def parse_command_line():
parser = argparse.ArgumentParser(
description="""Train, validate, and test a classifier that will detect clouds in
remote sensing data.""")
parser.add_argument("-p", "--prepare-data", help="Prepare training and validation data.",
action="store_true")
parser.add_argument("-t", "--train", help="""Train classifier. Use --graph to generate quality
graphs""", action="store_true")
parser.add_argument("-g", "--graph", help="Generate training graphs.", action="store_true")
parser.add_argument("--weights", help="""The trained model weights to use; if not provided
defaults to the network that was just trained""", type=str, default=None)
parser.add_argument("--note", help="Adds extra note onto generated quality graph.", type=str)
args = vars(parser.parse_args())
if os.environ.get("CAFFE_HOME") == None:
print "You must set CAFFE_HOME to point to where Caffe is installed. Example:"
print "export CAFFE_HOME=/usr/local/caffe"
exit(1)
# Ensure the random number generator always starts from the same place for consistent tests.
random.seed(0)
if args["prepare_data"] == True:
prepare_data();
if args["train"] == True:
train(args["graph"], weight_file=args["weights"], note=args["note"])
def main(params):
# c ^= coarse, f ^= fine
(Xtr, ytr_c, ytr_f), (Xva, yva_c, yva_f), (Xte, yte_c, yte_f), (le_c, le_f) = df.datasets.cifar100.data()
model = cnn(32,
df.Parallel(
df.Sequential(df.Linear(1024, 20, init=df.init.const(0)), df.SoftMax()),
df.Sequential(df.Linear(1024,100, init=df.init.const(0)), df.SoftMax()),
)
)
crit = df.ParallelCriterion(
(0.75, df.ClassNLLCriterion()),
df.ClassNLLCriterion() # No weight defaults to one.
)
crit.add_penalty(1e-4, df.L1WeightDecay(model))
crit.add_penalty(1e-5, df.L2WeightDecay(model))
optim = df.AdaDelta(params['adadelta_rho'])
for epoch in range(100):
model.training()
train(Xtr, ytr_c, ytr_f, model, optim, crit, epoch, params['batch_size'], 'train')
train(Xtr, ytr_c, ytr_f, model, optim, crit, epoch, params['batch_size'], 'stats')
model.evaluate()
validate(Xva, yva_c, yva_f, model, epoch, params['batch_size'])
def load_train_save(options, adt, pbt, sfx, save_dir):
options_path = os.path.join(save_dir, "options")
save_dict_csv(options_path, options)
if options['load_params'] is True:
adt.load_params(options['params_file'])
if options['save_params'] is True:
path = os.path.join(save_dir, "final" + sfx)
adt.save_params(path)
if options['train'] is True:
train(adt, pbt, num_epochs=options['num_epochs'],
sfx=sfx, save_dir=save_dir, save_every=options['save_every'])
def test_arch(p):
acc = []
auc = []
prec = []
rec = []
f1 = []
plot_data = []
filename = ''.join(choice(ascii_uppercase) for i in range(6))
with open('arch/'+root_filename+'_'+str(p['test_fraction'])+'_'+filename+'.txt', 'w') as outfile:
json.dump(p, outfile)
for i in range(0,nb_rand):
result = train(batch_size=p['batch_size'], dropout=p['dropout'], nb_hidden=p['layers'], path=p['data'][0], test_fraction=p['test_fraction'],
classifier=p['classifier'][0], nb_epoch=p['nb_epoch'], bidirectional=p['bidirectional'], rnn_type=p['rnn'], save_model=p['save_model'],
plot_loss=p['plot_loss'], filename=root_filename+'_'+filename, plot_data=plot_data, optimizer=p['optimizer'], nb_augment=p['nb_augment'])
acc.append(result[0])
auc.append(result[1])
prec.append(result[2].tolist())
rec.append(result[3].tolist())
f1.append(result[4])
plot_data = result[5]
p['acc'] = acc
p['auc'] = auc
p['prec'] = prec
p['rec'] = rec
p['f1'] = f1
with open('arch/'+root_filename+'_'+str(p['test_fraction'])+'_'+filename+'.txt', 'w') as outfile:
json.dump(p, outfile)
def main(params):
train_set, valid_set, test_set = df.datasets.mnist.data()
train_set_x, train_set_y = train_set
test_set_x, test_set_y = test_set
model = lenet()
criterion = df.ClassNLLCriterion()
optimiser = df.SGD(lr=params['lr'])
for epoch in range(100):
model.training()
train(train_set_x, train_set_y, model, optimiser, criterion, epoch, params['batch_size'], 'train')
train(train_set_x, train_set_y, model, optimiser, criterion, epoch, params['batch_size'], 'stats')
model.evaluate()
validate(test_set_x, test_set_y, model, epoch, params['batch_size'])
def main():
args = parse_arguments()
if args.sub == "crawl":
lc = LawCrawler(args.domain, args.num_process, args.verbose)
data = lc.start()
with open(args.output, "w") as w:
w.write("{0}" .format(data))
if args.sub == "word2vec":
# To store result files
if not os.path.exists("result/"):
os.makedirs("result/")
# Store/load preprocessed data
try:
with open("result/preprocess", "rb") as pre_data:
print("Use preprocessd data.")
word_dict = pickle.load(pre_data)
word2idx = pickle.load(pre_data)
except IOError as e:
word2idx, word_dict = preprocess(args)
with open("result/preprocess", "wb") as pre_data:
pickle.dump(word_dict, pre_data)
pickle.dump(word2idx, pre_data)
start_time = time.time()
result = train(word2idx, args)
end_time = time.time()
print("Train word2vec done. {0:.2f} sec." .format(end_time-start_time))
# Store trained data (word vector).
with open("result/"+args.result, "wb") as w2v_data:
pickle.dump(result, w2v_data)
if args.sub == "vis":
try:
with open("result/result", "rb") as w2v_data:
result = pickle.load(w2v_data)
with open("result/preprocess", "rb") as pre_data:
word_dict = pickle.load(pre_data)
except IOError as e:
print("No result files.")
log.close()
start_time = time.time()
visualization(result, word_dict)
end_time = time.time()
print("t-SNE and visualization done. {0:.2f} sec." .format(end_time-start_time))
def main():
i = neurons(3)
h = neurons(5)
o = neurons(1)
n = network()
n += e_hidden(i, h)
n += e_out(h, o)
x = xor_dataset()
t = train(n)
while True:
print t.train(x)
def run(template, path):
cifar = prepare_cifar10()
cifar_train = cifar.train
cifar_train_stream = cifar.train_stream
cifar_validation = cifar.validation
cifar_validation_stream = cifar.validation_stream
cifar_test = cifar.test
cifar_test_stream = cifar.test_stream
print("Compiling...", end=" ")
sys.stdout.flush()
network = compile(template)
print("DONE")
sys.stdout.flush()
train(network, cifar_train_stream, cifar_validation_stream, 1e-3, 0.7, path)
print("Test error rate is %f%%" % (compute_error_rate(cifar_test_stream, network.predict) * 100.0,))
def main():
# args should include the following in sequence:
# file path for negative words
# file path for positive words
# directory of negative training/testing examples
# directory of positive training/testing examples
# partial=True means model test on testing split; otherwise on training split
# neg_hand=True means negation handling is enabled
# bigram_hand=True means bigram handling is enabled
partial = False
neg_hand = False
bigram_hand = False
args = ['/Users/tianshuren/Google Drive/10701/Homework 1/opinion-lexicon-English/negative-words.txt',
'/Users/tianshuren/Google Drive/10701/Homework 1/opinion-lexicon-English/positive-words.txt',
'/Users/tianshuren/Google Drive/10701/Homework 1/review_polarity/txt_sentoken/neg/*',
'/Users/tianshuren/Google Drive/10701/Homework 1/review_polarity/txt_sentoken/pos/*',
partial, neg_hand, bigram_hand]
N_Voc = args[0]
P_Voc = args[1]
N_Train = args[2]
P_Train = args[3]
N_Test = N_Train
P_Test = P_Train
config = [partial, neg_hand, bigram_hand]
bigrams = ['extremely', 'quite', 'just', 'almost', 'very', 'too', 'enough']
print 'building vocabulary...'
vocab = build(N_Voc, P_Voc)
print 'training neg files at ' + N_Train + '; pos files at ' + P_Train + '...'
(num_neg, num_pos, vocab) = train(N_Train, P_Train, vocab, bigrams, config)
print 'testing on negative files at ' + N_Test + '...'
(neg_correct, neg_total) = test(N_Test, 'negative', vocab, bigrams, num_neg, num_pos, config)
print 'testing on positive files at ' + P_Test + '...'
