def test_cnn(root_path, step_max_devacc, max_devacc):
dropout = 0.5
l2_reg = 0.0001
learning_rate = 0.001
gpu = 0
word_lookup_file = './data/word_vectors_pruned_300.txt'
label_id_file = './data/nlu.label_id.txt'
test_feature_file = './data_vds_id_test'
sentence_length = 20
reg_length = 8
word_embed = 300
dh.embedding_size = word_embed
dh.reg_length = reg_length
dh.sentence_length = sentence_length
print('Reading word lookup table...')
id2vect, word2id, id2word = dh.read_word_lookup_table(word_lookup_file)
id2vect = np.asarray(id2vect, dtype=np.float32)
print('Reading label id...')
label2id, id2label = dh.read_label(label_id_file)
print('Reading test data...')
reg2id = {'N/A': 0}
test_word, test_vds, test_reg, test_y = dh.read_data(test_feature_file, label2id, word2id, reg2id)
vbs_size = 308
vocab_size = len(word2id)
reg_size = len(reg2id)
num_class = len(label2id)
cnn = CNN(num_class, id2vect, gpu, l2_reg, dropout, learning_rate, vocab_size, vbs_size, reg_size)
test_acc = cnn.test(root_path, step_max_devacc, max_devacc, test_word, test_vds, test_reg, test_y)
print('Test accuracy: %.3f' % test_acc)
def main():
# Data loaders
tr_loader, va_loader, te_loader, _ = get_train_val_test_loaders(
num_classes=config('cnn.num_classes'))
# Model
model = CNN()
# TODO: define loss function, and optimizer
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(params=model.parameters(), lr=1e-4)
#
print('Number of float-valued parameters:', count_parameters(model))
# Attempts to restore the latest checkpoint if exists
print('Loading cnn...')
model, start_epoch, stats = restore_checkpoint(model,
config('cnn.checkpoint'))
axes = utils.make_cnn_training_plot()
# Evaluate the randomly initialized model
_evaluate_epoch(axes, tr_loader, va_loader, model, criterion, start_epoch,
stats)
# Loop over the entire dataset multiple times
for epoch in range(start_epoch, config('cnn.num_epochs')):
# Train model
_train_epoch(tr_loader, model, criterion, optimizer)
# Evaluate model
_evaluate_epoch(axes, tr_loader, va_loader, model, criterion,
epoch + 1, stats)
# Save model parameters
save_checkpoint(model, epoch + 1, config('cnn.checkpoint'), stats)
print('Finished Training')
y_true, y_pred = [], []
correct, total = 0, 0
running_loss = []
for X, y in va_loader:
with torch.no_grad():
output = model(X)
predicted = predictions(output.data)
y_true.extend(y)
y_pred.extend(predicted)
total += y.size(0)
correct += (predicted == y).sum().item()
running_loss.append(criterion(output, y).item())
print("Validation data accuracies:")
print(confusion_matrix(y_true, y_pred))
# Save figure and keep plot open
utils.save_cnn_training_plot()
utils.hold_training_plot()
def __init__(self, config):
# get config args
self.out_dir = config['output_directory']
self.batch_size = config['batch_size']
self.input_dims = config['input_dimensions']
self.learn_rate = config['learning_rate']
self.z_dim = config['z_dimension']
self.name = config['model_name']
self.verbosity = config['verbosity']
self.num_train = config['number_train']
self.num_test = config['number_test']
# initialize logging to create new log file and log any level event
logging.basicConfig(format='%(asctime)s %(message)s',
datefmt='%m/%d/%Y %I:%M:%S %p',
filename='{}{}.log'.format(self.out_dir,
self.name),
filemode='w',
level=logging.INFO)
# try to get gpu device, if not just use cpu
self.device = \
torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# initialize critic (CNN) model
self.critic = CNN(in_chan=self.input_dims[-1], out_dim=1, out_act=None)
# initialize generator (TransposeCNN) model
self.generator = TransposeCNN(in_dim=self.z_dim,
out_chan=self.input_dims[-1],
out_act=torch.nn.Tanh())
# initialize zdim dimensional normal distribution to sample generator
# inputs
self.z_dist = torch.distributions.normal.Normal(
torch.zeros(self.batch_size, self.z_dim),
torch.ones(self.batch_size, self.z_dim))
# initialize bs dimensional uniform distribution to sample eps vals for
# creating interpolations
self.eps_dist = torch.distributions.uniform.Uniform(
torch.zeros(self.batch_size, 1, 1, 1),
torch.ones(self.batch_size, 1, 1, 1))
# sample a batch of z to have constant set of generator inputs
# as model trains
self.z_const = self.z_dist.sample()[:64].to(self.device)
# initialize critic and generator optimizers
self.crit_opt = torch.optim.Adam(self.critic.parameters(),
lr=self.learn_rate)
self.gen_opt = torch.optim.Adam(self.generator.parameters(),
lr=self.learn_rate)
# move tensors to device
self.critic.to(self.device)
self.generator.to(self.device)
def build_model(input_tensor, params, is_training, reuse):
'''Creates the corresponding TensorFlow graph that needs to be executed inside a tf.Session().
