def setup(self, bottom, top):
param = eval(self.param_str)
self.batch = int(param['batch'])
self.batch_loader = BatchLoader(param, "train")
self.ohem_batch_loader = BatchLoader(param, 'ohem')
self.train_ratio = 1.
self.ohem_ratio = 0
if self.ohem_batch_loader.is_loaded():
self.train_ratio = 7 / 8.
self.ohem_ratio = 1. - self.train_ratio
top[0].reshape(self.batch, 3, self.img_size, self.img_size) # data
top[1].reshape(self.batch, 72) # landmark
def inference(self, input_):
self.loader = BatchLoader(
self.config.data_dir, self.config.dataset_name,
self.config.batch_size, self.config.num_classes,
self.config.preprocessor, self.config.epoch,
self.config.specialist, self.config.forward_only)
content, filename = self.loader.prepare_inference()
#with tf.control_dependencies([self.loader.enqueue]):
logits = self.build_model(content)
softmax = tf.nn.softmax(logits)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
self.sess.run(init_op)
self.saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(self.config.checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
self.saver.restore(self.sess, ckpt.model_checkpoint_path)
else:
print('no checkpoint found...')
self.sess.run(self.loader.enqueue,
feed_dict={self.loader.filenames: input_})
m_logits, m_softmax, m_filename = self.sess.run(
[logits, softmax, filename])
print(m_softmax, m_filename)
def setup(self, bottom, top):
param = eval(self.param_str)
self.batch = int(param['batch'])
self.net = param['net']
self.img_size = config.NET_IMG_SIZES[self.net]
self.batch_loader = BatchLoader(param, "train")
self.ohem_batch_loader = BatchLoader(param, 'ohem')
self.train_ratio = 1.
self.ohem_ratio = 0
if self.ohem_batch_loader.is_loaded():
self.train_ratio = 7 / 8.
self.ohem_ratio = 1 / 8.
top[0].reshape(self.batch, 3, self.img_size, self.img_size) # data
top[1].reshape(self.batch, 1) # label
top[2].reshape(self.batch, 4) # bbox
if self.net != 'pnet':
top[3].reshape(self.batch, config.LANDMARK_SIZE * 2)
def train(args):
yaml_file = yaml.safe_load(args.cfg)
print(yaml.dump(yaml_file, default_flow_style=False))
cfg = TrainingConfig(yaml_file)
output_dir = cfg.output
if not os.path.exists(output_dir):
os.makedirs(output_dir)
chunks = get_all_chunks(cfg.input)
print("Found {} files".format(len(chunks)))
num_train_chunks = int(len(chunks) * cfg.train_ratio)
training_chunks = chunks[:num_train_chunks]
test_chunks = chunks[num_train_chunks:]
print("Chunks Training({}) Testing({})".format(len(training_chunks),
len(test_chunks)))
train_loader = BatchLoader(training_chunks, cfg)
test_loader = BatchLoader(test_chunks, cfg)
worker = TensorWorker(cfg, train_loader, test_loader)
print()
def __init__(self,
batch_size=params.BATCH_SIZE,
nb_epochs=params.NB_EPOCHS,
mask=None,
experiment_path=params.EXPERIMENT_PATH,
use_adversarial_loss=params.USE_ADVERSARIAL_LOSS,
lambda_decay=params.LAMBDA_DECAY,
lambda_adversarial=params.LAMBDA_ADVERSARIAL,
patience=params.PATIENCE,
discr_whole_image=params.DISCR_WHOLE_IMAGE,
discr_loss_limit=params.DISCR_LOSS_LIMIT,
use_dropout=params.USE_DROPOUT):
self.batch_size = batch_size
self.nb_epochs = nb_epochs
self.experiment_path = experiment_path
self.save_path = os.path.join(self.experiment_path, "model/")
self.save_best_path = os.path.join(self.experiment_path, "best_model/")
self.logs_path = os.path.join(self.experiment_path, "logs")
create_dir(self.save_path)
create_dir(self.logs_path)
self.global_step = tf.Variable(0, trainable=False, name='global_step')
self.phase = tf.placeholder(tf.bool, name='phase')
self.patience = patience
# parameters for adversarial loss
self.use_adversarial_loss = use_adversarial_loss
self.lambda_adversarial = lambda_adversarial
if lambda_decay:
self.lambda_adversarial = 1 - tf.train.exponential_decay(
