av_train_perf = {"acc": 0, "prec": 0, "rec": 0, "f1": 0}
av_val_perf = {"acc": 0, "prec": 0, "rec": 0, "f1": 0}
for i in range(args.ensemble_size):
print("Building model")
model = ShallowNet(Xs["train"].shape[1], dropout, dense_layers,
dense_layer_units, args.weights)
model.compile(optimizer=Adam(lr=lr), loss="binary_crossentropy")
print("Model built")
history = model.fit(
X=Xs["train"],
y=ys["train"],
batch_size=batch_size,
nb_epoch=epochs,
verbose=1,
validation_data=(Xs["val"], ys["val"]),
shuffle=True,
show_accuracy=True,
callbacks=[
LearningRateScheduler(lambda e: lr_schedule(epochs, lr, e))
])
print("\n".join(map(str, history.history["acc"])),
file=open(
os.path.join(save_path, "train_accs{}.txt".format(i)),
"w"))
print("\n".join(map(str, history.history["loss"])),
file=open(
os.path.join(save_path, "train_losses{}.txt".format(i)),
"w"))
print("\n".join(map(str, history.history["val_acc"])),
save_path = os.path.join(base_save_dir, "lr{};epochs{};dropout{};dense_layers{};dense_layer_units{};batch_size{}".format(*params))
os.makedirs(save_path)
train_preds = np.zeros((Xs["train"].shape[0], args.ensemble_size))
val_preds = np.zeros((Xs["val"].shape[0], args.ensemble_size))
print("Building model")
model = ShallowNet(Xs["train"].shape[1], dropout, dense_layers, dense_layer_units, args.weights)
model.compile(optimizer=Adam(lr=lr), loss="binary_crossentropy")
print("Model built")
history = model.fit(
X=Xs["train"],
y=ys["train"],
batch_size=batch_size,
nb_epoch=epochs,
verbose=1,
validation_data=(Xs["val"], ys["val"]),
shuffle=True,
show_accuracy=True,
)
model.layers.pop()
model.compile(optimizer=Adam(lr=lr), loss="binary_crossentropy")
train_pred = model.predict(X=Xs["train"], batch_size=batch_size, verbose=0)
val_pred = model.predict(X=Xs["val"], batch_size=batch_size, verbose=0)
cou = 0
with open(args.last_layer_file, "w") as layer_file:
for vec in train_pred:
layer_file.write(",".join([str(i) for i in vec]))
layer_file.write("\n")
cou += 1
os.makedirs(save_path)
train_preds = np.zeros((Xs["train"].shape[0], args.ensemble_size))
val_preds = np.zeros((Xs["val"].shape[0], args.ensemble_size))
print("Building model")
model = ShallowNet(Xs["train"].shape[1], dropout, dense_layers,
dense_layer_units, args.weights)
model.compile(optimizer=Adam(lr=lr), loss="binary_crossentropy")
print("Model built")
history = model.fit(
X=Xs["train"],
y=ys["train"],
batch_size=batch_size,
nb_epoch=epochs,
verbose=1,
validation_data=(Xs["val"], ys["val"]),
shuffle=True,
show_accuracy=True,
)
model.layers.pop()
model.compile(optimizer=Adam(lr=lr), loss="binary_crossentropy")
train_pred = model.predict(X=Xs["train"],
batch_size=batch_size,
verbose=0)
val_pred = model.predict(X=Xs["val"], batch_size=batch_size, verbose=0)
cou = 0
with open(args.last_layer_file, "w") as layer_file:
for vec in train_pred:
layer_file.write(",".join([str(i) for i in vec]))
os.makedirs(save_path)
av_train_perf = {"acc": 0, "prec": 0, "rec": 0, "f1": 0}
av_val_perf = {"acc": 0, "prec": 0, "rec": 0, "f1": 0}
for i in range(args.ensemble_size):
print("Building model")
model = ShallowNet(Xs["train"].shape[1], dropout, dense_layers, dense_layer_units, args.weights)
model.compile(optimizer=Adam(lr=lr), loss="binary_crossentropy")
print("Model built")
history = model.fit(
X=Xs["train"],
y=ys["train"],
batch_size=batch_size,
nb_epoch=epochs,
verbose=1,
validation_data=(Xs["val"], ys["val"]),
shuffle=True,
show_accuracy=True,
callbacks=[LearningRateScheduler(lambda e: lr_schedule(epochs, lr, e))]
)
print("\n".join(map(str, history.history["acc"])), file=open(os.path.join(save_path, "train_accs{}.txt".format(i)), "w"))
print("\n".join(map(str, history.history["loss"])), file=open(os.path.join(save_path, "train_losses{}.txt".format(i)), "w"))
print("\n".join(map(str, history.history["val_acc"])), file=open(os.path.join(save_path, "val_accs{}.txt".format(i)), "w"))
print("\n".join(map(str, history.history["val_loss"])), file=open(os.path.join(save_path, "val_losses{}.txt".format(i)), "w"))
train_pred = model.predict_classes(X=Xs["train"], batch_size=batch_size, verbose=0)
train_perf = eval_pred(ys["train"], train_pred)
for met in ["acc", "prec", "rec", "f1"]:
av_train_perf[met] += float(train_perf[met]) / args.ensemble_size
