Fine tuning support (#219)

* moved fine tuning to meta tools

* fix error messages

* move get_diff and add use_florence2_fine_tuning

* add fine tuning arg to florence2

* set notebook execute path to remote path'

* remove comments

* fix bug exec isolation wasn't setting resources

* ensure agent uses print to view results

* fixed bug with edit code errors

* fixed bug with edit code errors, and fixed replace code for fine tune

* add imports for new meta tools

* fixed type errors

* fix format issue

* fixed regex

* fix bug with upload return path

vision_agent/agent/vision_agent.py

@@ -30,7 +30,7 @@ class BoilerplateCode:
  pre_code = [
  "from typing import *",
  "from vision_agent.utils.execute import CodeInterpreter",
- "from vision_agent.tools.meta_tools import Artifacts, open_code_artifact, create_code_artifact, edit_code_artifact, get_tool_descriptions, generate_vision_code, edit_vision_code, write_media_artifact",
+ "from vision_agent.tools.meta_tools import Artifacts, open_code_artifact, create_code_artifact, edit_code_artifact, get_tool_descriptions, generate_vision_code, edit_vision_code, write_media_artifact, florence2_fine_tuning, use_florence2_fine_tuning",
  "artifacts = Artifacts('{remote_path}')",
  "artifacts.load('{remote_path}')",
  ]
@@ -76,11 +76,16 @@ def run_conversation(orch: LMM, chat: List[Message]) -> Dict[str, Any]:
 
 def run_code_action(
  code: str, code_interpreter: CodeInterpreter, artifact_remote_path: str
-) -> Execution:
- return code_interpreter.exec_isolation(
+) -> Tuple[Execution, str]:
+ result = code_interpreter.exec_isolation(
  BoilerplateCode.add_boilerplate(code, remote_path=artifact_remote_path)
  )
 
+ obs = str(result.logs)
+ if result.error:
+ obs += f"\n{result.error}"
+ return result, obs
+
 
 def parse_execution(response: str) -> Optional[str]:
  code = None
@@ -192,7 +197,7 @@ def chat_with_code(
  artifacts = Artifacts(WORKSPACE / "artifacts.pkl")
 
  with CodeInterpreterFactory.new_instance(
- code_sandbox_runtime=self.code_sandbox_runtime
+ code_sandbox_runtime=self.code_sandbox_runtime,
  ) as code_interpreter:
  orig_chat = copy.deepcopy(chat)
  int_chat = copy.deepcopy(chat)
@@ -260,10 +265,9 @@ def chat_with_code(
  code_action = parse_execution(response["response"])
 
  if code_action is not None:
- result = run_code_action(
+ result, obs = run_code_action(
  code_action, code_interpreter, str(remote_artifacts_path)
  )
- obs = str(result.logs)
 
  if self.verbosity >= 1:
  _LOGGER.info(obs)

vision_agent/agent/vision_agent_coder.py

@@ -1,5 +1,4 @@
 import copy
-import difflib
 import logging
 import os
 import sys
@@ -29,6 +28,7 @@
  USER_REQ,
 )
 from vision_agent.lmm import LMM, AzureOpenAILMM, Message, OllamaLMM, OpenAILMM
+from vision_agent.tools.meta_tools import get_diff
 from vision_agent.utils import CodeInterpreterFactory, Execution
 from vision_agent.utils.execute import CodeInterpreter
 from vision_agent.utils.image_utils import b64_to_pil
@@ -63,14 +63,6 @@ def prepend_imports(code: str) -> str:
  return DefaultImports.to_code_string() + "\n\n" + code
 
 
-def get_diff(before: str, after: str) -> str:
- return "".join(
- difflib.unified_diff(
- before.splitlines(keepends=True), after.splitlines(keepends=True)
- )
- )
-
-
 def format_memory(memory: List[Dict[str, str]]) -> str:
  output_str = ""
  for i, m in enumerate(memory):

vision_agent/agent/vision_agent_prompts.py

@@ -48,7 +48,7 @@
 4| return dogs
 [End of artifact]
 
-AGENT: {"thoughts": "I have generated the code to detect the dogs in the image, I must now run the code to get the output.", "response": "<execute_python>from dog_detector import detect_dogs\n print(detect_dogs('/example/workspace/dog.jpg'))</execute_python>", "let_user_respond": false}
+AGENT: {"thoughts": "I have generated the code to detect the dogs in the image, I must now run the code and print the results to get the output.", "response": "<execute_python>from dog_detector import detect_dogs\n print(detect_dogs('/example/workspace/dog.jpg'))</execute_python>", "let_user_respond": false}
 
 OBSERVATION:
 ----- stdout -----
@@ -75,7 +75,7 @@
 4| return dogs
 [End of artifact]
 
