Merge pull request #296 from derkmed/intent-detection-dev

Add Intent Detection Node

ros/angel_system_nodes/angel_system_nodes/base_intent_detector.py

@@ -0,0 +1,164 @@
+import queue
+import rclpy
+from rclpy.node import Node
+from termcolor import colored
+import threading
+
+from angel_msgs.msg import InterpretedAudioUserIntent, Utterance
+from angel_utils import declare_and_get_parameters
+
+NEXT_STEP_KEYPHRASES = ["skip", "next", "next step"]
+PREV_STEP_KEYPHRASES = ["previous", "previous step", "last step", "go back"]
+QUESTION_KEYPHRASES = ["question"]
+OVERRIDE_KEYPHRASES = ["angel", "angel system"]
+
+# TODO(derekahmed): Please figure out how to keep this sync-ed with
+# config/angel_system_cmds/user_intent_to_sys_cmd_v1.yaml.
+# Please refer to labels defined in
+# https://docs.google.com/document/d/1uuvSL5de3LVM9c0tKpRKYazDxckffRHf7IAcabSw9UA .
+INTENT_LABELS = ["next_step", "prev_step", "inquiry", "other"]
+
+UTTERANCES_TOPIC = "utterances_topic"
+PARAM_EXPECT_USER_INTENT_TOPIC = "expect_user_intent_topic"
+PARAM_INTERP_USER_INTENT_TOPIC = "interp_user_intent_topic"
+
+
+class BaseIntentDetector(Node):
+    def __init__(self):
+        super().__init__(self.__class__.__name__)
+        self.log = self.get_logger()
+
+        # Handle parameterization.
+        param_values = declare_and_get_parameters(
+            self,
+            [
+                (UTTERANCES_TOPIC,),
+                (PARAM_EXPECT_USER_INTENT_TOPIC,),
+                (PARAM_INTERP_USER_INTENT_TOPIC,),
+            ],
+        )
+        self._utterances_topic = param_values[UTTERANCES_TOPIC]
+        self._expect_uintent_topic = param_values[PARAM_EXPECT_USER_INTENT_TOPIC]
+        self._interp_uintent_topic = param_values[PARAM_INTERP_USER_INTENT_TOPIC]
+
+        # Handle subscription/publication topics.
+        self.subscription = self.create_subscription(
+            Utterance, self._utterances_topic, self.utterance_callback, 1
+        )
+        self._expected_publisher = self.create_publisher(
+            InterpretedAudioUserIntent, self._expect_uintent_topic, 1
+        )
+        self._interp_publisher = self.create_publisher(
+            InterpretedAudioUserIntent, self._interp_uintent_topic, 1
+        )
+
+        self.utterance_message_queue = queue.Queue()
+        self.handler_thread = threading.Thread(
+            target=self.process_utterance_message_queue
+        )
+        self.handler_thread.start()
+
+    def utterance_callback(self, msg):
+        """
+        This is the main ROS node listener callback loop that will process all messages received
+        via subscribed topics.
+        """
+        self.log.debug(f'Received message:\n\n"{msg.value}"')
+        self.utterance_message_queue.put(msg)
+
+    def process_utterance_message_queue(self):
+        """
+        Constant loop to process received messages.
+        """
+        while True:
+            msg = self.utterance_message_queue.get()
+            self.log.debug(f'Processing message:\n\n"{msg.value}"')
+            intent, score = self.detect_intents(msg)
+            if not intent:
+                continue
+            self.publish_msg(msg.value, intent, score)
+
+    def detect_intents(self, msg):
+        """
+        Keyphrase search for intent detection. This implementation does simple
