I am serializing a trained model using
TradNaiveBayesClassifier classifier = new TradNaiveBayesClassifier(categories,tokenizerFactory,categoryPrior,tokenInCategoryPrior,lengthNorm);
then I trained it and compiled it using
AbstractExternalizable.compileTo(classifier,new File(modelPath));
When I read in the model using
TradNaiveBayesClassifier decompClassifier = (TradNaiveBayesClassifier)AbstractExternalizable.readObject(new File(modelPath));{
I get a ClassCastException. Any ideas?
I got it working. I had to upcast to BaseClassifier:
BaseClassifier<CharSequence> eval = (BaseClassifier<CharSequence>)AbstractExternalizable.readObject(new File(modelPath));
evaluator = new BaseClassifierEvaluator<CharSequence>(eval, cat, storeInputs);
Then I couldn't use the JointClassifierEvaluator any more, I had to use the BaseClassifierEvaluator.
Related
I'm struggling to correctly set Vertex AI pipeline which does the following:
read data from API and store to GCS and as as input for batch prediction.
get an existing model (Video classification on Vertex AI)
create Batch prediction job with input from point 1.
As it will be seen, I don't have much experience with Vertex Pipelines/Kubeflow thus I'm asking for help/advice, hope it's just some beginner mistake.
this is the gist of the code I'm using as pipeline
from google_cloud_pipeline_components import aiplatform as gcc_aip
from kfp.v2 import dsl
from kfp.v2.dsl import component
from kfp.v2.dsl import (
Output,
Artifact,
Model,
)
PROJECT_ID = 'my-gcp-project'
BUCKET_NAME = "mybucket"
PIPELINE_ROOT = "{}/pipeline_root".format(BUCKET_NAME)
#component
def get_input_data() -> str:
# getting data from API, save to Cloud Storage
# return GS URI
gcs_batch_input_path = 'gs://somebucket/file'
return gcs_batch_input_path
#component(
base_image="python:3.9",
packages_to_install=['google-cloud-aiplatform==1.8.0']
)
def load_ml_model(project_id: str, model: Output[Artifact]):
"""Load existing Vertex model"""
import google.cloud.aiplatform as aip
model_id = '1234'
model = aip.Model(model_name=model_id, project=project_id, location='us-central1')
#dsl.pipeline(
name="batch-pipeline", pipeline_root=PIPELINE_ROOT,
)
def pipeline(gcp_project: str):
input_data = get_input_data()
ml_model = load_ml_model(gcp_project)
gcc_aip.ModelBatchPredictOp(
project=PROJECT_ID,
job_display_name=f'test-prediction',
model=ml_model.output,
gcs_source_uris=[input_data.output], # this doesn't work
# gcs_source_uris=['gs://mybucket/output/'], # hardcoded gs uri works
gcs_destination_output_uri_prefix=f'gs://{PIPELINE_ROOT}/prediction_output/'
)
if __name__ == '__main__':
from kfp.v2 import compiler
import google.cloud.aiplatform as aip
pipeline_export_filepath = 'test-pipeline.json'
compiler.Compiler().compile(pipeline_func=pipeline,
package_path=pipeline_export_filepath)
# pipeline_params = {
# 'gcp_project': PROJECT_ID,
# }
# job = aip.PipelineJob(
# display_name='test-pipeline',
# template_path=pipeline_export_filepath,
# pipeline_root=f'gs://{PIPELINE_ROOT}',
# project=PROJECT_ID,
# parameter_values=pipeline_params,
# )
# job.run()
When running the pipeline it throws this exception when running Batch prediction:
details = "List of found errors: 1.Field: batch_prediction_job.model; Message: Invalid Model resource name.
so I'm not sure what could be wrong. I tried to load model in the notebook (outside of component) and it correctly returns.
Second issue I'm having is referencing GCS URI as output from component to batch job input.
input_data = get_input_data2()
gcc_aip.ModelBatchPredictOp(
project=PROJECT_ID,
job_display_name=f'test-prediction',
model=ml_model.output,
gcs_source_uris=[input_data.output], # this doesn't work
# gcs_source_uris=['gs://mybucket/output/'], # hardcoded gs uri works
gcs_destination_output_uri_prefix=f'gs://{PIPELINE_ROOT}/prediction_output/'
)
During compilation, I get following exception TypeError: Object of type PipelineParam is not JSON serializable, though I think this could be issue of ModelBatchPredictOp component.
