def check_collective_communication(param, use_gpu):
communicator = create_communicator(param, use_gpu)
mpi_comm.barrier()
model = ExampleModel(param.model_dtype)
if use_gpu:
model.to_gpu()
check_bcast_data(communicator, model)
check_allreduce_grad(communicator, model)
check_allreduce_grad_empty(communicator, model)
# Check allreduce debug mode
model = ExampleModel()
if use_gpu:
model.to_gpu()
# The example model includes some nan parameters so the debug mode
# must detect it.
chainer.set_debug(True)
with pytest.raises(ValueError, match=r'.* diverged .*'):
check_allreduce_grad(communicator, model)
chainer.set_debug(False)
# barrier() requires before destructor of PureNcclCommunicator
# because communication may not be finished.
mpi_comm.barrier()
destroy_communicator(communicator)
def main(args):
chainer.set_debug(True)
# Initialize the model to train
model = models.archs[args.arch]()
if args.finetune and hasattr(model, 'finetuned_model_path'):
utils.finetuning.load_param(model.finetuned_model_path, model)
if args.initmodel:
print('Load model from', args.initmodel)
chainer.serializers.load_npz(args.initmodel, model)
outputdir = os.path.join('.', args.out, args.arch)
if args.initmodel:
outputdir = os.path.dirname(args.initmodel)
if not os.path.exists(outputdir):
os.makedirs(outputdir)
for layer in args.layer:
outputpath = os.path.join(
outputdir, "{0}_distance_{1}.npy".format(layer, args.dis))
similar_path = os.path.join(
outputdir, "{0}_similar_{1}.tsv".format(layer, args.dis))
best_path = os.path.join(
outputdir, "{0}_best{1}_{2}.tsv".format(layer, args.best,
args.dis))
W, b = model[layer].W.data, model[layer].b.data
if args.withbias:
params = np.hstack((W.reshape(
(W.shape[0], -1)), b.reshape((b.shape[0], -1))))
else:
params = W.reshape((W.shape[0], -1))
print("{0}\tW: {1}\tb: {2}".format(layer, W.shape, b.shape))
N = params.shape[0]
#d = np.zeros(N**2).reshape((N, N))
'''for i, p in enumerate(tqdm(itertools.product(params, repeat=2))):
w1, w2 = p
if args.dis == 'cos':
d[i] = dis.cosine(w1, w2)
elif args.dis == 'euclidean':
d[i] = dis.euclidean(w1, w2)'''
if args.dis == 'cos':
d = cos_distance_matrix(params)
elif args.dis == 'euclidean':
d = euclidean_distance_matrix(params)
#d = d.reshape((N, N))
np.savetxt(outputpath, d, delimiter=",")
d_list = [(int(x[0]), int(x[1]), d[x[0], x[1]])
for x in itertools.combinations(six.moves.range(N), 2)]
if args.threshold:
similar_list = filter((lambda x: x[2] < args.threshold), d_list)
save_list(similar_path, similar_list)
if args.best:
order_by_disance_ascending = sorted(d_list, key=lambda x: x[2])
save_list(best_path, order_by_disance_ascending[0:args.best])
def setUp(self):
self.original_debug = chainer.is_debug()
chainer.set_debug(True)
self.one = numpy.array(1, numpy.float32)
self.f = chainer.FunctionNode()
self.return_value = tuple(None if x is None else chainer.Variable(x)
for x in self.return_data)
def setUp(self):
self.original_debug = chainer.is_debug()
chainer.set_debug(True)
self.one = numpy.array(1, numpy.float32)
self.f = chainer.FunctionNode()
self.return_value = tuple(None if x is None else chainer.Variable(x)
for x in self.return_data)
def check_collective_communication(param, use_gpu):
communicator = create_communicator(param, use_gpu)
mpi_comm.barrier()
model = ExampleModel(param.model_dtype)
if use_gpu:
model.to_gpu()
check_bcast_data(communicator, model)
check_allreduce_grad(communicator, model)
check_allreduce_grad_empty(communicator, model)
# Check allreduce debug mode
model = ExampleModel()
if use_gpu:
model.to_gpu()
# The example model includes some nan parameters so the debug mode
# must detect it.
chainer.set_debug(True)
with pytest.raises(ValueError, match=r'.* diverged .*'):
check_allreduce_grad(communicator, model)
chainer.set_debug(False)
# barrier() requires before destructor of PureNcclCommunicator
# because communication may not be finished.
mpi_comm.barrier()
destroy_communicator(communicator)
def setUp(self):
chainer.set_debug(True)
np.random.seed(0)
dataset = VOC('train')
img, im_info, bbox = dataset[1]
self.x = Variable(img[None, ...])
self.im_info = Variable(im_info[None, ...])
self.gt_boxes = Variable(bbox[None, ...])
def setUp(self):
chainer.set_debug(True)
np.random.seed(0)
dataset = VOC('train')
img, im_info, bbox = dataset[1]
self.x = Variable(img[None, ...])
self.im_info = Variable(im_info[None, ...])
self.gt_boxes = Variable(bbox[None, ...])
def main(args):
chainer.set_debug(True)
# Initialize the model to train
model = models.archs[args.arch]()
if args.finetune and hasattr(model, 'finetuned_model_path'):
utils.finetuning.load_param(model.finetuned_model_path, model)
if args.initmodel:
print('Load model from', args.initmodel)
chainer.serializers.load_npz(args.initmodel, model)
outputdir = os.path.join('.', args.out, args.arch)
if args.initmodel:
outputdir = os.path.dirname(args.initmodel)
if not os.path.exists(outputdir):
os.makedirs(outputdir)
for layer in args.layer:
outputpath = os.path.join(outputdir,
"{0}_distance_{1}.npy".format(layer, args.dis))
similar_path = os.path.join(outputdir,
"{0}_similar_{1}.tsv".format(layer, args.dis))
best_path = os.path.join(outputdir,
"{0}_best{1}_{2}.tsv".format(layer, args.best, args.dis))
W, b = model[layer].W.data, model[layer].b.data
if args.withbias:
params = np.hstack((W.reshape((W.shape[0], -1)), b.reshape((b.shape[0], -1))))
else:
params = W.reshape((W.shape[0], -1))
print("{0}\tW: {1}\tb: {2}".format(layer, W.shape, b.shape))
N = params.shape[0]
#d = np.zeros(N**2).reshape((N, N))
'''for i, p in enumerate(tqdm(itertools.product(params, repeat=2))):
w1, w2 = p
if args.dis == 'cos':
d[i] = dis.cosine(w1, w2)
elif args.dis == 'euclidean':
d[i] = dis.euclidean(w1, w2)'''
if args.dis == 'cos':
d = cos_distance_matrix(params)
elif args.dis == 'euclidean':
d = euclidean_distance_matrix(params)
#d = d.reshape((N, N))
np.savetxt(outputpath, d, delimiter=",")
d_list = [(int(x[0]), int(x[1]), d[x[0],x[1]]) for x in
itertools.combinations(six.moves.range(N), 2)]
if args.threshold:
similar_list = filter((lambda x: x[2] < args.threshold), d_list)
save_list(similar_path, similar_list)
if args.best:
order_by_disance_ascending = sorted(d_list, key=lambda x: x[2])
save_list(best_path, order_by_disance_ascending[0:args.best])
def setUp(self):
chainer.set_debug(True)
np.random.seed(0)
x = np.random.randint(0, 255, size=(224, 224, 3)).astype(np.float)
x -= np.array([[[102.9801, 115.9465, 122.7717]]])
self.x = np.expand_dims(x, 0).transpose(0, 3, 1, 2).astype(np.float32)
self.im_info = np.array([[224, 224, 1.6]])
self.gt_boxes = np.array([[10, 10, 60, 200, 0], [50, 100, 210, 210, 1],
[160, 40, 200, 70, 2]])
def main(args):
chainer.set_debug(True)
# Initialize the model to train
model = models.archs[args.arch]()
if hasattr(model, 'finetuned_model_path'):
finetuning.load_param(model.finetuned_model_path, model)
if args.initmodel:
print('Load model from', args.initmodel)
chainer.serializers.load_npz(args.initmodel, model)
layer = model[args.layer]
outputdir = os.path.join('.', args.out, args.arch)
outputpath = os.path.join(outputdir, args.layer + '_distance.npy')
similar_path = os.path.join(outputdir, args.layer + '_similar.tsv')
best_path = os.path.join(outputdir, args.layer + '_best.tsv')
if not os.path.exists(outputdir):
os.makedirs(outputdir)
W = layer.W.data
b = layer.b.data
if args.withbias:
params = np.hstack((W.reshape(
(W.shape[0], -1)), b.reshape((b.shape[0], -1))))
else:
params = W.reshape((W.shape[0], -1))
print('W', W.shape)
print('b', b.shape)
N = params.shape[0]
d = np.zeros(N**2)
for i, p in enumerate(itertools.product(params, repeat=2)):
w1, w2 = p
if args.dis == 'cos':
d[i] = dis.cosine(w1, w2)
elif args.dis == 'euclidean':
d[i] = dis.euclidean(w1, w2)
d = d.reshape((N, N))
np.savetxt(outputpath, d, delimiter=",")
d_list = [(int(x[0]), int(x[1]), d[x[0], x[1]])
for x in itertools.combinations_with_replacement(
six.moves.range(N), 2)]
d_list = filter((lambda x: x[0] != x[1]), d_list)
if args.threshold:
similar_list = filter((lambda x: x[2] < args.threshold), d_list)
save_list(similar_path, similar_list)
if args.best:
order_by_disance_ascending = sorted(d_list, key=lambda x: x[2])
save_list(best_path, order_by_disance_ascending[0:args.best])
return d
def main():
parser = argparse.ArgumentParser(description='Dynamic SGM Net')
parser.add_argument('--gpu', '-g', type=int, default=-1, help='GPU ID (negative value indicates CPU)')
parser.add_argument('--out', '-o', default='output', help='Directory to output the result')
parser.add_argument('--vol', '-v', type=str2bool, default=False, help='Save cost volume data')
args = parser.parse_args()
outdir = args.out
print('cuda:' + str(chainer.cuda.available))
print('cudnn:' + str(chainer.cuda.cudnn_enabled))
print('GPU: {}'.format(args.gpu))
print('outdir: ', outdir)
print('')
chainer.config.train = False
chainer.set_debug(False)
chainer.using_config('use_cudnn', 'auto')
# Load MC-CNN pre-trained models from
# kitti_fast, kitti_slow, kitti2015_fast, kitti2015_slow, mb_fast, mb_slow
model_kitti = mcnet.MCCNN_pretrained('mccnn/kitti_fast')
model_mb = mcnet.MCCNN_pretrained('mccnn/mb_slow')
if args.gpu >= 0:
# Make a specified GPU current
chainer.cuda.get_device_from_id(args.gpu).use()
model_kitti.to_gpu() # Copy the model to the GPU
model_mb.to_gpu() # Copy the model to the GPU
samples = []
sattellite_flag = 9999
#samples.append((model_mb, 'mb2014', 145))
#samples.append((model_kitti, 'kitti', 70))
samples.append((model_mb, 'satellite', sattellite_flag))
for sample in samples:
model, target, ndisp = sample
print('Processing ' + target)
im0 = load_image(os.path.join('input', target, 'im0.png')).astype(np.float32)
im1 = load_image(os.path.join('input', target, 'im1.png')).astype(np.float32)
inputs = (im0, im1, np.array([ndisp]))
batch = chainer.dataset.concat_examples([inputs], args.gpu)
with chainer.no_backprop_mode():
vol = model(*batch)[0].array
if target == 'satellite':
dmin, dmax = load_ndisp(os.path.join('input', target, 'ndisp.txt'))
ndisp = dmax-dmin
disp = vol.argmin(0).astype(np.float32) * (255 / ndisp)
os.makedirs(os.path.join(args.out, target), exist_ok=True)
cv2.imwrite(os.path.join(args.out, target, 'disp0.png'), chainer.cuda.to_cpu(disp))
if args.vol:
vol.tofile(os.path.join(args.out, target, 'im0.bin'))
def setUp(self):
chainer.set_debug(True)
np.random.seed(0)
x = np.random.randint(0, 255, size=(224, 224, 3)).astype(np.float)
x -= np.array([[[102.9801, 115.9465, 122.7717]]])
self.x = np.expand_dims(x, 0).transpose(0, 3, 1, 2).astype(np.float32)
self.im_info = np.array([[224, 224, 1.6]])
self.gt_boxes = np.array([
[10, 10, 60, 200, 0],
[50, 100, 210, 210, 1],
[160, 40, 200, 70, 2]
])
def main():
chainer.set_debug(True)
parser = get_parser()
