def main():
if os.path.exists(model_path):
model = keras.models.load_model(model_path)
else:
model = train_model()
graph = KerasConverter(batch_size=1).convert(model)
generate_descriptor('webgl', graph).save('./output')
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument("--model", default="fc", choices=["fc", "conv"])
parser.add_argument(
'--out',
'-o',
default='output_chainer',
help='Directory to output the graph descriptor and sample test data')
parser.add_argument("--backend",
default="webgpu,webgl,webassembly,fallback")
args = parser.parse_args()
output_dir = os.path.join(args.out, f"./chainer_model")
os.makedirs(output_dir, exist_ok=True)
model, test, graph = generate_graph(args.model, output_dir)
for backend in args.backend.split(","):
exec_info = generate_descriptor(backend, graph)
exec_info.save(args.out)
test_samples_json = []
for i in range(10):
image, label = test[i]
test_samples_json.append({
'x': image.flatten().tolist(),
'y': int(label)
})
with open(os.path.join(args.out, 'test_samples.json'), 'w') as f:
json.dump(test_samples_json, f)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--backend",
default="webgpu,webgl,webassembly,fallback")
parser.add_argument("--encoding")
parser.add_argument('--out',
'-o',
default='output_chainer',
help='Directory to output the graph descriptor')
_, _, _, graph = generate_graph()
args = parser.parse_args()
os.makedirs(args.out, exist_ok=True)
any_backend_failed = False
last_backend_exception = None
for backend in args.backend.split(","):
try:
graph_exec_data = generate_descriptor(
backend, graph, constant_encoder_name=args.encoding)
graph_exec_data.save(args.out)
except Exception as ex:
any_backend_failed = True
last_backend_exception = ex
console.error(
f"Failed generating descriptor for backend {backend}: {str(ex)}\n"
)
if any_backend_failed:
raise last_backend_exception
def main():
sys.setrecursionlimit(10000) # workaround for deep copying large graph
parser = argparse.ArgumentParser()
parser.add_argument("--model",
default="resnet50",
choices=["vgg16", "resnet50"])
parser.add_argument("--backend", default="webgpu,webassembly,fallback")
parser.add_argument("--encoding")
parser.add_argument('--out',
'-o',
default='output_chainer',
help='Directory to output the graph descriptor')
args = parser.parse_args()
os.makedirs(args.out, exist_ok=True)
sample_image = np.zeros((224, 224, 3),
dtype=np.uint8) # PIL.Image.open("")
if args.model == "vgg16":
link = chainer.links.model.vision.vgg.VGG16Layers()
prepared_image = chainer.links.model.vision.vgg.prepare(
sample_image) # BGR, CHW
elif args.model == "resnet50":
link = chainer.links.model.vision.resnet.ResNet50Layers()
prepared_image = chainer.links.model.vision.resnet.prepare(
sample_image)
else:
raise NotImplementedError
nn_input = chainer.Variable(np.array([prepared_image], dtype=np.float32))
if chainer.__version__ >= "2.":
with chainer.using_config('train', False):
nn_output = link(nn_input, layers=['prob'])['prob']
else:
nn_output = link(nn_input, layers=['prob'], test=True)['prob']
graph = ChainerConverter().convert([nn_input], [nn_output]) # type: Graph
any_backend_failed = False
last_backend_exception = None
for backend in args.backend.split(","):
try:
graph_exec_data = generate_descriptor(
backend, graph, constant_encoder_name=args.encoding)
graph_exec_data.save(args.out)
except Exception as ex:
any_backend_failed = True
last_backend_exception = ex
console.error(
f"Failed generating descriptor for backend {backend}: {str(ex)}\n"
)
if any_backend_failed:
raise last_backend_exception
def main():
sys.setrecursionlimit(10000)
parser = argparse.ArgumentParser()
parser.add_argument("--model", default="resnet50", choices=["resnet50"])
parser.add_argument('--out',
'-o',
default='output_tensorflow',
help='Directory to output the graph descriptor')
parser.add_argument("--encoding", help="name of weight encoder")
parser.add_argument("--backend",
default="webgpu,webgl,webassembly,fallback",
help="backend")
args = parser.parse_args()
os.makedirs(args.out, exist_ok=True)
slim_dir = os.path.join(args.out, "models/slim")
if not os.path.exists(slim_dir):
clone_slim(args.out)
model_path = download_model(args.out)
sys.path.append(slim_dir)
from nets import resnet_v1
image_size = resnet_v1.resnet_v1.default_image_size
checkpoints_dir = args.out
sess = tf.Session()
processed_images = tf.placeholder(tf.float32,
[1, image_size, image_size, 3])