Again any help/advice appreciated, I'm dealing with this from yesterday, so maybe I missed something obvious.
libraries I'm using:
google-cloud-aiplatform==1.8.0
google-cloud-pipeline-components==0.2.0
kfp==1.8.10
kfp-pipeline-spec==0.1.13
kfp-server-api==1.7.1
UPDATE
After comments, some research and tuning, for referencing model this works:
#component
def load_ml_model(project_id: str, model: Output[Artifact]):
region = 'us-central1'
model_id = '1234'
model_uid = f'projects/{project_id}/locations/{region}/models/{model_id}'
model.uri = model_uid
model.metadata['resourceName'] = model_uid
and then I can use it as intended:
batch_predict_op = gcc_aip.ModelBatchPredictOp(
project=gcp_project,
job_display_name=f'batch-prediction-test',
model=ml_model.outputs['model'],
gcs_source_uris=[input_batch_gcs_path],
gcs_destination_output_uri_prefix=f'gs://{BUCKET_NAME}/prediction_output/test'
)
UPDATE 2
regarding GCS path, a workaround is to define path outside of the component and pass it as an input parameter, for example (abbreviated):
#dsl.pipeline(
name="my-pipeline",
pipeline_root=PIPELINE_ROOT,
)
def pipeline(
gcp_project: str,
region: str,
bucket: str
):
ts = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
gcs_prediction_input_path = f'gs://{BUCKET_NAME}/prediction_input/video_batch_prediction_input_{ts}.jsonl'
batch_input_data_op = get_input_data(gcs_prediction_input_path) # this loads input data to GCS path
batch_predict_op = gcc_aip.ModelBatchPredictOp(
project=gcp_project,
model=training_job_run_op.outputs["model"],
job_display_name='batch-prediction',
# gcs_source_uris=[batch_input_data_op.output],
gcs_source_uris=[gcs_prediction_input_path],
gcs_destination_output_uri_prefix=f'gs://{BUCKET_NAME}/prediction_output/',
).after(batch_input_data_op) # we need to add 'after' so it runs after input data is prepared since get_input_data doesn't returns anything
still not sure, why it doesn't work/compile when I return GCS path from get_input_data component
I'm glad you solved most of your main issues and found a workaround for model declaration.
For your input.output observation on gcs_source_uris, the reason behind it is because the way the function/class returns the value. If you dig inside the class/methods of google_cloud_pipeline_components you will find that it implements a structure that will allow you to use .outputs from the returned value of the function called.
If you go to the implementation of one of the components of the pipeline you will find that it returns an output array from convert_method_to_component function. So, in order to have that implemented in your custom class/function your function should return a value which can be called as an attribute. Below is a basic implementation of it.
class CustomClass():
def __init__(self):
self.return_val = {'path':'custompath','desc':'a desc'}
#property
def output(self):
return self.return_val
hello = CustomClass()
print(hello.output['path'])
If you want to dig more about it you can go to the following pages:
convert_method_to_component, which is the implementation of convert_method_to_component
Properties, basics of property in python.
I am beginner to Julia though I have experience with Python and some other languages. I get that this is probably a very simple/beginner issue, but I fail to understand how it should work in Julia.
I want to create a Julia module. I saw recommendations to create it with PkgTemplates, so that is exactly what I have done. My directory structure is thus:
It is located at the default path proposed by PkgTemplates: /home/username/.julia/dev/Keras2Flux.
I want to develop it with Revise package due to the slow start-up time of the Julia REPL. However, I fail to import my module to the Julia REPL in the terminal.
So, I cd to the directory mentioned above, use julia command and try using Keras2Flux. I get the error:
ERROR: ArgumentError: Package Keras2Flux not found in current path:
I tried both using Keras2Flux and using Keras2Flux.jl, and I also tried to call it from one level above in my directory structure (i.e. /home/username/.julia/dev). All has the same problem.
What is wrong (more importantly, why?) and how to fix it?
Current contents of the module (not really relevant to the question but still):
module Keras2Flux
import JSON
using Flux
export convert
function create_dense(config)
in = config["input_dim"]
out = config["output_dim"]
dense = Dense(in, outŠ¾)
return dense
end
function create_dropout(config)
p = config["p"]
dropout = Dropout(p)
return dropout
end
function create_model(model_config)
layers = []
for layer_config in model_config
if layer_config["class_name"] == "Dense"
layer = create_dense(layer_config["config"])
elseif layer_config["class_name"] == "Dropout"
layer = create_dropout(layer_config["config"])
else
println(layer_config["class_name"])
throw("unimplemented")
end
push!(layers, layer)
end
model = Chain(layers)
end
function convert(filename)
jsontxt = ""
open(filename, "r") do f
jsontxt = read(f, String)
end
model_params = JSON.parse(jsontxt)
if model_params["keras_version"] == "1.1.0"
create_model(model_params["config"])
else
throw("unimplemented")
end
end
end
Here is a full recipe to get you going:
cd("/home/username/.julia/dev")
using Pkg
pkg"generate Keras2Flux"
cd("Keras2Flux")
pkg"activate ."
