=neg_cls_weight,
nb_worker=nb_worker,
weight_decay2=weight_decay2,
hidden_dropout2=hidden_dropout2)
# Training report.
if len(loss_hist) > 0:
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
if final_model != "NOSAVE":
model.save(final_model)
# ==== Predict on test set ==== #
print "\n==== Predicting on test set ===="
test_generator = test_imgen.flow_from_directory(
test_dir,
target_size=img_size,
target_scale=img_scale,
rescale_factor=rescale_factor,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
class_mode='categorical',
batch_size=batch_size,
preprocess=preprocess_input,
shuffle=False)
if test_generator.nb_sample:
print "Test samples =", test_generator.nb_sample
print "Load saved best model:", best_model + '.',
sys.stdout.flush()
org_model.load_weights(best_model)
print "Done."
test_steps = int(test_generator.nb_sample / batch_size)
#### DEBUG ####
# test_samples = 10
#### DEBUG ####
test_res = model.evaluate_generator(
test_generator,
test_steps,
nb_worker=nb_worker,
pickle_safe=True if nb_worker > 1 else False)
print "Evaluation result on test set:", test_res
else:
print "Skip testing because no test sample is found."
def run(train_dir,
val_dir,
test_dir,
img_size=[256, 256],
img_scale=255.,
featurewise_center=True,
featurewise_mean=59.6,
equalize_hist=True,
augmentation=False,
class_list=['background', 'malignant', 'benign'],
batch_size=64,
train_bs_multiplier=.5,
nb_epoch=5,
top_layer_epochs=10,
all_layer_epochs=20,
load_val_ram=False,
load_train_ram=False,
net='resnet50',
use_pretrained=True,
nb_init_filter=32,
init_filter_size=5,
init_conv_stride=2,
pool_size=2,
pool_stride=2,
weight_decay=.0001,
weight_decay2=.0001,
bias_multiplier=.1,
alpha=.0001,
l1_ratio=.0,
inp_dropout=.0,
hidden_dropout=.0,
hidden_dropout2=.0,
optim='sgd',
init_lr=.01,
lr_patience=10,
es_patience=25,
resume_from=None,
auto_batch_balance=False,
pos_cls_weight=1.0,
neg_cls_weight=1.0,
top_layer_nb=None,
top_layer_multiplier=.1,
all_layer_multiplier=.01,
best_model='./modelState/patch_clf.h5',
final_model="NOSAVE"):
'''Train a deep learning model for patch classifications
'''
# ======= Environmental 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))
# ========= Image generator ============== #
# if use_pretrained: # use pretrained model's preprocessing.
# train_imgen = DMImageDataGenerator()
# val_imgen = DMImageDataGenerator()
if featurewise_center:
# fitgen = DMImageDataGenerator()
# # Calculate pixel-level mean and std.
# print "Create generator for mean and std fitting"
# fit_patch_generator = fitgen.flow_from_directory(
# train_dir, target_size=img_size, target_scale=img_scale,
# classes=class_list, class_mode=None, batch_size=batch_size,
# shuffle=True, seed=random_seed)
# sys.stdout.flush()
# fit_X_lst = []
# patches_seen = 0
# while patches_seen < fit_size:
# X = fit_patch_generator.next()
# fit_X_lst.append(X)
# patches_seen += len(X)
# fit_X_arr = np.concatenate(fit_X_lst)
train_imgen = DMImageDataGenerator(featurewise_center=True)
# featurewise_std_normalization=True)
val_imgen = DMImageDataGenerator(featurewise_center=True)
test_imgen = DMImageDataGenerator(featurewise_center=True)
# featurewise_std_normalization=True)
# train_imgen.fit(fit_X_arr)
# print "Found mean=%.2f, std=%.2f" % (train_imgen.mean, train_imgen.std)
# sys.stdout.flush()
train_imgen.mean = featurewise_mean
val_imgen.mean = featurewise_mean
test_imgen.mean = featurewise_mean
# del fit_X_arr, fit_X_lst
else:
train_imgen = DMImageDataGenerator()
val_imgen = DMImageDataGenerator()
test_imgen = DMImageDataGenerator()
# train_imgen = DMImageDataGenerator(
# samplewise_center=True,
# samplewise_std_normalization=True)
# val_imgen = DMImageDataGenerator(
# samplewise_center=True,
# samplewise_std_normalization=True)
# Add augmentation options.
if augmentation:
train_imgen.horizontal_flip = True
train_imgen.vertical_flip = True
train_imgen.rotation_range = 45.
train_imgen.shear_range = np.pi / 8.