-AGENT: {"thoughts": "I have edited the code to detect only one dog, I must now run the code to get the output.", "response": "<execute_python>from dog_detector import detect_dogs\n print(detect_dogs('/path/to/images/dog.jpg'))</execute_python>", "let_user_respond": false}
+AGENT: {"thoughts": "I have edited the code to detect only one dog, I must now run the code and print the results to get the output.", "response": "<execute_python>from dog_detector import detect_dogs\n print(detect_dogs('/path/to/images/dog.jpg'))</execute_python>", "let_user_respond": false}
 
 OBSERVATION:
 ----- stdout -----
@@ -126,7 +126,7 @@
 15| return count
 [End of artifact]
 
-AGENT: {"thoughts": "I have generated the code to count the workers with helmets in the image, I must now run the code to get the output and write the visualization to the artifacts so the user can see it.", "response": "<execute_python>from code import count_workers_with_helmets\n print(count_workers_with_helmets('/path/to/images/workers.png', 'workers_viz.png'))\n write_media_artifact(artifacts, 'workers_viz.png')</execute_python>", "let_user_respond": false}
+AGENT: {"thoughts": "I have generated the code to count the workers with helmets in the image, I must now run the code and print the output and write the visualization to the artifacts so I can see the result and the user can see the visaulization.", "response": "<execute_python>from code import count_workers_with_helmets\n print(count_workers_with_helmets('/path/to/images/workers.png', 'workers_viz.png'))\n write_media_artifact(artifacts, 'workers_viz.png')</execute_python>", "let_user_respond": false}
 
 OBSERVATION:
 ----- stdout -----

vision_agent/tools/meta_tools.py

@@ -1,5 +1,7 @@
+import difflib
 import os
 import pickle as pkl
+import re
 import subprocess
 import tempfile
 from pathlib import Path
@@ -8,10 +10,13 @@
 from IPython.display import display
 
 import vision_agent as va
+from vision_agent.clients.landing_public_api import LandingPublicAPI
 from vision_agent.lmm.types import Message
 from vision_agent.tools.tool_utils import get_tool_documentation
 from vision_agent.tools.tools import TOOL_DESCRIPTIONS
+from vision_agent.tools.tools_types import BboxInput, BboxInputBase64, PromptTask
 from vision_agent.utils.execute import Execution, MimeType
+from vision_agent.utils.image_utils import convert_to_b64
 
 # These tools are adapted from SWE-Agent https://github.com/princeton-nlp/SWE-agent
 
@@ -99,13 +104,14 @@ def load(self, file_path: Union[str, Path]) -> None:
 
  def show(self) -> str:
  """Shows the artifacts that have been loaded and their remote save paths."""
- out_str = "[Artifacts loaded]\n"
+ output_str = "[Artifacts loaded]\n"
  for k in self.artifacts.keys():
- out_str += (
+ output_str += (
  f"Artifact {k} loaded to {str(self.remote_save_path.parent / k)}\n"
  )
- out_str += "[End of artifacts]\n"
- return out_str
+ output_str += "[End of artifacts]\n"
+ print(output_str)
+ return output_str
 
  def save(self, local_path: Optional[Union[str, Path]] = None) -> None:
  save_path = (
@@ -135,7 +141,12 @@ def format_lines(lines: List[str], start_idx: int) -> str:
 
 
 def view_lines(
- lines: List[str], line_num: int, window_size: int, name: str, total_lines: int
+ lines: List[str],
+ line_num: int,
+ window_size: int,
+ name: str,
+ total_lines: int,
+ print_output: bool = True,
 ) -> str:
  start = max(0, line_num - window_size)
  end = min(len(lines), line_num + window_size)
@@ -148,7 +159,9 @@ def view_lines(
  else f"[{len(lines) - end} more lines]"
  )
  )
- print(return_str)
+
+ if print_output:
+ print(return_str)
  return return_str
 
 
@@ -231,7 +244,7 @@ def edit_code_artifact(
  new_content_lines = [
  line if line.endswith("\n") else line + "\n" for line in new_content_lines
  ]
- lines = artifacts[name].splitlines()
+ lines = artifacts[name].splitlines(keepends=True)
  edited_lines = lines[:start] + new_content_lines + lines[end:]
 
  cur_line = start + len(content.split("\n")) // 2
@@ -261,13 +274,20 @@ def edit_code_artifact(
  DEFAULT_WINDOW_SIZE,
  name,
  total_lines,
+ print_output=False,
  )
  total_lines_edit = sum(1 for _ in edited_lines)
  edited_view = view_lines(
- edited_lines, cur_line, DEFAULT_WINDOW_SIZE, name, total_lines_edit
+ edited_lines,
+ cur_line,
+ DEFAULT_WINDOW_SIZE,
+ name,
+ total_lines_edit,
+ print_output=False,
  )
 
  error_msg += f"\n[This is how your edit would have looked like if applied]\n{edited_view}\n\n[This is the original code before your edit]\n{original_view}"
+ print(error_msg)
  return error_msg
 
  artifacts[name] = "".join(edited_lines)
@@ -390,13 +410,122 @@ def write_media_artifact(artifacts: Artifacts, local_path: str) -> str:
  return f"[Media {Path(local_path).name} saved]"
 
