checkpointer = ModelCheckpoint(filepath=CHECKPOINTED_WEIGHTS, verbose=1, save_best_only=True, monitor='oneshot_acc') # In[33]: STEPS_PER_EPOCH = 1 + (NUM_TRAIN_TRIPLETS//BATCH_SIZE) VALIDATION_STEPS = (NUM_VAL_TRIPLETS//BATCH_SIZE) + 1 # In[34]: import custom_callbacks; reload(custom_callbacks) from custom_callbacks import LossHistory loss_history = LossHistory(X_val, y_val, test_alphabet_to_index) # In[35]: from keras.optimizers import Adam adam = Adam(1e-3) triplet_net.compile(loss=triplet_loss, optimizer=adam) triplet_net.load_weights(INIT_WEIGHTS) # In[36]: # triplet_net.load_weights(CHECKPOINTED_WEIGHTS)
def train(self,
X_train,
y_train,
X_valid,
y_valid,
batch_size=128,
epochs=50,
class_weight=False,
loss_per_batch=True,
early_stopping_metric='val_loss',
value_cutoff=2.1,
patience=5,
min_delta=0.00001,
deadrelu_filepath=None,
only_first_epoch=True,
best_model_filepath=None,
log_filepath=None,
epochtime_filepath=None):
callbacks = []
early_stopping = EarlyStoppingCustomed(value_cutoff=value_cutoff,
monitor=early_stopping_metric,
min_delta=min_delta,
patience=patience,
verbose=1)
callbacks.append(early_stopping)
if loss_per_batch:
batch_loss_history = LossHistory()
callbacks.append(batch_loss_history)
if deadrelu_filepath:
dead_relu = DeadReluDetector(X_train,
deadrelu_filepath,
only_first_epoch,
verbose=True)
callbacks.append(dead_relu)
if best_model_filepath:
best_model = ModelCheckpoint(filepath=best_model_filepath,
monitor=early_stopping_metric,
verbose=0,
save_best_only=True,
save_weights_only=False,
mode='auto')
callbacks.append(best_model)
if log_filepath:
csv_logger = CSVLoggerCustomed(log_filepath)
callbacks.append(csv_logger)
if epochtime_filepath:
time_callback = TimeHistory(epochtime_filepath)
callbacks.append(time_callback)
if class_weight:
num_pos = y_train.sum()
num_seq = len(y_train)
num_neg = num_seq - num_pos
self.hist = self.model.fit(X_train,
y_train,
class_weight={
True: num_seq / num_pos,
False: num_seq / num_neg
},
batch_size=batch_size,
epochs=epochs,
verbose=2,
validation_data=(X_valid, y_valid),
shuffle=True,
callbacks=callbacks)
else:
self.hist = self.model.fit(X_train,
y_train,
batch_size=batch_size,
epochs=epochs,
verbose=2,
validation_data=(X_valid, y_valid),
shuffle=True,
callbacks=callbacks)
return self.hist, batch_loss_history
def question_1():
global input_shape, X_train, y_train_labels, y_train, X_test, y_test_labels, y_test
print(
"------------------------------------------------------------------------"
)
print("Baseline Model")
print(
"------------------------------------------------------------------------"
)
model1 = baseline_model(input_shape, num_classes)
loss_callback_1 = LossHistory((X_test, y_test))
model1.fit(X_train,
y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_data=(X_test, y_test),
callbacks=[loss_callback_1])
model1.save('model1.h5')
plot_learning_curve(
[loss_callback_1.train_indices, loss_callback_1.test_indices],
[loss_callback_1.train_losses, loss_callback_1.test_losses],
colors=['g-', 'm-'],
labels=['Train loss', 'Test loss'],
title="Loss evolution for Baseline Model",
path="../outputs/q1/plots/train_test_loss_baseline.png",
axlabels=["Iterations", "Loss"])
plot_learning_curve([loss_callback_1.test_indices],
[loss_callback_1.test_acc],
colors=['c-'],
labels=['Test Accuracy'],
title="Accuracy evolution for Baseline Model",
path="../outputs/q1/plots/test_acc_baseline.png",
axlabels=["Iterations", "Accuracy"])
print(
"------------------------------------------------------------------------"
)
print("2 conv layer model")
print(
"------------------------------------------------------------------------"
)
model2 = two_conv_layer_model(input_shape, num_classes)
loss_callback_2 = LossHistory((X_test, y_test))
model2.fit(X_train,
y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_data=(X_test, y_test),
callbacks=[loss_callback_2])
model2.save('model2.h5')
plot_learning_curve(
[loss_callback_2.train_indices, loss_callback_2.test_indices],
[loss_callback_2.train_losses, loss_callback_2.test_losses],
colors=['g-', 'm-'],
labels=['Train loss', 'Test loss'],
title="Loss evolution for 2 conv layered Model",
path="../outputs/q1/plots/train_test_loss_2_conv.png",
axlabels=["Iterations", "Loss"])
plot_learning_curve([loss_callback_1.test_indices],
[loss_callback_1.test_acc],
colors=['c-'],
labels=['Test Accuracy'],
title="Accuracy evolution for 2 conv layered Model",
path="../outputs/q1/plots/test_acc_2_conv.png",
axlabels=["Iterations", "Accuracy"])
print(
"------------------------------------------------------------------------"
)
print("2 conv layer + 1 hidden dense layer model")
print(
"------------------------------------------------------------------------"
)
model3 = two_conv_one_dense_layer_model(input_shape, num_classes)
loss_callback_3 = LossHistory((X_test, y_test))
model3.fit(X_train,
y_train,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_data=(X_test, y_test),
callbacks=[loss_callback_3])
model3.save('model3.h5')
plot_learning_curve(
[loss_callback_3.train_indices, loss_callback_3.test_indices],
[loss_callback_3.train_losses, loss_callback_3.test_losses],
colors=['g-', 'm-'],
labels=['Train loss', 'Test loss'],
title="Loss evolution for 2 Conv + 1 Dense layer config",
path="../outputs/q1/plots/train_test_loss_2_conv_1_dense.png",
axlabels=["Iterations", "Loss"])
plot_learning_curve([loss_callback_3.test_indices],
[loss_callback_3.test_acc],
colors=['c-'],
labels=['Test Accuracy'],
title="Accuracy evolution for 2 conv + 1 dense config",
path="../outputs/q1/plots/test_acc_2_conv_1_dense.png",
axlabels=["Iterations", "Accuracy"])
ids = np.random.choice(X_test.shape[0], 20)
X_samples = X_train[ids]
pred_samples_1 = model1.predict(X_samples)
generate_image_outputs(X_samples,
np.argmax(pred_samples_1, axis=1),
path="../outputs/q1/predictions/baseline")
pred_samples_2 = model2.predict(X_samples)
generate_image_outputs(X_samples,
np.argmax(pred_samples_2, axis=1),
path="../outputs/q1/predictions/2_conv")
pred_samples_3 = model3.predict(X_samples)
generate_image_outputs(X_samples,
np.argmax(pred_samples_3, axis=1),
path="../outputs/q1/predictions/2_conv_1_dense")