(pos_correct, pos_total) = test(P_Test, 'positive', vocab, bigrams, num_neg, num_pos, config)
total_accuracy = float(neg_correct + pos_correct) / (neg_total + pos_total)
print 'total accuracy is %f' % total_accuracy
def getanswer(msg,query,orig_query):
if msg == "<train status>":
ans = train(query)
elif msg == "<weather status>":
ans = weather(query)
elif msg == "<cricinfo module>":
ans = cric_info(orig_query)
elif msg == "<festival module>":
ans = festival(orig_query)
elif msg == "<tennis module>":
ans = tennis(orig_query)
elif msg == "<car module>":
ans=car(orig_query)
elif msg == "<tank module>":
ans=tank(orig_query)
elif msg == "<disease module>":
ans=disease(orig_query)
elif msg == "<flight module>":
ans = flight(orig_query)
elif msg == "<recipe module>":
ans = recipe(orig_query)
elif msg=="<discography module>":
ans=discography(orig_query)
elif msg=="<electronic module>":
ans=electronic(orig_query)
elif msg == "<wiki module>":
ans = wiki_module(orig_query)
elif msg == "<bank module>":
ans = bank_module(orig_query)
#elif msg == "<restaurant module>":
#ans = restaurant(orig_query)
#elif msg == "<website module>":
# ans = get_website(orig_query)
elif msg == "<stock status>":
ans = stock(query)
elif msg == "<mineral status>":
ans = mineral(query)
elif msg == "<sports status>":
ans = sports(query)
elif msg == "<movie review>":
ans = movie(query,orig_query)
elif msg == "<exam status>":
ans = exam(query)
elif msg == "<locationcentric module>":
ans = location(query)
elif msg == "<highway module>":
ans = highway(query)
elif msg == "<differences module>":
ans = differences(orig_query)
elif msg == "<currency module>":
ans = currency(query)
elif msg == "<meanings module>":
ans= meanings(orig_query)
elif msg == "<theatre module>":
ans= theatre(orig_query)
elif msg == "<minister module>":
ans= minister(orig_query)
elif msg == "<std module>":
ans= std(query)
elif msg == "<bday module>":
ans= [{"bday":[]}]
elif msg == "<highcourt module>":
ans = highcourt(orig_query)
#print "here"
return ans
def main(root_dir, logdir):
#-------------------------------------------------------------------------------
# SETUP
#-------------------------------------------------------------------------------
global_step = tf.Variable(0,trainable=False,name="global_step")
keep_prob = tf.placeholder(tf.float32, name="drop_prob")
l_r = tf.placeholder(tf.float32, name="learning_rate")
images_ = tf.placeholder(tf.float32, shape=[BATCH_SIZE,ROWS,COLS,DEPTH],name="image_placeholder")
labels_ = tf.placeholder(tf.int64, shape=[None], name="labels_plaeholder")
#read inputs
filenames_list = [os.path.join(root_dir, f) for f in os.listdir(root_dir) if os.path.isfile(os.path.join(root_dir, f)) and f[-8:] == 'tfrecord']
filenames_train = [f for f in filenames_list if "39_42" not in f]
random.shuffle(filenames_train)
filenames_val = [f for f in filenames_list if "39_42" in f]
random.shuffle(filenames_val)
print('Going to train on ', filenames_train)
print('Going to test on ', filenames_val)
#train_batch
images,labels = read_and_decode_single_example(filenames_train, IMAGE_SHAPE, ONLY_DEPTH)
#images = randomFlips(images)
images_batch,labels_batch = getShuffledMiniBatch(BATCH_SIZE,images,labels)
#val batch
images_v,labels_v = read_and_decode_single_example(filenames_val, IMAGE_SHAPE, ONLY_DEPTH)
images_batch_v, labels_batch_v = getShuffledMiniBatch(BATCH_SIZE,images_v,labels_v)
#visualize image in tensorboard
tf.image_summary('images',images_batch)
#models
model = buildDeepNet(images_,keep_prob, BATCH_SIZE,CLASSES,WORKING_SHAPE)
#weights to handle unbalanced training set
#weights = tf.constant([NEG_EXAMPLE_TRAIN/TOTAL_EXAMPLE,POS_EXAMPLE_TRAIN/TOTAL_EXAMPLE])
weights = tf.ones([1,CLASSES],tf.float32)
#loss
loss = computeLoss(model,labels_,weights,False)
#train
train_op = train(loss,global_step,l_r)
with tf.device("/cpu:0"):
#create a saver to save and restore models
saver = tf.train.Saver(tf.all_variables())
#accuracy op
softmaxed = tf.nn.softmax(model)
correct_prediction = tf.equal(tf.arg_max(softmaxed,1), labels_)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
#placeholder to print average_loss
average_pl = tf.placeholder(tf.float32)
average_summary = tf.scalar_summary("average_loss", average_pl)
#placeholder to print average validation accuracy
average_val_acc = tf.placeholder(tf.float32)
val_summary = tf.scalar_summary("average_validation_accuracy",average_val_acc)
#placeholder to print average train accuracy
average_train_acc = tf.placeholder(tf.float32)
val_summary = tf.scalar_summary("average_train_accuracy",average_train_acc)
#init operation
init = tf.initialize_all_variables()
#summary operation for visualization
summary_op = tf.merge_all_summaries()
#-----------------------------------------------------------------------------------------------
# TRAINING
#-----------------------------------------------------------------------------------------------
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
#logging folder
if not os.path.exists(logdir+'/log/'):
os.makedirs(logdir+'/log/')
if not os.path.exists(logdir+'/weights/'):
os.makedirs(logdir+'/weights/')
sum_writer = tf.train.SummaryWriter(logdir+'/log/',sess.graph)
step = 0
ckpt = tf.train.get_checkpoint_state(logdir+'/weights/')
if ckpt and ckpt.model_checkpoint_path:
print('Founded valid checkpoint file at: '+ckpt.model_checkpoint_path)
#restore variable
saver.restore(sess, ckpt.model_checkpoint_path)
#restore step
step = int(ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1])
print('Restored %d step model'%(step))
else:
print('No checkpoint file found, initialization instead')
#init all variables
sess.run(init)
#start queue
tf.train.start_queue_runners(sess=sess)
best_loss = np.inf
lr_countdown = LR_COUNTDOWN
working_lr = LEARNING_RATE
losses_history = []
val_acc_history = []
train_acc_history = []
while step < 40000:
start_time = time.time()
train_images, train_labels = sess.run([images_batch,labels_batch])
_, loss_value = sess.run([train_op, loss],
feed_dict={
keep_prob:0.5,
l_r:working_lr,
images_:train_images,
labels_:train_labels
})
losses_history.append(loss_value)
duration = time.time() - start_time
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step % 10 == 0:
num_examples_per_step = BATCH_SIZE
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
#print(losses_history,'\n',sum(losses_history),'\n',len(losses_history),'\n')
format_str = ('%s: epoch %d, step %d, loss = %.2f (%.1f examples/sec; %.3f sec/batch)')
print (format_str % (datetime.now(), step/STEP_PER_EPOCH, step, loss_value,examples_per_sec, sec_per_batch))
#compute validation score, save average loss
val_images, val_labels = sess.run([images_batch_v,labels_batch_v])
val_accuracy = sess.run(accuracy,
feed_dict={
keep_prob:1.0,
l_r:working_lr,
images_:val_images,
labels_:val_labels
})
train_accuracy=sess.run(accuracy,
feed_dict={
keep_prob:1.0,
l_r:working_lr,
images_:train_images,
labels_:train_labels
})
val_acc_history.append(val_accuracy)
train_acc_history.append(train_accuracy)
if step % 100 == 0:
#print('Losses_avg: ',sum(losses_history),'\n',len(losses_history),'\n')
avg_loss = sum(losses_history)/len(losses_history)
avg_val = sum(val_acc_history)/len(val_acc_history)
avg_train = sum(train_acc_history)/len(train_acc_history)
if avg_loss<best_loss:
best_loss = avg_loss
lr_countdown=LR_COUNTDOWN
else:
lr_countdown-=1
if lr_countdown==0:
best_loss = avg_loss
working_lr = working_lr/2
lr_countdown = LR_COUNTDOWN
summary_str = sess.run(summary_op,
feed_dict={
keep_prob:1.0,
l_r:working_lr,
images_:val_images,
labels_:val_labels,
average_pl: avg_loss,
average_val_acc: avg_val,
average_train_acc: avg_train
})
print("step %d, avg validation accuracy %g, avg train accuracy %g, avg loss %g, learning rate %g, countdown %d"%(step,avg_val,avg_train, avg_loss, working_lr, lr_countdown))
sum_writer.add_summary(summary_str, step)
losses_history = []
val_acc_history = []
train_acc_history = []