+        string matching to assign a detected label. When multiple intents are
+        detected, the message is classified as the first intent or as an
+        'inquiry' if 'inquiry' is one of the classifications.
+        """
+
+        def _tiebreak_intents(intents, confidences):
+            classification = intents[0]
+            score = confidences[0]
+            if len(intents) > 1:
+                for i, intent in enumerate(intents):
+                    if intent == INTENT_LABELS[2]:
+                        classification, score = intent, confidences[i]
+                self.log.info(
+                    f'Detected multiple intents: {intents}. Selected "{classification}".'
+                )
+            return classification, score
+
+        lower_utterance = msg.value.lower()
+        intents = []
+        confidences = []
+        if self._contains_phrase(lower_utterance, NEXT_STEP_KEYPHRASES):
+            intents.append(INTENT_LABELS[0])
+            confidences.append(0.5)
+        if self._contains_phrase(lower_utterance, PREV_STEP_KEYPHRASES):
+            intents.append(INTENT_LABELS[1])
+            confidences.append(0.5)
+        if self._contains_phrase(lower_utterance, QUESTION_KEYPHRASES):
+            intents.append(INTENT_LABELS[2])
+            confidences.append(0.5)
+        if not intents:
+            colored_utterance = colored(msg.value, "light_blue")
+            self.log.info(f'No intents detected for:\n>>> "{colored_utterance}":')
+            return None, -1.0
+
+        classification, confidence = _tiebreak_intents(intents, confidences)
+        classification = colored(classification, "light_green")
+        return classification, confidence
+
+    def publish_msg(self, utterance, intent, score):
+        """
+        Handles message publishing for an utterance with a detected intent.
+        """
+        intent_msg = InterpretedAudioUserIntent()
+        intent_msg.header.frame_id = "Intent Detection"
+        intent_msg.header.stamp = self.get_clock().now().to_msg()
+        intent_msg.utterance_text = utterance
+        intent_msg.user_intent = intent
+        intent_msg.confidence = score
+        published_topic = None
+        if self._contains_phrase(utterance.lower(), OVERRIDE_KEYPHRASES):
+            intent_msg.confidence = 1.0
+            self._expected_publisher.publish(intent_msg)
+            published_topic = PARAM_EXPECT_USER_INTENT_TOPIC
+        else:
+            self._interp_publisher.publish(intent_msg)
+            published_topic = PARAM_INTERP_USER_INTENT_TOPIC
+
+        colored_utterance = colored(utterance, "light_blue")
+        colored_intent = colored(intent_msg.user_intent, "light_green")
+        self.log.info(
+            f'Publishing {{"{colored_intent}": {score}}} to {published_topic} '
+            + f'for:\n>>> "{colored_utterance}"'
+        )
+
+    def _contains_phrase(self, utterance, phrases):
+        for phrase in phrases:
+            if phrase in utterance:
+                return True
+        return False
+
+
+def main():
+    rclpy.init()
+    intent_detector = BaseIntentDetector()
+    rclpy.spin(intent_detector)
+    intent_detector.destroy_node()
+    rclpy.shutdown()
+
+
+if __name__ == "__main__":
+    main()