.1, self.global_step, 10000, .5, staircase=True)
self.discr_whole_image = discr_whole_image
self.discr_loss_limit = discr_loss_limit
self.num_discr_trained = tf.Variable(tf.constant(0, dtype=tf.int32),
trainable=False)
self.use_dropout = use_dropout
self.batch_loader = BatchLoader(self.batch_size)
if mask is not None:
self.np_mask = mask
else:
self.np_mask = np.zeros((1, 64, 64, 1))
self.np_mask[:, 16:48, 16:48, :] = 1
self._sess = tf.Session()
tf.summary.scalar("lambda_adversarial", self.lambda_adversarial)
tf.summary.scalar("num discr trained", self.num_discr_trained)
def __init__(self, vocab_files, \
generator_file, \
batch_size=32, load_generator=True):
self.batch_size = batch_size
self.generator_data = None
if generator_file:
with open(generator_file, 'rb') as f:
self.generator_data = [pkl.load(f)]
self.gen_batch_loader = BatchLoader(self.generator_data,
vocab_files,
sentence_array=True)
self.vocab_size = self.gen_batch_loader.words_vocab_size
def pretrain_generator(gen, gen_opt, epochs):
global data_df
n_iter = 0
loader = BatchLoader(data_df)
for epoch in range(epochs):
# print(f'epoch = {epoch}\n --------------------------------')
total_loss = 0
n_iter += 1
for batch in tqdm(loader.load_action_batch(MAX_SEQ_LEN, BATCH_SIZE,
CUDA),
total=int(NUM_SAMPLES / BATCH_SIZE / MAX_SEQ_LEN)):
gen_opt.zero_grad()
loss = gen.batchNLLLoss(batch)
loss.backward()
gen_opt.step()
total_loss += loss.data.item()
total_loss /= NUM_SAMPLES / BATCH_SIZE / MAX_SEQ_LEN
print('iteration = %d, NLL loss = %.4f' % (n_iter, total_loss))
def main():
np.random.seed(233)
torch.manual_seed(233)
data_loader = BatchLoader(10000)
model = Seq2SeqModel(10, NUM_CLASSES, WINDOW_SIZE)
trainval(model, data_loader)
# Should print 'polo '
inference(model, "nor marco I", "polo ")
inference(model, "marco nor I", "polo ")
inference(model, "nor I marco", "polo ")
# Should print ' '
inference(model, "nor I neither", " ")
# More difficult task
inference(model, "nor ma rco I", " ")
inference(model, "ma rco nor I", " ")
pass
def main():
LAMBDA = 0.0
num_class = 526
checkpoint_dir = "../model/"
with tf.name_scope('input'):
input_images = tf.placeholder(tf.float32,
shape=(None, 100, 100, 3),
name='input_images')
labels = tf.placeholder(tf.int64, shape=(None), name='labels')
logits, features, total_loss, accuracy, centers_update_op, center_loss, softmax_loss = build_network(
input_images, labels, num_class, ratio=LAMBDA)
global_step = tf.Variable(0, trainable=False, name='global_step')
train_batch_loader = BatchLoader("../data/facescrub_train.list", 16)
test_batch_loader = BatchLoader("../data/facescrub_val.list", 16)
optimizer = tf.train.AdamOptimizer(0.001)
with tf.control_dependencies([centers_update_op]):
train_op = optimizer.minimize(total_loss, global_step=global_step)
summary_op = tf.summary.merge_all()
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter('../tmp/face_log', sess.graph)
saver = tf.train.Saver()
step = sess.run(global_step)
while step <= 80000:
batch_images, batch_labels = train_batch_loader.next_batch()
# print batch_images.shape
# print batch_labels.shape
_, summary_str, train_acc, Center_loss, Softmax_loss = sess.run(
[train_op, summary_op, accuracy, center_loss, softmax_loss],
feed_dict={
input_images:
(batch_images - 127.5) * 0.0078125, # - mean_data,
labels: batch_labels,
})
step += 1
print("step", step)
if step % 100 == 0:
print("********* Step %s: ***********" % str(step))
print("center loss: %s" % str(Center_loss))
print("softmax_loss: %s" % str(Softmax_loss))
print("train_acc: %s" % str(train_acc))
print("*******************************")
if step % 10000 == 0:
saver.save(sess,
checkpoint_dir + 'model.ckpt',
global_step=step)
writer.add_summary(summary_str, global_step=step)
if step % 2000 == 0:
batch_images, batch_labels = test_batch_loader.next_batch()
