def convert_to_tfjs(exported_graph_path, output_names, output_path):
from tensorflowjs.converters import convert_tf_saved_model
saved_model_path = os.path.join(exported_graph_path, 'saved_model')
output_names_str = ','.join(output_names)
try:
convert_tf_saved_model(saved_model_path,
output_names_str,
output_path,
saved_model_tags='serve',
quantization_dtype=None,
skip_op_check=False,
strip_debug_ops=True)
except Exception as err:
print('Error: {}'.format(err))
# Move the labels to the model directory.
json_labels = os.path.join(exported_graph_path, 'labels.json')
text_labels = os.path.join(exported_graph_path, 'labels.txt')
if not os.path.isfile(json_labels):
with open(text_labels, 'r') as f:
labels = f.read()
labels = list(filter(bool, [s.strip() for s in labels.splitlines()]))
with open(os.path.join(output_path, 'labels.json'), 'w') as f:
json.dump(labels, f)
else:
shutil.copy2(json_labels, output_path)
def convert_to_tfjs(exported_graph_path, output_names, output_path):
from tensorflowjs.converters import convert_tf_saved_model
saved_model_path = os.path.join(exported_graph_path, 'saved_model')
output_names_str = ','.join(output_names)
try:
convert_tf_saved_model(
saved_model_path,
output_names_str,
output_path,
saved_model_tags='serve',
quantization_dtype=None,
skip_op_check=False,
strip_debug_ops=True)
except Exception as err:
print('Error: {}'.format(err))
# Move the labels to the model directory.
json_labels = os.path.join(exported_graph_path, 'labels.json')
text_labels = os.path.join(exported_graph_path, 'labels.txt')
if not os.path.isfile(json_labels):
with open(text_labels, 'r') as f:
labels = f.read()
labels = list(filter(bool, [s.strip() for s in labels.splitlines()]))
with open(os.path.join(output_path, 'labels.json'), 'w') as f:
json.dump(labels, f)
else:
shutil.copy2(json_labels, output_path)
def convert_to_tfjs():
from tensorflowjs.converters import convert_tf_saved_model
output_node_names = ','.join(args.output_names)
convert_tf_saved_model(
args.tf_model_path,
output_node_names,
args.tfjs_path,
saved_model_tags='serve',
quantization_dtype=None,
skip_op_check=False,
strip_debug_ops=True)
def main(*args):
base_model_path = "/tmp/jax2tf/tf_js_quickdraw"
dataset_path = os.path.join(base_model_path, "data")
classes = utils.download_dataset(dataset_path, NB_CLASSES)
assert len(classes) == NB_CLASSES, classes
print(f"Classes are: {classes}")
train_ds, test_ds = utils.load_classes(dataset_path, classes)
flax_params = train(train_ds, test_ds, classes)
model_dir = os.path.join(base_model_path, "saved_models")
# the model must be converted with with_gradient set to True to be able to
# convert the saved model to TF.js, as "PreventGradient" is not supported
saved_model_lib.convert_and_save_model(predict, flax_params, model_dir,
input_signatures=[tf.TensorSpec([1, 28, 28, 1])],
with_gradient=True, compile_model=False,
enable_xla=False)
conversion_dir = os.path.join(base_model_path, 'tfjs_models')
convert_tf_saved_model(model_dir, conversion_dir)
def to_tfjs():
import tensorflow as tf
from jax.experimental.jax2tf import convert
from tensorflowjs.converters import convert_tf_saved_model
import os
# tf_areas = convert(areas, polymorphic_shapes=["(a, 2), (b, 2), (c, 2)"])
# tf_areas = convert(areas, polymorphic_shapes=['(a, 2)', '(b, 2)', '(c, 2)'], with_gradient=False)
# f_tf_areas = tf.function(tf_areas, autograph=False)
tf_areas = convert(areas, with_gradient=True, enable_xla=False)
f_tf_areas = tf.function(tf_areas,
autograph=False,
input_signature=[
tf.TensorSpec([1, 2]),
tf.TensorSpec([1, 2]),
tf.TensorSpec([1, 2]),
])
# tf_areas = convert(areas, polymorphic_shapes=['(a, d)', '(b, d)', '(c, d)'], with_gradient=True, enable_xla=False)
# f_tf_areas = tf.function(tf_areas, autograph=False,
# input_signature=[tf.TensorSpec([None, 2]), tf.TensorSpec([None, 2]),
# tf.TensorSpec([None, 2]), ])
# f_tf_areas = tf.function(tf_areas, autograph=False)
# f_tf_areas.get_concrete_function(tf.TensorSpec([None, 2]), tf.TensorSpec([None, 2]), tf.TensorSpec([None, 2]))
# f_tf_areas.get_concrete_function(tf.TensorSpec([None, 2]), tf.TensorSpec([None, 2]), tf.TensorSpec([None, 2]))
model = tf.Module()
model.f = f_tf_areas
tf.saved_model.save(
model,
'./scripts/simplex_volume_tf',
options=tf.saved_model.SaveOptions(experimental_custom_gradients=True))
restored_model = tf.saved_model.load('./scripts/simplex_volume_tf')
restored_model.f(tf.convert_to_tensor([[0., 0.]]),
tf.convert_to_tensor([[1., 0.]]),
tf.convert_to_tensor([[0., 1.]]))
convert_tf_saved_model(
'./scripts/simplex_volume_tf',
os.path.expanduser(
'~/Work/Projects/JavaScript/understanding-modules/model'))
def main():
# load huggingface model
model = TFDistilBertForFakeNewsClassification.from_pretrained(
'./model', num_labels_aggregate=3, num_labels_category=8)
# define tensorflow function for faster inference
callable = tf.function(model.call)
max_sequence_length = 512
concrete_function = callable.get_concrete_function([
tf.TensorSpec([None, max_sequence_length], tf.int32, name="input_ids"),
tf.TensorSpec([None, max_sequence_length],
tf.int32,
name="attention_mask")
])
print(concrete_function)
# save huggingface model as tf saved_model
tf.saved_model.save(model, 'distilbert_nela', signatures=concrete_function)
# convert saved_model to tensorflow.js web format
convert_tf_saved_model('distilbert_nela', 'distilbert_nela_js')
# convert tokenizer vocab
vocab = []
with open('model/vocab.txt', 'r', encoding='utf-8') as f:
for line in f:
line = line.rstrip()
# idk why but the tokenizer library we're using does it the opposite of huggingface
if line.startswith('##'):
vocab.append(line.strip('##'))
elif line.startswith('['):
vocab.append(line)
else:
vocab.append(u'\u2581' + line)
# jsonString = json.dumps(vocab)
# print(jsonString[:50])
with open('../src/detection/tokenizer/vocab.json', 'w+',
encoding='utf-8') as f:
json.dump(vocab, f, ensure_ascii=False)