ros/angel_system_nodes/angel_system_nodes/gpt_intent_detector.py

@@ -0,0 +1,100 @@
+from langchain import PromptTemplate, FewShotPromptTemplate
+from langchain.chains import LLMChain
+from langchain.chat_models import ChatOpenAI
+from langchain.llms import OpenAI
+import openai
+import os
+import rclpy
+
+from angel_system_nodes.base_intent_detector import BaseIntentDetector, INTENT_LABELS
+
+openai.organization = os.getenv("OPENAI_ORG_ID")
+openai.api_key = os.getenv("OPENAI_API_KEY")
+
+# The following are few shot examples when prompting GPT.
+FEW_SHOT_EXAMPLES = [
+    {"utterance": "Go back to the previous step!", "label": "prev_step."},
+    {"utterance": "Next step, please.", "label": "next_step"},
+    {"utterance": "How should I wrap this tourniquet?", "label": "inquiry"},
+    {"utterance": "The sky is blue", "label": "other"},
+]
+
+
+class GptIntentDetector(BaseIntentDetector):
+    def __init__(self):
+        super().__init__()
+        self.log = self.get_logger()
+
+        # This node additionally includes fields for interacting with OpenAI
+        # via LangChain.
+        if not os.getenv("OPENAI_API_KEY"):
+            self.log.info("OPENAI_API_KEY environment variable is unset!")
+        else:
+            self.openai_api_key = os.getenv("OPENAI_API_KEY")
+        if not os.getenv("OPENAI_ORG_ID"):
+            self.log.info("OPENAI_ORG_ID environment variable is unset!")
+        else:
+            self.openai_org_id = os.getenv("OPENAI_ORG_ID")
+        if not bool(self.openai_api_key and self.openai_org_id):
+            raise ValueError("Please configure OpenAI API Keys.")
+        self.chain = self._configure_langchain()
+
+    def _configure_langchain(self):
+        def _labels_list_parenthetical_str(labels):
+            concat_labels = ", ".join(labels)
+            return f"({concat_labels})"
+
+        def _labels_list_str(labels):
+            return ", ".join(labels[:-1]) + f" or {labels[-1]}"
+
+        all_intents_parenthetical = _labels_list_parenthetical_str(INTENT_LABELS)
+        all_intents = _labels_list_str(INTENT_LABELS)
+
+        # Define the few shot template.
+        template = (
+            f"Utterance: {{utterance}}\nIntent {all_intents_parenthetical}: {{label}}"
+        )
+        example_prompt = PromptTemplate(
+            input_variables=["utterance", "label"], template=template
+        )
+        prompt_instructions = f"Classify each utterance as {all_intents}.\n"
+        inference_sample = (
+            f"Utterance: {{utterance}}\nIntent {all_intents_parenthetical}:"
+        )
+        few_shot_prompt = FewShotPromptTemplate(
+            examples=FEW_SHOT_EXAMPLES,
+            example_prompt=example_prompt,
+            prefix=prompt_instructions,
+            suffix=inference_sample,
+            input_variables=["utterance"],
+            example_separator="\n",
+        )
+
+        # Please refer to https://github.com/hwchase17/langchain/blob/master/langchain/llms/openai.py
+        openai_llm = ChatOpenAI(
+            model_name="gpt-3.5-turbo",
+            openai_api_key=self.openai_api_key,
+            temperature=0.0,
+            # Only 2 tokens needed for classification (tokens are delimited by use of '_', i.e.
+            # 'next_step' counts as 2 tokens).
+            max_tokens=2,
+        )
+        return LLMChain(llm=openai_llm, prompt=few_shot_prompt)
+
+    def detect_intents(self, msg):
+        """
+        Detects the user intent via langchain execution of GPT.
+        """
+        return self.chain.run(utterance=msg), 0.5
+
+
+def main():
+    rclpy.init()
+    intent_detector = GptIntentDetector()
+    rclpy.spin(intent_detector)
+    intent_detector.destroy_node()
+    rclpy.shutdown()
+
+
+if __name__ == "__main__":
+    main()

ros/angel_system_nodes/setup.py

@@ -22,6 +22,8 @@
             "video_listener = angel_system_nodes.video_subscriber:main",
             "asr = angel_system_nodes.asr:main",
             "voice_activity_detector = angel_system_nodes.voice_activity_detector:main",
+            "base_intent_detector = angel_system_nodes.base_intent_detector:main",
+            "gpt_intent_detector = angel_system_nodes.gpt_intent_detector:main",
             "intent_detector = angel_system_nodes.intent_detector:main",
             "audio_player = angel_system_nodes.audio_player:main",
             "frame_publisher = angel_system_nodes.ros_publisher:main",