# Save the model checkpoint periodically.
if step % 1000 == 0 or (step + 1) == 10000 or (step<1000 and step%100==0):
checkpoint_path = os.path.join(logdir+'/weights/', 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
# Increase step
step+=1
ret5 = fully(ret4, [1024, 1], [1])
with tf.variable_scope("q1"):
retq1 = fully(ret4, [1024, 128], [128])
with tf.variable_scope("q_u"):
retq_u = tf.tanh(fully(retq1, [128, 1], [1]))
with tf.variable_scope("q_c"):
retq_c = tf.nn.softmax(fully(retq1, [128, cat_num], [cat_num]))
retq = [retq_u, retq_c]
return ret5, retq
def q_loss_computation(noise, retq):
[retq_u, retq_c] = retq
global cat_num
loss_u = 2 * tf.reduce_mean(
tf.square(noise[:, -1 - cat_num:-cat_num] - retq_u))
noise_oh = noise[:, -cat_num:]
q_relevant = tf.reduce_max(noise_oh * retq_c, axis=1)
loss_c = tf.reduce_mean(-tf.log(q_relevant))
return (loss_c + loss_u) / 2
train(generator, discriminator, tf_noise_generator, cpu_noise_generator,
cpu_noise_generator_tricky, q_loss_computation)
DEVICE = t.device('cuda')
map_location = None
else:
DEVICE = t.device('cpu')
map_location = 'cpu'
args.DEVICE = DEVICE
save_dir = os.path.join(os.getcwd(), args.output_dir)
# 构建两个相同结构的net,参数定期同步
RolloutNet = AttentionModel(args)
RolloutNet = RolloutNet.to(DEVICE)
baseNet = AttentionModel(args)
baseNet = baseNet.to(DEVICE)
baseNet.load_state_dict(RolloutNet.state_dict())
is_train = True # 是
if is_train:
if args.optimizer == 'adam':
optimizer = optim.Adam(RolloutNet.parameters(), lr=args.lr)
elif args.optimizer == 'sgd':
optimizer = optim.SGD(RolloutNet.parameters(), lr=args.lr)
elif args.optimizer == 'rmsprop':
optimizer = optim.RMSprop(RolloutNet.parameters(), lr=args.lr)
else:
raise ValueError('optimizer undefined: ', args.optimizer)
# 训练部分
train(args, optimizer, baseNet, RolloutNet)
else:
# 测试部分
evaluate(args, RolloutNet, map_location)
def main():
args = arg_parse()
print_args(args)
print("==> Creating dataloader...")
data_dir = args.data_path
train_list = './list/image/train.txt'
train_set = get_train_set(data_dir, train_list, args)
train_list1 ='./list/video/train.txt'
train_set1 = get_train_set(data_dir, train_list1, args)
train_list2 = './list/audio/train.txt'
train_set2 = get_train_set(data_dir, train_list2, args)
train_list3 = './list/text/train.txt'
train_set3 = get_text_set(data_dir, train_list3, args, 'train')
test_list = './list/image/test.txt'
test_set = get_test_set(data_dir, test_list, args)
test_list1 = './list/video/test.txt'
test_set1 = get_test_set(data_dir, test_list1, args)
test_list2 = './list/audio/test.txt'
test_set2 = get_test_set(data_dir, test_list2, args)
test_list3 = './list/text/test.txt'
test_set3 = get_text_set(data_dir, test_list3, args, 'test')
test_loader=DataLoader(dataset=test_set, num_workers=args.workers, batch_size=args.batch_size, shuffle=False)
test_loader1=DataLoader(dataset=test_set1, num_workers=args.workers, batch_size=args.batch_size, shuffle=False)
test_loader2=DataLoader(dataset=test_set2, num_workers=args.workers, batch_size=args.batch_size, shuffle=False)
test_loader3=DataLoader(dataset=test_set3, num_workers=args.workers, batch_size=args.batch_size, shuffle=False)
print("==> Loading the network ...")
model = resnet50(num_classes=200)
center_loss = CenterLoss(200, 200, True)
if args.gpu is not None:
# model = nn.DataParallel(model, device_ids=range(args.gpu))
model = model.cuda()
cudnn.benchmark = True
if True: # os.path.isfile(args.snapshot):
print("==> loading checkpoint '{}'".format(args.snapshot))
checkpoint = torch.load(args.snapshot)
model_dict = model.state_dict()
restore_param = {k: v for k, v in checkpoint.items() if k in model_dict}
model_dict.update(restore_param)
model.load_state_dict(model_dict)
print("==> loaded checkpoint '{}'".format(args.snapshot))
else:
print("==> no checkpoint found at '{}'".format(args.snapshot))
# exit()
criterion = nn.CrossEntropyLoss()
# '''只训练指定层'''
#for i, v in center_loss.named_parameters():
# v.requires_grad = False
# for i, v in model.named_parameters():
# if i != 'embed.weight' and i != "conv01.weight" \
# and i != "conv01.bias" and i != "conv02.weight" \
# and i != "conv02.bias" and i != "conv0.weight" \
# and i != 'conv0.bias'and i != 'fc1.weight'and i != 'fc1.bias':
# v.requires_grad = False
params = list(model.parameters()) + list(center_loss.parameters())
optimizer = optim.SGD(filter(lambda p: p.requires_grad,params),
lr=0.001, momentum=0.9)
for name, param in model.named_parameters():
if param.requires_grad:
print(name)
for name, param in center_loss.named_parameters():
if param.requires_grad:
print(name)
scheduler = lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.8)
savepath = args.model_path
if not os.path.exists(savepath):
os.makedirs(savepath)
if True:
print('-' * 20)
print("Video Acc:")
text_acc = validate(test_loader1, model, args,'v')
exit()
for epoch in range(args.epochs):
scheduler.step()
train_loader = DataLoader(dataset=train_set, num_workers=args.workers, batch_size=args.batch_size, shuffle=True)
train_loader1 = DataLoader(dataset=train_set1, num_workers=args.workers, batch_size=args.batch_size,
shuffle=True)
train_loader2 = DataLoader(dataset=train_set2, num_workers=args.workers, batch_size=args.batch_size,
shuffle=True)
train_loader3 = DataLoader(dataset=train_set3, num_workers=args.workers, batch_size=args.batch_size,
shuffle=True)
train(train_loader, train_loader1, train_loader2, train_loader3, args, model, criterion, center_loss, optimizer,
epoch, args.epochs)
if(epoch%1==0):
print("Image Acc:")
image1_acc = validate(test_loader, model, args, 'i')
print("Audio Acc:")
image_acc = validate(test_loader2, model, args, 'a')
print("Video Acc:")
text_acc = validate(test_loader1, model, args,'v')
print("Text Acc:")
image_acc = validate(test_loader3, model, args, 't')
save_model_path = savepath + 'epoch_' + str(epoch) + '_' + str(image_acc) + '.pkl'
torch.save(model.state_dict(), save_model_path)
from train import *
if __name__ == '__main__':
train = Train("E:\mtcnn_data", 12)
train(10000)
# train = Train("D:\celeba_3", 24)
# train(10000)
print("===> Building model")
model = MODEL(args)
criterion = nn.MSELoss()
print("===> Setting GPU")
if cuda:
model = model.cuda()
criterion = criterion.cuda()
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint["epoch"] + 1
model.load_state_dict(checkpoint["model"].state_dict())
else:
print("=> no checkpoint found at '{}'".format(args.resume))
print("===> Setting Optimizer")
optimizer = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=args.lr,
betas=(0.9, 0.99),
eps=1e-08)
print("===> Training")
for epoch in range(args.start_epoch, args.nEpochs + 1):
train(data_loader, optimizer, model, criterion, epoch, args)
save_checkpoint(model, epoch, 1)
val_ds = logSpecData(val_path, classes, sample_ratio=subsample)
print(f'{len(train_ds.img_paths)} training samples, {len(val_ds.img_paths)} validation samples')
print(f'classes {classes}')
weights, weight_dict = get_sampler_weights(train_ds)
sampler = WeightedRandomSampler(weights, len(weights))
train_dl = DataLoader(train_ds, 128, sampler=sampler, drop_last=True, num_workers=8)
valid_dl = DataLoader(val_ds, 128, num_workers=8)
#define network
#distribute module to gpu to work on parallel
net = cnn_trad_pool2_net()
net.cuda()
print(f'Training...')