:param input_x
:param labels
:param dropout_rate
:param learning_rate
:return: Compute logits of the model (output distribution)
'''
# Feeding the 'inputs' data to the model
cnn = CNN()
cnn.build(input_tensor, params, is_training, reuse)
logits = cnn.logits
return logits
def main():
# Data loaders
tr_loader, va_loader, te_loader, _ = get_train_val_test_loaders(
num_classes=config('cnn.num_classes'))
# Model
model = CNN()
# TODO: define loss function, and optimizer
import torch.optim as op
import torch.nn as nn
criterion = nn.CrossEntropyLoss()
optimizer = op.Adam(model.parameters(), lr=config('cnn.learning_rate'))
#
print('Number of float-valued parameters:', count_parameters(model))
# Attempts to restore the latest checkpoint if exists
print('Loading cnn...')
model, start_epoch, stats = restore_checkpoint(model,
config('cnn.checkpoint'))
axes = utils.make_cnn_training_plot()
# Evaluate the randomly initialized model
_evaluate_epoch(axes, tr_loader, va_loader, model, criterion, start_epoch,
stats)
# Loop over the entire dataset multiple times
for epoch in range(start_epoch, config('cnn.num_epochs')):
# Train model
_train_epoch(tr_loader, model, criterion, optimizer)
# Evaluate model
_evaluate_epoch(axes, tr_loader, va_loader, model, criterion, epoch+1,
stats)
# Save model parameters
save_checkpoint(model, epoch+1, config('cnn.checkpoint'), stats)
print('Finished Training')
# Save figure and keep plot open
utils.save_cnn_training_plot()
utils.hold_training_plot()
def __init__(self, vocab_size, config):
super(OCR, self).__init__()
self.cnn = CNN()
self.config = config
self.transformer = Seq2Seq(
vocab_size,
encoder_hidden=config['seq_parameters']['encoder_hidden'],
decoder_hidden=config['seq_parameters']['decoder_hidden'],
img_channel=config['seq_parameters']['img_channel'],
decoder_embedded=config['seq_parameters']['decoder_embedded'],
dropout=config['seq_parameters']['dropout'])
def main():
# Data loaders
tr_loader, va_loader, te_loader, get_semantic_labels = get_train_val_test_loaders(
num_classes=config('cnn.num_classes'))
# Model
model = CNN()
# TODO: define loss function, and optimizer
params = list(model.conv1.parameters()) + list(
model.conv2.parameters()) + list(model.conv3.parameters())
params = params + list(model.fc1.parameters()) + list(
model.fc2.parameters()) + list(model.fc3.parameters())
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(params, lr=0.0001)
#
print('Number of float-valued parameters:', count_parameters(model))
# Attempts to restore the latest checkpoint if exists
print('Loading cnn...')
model, start_epoch, stats = restore_checkpoint(model,
config('cnn.checkpoint'))
fig, axes = utils.make_cnn_training_plot()
# Evaluate the randomly initialized model
_evaluate_epoch(axes, tr_loader, va_loader, model, criterion, start_epoch,
stats)
# Loop over the entire dataset multiple times
for epoch in range(start_epoch, config('cnn.num_epochs')):
# Train model
_train_epoch(tr_loader, model, criterion, optimizer)
# Evaluate model
_evaluate_epoch(axes, tr_loader, va_loader, model, criterion,
epoch + 1, stats)
# Save model parameters
save_checkpoint(model, epoch + 1, config('cnn.checkpoint'), stats)
print('Finished Training')
model, _, _ = restore_checkpoint(model, config('cnn.checkpoint'))
dataset = get_data_by_label(va_loader)
evaluate_cnn(dataset, model, criterion, get_semantic_labels)
# Save figure and keep plot open
utils.save_cnn_training_plot(fig)
utils.hold_training_plot()
def __init__(self, param, margin=0.5):
super(MemCnnSim, self).__init__()
self.param = param
self.max_grad_norm = param['max_grad_norm']
# self.memn2n = CNN(param)
self.memn2n = MemN2N(param)
self.cnn = CNN(param)
self.criterion = nn.CosineEmbeddingLoss(margin=margin)
self.optimizer = optim.Adam(self.parameters(), lr=0.00005)
def __init__(self, param):
super(MemN2N, self).__init__()
self.param = param
self.hops = self.param['hops']
self.vocab_size = param['vocab_size']
self.embedding_size = param['embedding_size']
self.embedding = nn.Embedding(self.vocab_size, self.embedding_size)
self.cnn = CNN(param, embedding=self.embedding)
self.linear = nn.Linear(self.embedding_size, self.embedding_size)
self.softmax = nn.Softmax()
self.weights_init()
def main():
# data loaders
_, va_loader, _, get_semantic_label = get_train_val_test_loaders(
num_classes=config('autoencoder.num_classes'))
dataset = get_data_by_label(va_loader)
model = Autoencoder(config('autoencoder.ae_repr_dim'))
criterion = torch.nn.MSELoss()
# Attempts to restore the latest checkpoint if exists
print('Loading autoencoder...')
#model, start_epoch, _ = restore_checkpoint(model,
#config('autoencoder.checkpoint'))
#evaluate_autoencoder(dataset, get_semantic_label, model, criterion)
# Evaluate model
model = CNN()
model, start_epoch, _ = restore_checkpoint(model, config('cnn.checkpoint'))
evaluate_autoencoder(dataset, get_semantic_label, model, criterion)
def __init__(self, args, train_dataset, device, input_channel, num_classes):
# Hyper Parameters
self.batch_size = 128
learning_rate = args.lr
if args.forget_rate is None:
if args.noise_type == "asymmetric":