# Create the model, use the default arg scope to configure the batch norm parameters.
with slim.arg_scope(resnet_v1.resnet_arg_scope()):
logits, _ = resnet_v1.resnet_v1_50(processed_images,
num_classes=1000,
is_training=False)
probabilities = tf.nn.softmax(logits)
init_fn = slim.assign_from_checkpoint_fn(model_path,
slim.get_model_variables())
init_fn(sess)
graph = TensorFlowConverter(sess, batch_size=1).convert([processed_images],
[probabilities])
from webdnn.graph import traverse
traverse.dump(graph)
for backend in args.backend.split(","):
graph_exec_data = generate_descriptor(
backend, graph, constant_encoder_name=args.encoding)
graph_exec_data.save(args.out)
console.stderr("Done.")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--model", default="fc", choices=["fc", "conv"])
parser.add_argument("--out", default="output_tensorflow")
parser.add_argument("--backend", default=",".join(backend_names))
args = parser.parse_args()
_, _, _, graph = generate_graph(args.model, args.out)
for backend in args.backend.split(","):
desc = generate_descriptor(backend, graph)
desc.save(args.out)
def main():
args = arg()
model = AutoEncoder(F.mean_squared_error)
chainer.serializers.load_npz(args.model, model)
example_input = np.zeros((1, 1, 28, 28)).astype("f")
x = chainer.Variable(example_input)
y = model.predict(x)
graph = ChainerConverter().convert([x], [y])
for backend in ["webgpu", "webassembly"]:
exec_info = generate_descriptor(backend, graph)
exec_info.save(args.out)
def main():
model_descriptions = []
for model_config in model_configs:
model = load_model(model_config)
graph = get_graph(model)
exec_info = generate_descriptor("webassembly", graph, constant_encoder_name="eightbit")
output_dir = f"../docs/webdnn_model/{model_config['name']}"
exec_info.save(output_dir)
data_size = os.path.getsize(f"{output_dir}/weight_webassembly.bin") + 65536 # 重み以外の関連ファイルサイズ加算
model_descriptions.append({"name": model_config["name"],
"description": f"{model_config['name']} {data_size // 1024}kB"})
export_model_metadata(model_descriptions, "../src/models.ts")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--model", default="fc", choices=["fc", "conv"])
parser.add_argument("--out", default="output_tensorflow")
parser.add_argument("--backends",
action="append",
default=["webgpu", "webgl", "webassembly", "fallback"])
args = parser.parse_args()
_, _, _, graph = generate_graph(args.model, args.out)
for backend in args.backends:
desc = generate_descriptor(backend, graph)
desc.save(args.out)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--out', '-o', default='output_keras', help='Directory to output the graph descriptor')
parser.add_argument("--encoding", help="name of weight encoder")
parser.add_argument("--backend", default="webgpu,webgl,webassembly,fallback", help="backend")
args = parser.parse_args()
_, graph = generate_graph()
for backend in args.backend.split(","):
graph_exec_data = generate_descriptor(backend, graph, constant_encoder_name=args.encoding)
graph_exec_data.save(args.out)
console.stderr("Done.")
def export_feature_extractor(output_dir: str):
# model = chainercv.links.model.resnet.ResNet50(pretrained_model="imagenet", arch="he")
# model.pick = "pool5" # =>2048dim
model = SqueezeNetFeatureExtactor()
with chainer.using_config("train", False):
with chainer.using_config("enable_backprop", True):
nn_input = chainer.Variable(
np.zeros((1, 3, 227, 227), dtype=np.float32))
nn_output = model(nn_input)
graph = ChainerConverter().convert([nn_input], [nn_output])
for backend in ["webgpu", "webgl", "webassembly"]:
graph_exec_data = generate_descriptor(backend,
graph,
constant_encoder_name="eightbit")
graph_exec_data.save(output_dir)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--model",
default="fc",
choices=["fc", "conv", "dilated_conv", "residual", "complex"])
parser.add_argument("--out", default="output_keras")
parser.add_argument("--backend", default=",".join(backend_names))
args = parser.parse_args()
model, graph = generate_graph(args.model, args.out)
for backend in args.backend.split(","):
exec_info = generate_descriptor(backend, graph)
exec_info.save(args.out)
def main():
z_dim = 100
device = -1 #CPU
batch_size = 1
model = Generator(z_dim)
model.to_gpu()
chainer.serializers.load_npz('result-dcgan/gen_snapshot_epoch-200.npz',
model)
model.to_cpu()
x, _ = model.generate_noise(device, batch_size)
y = model(x)
graph = ChainerConverter().convert([x], [y])
exec_info = generate_descriptor("webassembly", graph)