vali_image = (batch_images - 127.5) * 0.0078125
vali_acc = sess.run(accuracy,
feed_dict={
input_images: vali_image,
labels: batch_labels
})
print(("step: {}, train_acc:{:.4f}, vali_acc:{:.4f}".format(
step, train_acc, vali_acc)))
sess.close()
def main():
params = args()
model_prefix = params.model_prefix
load_epoch = params.load_epoch
batch_size = params.batch_size
img1_path = params.image1
img2_path = params.image2
img_shape = params.image_shape
img_dir = params.img_dir
txt_path = params.file_txt
# in out files
file_ = open(txt_path, 'r')
lines_ = file_.readlines()
result_ = open("result.txt", 'w')
#
if Test_Img == 'True':
img1 = cv2.imread(img1_path)
img2 = cv2.imread(img2_path)
img1 = cv2.resize(img1, (img_shape, img_shape))
img2 = cv2.resize(img2, (img_shape, img_shape))
img1 = np.expand_dims(img1, 0)
img2 = np.expand_dims(img2, 0)
test_batch_loader = BatchLoader("../data/facescrub_val.list", batch_size,
img_shape)
tf.reset_default_graph()
with tf.name_scope('input'):
input_images = tf.placeholder(tf.float32,
shape=(batch_size, img_shape, img_shape,
3),
name='input_images')
labels = tf.placeholder(tf.int32, shape=(batch_size), name='labels')
features, accuracy, pred_class, res1 = build_network(
input_images, labels, 526, 'test')
check_ckpt(model_prefix + '-' + str(load_epoch))
#detect
detector = face_dect()
#
with tf.Session() as sess:
restore_model = tf.train.Saver()
#restore_model = tf.train.import_meta_graph(model_prefix+'-'+str(load_epoch) +'.meta')
restore_model.restore(sess, model_prefix + '-' + str(load_epoch))
print("face model restore over")
'''
all_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
key_list = []
var_dic = dict()
for v_name in tf.global_variables():
i=0
print("name : ",v_name.name[:-2],v_name.shape)
print("shape",all_vars[i])
key_list.append(v_name.name[:-2])
i+=1
#print(tf.get_variable_scope())
#all_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
vas= sess.run([all_vars])
print(len(vas))
for i in range(len(vas)):
cur_name = key_list[i]
cur_var = vas[i]
print("name ,shape : ",cur_name,np.shape(cur_var))
var_dic[cur_name] = cur_var
'''
#restore_model = tf.train.import_meta_graph(model_prefix+'-'+str(load_epoch)+'.meta')
#restore_model.restore(sess,model_prefix+'-'+str(load_epoch))
if Test_Img == 'False':
iter_num = 0
accuracy_sum = 0
while iter_num < test_batch_loader.batch_num:
batch_images, batch_labels = test_batch_loader.next_batch()
images_in = (batch_images - 127.5) * 0.0078125
feat, batch_accuracy = sess.run([features, accuracy],
feed_dict={
input_images: images_in,
labels: batch_labels
})
accuracy_sum += batch_accuracy
iter_num += 1
if iter_num % 10 == 0:
print("step ", iter_num, batch_accuracy)
print("image num: ", test_batch_loader.data_num,
"the test accuracy: ",
accuracy_sum / (test_batch_loader.batch_num))
elif Test_Img == 'True':
with tf.name_scope('valdata'):
label_t = np.zeros([1])
feat1 = sess.run([features],
feed_dict={
input_images: img1,
labels: label_t
})
feat2 = sess.run([features],
feed_dict={
input_images: img2,
labels: label_t
})
distance = L2_distance(feat1, feat2, 512)
print("the 2 image dis: ", distance)
for rot, fdir, fname in os.walk(img_dir):
if len(fname) != 0:
break
img_list = []
print(fname)
for i in range(len(fname)):
org_img = cv2.imread(os.path.join(rot, fname[i]))
img_org = cv2.resize(org_img, (img_shape, img_shape))
img_org = np.expand_dims(img_org, 0)
img_list.append(img_org)
for i in range(len(fname)):
img1 = img_list[i]
feat1 = sess.run([features],
feed_dict={
input_images: img1,
labels: label_t
})
j = i + 1
while j < len(fname):
img2 = img_list[j]
t1 = time.time()
feat2 = sess.run([features],
feed_dict={
input_images: img2,
labels: label_t
})
t2 = time.time()
print("one image time ", t2 - t1)