forget_rate = args.noise_rate / 2
else:
forget_rate = args.noise_rate
else:
forget_rate = args.forget_rate
self.noise_or_not = train_dataset.noise_or_not
# Adjust learning rate and betas for Adam Optimizer
mom1 = 0.9
mom2 = 0.1
self.alpha_plan = [learning_rate] * args.n_epoch
self.beta1_plan = [mom1] * args.n_epoch
for i in range(args.epoch_decay_start, args.n_epoch):
self.alpha_plan[i] = float(args.n_epoch - i) / (args.n_epoch - args.epoch_decay_start) * learning_rate
self.beta1_plan[i] = mom2
# define drop rate schedule
self.rate_schedule = np.ones(args.n_epoch) * forget_rate
self.rate_schedule[:args.num_gradual] = np.linspace(0, forget_rate ** args.exponent, args.num_gradual)
self.device = device
self.num_iter_per_epoch = args.num_iter_per_epoch
self.print_freq = args.print_freq
self.co_lambda = args.co_lambda
self.n_epoch = args.n_epoch
self.train_dataset = train_dataset
if args.model_type == "cnn":
if args.dataset == 'mnist':
self.model1 = CNN(input_channel=input_channel, n_outputs=num_classes, linear_num=144)
self.model2 = CNN(input_channel=input_channel, n_outputs=num_classes, linear_num=144)
else:
self.model1 = CNN(input_channel=input_channel, n_outputs=num_classes)
self.model2 = CNN(input_channel=input_channel, n_outputs=num_classes)
elif args.model_type == "mlp":
self.model1 = MLPNet()
self.model2 = MLPNet()
elif args.model_type == 'resnet50':
self.model1 = resnet50(pretrained=True, num_classes=num_classes)
self.model2 = resnet50(pretrained=True, num_classes=num_classes)
self.model1.to(device)
print(self.model1.parameters)
self.model2.to(device)
print(self.model2.parameters)
if args.optimizer == 'Adam' or args.optimizer == 'adam':
self.optimizer = torch.optim.Adam(list(self.model1.parameters()) + list(self.model2.parameters()),
lr=learning_rate)
elif args.optimizer == "SGD" or args.optimizer == 'sgd':
self.optimizer = torch.optim.SGD(list(self.model1.parameters()) + list(self.model2.parameters()),
lr=learning_rate, momentum=0.9, weight_decay=1e-3)
else:
raise NotImplementedError("ERROR: Optimizer {} not been implemented!".format(args.optimizer))
self.loss_fn = loss_jocor
self.adjust_lr = args.adjust_lr
xi = xi.view((1,3,config('image_dim'),config('image_dim')))
zi = F.relu(model.conv1(xi))
zi = zi.detach().numpy()[0]
sort_mask = np.argsort(model.conv1.weight.detach().numpy().mean(axis=(1,2,3)))
zi = zi[sort_mask]
fig, axes = plt.subplots(4, 4, figsize=(10,10))
for i, ax in enumerate(axes.ravel()):
ax.axis('off')
im = ax.imshow(zi[i], cmap='gray')
fig.suptitle('Layer 1 activations, y={}'.format(yi))
fig.savefig('CNN_viz1_{}.png'.format(yi), dpi=200, bbox_inches='tight')
if __name__ == '__main__':
# Attempts to restore from checkpoint
print('Loading cnn...')
model = CNN()
model, start_epoch, _ = restore_checkpoint(model, config('cnn.checkpoint'),
force=True)
tr_loader, _, _, _ = get_train_val_test_loaders(
num_classes=config('cnn.num_classes'))
# Saving input images in original resolution
metadata = pd.read_csv(config('csv_file'))
for idx in [0, 4, 14, 15, 21]:
filename = os.path.join(
config('image_path'), metadata.loc[idx, 'filename'])
plt.imshow(imread(filename))
plt.axis('off')
plt.savefig('CNN_viz0_{}.png'.format(int(
metadata.loc[idx, 'numeric_label'])),
def __init__(self, args, train_dataset, device, input_channel,
num_classes):
# Hyper Parameters
self.batch_size = 128
learning_rate = args.lr
if args.forget_rate is None:
if args.noise_type == "asymmetric":
forget_rate = args.noise_rate / 2
else:
forget_rate = args.noise_rate
else:
forget_rate = args.forget_rate
self.noise_or_not = train_dataset.noise_or_not
# Adjust learning rate and betas for Adam Optimizer
mom1 = 0.9
mom2 = 0.1
self.alpha_plan = [learning_rate] * args.n_epoch
self.beta1_plan = [mom1] * args.n_epoch
for i in range(args.epoch_decay_start, args.n_epoch):
self.alpha_plan[i] = float(args.n_epoch - i) / (
args.n_epoch - args.epoch_decay_start) * learning_rate
self.beta1_plan[i] = mom2
# define drop rate schedule
self.rate_schedule = np.ones(args.n_epoch) * forget_rate
self.rate_schedule[:args.num_gradual] = np.linspace(
0, forget_rate**args.exponent, args.num_gradual)
self.device = device
self.num_iter_per_epoch = args.num_iter_per_epoch
self.print_freq = args.print_freq
self.co_lambda = args.co_lambda
self.n_epoch = args.n_epoch
self.train_dataset = train_dataset
if args.model_type == "cnn":
self.model1 = CNN(input_channel=input_channel,
n_outputs=num_classes)
self.model2 = CNN(input_channel=input_channel,
n_outputs=num_classes)
elif args.model_type == "mlp":
self.model1 = MLPNet()
self.model2 = MLPNet()
self.model1.to(device)
print(self.model1.parameters)
self.model2.to(device)
print(self.model2.parameters)
self.optimizer = torch.optim.Adam(list(self.model1.parameters()) +
list(self.model2.parameters()),
lr=learning_rate)
self.loss_fn = loss_jocor
self.adjust_lr = args.adjust_lr
def __init__(self):
img_width_range = cfg.img_width_range
word_len = cfg.word_len
self.batch_size = cfg.batch_size
self.visualize = cfg.visualize
gpu_device_id = '/gpu:' + str(cfg.gpu_id)
if cfg.gpu_id == -1:
gpu_device_id = '/cpu:0'
print("Using CPU model!")