# print(f'Network ready')
# if torch.cuda.is_available():
# net.cuda()
# n_devices = torch.cuda.device_count()
# print(f'Running model on {n_devices} GPUs\n')
# net = nn.DataParallel(net)
# else: print(f'Running model on cpu')
# define optimizer and criterion
criterion = F.cross_entropy
optimizer = optim.SGD(net.parameters(),lr=0.001, momentum=0.9, weight_decay=0.0001)
#optimizer = optim.Adam(net.parameters(), lr=0.01)
training_loss = train(net, train_dl, valid_dl, criterion, optimizer, epochs)
def run_seq2seq(train_data, test_data, word2index, index2word, word_embeddings,
out_fname, encoder_params_file, decoder_params_file,
max_target_length, n_epochs, batch_size, n_layers):
# Initialize q models
print 'Initializing models'
encoder = EncoderRNN(HIDDEN_SIZE,
word_embeddings,
n_layers,
dropout=DROPOUT)
decoder = AttnDecoderRNN(HIDDEN_SIZE, len(word2index), word_embeddings,
n_layers)
# Initialize optimizers
encoder_optimizer = optim.Adam(
[par for par in encoder.parameters() if par.requires_grad],
lr=LEARNING_RATE)
decoder_optimizer = optim.Adam(
[par for par in decoder.parameters() if par.requires_grad],
lr=LEARNING_RATE * DECODER_LEARNING_RATIO)
# Move models to GPU
if USE_CUDA:
encoder.cuda()
decoder.cuda()
# Keep track of time elapsed and running averages
start = time.time()
print_loss_total = 0 # Reset every print_every
# epoch = 0.0
epoch = 60.0
print 'Loading encoded, decoder params'
encoder.load_state_dict(
torch.load(encoder_params_file + '.epoch%d' % epoch))
decoder.load_state_dict(
torch.load(decoder_params_file + '.epoch%d' % epoch))
input_seqs, input_lens, output_seqs, output_lens = train_data
n_batches = len(input_seqs) / batch_size
teacher_forcing_ratio = 1.0
# decr = teacher_forcing_ratio/n_epochs
while epoch < n_epochs:
epoch += 1
for input_seqs_batch, input_lens_batch, \
output_seqs_batch, output_lens_batch in \
iterate_minibatches(input_seqs, input_lens, output_seqs, output_lens, batch_size):
start_time = time.time()
# Run the train function
loss = train(input_seqs_batch, input_lens_batch, output_seqs_batch,
output_lens_batch, encoder, decoder,
encoder_optimizer, decoder_optimizer,
word2index[SOS_token], max_target_length, batch_size,
teacher_forcing_ratio)
# Keep track of loss
print_loss_total += loss
# teacher_forcing_ratio = teacher_forcing_ratio - decr
print_loss_avg = print_loss_total / n_batches
print_loss_total = 0
print_summary = '%s %d %.4f' % (time_since(
start, epoch / n_epochs), epoch, print_loss_avg)
print(print_summary)
print 'Epoch: %d' % epoch
if epoch % 10 == 0:
print 'Saving model params'
torch.save(encoder.state_dict(),
encoder_params_file + '.epoch%d' % epoch)
torch.save(decoder.state_dict(),
decoder_params_file + '.epoch%d' % epoch)
if (epoch > n_epochs - 10) or (epoch % 10 == 0):
out_file = open(out_fname + '.epoch%d' % int(epoch), 'w')
evaluate(word2index, index2word, encoder, decoder, test_data,
batch_size, out_file)
def trainIters(encoder,
decoder,
vocab,
n_iters,
print_every=1000,
plot_every=100,
batch_size=32,
learning_rate=0.01,
teacher_forcing_ratio=1.0):
start = time.time()
plot_losses = []
plot_scores = []
print_loss_total = 0 # Reset every print_every
plot_loss_total = 0 # Reset every plot_every
encoder_optimizer = optim.SGD(encoder.parameters(), lr=learning_rate)
decoder_optimizer = optim.SGD(decoder.parameters(), lr=learning_rate)
# create dataloader
trainset = SpellingLoader('train', vocab)
trainloader = data.DataLoader(trainset, batch_size, shuffle=True)
criterion = nn.CrossEntropyLoss()
max_score = 0.8785
for iter in range(1, n_iters + 1):
for input_tensor, target_tensor in trainloader:
input_tensor = input_tensor.to(device)
target_tensor = target_tensor.to(device)
loss = train(input_tensor, target_tensor, encoder, decoder,
encoder_optimizer, decoder_optimizer, criterion,
teacher_forcing_ratio)
print_loss_total += loss
plot_loss_total += loss
# evaluate and save model
_, avg_bleu = evaluate(encoder, decoder, vocab)
plot_scores.append(avg_bleu)
if avg_bleu > max_score:
max_score = avg_bleu
print("Model save...")
torch.save(encoder,
"./models/encoder_{:.4f}.ckpt".format(avg_bleu))
torch.save(decoder,
"./models/decoder_{:.4f}.ckpt".format(avg_bleu))
if iter % print_every == 0:
print_loss_avg = print_loss_total / print_every
print_loss_total = 0
print('%s (%d %d%%) Loss: %.4f BLEU: %.4f' %
(timeSince(start, iter / n_iters), iter,
iter / n_iters * 100, print_loss_avg, avg_bleu))
plot_losses.append(plot_loss_total)
plot_loss_total = 0
print("The highest score is %s" % max_score)
return plot_scores, plot_losses
help='then crop to this size')
parser.add_argument('--n_epoch_init',
type=int,
default=10,
help='# of iter at pretrained')
parser.add_argument('--n_epoch',
type=int,
default=2000,
help='# of iter at training')
parser.add_argument('--beta1',
type=float,
default=0.9,
help='momentum term of adam')
parser.add_argument('--lr_init',
type=float,
default=1e-4,
help='initial learning rate for pretrain')
parser.add_argument('--lr_decay',
type=float,
default=0.1,
help='initial learning rate for adam')
parser.add_argument('--train', type=bool, default=True, help='train or valid')
opt = parser.parse_args()
print(opt)
if opt.train:
train(opt)
else:
valid(opt)
from time import localtime, strftime
import logging, scipy
import tensorflow as tf
import tensorlayer as tl
from model import *
from train import *
from evaluate import *
from utils import *
from config import config, log_config
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--mode',
type=str,
default='srgan',
help='srgan, evaluate')
args = parser.parse_args()
tl.global_flag['mode'] = args.mode
if tl.global_flag['mode'] == 'srgan':
train()
elif tl.global_flag['mode'] == 'evaluate':
evaluate()
else:
raise Exception("Unknow --mode")
for p in model.parameters()) / 1000000.0))
print('Total params: %.2fk' % (sum(p.numel()
for p in model.parameters()) / 1000.0))
print('************')
print(model)
#==============================================================================
# Start training
#==============================================================================
model, optimizer, epoch_hist = train(model,
train_loader,
lr=args.lr,
weight_decay=args.wd,
lamb=args.lamb,
num_epochs=args.epochs,
learning_rate_change=args.lr_decay,
epoch_update=args.lr_update,
nu=args.nu,
eta=args.eta,
backward=args.backward,
steps=args.steps,
steps_back=args.steps_back,
gradclip=args.gradclip)
torch.save(model.state_dict(), args.folder + '/model' + '.pkl')
#******************************************************************************
# Prediction
#******************************************************************************
Xinput, Xtarget = Xtest[:-1], Xtest[1:]
_, Xtarget = Xtest_clean[:-1], Xtest_clean[1:]
from train import *
from sklearn.cross_validation import KFold
from tensorflow.python.framework import ops
df = df.sample(frac=1)
kf = KFold(df.shape[0], n_folds=10, shuffle=True)
epoch_number = 1
i = 0
print(df.shape)
logger.info('start validation')
step = 0
for train_index, test_index in kf:
train_set = df.iloc[train_index, :]
test_set = df.iloc[test_index, :]
ops.reset_default_graph()
with tf.Session() as sess:
graphs = build_graph(sess)
train(train_set, './models/validate', epoch_number, graphs, sess)
test(test_set, './models/validate', graphs, sess, str(step))
step += 1
def test_rmse_should_be_less_than_1900000(self):
score = train()
self.assertLess(score, 1900000)
def main():
train()
weights, weight_dict = get_sampler_weights(train_ds)
sampler = WeightedRandomSampler(weights, len(weights))
train_dl = DataLoader(train_ds,
128,
sampler=sampler,
drop_last=True,
num_workers=8)
valid_dl = DataLoader(val_ds, 128, num_workers=8)
#define network
#distribute module to gpu to work on parallel
net = cnn_trad_pool2_net()
net.cuda()
print(f'Training...')