args = parser.parse_args()
reset_seed(args.seed)
#load vocabulary
source0_ids = load_vocabulary(args.SOURCE_VOCAB0)
source1_ids = load_vocabulary(args.SOURCE_VOCAB1)
target_ids = load_vocabulary(args.TARGET_VOCAB)
corpus = make_data_tuple(source0=(source0_ids, args.SOURCE0),
source1=(source1_ids, args.SOURCE1),
target=(target_ids, args.TARGET))
source0_words = {i: w for w, i in source0_ids.items()}
source1_words = {i: w for w, i in source1_ids.items()}
target_words = {i: w for w, i in target_ids.items()}
# Setup model
model = Seq2seq(args.layer, len(source0_ids), len(source1_ids),
len(target_ids), args.unit)
if args.resume:
# Resume from a snapshot
print("Load Model")
chainer.serializers.load_npz(args.resume, model)
if args.gpu >= 0:
chainer.backends.cuda.get_device(args.gpu).use()
model.to_gpu(args.gpu)
for i in range(len(corpus)):
source0, source1, target = corpus[i]
result = model.translate([model.xp.array(source0)],
[model.xp.array(source1)])[0]
source0_sentence = ' '.join([source0_words[x] for x in source0])
source1_sentence = ' '.join([source1_words[x] for x in source1])
target_sentence = ' '.join([target_words[y] for y in target])
result_sentence = ' '.join([target_words[y] for y in result])
print('# source0 : ' + source0_sentence)
print('# source1 : ' + source1_sentence)
print('# result : ' + result_sentence)
print('# expect : ' + target_sentence)
print("")
def check_collective_communication(param, use_gpu, use_chx):
communicator = create_communicator(param, use_gpu, use_chx)
mpi_comm.barrier()
model = ExampleModel(param.model_dtype)
if use_gpu:
device = chainermn.testing.get_device(communicator.intra_rank, use_chx)
else:
device = chainermn.testing.get_device(use_chainerx=use_chx)
model.to_device(device)
check_bcast_data(communicator, model)
model = ExampleModel(param.model_dtype)
model.to_device(device)
check_multi_node_mean_grad(communicator, model)
model = ExampleModel(param.model_dtype)
model.to_device(device)
check_multi_node_mean_grad_empty(communicator, model)
model = ExampleModel(param.model_dtype)
model.to_device(device)
check_multi_node_mean_grad_empty_half(communicator, model)
# Check allreduce debug mode
model = ExampleModel()
model.to_device(device)
# The example model includes some nan parameters so the debug mode
# must detect it.
chainer.set_debug(True)
with pytest.raises(ValueError, match=r'.* diverged .*'):
check_multi_node_mean_grad(communicator, model)
chainer.set_debug(False)
# barrier() requires before destructor of PureNcclCommunicator
# because communication may not be finished.
mpi_comm.barrier()
communicator.finalize()
def setUp(self):
self.original_debug = chainer.is_debug()
chainer.set_debug(True)
self.one = numpy.array([1], numpy.float32)
self.f = chainer.Function()
def setUp(self):
self.default_debug = chainer.is_debug()
chainer.set_debug(True)
self.a_data = numpy.random.uniform(-1, 1, (4, 3, 2))
self.b_data = numpy.random.uniform(-1, 1, (2, 2))
def tearDown(self):
chainer.set_debug(False)
def tearDown(self):
chainer.set_debug(False)
def main(args):
chainer.set_debug(True)
# Initialize the model to train
model = models.archs[args.arch]()
if args.finetune and hasattr(model, 'finetuned_model_path'):
utils.finetuning.load_param(model.finetuned_model_path, model, args.ignore)
if args.initmodel:
print('Load model from', args.initmodel)
chainer.serializers.load_npz(args.initmodel, model)
if args.gpu >= 0:
chainer.cuda.get_device(args.gpu).use() # Make the GPU current
#cuda.cudnn_enabled = False
model.to_gpu()
nowt = datetime.datetime.today()
outputdir = os.path.join(args.out, args.arch, 'extract')
if args.initmodel is not None:
outputdir = os.path.dirname(args.initmodel)
# Load the datasets and mean file
mean = None
if hasattr(model, 'mean_value'):
mean = makeMeanImage(model.mean_value)
else:
mean = np.load(args.mean)
assert mean is not None
val = ppds.PreprocessedDataset(args.val, args.root, mean, model.insize, False)
val_iter = chainer.iterators.SerialIterator(
val, args.val_batchsize, repeat=False, shuffle=False)
# Set up an optimizer
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(model)
# Set up a trainer
updater = training.StandardUpdater(val_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (1, 'epoch'), outputdir)
#val_interval = (10 if args.test else int(len(train) / args.batchsize)), 'iteration'
val_interval = (1, 'iteration')
#snapshot_interval = (10, 'iteration') if args.test else (2, 'epoch')
#log_interval = (10, 'iteration')
# Copy the chain with shared parameters to flip 'train' flag only in test
eval_model = model.copy()
eval_model.train = False
val_extractor = utils.Extractor(val_iter, eval_model, device=args.gpu)
val_extractor.layer_rank = eval_model.layer_rank[args.layer]
val_extractor.layer_name = args.layer
val_extractor.operation = args.operation
val_extractor.save_features = args.savefeatures
val_extractor.top = args.top
if 'googlenet' in args.arch:
val_extractor.lastname = 'validation/main/loss3'
trainer.extend(val_extractor, trigger=val_interval)
#trainer.extend(extensions.PrintReport([
# 'epoch', 'iteration', 'main/loss', 'validation/main/loss',
# 'main/accuracy', 'validation/main/accuracy',
#]), trigger=log_interval)
#trainer.extend(extensions.ProgressBar(update_interval=10))
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
results = val_extractor(trainer)
results['outputdir'] = outputdir
return results
def main(args):
chainer.set_debug(True)
# Initialize the model to train
model = models.archs[args.arch]()
if args.finetune and hasattr(model, 'finetuned_model_path'):
utils.finetuning.load_param(model.finetuned_model_path, model, args.ignore)
if args.initmodel:
print('Load model from', args.initmodel)
chainer.serializers.load_npz(args.initmodel, model)
if args.gpu >= 0:
chainer.cuda.get_device(args.gpu).use() # Make the GPU current
#if args.test:
#cuda.cudnn_enabled = False
model.to_gpu()
nowt = datetime.datetime.today()
outputdir = os.path.join(args.out, args.arch, 'extract')
if args.initmodel is not None:
outputdir = os.path.dirname(args.initmodel)
#if args.indices is None:
# args.indices = os.path.join(outputdir, 'features', 'top_' + args.layer + '.txt')
# Load the datasets and mean file
mean = None
if hasattr(model, 'mean_value'):
mean = makeMeanImage(model.mean_value)
else:
mean = np.load(args.mean)
assert mean is not None
#if args.indices is None:
# args.indices = os.path.join(args.out, args.arch, 'extract', 'top_' + args.layer + '.txt')
#top_path = os.path.join(args.out, args.arch, 'extract', 'top_' + args.layer + '.txt')
#train = ppds.PreprocessedDataset(args.train, args.root, mean, model.insize)
val = ppds.PreprocessedDataset(args.val, args.root, mean, model.insize, False)
# These iterators load the images with subprocesses running in parallel to
# the training/validation.
#train_iter = chainer.iterators.MultiprocessIterator(
# train, args.batchsize, shuffle=False, n_processes=args.loaderjob)
#val_iter = chainer.iterators.MultiprocessIterator(
# val, args.val_batchsize, repeat=False, shuffle=False, n_processes=args.loaderjob)
val_iter = chainer.iterators.SerialIterator(
val, args.val_batchsize, repeat=False, shuffle=False)
# Set up an optimizer
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(model)
# Set up a trainer
updater = training.StandardUpdater(val_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (1, 'epoch'), outputdir)
#val_interval = (10 if args.test else int(len(train) / args.batchsize)), 'iteration'
val_interval = (1, 'iteration')
#snapshot_interval = (10, 'iteration') if args.test else (2, 'epoch')
log_interval = (10, 'iteration')
# Copy the chain with shared parameters to flip 'train' flag only in test
eval_model = model.copy()
eval_model.train = False
val_acquirer = utils.SaliencyMapAcquirer(val_iter, eval_model, device=args.gpu)
val_acquirer.mean = mean
val_acquirer.layer_rank = eval_model.layer_rank[args.layer]
val_acquirer.layer_name = args.layer
val_acquirer.operation = args.operation
val_acquirer.fixed_RMS = args.rms
#val_acquirer.top = args.top
val_acquirer.target = args.target
val_acquirer.n_features = model.labelsize
if 'googlenet' in args.arch:
val_acquirer.lastname = 'validation/main/loss3'
trainer.extend(val_acquirer, trigger=val_interval)
#trainer.extend(extensions.dump_graph('main/loss'))
#trainer.extend(extensions.snapshot(), trigger=snapshot_interval)
#trainer.extend(extensions.snapshot_object(
# model, 'model_iter_{.updater.iteration}'), trigger=snapshot_interval)
# Be careful to pass the interval directly to LogReport
# (it determines when to emit log rather than when to read observations)
#trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'main/loss', 'validation/main/loss',
'main/accuracy', 'validation/main/accuracy',
]), trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
#trainer.extend(extensions.ExponentialShift('lr', args.gamma),
# trigger=(1, 'epoch'))
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
#if not args.test:
# trainer.run()
# chainer.serializers.save_npz(outputdir + '/model', model)
results = val_acquirer(trainer)
results['outputdir'] = outputdir
#if eval_model.layer_rank[args.layer] == 1:
# save_first_conv_filter(os.path.join(outputdir, args.layer),
# model[args.layer].W.data, cols = args.cols, pad = args.pad,
# scale = args.scale, gamma = args.gamma)
#if args.test:
#print(val_acquirer.confmat)
#categories = utils.io.load_categories(args.categories)
#confmat_csv_name = args.initmodel + '.csv'
#confmat_fig_name = args.initmodel + '.eps'
#utils.io.save_confmat_csv(confmat_csv_name, val_acquirer.confmat, categories)
#utils.io.save_confmat_fig(confmat_fig_name, val_acquirer.confmat, categories,
# mode="rate", saveFormat="eps")
return results
def main(args):
chainer.set_debug(True)
# Initialize the model to train
model = models.archs[args.arch]()
if args.finetune and hasattr(model, 'finetuned_model_path'):
finetuning.load_param(model.finetuned_model_path, model, args.ignore)
if args.initmodel:
print('Load model from', args.initmodel)
chainer.serializers.load_npz(args.initmodel, model)
if args.gpu >= 0:
chainer.cuda.get_device(args.gpu).use() # Make the GPU current
#if args.test:
#cuda.cudnn_enabled = False
model.to_gpu()
nowt = datetime.datetime.today()
outputdir = args.out + '/' + args.arch + '/' + nowt.strftime(
"%Y%m%d-%H%M") + '_bs' + str(args.batchsize)
if args.test and args.initmodel is not None:
outputdir = os.path.dirname(args.initmodel)
# Load the datasets and mean file
mean = None
if hasattr(model, 'mean_value'):
mean = makeMeanImage(model.mean_value)
else:
mean = np.load(args.mean)
assert mean is not None
train = ppds.PreprocessedDataset(args.train, args.root, mean, model.insize)
val = ppds.PreprocessedDataset(args.val, args.root, mean, model.insize,
False)
