helloeve/universal-sentence-encoder-fine-tune

universal-sentence-encoder-fine-tune's Issues

When fine tuning starts, it eats up my ram

Thanks for the contribution. As i start fine tuning, my ram utilization goes upto 10Gb, what might be the cause?

any idea how to use sentence encoder as Siamese architecture?

how can we use sentence encoder as siamese architecture to pass two sentences as input?

Problem with get_operation_by_name('finetune/init_all_tables')

I've been trying to run the code on Google Colab and my local computer, both using Tensorlow 1.15.

When trying to graph the sentences before fine-tuning, the following error code:
The name 'finetune/init_all_tables' refers to an Operation not in the graph.

I downloaded the model directly from the hub and imported it on a new folder:
`from tensorflow.python.saved_model import tag_constants

scope = 'finetune'

graph=tf.Graph()

with tf.Session(graph=graph) as sess:
model_path = 'D:/Users/GermanEBR/Glite/ITESM/DCI/Tesis/Databases/USE/universal-sentence-encoder-4'
tf.saved_model.loader.load(sess, [tag_constants.SERVING], model_path)

sess.run(tf.global_variables_initializer())
sess.run(tf.get_default_graph().get_operation_by_name('finetune/init_all_tables'))

in_tensor = tf.get_default_graph().get_tensor_by_name(scope + '/module/fed_input_values:0')
ou_tensor = tf.get_default_graph().get_tensor_by_name(scope + '/module/Encoder_en/hidden_layers/l2_normalize:0')

run_and_plot(sess, in_tensor, X, ou_tensor)`

I would appreciate if someone knows what the problem is. Thanks in advance

will this also work for "https://tfhub.dev/google/universal-sentence-encoder/2"

will this finetuning method also work for transformer based encoder available at : https://tfhub.dev/google/universal-sentence-encoder/2 ?

tensorflow version compatibility

which version of tensorflow and tensorflow_hub does the code require ?

Thanks!

Running into: `ValueError: cannot add an op with id 133 as it already exists in the graph`

Tensorflow versions:

tensorflow               1.15.2
tensorflow-estimator     1.15.1
tensorflow-hub           0.9.0
tensorflow-text          1.15.1

I'm running the code as is and I'm consistently running into the following error:

/tmp/ipykernel_1454/3669827002.py in main()
     24             ops_list = g1.get_operations()
     25             for op in g1.get_operations():
---> 26                 copy_op_to_graph(op, g2, variables, scope)
     27             # copy table initilization
     28             copy_op_to_graph(tf.tables_initializer(), g2, variables, scope)

/tmp/ipykernel_1454/1749896782.py in copy_op_to_graph(org_instance, to_graph, variables, scope)
    190                            op_def)
    191     #Use Graph's hidden methods to add the op
--> 192     to_graph._add_op(new_op)  # pylint: disable=protected-access
    193     to_graph._record_op_seen_by_control_dependencies(new_op)
    194     for device_function in reversed(to_graph._device_function_stack):

/opt/miniconda/envs/product_recos_description_model/lib/python3.7/site-packages/tensorflow_core/python/framework/ops.py in _add_op(self, op)
   3015       if op._id in self._nodes_by_id:
   3016         raise ValueError("cannot add an op with id %d as it already "
-> 3017                          "exists in the graph" % op._id)
   3018       if op.name in self._nodes_by_name:
   3019         raise ValueError("cannot add op with name %s as that name "

ValueError: cannot add an op with id 133 as it already exists in the graph

I'm wondering if anyone else has run into this error.

Issue in loading the saved model in Java

I saved the trained model as follows (similar to how u do in convert_use.py)

tf.saved_model.simple_save(sess, save_path, inputs={'input':in_tensor}, outputs={'output':ou_tensor},legacy_init_op=tf.tables_initializer())

But when I load the resulting model in Java, I get the following exception. Can you please help?

W tensorflow/core/framework/op_kernel.cc:1318] OP_REQUIRES failed at lookup_table_op.cc:675 : Failed precondition: Table not initialized.
Exception in thread "main" java.lang.IllegalStateException: Table not initialized.
	 [[Node: finetune/module/string_to_index_Lookup/hash_table_Lookup = LookupTableFindV2[Tin=DT_STRING, Tout=DT_INT64, _output_shapes=[<unknown>], _device="/job:localhost/replica:0/task:0/device:CPU:0"](finetune/module/string_to_index/hash_table, finetune/module/compound_bigrams/boolean_mask/Gather, finetune/module/string_to_index/hash_table/Const)]]
	at org.tensorflow.Session.run(Native Method)

Trainable model size doubles

hi @helloeve , I was trying to use this model with module wrapper but it gives error that "graph def should be less than 2GB". While further inspecting the issue I noted that output of convert_use is 1.9GB while the original model is ~ 812 MB. Do you happen to know why this might be happening? we are not adding any new tensors or operations in graph copy.

Need help with packaging this as module

Hi can you please help me with creating a trainable module. This is the code I wrote.

def module_fn():
    scope = 'finetune'
    with tf.Session() as sess:
        model_path = 'model/'
        tf.saved_model.loader.load(sess, [tag_constants.SERVING], model_path)
        sess.run(tf.global_variables_initializer())
        sess.run(tf.get_default_graph().get_operation_by_name('finetune/init_all_tables'))
        in_tensor = tf.get_default_graph().get_tensor_by_name(scope + '/module/fed_input_values:0')
        ou_tensor = tf.get_default_graph().get_tensor_by_name(scope + '/module/Encoder_en/hidden_layers/l2_normalize:0')
        hub.add_signature(inputs=in_tensor, outputs=ou_tensor)

spec = hub.create_module_spec(module_fn)
embed = hub.Module(spec)
scope = 'finetune'

with tf.Session() as session:
    model_path = 'model/'
    tf.saved_model.loader.load(session, [tag_constants.SERVING], model_path)
    session.run([tf.global_variables_initializer(),tf.tables_initializer()])
    session.run(tf.get_default_graph().get_operation_by_name(scope + '/init_all_tables'))
    embed.export("module/",session)

But when I try to import it and find embedding, it results to an error 'FailedPreconditionError: Table not initialized.'

with tf.Session() as sess:
    embed = hub.Module("module/", trainable=True)
    sess.run([tf.global_variables_initializer(),tf.tables_initializer()])
    query_embeddings = sess.run(embed(["t shirts men"]))
    print(query_embeddings)

How to add a new dataset for finetuning?

How to add a new dataset for finetuning?

I have corpus without labels

I want to fine tune but i dont have labels

Error in Loading Model

Hi @helloeve , thanks for the super helpful implementation. Despite V2 of USE having trainable variables, I have been unable to use them.

On running your code, I keep running into this error :
Code : tf.saved_model.loader.load(sess, [tag_constants.SERVING], model_path)
Error : MetaGraphDef associated with tags 'serve' could not be found in SavedModel. To inspect available tag-sets in the SavedModel, please use the SavedModel CLI: 'saved_model_cli'

Any workaround for this? Thanks!