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,
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(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,
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,
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(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