exec_info.save("./model")
def main():
# construct model object and load weights
model = chainer.links.Classifier(CNN())
chainer.serializers.load_npz('chainer_output/chainer_model.npz', model)
# run model with dummy variable
input_variable = chainer.Variable(np.zeros((1, 1, 28, 28), dtype=np.float32))
prediction_raw_variable = model.predictor(input_variable) # raw activation before softmax
prediction_with_softmax_variable = chainer.functions.softmax(prediction_raw_variable)
# convert graph to intermediate representation
graph = ChainerConverter().convert([input_variable], [
prediction_with_softmax_variable])
# generate graph descriptor
backend = 'webgl'
exec_info = generate_descriptor(backend, graph)
exec_info.save('webdnn_graph_descriptor')
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--model", default="resnet50", choices=["resnet50"])
parser.add_argument('--out', '-o', default='output_keras',
help='Directory to output the graph descriptor')
parser.add_argument("--encoding", help="name of weight encoder")
parser.add_argument("--backend", default="webgpu,webassembly,fallback", help="backend")
args = parser.parse_args()
model = resnet50.ResNet50(include_top=True, weights='imagenet')
sys.setrecursionlimit(10000)
graph = KerasConverter(batch_size=1).convert(model)
for backend in args.backend.split(","):
graph_exec_data = generate_descriptor(backend, graph, constant_encoder_name=args.encoding)
graph_exec_data.save(args.out)
console.stderr("Done.")
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--backend',
choices=('webgl', 'webassembly'),
nargs='+')
parser.add_argument('--out', default='model')
args = parser.parse_args()
model = SSD(chainercv.links.SSD300(pretrained_model='voc0712'))
x = chainer.Variable(
np.empty((1, 3, model.insize, model.insize), dtype=np.float32))
ys = model(x)
print(x.shape, '->', ', '.join('{}'.format(y.shape) for y in ys))
graph = ChainerConverter().convert([x], ys)
for backend in args.backend:
print('backend:', backend)
desc = generate_descriptor(backend, graph)
desc.save(args.out)
def main():
parser = argparse.ArgumentParser(
description='chainer implementation of pix2pix')
parser.add_argument('--batchsize',
'-b',
type=int,
default=1,
help='Number of images in each mini-batch')
parser.add_argument('--epoch',
'-e',
type=int,
default=200,
help='Number of sweeps over the dataset to train')
parser.add_argument('--gpu',
'-g',
type=int,
default=-1,
help='GPU ID (negative value indicates CPU)')
parser.add_argument('--out',
'-o',
default='result',
help='Directory to output the result')
parser.add_argument('--enc-npz', default='', required=True)
parser.add_argument('--dec-npz', default='', required=True)
args = parser.parse_args()
# Set up a neural network to train
enc = Encoder(in_ch=3)
dec = Decoder(out_ch=3)
chainer.serializers.load_npz(args.enc_npz, enc)
chainer.serializers.load_npz(args.dec_npz, dec)
# graph
input = np.zeros((1, 3, 256, 256), dtype=np.float32)
x = chainer.Variable(input)
z = enc(x)
y = dec(z)
graph = ChainerConverter().convert([x], [y])
exec_info = generate_descriptor("webassembly", graph)
exec_info.save(args.out)
def main():
FLAGS(sys.argv)
# 0. load dataset
char_list = [
line.strip().split('\t')[0] for line in open(FLAGS.vocab_file)
]
charset_size = len(char_list) + 1
# 1. build model
assert FLAGS.model in ('rnnlm')
if FLAGS.model == 'rnnlm':
model = Decoder(charset_size=charset_size,
hidden_size=FLAGS.hidden_size,
n_layers=FLAGS.n_layers,
dropout=FLAGS.dropout)
ins, outs = model.webdnn_anchor()
graph = ChainerConverter().convert(ins, outs)
exec_info = generate_descriptor("webgpu", graph)
exec_info.save("./output")
def main():
sys.setrecursionlimit(10000) # workaround for deep copying large graph
parser = argparse.ArgumentParser()
# default is Caffenet of Caffe example
parser.add_argument("caffemodel")
parser.add_argument("--backend", default="webgpu,webassembly,fallback",
help="comma-separated list of backends")
parser.add_argument("--input_name",
help="blob name for input (mandatory)")
parser.add_argument("--input_shape",
help="shape of blobs for inputs (example: '(1,3,224,224)')")
parser.add_argument("--input_npy",
help="npy file containing sample inputs")
parser.add_argument("--output_names", required=True,
help="comma-separated blob name for output (mandatory)")
parser.add_argument("--out",
help="output directory (default: <model>/webdnn_graph_descriptor)")
parser.add_argument("--encoding", help="name of weight encoder")
args = parser.parse_args()
