def test_overfit(self):
print("\n======== TestCaptioningRNN.test_overfit:")
small_data = coco_utils.load_coco_data(max_train=50)
small_rnn_model = captioning_rnn.CaptioningRNN(
cell_type='rnn',
word_to_idx=small_data['word_to_idx'],
input_dim=small_data['train_features'].shape[1],
hidden_dim=512,
wordvec_dim=256)
solv = captioning_solver.CaptioningSolver(
small_rnn_model,
small_data,
update_rule='adam',
num_epochs=50,
batch_size=25,
optim_config={'learning_rate': 5e-3},
lr_decay=0.95,
verbose=self.verbose,
print_every=1)
solv.train()
if self.draw_figures:
plt.plot(solv.loss_history)
plt.xlabel('Iteration')
plt.ylabel('Loss')
plt.title("Training loss history")
plt.show()
print("======== TestCaptioningRNN.test_overfit: <END> ")
def ex_caption_rnn(verbose=False, plot_figures=False):
test_data = coco_utils.load_coco_data(max_train=50000)
small_rnn_model = captioning_rnn.CaptioningRNN(
cell_type='rnn',
word_to_idx=test_data['word_to_idx'],
input_dim=test_data['train_features'].shape[1],
hidden_dim=512,
wordvec_dim=256
)
solv = captioning_solver.CaptioningSolver(
small_rnn_model,
test_data,
update_rule='adam',
num_epochs=50,
batch_size=25,
optim_config={'learning_rate': 5e-3},
lr_decay=0.95,
verbose=verbose,
print_every=100
)
solv.train()
test_time_sampling(test_data,
small_rnn_model,
plot_figures=plot_figures)
# Plot the loss...
if plot_figures:
plt.plot(solv.loss_history, 'o')
plt.xlabel('Iterations')
plt.ylabel('Loss')
plt.title('Captioning RNN Loss')
plt.show()
def load_data(verbose=False):
data = coco_utils.load_coco_data(pca_features=True)
if verbose:
for k, v in data.items():
if type(v) == np.ndarray:
print('%s : %s, %s, %%d' % (k, type(v), v.shape, v.dtype))
else:
print('%s : %s (%s)' % (k, type(v), len(v)))
return data
def test_sampling(self):
print("\n======== TestCaptioningRNN.test_sampling:")
small_data = coco_utils.load_coco_data(max_train=50)
small_rnn_model = captioning_rnn.CaptioningRNN(
cell_type='rnn',
word_to_idx=small_data['word_to_idx'],
input_dim=small_data['train_features'].shape[1],
hidden_dim=512,
wordvec_dim=256)
solv = captioning_solver.CaptioningSolver(
small_rnn_model,
small_data,
update_rule='adam',
num_epochs=50,
batch_size=25,
optim_config={'learning_rate': 5e-3},
lr_decay=0.95,
verbose=self.verbose,
print_every=1)
# Train the model
solv.train()
for split in ['train', 'val']:
minibatch = coco_utils.sample_coco_minibatch(small_data,
split=split,
batch_size=8)
gt_captions, features, urls = minibatch
gt_captions = coco_utils.decode_captions(gt_captions,
small_data['idx_to_word'])
sample_captions = small_rnn_model.sample(features)
sample_captions = sample_captions.astype(np.int32)
sample_captions = coco_utils.decode_captions(
sample_captions, small_data['idx_to_word'])
for gt_capt, samp_capt, url in zip(gt_captions, sample_captions,
urls):
plt.imshow(image_utils.image_from_url(url))
plt.title('Split : %s\n%5s\nGT :%s' %
(split, samp_capt, gt_capt))
plt.axis('off')
plt.show()
print("======== TestCaptioningRNN.test_sampling: <END> ")
def test_overfit_model(self):
print("\n======== TestCaptioningLSTM.test_overfit_model:")
small_data = coco_utils.load_coco_data(max_train=50)
small_lstm_model = captioning_rnn.CaptioningRNN(
cell_type='lstm',
word_to_idx=small_data['word_to_idx'],
input_dim=small_data['train_features'].shape[1],
hidden_dim=512,
wordvec_dim=256,
dtype=np.float32)
small_lstm_solv = captioning_solver.CaptioningSolver(
small_lstm_model,
small_data,
update_rule='adam',
num_epochs=50,
batch_size=25,
optim_config={'learning_rate': 5e-3},
lr_decay=0.95,
print_every=10,
verbose=self.verbose)
# Train
small_lstm_solv.train()
if self.draw_figures:
#fig, ax = vis_solver.get_train_fig()
#vis_solver.plot_solver(ax, small_lstm_solv)
plt.plot(small_lstm_solv.loss_history)
plt.xlabel('Iteration')
plt.ylabel('Loss')
plt.title('Training loss history')
plt.show()
print("\n======== TestCaptioningLSTM.test_overfit_model: <END>")