# ================= Model creation ============== #
model, preprocess_input, top_layer_nb = get_dl_model(
net,
nb_class=len(class_list),
use_pretrained=use_pretrained,
resume_from=resume_from,
img_size=img_size,
top_layer_nb=top_layer_nb,
weight_decay=weight_decay,
bias_multiplier=bias_multiplier,
hidden_dropout=hidden_dropout,
nb_init_filter=nb_init_filter,
init_filter_size=init_filter_size,
init_conv_stride=init_conv_stride,
pool_size=pool_size,
pool_stride=pool_stride,
alpha=alpha,
l1_ratio=l1_ratio,
inp_dropout=inp_dropout)
if featurewise_center:
preprocess_input = None
if gpu_count > 1:
model, org_model = make_parallel(model, gpu_count)
else:
org_model = model
# ============ Train & validation set =============== #
train_bs = int(batch_size * train_bs_multiplier)
if use_pretrained:
dup_3_channels = True
else:
dup_3_channels = False
if load_train_ram:
raw_imgen = DMImageDataGenerator()
print "Create generator for raw train set"
raw_generator = raw_imgen.flow_from_directory(
train_dir,
target_size=img_size,
target_scale=img_scale,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
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=auto_batch_balance,
preprocess=preprocess_input,
shuffle=True,
seed=random_seed)
else:
print "Create generator for train set"
train_generator = train_imgen.flow_from_directory(
train_dir,
target_size=img_size,
target_scale=img_scale,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
class_mode='categorical',
auto_batch_balance=auto_batch_balance,
batch_size=train_bs,
preprocess=preprocess_input,
shuffle=True,
seed=random_seed)
# import pdb; pdb.set_trace()
print "Create generator for val set"
validation_set = val_imgen.flow_from_directory(
val_dir,
target_size=img_size,
target_scale=img_scale,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
class_mode='categorical',
batch_size=batch_size,
preprocess=preprocess_input,
shuffle=False)
sys.stdout.flush()
if load_val_ram:
print "Loading validation set into RAM.",
sys.stdout.flush()
validation_set = load_dat_ram(validation_set, validation_set.nb_sample)
print "Done."
sys.stdout.flush()
# ==================== Model training ==================== #
# Callbacks and class weight.
early_stopping = EarlyStopping(monitor='val_loss',
patience=es_patience,
verbose=1)
checkpointer = ModelCheckpoint(best_model,
monitor='val_acc',
verbose=1,
save_best_only=True)
stdout_flush = DMFlush()
callbacks = [early_stopping, checkpointer, stdout_flush]
if optim == 'sgd':
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.5,
patience=lr_patience,
verbose=1)
callbacks.append(reduce_lr)
if auto_batch_balance:
class_weight = None
elif len(class_list) == 2:
class_weight = {0: 1.0, 1: pos_cls_weight}
elif len(class_list) == 3:
class_weight = {0: 1.0, 1: pos_cls_weight, 2: neg_cls_weight}
else:
class_weight = None
# Do 3-stage training.
train_batches = int(train_generator.nb_sample / train_bs) + 1
samples_per_epoch = train_bs * train_batches
#### DEBUG ####
# samples_per_epoch = train_bs*10
#### DEBUG ####
if isinstance(validation_set, tuple):
val_samples = len(validation_set[0])
else:
val_samples = validation_set.nb_sample
#### DEBUG ####
# val_samples = 100
#### DEBUG ####
model, loss_hist, acc_hist = do_3stage_training(
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=use_pretrained,
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=auto_batch_balance,
pos_cls_weight=pos_cls_weight,
neg_cls_weight=neg_cls_weight,
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
if final_model != "NOSAVE":
model.save(final_model)
# ==== Predict on test set ==== #
print "\n==== Predicting on test set ===="
test_generator = test_imgen.flow_from_directory(
test_dir,
target_size=img_size,
target_scale=img_scale,
equalize_hist=equalize_hist,
dup_3_channels=dup_3_channels,
classes=class_list,
class_mode='categorical',
batch_size=batch_size,
preprocess=preprocess_input,
shuffle=False)
print "Test samples =", test_generator.nb_sample
print "Load saved best model:", best_model + '.',
sys.stdout.flush()
org_model.load_weights(best_model)
print "Done."
test_samples = test_generator.nb_sample
#### DEBUG ####
# test_samples = 10
#### DEBUG ####
test_res = 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