tmux/eval_vocalized_intent_detection.yml

@@ -0,0 +1,96 @@
+#
+# Used to evaluate Intent Detection with vocal processing for a specified ROS bag of data 
+# This configuration should be run by itself (e.g. not in combination with
+# another tmuxinator launch).
+#
+# NOTE: In order to query GPT, you will need to execute
+# ```
+# export OPENAI_API_KEY="YOUR API KEY"
+# export OPENAI_ORG_ID="YOUR ORG ID"
+# ```
+#
+
+name: Intent Detection with VAD
+root: <%= ENV["ANGEL_WORKSPACE_DIR"] %>
+
+# Optional tmux socket
+# socket_name: foo
+
+# Note that the pre and post options have been deprecated and will be replaced by
+# project hooks.
+
+# Project hooks
+
+# Runs on project start, always
+# on_project_start: command
+on_project_start: |
+  export ROS_NAMESPACE=${ROS_NAMESPACE:-/debug}
+  export CONFIG_DIR=${ANGEL_WORKSPACE_DIR}/src/angel_system_nodes/configs
+  export NODE_RESOURCES_DIR=${ANGEL_WORKSPACE_DIR}/src/angel_system_nodes/resource
+# Run on project start, the first time
+# on_project_first_start: command
+
+# Run on project start, after the first time
+# on_project_restart: command
+
+# Run on project exit ( detaching from tmux session )
+# on_project_exit: command
+
+# Run on project stop
+# on_project_stop: command
+
+# Runs in each window and pane before window/pane specific commands. Useful for setting up interpreter versions.
+# pre_window: rbenv shell 2.0.0-p247
+
+# Pass command line options to tmux. Useful for specifying a different tmux.conf.
+# tmux_options: -f ~/.tmux.mac.conf
+tmux_options: -f <%= ENV["ANGEL_WORKSPACE_DIR"] %>/tmux/tmux.conf
+
+# Change the command to call tmux.  This can be used by derivatives/wrappers like byobu.
+# tmux_command: byobu
+
+# Specifies (by name or index) which window will be selected on project startup. If not set, the first window is used.
+# startup_window: editor
+
+# Specifies (by index) which pane of the specified window will be selected on project startup. If not set, the first pane is used.
+# startup_pane: 1
+
+# Controls whether the tmux session should be attached to automatically. Defaults to true.
+# attach: false
+
+windows:
+  # - ros_bag_play: ros2 bag play <<PATH_TO_BAG_FILE>>
+  - ros_bag_play: sleep 5; ros2 bag play /angel_workspace/ros_bags/rosbag2_2023_03_01-17_28_00_0.db3
+  - vocal:
+      layout: even-vertical
+      panes:
+        - vad: ros2 run angel_system_nodes voice_activity_detector --ros-args
+            -r __ns:=${ROS_NAMESPACE}
+            -p input_audio_topic:=HeadsetAudioData
+            -p output_voice_activity_topic:=DetectedVoiceData
+            -p vad_server_url:=http://communication.cs.columbia.edu:55667/vad
+            -p vad_cadence:=3
+            -p vad_margin:=0.20
+            -p max_accumulation_length:=10
+            -p debug_mode:=True
+        - asr: ros2 run angel_system_nodes asr --ros-args
+            -r __ns:=${ROS_NAMESPACE}
+            -p audio_topic:=DetectedVoiceData
+            -p utterances_topic:=utterances_topic
+            -p asr_server_url:=http://communication.cs.columbia.edu:55667/asr
+            -p asr_req_segment_duration:=1
+            -p is_sentence_tokenize:=False
+            -p debug_mode:=True
+  - intent:
+      layout: even-vertical
+      panes:
+        - base_intent_detection: ros2 run angel_system_nodes base_intent_detector --ros-args
+           -r __ns:=${ROS_NAMESPACE}
+           -p utterances_topic:=utterances_topic
+           -p expect_user_intent_topic:=expect_user_intent_topic
+           -p interp_user_intent_topic:=interp_user_intent_topic
+        - gpt_intent_detection: ros2 run angel_system_nodes gpt_intent_detector --ros-args
+           -r __ns:=${ROS_NAMESPACE}
+           -p utterances_topic:=utterances_topic
+           -p expect_user_intent_topic:=expect_user_intent_topic
+           -p interp_user_intent_topic:=interp_user_intent_topic