with tf.device(gpu_device_id):
self.img_data = tf.placeholder(tf.float32,
shape=(None, 1, 32, None),
name='img_data')
self.zero_paddings = tf.placeholder(tf.float32,
shape=(None, None, 512),
name='zero_paddings')
self.bucket_specs = [(int(math.floor(64 / 4)), int(word_len + 2)),
(int(math.floor(108 / 4)), int(word_len + 2)),
(int(math.floor(140 / 4)), int(word_len + 2)),
(int(math.floor(256 / 4)), int(word_len + 2)),
(int(math.floor(img_width_range[1] / 4)),
int(word_len + 2))]
buckets = self.buckets = self.bucket_specs
self.decoder_inputs = []
self.encoder_masks = []
self.target_weights = []
with tf.device(gpu_device_id):
for i in xrange(int(buckets[-1][0] + 1)):
self.encoder_masks.append(
tf.placeholder(tf.float32,
shape=[None, 1],
name="encoder_mask{0}".format(i)))
for i in xrange(buckets[-1][1] + 1):
self.decoder_inputs.append(
tf.placeholder(tf.int32,
shape=[None],
name="decoder{0}".format(i)))
self.target_weights.append(
tf.placeholder(tf.float32,
shape=[None],
name="weight{0}".format(i)))
self.bucket_min_width, self.bucket_max_width = img_width_range
self.image_height = cfg.img_height
self.valid_target_len = cfg.valid_target_len
self.forward_only = True
self.bucket_data = {
i: BucketData()
for i in range(self.bucket_max_width + 1)
}
with tf.device(gpu_device_id):
cnn_model = CNN(self.img_data, True) #(not self.forward_only))
self.conv_output = cnn_model.tf_output()
self.concat_conv_output = tf.concat(
axis=1, values=[self.conv_output, self.zero_paddings])
self.perm_conv_output = tf.transpose(self.concat_conv_output,
perm=[1, 0, 2])
with tf.device(gpu_device_id):
self.attention_decoder_model = Seq2SeqModel(
encoder_masks=self.encoder_masks,
encoder_inputs_tensor=self.perm_conv_output,
decoder_inputs=self.decoder_inputs,
target_weights=self.target_weights,
target_vocab_size=cfg.target_vocab_size,
buckets=self.buckets,
target_embedding_size=cfg.target_embedding_size,
attn_num_layers=cfg.attn_num_layers,
attn_num_hidden=cfg.attn_num_hidden,
forward_only=self.forward_only,
use_gru=cfg.use_gru)
#gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=1.0)
self.sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True))
self.saver_all = tf.train.Saver(tf.global_variables())
self.saver_all.restore(self.sess, cfg.ocr_model_path)
n_data, n_input = X.shape
n_class = np.unique(t).size
T = create_label(t, n_data, n_class)
print 'make train/test data'
n_train, n_test = 1000, 50
i = np.random.permutation(n_data)[:n_train+n_test]
i_train, i_test = np.hsplit(i, [n_train])
X_train, X_test = X[i_train, :].reshape(n_train, 1, 28, 28), X[i_test, :].reshape(n_test, 1, 28, 28)
T_train, T_test = T[i_train, :], T[i_test, :]
print 'initialize...'
linear, sigmoid, softmax, relu = act.linear(), act.sigmoid(), act.softmax(), act.relu()
conv1, conv2 = conv(20, 1, 5, 5, relu), conv(50, 20, 5, 5, relu)
pool1, pool2 = pool(2, 2, 2), pool(2, 2, 2)
neural = NN(800, 500, 10, linear, sigmoid, softmax)
error = err.cross_entropy()
cnn = CNN(conv1, pool1, conv2, pool2, neural, error)
print 'train...'
cnn.train(X_train, T_train, epsilon = 0.005, lam = 0.0001, gamma = 0.9, s_batch = 5, epochs = 50)
print 'predict...'
Y_test = cnn.predict(X_test)
accuracy = cnn.accuracy(Y_test, T_test)
print 'accuracy: {0}'.format(accuracy)
print 'save figure of loss...'
cnn.save_lossfig()
def __init__(self,
phase,
visualize,
output_dir,
batch_size,
initial_learning_rate,
steps_per_checkpoint,
model_dir,
target_embedding_size,
attn_num_hidden,
attn_num_layers,
clip_gradients,
max_gradient_norm,
session,
load_model,
gpu_id,
use_gru,
use_distance=True,
max_image_width=160,
max_image_height=60,
max_prediction_length=8,
channels=1,
reg_val=0):
self.use_distance = use_distance
# We need resized width, not the actual width
max_resized_width = 1. * max_image_width / max_image_height * DataGen.IMAGE_HEIGHT
self.max_original_width = max_image_width
self.max_width = int(math.ceil(max_resized_width))
self.encoder_size = int(math.ceil(1. * self.max_width / 4))
self.decoder_size = max_prediction_length + 2
self.buckets = [(self.encoder_size, self.decoder_size)]
if gpu_id >= 0:
device_id = '/gpu:' + str(gpu_id)
else:
device_id = '/cpu:0'
self.device_id = device_id
if not os.path.exists(model_dir):
os.makedirs(model_dir)
if phase == 'test':
batch_size = 1
logging.info('phase: %s', phase)
logging.info('model_dir: %s', model_dir)
logging.info('load_model: %s', load_model)
logging.info('output_dir: %s', output_dir)
logging.info('steps_per_checkpoint: %d', steps_per_checkpoint)
logging.info('batch_size: %d', batch_size)
logging.info('learning_rate: %f', initial_learning_rate)
logging.info('reg_val: %d', reg_val)
logging.info('max_gradient_norm: %f', max_gradient_norm)
logging.info('clip_gradients: %s', clip_gradients)
logging.info('max_image_width %f', max_image_width)
logging.info('max_prediction_length %f', max_prediction_length)
logging.info('channels: %d', channels)
logging.info('target_embedding_size: %f', target_embedding_size)
logging.info('attn_num_hidden: %d', attn_num_hidden)
logging.info('attn_num_layers: %d', attn_num_layers)
logging.info('visualize: %s', visualize)
if use_gru:
logging.info('using GRU in the decoder.')