pkg"add JSON Flux"
# now copy-paste whatever you need to Keras2Flux\src\Keras2Flux.jl
using Revise
using Keras2Flux
# happy development!
I am currenly attempting to train a NER model centered around Property Descriptions. I could get a fully trained model to function to my liking however, I now want to add a retokenize pipe to the model so that I can set up the model to train other things.
From here, I am having issues getting the retokenize pipe to actually work. Here is the definition:
def retok(doc):
ents = [(ent.start, ent.end, ent.label) for ent in doc.ents]
with doc.retokenize() as retok:
string_store = doc.vocab.strings
for start, end, label in ents:
retok.merge(
doc[start: end],
attrs=intify_attrs({'ent_type':label},string_store))
return doc
i am adding it into my training like this:
nlp.add_pipe(retok, after="ner")
and I am adding it into the Language Factories like this:
Language.factories['retok'] = lambda nlp, **cfg: retok(nlp)
The issue I keep getting is "AttributeError: 'English' object has no attribute 'ents'". Now I am assuming I am getting this error because the parameter that is being passed through this function is not a doc but actually the NLP model itself. I am not really sure to get a doc to flow into this pipe during training. At this point I don't really know where to go from here to get the pipe to function the way I want.
Any help is appreciated, thanks.
You can potentially use the built-in merge_entities pipeline component: https://spacy.io/api/pipeline-functions#merge_entities
The example copied from the docs:
texts = [t.text for t in nlp("I like David Bowie")]
assert texts == ["I", "like", "David", "Bowie"]
merge_ents = nlp.create_pipe("merge_entities")
nlp.add_pipe(merge_ents)
texts = [t.text for t in nlp("I like David Bowie")]
assert texts == ["I", "like", "David Bowie"]
If you need to customize it further, the current implementation of merge_entities (v2.2) is a good starting point:
def merge_entities(doc):
"""Merge entities into a single token.
doc (Doc): The Doc object.
RETURNS (Doc): The Doc object with merged entities.
DOCS: https://spacy.io/api/pipeline-functions#merge_entities
"""
with doc.retokenize() as retokenizer:
for ent in doc.ents:
attrs = {"tag": ent.root.tag, "dep": ent.root.dep, "ent_type": ent.l
abel}
retokenizer.merge(ent, attrs=attrs)
return doc
P.S. You are passing nlp to retok() below, which is where the error is coming from:
Language.factories['retok'] = lambda nlp, **cfg: retok(nlp)
See a related question: Spacy - Save custom pipeline
After loading the wide and deep model, i was able to make prediction for one request object using the map of features and then serializing it to string for predictions as shown below-
is there a way we can create a batch of requests objects and send them for prediction to tensorflow server?
Code for single prediction looks like this-
for (each feature in feature list) {
Feature feature = null;
feature = Feature.newBuilder().setBytesList(BytesList.newBuilder().addValue(ByteString.copyFromUtf8("dummy string"))).build();
if (feature != null) {
inputFeatureMap.put(name, feature);
}
}
//Converting features(in inputFeatureMap) corresponding to one request into 'Features' Proto object
Features features = Features.newBuilder().putAllFeature(inputFeatureMap).build();
inputStr = Example.newBuilder().setFeatures(features).build().toByteString();
}
TensorProto proto = TensorProto.newBuilder()
.addStringVal(inputStr)
.setTensorShape(TensorShapeProto.newBuilder().addDim(TensorShapeProto.Dim.newBuilder().setSize(1).build()).build())
.setDtype(DataType.DT_STRING)
.build();
PredictRequest req = PredictRequest.newBuilder()
.setModelSpec(ModelSpec.newBuilder()
.setName("your serving model name")
.setSignatureName("serving_default")
.setVersion(Int64Value.newBuilder().setValue(modelVer)))
.putAllInputs(ImmutableMap.of("inputs", proto))
.build();
PredictResponse response = stub.predict(req);
System.out.println(response.getOutputsMap());
Is there a way we can send the list of Features Object for predictions, something similar to this-
List<Features> = {someway to create array/list of inputFeatureMap's which can be converted to serialized string.}
For anyone stumbling here, I found a simple workaround with Example proto to do batch request. I will borrow code from this question and modify it for the batch.
