output_layer, train_func, test_func, predict_func, get_hidden_func = word_prediction_network(BATCH_SIZE, word_embedding_size, num_words, MAX_SEQ_LEN, WEIGHTS, NUM_UNITS_GRU, learning_rate)
estimator = LasagneNet(output_layer, train_func, test_func, predict_func, get_hidden_func, on_epoch_finished=[SaveParams('save_params','word_embedding', save_interval = 1)])
# estimator.draw_network() # requires networkx package
X_train = {'X': encoded_sequences[:train_split], 'X_mask': masks[:train_split]}
y_train = y[:train_split]
X_test = {'X': encoded_sequences[train_split:test_split], 'X_mask': masks[train_split:test_split]}
y_test = y[train_split:test_split]
train = False
if train:
estimator.fit(X_train, y_train)
else:
estimator.load_weights_from('saved_params')
word2vec_vocab_rev = dict(zip(word2vec_vocab.values(), word2vec_vocab.keys())) # Maps indeces to words.
predictions = estimator.predict(X_test)
predictions = predictions.reshape(-1, num_words + 1) # Reshape into #samples x #words.
n_pred = predictions.shape[0]
for row in range(n_pred):
# Sample a word from output probabilities.
sample = np.random.multinomial(n=1, pvals=predictions[row,:]).argmax()
line = X_test['X'][row] # Get the input line.
line = line[X_test['X_mask'][row].astype('bool')] # Apply mask.
print 'Input: %r' %([word2vec_vocab_rev[w] for w in line]) # Print words in line.
print 'Guess:: %r' %(word2vec_vocab_rev[sample]) # Print predicted word.
print 'Output: %r' %(word2vec_vocab_rev[y_test[row]]) # Print the correct word.
pdb.set_trace()
return train_func(X, y.reshape((-1,1)))
def test_function(X, y):
return test_func(X, y.reshape((-1,1)))
def predict_function(X):
return predict_func(X)
return l_out, train_function, test_function, predict_function, learning_rate
from sklearn import datasets
boston = datasets.load_boston()
num_epochs = 20
train_x, test_x, train_y, test_y = train_test_split(boston['data'], boston['target'], test_size=0.05)
l_output, train_function, test_function, predict_function, learning_rate = create_lasagne_network(train_x.shape[1])
estimator = LasagneNet(l_output, train_function, test_function, predict_function,is_regression=True,batch_iterator_train=BatchIterator(256),max_epochs=num_epochs,
on_epoch_finished=[AdjustVariable('learning_rate',start=0.03, stop=0.000001,end_epoch=num_epochs),
SaveParams('save_params','C:/params/rossman/', save_interval = 50)])
X = {'X' : train_x.astype('float32')}
estimator.fit(X, train_y.astype('float32'))
X = {'X' : test_x.astype('float32')}
pred = estimator.predict(X).reshape((-1,1))
print np.mean((pred-test_y)**2)
pred_sents = []
# For each test example, predict the response.
for idx in xrange(X_test["X"].shape[0]):
pred_words = []
temp = X_test["X"][idx].reshape(1, MAX_SEQ_LEN)
X_new = np.empty_like(temp)
X_new[:] = temp
temp = X_test["X_mask"][idx].reshape(1, MAX_SEQ_LEN)
X_mask_new = np.empty_like(temp)
X_mask_new[:] = temp
# Predict one word at a time, based on the previous predicted words and the query.
for pred_iter in range(10): # FIXME: predicting 10 words. To stop predicting, make the model predict <EOS>.
zero_found = False
prediction = estimator.predict({"X": X_new, "X_mask": X_mask_new}) # Predict current sequence.
prediction = prediction.reshape(-1, num_words + 1).argmax(axis=-1)[
0
] # Take the argmax of the softmax output.
pred_words.append(prediction)
# Find where to place the prediction in X and X_mask.
for i, elem in enumerate(X_mask_new[0]):
if elem == 0:
zero_found = True
X_mask_new[0, i] = 1
X_new[0, i] = prediction
break
if zero_found == False:
# Can't predict more because of MAX_SEQ_LEN.
break
pred_sents.append(pred_words)