# print(f'Network ready')
# if torch.cuda.is_available():
# net.cuda()
# n_devices = torch.cuda.device_count()
# print(f'Running model on {n_devices} GPUs\n')
# net = nn.DataParallel(net)
# else: print(f'Running model on cpu')
# define optimizer and criterion
criterion = F.cross_entropy
optimizer = optim.SGD(net.parameters(),
lr=0.001,
momentum=0.9,
weight_decay=0.0001)
#optimizer = optim.Adam(net.parameters(), lr=0.01)
training_loss = train(net, train_dl, valid_dl, criterion, optimizer, epochs)
print("\tconstructing datasets and network...")
training_tweets, training_users, training_seq_lengths, valid_tweets, valid_users, valid_seq_lengths, test_tweets, test_users, test_seq_lengths = partite_dataset(tweets, users, seq_lengths)
#hyperparameter optimization if it is set
if FLAGS.optimize == False:
#print specs
print("---TRAINING STARTED---")
model_specs = "with parameters: Learning Rate:" + str(FLAGS.learning_rate) + ", Regularization parameter:" + str(FLAGS.l2_reg_lambda) + ", fully connected size:"
model_specs+= str(FLAGS.fc_size) + ", language:" + FLAGS.lang
print(model_specs)
#run the network
tf.reset_default_graph()
net = network(embeddings_char, embeddings_word)
train(net, training_tweets, training_users, training_seq_lengths, valid_tweets, valid_users, valid_seq_lengths, \
target_values, vocabulary_word, vocabulary_char, embeddings_char, embeddings_word)
else:
for i in range(len(FLAGS.rnn_cell_sizes)):
for learning_rate in FLAGS.l_rate:
for regularization_param in FLAGS.reg_param:
#prep the network
tf.reset_default_graph()
FLAGS.learning_rate = learning_rate
FLAGS.l2_reg_lambda = regularization_param
FLAGS.rnn_cell_size = FLAGS.rnn_cell_sizes[i]
FLAGS.num_filters = FLAGS.cnn_filter_counts[i]
net = network(embeddings_char, embeddings_word)
#print specs
model.cuda()
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
def callback(model, epoch):
torch.save(
model.state_dict(),
"./checkpoints/{:s}_trial{:d}_epoch{:d}.pt".format(
name, trial + 1, epoch + 1))
train_accuracy[:,
trial], test_accuracy[:,
trial], train_loss[:,
trial] = train(
model,
criterion,
optimizer,
trainloader,
testloader,
n_epochs,
device,
callback)
print("Trial took {:.1f} seconds".format(time.time() - trial_start))
torch.save(train_accuracy,
'./{:s}_train_accuracy_t{:d}.pt'.format(name, n_trials))
torch.save(test_accuracy,
'./{:s}_test_accuracy_t{:d}.pt'.format(name, n_trials))
torch.save(train_loss, './{:s}_train_loss_t{:d}.pt'.format(name, n_trials))
#
def main():
args = arg_parse()
print_args(args)
print("==> Creating dataloader...")
data_dir = args.data_path
train_list = './list/image/train.txt'
train_loader = get_train_set(data_dir, train_list, args)
train_list1 = './list/video/train.txt'
train_loader1 = get_train_set(data_dir, train_list1, args)
train_list = './list/audio/train.txt'
train_loader2 = get_train_set(data_dir, train_list, args)
train_list3 = './list/text/train.txt'
train_loader3 = get_text_set(data_dir, train_list3, args, 'train')
test_list = './list/image/test.txt'
test_loader = get_test_set(data_dir, test_list, args)
test_list1 = './list/video/test.txt'
test_loader1 = get_test_set(data_dir, test_list1, args)
test_list = './list/audio/test.txt'
test_loader2 = get_test_set(data_dir, test_list, args)
test_list3 = './list/text/test.txt'
test_loader3 = get_text_set(data_dir, test_list3, args, 'test')
print("==> Loading the network ...")
model = resnet50(num_classes=200)
if args.gpu is not None:
model = nn.DataParallel(model, device_ids=range(args.gpu))
model = model.cuda()
cudnn.benchmark = True
if os.path.isfile(args.snapshot):
print("==> loading checkpoint '{}'".format(args.snapshot))
checkpoint = torch.load(args.snapshot)
model_dict = model.state_dict()
restore_param = {k: v for k, v in checkpoint.items() if k in model_dict}
model_dict.update(restore_param)
model.load_state_dict(model_dict)
print("==> loaded checkpoint '{}'".format(args.snapshot))
else:
print("==> no checkpoint found at '{}'".format(args.snapshot))
exit()
criterion = nn.CrossEntropyLoss()
center_loss = CenterLoss(200, 200, True)
params = list(model.parameters()) + list(center_loss.parameters())
optimizer = optim.SGD(params, lr=0.001, momentum=0.9)
scheduler = lr_scheduler.StepLR(optimizer, step_size=3, gamma=0.5)
savepath = args.model_path
if not os.path.exists(savepath):
os.makedirs(savepath)
for epoch in range(args.epochs):
scheduler.step()
train(train_loader, train_loader1, train_loader2, train_loader3, args, model, criterion, center_loss, optimizer, epoch, args.epochs)
print('-' * 20)
print("Image Acc:")
image_acc = validate(test_loader, model, args, False)
print("Text Acc:")
text_acc = validate(test_loader3, model, args, True)
save_model_path = savepath + 'epoch_' + str(epoch) + '_' + str(image_acc) +'.pkl'
torch.save(model.state_dict(), save_model_path)
parser.add_argument("--ny", default=256, type=int, dest="ny")
parser.add_argument("--nx", default=256, type=int, dest="nx")
parser.add_argument("--nch", default=3, type=int, dest="nch")
parser.add_argument("--nker", default=64, type=int, dest="nker")
parser.add_argument("--wgt_cycle", default=1e1, type=float, dest="wgt_cycle")
parser.add_argument("--wgt_ident", default=5e-1, type=float, dest="wgt_ident")
parser.add_argument("--norm", default='inorm', type=str, dest="norm")
parser.add_argument("--network", default="PGGAN", choices=['DCGAN', 'pix2pix', 'CycleGAN','PGGAN'], type=str, dest="network")
parser.add_argument("--learning_type", default="plain", choices=["plain", "residual"], type=str, dest="learning_type")
args = parser.parse_args()
if __name__ == "__main__":
if args.mode == "train_single":
train(0,1,args)
elif args.mode == "train_multi":
gpu_devices = ','.join([str(id) for id in args.gpu_devices])
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_devices
ngpus_per_node = torch.cuda.device_count()
args.world_size = ngpus_per_node * args.world_size
mp.spawn(train, nprocs=ngpus_per_node, args=(ngpus_per_node, args))
elif args.mode == "test":
test(args)
args['class_num'] = len(LABEL.vocab)
args['embedding_matrix'] = TEXT.vocab.vectors
args["lr"] = 1e-3
args["log"] = False
args['epochs'] = 200
args["val_interval"] = 100
args["features_per_size"] = [100, 100, 100]
args["kernel_sizes"] = [3, 4, 5]
args["dropout_rate"] = 0.5
args["log_interval"] = 100
args["cuda"] = torch.cuda.is_available()
args["static"] = False
args["rand"] = True
rand_model = textCNN(args)
train(rand_model, train_iter, val_iter, args)
rand_model.load_state_dict(torch.load("rand_model.pt"))
_, test_acc = eval(rand_model, val_iter, args)
print("accuracy over test set:{:.6f}".format(test_acc))
args["rand"] = False
args["static"] = True
static_model = textCNN(args)
train(static_model, train_iter, val_iter, args)
static_model.load_state_dict(torch.load("static_model.pt"))
_, test_acc = eval(static_model, val_iter, args)
print("Accuracy on test set:{:.6f}".format(test_acc))
args["static"] = False
non_static_model = textCNN(args)
train(non_static_model, train_iter, val_iter, args)
def main():
if not torch.cuda.is_available():
print('no gpu device available')
sys.exit(1)
num_gpus = torch.cuda.device_count()
np.random.seed(args.seed)
args.gpu = args.local_rank % num_gpus
torch.cuda.set_device(args.gpu)
cudnn.benchmark = True
cudnn.deterministic = True
torch.manual_seed(args.seed)
cudnn.enabled = True
torch.cuda.manual_seed(args.seed)
group_name = 'search_space_shrinking'
print('gpu device = %d' % args.gpu)
print("args = %s", args)
torch.distributed.init_process_group(backend='nccl',
init_method='env://',
group_name=group_name)
args.world_size = torch.distributed.get_world_size()
args.batch_size = args.batch_size // args.world_size
criterion_smooth = CrossEntropyLabelSmooth(args.classes,
args.label_smooth).cuda()
total_iters = args.epochs * per_epoch_iters
# Max shrinking iterations
iters = config.op_num
# Prepare data
train_loader = get_train_dataloader(args.train_dir, args.batch_size,
args.local_rank, total_iters)
train_dataprovider = DataIterator(train_loader)
operations = []
for _ in range(config.edges):
operations.append(list(range(config.op_num)))
print('operations={}'.format(operations))
# Prepare model
base_model = Network_ImageNet().cuda(args.gpu)
model, seed = get_warmup_model(train_dataprovider, criterion_smooth,
operations, per_epoch_iters, args.seed,
args)
print('arch = {}'.format(model.module.architecture()))
optimizer, scheduler = get_optimizer_schedule(model, args, total_iters)
start_iter, ops_dim = 0, 0
checkpoint_tar = config.checkpoint_cache
if os.path.exists(checkpoint_tar):
checkpoint = torch.load(
checkpoint_tar,
map_location={'cuda:0': 'cuda:{}'.format(args.local_rank)})
start_iter = checkpoint['iter'] + 1
seed = checkpoint['seed']
operations = checkpoint['operations']
model.load_state_dict(checkpoint['state_dict'])
now = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
print('{} load checkpoint..., iter = {}, operations={}'.format(
now, start_iter, operations))
# Reset the scheduler
cur_iters = (config.first_stage_epochs +
(start_iter - 1) * config.other_stage_epochs
) * per_epoch_iters if start_iter > 0 else 0
for _ in range(cur_iters):
if scheduler.get_lr()[0] > args.min_lr:
scheduler.step()
# Save the base weights for computing angle
if start_iter == 0 and args.local_rank == 0:
torch.save(model.module.state_dict(), config.base_net_cache)
print('save base weights ...')