distance = L2_distance(feat1, feat2, 512)
t3 = time.time()
print("compere time ", t3 - t2)
print(i, j, distance)
j += 1
elif Test_Img == 'File':
label_t = np.ones([1])
for i in range(len(lines_)):
feat_vec = []
feat1_fg = 0
feat2_fg = 0
line_1 = lines_[i]
line_1 = string.strip(line_1)
line_s = line_1.split(',')
dir_path_save = line_s[0][:-4]
dir_path_save = "../cropface/" + dir_path_save
mkdir_(dir_path_save)
for j in range(len(line_s)):
feat_vec2 = []
if j == 0:
#print("line ",line_s)
img1_pic = line_s[0]
img1_path = os.path.join(img_dir, img1_pic)
img1 = cv2.imread(img1_path)
bboxes_1 = detector.get_face(img1)
if bboxes_1 is not None:
for k in range(bboxes_1.shape[0]):
crop_img1 = get_img(img1, bboxes_1[k],
img_shape)
if k == 0 and SV_IMG:
img_save_path = dir_path_save + '/' + line_s[
0][:-4] + ".jpg"
save_image(img_save_path, crop_img1)
crop_img1 = (crop_img1 - 127.5) * 0.0078125
crop_img1 = np.expand_dims(crop_img1, 0)
feat1 = sess.run([features],
feed_dict={
input_images: crop_img1,
labels: label_t
})
print("a feature shape ", np.shape(feat1))
feat_vec.append(feat1)
feat_fg = 1
else:
print("a no face detect ")
break
else:
img2_pic = line_s[j]
img2_path = os.path.join(img_dir, img2_pic)
img2 = cv2.imread(img2_path)
bboxes_2 = detector.get_face(img2)
if bboxes_2 is not None:
for k in range(bboxes_2.shape[0]):
crop_img2 = get_img(img2, bboxes_2[k],
img_shape)
if SV_IMG:
img_save_path = dir_path_save + '/' + line_s[
j][:-4] + "-" + str(k) + ".jpg"
save_image(img_save_path, crop_img2)
crop_img2 = (crop_img2 - 127.5) * 0.0078125
crop_img2 = np.expand_dims(crop_img2, 0)
feat2 = sess.run([features],
feed_dict={
input_images: crop_img2,
labels: label_t
})
print("b feature shape ", np.shape(feat2))
feat_vec2.append(feat2)
feat2_fg = 1
else:
print("b no face detect ")
continue
if j > 0:
t2 = time.time()
distance = L2_distance(feat_vec[0], feat_vec2[0], 512)
print("distance is ", distance)
t3 = time.time()
print("compere time ", t3 - t2)
result_.write("{} ".format(distance))
result_.write("\n")
print(feat2)
file_.close()
result_.close()
def main():
LAMBDA = 1e-8
center_alpha = 0.9
num_class = 10000
embd_size = 512
args = argument()
checkpoint_dir = args.save_model_name
lr = args.lr
batch_size = args.batch_size
epoch_num = args.epoch_num
sta = args.sta
img_shape = args.img_size
train_file = args.train_file
val_file = args.val_file
train_batch_loader = BatchLoader(train_file, batch_size, img_shape)
test_batch_loader = BatchLoader(val_file, batch_size, img_shape)
#(Height,Width) = (train_batch_loader.height,train_batch_loader.width)
#train_batch_loader = mnist_data(batch_size)
print("img shape", img_shape)
with tf.name_scope('input'):
input_images = tf.placeholder(tf.float32,
shape=(batch_size, img_shape[0],
img_shape[1], 3),
name='input_images')
labels = tf.placeholder(tf.int32, shape=(batch_size), name='labels')
learn_rate = tf.placeholder(tf.float32,
shape=(None),
name='learn_rate')
with tf.name_scope('var'):
global_step = tf.Variable(0, trainable=False, name='global_step')
loss_op = CenterLoss(center_alpha, num_class, embd_size)
#with tf.device('/gpu:0'):
total_loss, accuracy, centers_update_op, center_loss, softmax_loss, pred_class, res1 = build_network(
input_images, labels, num_class, sta, loss_op, ratio=LAMBDA)
optimizer = tf.train.AdamOptimizer(learn_rate)
#optimizer = tf.train.GradientDescentOptimizer(learn_rate)
with tf.control_dependencies([centers_update_op]):
train_op = optimizer.minimize(total_loss, global_step=global_step)
#train_op = optimizer.minimize(total_loss, global_step=global_step)
summary_op = tf.summary.merge_all()
with tf.Session(config=tf.ConfigProto(log_device_placement=False)) as sess:
#sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter('../tmp/face_log', sess.graph)