# multiple blob input can be easily implemented, but command-line arguments becomes complicated.
input_blob, input_filled = parse_input_blob(args)
output_names = args.output_names.split(",")
console.stderr("[convert_caffe] Loading caffe model... (usually takes several minutes)")
link = chainer.links.caffe.CaffeFunction(args.caffemodel)
console.stderr("[convert_caffe] Generating feedforward graph")
if chainer.__version__ >= "2.":
chainer.using_config("train", False)
output_blobs = list(
link(inputs={args.input_name: input_blob}, outputs=output_names)) # list of Variable
else:
output_blobs = list(
link(inputs={args.input_name: input_blob}, outputs=output_names, train=False)) # list of Variable
chainer_cg = chainer.computational_graph.build_computational_graph(output_blobs)
converter = ChainerConverter()
graph = converter.convert(chainer_cg, [input_blob], output_blobs) # type: Graph
if args.out:
output_dir = args.out
else:
output_dir = path.join(path.dirname(args.caffemodel), "webdnn_graph_descriptor")
os.makedirs(output_dir, exist_ok=True)
if input_filled:
# save output of Caffe Network (not required for inference)
output_arrays = {output_name: output_blob.data for output_name, output_blob in zip(output_names, output_blobs)}
np.savez(path.join(output_dir, "example_output.npz"), **output_arrays)
console.stderr("[convert_caffe] Generating descriptors")
any_backend_failed = False
for backend in args.backend.split(","):
try:
graph_exec_data = generate_descriptor(backend, graph, constant_encoder_name=args.encoding)
graph_exec_data.save(output_dir)
except Exception as ex:
any_backend_failed = True
console.error(f"[convert_caffe] Failed generating descriptor for backend {backend}: {str(ex)}")
if any_backend_failed:
sys.exit(1)
def save_generated_image(image, name):
Imag = combine_images(image)
save_images(Imag, name)
z_size = 128
noise = np.random.normal(0, 0.5, [1, z_size])
#model_set
g = generator(512, 512, z_size)
serializers.load_npz("generator.model", g)
x = chainer.Variable(np.zeros((1, z_size), dtype=np.float32))
y = g(x, np.zeros(NUMBER_OF_TAG), 5, 1)
noise = np.random.normal(0, 0.5, [1, z_size]).astype(np.float32)
image = g(noise, np.zeros(NUMBER_OF_TAG), 5, 1)
image = image.data[0]
image = image.transpose(1, 2, 0)
save_images((image * 127.5) + 127.5, "test")
from webdnn.frontend.chainer import ChainerConverter
graph = ChainerConverter().convert([x], [y])
from webdnn.backend import generate_descriptor
exec_info = generate_descriptor(
"webgpu", graph) # also "webassembly", "webgl", "fallback" are available.
exec_info.save("./output")
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument("--model", default="mlp", choices=["mlp", "conv"])
parser.add_argument('--batchsize',
'-b',
type=int,
default=100,
help='Number of images in each mini-batch')
parser.add_argument('--epoch',
'-e',
type=int,
default=5,
help='Number of sweeps over the dataset to train')
parser.add_argument('--frequency',
'-f',
type=int,
default=-1,
help='Frequency of taking a snapshot')
parser.add_argument('--gpu',
'-g',
type=int,
default=-1,
help='GPU ID (negative value indicates CPU)')
parser.add_argument(
'--out',
'-o',
default='output_chainer',
help='Directory to output the graph descriptor and sample test data')
parser.add_argument('--resume',
'-r',
default='',
help='Resume the training from snapshot')
args = parser.parse_args()
print('GPU: {}'.format(args.gpu))
print('# Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('')
os.makedirs(args.out, exist_ok=True)
# Set up a neural network to train
# Classifier reports softmax cross entropy loss and accuracy at every
# iteration, which will be used by the PrintReport extension below.
model = L.Classifier(models[args.model](10))
if args.gpu >= 0:
# Make a specified GPU current
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu() # Copy the model to the GPU
# Setup an optimizer
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
# Load the MNIST dataset
train, test = chainer.datasets.get_mnist(ndim=3)
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test_iter = chainer.iterators.SerialIterator(test,
args.batchsize,
repeat=False,
shuffle=False)
# Set up a trainer
updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (args.epoch, 'epoch'),
out=os.path.join(args.out, 'chainer_model'))
# Evaluate the model with the test dataset for each epoch
trainer.extend(extensions.Evaluator(test_iter, model, device=args.gpu))
