BERT + sentiment classifier example

examples/sentiment/cortex.yaml

@@ -0,0 +1,7 @@
+- kind: deployment
+  name: sentiment
+
+- kind: api
+  name: sentiment
+  model: s3://cortex-examples/sentiment/model.zip
+  request_handler: sentiment.py

examples/sentiment/requirements.txt

@@ -0,0 +1,2 @@
+tensorflow_hub==0.5.0
+bert-tensorflow==1.0.1

examples/sentiment/samples.json

@@ -0,0 +1,7 @@
+{
+  "samples": [
+    {
+      "input": "Great movie!"
+    }
+  ]
+}

examples/sentiment/sentiment.py

@@ -0,0 +1,41 @@
+import numpy as np
+import tensorflow as tf
+import tensorflow_hub as hub
+import bert
+from bert import run_classifier
+from bert import tokenization
+
+
+def create_tokenizer_from_hub_module():
+    """Get the vocab file and casing info from the Hub module."""
+    with tf.Graph().as_default():
+        bert_module = hub.Module("https://tfhub.dev/google/bert_uncased_L-12_H-768_A-12/1")
+        tokenization_info = bert_module(signature="tokenization_info", as_dict=True)
+        with tf.Session() as sess:
+            vocab_file, do_lower_case = sess.run(
+                [tokenization_info["vocab_file"], tokenization_info["do_lower_case"]]
+            )
+
+    return bert.tokenization.FullTokenizer(vocab_file=vocab_file, do_lower_case=do_lower_case)
+
+
+tokenizer = create_tokenizer_from_hub_module()
+labels = ["Negative", "Positive"]
+
+
+def pre_inference(sample, metadata):
+    input_feature = run_classifier.convert_single_example(
+        0,  # example Idx
+        run_classifier.InputExample(guid="", text_a=sample["input"], text_b=None, label=0),
+        [0, 1],  # label IDs
+        128,  # max len
+        tokenizer,
+    )
+    return {"input_ids": input_feature.input_ids}
+
+
+def post_inference(prediction, metadata):
+    return {
+        "sentiment": labels[prediction["response"]["labels"][0]],
+        "probabilities": prediction["response"]["probabilities"],
+    }