for i in range(start_iter, iters):
print('search space size: {}'.format(
get_search_space_size(operations)))
# ABS finishes when the size of search space is less than the threshold
if get_search_space_size(
operations) <= config.shrinking_finish_threshold:
# Save the shrunk search space
pickle.dump(operations, open(args.operations_path, 'wb'))
break
per_stage_iters = config.other_stage_epochs * per_epoch_iters if i > 0 else config.first_stage_epochs * per_epoch_iters
seed = train(train_dataprovider, optimizer, scheduler, model,
criterion_smooth, operations, i, per_stage_iters, seed,
args)
if args.local_rank == 0:
# Search space shrinking
load(base_model, config.base_net_cache)
operations = ABS(base_model, model.module, operations, i)
now = time.strftime('%Y-%m-%d %H:%M:%S',
time.localtime(time.time()))
print('{} |=> iter = {}, operations={}, seed={}'.format(
now, i + 1, operations, seed))
save_checkpoint(
{
'operations': operations,
'iter': i,
'state_dict': model.state_dict(),
'seed': seed
}, config.checkpoint_cache)
operations = merge_ops(operations)
ops_dim = len(operations)
# Synchronize variable cross multiple processes
ops_dim = broadcast(obj=ops_dim, src=0)
if not args.local_rank == 0:
operations = np.zeros(ops_dim, dtype=np.int)
operations = broadcast(obj=operations, src=0)
operations = split_ops(operations)
def main():
train(0.01)
patience=3,
verbose=True)
min_valid_loss = 2147483647.0
train_losses = []
valid_losses = []
train_acs = []
valid_acs = []
for ep in range(epochs):
print("Epoch {}".format(ep + 1))
train_loss, train_ac = train(model, optimizer, train_dataloader)
print("Training Loss:{}\tTraining Accuracy:{}".format(
train_loss, train_ac))
train_losses.append(train_loss)
train_acs.append(train_ac)
valid_loss, valid_ac = evaluate(model, valid_dataloader)
print("Validaiton Loss:{}\tValidation Accuracy:{}".format(
valid_loss, valid_ac))
if valid_loss < min_valid_loss:
min_valid_loss = valid_loss
torch.save(model.state_dict(), "../../model/Task1/best_model")
valid_losses.append(valid_loss)
def main():
train(0.0001)
# print predict([55.48216114069585,35.57070347228866])
predict()
parser.add_argument('--lr', type=float, default=2e-4)
args = parser.parse_args()
if __name__ == "__main__":
# devie---------------------------------------------------------------------------
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Fix random seed---------------------------------------------------------------------------
torch.manual_seed(1)
torch.cuda.manual_seed_all(1)
# dataset------------------------------------------------------------------------------------
train_dataloader = getDataLoader(args.dataset, args.batch_size,
args.dataset_path)
# model------------------------------------------------------------------------------------
model = build_model(args.experiment,
image_size=args.image_size,
lr=args.lr)
model = checkpointNet.load_part_network(model, args.checkpoint,
args.which_epoch)
model = model.to(device)
# save_dir_path-----------------------------------------------------------------------------------
save_dir_path = os.path.join(args.save_path, args.dataset)
os.makedirs(save_dir_path, exist_ok=True)
# train -----------------------------------------------------------------------------------
train(train_dataloader, model, device, save_dir_path, args)
from __future__ import print_function
from train import *
import unet
K.set_image_dim_ordering('th') # Theano dimension ordering in this code
history, model_weights = train(50)
model = unet.get_unet(dropout=True)
if inner:
model = unet.get_unet_inner(dropout=True)
model.load_weights(model_weights)
plot_loss_history(history)
test, uncertainty = evaluate_model(model)
target_optim = optim.Adam(target_net.parameters(), lr=lr)
shadow_net = shadow_net_type().to(device)
shadow_net.apply(models.weights_init)
shadow_loss = nn.CrossEntropyLoss()
shadow_optim = optim.Adam(shadow_net.parameters(), lr=lr)
attack_net = models.mlleaks_mlp(n_in=k).to(device)
attack_net.apply(models.weights_init)
attack_loss = nn.BCELoss()
attack_optim = optim.Adam(attack_net.parameters(), lr=lr)
train(shadow_net,
shadow_train_loader,
cifar10_testloader,
shadow_optim,
shadow_loss,
n_epochs,
classes=classes)
train_attacker(attack_net,
shadow_net,
shadow_train_loader,
shadow_out_loader,
attack_optim,
0.01,
n_epochs=1,
k=k)
# train_attacker(attack_net, shadow_net, shadow_train_loader, shadow_out_loader, attack_optim, attack_loss, n_epochs=1, k=k)
train(target_net,
target_train_loader,
cifar10_testloader,
from train import *
from tools import *
from libsvm.svmutil import svm_train, svm_save_model
from ml_ops import search, train, bootstrap
import argparse, sys, datetime
modes = ['train', 'bootstrap', 'search']
parser = argparse.ArgumentParser()
parser.add_argument("-m", "--mode", help="specify the mode [-m train|bootstrap|search]", required=True, choices=modes)
args = parser.parse_args()
if args.mode == 'train':
hists, labels = train()
svm = svm_train(labels, hists, '-s 0 -t 0 -c 1')
model_name = 'svm_trained_' + str(datetime.datetime.now()).replace(' ', '_') + '.dat'
svm_save_model(model_name, svm)
if args.mode == 'bootstrap':
hists, labels = train()
hists, labels = bootstrap(hists, labels)
svm = svm_train(labels, hists, '-s 0 -t 0 -c 1')
model_name = 'svm_bootstrapped_' + str(datetime.datetime.now()).replace(' ', '_') + '.dat'
svm_save_model(model_name, svm)
if args.mode == 'search':
search()
errors[m + t] = []
for i in range(args['repeats']):
print 'Model:', m
print 'Augmentation:', t
print 'Repeat:', i + 1
modelfn = os.path.join(args['modelpath'], m)
err, _, _ = train(
modelfn=modelfn,
trainfn=args['trainfn'],
valfn=args['valfn'],
epochs=model_epochs[m],
batchsize=128,
opt='Adam', opt_kwargs={'alpha': 0.001},
net_kwargs={},
transformations=t,
val_freq=25,
save_freq=25,
seed=i,
gpu=0,
silent=False,
logme=None
)
errors[m + t].append(err)
print '-- Evaluation run complete --'
print 'Model:', m,
print 'augmentations:', t
print 'error rates:', errors[m + t]
print 'mean error rate:', np.mean(errors[m + t]), '+/-', np.std(errors[m + t])
train_file_path = os.path.expanduser('~/.keras/datasets/VOC2012/combined_imageset_train.txt') #Data/VOClarge/VOC2012/ImageSets/Segmentation
# train_file_path = os.path.expanduser('~/.keras/datasets/oneimage/train.txt') #Data/VOClarge/VOC2012/ImageSets/Segmentation
val_file_path = os.path.expanduser('~/.keras/datasets/VOC2012/combined_imageset_val.txt')
data_dir = os.path.expanduser('~/.keras/datasets/VOC2012/VOCdevkit/VOC2012/JPEGImages')
label_dir = os.path.expanduser('~/.keras/datasets/VOC2012/combined_annotations')
if dataset is 'VOC2012':
train_file_path = os.path.expanduser('~/.keras/datasets/VOC2012/VOCdevkit/VOC2012/ImageSets/Segmentation/train.txt') #Data/VOClarge/VOC2012/ImageSets/Segmentation
# train_file_path = os.path.expanduser('~/.keras/datasets/oneimage/train.txt') #Data/VOClarge/VOC2012/ImageSets/Segmentation
val_file_path = os.path.expanduser('~/.keras/datasets/VOC2012/VOCdevkit/VOC2012/ImageSets/Segmentation/val.txt')
data_dir = os.path.expanduser('~/.keras/datasets/VOC2012/VOCdevkit/VOC2012/JPEGImages')
label_dir = os.path.expanduser('~/.keras/datasets/VOC2012/VOCdevkit/VOC2012/SegmentationClass')
classes = 21
class_weight = None
elif dataset is 'COCO':
train_file_path = os.path.expanduser('~/.keras/datasets/coco/annotations/train2014.txt') #Data/VOClarge/VOC2012/ImageSets/Segmentation
# train_file_path = os.path.expanduser('~/.keras/datasets/oneimage/train.txt') #Data/VOClarge/VOC2012/ImageSets/Segmentation
val_file_path = os.path.expanduser('~/.keras/datasets/coco/annotations/test2014.txt')
data_dir = os.path.expanduser('~/.keras/datasets/coco/train2014')
label_dir = os.path.expanduser('~/.keras/datasets/coco/seg_mask/train2014')
stats_file = os.path.expanduser('~/.keras/datasets/coco/seg_mask/train2014/image_segmentation_class_stats.json')