# These iterators load the images with subprocesses running in parallel to
# the training/validation.
train_iter = chainer.iterators.MultiprocessIterator(
train, args.batchsize, shuffle=False, n_processes=args.loaderjob)
#val_iter = chainer.iterators.MultiprocessIterator(
# val, args.val_batchsize, repeat=False, shuffle=False, n_processes=args.loaderjob)
val_iter = chainer.iterators.SerialIterator(val,
args.val_batchsize,
repeat=False,
shuffle=False)
# Set up an optimizer
optimizer = chainer.optimizers.MomentumSGD(lr=args.baselr, momentum=0.9)
optimizer.setup(model)
# Set up a trainer
updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (args.epoch, 'epoch'), outputdir)
#val_interval = (10 if args.test else int(len(train) / args.batchsize)), 'iteration'
val_interval = (10, 'iteration') if args.test else (1, 'epoch')
snapshot_interval = (10, 'iteration') if args.test else (2, 'epoch')
log_interval = (10 if args.test else 200), 'iteration'
# Copy the chain with shared parameters to flip 'train' flag only in test
eval_model = model.copy()
eval_model.train = False
if not args.test:
val_evaluator = extensions.Evaluator(val_iter,
eval_model,
device=args.gpu)
else:
val_evaluator = evaluator_plus.EvaluatorPlus(val_iter,
eval_model,
device=args.gpu)
if 'googlenet' in args.arch:
val_evaluator.lastname = 'validation/main/loss3'
trainer.extend(val_evaluator, trigger=val_interval)
trainer.extend(extensions.dump_graph('main/loss'))
trainer.extend(extensions.snapshot(), trigger=snapshot_interval)
trainer.extend(extensions.snapshot_object(
model, 'model_iter_{.updater.iteration}'),
trigger=snapshot_interval)
# Be careful to pass the interval directly to LogReport
# (it determines when to emit log rather than when to read observations)
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.PrintReport([
'epoch',
'iteration',
'main/loss',
'validation/main/loss',
'main/accuracy',
'validation/main/accuracy',
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(extensions.ExponentialShift('lr', args.gamma),
trigger=(1, 'epoch'))
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
if not args.test:
trainer.run()
chainer.serializers.save_npz(outputdir + '/model', model)
results = val_evaluator(trainer)
results['outputdir'] = outputdir
if args.test:
print(val_evaluator.confmat)
categories = dataio.load_categories(args.categories)
confmat_csv_name = args.initmodel + '.csv'
confmat_fig_name = args.initmodel + '.eps'
dataio.save_confmat_csv(confmat_csv_name, val_evaluator.confmat,
categories)
dataio.save_confmat_fig(confmat_fig_name,
val_evaluator.confmat,
categories,
mode="rate",
saveFormat="eps")
return results
def setUp(self):
self.default_debug = chainer.is_debug()
chainer.set_debug(True)
self.x_data = numpy.random.uniform(-1, 1, (4, 3, 2))
def setUp(self):
self.x = numpy.arange(10).reshape((2, 5)).astype('f')
self.ind = numpy.array(self.indices, 'i')
self.debug = chainer.is_debug()
chainer.set_debug(True)
from multi_task.multi_task_300 import Multi_task_300
from multi_task.multi_task_512 import Multi_task_512
from datasets.multi_task_dataset_voc import Multi_task_VOC
from config.datasets import voc_experiments
from datasets.transforms import Transform
import chainer.functions as F
from multi_task.evaluator.multi_task_evaluator import MultitaskEvaluator
import numpy as np
from multi_task.segmentation.loss.losses import MultiLoss
chainer.set_debug(False)
class MultiboxTrainChain(chainer.Chain):
def __init__(self,
model,
gpu=False,
alpha=1,
k=3,
use_multi_task_loss=False,
use_dynamic_loss=False,
loss_split=0.5):
super(MultiboxTrainChain, self).__init__()
self.gpu = gpu
with self.init_scope():
def setUp(self):
self.default_debug = chainer.is_debug()
chainer.set_debug(True)
# training parameter
parser.add_argument('--batchsize', '-b', type=int, default=128)
parser.add_argument('--epoch',
'-e',
type=int,
default=10,
help='The number of training epoch')
# model parameter
parser.add_argument('--unit-num',
'-u',
type=int,
default=512,
help='The unit size of each layer in MLP')
args = parser.parse_args()
chainer.set_debug(True)
train, val = kaggle.get_kaggle()
train_iter = I.SerialIterator(train, args.batchsize)
val_iter = I.SerialIterator(val, args.batchsize, repeat=False, shuffle=False)
C = len(kaggle.task_names)
predictor = mlp.MLP(args.unit_num, C)
classifier = classifier.Classifier(predictor=predictor)
if args.gpu >= 0:
cuda.get_device(args.gpu).use()
classifier.to_gpu()
optimizer = O.SGD()
optimizer.setup(classifier)
def main():
chainer.set_debug(True)
parser = get_parser()
args = parser.parse_args()
reset_seed(args.seed)
#load vocabulary
source0_ids = load_vocabulary(args.SOURCE_VOCAB0)
source1_ids = load_vocabulary(args.SOURCE_VOCAB1)
target_ids = load_vocabulary(args.TARGET_VOCAB)
print('Source vocabulary size: %d' % len(source0_ids))
print('Source vocabulary size: %d' % len(source1_ids))
print('Target vocabulary size: %d' % len(target_ids))
train_data = make_data_tuple(source0=(source0_ids, args.SOURCE0),
source1=(source1_ids, args.SOURCE1),
target=(target_ids, args.TARGET))
source0_words = {i: w for w, i in source0_ids.items()}
source1_words = {i: w for w, i in source1_ids.items()}
target_words = {i: w for w, i in target_ids.items()}
# Setup model
model = Seq2seq(args.layer, len(source0_ids), len(source1_ids),
len(target_ids), args.unit)
if args.gpu >= 0:
chainer.backends.cuda.get_device(args.gpu).use()
model.to_gpu(args.gpu)
# Setup optimizer
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(args.l2))
# Setup iterator
train_iter = chainer.iterators.SerialIterator(train_data, args.batchsize)
# Setup updater and trainer
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
converter=convert,
device=args.gpu)
trainer = training.Trainer(updater, (args.epoch, 'epoch'), out=args.out)
trainer.extend(
extensions.LogReport(trigger=(args.log_interval, 'iteration')))
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'main/loss', 'validation/main/loss', 'main/perp',
'validation/main/perp', 'validation/main/bleu', 'test/main/bleu',
'elapsed_time'
]),
trigger=(args.log_interval, 'iteration'))
if args.validation_source0 and args.validation_source1 and args.validation_target:
valid_data = make_data_tuple(
source0=(source0_ids, args.validation_source0),
source1=(source1_ids, args.validation_source1),
target=(target_ids, args.validation_target))
@chainer.training.make_extension()
def translate(trainer):
source0, source1, target = valid_data[numpy.random.choice(
len(valid_data))]
result = model.translate([model.xp.array(source0)],
[model.xp.array(source1)])[0]
source0_sentence = ' '.join([source0_words[x] for x in source0])
source1_sentence = ' '.join([source1_words[x] for x in source1])
target_sentence = ' '.join([target_words[y] for y in target])
result_sentence = ' '.join([target_words[y] for y in result])
print('# source0 : ' + source0_sentence)
print('# source1 : ' + source1_sentence)
print('# result : ' + result_sentence)
print('# expect : ' + target_sentence)
trainer.extend(translate,
trigger=(args.validation_interval, 'iteration'))
trainer.extend(CalculateBleu(model,
valid_data,
'validation/main/bleu',
device=args.gpu),
trigger=(args.validation_interval, 'iteration'))
dev_iter = chainer.iterators.SerialIterator(valid_data,
args.batchsize,
repeat=False,
shuffle=False)
dev_eval = extensions.Evaluator(dev_iter,
model,
device=args.gpu,
converter=convert)
dev_eval.name = 'valid'
trainer.extend(dev_eval,
trigger=(args.validation_interval, 'iteration'))
if args.test_source0 and args.test_source1 and args.test_target:
test_data = make_data_tuple(source0=(source0_ids, args.test_source0),
source1=(source1_ids, args.test_source1),
target=(target_ids, args.test_target))
trainer.extend(CalculateBleu(model,
test_data,
'test/main/bleu',
device=args.gpu),
trigger=(args.test_interval, 'iteration'))
print('start training')
if args.resume:
# Resume from a snapshot
chainer.serializers.load_npz(args.resume, trainer)
save_args(args, args.out)
trainer.run()
if args.save:
# Save a snapshot
chainer.serializers.save_npz(args.out + "/trainer.npz", trainer)
chainer.serializers.save_npz(args.out + "/model.npz", model)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--outdir',
type=str,
default='results',
help='Directory path to save output files.'
' If it does not exist, it will be created.')
parser.add_argument(
'--env',
type=str,
choices=[
'Pendulum-v0', 'AntBulletEnv-v0', 'HalfCheetahBulletEnv-v0',
'HumanoidBulletEnv-v0', 'HopperBulletEnv-v0',
'Walker2DBulletEnv-v0'
],
help=
'OpenAI Gym env and Pybullet (roboschool) env to perform algorithm on.'
)
parser.add_argument('--num-envs',
type=int,
default=1,
help='Number of envs run in parallel.')