 
+def list_artifacts(artifacts: Artifacts) -> str:
+ """Lists all the artifacts that have been loaded into the artifacts object."""
+ output_str = artifacts.show()
+ print(output_str)
+ return output_str
+
+
 def get_tool_descriptions() -> str:
  """Returns a description of all the tools that `generate_vision_code` has access to.
  Helpful for answering questions about what types of vision tasks you can do with
  `generate_vision_code`."""
  return TOOL_DESCRIPTIONS
 
 
+def florence2_fine_tuning(bboxes: List[Dict[str, Any]], task: str) -> str:
+ """'florence2_fine_tuning' is a tool that fine-tune florence2 to be able to detect
+ objects in an image based on a given dataset. It returns the fine tuning job id.
+
+ Parameters:
+ bboxes (List[BboxInput]): A list of BboxInput containing the
+ image path, labels and bounding boxes.
+ task (str): The florencev2 fine-tuning task. The options are
+ 'phrase_grounding'.
+
+ Returns:
+ UUID: The fine tuning job id, this id will used to retrieve the fine
+ tuned model.
+
+ Example
+ -------
+ >>> fine_tuning_job_id = florencev2_fine_tuning(
+ [{'image_path': 'filename.png', 'labels': ['screw'], 'bboxes': [[370, 30, 560, 290]]},
+ {'image_path': 'filename.png', 'labels': ['screw'], 'bboxes': [[120, 0, 300, 170]]}],
+ "phrase_grounding"
+ )
+ """
+ bboxes_input = [BboxInput.model_validate(bbox) for bbox in bboxes]
+ task_type = PromptTask[task.upper()]
+ fine_tuning_request = [
+ BboxInputBase64(
+ image=convert_to_b64(bbox_input.image_path),
+ filename=Path(bbox_input.image_path).name,
+ labels=bbox_input.labels,
+ bboxes=bbox_input.bboxes,
+ )
+ for bbox_input in bboxes_input
+ ]
+ landing_api = LandingPublicAPI()
+ fine_tune_id = str(
+ landing_api.launch_fine_tuning_job("florencev2", task_type, fine_tuning_request)
+ )
+ print(f"[Florence2 fine tuning id: {fine_tune_id}]")
+ return fine_tune_id
+
+
+def get_diff(before: str, after: str) -> str:
+ return "".join(
+ difflib.unified_diff(
+ before.splitlines(keepends=True), after.splitlines(keepends=True)
+ )
+ )
+
+
+def use_florence2_fine_tuning(
+ artifacts: Artifacts, name: str, task: str, fine_tune_id: str
+) -> str:
+ """Replaces florence2 calls with the fine tuning id. This ensures that the code
+ utilizes the fined tuned florence2 model. Returns the diff between the original
+ code and the new code.
+
+ Parameters:
+ artifacts (Artifacts): The artifacts object to edit the code from.
+ name (str): The name of the artifact to edit.
+ task (str): The task to fine tune the model for. The options are
+ 'phrase_grounding'.
+ fine_tune_id (str): The fine tuning job id.
+
+ Examples
+ --------
+ >>> diff = use_florence2_fine_tuning(artifacts, "code.py", "phrase_grounding", "23b3b022-5ebf-4798-9373-20ef36429abf")
+ """
+
+ task_to_fn = {"phrase_grounding": "florence2_phrase_grounding"}
+
+ if name not in artifacts:
+ output_str = f"[Artifact {name} does not exist]"
+ print(output_str)
+ return output_str
+
+ code = artifacts[name]
+ if task.lower() == "phrase_grounding":
+ pattern = r"florence2_phrase_grounding\(\s*([^\)]+)\)"
+
+ def replacer(match: re.Match) -> str:
+ arg = match.group(1) # capture all initial arguments
+ return f'florence2_phrase_grounding({arg}, "{fine_tune_id}")'
+
+ else:
+ raise ValueError(f"Task {task} is not supported.")
+
+ new_code = re.sub(pattern, replacer, code)
+
+ if new_code == code:
+ output_str = (
+ f"[Fine tuning task {task} function {task_to_fn[task]} not found in code]"
+ )
+ print(output_str)
+ return output_str
+
+ artifacts[name] = new_code
+
+ diff = get_diff(code, new_code)
+ print(diff)
+ return diff
+
+
 META_TOOL_DOCSTRING = get_tool_documentation(
  [
  get_tool_descriptions,
@@ -406,5 +535,8 @@ def get_tool_descriptions() -> str:
  generate_vision_code,
  edit_vision_code,
  write_media_artifact,
+ florence2_fine_tuning,
+ use_florence2_fine_tuning,
+ list_artifacts,
  ]
 )