classes = 91
# class_weight = data_coco.class_weight(image_segmentation_stats_file=stats_file)
class_weight = None
config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
session = tf.Session(config=config)
K.set_session(session)
train(batch_size, epochs, lr_base, lr_power, weight_decay, classes, model_name, train_file_path, val_file_path,
data_dir, label_dir, target_size=target_size, batchnorm_momentum=batchnorm_momentum, resume_training=resume_training,
class_weight=class_weight, dataset=dataset)
def run_sim(params=None,parent=None):
if not params:
params=default_params()
elif isinstance(params,basestring):
fname=params
d=zpickle.load(fname)
params=d['params']
if parent:
save_sim_file=params['tmpfile']
params['display']=1
else:
save_sim_file=params['save_sim_file']
if params['load_sim_file']==params['tmpfile']:
params['load_sim_file']=''
if not save_sim_file:
save_sim_file='sims/untitled.dat'
# Deal the random number seed here
if params['random_seed']=='clock':
numpy.random.seed(None)
else:
numpy.random.seed(params['random_seed'])
params['actual_random_seed']=numpy.random.get_state()
params['random_seed2']=int(round(numpy.random.rand()*1e6))
start_time=time.time()
end_time=None
sim_time=None
num_channels=len(params['pattern_input'])
num_moments=get_num_moments(params)
layer=0
test_stimulus=params['test_stimulus'][layer]['type']
total_num_inputs=0
for p in params['pattern_input']:
total_num_inputs=total_num_inputs+p['num_inputs']
num_neurons=prod(params['num_neurons'])
if params['load_sim_file']:
d=zpickle.load(params['load_sim_file'])
load_sim_params=d['params']
weights=d['weights']
moments=d['moments']
initial_weights=d['weights']
initial_moments=d['moments']
t_mat=None
if params['continue']:
t_mat=d['t_mat']
moments_mat=d['moments_mat']
response_mat=d['response_mat']
weights_mat=d['weights_mat']
if not t_mat:
moments_mat=None
response_mat=None
else:
params['initial_weights']=weights
params['initial_moments']=moments
else:
params['continue']=0
full_mask=[]
for p in params['pattern_input']:
full_mask.append(get_circle_mask(p['num_inputs']))
full_mask=numpy.array(full_mask).ravel()
if not params['initial_weights']:
params['initial_weights']=[numpy.asarray(numpy.random.rand(num_neurons,total_num_inputs)*(params['initial_weight_range'][1]-params['initial_weight_range'][0])+params['initial_weight_range'][0],numpy.float64)]
params['initial_weights'][layer]=params['initial_weights'][layer]*(full_mask.repeat(num_neurons))
initial_weights=params['initial_weights']
if not params['initial_moments']:
params['initial_moments']=[numpy.asarray(numpy.random.rand(num_moments,num_neurons)*(params['initial_moment_range'][1]-params['initial_moment_range'][0])+params['initial_moment_range'][0],numpy.float64)]
initial_moments=params['initial_moments']
moments_mat=numpy.dstack( (initial_moments[0],) )
t_mat=None
if not t_mat: # not loaded from file
t_mat=[0]
start_epoch=1
moments_mat=numpy.dstack( (initial_moments[0],) )
if params['keep_every_epoch']:
weights_mat=[initial_weights]
else:
weights_mat=[]
response_mat=[]
response_var=[]
test_stimulus_type=params['test_stimulus'][layer]['type']
if test_stimulus_type==1: # test OR single
response_var=test_OR_single(params,initial_weights)
if not response_var:
params['test_stimulus'][layer]['type']=0
else:
for r in response_var:
response_mat.append([r[0]])
else:
start_epoch=len(t_mat)
test_stimulus_type=params['test_stimulus'][layer]['type']
response_var=[]
if test_stimulus_type==1: # test OR single
response_var=test_OR_single(params,initial_weights)
if not response_var:
params['test_stimulus'][layer]['type']=0
else:
for r,m in zip(response_var,response_mat):
m.append(r[0])
weights=deepcopy(initial_weights)
moments=deepcopy(initial_moments)
response_var_list=[response_var]
extra_mat=[]
sim={'params':params,
'weights':weights,
'moments':moments,
'moments_mat':moments_mat,
'weights_mat':weights_mat,
'response_mat':response_mat,
'response_var_list':response_var_list,
'initial_weights':weights,
'initial_moments':moments,
't_mat':t_mat,
'start_time':start_time,
'end_time':end_time,
'sim_time':sim_time,
'extra_mat':extra_mat}
params_with_images=deepcopy(params)
# make these point to the same thing
params_with_images['saved_input_vectors']=params['saved_input_vectors']
fix_images_directories(params_with_images)
status('Preemtive Save: %s' % save_sim_file,parent)
zpickle.save(sim,save_sim_file)
w=weights[layer]
m=moments[layer]
t0=time.time()
for epoch in range(start_epoch,start_epoch+params['epoch_number']):
t1=time.time()
extra=train(epoch-start_epoch,params_with_images,w,m)
if extra:
extra_mat.append(extra)
# copy over mask stuff
if epoch==start_epoch:
for i,p in enumerate(params['pattern_input']):
params['pattern_input'][i]['mask']=params_with_images['pattern_input'][i]['mask']
dt1=time.time()-t1
dt0=time.time()-t0
frac=(epoch-start_epoch+1.0)/params['epoch_number']
eta=sec2str(round(dt0/frac-dt0))
status("%.4f...ETA %s" % (dt1,eta),parent)
sim['moments_mat']=numpy.dstack( (sim['moments_mat'],m) )
test_stimulus_type=params['test_stimulus'][layer]['type']
if test_stimulus_type==1: # test OR single
response_var=test_OR_single(params,weights)
response_var_list[0]=response_var
if not response_var:
params['test_stimulus'][layer]['type']=0
else:
for r,mat in zip(response_var,response_mat):
mat.append(r[0])
t_mat.append(t_mat[-1]+params['iter_per_epoch'])
if params['keep_every_epoch']:
weights_mat.append(deepcopy(weights))
if params['display'] and epoch%params['epoch_per_display']==0:
if parent:
parent.Plot(sim)
else:
Plot(sim)
if parent:
parent.Yield()
if parent.Stopping():
break
if dt0>(60*60*20): # every 20 minutes
status('Incremental Save: %s' % (save_sim_file),parent)
zpickle.save(sim,save_sim_file)
end_time=time.time()
sim_time=end_time-start_time
if params['display']:
if parent:
parent.Plot(sim)
else:
Plot(sim)
status('Save: %s' % (save_sim_file),parent)
zpickle.save(sim,save_sim_file)
nb_epoch = config.getint('Train','num_epochs')
batch_size = config.getint('Train','batch_size')
nb_augment = config.getint('Data','num_augment')
path = config.get('Data','data')
classifier_type = config.get('Data','classifier')
if classifier_type == 'SN1a':
classifier = sn1a_classifier
elif classifier_type == '123':
classifier = type_classifier
elif classifier_type.lower() == 'Sub':
classifier = subtype_classifier
else:
raise Exception('Incorrect classifier')
root_filename = config.get('Options','filename')
plot_loss = config.getboolean('Options','plot_loss')
save_model = config.getboolean('Options','save_model')
with open('save/' + root_filename + '.ini', 'w') as f:
config.write(f)
start_time = time.time()
result = train(batch_size=batch_size, dropout=dropout, nb_hidden=nb_hidden, path=path, test_fraction=test_fraction,
classifier=classifier, nb_epoch=nb_epoch, bidirectional=bidirectional, rnn_type=rnn_type, save_model=save_model,
plot_loss=plot_loss, filename=root_filename, optimizer=optimizer, nb_augment=nb_augment,
consensus=consensus)
print("--- %s seconds ---" % (time.time() - start_time))
criterion = nn.NLLLoss(reduce=False)
#Define a learning rate optimizer
encoder_lr_decay_scheduler= optim.lr_scheduler.ReduceLROnPlateau(encoder_optimizer, factor=0.5,patience=5)
decoder_lr_decay_scheduler= optim.lr_scheduler.ReduceLROnPlateau(decoder_optimizer,factor=0.5,patience=5)
#Start Training
start = time.time()
print_loss_total = 0 #Reset every print_every
iter_count = 0
for epoch in range(1,n_epochs + 1):
for batch_x,batch_y,batch_x_lengths,batch_y_lengths,_ in data_generator(train_data_index,batch_size):
input_variable = batch_x #B*W_x
target_variable = batch_y #B*W_y
#Run the train function
loss = train(input_variable, target_variable, batch_x_lengths,batch_y_lengths,encoder, decoder, encoder_optimizer, decoder_optimizer, criterion)