self.reg_val = reg_val
self.sess = session
self.steps_per_checkpoint = steps_per_checkpoint
self.model_dir = model_dir
self.output_dir = output_dir
self.batch_size = batch_size
self.global_step = tf.Variable(0, trainable=False)
self.phase = phase
self.visualize = visualize
self.learning_rate = initial_learning_rate
self.clip_gradients = clip_gradients
self.channels = channels
if phase == 'train':
self.forward_only = False
else:
self.forward_only = True
with tf.device(device_id):
self.height = tf.constant(DataGen.IMAGE_HEIGHT, dtype=tf.int32)
self.height_float = tf.constant(DataGen.IMAGE_HEIGHT,
dtype=tf.float64)
self.img_pl = tf.placeholder(tf.string,
name='input_image_as_bytes')
self.img_data = tf.cond(tf.less(tf.rank(self.img_pl), 1),
lambda: tf.expand_dims(self.img_pl, 0),
lambda: self.img_pl)
self.img_data = tf.map_fn(self._prepare_image,
self.img_data,
dtype=tf.float32)
num_images = tf.shape(self.img_data)[0]
# TODO: create a mask depending on the image/batch size
self.encoder_masks = []
for i in xrange(self.encoder_size + 1):
self.encoder_masks.append(tf.tile([[1.]], [num_images, 1]))
self.decoder_inputs = []
self.target_weights = []
for i in xrange(self.decoder_size + 1):
self.decoder_inputs.append(tf.tile([1], [num_images]))
if i < self.decoder_size:
self.target_weights.append(tf.tile([1.], [num_images]))
else:
self.target_weights.append(tf.tile([0.], [num_images]))
cnn_model = CNN(self.img_data, not self.forward_only)
self.conv_output = cnn_model.tf_output()
self.perm_conv_output = tf.transpose(self.conv_output,
perm=[1, 0, 2])
self.attention_decoder_model = Seq2SeqModel(
encoder_masks=self.encoder_masks,
encoder_inputs_tensor=self.perm_conv_output,
decoder_inputs=self.decoder_inputs,
target_weights=self.target_weights,
target_vocab_size=len(DataGen.CHARMAP),
buckets=self.buckets,
target_embedding_size=target_embedding_size,
attn_num_layers=attn_num_layers,
attn_num_hidden=attn_num_hidden,
forward_only=self.forward_only,
use_gru=use_gru)
table = tf.contrib.lookup.MutableHashTable(
key_dtype=tf.int64,
value_dtype=tf.string,
default_value="",
checkpoint=True,
)
insert = table.insert(
tf.constant(list(range(len(DataGen.CHARMAP))), dtype=tf.int64),
tf.constant(DataGen.CHARMAP),
)
with tf.control_dependencies([insert]):
num_feed = []
prb_feed = []
for line in xrange(len(self.attention_decoder_model.output)):
guess = tf.argmax(
self.attention_decoder_model.output[line], axis=1)
proba = tf.reduce_max(tf.nn.softmax(
self.attention_decoder_model.output[line]),
axis=1)
num_feed.append(guess)
prb_feed.append(proba)
# Join the predictions into a single output string.
trans_output = tf.transpose(num_feed)
trans_output = tf.map_fn(
lambda m: tf.foldr(
lambda a, x: tf.cond(
tf.equal(x, DataGen.EOS_ID),
lambda: '',
lambda: table.lookup(x) + a # pylint: disable=undefined-variable
),
m,
initializer=''),
trans_output,
dtype=tf.string)
# Calculate the total probability of the output string.
trans_outprb = tf.transpose(prb_feed)
trans_outprb = tf.gather(trans_outprb,
tf.range(tf.size(trans_output)))
trans_outprb = tf.map_fn(lambda m: tf.foldr(
lambda a, x: tf.multiply(tf.cast(x, tf.float64), a),
m,
initializer=tf.cast(1, tf.float64)),
trans_outprb,
dtype=tf.float64)
self.prediction = tf.cond(
tf.equal(tf.shape(trans_output)[0], 1),
lambda: trans_output[0],
lambda: trans_output,
)
self.probability = tf.cond(
tf.equal(tf.shape(trans_outprb)[0], 1),
lambda: trans_outprb[0],
lambda: trans_outprb,
)
self.prediction = tf.identity(self.prediction,
name='prediction')
self.probability = tf.identity(self.probability,
name='probability')
if not self.forward_only: # train
self.updates = []
self.summaries_by_bucket = []
params = tf.trainable_variables()
opt = tf.train.AdadeltaOptimizer(
learning_rate=initial_learning_rate)
loss_op = self.attention_decoder_model.loss
if self.reg_val > 0:
reg_losses = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
logging.info('Adding %s regularization losses',
len(reg_losses))
logging.debug('REGULARIZATION_LOSSES: %s', reg_losses)
loss_op = self.reg_val * tf.reduce_sum(
reg_losses) + loss_op
gradients, params = list(
zip(*opt.compute_gradients(loss_op, params)))
if self.clip_gradients:
gradients, _ = tf.clip_by_global_norm(
gradients, max_gradient_norm)
# Summaries for loss, variables, gradients, gradient norms and total gradient norm.
summaries = [
tf.summary.scalar("loss", loss_op),
tf.summary.scalar("total_gradient_norm",
tf.global_norm(gradients))
]
all_summaries = tf.summary.merge(summaries)
self.summaries_by_bucket.append(all_summaries)
# update op - apply gradients
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
self.updates.append(
opt.apply_gradients(list(zip(gradients, params)),
global_step=self.global_step))
self.saver_all = tf.train.Saver(tf.all_variables())
self.checkpoint_path = os.path.join(self.model_dir, "model.ckpt")
ckpt = tf.train.get_checkpoint_state(model_dir)
if ckpt and load_model:
# pylint: disable=no-member
logging.info("Reading model parameters from %s",
ckpt.model_checkpoint_path)
self.saver_all.restore(self.sess, ckpt.model_checkpoint_path)
else:
logging.info("Created model with fresh parameters.")