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--gpu',
type=int,
default=0,
help='GPU to use, set to -1 if no GPU.')
parser.add_argument('--load',
type=str,
default='',
help='Directory to load agent from.')
parser.add_argument('--load-demo',
type=str,
default='',
help='Directory to load replay buffer of demo from.')
parser.add_argument(
'--expert-num-episode',
type=int,
default=0,
help='the number of expert trajectory, if 0, no create demo mode.')
parser.add_argument('--absorb',
action='store_true',
help='Add absorb state or do not.')
parser.add_argument('--lamda',
type=float,
default=1,
help='reguralized lambda')
parser.add_argument('--reward-func',
action='store_true',
help='Add absorb state or do not.')
parser.add_argument('--steps',
type=int,
default=10**6,
help='Total number of timesteps to train the agent.')
parser.add_argument('--eval-n-runs',
type=int,
default=10,
help='Number of episodes run for each evaluation.')
parser.add_argument('--eval-interval',
type=int,
default=5000,
help='Interval in timesteps between evaluations.')
parser.add_argument('--replay-start-size',
type=int,
default=10000,
help='Minimum replay buffer size before ' +
'performing gradient updates.')
parser.add_argument('--batch-size',
type=int,
default=100,
help='Minibatch size')
parser.add_argument('--learning-rate',
type=float,
default=1e-3,
help='Learning rate of optimizers')
parser.add_argument('--render',
action='store_true',
help='Render env states in a GUI window.')
parser.add_argument('--demo',
action='store_true',
help='Just run evaluation, not training.')
parser.add_argument('--pretrain',
type=int,
default=0,
help='The number of pretrain iterations')
parser.add_argument('--monitor',
action='store_true',
help='Wrap env with gym.wrappers.Monitor.')
parser.add_argument('--log-interval',
type=int,
default=1000,
help='Interval in timesteps between outputting log'
' messages during training')
parser.add_argument('--logger-level',
type=int,
default=logging.INFO,
help='Level of the root logger.')
parser.add_argument('--policy-output-scale',
type=float,
default=1.,
help='Weight initialization scale of polity output.')
parser.add_argument('--debug', action='store_true', help='Debug mode.')
args = parser.parse_args()
if args.expert_num_episode > 0:
args.absorb = True
logging.basicConfig(level=args.logger_level)
if args.debug:
chainer.set_debug(True)
dir_name = f'{args.env}_{args.seed}'
dir_name += '_demo' if args.demo else ''
dir_name += '_absorb' if args.absorb else ''
dir_name += '_reward' if args.reward_func else ''
# dir_name += f'_{args.lamda}'
args.outdir = experiments.prepare_output_dir(args,
args.outdir,
argv=sys.argv,
time_format=dir_name)
print('Output files are saved in {}'.format(args.outdir))
# Set a random seed used in ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# Set different random seeds for different subprocesses.
# If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
# If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
assert process_seeds.max() < 2**32
def make_env(process_idx, test):
env = gym.make(args.env)
# Unwrap TimiLimit wrapper
assert isinstance(env, gym.wrappers.TimeLimit)
env = env.env
# Use different random seeds for train and test envs
process_seed = int(process_seeds[process_idx])
env_seed = 2**32 - 1 - process_seed if test else process_seed
env.seed(env_seed)
if isinstance(env.observation_space, Box):
# Cast observations to float32 because our model uses float32
env = chainerrl.wrappers.CastObservationToFloat32(env)
else:
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(
args.env, max_frames=None),
episode_life=not test,
clip_rewards=not test)
# if args.absorb:
# if isinstance(env.observation_space, Box):
# env = wrappers.AbsorbingWrapper(env)
# else:
# raise NotImplementedError('Currently, my AbsorobingWrapper is not available for Discrete observation.')
if isinstance(env.action_space, Box):
# Normalize action space to [-1, 1]^n
env = wrappers.NormalizeActionSpace(env)
if args.monitor:
env = gym.wrappers.Monitor(env, args.outdir)
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def make_batch_env(test):
return chainerrl.envs.MultiprocessVectorEnv([
functools.partial(make_env, idx, test)
for idx, env in enumerate(range(args.num_envs))
])
sample_env = make_env(process_idx=0, test=False)
timestep_limit = sample_env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = sample_env.observation_space
action_space = sample_env.action_space
print('Observation space:', obs_space)
print('Action space:', action_space)
if isinstance(obs_space, Box):
head = network_sqil.FCHead()
phi = lambda x: x
else:
head = network_sqil.CNNHead(n_input_channels=4)
phi = lambda x: np.asarray(x, dtype=np.float32) / 255
if isinstance(action_space, Box):
action_size = action_space.low.size
policy = network_sqil.GaussianPolicy(copy.deepcopy(head), action_size)
q_func1 = network_sqil.QSAFunction(copy.deepcopy(head), action_size)
q_func2 = network_sqil.QSAFunction(copy.deepcopy(head), action_size)
def burnin_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(action_space.low,
action_space.high).astype(np.float32)
else:
action_size = action_space.n
policy = network_sqil.SoftmaxPolicy(copy.deepcopy(head), action_size)
q_func1 = network_sqil.QSFunction(copy.deepcopy(head), action_size)
q_func2 = network_sqil.QSFunction(copy.deepcopy(head), action_size)
def burnin_action_func():
return np.random.randint(0, action_size)
policy_optimizer = optimizers.Adam(args.learning_rate).setup(policy)
policy_optimizer.add_hook(GradientClipping(40))
q_func1_optimizer = optimizers.Adam(args.learning_rate).setup(q_func1)
q_func2_optimizer = optimizers.Adam(args.learning_rate).setup(q_func2)
if args.reward_func:
reward_func = network_sqil.FCRewardFunction(action_size)
reward_func_optimizer = optimizers.Adam(
args.learning_rate).setup(reward_func)
else:
reward_func = None
reward_func_optimizer = None
# Draw the computational graph and save it in the output directory.
fake_obs = chainer.Variable(policy.xp.zeros_like(obs_space.low,
dtype=np.float32)[None],
name='observation')
fake_action = chainer.Variable(policy.xp.zeros_like(
action_space.low, dtype=np.float32)[None],
name='action')
fake_absorb = chainer.Variable(policy.xp.zeros_like(
[1], dtype=np.float32)[None],
name='absorb')
chainerrl.misc.draw_computational_graph(
[policy(fake_obs, fake_absorb).sample()],
os.path.join(args.outdir, 'policy'))
chainerrl.misc.draw_computational_graph(
[q_func1(fake_obs, fake_absorb, fake_action)],
os.path.join(args.outdir, 'q_func1'))
chainerrl.misc.draw_computational_graph(
[q_func2(fake_obs, fake_absorb, fake_action)],
os.path.join(args.outdir, 'q_func2'))
if args.absorb:
absorb_state = sample_env.observation_space.sample() * 0
absorb_action = sample_env.action_space.sample() * 0
rbuf = AbsorbReplayBuffer(5 * 10**5, absorb_state, absorb_action)
else:
rbuf = replay_buffer.ReplayBuffer(5 * 10**5)
rbuf_demo = replay_buffer.ReplayBuffer(5 * 10**5)
if len(args.load_demo) > 0:
rbuf_demo.load(os.path.join(args.load_demo, 'replay'))
if args.absorb:
from convert_to_absorb_replay import convert
rbuf_demo = convert(rbuf_demo)
assert isinstance(rbuf_demo, AbsorbReplayBuffer)
else:
assert isinstance(rbuf_demo, replay_buffer.ReplayBuffer)
# Hyperparameters in http://arxiv.org/abs/1802.09477
agent = sqil.SQIL(policy,
q_func1,
q_func2,
reward_func,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
reward_func_optimizer,
rbuf,
rbuf_demo,
gamma=0.99,
is_discrete=isinstance(action_space, Discrete),
replay_start_size=args.replay_start_size,
gpu=args.gpu,
minibatch_size=args.batch_size,
phi=phi,
burnin_action_func=burnin_action_func,
entropy_target=-action_size if isinstance(
action_space, Box) else -np.log(
(1.0 / action_size)) * 0.98,
temperature_optimizer=chainer.optimizers.Adam(3e-4),
lamda=args.lamda)
if args.load:
agent.load(args.load, args.expert_num_episode == 0)
if args.demo:
eval_stats = experiments.eval_performance(
env=make_env(process_idx=0, test=True),
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit,
)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=make_env(process_idx=0, test=False),
eval_env=make_env(process_idx=0, test=True),
outdir=args.outdir,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
# log_interval=args.log_interval,
train_max_episode_len=timestep_limit,
eval_max_episode_len=timestep_limit,
)
def tearDown(self):
chainer.set_debug(self.default_debug)
def main():
logging.basicConfig(
format='%(asctime)s : %(threadName)s : %(levelname)s : %(message)s',
level=logging.INFO)
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', '-g', default=-1, type=int,
help='GPU ID (negative value indicates CPU)')
parser.add_argument('--batchsize', dest='batchsize', type=int,
default=32, help='learning minibatch size')
parser.add_argument('--batchsize_semi', dest='batchsize_semi', type=int,
default=64, help='learning minibatch size')
parser.add_argument('--n_epoch', dest='n_epoch', type=int, default=30,
help='n_epoch')
parser.add_argument('--pretrained_model', dest='pretrained_model',
type=str, default='', help='pretrained_model')
parser.add_argument('--use_unlabled_to_vocab', dest='use_unlabled_to_vocab',
type=int, default=1, help='use_unlabled_to_vocab')
parser.add_argument('--use_rational', dest='use_rational',
type=int, default=0, help='use_rational')
parser.add_argument('--save_name', dest='save_name', type=str,
default='sentiment_model', help='save_name')
parser.add_argument('--n_layers', dest='n_layers', type=int,
default=1, help='n_layers')
parser.add_argument('--alpha', dest='alpha',
type=float, default=0.001, help='alpha')
parser.add_argument('--alpha_decay', dest='alpha_decay',
type=float, default=0.0, help='alpha_decay')
parser.add_argument('--clip', dest='clip',
type=float, default=5.0, help='clip')
parser.add_argument('--debug_mode', dest='debug_mode',
type=int, default=0, help='debug_mode')
parser.add_argument('--use_exp_decay', dest='use_exp_decay',
type=int, default=1, help='use_exp_decay')
parser.add_argument('--load_trained_lstm', dest='load_trained_lstm',
type=str, default='', help='load_trained_lstm')
parser.add_argument('--freeze_word_emb', dest='freeze_word_emb',
type=int, default=0, help='freeze_word_emb')
parser.add_argument('--dropout', dest='dropout',
type=float, default=0.50, help='dropout')
parser.add_argument('--use_adv', dest='use_adv',
type=int, default=0, help='use_adv')
parser.add_argument('--xi_var', dest='xi_var',
type=float, default=1.0, help='xi_var')
parser.add_argument('--xi_var_first', dest='xi_var_first',
type=float, default=1.0, help='xi_var_first')
parser.add_argument('--lower', dest='lower',
type=int, default=1, help='lower')