saver = tf.train.Saver()
if args.pretrained is not None:
model_path = args.save_model_name + '-' + str(args.pretrained)
#saver.restore(sess,'./face_model/high_score-60')
saver.restore(sess, model_path)
else:
init = tf.global_variables_initializer()
sess.run(init)
step = sess.run(global_step)
epoch_idx = 0
graph_step = 0
while epoch_idx <= epoch_num:
step = 0
ckpt_fg = 'True'
ps_loss = 0.0
pc_loss = 0.0
acc_sum = 0.0
while step < train_batch_loader.batch_num:
batch_images, batch_labels = train_batch_loader.next_batch()
#batch_images, batch_labels = train_batch_loader.get_batchdata()
in_imgs = (batch_images - 127.5) * 0.0078125
#print("data in ",in_img[0,:2,:2,0])
_, summary_str, train_acc, Center_loss, Softmax_loss, Pred_class, res1_o = sess.run(
[
train_op, summary_op, accuracy, center_loss,
softmax_loss, pred_class, res1
],
feed_dict={
input_images: in_imgs,
labels: batch_labels,
learn_rate: lr
})
step += 1
#print("step",step, str(Softmax_loss),str(Center_loss))
#print("step label",step, str(batch_labels))
graph_step += 1
if step % 100 == 0:
writer.add_summary(summary_str, global_step=graph_step)
pc_loss += Center_loss
ps_loss += Softmax_loss
acc_sum += train_acc
if step % 1000 == 0:
#lr = lr*0.1
#c_loss+=c_loss
#s_loss+=s_loss
print("****** Epoch {} Step {}: ***********".format(
str(epoch_idx), str(step)))
print("center loss: {}".format(pc_loss / 1000.0))
print("softmax_loss: {}".format(ps_loss / 1000.0))
print("train_acc: {}".format(acc_sum / 1000.0))
print("centers", res1_o[0, :5])
print("*******************************")
#if (acc_sum/100.0) >= 0.98 and (pc_loss/100.0)<40 and (ps_loss/100.0) <0.1 and ckpt_fg=='True':
if ckpt_fg == 'True':
print(
"******************************************************************************"
)
saver.save(sess, checkpoint_dir, global_step=epoch_idx)
ckpt_fg = 'False'
ps_loss = 0.0
pc_loss = 0.0
acc_sum = 0.0
epoch_idx += 1
if epoch_idx % 10 == 0:
print(
"******************************************************************************"
)
saver.save(sess, checkpoint_dir, global_step=epoch_idx)
#writer.add_summary(summary_str, global_step=step)
if epoch_idx % 5 == 0:
lr = lr * 0.5
if epoch_idx:
batch_images, batch_labels = test_batch_loader.next_batch()
#batch_images,batch_labels = train_batch_loader.get_valdata()
vali_image = (batch_images - 127.5) * 0.0078125
vali_acc = sess.run(accuracy,
feed_dict={
input_images: vali_image,
labels: batch_labels
})
print(("epoch: {}, train_acc:{:.4f}, vali_acc:{:.4f}".format(
epoch_idx, train_acc, vali_acc)))
sess.close()
def train(self,
train_image_indices,
batch_size,
num_epochs=50,
train_method='normal',
lambda_1=0,
lambda_2=0,
start_from_pretrained_model=True,
learning_rate=0.01,
optimizer='SGD'):
if os.path.exists(self.checkpoint_path):
os.remove(self.checkpoint_path)
model = self.initialize_model(
start_from_pretrained_model=start_from_pretrained_model)
model = model.to(self.device)
criterion = nn.CrossEntropyLoss()
if optimizer == 'SGD':
optimizer = optim.SGD(model.parameters(),
lr=learning_rate,
momentum=0.9,
weight_decay=5e-4)
elif optimizer == 'Adam':
optimizer = optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=5e-4)
else:
optimizer = optim.SGD(model.parameters(),
lr=learning_rate,
momentum=0.9,
weight_decay=5e-4)
train_batch_loader = BatchLoader(self.train_folder_path,
train_image_indices)
n_images = len(train_image_indices)
if n_images % batch_size == 0:
num_batches = n_images // batch_size
else:
num_batches = (n_images // batch_size) + 1
penalty_inside_list = []
penalty_outside_list = []
train_acc_list = []
train_loss_list = []
val_loss_list = []
val_acc_list = []
best_acc = 0.0
for epoch in range(num_epochs):
model.train()
train_batch_loader.reset()
print('Epoch: {}/{}'.format(epoch + 1, num_epochs))