Features features =
Features.newBuilder()
.putFeature("Attribute1", feature("A12"))
.putFeature("Attribute2", feature(12))
.putFeature("Attribute3", feature("A32"))
.putFeature("Attribute4", feature("A40"))
.putFeature("Attribute5", feature(7472))
.putFeature("Attribute6", feature("A65"))
.putFeature("Attribute7", feature("A71"))
.putFeature("Attribute8", feature(1))
.putFeature("Attribute9", feature("A92"))
.putFeature("Attribute10", feature("A101"))
.putFeature("Attribute11", feature(2))
.putFeature("Attribute12", feature("A121"))
.putFeature("Attribute13", feature(24))
.putFeature("Attribute14", feature("A143"))
.putFeature("Attribute15", feature("A151"))
.putFeature("Attribute16", feature(1))
.putFeature("Attribute17", feature("A171"))
.putFeature("Attribute18", feature(1))
.putFeature("Attribute19", feature("A191"))
.putFeature("Attribute20", feature("A201"))
.build();
Example example = Example.newBuilder().setFeatures(features).build();
String pfad = System.getProperty("user.dir") + "\\1511523781";
try (SavedModelBundle model = SavedModelBundle.load(pfad, "serve")) {
Session session = model.session();
final String xName = "input_example_tensor";
final String scoresName = "dnn/head/predictions/probabilities:0";
try (Tensor<String> inputBatch = Tensors.create(new byte[][] {example.toByteArray(), example.toByteArray(), example.toByteArray(), example.toByteArray()});
Tensor<Float> output =
session
.runner()
.feed(xName, inputBatch)
.fetch(scoresName)
.run()
.get(0)
.expect(Float.class)) {
System.out.println(Arrays.deepToString(output.copyTo(new float[4][2])));
}
}
Essentially you can pass each example as an object in byte[4][] and you will receive the result in the same shape float[4][2]
I want to create and train a model, export it and run inference in C++.
I'm following the tutorial listed here: https://www.tensorflow.org/tutorials/wide_and_deep
I'm also trying to use the SavedModel approach as described here since this is the canonical way to export TensorFlow graphs for serving:
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md.
At the very end, I export the saved model as follows:
feature_spec = tf.contrib.layers.create_feature_spec_for_parsing(feature_columns)
serving_input_fn = input_fn_utils.build_parsing_serving_input_fn(feature_spec)
output = model.export_savedmodel(model_dir, serving_input_fn, as_text=True)
print('Model saved to {}'.format(output))
I see the saved_model.pbtxt has the following signature definition.
signature_def {
key: "serving_default"
value {
inputs {
key: "inputs"
value {
name: "input_example_tensor:0"
dtype: DT_STRING
tensor_shape {
dim {
size: -1
}
}
}
}
outputs {
...
I can load the saved model on the C++ side
SavedModelBundle bundle;
const std::string graph_path = "models/1498572863";
const std::unordered_set<std::string> tags = {"serve"};
Status status = LoadSavedModel(session_options,
run_options, graph_path,
tags, &bundle);
I'm stuck at the last part where I need to feed the input into this model.
The Run function expects the input parameter to be of the form: std::vector<std::pair<string, Tensor>>.
I would have expected this to be a vector of pairs where the key is the feature name used in the python code and the Tensor is multiple values for that feature.
However, it seems to expect the string to be "input_example_tensor".
I'm not sure how I'm supposed to now feed the model with different features using a single Tensor.
std::vector<string> output_tensor_names = {
"binary_logistic_head/_classification_output_alternatives/classes_tensor"};
// How do I create input_tensor?
status = bundle.session->Run({{"input_example_tensor", input_tensor}}
output_tensor_names, {}, &outputs);
Solution
I did something like this
tensorflow::Example example;
auto& tf_feature_map = *(example.mutable_features()->mutable_feature());
tf_feature_map["name"].mutable_int64_list()->add_value(15);
const std::string& serialized = example.SerializeAsString();
tensorflow::Input input({serialized});
status = bundle.session->Run({{"input_example_tensor", input.tensor()}}
output_tensor_names, {}, &outputs);
Your model signature suggests that it is expecting a DT_STRING tensor as input. When using tensorflow::Example, this typically means that the protocol buffer needs to be serialized into a tensor with a string as the type of its elements.
To convert the tensorflow::Example object to a string, you can use the protocol buffer methods such as SerializeToString, SerializeAsString etc.
Hope that helps.