parser.add_argument('--nl_factor', dest='nl_factor', type=float,
default=1.0, help='nl_factor')
parser.add_argument('--min_count', dest='min_count', type=int,
default=1, help='min_count')
parser.add_argument('--ignore_unk', dest='ignore_unk', type=int,
default=0, help='ignore_unk')
parser.add_argument('--use_semi_data', dest='use_semi_data',
type=int, default=0, help='use_semi_data')
parser.add_argument('--add_labeld_to_unlabel', dest='add_labeld_to_unlabel',
type=int, default=1, help='add_labeld_to_unlabel')
parser.add_argument('--norm_sentence_level', dest='norm_sentence_level',
type=int, default=1, help='norm_sentence_level')
parser.add_argument('--dataset', default='imdb',
choices=['imdb', 'elec', 'rotten', 'dbpedia', 'rcv1'])
parser.add_argument('--eval', dest='eval', type=int, default=0, help='eval')
parser.add_argument('--emb_dim', dest='emb_dim', type=int,
default=256, help='emb_dim')
parser.add_argument('--hidden_dim', dest='hidden_dim', type=int,
default=1024, help='hidden_dim')
parser.add_argument('--hidden_cls_dim', dest='hidden_cls_dim', type=int,
default=30, help='hidden_cls_dim')
parser.add_argument('--adaptive_softmax', dest='adaptive_softmax',
type=int, default=1, help='adaptive_softmax')
parser.add_argument('--random_seed', dest='random_seed', type=int,
default=1234, help='random_seed')
parser.add_argument('--n_class', dest='n_class', type=int,
default=2, help='n_class')
parser.add_argument('--word_only', dest='word_only', type=int,
default=0, help='word_only')
args = parser.parse_args()
batchsize = args.batchsize
batchsize_semi = args.batchsize_semi
print(args)
random.seed(args.random_seed)
np.random.seed(args.random_seed)
os.environ["CHAINER_SEED"] = str(args.random_seed)
os.makedirs("models", exist_ok=True)
if args.debug_mode:
chainer.set_debug(True)
use_unlabled_to_vocab = args.use_unlabled_to_vocab
lower = args.lower == 1
n_char_vocab = 1
n_class = 2
if args.dataset == 'imdb':
vocab_obj, dataset, lm_data, t_vocab = utils.load_dataset_imdb(
include_pretrain=use_unlabled_to_vocab, lower=lower,
min_count=args.min_count, ignore_unk=args.ignore_unk,
use_semi_data=args.use_semi_data,
add_labeld_to_unlabel=args.add_labeld_to_unlabel)
(train_x, train_x_len, train_y,
dev_x, dev_x_len, dev_y,
test_x, test_x_len, test_y) = dataset
vocab, vocab_count = vocab_obj
n_class = 2
if args.use_semi_data:
semi_train_x, semi_train_x_len = lm_data
print('train_vocab_size:', t_vocab)
vocab_inv = dict([(widx, w) for w, widx in vocab.items()])
print('vocab_inv:', len(vocab_inv))
xp = cuda.cupy if args.gpu >= 0 else np
if args.gpu >= 0:
cuda.get_device(args.gpu).use()
xp.random.seed(args.random_seed)
n_vocab = len(vocab)
model = net.uniLSTM_VAT(n_vocab=n_vocab, emb_dim=args.emb_dim,
hidden_dim=args.hidden_dim,
use_dropout=args.dropout, n_layers=args.n_layers,
hidden_classifier=args.hidden_cls_dim,
use_adv=args.use_adv, xi_var=args.xi_var,
n_class=n_class, args=args)
if args.pretrained_model != '':
# load pretrained LM model
pretrain_model = lm_nets.RNNForLM(n_vocab, 1024, args.n_layers, 0.50,
share_embedding=False,
adaptive_softmax=args.adaptive_softmax)
serializers.load_npz(args.pretrained_model, pretrain_model)
pretrain_model.lstm = pretrain_model.rnn
model.set_pretrained_lstm(pretrain_model, word_only=args.word_only)
if args.load_trained_lstm != '':
serializers.load_hdf5(args.load_trained_lstm, model)
if args.gpu >= 0:
model.to_gpu()
def evaluate(x_set, x_length_set, y_set):
chainer.config.train = False
chainer.config.enable_backprop = False
iteration_list = range(0, len(x_set), batchsize)
correct_cnt = 0
total_cnt = 0.0
predicted_np = []
for i_index, index in enumerate(iteration_list):
x = [to_gpu(_x) for _x in x_set[index:index + batchsize]]
x_length = x_length_set[index:index + batchsize]
y = to_gpu(y_set[index:index + batchsize])
output = model(x, x_length)
predict = xp.argmax(output.data, axis=1)
correct_cnt += xp.sum(predict == y)
total_cnt += len(y)
accuracy = (correct_cnt / total_cnt) * 100.0
chainer.config.enable_backprop = True
return accuracy
def get_unlabled(perm_semi, i_index):
index = i_index * batchsize_semi
sample_idx = perm_semi[index:index + batchsize_semi]
x = [to_gpu(semi_train_x[_i]) for _i in sample_idx]
x_length = [semi_train_x_len[_i] for _i in sample_idx]
return x, x_length
base_alpha = args.alpha
opt = optimizers.Adam(alpha=base_alpha)
opt.setup(model)
opt.add_hook(chainer.optimizer.GradientClipping(args.clip))
if args.freeze_word_emb:
model.freeze_word_emb()
prev_dev_accuracy = 0.0
global_step = 0.0
adv_rep_num_statics = {}
adv_rep_pos_statics = {}
if args.eval:
dev_accuracy = evaluate(dev_x, dev_x_len, dev_y)
log_str = ' [dev] accuracy:{}, length:{}'.format(str(dev_accuracy))
logging.info(log_str)
# test
test_accuracy = evaluate(test_x, test_x_len, test_y)
log_str = ' [test] accuracy:{}, length:{}'.format(str(test_accuracy))
logging.info(log_str)
for epoch in range(args.n_epoch):
logging.info('epoch:' + str(epoch))
# train
model.cleargrads()
chainer.config.train = True
iteration_list = range(0, len(train_x), batchsize)
# iteration_list_semi = range(0, len(semi_train_x), batchsize)
perm = np.random.permutation(len(train_x))
if args.use_semi_data:
perm_semi = [np.random.permutation(len(semi_train_x)) for _ in range(2)]
perm_semi = np.concatenate(perm_semi, axis=0)
# print 'perm_semi:', perm_semi.shape
def idx_func(shape):
return xp.arange(shape).astype(xp.int32)
sum_loss = 0.0
sum_loss_z = 0.0
sum_loss_z_sparse = 0.0
sum_loss_label = 0.0
avg_rate = 0.0
avg_rate_num = 0.0
correct_cnt = 0
total_cnt = 0.0
N = len(iteration_list)
is_adv_example_list = []
is_adv_example_disc_list = []
is_adv_example_disc_craft_list = []
y_np = []
predicted_np = []
save_items = []
for i_index, index in enumerate(iteration_list):
global_step += 1.0
model.set_train(True)
sample_idx = perm[index:index + batchsize]
x = [to_gpu(train_x[_i]) for _i in sample_idx]
x_length = [train_x_len[_i] for _i in sample_idx]
y = to_gpu(train_y[sample_idx])
d = None
# Classification loss
output = model(x, x_length)
output_original = output
loss = F.softmax_cross_entropy(output, y, normalize=True)
if args.use_adv or args.use_semi_data:
# Adversarial Training
if args.use_adv:
output = model(x, x_length, first_step=True, d=None)
# Adversarial loss (First step)
loss_adv_first = F.softmax_cross_entropy(output, y, normalize=True)
model.cleargrads()
loss_adv_first.backward()
if args.use_adv:
d = model.d_var.grad
d_data = d.data if isinstance(d, chainer.Variable) else d
output = model(x, x_length, d=d)
# Adversarial loss
loss_adv = F.softmax_cross_entropy(output, y, normalize=True)
loss += loss_adv * args.nl_factor
# Virtual Adversarial Training
if args.use_semi_data:
x, length = get_unlabled(perm_semi, i_index)
output_original = model(x, length)
output_vat = model(x, length, first_step=True, d=None)
loss_vat_first = net.kl_loss(xp, output_original.data, output_vat)
model.cleargrads()
loss_vat_first.backward()
d_vat = model.d_var.grad
output_vat = model(x, length, d=d_vat)
loss_vat = net.kl_loss(xp, output_original.data, output_vat)
loss += loss_vat
predict = xp.argmax(output.data, axis=1)
correct_cnt += xp.sum(predict == y)
total_cnt += len(y)
# update
model.cleargrads()
loss.backward()
opt.update()
if args.alpha_decay > 0.0:
if args.use_exp_decay:
opt.hyperparam.alpha = (base_alpha) * (args.alpha_decay**global_step)
else:
opt.hyperparam.alpha *= args.alpha_decay # 0.9999
sum_loss += loss.data
accuracy = (correct_cnt / total_cnt) * 100.0
logging.info(' [train] sum_loss: {}'.format(sum_loss / N))
logging.info(' [train] apha:{}, global_step:{}'.format(opt.hyperparam.alpha, global_step))
logging.info(' [train] accuracy:{}'.format(accuracy))
model.set_train(False)
# dev
dev_accuracy = evaluate(dev_x, dev_x_len, dev_y)
log_str = ' [dev] accuracy:{}'.format(str(dev_accuracy))
logging.info(log_str)
# test
test_accuracy = evaluate(test_x, test_x_len, test_y)
log_str = ' [test] accuracy:{}'.format(str(test_accuracy))
logging.info(log_str)
last_epoch_flag = args.n_epoch - 1 == epoch
if prev_dev_accuracy < dev_accuracy:
logging.info(' => '.join([str(prev_dev_accuracy), str(dev_accuracy)]))
result_str = 'dev_acc_' + str(dev_accuracy)
result_str += '_test_acc_' + str(test_accuracy)
model_filename = './models/' + '_'.join([args.save_name,
str(epoch), result_str])
serializers.save_hdf5(model_filename + '.model', model)
prev_dev_accuracy = dev_accuracy
def setUp(self):
self.link = links.EmbedID(2, 2)
self.t = numpy.array([self.t_value], dtype=numpy.int32)
self.original_debug = chainer.is_debug()
chainer.set_debug(True)
def setUp(self):
self.default_debug = chainer.is_debug()
chainer.set_debug(True)
def main():
parser = argparse.ArgumentParser(
description='Condtitional WGAN in Chainer')
parser.add_argument('--batchsize',
'-b',
type=int,
default=64,
help='Integer of images in each mini-batch')
parser.add_argument('--epoch',
'-e',
type=int,
default=1,
help='Integer of Epochs')
parser.add_argument('--out',
'-o',
default='result',
help='Directory of output result')
parser.add_argument('--gpu', '-g', type=int, default=-1, help='Gpu number')
parser.add_argument('--resume',
'-r',
default='',
help='start training from snapshot')
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed of z at visualization stage')
parser.add_argument('--imagedir',
'-dir',
default=None,
help="Directory of image dir")
parser.add_argument('--ncritic', '-nc', default=5, help='n_critic')
parser.add_argument('--clamp',
default=0.01,
help='bound of weight clipping in critic')
parser.add_argument('--debug', default=False, help='chainer debug mode')
parser.add_argument('--distribution',
'-dist',
default='uniform',
help='noise z sampling distribution')
args = parser.parse_args()
if args.debug:
chainer.set_debug(True)
# model setup
generator = Generator(n_hidden=110,
label_num=10,
distribution=args.distribution)
critic = Critic(batch_num=args.batchsize, label_num=10)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
generator.to_gpu()
critic.to_gpu()
# optimizer setup
# optimizer is need quite tough decision.
def make_optimizer(model, alpha=0.0002, beta1=0.5):
optimizer = optimizers.Adam(alpha=alpha, beta1=beta1)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(0.0001), 'hook_dec')
return optimizer
def make_optimizer_SGD(model, lr=0.01):
optimizer = optimizers.SGD(lr)
optimizer.setup(model)
return optimizer
def make_optimizer_RMS(model, lr=0.0002):
optimizer = optimizers.RMSprop(lr)
optimizer.setup(model)
return optimizer
opt_gen = make_optimizer_RMS(generator)
opt_critic = make_optimizer_RMS(critic)
# dataset setup
if args.imagedir is None:
# if imagedir not given, use cifar-10
train, _ = chainer.datasets.get_cifar10(withlabel=True, scale=255.)