print('-' * 50)
train_correct = 0.0
train_loss = 0.0
penalty_inside = 0.0
penalty_outside = 0.0
for batch in range(num_batches):
batch_indices = train_batch_loader.get_batch_indices(
batch_size)
inputs = self.x_train[batch_indices]
labels = self.y_train[batch_indices]
inputs, labels = inputs.to(self.device), labels.to(self.device)
if train_method == 'bbox':
inputs.requires_grad_()
outputs = model(inputs)
preds = torch.argmax(outputs, dim=1)
# cross entropy loss
loss = criterion(outputs, labels)
input_gradient = torch.autograd.grad(loss,
inputs,
create_graph=True)[0]
penalty_inside_box, penalty_outside_box = self.calculate_penalty_box(
batch_indices, input_gradient)
new_loss = loss + lambda_1 * penalty_inside_box + lambda_2 * penalty_outside_box
optimizer.zero_grad()
new_loss.backward()
optimizer.step()
else:
outputs = model(inputs)
preds = torch.argmax(outputs, dim=1)
# cross entropy loss
loss = criterion(outputs, labels)
optimizer.zero_grad()
loss.backward()
optimizer.step()
penalty_inside_box = torch.tensor(0).to(self.device)
penalty_outside_box = torch.tensor(0).to(self.device)
train_loss += loss.item()
train_correct += torch.sum(preds == labels).float().item()
penalty_inside += penalty_inside_box.item() * lambda_1
penalty_outside += penalty_outside_box.item() * lambda_2
train_loss = train_loss / self.train_dataset_length
train_loss_list.append(train_loss)
train_acc = (train_correct / self.train_dataset_length) * 100.0
train_acc_list.append(train_acc)
penalty_inside = penalty_inside / self.train_dataset_length
penalty_outside = penalty_outside / self.train_dataset_length
penalty_inside_list.append(penalty_inside)
penalty_outside_list.append(penalty_outside)
print('Train Loss: {:.4f} Acc: {:.4f} % '.format(
train_loss, train_acc))
print(f'Penalty Inside Box: {round(penalty_inside, 4)}')
print(f'Penalty Outside Box: {round(penalty_outside, 4)}')
# validate after each epoch
val_correct = 0.0
val_loss = 0.0
model.eval()
with torch.no_grad():
for inputs_val, labels_val in self.val_loader:
inputs_val, labels_val = inputs_val.to(
self.device), labels_val.to(self.device)
outputs_val = model(inputs_val)
preds_val = torch.argmax(outputs_val, dim=1)
loss_test = criterion(outputs_val, labels_val)
val_loss += loss_test.item()
val_correct += torch.sum(
preds_val == labels_val).float().item()
val_loss = val_loss / self.val_dataset_length
val_loss_list.append(val_loss)
val_acc = (val_correct / self.val_dataset_length) * 100.0
val_acc_list.append(val_acc)
print('Val Loss: {:.4f} Acc: {:.4f} % \n'.format(
val_loss, val_acc))
# save the best model
if val_acc > best_acc:
best_acc = val_acc
model.state_dict()
if os.path.exists(self.checkpoint_path):
os.remove(self.checkpoint_path)
torch.save(model.state_dict(), self.checkpoint_path)
return_dict = {
'train_acc_list': train_acc_list,
'train_loss_list': train_loss_list,
'penalty_inside_list': penalty_inside_list,
'penalty_outside_list': penalty_outside_list,
'val_loss_list': val_loss_list,
'val_acc_list': val_acc_list,
'best_acc': best_acc
}
return return_dict
def __init__(self, sess, config):
self.sess = sess
self.config = config
self.tb_dir = "/data/tensorboard_log_dict/TDNN_white"
self.loader = BatchLoader(self.config)
def main():
# setting parameters
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--num_factors',
type=float,
default=10,
help='embedding size')
parser.add_argument('--model', type=str, default='mlp_bpr', help='model')
parser.add_argument('--epoches', type=str, default=1, help='epoches')
parser.add_argument('--learning_rate',
type=float,
default=0.01,
help='learning rate')
parser.add_argument('--reg_lambda',
type=float,
default=1.0,
help='l2_regularizer lambda')
parser.add_argument('--layers',
nargs='?',
default='[10,1]',
help="Size of each layer.")