# Dump a computational graph from 'loss' variable at the first iteration
# The "main" refers to the target link of the "main" optimizer.
trainer.extend(extensions.dump_graph('main/loss'))
# Take a snapshot for each specified epoch
frequency = args.epoch if args.frequency == -1 else max(1, args.frequency)
trainer.extend(extensions.snapshot(), trigger=(frequency, 'epoch'))
# Write a log of evaluation statistics for each epoch
trainer.extend(extensions.LogReport())
# Save two plot images to the result dir
if extensions.PlotReport.available():
trainer.extend(
extensions.PlotReport(['main/loss', 'validation/main/loss'],
'epoch',
file_name='loss.png'))
trainer.extend(
extensions.PlotReport(
['main/accuracy', 'validation/main/accuracy'],
'epoch',
file_name='accuracy.png'))
# Print selected entries of the log to stdout
# Here "main" refers to the target link of the "main" optimizer again, and
# "validation" refers to the default name of the Evaluator extension.
# Entries other than 'epoch' are reported by the Classifier link, called by
# either the updater or the evaluator.
trainer.extend(
extensions.PrintReport([
'epoch', 'main/loss', 'validation/main/loss', 'main/accuracy',
'validation/main/accuracy', 'elapsed_time'
]))
# Print a progress bar to stdout
trainer.extend(extensions.ProgressBar())
if args.resume:
# Resume from a snapshot
chainer.serializers.load_npz(args.resume, trainer)
# Run the training
trainer.run()
# conversion
print('Transpiling model to WebDNN graph descriptor')
example_input = numpy.expand_dims(
train[0][0], axis=0) # example input (anything ok, (batch_size, 784))
x = chainer.Variable(example_input)
y = F.softmax(model.predictor(x)) # run model
graph = ChainerConverter().convert_from_inout_vars(
[x], [y]) # convert graph to intermediate representation
for backend in ["webgpu", "webassembly", "fallback"]:
try:
exec_info = generate_descriptor(backend, graph)
exec_info.save(args.out)
except Exception as ex:
print(
f"Failed generating descriptor for backend {backend}: {str(ex)}\n"
)
else:
print(f"Backend {backend} ok\n")
print('Exporting test samples (for demo purpose)')
test_samples_json = []
for i in range(10):
image, label = test[i]
test_samples_json.append({
'x': image.flatten().tolist(),
'y': int(label)
})
with open(os.path.join(args.out, 'test_samples.json'), 'w') as f:
json.dump(test_samples_json, f)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--out", default="output_tensorflow")
parser.add_argument("--backends",
action="append",
default=["webgpu", "webassembly", "fallback"])
args = parser.parse_args()
session_path = os.path.join(args.out, "session")
sample_path = os.path.join(args.out, "test_samples.json")
data_path = os.path.join(args.out, "data")
x = tf.placeholder(tf.float32, [None, 784])
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
y = tf.nn.softmax(tf.matmul(x, W) + b)
t = tf.placeholder(tf.float32, [None, 10])
loss = tf.reduce_mean(-tf.reduce_sum(t * tf.log(y), reduction_indices=[1]))
accuracy = tf.reduce_mean(
tf.cast(tf.equal(tf.argmax(y, 1), tf.argmax(t, 1)), tf.float32))
optimizer = tf.train.GradientDescentOptimizer(0.05).minimize(loss)
sess = tf.Session()
if os.path.exists(session_path):
# -------------------------------------------------------------------------------
# Load pretrained model
saver = tf.train.Saver()
saver.restore(sess, os.path.join(session_path, "session"))
else:
# -------------------------------------------------------------------------------
# Train model
mnist = input_data.read_data_sets(data_path, one_hot=True)
tf.global_variables_initializer().run()
for step in range(1000):
batch_xs, batch_ys = mnist.train.next_batch(100)
_, loss_val = sess.run([optimizer, loss],
feed_dict={
x: batch_xs,
t: batch_ys
})
if step % 100 == 0:
print(f"Step {step}: loss = {loss_val}")
print(
f"accuracy: {sess.run(accuracy, feed_dict={x: mnist.test.images, t: mnist.test.labels})}"
)
saver = tf.train.Saver()
saver.save(sess, os.path.join(session_path, "session"))
with open(sample_path, "w") as f:
json.dump([{
"x": mnist.test.images[i].flatten().tolist(),
"y": int(mnist.test.labels[i].argmax())
} for i in range(10)], f)
# -------------------------------------------------------------------------------
# Convert
webdnn_graph = TensorFlowConverter(sess, batch_size=1).convert([x], [y])
