def get_accuracy_iterator(mod, Xdatagen, k=1):
acc = 0
npoints = 0
for i in range(len(Xdatagen)):
Xfr, yfr = next(Xdatagen)
ypred = mod.predict(Xfr)
npoints += len(yfr)
acc += np.sum(top_k_categorical_accuracy(yfr, ypred, k=k))
del Xfr
del yfr
del ypred
gc.collect()
return acc / npoints
batch_size=train_batch_size)
valid_batches = ImageDataGenerator(
preprocessing_function=
tensorflow.keras.applications.xception.preprocess_input).flow_from_directory(
valid_path,
target_size=(image_size, image_size),
batch_size=val_batch_size)
# Train the model
# Define Top2 and Top3 Accuracy
from tensorflow.keras.metrics import categorical_accuracy, top_k_categorical_accuracy
def top_3_accuracy(y_true, y_pred)
return top_k_categorical_accuracy(y_true, y_pred, k=3)
def top_2_accuracy(y_true, y_pred)
return top_k_categorical_accuracy(y_true, y_pred, k=2)
# Compile the model
model.compile(Adam(lr=0.01), loss='categorical_crossentropy', metrics=[categorical_accuracy, top_2_accuracy, top_3_accuracy])
# Add weights to make the model more sensitive to melanoma
# Declare the filepath for the saved model
filepath = model_xception_crop.h5
# Declare a checkpoint to save the best version of the model
checkpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=1,
save_best_only=True, mode='min')
def top_2_accuracy(y_true, y_pred):
"""Calculate Top2 Accuracy
return: Top2 accuracy
"""
return top_k_categorical_accuracy(y_true, y_pred, k=2)
def top_3_accuracy(y_true, y_pred):
"""Calculate Top3 Accuracy
return: top3 accuracy
"""
return top_k_categorical_accuracy(y_true, y_pred, k=3)
def top3_accuracy(y, predic):
return metrics.top_k_categorical_accuracy(y, predic, k=3)
def top_2_accuracy(x, y):
return top_k_categorical_accuracy(x, y, k=2)
def top_5_accuracy(x, y): return top_k_categorical_accuracy(x, y, 5) def create_model_fn(params, is_optimizer_adam=True):
def top_3_accuracy(x, y):
t3 = top_k_categorical_accuracy(x, y, 3)
return t3
def top_5_categorical_error(y_true, y_pred):
return 1.0 - metrics.top_k_categorical_accuracy(y_true, y_pred, 5)
def top_1_categorical_accuracy(y_true, y_pred):
s = K.prod(K.shape(y_true)[1:])
y_true = K.reshape(y_true, [-1, s])
y_pred = K.reshape(y_pred, [-1, s])
return metrics.top_k_categorical_accuracy(y_true, y_pred, k=1)
def train_model_with_augmentations(model,
datagen,
X_test,
Y_test,
epochs=100,
batch_size=256,
batches_per_epoch=4,
patience=15,
savename="model.h5"):
""" Train the model until it can not be trained anymore! """
train_loss = []
valid_loss = []
train_top1_acc = []
valid_top1_acc = []
train_top5_acc = []
valid_top5_acc = []
down_rounds = 0
best_loss = np.inf
Y_test_cat = to_categorical(Y_test)
# Create a data flow
data_flow = datagen.flow(X_train,
Y_train,
batch_size=batch_size,
shuffle=True)
for epoch in range(epochs):
batch_loss = []
batch_top1 = []
batch_top5 = []
for batch in range(batches_per_epoch):
X_batch, Y_batch = next(data_flow)
step_loss = model.train_on_batch(X_batch, Y_batch)
batch_loss.append(step_loss)
batch_top1.append(
np.mean(
top_k_categorical_accuracy(Y_batch,
model.predict(X_batch),
k=1)))
batch_top5.append(
np.mean(
top_k_categorical_accuracy(Y_batch,
model.predict(X_batch),
k=5)))
# End of epoch logging
Y_pred = model.predict(X_test)
valid_top1_acc.append(
np.mean(
top_k_categorical_accuracy(Y_test_cat,
model.predict(X_test),
k=1)))
valid_top5_acc.append(
np.mean(
top_k_categorical_accuracy(Y_test_cat,
model.predict(X_test),
k=5)))
train_top1_acc.append(np.mean(batch_top1))
train_top5_acc.append(np.mean(batch_top5))
train_loss.append(np.mean(batch_loss))
valid_loss.append(model.evaluate(X_test, Y_test_cat))
# Model checkpoint and early stopping.
if valid_loss[epoch] < best_loss:
best_loss = valid_loss[epoch]
down_rounds = 0
model.save(savename)
print(
"[New Best] Epoch {0}, Training Loss: {1:8.4f}, Testing Loss: {2:8.4f}, Training Top-1: {3:8.4f}, Testing Top-1: {4:8.4f}"
.format(epoch, train_loss[epoch], valid_loss[epoch],
train_top1_acc[epoch], valid_top1_acc[epoch]))
else:
down_rounds += 1
if down_rounds >= patience:
print("Earlying stopping at epoch {}.".format(epoch))
break
return train_loss, valid_loss, train_top1_acc, valid_top1_acc, train_top5_acc, valid_top5_acc
def top_3_accuracy(y_test, y_pred):
return top_k_categorical_accuracy(y_test, y_pred, k=3)
def top_n_categorical_accuracy(y_pred, y_label):
return metrics.top_k_categorical_accuracy(y_pred, y_label, k=TOP_N)
def top_3_metrics(y_true, y_pred):
return top_k_categorical_accuracy(y_true, y_pred, k=3)
def top_5_accuracy(self, y_true, y_pred): y_true = y_true[:, :self.num_class] y_pred = y_pred[:, :self.num_class] return top_k_categorical_accuracy(y_true, y_pred)
def top_2_accuracy(y_true, y_pred):
return top_k_categorical_accuracy(y_true, y_pred, k=2)
def top_5_accuracy(self, y_true, y_pred):
return top_k_categorical_accuracy(y_true, y_pred, k=5)
def top_50_accuracy(y_true,y_pred):
return metrics.top_k_categorical_accuracy(y_true,y_pred,k=50)
def top_2_accuracy(in_gt, in_pred):
return top_k_categorical_accuracy(in_gt, in_pred, k=2)