parser.add_argument('--batch_size',
type=int,
default=10000,
help='minibatch size')
parser.add_argument(
'--recom_mode',
type=str,
default='p_u',
help='recommendation mode, u_p: users to items, p_u: items to users')
parser.add_argument('--decay_rate',
type=float,
default=0.99,
help='decay rate for Adam')
parser.add_argument('--keep_prob',
type=float,
default=0.2,
help='dropout probility')
parser.add_argument(
'--uti_k',
type=int,
default=30,
help='top-k recommendation for recommending items to user')
parser.add_argument(
'--itu_k',
type=int,
default=100,
help='top-k recommendation for recommending users to item')
parser.add_argument('--log_dir',
type=str,
default='logs',
help='directory to store tensorboard logs')
parser.add_argument(
'--mode',
type=str,
default='validation',
help=
'train: only train the model, validation: train the model and test it with test data, predict: predict new data'
)
parser.add_argument(
'--dev',
type=str,
default='cpu',
help=
'training by CPU or GPU, input cpu or gpu:0 or gpu:1 or gpu:2 or gpu:3'
)
parser.add_argument(
'--pIter',
type=int,
default=2,
help='how many rounds of iterations show the effect on the test set')
args = parser.parse_args()
if FLAGS.mode == 'validation':
# read data from file
train_file = 'data/train/train_u.csv'
test_file = 'data/test/test_u.csv'
test_count1_file = 'data/test/test_u_count1.csv'
train_X, train_y, test_X, test_y, test_count1_X, test_count1_y, user_pro_dict, pro_user_dict, user_index_map = utils.read_data_val(
train_file, test_file, test_count1_file)
print(test_count1_X)
# read feature data from file
user_file = 'data/mid/user_features.csv'
item_file = 'data/mid/item_features.csv'
user_feature_df = utils.read_user_data(user_file, user_index_map)
item_feature_df = utils.read_item_data(item_file)
# generate batches
batch_loader = BatchLoader(args.batch_size, train_X, train_y)
args.num_users = np.max([np.max(train_X['uid'])]) + 1
args.num_items = np.max([np.max(train_X['pid'])]) + 1
model = Model(args, args.recom_mode, batch_loader)
model.train_val(test_X, user_pro_dict, pro_user_dict, user_feature_df,
item_feature_df)
#pred_dict = model.test(test_X, user_pro_dict, pro_user_dict, user_feature_df, item_feature_df)
#if args.recom_mode=='u_p':
# print(model.cal_MRR(pred_dict, test_X))
#elif args.recom_mode=='p_u':
# print(model.cal_pre_k(pred_dict, test_X))
pred_dict = model.cold_start_test(test_count1_X, user_feature_df,
item_feature_df)
print(model.cal_pre_k(pred_dict, test_count1_X))
elif FLAGS.mode == 'train':
# read data from file
train_file = 'data/train/train_u.csv'
train_X, train_y, user_pro_dict, pro_user_dict, user_index_map = utils.read_data(
train_file)
# read feature data from file
user_file = 'data/mid/user_features.csv'
item_file = 'data/mid/item_features.csv'
user_feature_df = utils.read_user_data(user_file, user_index_map)
item_feature_df = utils.read_item_data(item_file)
# generate batches
batch_loader = BatchLoader(args.batch_size, train_X, train_y)
args.num_users = np.max(train_X['uid']) + 1
args.num_items = len(item_feature_df) + 1
model = Model(args, args.recom_mode, batch_loader)
# train and save models
model.train(user_pro_dict, pro_user_dict, user_feature_df,
item_feature_df)
# save the user and index
utils.write_file(user_index_map['user'], "save/user_index_map")
elif FLAGS.mode == 'predict':
user_index_file = 'save/user_index_map'
user_index_map = pd.read_csv(user_index_file, names=['user'])
args.num_users = len(user_index_map)
user_file = 'data/data_pred/user_file.txt'
#user_file = 'data/mid/user_features.csv.liaoning'
user_df = pd.read_csv(user_file, names=['user'])
#user_df = pd.read_csv(user_file,header=None)
#user_df.rename(columns={0: 'user'}, inplace=True)
#user_df = user_df['user']
item_file = 'data/data_pred/item_file.txt'
item_df = pd.read_csv(item_file, names=['pid'])
user_index_map = user_index_map.append(
user_df[['user']]).drop_duplicates(['user']).reset_index(drop=True)
user_index_map['uid'] = user_index_map.index
user_df = pd.merge(user_df,
user_index_map,
left_on='user',
right_on='user',
how='left')
#del user_df['user']
# read feature data from file
user_file = 'data/mid/user_features.csv'
item_file = 'data/mid/item_features.csv'
user_feature_df = utils.read_user_data(user_file, user_index_map)
item_feature_df = utils.read_item_data(item_file)
args.num_items = len(item_feature_df) + 1
out_file = 'data/predict/predictions'
model = Model(args, args.recom_mode)
result_list = model.predict(user_df['uid'].values.tolist(),
item_df['pid'].values.tolist(),
user_feature_df, item_feature_df, out_file)
user_list = user_df['user'].values.tolist()
head_line = "cate_id,"
head_line += (',').join(user_list)
result_list.insert(0, head_line)
with codecs.open(out_file, 'w', encoding='utf-8') as f:
for val in result_list:
f.write(val + "\n")
else:
print('incorrect mode input...')