# # When you try to convert more complex model, maybe WebDNN failed to infer the data order.
# # In this case, you can give "data-order hints" to WebDNN graph converter.
#
# webdnn_graph = TensorFlowConverter(sess, batch_size=1).convert([x], [y], order_hints={
# x: OrderNC,
# W: OrderCN,
# b: OrderC,
# y: OrderNC
# })
for backend in args.backends:
desc = generate_descriptor(backend, webdnn_graph)
desc.save(args.out)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--model", default="fc", choices=["fc", "conv"])
parser.add_argument("--out", default="output_tensorflow")
parser.add_argument("--backends",
action="append",
default=["webgpu", "webgl", "webassembly", "fallback"])
args = parser.parse_args()
session_path = os.path.join(args.out, "session")
sample_path = os.path.join(args.out, "test_samples.json")
data_path = os.path.join(args.out, "data")
x, y, t, loss, accuracy, optimizer = setup_model(args.model)
sess = tf.Session()
if os.path.exists(session_path):
# -------------------------------------------------------------------------------
# Load pretrained model
saver = tf.train.Saver()
saver.restore(sess, os.path.join(session_path,
f"session_{args.model}"))
else:
# -------------------------------------------------------------------------------
# Train model
mnist = input_data.read_data_sets(data_path, one_hot=True)
with sess.as_default():
tf.global_variables_initializer().run()
for step in range(1000):
batch_xs, batch_ys = mnist.train.next_batch(100)
_, loss_val = sess.run([optimizer, loss],
feed_dict={
x: batch_xs,
t: batch_ys
})
if step % 100 == 0:
print(f"Step {step}: loss = {loss_val}")
print(
f"accuracy: {sess.run(accuracy, feed_dict={x: mnist.test.images, t: mnist.test.labels})}"
)
saver = tf.train.Saver()
saver.save(sess, os.path.join(session_path,
f"session_{args.model}"))
with open(sample_path, "w") as f:
json.dump([{
"x": mnist.test.images[i].flatten().tolist(),
"y": int(mnist.test.labels[i].argmax())
} for i in range(10)], f)
# -------------------------------------------------------------------------------
# Convert
webdnn_graph = TensorFlowConverter(sess, batch_size=1).convert([x], [y])
# # When you try to convert more complex model, maybe WebDNN failed to infer the data order.
# # In this case, you can give "data-order hints" to WebDNN graph converter.
#
# webdnn_graph = TensorFlowConverter(sess, batch_size=1).convert([x], [y], order_hints={
# x: OrderNC,
# W: OrderCN,
# b: OrderC,
# y: OrderNC
# })
for backend in args.backends:
desc = generate_descriptor(backend, webdnn_graph)
desc.save(args.out)
def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument(
'--out',
'-o',
default='output_pytorch',
help='Directory to output the graph descriptor and sample test data')
parser.add_argument("--backend",
default="webgpu,webgl,webassembly,fallback")
args = parser.parse_args()
training_dir = os.path.join(args.out, "pytorch_model")
os.makedirs(training_dir, exist_ok=True)
model_path = os.path.join(training_dir, "model.proto")
if not os.path.exists(model_path):
# model training part (as usual)
torch.manual_seed(1)
device = torch.device("cpu")
train_loader = torch.utils.data.DataLoader(
datasets.MNIST(
os.path.join(args.out, 'data'),
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
# If input is normalized, input for WebDNN also have to be normalized with same parameter.
# Default of datasets.MNIST is 0=black, 1=white
# transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=64,
shuffle=True)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST(
os.path.join(args.out, 'data'),
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
# transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=1000,
shuffle=True)
model = Net().to(device)
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)
for epoch in range(1, 4):
train(args, model, device, train_loader, optimizer, epoch)
test(args, model, device, test_loader)
# export model as ONNX format
dummy_input = torch.autograd.Variable(torch.randn(1, 1, 28, 28))
torch.onnx.export(model, dummy_input, model_path, verbose=True)
# model conversion using WebDNN
onnx_model = onnx.load(model_path)
graph = ONNXConverter().convert(onnx_model)
for backend in args.backend.split(","):
exec_info = generate_descriptor(backend, graph)
exec_info.save(args.out)
# test data for demo
output_test_samples(
datasets.MNIST(
os.path.join(args.out, 'data'),
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
# transforms.Normalize((0.1307,), (0.3081,))
])),
os.path.join(args.out, "test_samples.json"))
print(f"model: {args.model}")
print(f"backend: {args.backend}")
print(f"encoding: {args.encoding}")
# Load chainer pre-trained model
model = FastStyleNet()
model_path = NSTModelPath[args.model].value
if not path.exists(model_path):
raise FileNotFoundError(f"Model data ({model_path}) is not found. Please clone " +
"'https://github.com/gafr/chainer-fast-neuralstyle-models' under the resource directory. " +
"Clone command takes about a few minute, the repository size is about 200MB.")