print_loss_total += loss
if iter_count == 0:
iter_count += 1
continue
if iter_count%print_every == 0:
print_loss_avg = print_loss_total / print_every
print_loss_total = 0
print_summary = '%s (%d %d%%) %.4f' % (time_since(start, iter_count / n_epochs / batch_num), iter_count, iter_count / n_epochs / batch_num * 100, print_loss_avg)
print(print_summary)
if iter_count%eval_every == 0:
#Print the Bleu Score and loss for Dev Dataset
val_losses = 0
textCreate(pygame, screen, "Your average pace: %s, Your standard deviation: %s" %(mean_now, int(std_now)), white)
pygame.time.delay(4000)
textCreate(pygame, screen,"You are done! Thank you!", white)
elif trial == 10 and first == 0:
textCreate(pygame, screen, "Your average pace: %s, Your standard deviation: %s" %(mean_now, int(std_now)), white)
pygame.time.delay(4000)
textCreate(pygame, screen, "The pace will now change to 550ms", white)
pygame.time.delay(2000)
textCreate(pygame, screen,"Trial complete. Prepare for next trial.", white)
elif trial == 10 and first == 1:
textCreate(pygame, screen, "Your average pace: %s, Your standard deviation: %s" %(mean_now, int(std_now)), white)
pygame.time.delay(4000)
textCreate(pygame, screen, "The pace will now change to 350ms", white)
pygame.time.delay(2000)
textCreate(pygame, screen,"Trial complete. Prepare for next trial.", white)
else:
textCreate(pygame, screen, "Your average pace: %s, Your standard deviation: %s" %(mean_now, int(std_now)), white)
pygame.time.delay(4000)
textCreate(pygame, screen,"Trial complete. Prepare for next trial.", white)
pygame.time.delay(3000)
if __name__ == '__main__':
train(pygame, screen)
run()
createFile(name, trial_type, session, player_times, std, m, r_std, r_m)
# Mouse response rate
# 500hz - 2ms
# plot covariance Fn
# if person speeds over time shows + corr need to detrend data
# Wing analysis
parser.add_argument(
'-ee',
'--eval-every',
type=int,
default=10,
help='The number of epochs after which to print the validation loss')
parser.add_argument('--cuda',
type=bool,
default=False,
help='Whether to use cuda or not')
parser.add_argument('--mode',
choices=['train', 'test'],
default='train',
help='Whether to train or test')
FLAGS, unparsed = parser.parse_known_args()
FLAGS.cuda = torch.device('cuda:0' if torch.cuda.is_available() and FLAGS.cuda \
else 'cpu')
if FLAGS.mode.lower() == 'train':
train(FLAGS)
elif FLAGS.mode.lower() == 'test':
test(FLAGS)
else:
raise RuntimeError(
'Unknown mode passed. \n Mode passed should be either \
of "train" or "test"')
criterion = BCEDiceLoss
if config["optimizer"] == "Adam":
optimizer = optim.Adam(params,
lr=config["lr"],
weight_decay=config["weight_decay"])
else:
optimizer = optim.SGD(params,
lr=config["lr"],
momentum=0.9,
weight_decay=config["weight_decay"])
scheduler = CosineAnnealingLR(optimizer,
T_max=config["epochs"],
eta_min=config["min_lr"])
train_dl, test_dl = get_loader(config["batch_size"], config["num_workers"])
log = train(config,
train_dl,
test_dl,
model,
optimizer,
scheduler,
criterion,
metric=iou)
# analysis
plot_log(log)
#! /usr/local/bin/python3
import os
import json
_train = True
_inference = True
from train import *
os.environ["SM_MODEL_DIR"] = "./mock/out"
os.environ["SM_CHANNEL_TRAIN"] = "./mock/data"
if _train:
train({"team": "jmc"}, [],
0,
out_dir=os.environ["SM_MODEL_DIR"],
current_host="local")
if _inference:
from host import *
model = model_fn("./mock/out")
body = json.dumps({
"opponent": "test",
"width": 10,
"height": 10,
"ruleSet": {
"shipCells": 11,
"connectedShipCells": 5
}
})
response = transform_fn(model, body, 'image/jpeg', 'image/jpeg')
print(response)
# lstm = LSTM_plain(input_size=args.max_prev_node, embedding_size=args.embedding_size_lstm,
# hidden_size=args.hidden_size, num_layers=args.num_layers).cuda()
if 'GraphRNN_VAE_conditional' in args.note:
rnn = GRU_plain(input_size=args.max_prev_node, embedding_size=args.embedding_size_rnn,
hidden_size=args.hidden_size_rnn, num_layers=args.num_layers, has_input=True,
has_output=False).cuda()
output = MLP_VAE_conditional_plain(h_size=args.hidden_size_rnn, embedding_size=args.embedding_size_output, y_size=args.max_prev_node).cuda()
elif 'GraphRNN_MLP' in args.note:
rnn = GRU_plain(input_size=args.max_prev_node, embedding_size=args.embedding_size_rnn,
hidden_size=args.hidden_size_rnn, num_layers=args.num_layers, has_input=True,
has_output=False).cuda()
output = MLP_plain(h_size=args.hidden_size_rnn, embedding_size=args.embedding_size_output, y_size=args.max_prev_node).cuda()
elif 'GraphRNN_RNN' in args.note:
rnn = GRU_plain(input_size=args.max_prev_node, embedding_size=args.embedding_size_rnn,
hidden_size=args.hidden_size_rnn, num_layers=args.num_layers, has_input=True,
has_output=True, output_size=args.hidden_size_rnn_output).cuda()
output = GRU_plain(input_size=1, embedding_size=args.embedding_size_rnn_output,
hidden_size=args.hidden_size_rnn_output, num_layers=args.num_layers, has_input=True,
has_output=True, output_size=1).cuda()
### start training
train(args, dataset_loader, rnn, output)
### graph completion
# train_graph_completion(args,dataset_loader,rnn,output)
### nll evaluation
# train_nll(args, dataset_loader, dataset_loader, rnn, output, max_iter = 200, graph_validate_len=graph_validate_len,graph_test_len=graph_test_len)
opti = optim.Adam(net.parameters(), lr=lr)
scheduler = torch.optim.lr_scheduler.StepLR(
opti, step_size=step_size)
print('Start training, with batch size ', batch_size,
' maxlen ', maxlen, ' random sample ',
doRandom_train)
print('Proportion of random ', proportionRandom, ', lr ',
lr, 'stepSize', step_size, ' Epochs ', epochs)
net, train_error_all, val_error_all = train(
net,
criterion,
opti,
scheduler,
train_loader,
val_loader,
epochs=epochs,
use_cuda=use_cuda)
print(
"Finished training, best RMSE on the validation was ",
np.sqrt(np.max(val_error_all)))
res[count] = {
'batch_size': batch_size,
'maxlen': maxlen,
'propRandom': proportionRandom,
'lr': lr,
'step_size': step_size,
def main(train_file, model_file, test_file, test_predict, params):
targets, features = read_stack_data(train_file)
model = train(targets, features, model_file, params)
predict_stacked(model, test_file, test_predict)
gamma = 0.99
env = make_env()
obs_shape = env.observation_space.shape
n_actions = env.action_space.n
print("Observation shape:", obs_shape)
print("Num actions:", n_actions)
print("Action names:", env.env.env.get_action_meanings())
n_parallel_games = n_parallel_games
gamma = gamma
from agent import Agent
agent = Agent(obs_shape, n_actions, n_parallel_games)
if cuda:
agent.cuda()
chkpt_dir = "./chkpt"
pool = EnvPool(agent, make_env, n_parallel_games)
train(agent,
pool,
50000,
n_parallel_games,
gamma,
save_path=chkpt_dir,
curiosity=True)
#!/usr/bin/env python
#coding=utf-8
from parameters import *
from model import *
from train import *
if __name__ == '__main__':
net = ResNet(ResidualBlock)
net.loadIfExist()
'''
for elem in net.named_parameters():
print(elem)
'''
train(net)
'~/.keras/datasets/coco/annotations/test2014.txt')
data_dir = os.path.expanduser('~/.keras/datasets/coco/train2014')
label_dir = os.path.expanduser(
'~/.keras/datasets/coco/seg_mask/train2014')
stats_file = os.path.expanduser(
'~/.keras/datasets/coco/seg_mask/train2014/image_segmentation_class_stats.json'
)
classes = 91
# class_weight = data_coco.class_weight(image_segmentation_stats_file=stats_file)
class_weight = None
config = tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))
session = tf.Session(config=config)
K.set_session(session)
train(batch_size,
epochs,
lr_base,
lr_power,
weight_decay,
classes,
model_name,
train_file_path,
val_file_path,
data_dir,
label_dir,
target_size=target_size,
batchnorm_momentum=batchnorm_momentum,
resume_training=resume_training,
class_weight=class_weight,
dataset=dataset)