else:
# TODO: change from ImageDataset to Labeled Dataset.
p = Path(args.imagedir)
img_tuple = {}
label_num = 0
datapathstore = {}
labeldirs = [x for x in p.iterdir() if x.is_dir()]
for label in labeldirs:
img_tuple[label.name] = label_num
datapathstore[label.name] = map(
lambda x: (x, label_num),
[str for str in label.iterdir() if str.is_file()])
datalist = []
for str in datapathstore.items():
datalist = datalist + str
train = chainer.datasets.ImageDataset(datalist)
# train *= 1. / 255.
# *er setup
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
updater = Updater(models=(generator, critic),
iterator=train_iter,
optimizer={
'gen': opt_gen,
'critic': opt_critic
},
n_critic=args.ncritic,
device=args.gpu)
trainer = training.Trainer(updater, (args.epoch, 'epoch'), out=args.out)
# extentions
snapshot_interval = (1000, 'iteration')
display_interval = (100, 'iteration')
trainer.extend(
extensions.snapshot(filename='snapshot_iter_{.updater.iteration}.npz'),
trigger=snapshot_interval)
trainer.extend(extensions.snapshot_object(
generator, 'gen_iter_{.updater.iteration}.npz'),
trigger=snapshot_interval)
trainer.extend(extensions.snapshot_object(
critic, 'critic_iter_{.updater.iteration}.npz'),
trigger=snapshot_interval)
trainer.extend(extensions.LogReport(trigger=display_interval))
trainer.extend(extensions.PrintReport(
['epoch', 'iteration', 'gen_loss', 'critic_loss']),
trigger=display_interval)
trainer.extend(extensions.ProgressBar(update_interval=20))
trainer.extend(out_generated_random_image(generator,
critic,
10,
10,
10,
dst='uniform'),
trigger=snapshot_interval)
# trainer.extend(
# extensions.PlotReport(['gen/loss', 'critic/loss', 'classifier/loss'],
# 'epoch', file_name='plot_{.updater.epoch}.png', grid=True)
# )
# FLY TO THE FUTURE!
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
trainer.run()
def main():
logging.basicConfig(
format='%(asctime)s : %(threadName)s : %(levelname)s : %(message)s',
level=logging.INFO)
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--gpu',
'-g',
default=-1,
type=int,
help='GPU ID (negative value indicates CPU)')
parser.add_argument('--batchsize',
dest='batchsize',
type=int,
default=32,
help='learning minibatch size')
parser.add_argument('--batchsize_semi',
dest='batchsize_semi',
type=int,
default=64,
help='learning minibatch size')
parser.add_argument('--n_epoch',
dest='n_epoch',
type=int,
default=30,
help='n_epoch')
parser.add_argument('--pretrained_model',
dest='pretrained_model',
type=str,
default='',
help='pretrained_model')
parser.add_argument('--use_unlabled_to_vocab',
dest='use_unlabled_to_vocab',
type=int,
default=1,
help='use_unlabled_to_vocab')
parser.add_argument('--use_rational',
dest='use_rational',
type=int,
default=0,
help='use_rational')
parser.add_argument('--save_name',
dest='save_name',
type=str,
default='sentiment_model',
help='save_name')
parser.add_argument('--n_layers',
dest='n_layers',
type=int,
default=1,
help='n_layers')
parser.add_argument('--alpha',
dest='alpha',
type=float,
default=0.001,
help='alpha')
parser.add_argument('--alpha_decay',
dest='alpha_decay',
type=float,
default=0.0,
help='alpha_decay')
parser.add_argument('--clip',
dest='clip',
type=float,
default=5.0,
help='clip')
parser.add_argument('--debug_mode',
dest='debug_mode',
type=int,
default=0,
help='debug_mode')
parser.add_argument('--use_exp_decay',
dest='use_exp_decay',
type=int,
default=1,
help='use_exp_decay')
parser.add_argument('--load_trained_lstm',
dest='load_trained_lstm',
type=str,
default='',
help='load_trained_lstm')
parser.add_argument('--freeze_word_emb',
dest='freeze_word_emb',
type=int,
default=0,
help='freeze_word_emb')
parser.add_argument('--dropout',
dest='dropout',
type=float,
default=0.50,
help='dropout')
parser.add_argument('--use_adv',
dest='use_adv',
type=int,
default=0,
help='use_adv')
parser.add_argument('--xi_var',
dest='xi_var',
type=float,
default=1.0,
help='xi_var')
parser.add_argument('--xi_var_first',
dest='xi_var_first',
type=float,
default=1.0,
help='xi_var_first')
parser.add_argument('--lower',
dest='lower',
type=int,
default=1,
help='lower')
parser.add_argument('--nl_factor',
dest='nl_factor',
type=float,
default=1.0,
help='nl_factor')
parser.add_argument('--min_count',
dest='min_count',
type=int,
default=1,
help='min_count')
parser.add_argument('--ignore_unk',
dest='ignore_unk',
type=int,
default=0,
help='ignore_unk')
parser.add_argument('--use_semi_data',
dest='use_semi_data',
type=int,
default=0,
help='use_semi_data')
parser.add_argument('--add_labeld_to_unlabel',
dest='add_labeld_to_unlabel',
type=int,
default=1,
help='add_labeld_to_unlabel')
parser.add_argument('--norm_sentence_level',
dest='norm_sentence_level',
type=int,
default=1,
help='norm_sentence_level')
parser.add_argument('--dataset',
default='imdb',
choices=['imdb', 'elec', 'rotten', 'dbpedia', 'rcv1'])
parser.add_argument('--eval',
dest='eval',
type=int,
default=0,
help='eval')
parser.add_argument('--emb_dim',
dest='emb_dim',
type=int,
default=256,
help='emb_dim')
parser.add_argument('--hidden_dim',
dest='hidden_dim',
type=int,
default=1024,
help='hidden_dim')
parser.add_argument('--hidden_cls_dim',
dest='hidden_cls_dim',
type=int,
default=30,
help='hidden_cls_dim')
parser.add_argument('--adaptive_softmax',
dest='adaptive_softmax',
type=int,
default=1,
help='adaptive_softmax')
parser.add_argument('--random_seed',
dest='random_seed',
type=int,
default=1234,
help='random_seed')
parser.add_argument('--n_class',
dest='n_class',
type=int,
default=2,
help='n_class')
parser.add_argument('--word_only',
dest='word_only',
type=int,
default=0,
help='word_only')
args = parser.parse_args()
batchsize = args.batchsize
batchsize_semi = args.batchsize_semi
print(args)
random.seed(args.random_seed)
np.random.seed(args.random_seed)
os.environ["CHAINER_SEED"] = str(args.random_seed)
os.makedirs("models", exist_ok=True)
if args.debug_mode:
chainer.set_debug(True)
use_unlabled_to_vocab = args.use_unlabled_to_vocab
lower = args.lower == 1
n_char_vocab = 1
n_class = 2
if args.dataset == 'imdb':
vocab_obj, dataset, lm_data, t_vocab = utils.load_dataset_imdb(
include_pretrain=use_unlabled_to_vocab,
lower=lower,
min_count=args.min_count,
ignore_unk=args.ignore_unk,
use_semi_data=args.use_semi_data,
add_labeld_to_unlabel=args.add_labeld_to_unlabel)
(train_x, train_x_len, train_y, dev_x, dev_x_len, dev_y, test_x,
test_x_len, test_y) = dataset
vocab, vocab_count = vocab_obj
n_class = 2
if args.use_semi_data:
semi_train_x, semi_train_x_len = lm_data
print('train_vocab_size:', t_vocab)
vocab_inv = dict([(widx, w) for w, widx in vocab.items()])
print('vocab_inv:', len(vocab_inv))
xp = cuda.cupy if args.gpu >= 0 else np
if args.gpu >= 0:
cuda.get_device(args.gpu).use()
xp.random.seed(args.random_seed)
n_vocab = len(vocab)
model = net.uniLSTM_VAT(n_vocab=n_vocab,
emb_dim=args.emb_dim,
hidden_dim=args.hidden_dim,
use_dropout=args.dropout,
n_layers=args.n_layers,
hidden_classifier=args.hidden_cls_dim,
use_adv=args.use_adv,
xi_var=args.xi_var,
n_class=n_class,
args=args)
if args.pretrained_model != '':
# load pretrained LM model
pretrain_model = lm_nets.RNNForLM(
n_vocab,
1024,
args.n_layers,
0.50,
share_embedding=False,
adaptive_softmax=args.adaptive_softmax)
serializers.load_npz(args.pretrained_model, pretrain_model)
pretrain_model.lstm = pretrain_model.rnn
model.set_pretrained_lstm(pretrain_model, word_only=args.word_only)
if args.load_trained_lstm != '':
serializers.load_hdf5(args.load_trained_lstm, model)
if args.gpu >= 0:
model.to_gpu()
def evaluate(x_set, x_length_set, y_set):
chainer.config.train = False
chainer.config.enable_backprop = False
iteration_list = range(0, len(x_set), batchsize)
correct_cnt = 0
total_cnt = 0.0
predicted_np = []
for i_index, index in enumerate(iteration_list):
x = [to_gpu(_x) for _x in x_set[index:index + batchsize]]
x_length = x_length_set[index:index + batchsize]
y = to_gpu(y_set[index:index + batchsize])
output = model(x, x_length)
predict = xp.argmax(output.data, axis=1)
correct_cnt += xp.sum(predict == y)
total_cnt += len(y)
accuracy = (correct_cnt / total_cnt) * 100.0
chainer.config.enable_backprop = True
return accuracy
def get_unlabled(perm_semi, i_index):
index = i_index * batchsize_semi
sample_idx = perm_semi[index:index + batchsize_semi]
x = [to_gpu(semi_train_x[_i]) for _i in sample_idx]
x_length = [semi_train_x_len[_i] for _i in sample_idx]
return x, x_length
base_alpha = args.alpha
opt = optimizers.Adam(alpha=base_alpha)
opt.setup(model)
opt.add_hook(chainer.optimizer.GradientClipping(args.clip))
if args.freeze_word_emb:
model.freeze_word_emb()
prev_dev_accuracy = 0.0
global_step = 0.0
adv_rep_num_statics = {}
adv_rep_pos_statics = {}
if args.eval:
dev_accuracy = evaluate(dev_x, dev_x_len, dev_y)
log_str = ' [dev] accuracy:{}, length:{}'.format(str(dev_accuracy))
logging.info(log_str)
# test
test_accuracy = evaluate(test_x, test_x_len, test_y)
log_str = ' [test] accuracy:{}, length:{}'.format(str(test_accuracy))
logging.info(log_str)
for epoch in range(args.n_epoch):
logging.info('epoch:' + str(epoch))
# train
model.cleargrads()
chainer.config.train = True
iteration_list = range(0, len(train_x), batchsize)
# iteration_list_semi = range(0, len(semi_train_x), batchsize)
perm = np.random.permutation(len(train_x))
if args.use_semi_data:
perm_semi = [
np.random.permutation(len(semi_train_x)) for _ in range(2)
]
perm_semi = np.concatenate(perm_semi, axis=0)
# print 'perm_semi:', perm_semi.shape
def idx_func(shape):
return xp.arange(shape).astype(xp.int32)
sum_loss = 0.0
sum_loss_z = 0.0
sum_loss_z_sparse = 0.0
sum_loss_label = 0.0
avg_rate = 0.0
avg_rate_num = 0.0
correct_cnt = 0
total_cnt = 0.0
N = len(iteration_list)
is_adv_example_list = []
is_adv_example_disc_list = []
