abb_losses = [] #keeps track of loss values for ABB words
training_losses = [] #keeps track of loss on training data each run
make_ABB = lambda x: x[:1] + [1.0] + x[1 + 1:] #Adds semantic information
make_ABA = lambda x: x[:1] + [-1.0] + x[1 + 1:] #Adds semantic information
for rep in range(REPS):
print("Rep: " + str(rep))
#Erase the previous model:
keras.backend.clear_session()
#Build the new model:
model = Seq2Seq.seq2seq(input_dim=FEAT_NUM,
hidden_dim=FEAT_NUM,
output_length=2,
output_dim=FEAT_NUM,
batch_size=1,
learn_rate=0.001,
layer_type="lstm",
dropout=DROPOUT)
#PRETRAINING
if VOCAB_SIZE > 0:
print("Simulating real-life experience of infants...Rep=" + str(rep))
irl_X = []
irl_Y = []
for word in range(VOCAB_SIZE):
##Putting reduplication in training:
if np.random.rand() < REDUP_IN_PT:
syll_alpha = choice(all_sylls)
template = choice(["ABB", "ABA"])
if template == "ABB":
elif SCOPE == "segment":
feat_num, withheld_syll, syllables = novel_seg_data()
elif SCOPE == "syllable":
feat_num, withheld_syll, syllables = novel_syll_data()
else:
raise Exception(
"Wrong scope! Must be from the set {feature, segment, syllable}.")
X = np.array(syllables)
Y = np.array([syll + syll for syll in syllables])
#Build the model:
model = Seq2Seq.seq2seq(input_dim=feat_num,
hidden_dim=feat_num * 3,
output_length=Y.shape[1],
output_dim=Y.shape[2],
batch_size=1,
learn_rate=0.001,
layer_type="lstm",
dropout=DROPOUT)
#Train the model:
hist = model.train(X, Y, epoch_num=EPOCHS, print_every=10)
learning_curves.append(hist["Loss"])
#Test the model on trained data:
trained_IN = np.tile(X[0], (1, 1, 1))
trained_OUT = np.tile(Y[0], (1, 1, 1))
train_pred = model.predict(trained_IN)
#Test the model on withheld data:
withheld_IN = np.tile(np.array(withheld_syll), (1, 1, 1))
#Shuffle training data:
indexes = list(range(len(ordered_X)))
shuffle(indexes)
X = np.array([ordered_X[i] for i in indexes])
Y = np.array([ordered_Y[i] for i in indexes])
#Create the model object:
#model = Seq2Seq(input_dim=FEAT_NUM, hidden_dim=FEAT_NUM*3, output_length=3, output_dim=FEAT_NUM, depth=2)
#Build the new model:
model = Seq2Seq.seq2seq(
input_dim=FEAT_NUM,
hidden_dim=FEAT_NUM * 3,
output_length=3,
output_dim=FEAT_NUM,
batch_size=1,
learn_rate=0.005,
layer_type="lstm",
)
this_curve = []
for ep in range(EPOCHS):
#Train the model one epoch at a time,
#so we can give it a forced-choice task at each step:
print("Epoch: " + str(ep))
hist = model.train(X, Y, epoch_num=1, print_every=2)
this_curve.append(hist["Loss"][-1])
for trial_type in accuracies.keys():
corr_loss = np.square(
np.subtract(