self.sess.run(tf.initialize_all_variables())
from flask import Flask, request, jsonify, make_response
from flask_restful import Resource, Api
import pickle
import numpy as np
from model.cnn import CNN
import cv2
cnn = CNN("RGB")
app = Flask(__name__)
api = Api(app)
class Predict(Resource):
def post(self):
data = request.data
img = cv2.imdecode(np.fromstring(data, dtype=np.uint8),
cv2.IMREAD_COLOR)
results = cnn.raw_predict(img)
return make_response(jsonify(results), 201)
api.add_resource(Predict, '/predict')
if __name__ == '__main__':
app.run(host='0.0.0.0', debug=True, port=5000)
path = images.paths[20000]
image = images._getitem(path)
print ('read image {} of shape {}'.format(path, image.shape))
my_split=poses[0]
my_split=[path[:-4] for path in my_split]
"""Use SRGAN"""
srgan = generator()
srgan.load_weights('weights/srgan/gan_generator.h5')
"""Upload customed cnn model"""
cnn = CNN(256, 256, 3, 101)
cnn.load_weights('weights/custom/cnn_plus.h5')
plot_model(cnn, to_file='./model.png', show_shapes=True, show_layer_names=True)
train_model(2, 'cnn_plus', cnn, srgan)
#filepath="./cnn_weights.h5"
#checkpoint = ModelCheckpoint(filepath, monitor='accuracy', verbose=1, save_best_only=True, mode='max')
#callbacks_list = [checkpoint]
"""Prepare and train on a batch of data and labels, 10 iterations"""
for i in range(2):
train_set = devide(24, 2, 2)
X = tensor2numpy('./data/', train_set, srgan)
x = [X[i] for i in X.keys()]
def make_model(train_set_x, n_class):
input_dim = train_set_x.shape[1:]
layers = make_cnn(input_dim, n_class)
cnn = CNN(layers)
return cnn
import cv2
from model.cnn import CNN
import imutils
import math
if __name__ == "__main__":
cap = cv2.VideoCapture(0)
cap.set(cv2.CAP_PROP_FRAME_WIDTH, 720)
cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 720)
font = cv2.FONT_HERSHEY_SIMPLEX
model = CNN("gray")
while True:
ret, img = cap.read() # full image
gray, results = model.raw_predict(img)
print(results)
cv2.imshow('gray', gray)
k = cv2.waitKey(30) & 0xff
if k == 27:
break
cap.release()
cv2.destroyAllWindows()
class WassersteinGAN():
def __init__(self, config):
# get config args
self.out_dir = config['output_directory']
self.batch_size = config['batch_size']
self.input_dims = config['input_dimensions']
self.learn_rate = config['learning_rate']
self.z_dim = config['z_dimension']
self.name = config['model_name']
self.verbosity = config['verbosity']
self.num_train = config['number_train']
self.num_test = config['number_test']
# initialize logging to create new log file and log any level event
logging.basicConfig(format='%(asctime)s %(message)s',
datefmt='%m/%d/%Y %I:%M:%S %p',
filename='{}{}.log'.format(self.out_dir,
self.name),
filemode='w',
level=logging.INFO)
# try to get gpu device, if not just use cpu
self.device = \
torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# initialize critic (CNN) model
self.critic = CNN(in_chan=self.input_dims[-1], out_dim=1, out_act=None)
# initialize generator (TransposeCNN) model
self.generator = TransposeCNN(in_dim=self.z_dim,
out_chan=self.input_dims[-1],
out_act=torch.nn.Tanh())
# initialize zdim dimensional normal distribution to sample generator
# inputs
self.z_dist = torch.distributions.normal.Normal(
torch.zeros(self.batch_size, self.z_dim),
torch.ones(self.batch_size, self.z_dim))
# initialize bs dimensional uniform distribution to sample eps vals for
# creating interpolations
self.eps_dist = torch.distributions.uniform.Uniform(
torch.zeros(self.batch_size, 1, 1, 1),
torch.ones(self.batch_size, 1, 1, 1))
# sample a batch of z to have constant set of generator inputs
# as model trains
self.z_const = self.z_dist.sample()[:64].to(self.device)
# initialize critic and generator optimizers
self.crit_opt = torch.optim.Adam(self.critic.parameters(),
lr=self.learn_rate)
self.gen_opt = torch.optim.Adam(self.generator.parameters(),
lr=self.learn_rate)
# move tensors to device
self.critic.to(self.device)
self.generator.to(self.device)
def print_structure(self):
print('[INFO] \'{}\' critic structure \n{}'.format(
self.name, self.critic))
print('[INFO] \'{}\' generator structure \n{}'.format(
self.name, self.generator))
@staticmethod
def compute_losses(real_crit_out, fake_crit_out, crit_grad):
# compute wasserstein distance estimate
wass_dist = torch.mean(real_crit_out - fake_crit_out)
# compute mean of normed critic gradients
crit_grad_mean_norm = torch.mean(torch.norm(crit_grad, p=2,
dim=(2, 3)))
# lagrangian multiplier for critic gradient penalty
# (push crit_grad_mean_norm -> 1)
crit_grad_penalty = (crit_grad_mean_norm - 1.)**2
# compute generator loss
gen_loss = wass_dist
# compute critic loss with lambda=10 weighted critic gradient penalty
crit_loss = (10.0 * crit_grad_penalty) - wass_dist
return gen_loss, crit_loss, wass_dist, crit_grad_penalty
def generate_samples_and_tile(self, z):
# generator a batch of fak images from z input
fake_img_batch = self.generator(z)
# detach, move to cpu, and covert images to numpy
fake_img_batch = fake_img_batch.detach().cpu().numpy()
# move channel dim to last dim of tensor
fake_img_batch = np.transpose(fake_img_batch, [0, 2, 3, 1])
# construct tiled image (squeeze to remove channel dim for grayscale)
fake_img_tiled = np.squeeze(tile_images(fake_img_batch))
return fake_img_tiled
def train(self, dataloader, num_epochs):
# iterate through epochs
for e in range(num_epochs):
# accumulator for wasserstein distance over an epoch
running_w_dist = 0.0
# iterate through batches
for i, batch in enumerate(dataloader):
# get images from batch
real_img_batch = batch['image'].to(self.device)
# get number of samples in batch
bs = real_img_batch.shape[0]
# sample from z and eps distribution and clip based on number
# of samples in batch
z_sample = self.z_dist.sample()[:bs].to(self.device)
eps_sample = self.eps_dist.sample()[:bs].to(self.device)
# generate batch of fake images by feeding sampled z through
# generator
fake_img_batch = self.generator(z_sample)
# compute batch of images by interpolating eps_sample amount
# between real and fake
# (generated) images
int_img_batch = (eps_sample * real_img_batch) + \
((1. - eps_sample) * fake_img_batch)
# compute critic outputs from real, fake, and interpolated
# image batches
real_crit_out = self.critic(real_img_batch)
fake_crit_out = self.critic(fake_img_batch)
int_crit_out = self.critic(int_img_batch)
# compute gradient of critic output w.r.t interpolated image
# inputs
crit_grad = torch.autograd.grad(
outputs=int_crit_out,
inputs=int_img_batch,
grad_outputs=torch.ones_like(int_crit_out),
create_graph=True,
retain_graph=True,
only_inputs=True)[0]
# compute losses
gen_loss, crit_loss, w_dist, grad_pen = self.compute_losses(
real_crit_out, fake_crit_out, crit_grad)