chainer.serializers.load_npz(model_path, model)
# Execute forward propagation to construct computation graph
if chainer.__version__ >= "2.":
with chainer.using_config("train", False): # fixes batch normalization
x = chainer.Variable(np.zeros((1, 3, 144, 192), dtype=np.float32))
y = model(x)
else:
x = chainer.Variable(np.zeros((1, 3, 144, 192), dtype=np.float32))
y = model(x)
# Convert chainer computation graph into IR
graph = ChainerConverter().convert_from_inout_vars([x], [y])
# Generate graph descriptor
generate_descriptor(args.backend, graph, constant_encoder_name=args.encoding).save(path.join(path.dirname(__file__), "./output"))
version, blocks, filters, weights = read_net(sys.argv[1])
if data_format == 'NHWC':
planes = tf.placeholder(tf.float32, [None, BOARD_SIZE, BOARD_SIZE, FEATURES])
probs = tf.placeholder(tf.float32, [None, BOARD_SIZE * BOARD_SIZE + 1])
winner = tf.placeholder(tf.float32, [None, 1])
else:
planes = tf.placeholder(tf.float32, [None, FEATURES, BOARD_SIZE, BOARD_SIZE])
probs = tf.placeholder(tf.float32, [None, BOARD_SIZE * BOARD_SIZE + 1])
winner = tf.placeholder(tf.float32, [None, 1])
tfprocess = TFProcess()
tfprocess.INPUT_DIM = 2
tfprocess.DATA_FORMAT = data_format
tfprocess.BOARD_SIZE = BOARD_SIZE
tfprocess.FEATURES = FEATURES
tfprocess.RESIDUAL_FILTERS = filters
tfprocess.RESIDUAL_BLOCKS = blocks
training = tfprocess.training
tfprocess.training = False # batch normalizationをコンバートするため
tfprocess.init_net(planes, probs, winner)
tfprocess.replace_weights(weights)
graph = TensorFlowConverter(tfprocess.session).convert([planes], [tfprocess.y_conv, tfprocess.z_conv])
print("generating webgpu...")
exec_info = generate_descriptor("webgpu", graph)
exec_info.save("./ELF_OpenGo")
print("done")
print("generating webgl...")
exec_info = generate_descriptor("webgl", graph)
exec_info.save("./ELF_OpenGo")
def main():
sys.setrecursionlimit(10000) # workaround for deep copying large graph
parser = argparse.ArgumentParser()
parser.add_argument("kerasmodel")
parser.add_argument("--backend",
default="webgpu,webgl,webassembly,fallback",
help="comma-separated list of backends")
parser.add_argument(
"--input_shape",
required=True,
action="append",
help=
"shape of blobs for inputs (example: '(1,3,224,224)'), can be specified multiple times"
)
# parser.add_argument("--input_data_format", choices=["channels_first", "channels_last"])
parser.add_argument(
"--out",
help="output directory (default: <model>/webdnn_graph_descriptor)")
parser.add_argument("--encoding", help="name of weight encoder")
parser.add_argument("--visualize_ir", action="store_true")
parser.add_argument(
"--plugin",
action="append",
help="plugin python files which are imported before transpiling")
args = parser.parse_args()
console.stderr(f"[{path.basename(__file__)}] Generating feedforward graph")
class_list = []
if args.plugin:
for plugin_path in args.plugin:
class_list += _load_plugin(plugin_path)
custom_objects = {}
if len(class_list) > 0:
# custom_objects is a dictionary for load_model to load user-defined custom layers
for k, v in class_list:
custom_objects[k] = v
input_shapes = [
Shape.parse(input_shape)[0] for input_shape in args.input_shape
]
model = keras.models.load_model(args.kerasmodel,
custom_objects=custom_objects,
compile=False)
model.build(input_shape=None)
converter = KerasConverter(batch_size=Placeholder(label='N'))
graph = converter.convert(model)
traverse.dump(graph)
for graph_input, input_shape in zip(graph.inputs, input_shapes):
for p1, p2 in zip(graph_input.shape, input_shape):
if not Placeholder.check_resolved(
p1) and Placeholder.check_resolved(p2):
p1.value = Placeholder.force_int(p2)
elif Placeholder.check_resolved(
p1) and not Placeholder.check_resolved(p2):
raise ValueError(
f'Shape mismatch: expected:{input_shape}, real:{graph_input.shape}, {p1} != {p2}'
)
elif Placeholder.check_resolved(p1) and Placeholder.check_resolved(
p2):
assert p1 == p2, f'Shape mismatch: expected:{input_shape}, real:{graph_input.shape}, {p1} != {p2}'
if args.out:
output_dir = args.out
else:
output_dir = path.join(path.dirname(args.kerasmodel),