is_adv_example_disc_craft_list = []
y_np = []
predicted_np = []
save_items = []
for i_index, index in enumerate(iteration_list):
global_step += 1.0
model.set_train(True)
sample_idx = perm[index:index + batchsize]
x = [to_gpu(train_x[_i]) for _i in sample_idx]
x_length = [train_x_len[_i] for _i in sample_idx]
y = to_gpu(train_y[sample_idx])
d = None
# Classification loss
output = model(x, x_length)
output_original = output
loss = F.softmax_cross_entropy(output, y, normalize=True)
if args.use_adv or args.use_semi_data:
# Adversarial Training
if args.use_adv:
output = model(x, x_length, first_step=True, d=None)
# Adversarial loss (First step)
loss_adv_first = F.softmax_cross_entropy(output,
y,
normalize=True)
model.cleargrads()
loss_adv_first.backward()
if args.use_adv:
d = model.d_var.grad
d_data = d.data if isinstance(d,
chainer.Variable) else d
output = model(x, x_length, d=d)
# Adversarial loss
loss_adv = F.softmax_cross_entropy(output,
y,
normalize=True)
loss += loss_adv * args.nl_factor
# Virtual Adversarial Training
if args.use_semi_data:
x, length = get_unlabled(perm_semi, i_index)
output_original = model(x, length)
output_vat = model(x, length, first_step=True, d=None)
loss_vat_first = net.kl_loss(xp, output_original.data,
output_vat)
model.cleargrads()
loss_vat_first.backward()
d_vat = model.d_var.grad
output_vat = model(x, length, d=d_vat)
loss_vat = net.kl_loss(xp, output_original.data,
output_vat)
loss += loss_vat
predict = xp.argmax(output.data, axis=1)
correct_cnt += xp.sum(predict == y)
total_cnt += len(y)
# update
model.cleargrads()
loss.backward()
opt.update()
if args.alpha_decay > 0.0:
if args.use_exp_decay:
opt.hyperparam.alpha = (base_alpha) * (args.alpha_decay**
global_step)
else:
opt.hyperparam.alpha *= args.alpha_decay # 0.9999
sum_loss += loss.data
accuracy = (correct_cnt / total_cnt) * 100.0
logging.info(' [train] sum_loss: {}'.format(sum_loss / N))
logging.info(' [train] apha:{}, global_step:{}'.format(
opt.hyperparam.alpha, global_step))
logging.info(' [train] accuracy:{}'.format(accuracy))
model.set_train(False)
# dev
dev_accuracy = evaluate(dev_x, dev_x_len, dev_y)
log_str = ' [dev] accuracy:{}'.format(str(dev_accuracy))
logging.info(log_str)
# test
test_accuracy = evaluate(test_x, test_x_len, test_y)
log_str = ' [test] accuracy:{}'.format(str(test_accuracy))
logging.info(log_str)
last_epoch_flag = args.n_epoch - 1 == epoch
if prev_dev_accuracy < dev_accuracy:
logging.info(' => '.join(
[str(prev_dev_accuracy),
str(dev_accuracy)]))
result_str = 'dev_acc_' + str(dev_accuracy)
result_str += '_test_acc_' + str(test_accuracy)
model_filename = './models/' + '_'.join(
[args.save_name, str(epoch), result_str])
serializers.save_hdf5(model_filename + '.model', model)
prev_dev_accuracy = dev_accuracy
def setUp(self):
self.x = numpy.random.uniform(-1, 1, (1, 2)).astype(numpy.float32)
self.t = numpy.array([self.t_value], dtype=numpy.int32)
self.original_debug = chainer.is_debug()
chainer.set_debug(True)
def tearDown(self):
chainer.set_debug(self.original_debug)
parser.add_argument('--seed', default=0, type=int, help='random seed')
parser.add_argument('--batchsize', default=128, type=int, help='batchsize')
parser.add_argument('--iteration', default=100000, type=int, help='iteration')
parser.add_argument('--a-dim', default=100, type=int)
parser.add_argument('--z-dim', default=100, type=int)
parser.add_argument('--h-dim', default=500, type=int)
parser.add_argument('--beta', default=1, type=float)
parser.add_argument('--alpha', default=3e-4, type=float)
parser.add_argument('--debug', action='store_true')
parser.add_argument('--verbose', action='store_true')
parser.add_argument('--binarize', action='store_true')
parser.add_argument('--pruning', action='store_true')
args = parser.parse_args()
if args.debug:
chainer.set_debug(True)
np.random.seed(args.seed)
if args.gpu >= 0:
cuda.cupy.random.seed(args.seed)
(x_labeled, y_labeled, x_test, y_test,
x_unlabeled, D, T) = data.load_mnist(pruning=args.pruning)
labeled_data = feeder.DataFeeder((x_labeled, y_labeled))
test_data = feeder.DataFeeder((x_test, y_test))
unlabeled_data = feeder.DataFeeder(x_unlabeled)
N_labeled = len(labeled_data)
N_unlabeled = len(unlabeled_data)
def setUp(self):
self.x = np.array([1], np.float32)
chainer.set_debug(True)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--outdir',
type=str,
default='results',
help='Directory path to save output files.'
' If it does not exist, it will be created.')
parser.add_argument(
'--env',
type=str,
choices=[
'Pendulum-v0', 'AntBulletEnv-v0', 'HalfCheetahBulletEnv-v0',
'HumanoidBulletEnv-v0', 'HopperBulletEnv-v0',
'Walker2DBulletEnv-v0'
],
help=
'OpenAI Gym env and Pybullet (roboschool) env to perform algorithm on.'
)
parser.add_argument('--num-envs',
type=int,
default=1,
help='Number of envs run in parallel.')
parser.add_argument('--seed',
type=int,
default=0,
help='Random seed [0, 2 ** 32)')
parser.add_argument('--gpu',
type=int,
default=0,
help='GPU to use, set to -1 if no GPU.')
parser.add_argument('--load',
type=str,
default='',
help='Directory to load agent from.')
parser.add_argument(
'--expert-num-episode',
type=int,
default=0,
help='the number of expert trajectory, if 0, no create demo mode.')
parser.add_argument('--steps',
type=int,
default=10**6,
help='Total number of timesteps to train the agent.')
parser.add_argument('--eval-n-runs',
type=int,
default=10,
help='Number of episodes run for each evaluation.')
parser.add_argument('--eval-interval',
type=int,
default=5000,
help='Interval in timesteps between evaluations.')
parser.add_argument('--replay-start-size',
type=int,
default=10000,
help='Minimum replay buffer size before ' +
'performing gradient updates.')
parser.add_argument('--batch-size',
type=int,
default=256,
help='Minibatch size')
parser.add_argument('--render',
action='store_true',
help='Render env states in a GUI window.')
parser.add_argument('--demo',
action='store_true',
help='Just run evaluation, not training.')
parser.add_argument('--monitor',
action='store_true',
help='Wrap env with gym.wrappers.Monitor.')
parser.add_argument('--log-interval',
type=int,
default=1000,
help='Interval in timesteps between outputting log'
' messages during training')
parser.add_argument('--logger-level',
type=int,
default=logging.INFO,
help='Level of the root logger.')
parser.add_argument('--policy-output-scale',
type=float,
default=1.,
help='Weight initialization scale of polity output.')
parser.add_argument('--debug', action='store_true', help='Debug mode.')
args = parser.parse_args()
logging.basicConfig(level=args.logger_level)
if args.debug:
chainer.set_debug(True)
if args.expert_num_episode == 0:
args.outdir = experiments.prepare_output_dir(
args,
args.outdir,
argv=sys.argv,
time_format=f'{args.env}_{args.seed}')
else:
args.outdir = experiments.prepare_output_dir(
args,
args.outdir,
argv=sys.argv,
time_format=f'{args.env}_{args.expert_num_episode}expert')
args.replay_start_size = 1e8
print('Output files are saved in {}'.format(args.outdir))
# Set a random seed used in ChainerRL
misc.set_random_seed(args.seed, gpus=(args.gpu, ))
# Set different random seeds for different subprocesses.
# If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
# If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
assert process_seeds.max() < 2**32
def make_env(process_idx, test):
env = gym.make(args.env)
# Unwrap TimiLimit wrapper
assert isinstance(env, gym.wrappers.TimeLimit)
env = env.env
# Use different random seeds for train and test envs
process_seed = int(process_seeds[process_idx])
env_seed = 2**32 - 1 - process_seed if test else process_seed
env.seed(env_seed)
if isinstance(env.observation_space, Box):
# Cast observations to float32 because our model uses float32
env = chainerrl.wrappers.CastObservationToFloat32(env)
else:
env = atari_wrappers.wrap_deepmind(atari_wrappers.make_atari(
args.env, max_frames=None),
episode_life=not test,
clip_rewards=not test)
if isinstance(env.action_space, Box):
# Normalize action space to [-1, 1]^n
env = wrappers.NormalizeActionSpace(env)
if args.monitor:
env = gym.wrappers.Monitor(env, args.outdir)
if args.render:
env = chainerrl.wrappers.Render(env)
return env
def make_batch_env(test):
return chainerrl.envs.MultiprocessVectorEnv([
functools.partial(make_env, idx, test)
for idx, env in enumerate(range(args.num_envs))
])
sample_env = make_env(process_idx=0, test=False)
timestep_limit = sample_env.spec.tags.get(
'wrapper_config.TimeLimit.max_episode_steps')
obs_space = sample_env.observation_space
action_space = sample_env.action_space
print('Observation space:', obs_space)
print('Action space:', action_space)
if isinstance(obs_space, Box):
head = network.FCHead()
phi = lambda x: x
else:
head = network.CNNHead(n_input_channels=4)
phi = lambda x: np.asarray(x, dtype=np.float32) / 255
if isinstance(action_space, Box):
action_size = action_space.low.size
policy = network.GaussianPolicy(copy.deepcopy(head), action_size)
q_func1 = network.QSAFunction(copy.deepcopy(head), action_size)
q_func2 = network.QSAFunction(copy.deepcopy(head), action_size)
def burnin_action_func():
"""Select random actions until model is updated one or more times."""
return np.random.uniform(action_space.low,
action_space.high).astype(np.float32)
else:
action_size = action_space.n
policy = network.SoftmaxPolicy(copy.deepcopy(head), action_size)
q_func1 = network.QSFunction(copy.deepcopy(head), action_size)
q_func2 = network.QSFunction(copy.deepcopy(head), action_size)
def burnin_action_func():
return np.random.randint(0, action_size)
policy_optimizer = optimizers.Adam(3e-4).setup(policy)
q_func1_optimizer = optimizers.Adam(3e-4).setup(q_func1)
q_func2_optimizer = optimizers.Adam(3e-4).setup(q_func2)