# NOTE: Currently must update critic and generator separately.
# If both are updated within the same loop, either updating
# doesn't happen, or an inplace operator error occurs which
# prevents gradient computation, depending on the ordering of
# the zero_grad(), backward(), step() calls. ???
if i % 10 == 9:
# update just the generator (every 10th step)
self.gen_opt.zero_grad()
gen_loss.backward()
self.gen_opt.step()
if self.verbosity:
# generate const batch of fake samples and tile
fake_img_tiled = self.generate_samples_and_tile(
self.z_const)
# save tiled image
plt.imsave(
'{}{}_gen_step_{}.png'.format(
self.out_dir, self.name,
(e *
(int(self.num_train / self.batch_size) + 1)) +
i), fake_img_tiled)
else:
# update just the critic
self.crit_opt.zero_grad()
crit_loss.backward()
self.crit_opt.step()
# accumulate running wasserstein distance
running_w_dist += w_dist.item()
# done with current epoch
# compute average wasserstein distance over epoch
epoch_avg_w_dist = running_w_dist / i
# log epoch stats info
logging.info(
'| epoch: {:3} | wasserstein distance: {:6.2f} | gradient '
'penalty: {:6.2f} |'.format(e + 1, epoch_avg_w_dist, grad_pen))
# new sample from z dist
z_sample = self.z_dist.sample()[:64].to(self.device)
# generate const batch of fake samples and tile
fake_img_tiled = self.generate_samples_and_tile(z_sample)
if self.verbosity:
# save tiled image
plt.imsave(
'{}{}_gen_epoch_{}.png'.format(self.out_dir, self.name,
e + 1), fake_img_tiled)
# save current state of generator and critic
torch.save(self.generator.state_dict(),
'{}{}_generator.pt'.format(self.out_dir, self.name))
torch.save(self.critic.state_dict(),
'{}{}_critic.pt'.format(self.out_dir, self.name))
def main(args):
# Enter all arguments that you want to be in the filename of the saved output
ordered_args = [
'dataset',
'data_augmentation',
'seed',
'remove_percent',
'burn_in_epochs',
'remove_strategy',
'noise_percent',
'noise_labels',
'noise_pixels_percent',
'noise_pixels_std',
'optimizer',
'learning_rate',
]
save_fname = '__'.join('{}_{}'.format(arg, args_dict[arg])
for arg in ordered_args)
fname = os.path.join(args.output_dir, save_fname)
if os.path.exists(fname + '__stats_dict.pkl'):
redo = input(
"There exists experiment result already, continue? [yes/no] ")
if redo == 'no':
exit()
elif redo == 'yes':
pass
else:
raise ValueError('wrong answer')
os.makedirs(args.output_dir, exist_ok=True)
# Set appropriate devices
device = torch.device(args.device)
print('run on device: {0}'.format(device))
cudnn.benchmark = True # Should make training go faster for large models
# Set random seed for initialization
torch.manual_seed(args.seed)
if 'cuda' in args.device:
torch.cuda.manual_seed(args.seed)
npr.seed(args.seed)
train_ds, test_ds, num_classes = get_data(args.dataset)
if args.noise_percent > 0:
assert not (args.noise_labels and (args.noise_pixels_percent > 0))
if args.noise_labels:
train_ds, noise_indexes = noise_labels(train_ds,
args.noise_percent, fname)
if args.noise_pixels_percent:
train_ds, noise_indexes = noise_pixels(train_ds,
args.noise_percent,
args.noise_pixels_percent,
args.noise_pixels_std,
fname)
print('Training on ' + str(len(train_ds)) + ' examples')
# Setup model
if args.model == 'resnet18':
model = ResNet18(num_classes=num_classes)
elif args.model == 'wideresnet':
if args.dataset == 'svhn':
model = WideResNet(depth=16,
num_classes=num_classes,
widen_factor=8,
dropRate=0.4)
else:
model = WideResNet(depth=28,
num_classes=num_classes,
widen_factor=10,
dropRate=0.3)
elif args.model == 'cnn':
model = CNN(num_classes=num_classes)
else:
print(
'Specified model not recognized. Options are: resnet18 and wideresnet'
)
# Setup loss
model = model.to(args.device)
criterion = torch.nn.CrossEntropyLoss().cuda()
criterion.__init__(reduce=False)
# Setup optimizer
if args.optimizer == 'adam':
model_optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
elif args.optimizer == 'sgd':
model_optimizer = torch.optim.SGD(model.parameters(),
lr=args.learning_rate,
momentum=0.9,
nesterov=True,
weight_decay=5e-4)
scheduler = MultiStepLR(model_optimizer,
milestones=[60, 120, 160],
gamma=0.2)
elif args.optimizer == 'sgd-const-lr':
model_optimizer = torch.optim.SGD(model.parameters(),
lr=args.learning_rate,
momentum=0.9,
nesterov=True,
weight_decay=5e-4)
else:
print('Specified optimizer not recognized. Options are: adam and sgd')
save_point = os.path.join(args.output_dir, 'checkpoint', args.dataset)
os.makedirs(save_point, exist_ok=True)
checkpoint_fname = os.path.join(save_point, save_fname + '.t7')
# Initialize dictionary to save statistics for every example presentation
example_stats = {}
num_examples = len(train_ds)
example_weights = np.ones(num_examples)
elapsed_time = 0
# train_idx = np.array(range(0, len(train_ds)))
train_loader = DataLoader(train_ds,
batch_size=args.batch_size,
shuffle=True)
for epoch in range(args.epochs):
if args.remove_strategy != 'normal' and epoch >= args.burn_in_epochs:
if 'sampling' in args.remove_strategy:
# sampling by weight
normalized_weights = example_weights / example_weights.sum()
index_stats = example_stats.get('example_weights', [[], []])
index_stats[1].append(normalized_weights)
example_stats['example_weights'] = index_stats
choice_num = int(num_examples *
(1 - args.remove_percent / 100))
train_idx = np.random.choice(range(num_examples),
size=choice_num,
replace=False,
p=normalized_weights)
elif args.remove_strategy == 'low-acc':
remove_n = int(args.remove_percent * num_examples / 100)
losses = []
for idx in range(num_examples):
losses.append(example_stats[idx][0][epoch - 1])
losses = np.array(losses)
sorted_indexes = np.argsort(losses)
train_idx = sorted_indexes[:num_examples - remove_n]
elif args.remove_strategy == 'all-noise':
remove_n = int(args.remove_percent * num_examples / 100)
if args.remove_percent <= args.noise_percent_labels:
remove_indexes = npr.choice(noise_indexes,
remove_n,
replace=False)
train_idx = np.setdiff1d(range(num_examples),
remove_indexes)
else:
train_idx = np.setdiff1d(range(num_examples),
noise_indexes)
train_idx = npr.choice(train_idx,
num_examples - remove_n,
replace=False)
else:
# event method
_, unlearned_per_presentation, _, first_learned = compute_forgetting_statistics(
example_stats, epoch)
ordered_examples, ordered_values = sort_examples_by_forgetting(
[unlearned_per_presentation], [first_learned], epoch)
train_idx = sample_dataset_by_forgetting(
train_ds, ordered_examples, ordered_values,
args.remove_percent, args.remove_strategy)
sampler = torch.utils.data.SubsetRandomSampler(train_idx)
train_loader = DataLoader(train_ds,
batch_size=args.batch_size,
sampler=sampler)
start_time = time.time()
train(args, model, criterion, device, train_loader, model_optimizer,
epoch, example_stats)
test_loader = DataLoader(test_ds, batch_size=32, shuffle=True)
test(epoch, model, criterion, device, test_loader, example_stats,
checkpoint_fname)
if args.remove_strategy != 'normal' and epoch >= args.burn_in_epochs:
# evaluate on removed data
removed_idx = np.setdiff1d(range(num_examples), train_idx)
sampler = torch.utils.data.SubsetRandomSampler(removed_idx)
removed_loader = DataLoader(train_ds,
batch_size=args.batch_size,
sampler=sampler)
evaluate_on_removed(model, criterion, device, removed_loader,
epoch, example_stats)
if 'sampling' in args.remove_strategy:
example_weights = update_example_weights(example_weights,
example_stats, epoch,
args.remove_strategy)
epoch_time = time.time() - start_time
elapsed_time += epoch_time
print('| Elapsed time : %d:%02d:%02d' % (get_hms(elapsed_time)))
# Update optimizer step
if args.optimizer == 'sgd':
scheduler.step(epoch)
# Save the stats dictionary
fname = os.path.join(args.output_dir, save_fname)
with open(fname + "__stats_dict.pkl", "wb") as f:
pickle.dump(example_stats, f)
# Log the best train and test accuracy so far
with open(fname + "__best_acc.txt", "w") as f:
f.write('train test \n')
f.write(str(max(example_stats['train'][1])))
f.write(' ')
f.write(str(max(example_stats['test'][1])))
layers = [conv1, relu1, maxpool1, flatten, dense1]
return layers
if __name__ == '__main__':
if len(sys.argv) == 2:
action = sys.argv[1]
if action == '--debug':
# run gradient check on two random data points.
grad_check()
else:
train_set_x, train_set_y, test_set_x, test_set_y, classes = load_dataset(
)
num_of_classes = len(classes)
train_set_x = train_set_x / 255
test_set_x = test_set_x / 255
input_dim = train_set_x.shape[1:]
layers = make_cnn(input_dim, num_of_classes)
cnn = CNN(layers)
cnn, costs = Adam(model=cnn,
X_train=train_set_x,
y_train=train_set_y,
epoch=100,
learning_rate=0.001,
X_test=test_set_x,
y_test=test_set_y,
minibatch_size=64).minimize()
torch.IntTensor([l]),
raw=raw))
index += l
return texts
def label_dict(self):
return self.dict
def label_constant(self):
return self.alphabet
if __name__ == '__main__':
sys.path.append('../')
from model.cnn import CNN
img = Image.open('../test/ocr/test.jpg')
alphabet = process_alphabet('../weight/ocr/ocr.json')
model = CNN(1).cuda()
checkpoint = torch.load('../weight/ocr/ocr.pth.tar')
model.load_state_dict(checkpoint)
res = predict(model, img, alphabet)
print(res)
from main import text_ocr
from helper.image import read_url_img, base64_to_PIL, get_now
from config import ocrPath, textPath, GPU
# load model
import torch
from model.vgg import VGG
from model.cnn import CNN
textModel = VGG(3).cuda() if GPU else VGG(3)
checkpoint = torch.load(textPath)
textModel.load_state_dict(checkpoint)
print(textModel)
ocrModel = CNN(1).cuda() if GPU else CNN(1)
checkpoint = torch.load(ocrPath)
ocrModel.load_state_dict(checkpoint)
print(ocrModel)
billList = []
root = './test/'
timeOutTime = 5
def job(uid, url, imgString, iscut, isclass, billModel, ip):
now = get_now()
if url is not None:
img = read_url_img(url)
elif imgString is not None:
img = base64_to_PIL(imgString)