"webdnn_graph_descriptor")
os.makedirs(output_dir, exist_ok=True)
if args.visualize_ir:
ir_dot_path = path.join(output_dir, "ir.dot")
with open(ir_dot_path, "w") as f:
f.write(dump_dot(graph))
console.stderr(
f"IR graph can be visualized with graphviz command: 'dot {ir_dot_path} -T png -o output.png'"
)
console.stderr(f"[{path.basename(__file__)}] Generating graph descriptor")
any_backend_failed = False
backends = args.backend.split(",")
for i, backend in enumerate(backends):
console.stderr(
f"[{path.basename(__file__)}] BackendName: {console.colorize(backend, console.Color.Cyan)}"
)
try:
graph_exec_data = generate_descriptor(
backend, graph, constant_encoder_name=args.encoding)
graph_exec_data.save(output_dir)
except Exception as ex:
if flags.DEBUG:
raise ex
any_backend_failed = True
console.error(
f"[{path.basename(__file__)}] Failed generating descriptor for {backend} backend"
)
console.stderr(traceback.format_exc())
continue
if any_backend_failed:
exit(1)
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument("--model", default="fc", choices=["fc", "conv"])
parser.add_argument('--gpu',
'-g',
type=int,
default=-1,
help='GPU ID (negative value indicates CPU)')
parser.add_argument(
'--out',
'-o',
default='output_chainer',
help='Directory to output the graph descriptor and sample test data')
parser.add_argument("--backend",
default="webgpu,webgl,webassembly,fallback")
args = parser.parse_args()
output_dir = os.path.join(args.out, f"./chainer_model")
os.makedirs(output_dir, exist_ok=True)
# Set up a neural network to train
# Classifier reports softmax cross entropy loss and accuracy at every
# iteration, which will be used by the PrintReport extension below.
model = L.Classifier(models[args.model](10))
if args.gpu >= 0:
# Make a specified GPU current
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu() # Copy the model to the GPU
# Setup an optimizer
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
# Load the MNIST dataset
train, test = chainer.datasets.get_mnist(ndim=3)
train_iter = chainer.iterators.SerialIterator(train, 128)
test_iter = chainer.iterators.SerialIterator(test,
128,
repeat=False,
shuffle=False)
# Set up a trainer
updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (2, 'epoch'), out=output_dir)
# Evaluate the model with the test dataset for each epoch
trainer.extend(extensions.Evaluator(test_iter, model, device=args.gpu))
# Take a snapshot for each specified epoch
trainer.extend(extensions.snapshot(filename=args.model),
trigger=(2, 'epoch'))
# Write a log of evaluation statistics for each epoch
trainer.extend(extensions.LogReport())
# Print selected entries of the log to stdout
# Here "main" refers to the target link of the "main" optimizer again, and
# "validation" refers to the default name of the Evaluator extension.
# Entries other than 'epoch' are reported by the Classifier link, called by
# either the updater or the evaluator.
trainer.extend(
extensions.PrintReport([
'epoch', 'main/loss', 'validation/main/loss', 'main/accuracy',
'validation/main/accuracy', 'elapsed_time'
]))
# Print a progress bar to stdout
trainer.extend(extensions.ProgressBar())
snapshot_path = os.path.join(output_dir, args.model)
if os.path.exists(snapshot_path):
# Resume from a snapshot
chainer.serializers.load_npz(snapshot_path, trainer)
else:
# Run the training
trainer.run()
# conversion
print('Transpiling model to WebDNN graph descriptor')
if args.gpu >= 0:
model.to_cpu()
example_input = numpy.expand_dims(
train[0][0], axis=0) # example input (anything ok, (batch_size, 784))
x = chainer.Variable(example_input)
y = model.predictor(x)
graph = ChainerConverter().convert(
[x], [y]) # convert graph to intermediate representation
for backend in args.backend.split(","):
exec_info = generate_descriptor(backend, graph)
exec_info.save(args.out)
print('Exporting test samples (for demo purpose)')
test_samples_json = []
for i in range(10):
image, label = test[i]
test_samples_json.append({
'x': image.flatten().tolist(),
'y': int(label)
})
with open(os.path.join(args.out, 'test_samples.json'), 'w') as f:
json.dump(test_samples_json, f)