def setup(self, bottom, top):
layer_params = yaml.load(self.param_str)
self.batch_size = layer_params['batch_size']
self.image_file = layer_params['image_file']
self.batch_loader = BatchLoader(self.image_file, self.batch_size)
# ## Import modules
# In[2]:
import os
import cv2
import numpy as np
import tensorflow as tf
import tflearn
from batch_loader import BatchLoader
slim = tf.contrib.slim
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
train_batch_loader = BatchLoader("./train.txt", 120)
test_batch_loader = BatchLoader("./test.txt", 100)
# ## Construct network
# In[3]:
with tf.name_scope('input'):
input_images = tf.placeholder(tf.float32,
shape=(None, 100, 100, 3),
name='input_images')
labels = tf.placeholder(tf.int64, shape=(None), name='labels')
global_step = tf.Variable(0, trainable=False, name='global_step')
# In[4]:
if __name__ == '__main__':
# Use GPUs if available
args.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Set the random seed for reproducible experiments
random.seed(args.seed)
torch.manual_seed(args.seed)
# Set the logger
utils.set_logger(os.path.join(args.model_dir, 'train.log'))
logging.info('device: {}'.format(args.device))
logging.info('Hyper params:%r' % args.__dict__)
# Create the input data pipeline
logging.info('Loading the datasets...')
bl = BatchLoader(args)
## Load train and dev data
train_data = bl.load_data('train.json')
dev_data = bl.load_data('dev.json')
## Train data
ner_train_data, re_train_data = bl.build_data(train_data, is_train=True)
train_bls = bl.batch_loader(ner_train_data,
re_train_data,
args.ner_max_len,
args.re_max_len,
args.batch_size,
is_train=True)
num_batchs_per_task = [len(train_bl) for train_bl in train_bls]
logging.info(
'num of batch per task for train: {}'.format(num_batchs_per_task))
train_task_ids = sum([[i] * num_batchs_per_task[i]
def __init__(self,
sess,
batch_size=100,
rnn_size=650,
layer_depth=2,
word_embed_dim=650,
char_embed_dim=15,
feature_maps=[50, 100, 150, 200, 200, 200, 200],
kernels=[1, 2, 3, 4, 5, 6, 7],
seq_length=35,
max_word_length=65,
use_word=False,
use_char=True,
hsm=0,
max_grad_norm=5,
highway_layers=2,
dropout_prob=0.5,
use_batch_norm=True,
checkpoint_dir="checkpoint",
forward_only=False,
data_dir="data",
dataset_name="pdb",
use_progressbar=False):
"""Initialize the parameters for LSTM TDNN
Args:
rnn_size: the dimensionality of hidden layers
layer_depth: # of depth in LSTM layers
batch_size: size of batch per epoch
word_embed_dim: the dimensionality of word embeddings
char_embed_dim: the dimensionality of character embeddings
feature_maps: list of feature maps (for each kernel width)
kernels: list of kernel widths
seq_length: max length of a word
use_word: whether to use word embeddings or not
use_char: whether to use character embeddings or not
highway_layers: # of highway layers to use
dropout_prob: the probability of dropout
use_batch_norm: whether to use batch normalization or not
hsm: whether to use hierarchical softmax
"""
self.sess = sess
self.batch_size = batch_size
self.seq_length = seq_length
# RNN
self.rnn_size = rnn_size
self.layer_depth = layer_depth
# CNN
self.use_word = use_word
self.use_char = use_char
self.word_embed_dim = word_embed_dim
self.char_embed_dim = char_embed_dim
self.feature_maps = feature_maps
self.kernels = kernels
# General
self.highway_layers = highway_layers
self.dropout_prob = dropout_prob
self.use_batch_norm = use_batch_norm
# Training
self.max_grad_norm = max_grad_norm
self.max_word_length = max_word_length
self.hsm = hsm
self.data_dir = data_dir
self.dataset_name = dataset_name
self.checkpoint_dir = checkpoint_dir
self.forward_only = forward_only
self.use_progressbar = use_progressbar
self.loader = BatchLoader(self.data_dir, self.dataset_name,
self.batch_size, self.seq_length,
self.max_word_length)
print('Word vocab size: %d, Char vocab size: %d, Max word length (incl. padding): %d' % \
(len(self.loader.idx2word), len(self.loader.idx2char), self.loader.max_word_length))
self.max_word_length = self.loader.max_word_length
self.char_vocab_size = len(self.loader.idx2char)
self.word_vocab_size = len(self.loader.idx2word)
# build LSTMTDNN model
self.prepare_model()
# load checkpoints
if self.forward_only == True:
if self.load(self.checkpoint_dir, self.dataset_name):
print(" [*] SUCCESS to load model for %s." % self.dataset_name)
else:
print(" [!] Failed to load model for %s." % self.dataset_name)
sys.exit(1)
from batch_loader import BatchLoader
import config
import random
if __name__ == '__main__':
param = {"net": "pnet", "batch": 64}
batch_loader = BatchLoader(param)
task = random.choice(config.TRAIN_TASKS[param['net']])
data = batch_loader.next_batch(64, task)
for datum in data:
print(datum)