# Draw the computational graph and save it in the output directory.
# fake_obs = chainer.Variable(
# policy.xp.zeros_like(obs_space.low, dtype=np.float32)[None],
# name='observation')
# fake_action = chainer.Variable(
# policy.xp.zeros_like(action_space.low, dtype=np.float32)[None],
# name='action')
# chainerrl.misc.draw_computational_graph(
# [policy(fake_obs)], os.path.join(args.outdir, 'policy'))
# chainerrl.misc.draw_computational_graph(
# [q_func1(fake_obs, fake_action)], os.path.join(args.outdir, 'q_func1'))
# chainerrl.misc.draw_computational_graph(
# [q_func2(fake_obs, fake_action)], os.path.join(args.outdir, 'q_func2'))
rbuf = replay_buffer.ReplayBuffer(10**6)
# Hyperparameters in http://arxiv.org/abs/1802.09477
agent = sac.SoftActorCritic(
policy,
q_func1,
q_func2,
policy_optimizer,
q_func1_optimizer,
q_func2_optimizer,
rbuf,
gamma=0.99,
is_discrete=isinstance(action_space, Discrete),
replay_start_size=args.replay_start_size,
gpu=args.gpu,
minibatch_size=args.batch_size,
phi=phi,
burnin_action_func=burnin_action_func,
entropy_target=-action_size if isinstance(action_space, Box) else
-np.log((1.0 / action_size)) * 0.98,
temperature_optimizer=chainer.optimizers.Adam(3e-4),
)
if len(args.load) > 0:
agent.load(args.load, args.expert_num_episode == 0)
if args.demo:
eval_stats = experiments.eval_performance(
env=make_env(process_idx=0, test=True),
agent=agent,
n_steps=None,
n_episodes=args.eval_n_runs,
max_episode_len=timestep_limit,
)
print('n_runs: {} mean: {} median: {} stdev {}'.format(
args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
eval_stats['stdev']))
elif args.expert_num_episode > 0:
episode_r = 0
env = sample_env
episode_len = 0
t = 0
logger = logging.getLogger(__name__)
episode_results = []
try:
for ep in range(args.expert_num_episode):
obs = env.reset()
r = 0
while True:
# a_t
action = agent.act_and_train(obs, r)
# o_{t+1}, r_{t+1}
obs, r, done, info = env.step(action)
t += 1
episode_r += r
episode_len += 1
reset = (episode_len == timestep_limit
or info.get('needs_reset', False))
if done or reset:
agent.stop_episode_and_train(obs, r, done=done)
logger.info('outdir:%s step:%s episode:%s R:%s',
args.outdir, t, ep, episode_r)
episode_results.append(episode_r)
episode_r = 0
episode_len = 0
break
logger.info('mean: %s',
sum(episode_results) / len(episode_results))
except (Exception, KeyboardInterrupt):
raise
# Save
save_name = os.path.join(
os.path.join('demos', f'{args.expert_num_episode}_episode'),
args.env)
makedirs(save_name, exist_ok=True)
agent.replay_buffer.save(os.path.join(save_name, 'replay'))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=make_env(process_idx=0, test=False),
eval_env=make_env(process_idx=0, test=True),
outdir=args.outdir,
steps=args.steps,
eval_n_steps=None,
eval_n_episodes=args.eval_n_runs,
eval_interval=args.eval_interval,
# log_interval=args.log_interval,
train_max_episode_len=timestep_limit,
eval_max_episode_len=timestep_limit,
)
def setUp(self):
self.x = np.array([1], np.float32)
chainer.set_debug(True)
def setUp(self):
self.x = numpy.random.uniform(-1, 1, (2, 2)).astype(numpy.float32)
# `0` is required to avoid NaN
self.t = numpy.array([self.t_value, 0], dtype=numpy.int32)
self.original_debug = chainer.is_debug()
chainer.set_debug(True)
def setUp(self):
self.x = numpy.random.uniform(-1, 1, (2, 2)).astype(numpy.float32)
# `0` is required to avoid NaN
self.t = numpy.array([self.t_value, 0], dtype=numpy.int32)
self.original_debug = chainer.is_debug()
chainer.set_debug(True)
def main(args):
chainer.set_debug(True)
# Initialize the model to train
model = models.archs[args.arch]()
if args.finetune and hasattr(model, 'finetuned_model_path'):
utils.finetuning.load_param(model.finetuned_model_path, model,
args.ignore)
if args.initmodel:
print('Load model from', args.initmodel)
chainer.serializers.load_npz(args.initmodel, model)
if args.gpu >= 0:
chainer.cuda.get_device(args.gpu).use() # Make the GPU current
#if args.test:
#cuda.cudnn_enabled = False
model.to_gpu()
nowt = datetime.datetime.today()
outputdir = os.path.join(args.out, args.arch, 'extract')
if args.initmodel is not None:
outputdir = os.path.dirname(args.initmodel)
if args.indices is None:
args.indices = os.path.join(outputdir, 'features',
'top_' + args.layer + '.txt')
# Load the datasets and mean file
mean = None
if hasattr(model, 'mean_value'):
mean = makeMeanImage(model.mean_value)
else:
mean = np.load(args.mean)
assert mean is not None
if args.indices is None:
args.indices = os.path.join(args.out, args.arch, 'extract',
'top_' + args.layer + '.txt')
#top_path = os.path.join(args.out, args.arch, 'extract', 'top_' + args.layer + '.txt')
#train = ppds.PreprocessedDataset(args.train, args.root, mean, model.insize)
val = ppds.PreprocessedDataset(args.val, args.root, mean, model.insize,
False, args.indices)
# These iterators load the images with subprocesses running in parallel to
# the training/validation.
#train_iter = chainer.iterators.MultiprocessIterator(
# train, args.batchsize, shuffle=False, n_processes=args.loaderjob)
#val_iter = chainer.iterators.MultiprocessIterator(
# val, args.val_batchsize, repeat=False, shuffle=False, n_processes=args.loaderjob)
val_iter = chainer.iterators.SerialIterator(val,
args.val_batchsize,
repeat=False,
shuffle=False)
# Set up an optimizer
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(model)
# Set up a trainer
updater = training.StandardUpdater(val_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (1, 'epoch'), outputdir)
#val_interval = (10 if args.test else int(len(train) / args.batchsize)), 'iteration'
val_interval = (1, 'iteration')
#snapshot_interval = (10, 'iteration') if args.test else (2, 'epoch')
log_interval = (10, 'iteration')
# Copy the chain with shared parameters to flip 'train' flag only in test
eval_model = model.copy()
eval_model.train = False
val_acquirer = utils.DeconvAcquirer(val_iter, eval_model, device=args.gpu)
val_acquirer.mean = mean
val_acquirer.layer_rank = eval_model.layer_rank[args.layer]
val_acquirer.layer_name = args.layer
val_acquirer.operation = args.operation
val_acquirer.fixed_RMS = args.rms
val_acquirer.top = args.top
val_acquirer.n_features = val.cols
if 'googlenet' in args.arch:
val_acquirer.lastname = 'validation/main/loss3'
trainer.extend(val_acquirer, trigger=val_interval)
#trainer.extend(extensions.dump_graph('main/loss'))
#trainer.extend(extensions.snapshot(), trigger=snapshot_interval)
#trainer.extend(extensions.snapshot_object(
# model, 'model_iter_{.updater.iteration}'), trigger=snapshot_interval)
# Be careful to pass the interval directly to LogReport
# (it determines when to emit log rather than when to read observations)
#trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.PrintReport([
'epoch',
'iteration',
'main/loss',
'validation/main/loss',
'main/accuracy',
'validation/main/accuracy',
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
#trainer.extend(extensions.ExponentialShift('lr', args.gamma),
# trigger=(1, 'epoch'))
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
#if not args.test:
# trainer.run()
# chainer.serializers.save_npz(outputdir + '/model', model)
results = val_acquirer(trainer)
results['outputdir'] = outputdir
#if eval_model.layer_rank[args.layer] == 1:
# save_first_conv_filter(os.path.join(outputdir, args.layer),
# model[args.layer].W.data, cols = args.cols, pad = args.pad,
# scale = args.scale, gamma = args.gamma)
#if args.test:
#print(val_acquirer.confmat)
#categories = utils.io.load_categories(args.categories)
#confmat_csv_name = args.initmodel + '.csv'
#confmat_fig_name = args.initmodel + '.eps'
#utils.io.save_confmat_csv(confmat_csv_name, val_acquirer.confmat, categories)
#utils.io.save_confmat_fig(confmat_fig_name, val_acquirer.confmat, categories,
# mode="rate", saveFormat="eps")
return results
def tearDown(self):
chainer.set_debug(self.default_debug)
def setUp(self):
self.x = numpy.random.uniform(-1, 1, (1, 2)).astype(numpy.float32)
self.t = numpy.array([self.t_value], dtype=numpy.int32)
self.original_debug = chainer.is_debug()
chainer.set_debug(True)
def setUp(self):
self.original_debug = chainer.is_debug()
chainer.set_debug(True)
self.one = numpy.array([1], numpy.float32)
self.f = chainer.Function()
def setUp(self):
self.x = numpy.arange(10).reshape((2, 5)).astype('f')
self.ind = numpy.array(self.indices, 'i')
self.debug = chainer.is_debug()
chainer.set_debug(True)
def tearDown(self):
chainer.set_debug(self.original_debug)
def main():
args = parse_args()
if args.debug:
logging.basicConfig(level=logging.DEBUG)
else:
logging.basicConfig(level=logging.INFO)
config = configparser.ConfigParser()
logger.info("read {}".format(args.config_path))
config.read(args.config_path, "UTF-8")
logger.info("setup devices")
if chainer.backends.cuda.available:
devices = setup_devices(config["training_param"]["gpus"])
else:
# cpu run
devices = {"main": -1}
seed = config.getint("training_param", "seed")
logger.info("set random seed {}".format(seed))
set_random_seed(devices, seed)
result = os.path.expanduser(config["result"]["dir"])
destination = os.path.join(result, "pose")
logger.info("> copy code to {}".format(os.path.join(result, "src")))
save_files(result)
logger.info("> copy config file to {}".format(destination))
if not os.path.exists(destination):
os.makedirs(destination)
shutil.copy(args.config_path, os.path.join(destination, "config.ini"))
logger.info("{} chainer debug".format("enable" if args.debug else "disable"))
chainer.set_debug(args.debug)
chainer.global_config.autotune = True
chainer.cuda.set_max_workspace_size(11388608)
chainer.config.cudnn_fast_batch_normalization = True
logger.info("> get dataset")
train_set, val_set, hand_param = select_dataset(config, return_data=["train_set", "val_set", "hand_param"])
model = select_model(config, hand_param)
logger.info("> transform dataset")
train_set = TransformDataset(train_set, model.encode)
val_set = TransformDataset(val_set, model.encode)
logger.info("> size of train_set is {}".format(len(train_set)))
logger.info("> size of val_set is {}".format(len(val_set)))
logger.info("> create iterators")
batch_size = config.getint("training_param", "batch_size")
n_processes = config.getint("training_param", "n_processes")
train_iter = chainer.iterators.MultiprocessIterator(
train_set, batch_size,
n_processes=n_processes
)
test_iter = chainer.iterators.MultiprocessIterator(
val_set, batch_size,
repeat=False, shuffle=False,
n_processes=n_processes,
)
logger.info("> setup optimizer")
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(0.0005))
logger.info("> setup parallel updater devices={}".format(devices))
updater = training.updaters.ParallelUpdater(train_iter, optimizer, devices=devices)
logger.info("> setup trainer")
trainer = training.Trainer(
updater,
(config.getint("training_param", "train_iter"), "iteration"),
destination,
)
logger.info("> setup extensions")
trainer.extend(
extensions.LinearShift("lr",
value_range=(config.getfloat("training_param", "learning_rate"), 0),
time_range=(0, config.getint("training_param", "train_iter"))
),
trigger=(1, "iteration")
)
trainer.extend(extensions.Evaluator(test_iter, model, device=devices["main"]))
if extensions.PlotReport.available():
trainer.extend(extensions.PlotReport([
"main/loss", "validation/main/loss",
], "epoch", file_name="loss.png"))
trainer.extend(extensions.LogReport())
trainer.extend(extensions.observe_lr())
trainer.extend(extensions.PrintReport([
"epoch", "elapsed_time", "lr",
"main/loss", "validation/main/loss",
"main/loss_resp", "validation/main/loss_resp",
"main/loss_iou", "validation/main/loss_iou",
"main/loss_coor", "validation/main/loss_coor",
"main/loss_size", "validation/main/loss_size",
"main/loss_limb", "validation/main/loss_limb",
"main/loss_vect_cos", "validation/main/loss_vect_cos",
"main/loss_vect_norm", "validation/main/loss_vect_cos",
"main/loss_vect_square", "validation/main/loss_vect_square",
]))
trainer.extend(extensions.ProgressBar())
trainer.extend(extensions.snapshot(filename="best_snapshot"),
trigger=training.triggers.MinValueTrigger("validation/main/loss"))
trainer.extend(extensions.snapshot_object(model, filename="bestmodel.npz"),
trigger=training.triggers.MinValueTrigger("validation/main/loss"))
logger.info("> start training")
trainer.run()