[GSK-852] Change Tensorflow to latest version, to enable arm64 test again

Author: Hartorn

.github/workflows/build-images.yml

@@ -34,12 +34,17 @@ on:
 
   pull_request: # This will allow to trigger on PR only with a specific label
     types: [opened, reopened, synchronize, labeled, unlabeled]
-# https://docs.docker.com/build/ci/github-actions/multi-platform/#distribute-build-across-multiple-runners
+# Concurrency : auto-cancel "old" jobs ie when pushing again
+# https://docs.github.com/fr/actions/using-jobs/using-concurrency
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref || github.run_id }}
+  cancel-in-progress: true
 env:
   RUN_TESTS: false
   BUILD_ONLY: false
   REGISTRY_IMAGE: giskardai/giskard
   DOCKERHUB_USER: giskardai
+# https://docs.docker.com/build/ci/github-actions/multi-platform/#distribute-build-across-multiple-runners
 jobs:
   build-images:
     # Debug
@@ -130,13 +135,14 @@ jobs:
           username: ${{ env.DOCKERHUB_USER }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
 
-      - name: Run python test inside docker
+      - name: Install deps inside docker
         if: ${{ env.RUN_TESTS }}
         uses: docker/build-push-action@v5
         with:
           context: .
-          target: test-python
+          target: full-install-python
           push: false
+          load: true
           tags: ${{ steps.meta.outputs.tags }}
           labels: ${{ steps.meta.outputs.labels }}
           builder: ${{ steps.builder.outputs.name }}
@@ -145,20 +151,35 @@ jobs:
           cache-from: type=gha
           cache-to: type=gha,mode=max
 
-      - name: Run python integration test inside docker
+      - name: Run python test inside docker
         if: ${{ env.RUN_TESTS }}
         uses: docker/build-push-action@v5
         with:
           context: .
-          target: integration-test-python
+          target: test-python
           push: false
+          load: false
           tags: ${{ steps.meta.outputs.tags }}
           labels: ${{ steps.meta.outputs.labels }}
           builder: ${{ steps.builder.outputs.name }}
           platforms: |
             ${{ matrix.platform}}
           cache-from: type=gha
-          cache-to: type=gha,mode=max
+
+      # - name: Run python integration test inside docker
+      #   if: ${{ env.RUN_TESTS }}
+      #   uses: docker/build-push-action@v5
+      #   with:
+      #     context: .
+      #     target: integration-test-python
+      #     push: false
+      #     load: false
+      #     tags: ${{ steps.meta.outputs.tags }}
+      #     labels: ${{ steps.meta.outputs.labels }}
+      #     builder: ${{ steps.builder.outputs.name }}
+      #     platforms: |
+      #       ${{ matrix.platform}}
+      #     cache-from: type=gha
 
       - name: Build and push
         id: build

.github/workflows/build_backend.yml

@@ -14,50 +14,29 @@ on:
         required: true
         type: boolean
         default: false
+      is-dispatch:
+        description: 'Just to identify manual dispatch'
+        required: true
+        type: boolean
+        default: true  
   workflow_call:
     inputs:
       run-integration-tests:
         description: 'If integration test should be run'
         required: true
         type: boolean
         default: false
+# Concurrency : auto-cancel "old" jobs ie when pushing again
+# https://docs.github.com/fr/actions/using-jobs/using-concurrency
+concurrency:
+  group: ${{ github.workflow }}-${{ inputs.run-integration-tests }}-${{ inputs.is-dispatch }}-${{ github.ref || github.run_id }}
+  cancel-in-progress: true
 env:
   GSK_DISABLE_ANALYTICS: true
 defaults:
   run:
     shell: bash
 jobs:
-  pre-check:
-    name: Pre check
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 1 
-      # Inspired from https://blog.pantsbuild.org/skipping-github-actions-jobs-without-breaking-branch-protection/
-      - id: files
-        name: Get changed files outside python-client
-        uses: tj-actions/changed-files@v39
-        with:
-          files_ignore: python-client/**
-
-      - id: files-python
-        name: Get changed files in python-client
-        uses: tj-actions/changed-files@v39
-        with:
-          files: python-client/**
-  
-      - id: python_only
-        if: steps.files.outputs.any_changed != 'true'
-        name: Check for changes in python only 
-        run: echo 'python_only=PYTHON_ONLY' >> $GITHUB_OUTPUT
-
-      - id: python_at_least
-        if: steps.files-python.outputs.any_changed != 'true'
-        name: Check for changes in python
-        run: echo 'python_at_least=PYTHON_AT_LEAST' >> $GITHUB_OUTPUT
-
   sonar:
     if: ${{ github.actor != 'dependabot[bot]' && (github.event_name == 'pull_request' || github.event_name == 'push') }}
     name: Sonar
@@ -93,8 +72,6 @@ jobs:
         run: ./gradlew sonar --info --parallel
 
   build:
-    needs: pre-check
-    if: ${{ needs.pre-check.outputs.python_only != 'PYTHON_ONLY' }}
     name: Backend
     runs-on: ubuntu-latest
     steps:
@@ -147,8 +124,6 @@ jobs:
 
   build-python:
     name: Build Python
-    needs: pre-check
-    if: ${{ needs.pre-check.outputs.python_at_least != 'PYTHON_AT_LEAST' }}
     runs-on: ${{ matrix.os }}
     strategy:
       fail-fast: false # Do not stop when any job fails
@@ -228,4 +203,4 @@ jobs:
         working-directory: python-client
         env:
           PYTEST_XDIST_AUTO_NUM_WORKERS: ${{ matrix.os == 'windows-2019' && 1 || 2 }}
-        run: pdm run test
+        run: pdm run test -m "slow"

python-client/giskard/scanner/llm/detoxify.py

@@ -0,0 +1,29 @@
+import pandas as pd
+from transformers import pipeline
+
+
+# Re-implementation based on https://github.com/unitaryai/detoxify/issues/15#issuecomment-900443551
+class Detoxify:
+    def __init__(
+        self,
+    ):
+        super().__init__()
+        self.pipeline = pipeline(
+            "text-classification",
+            model="unitary/unbiased-toxic-roberta",
+            tokenizer="unitary/unbiased-toxic-roberta",
+            function_to_apply="sigmoid",
+            return_all_scores=True,
+        )
+
+    def predict(self, text) -> pd.DataFrame:
+        inputs = [text] if isinstance(text, str) else text
+        results = self.pipeline(inputs)
+        output = []
+        for one_result in results:
+            res = {}
+            for single_label_result in one_result:
+                res[single_label_result["label"]] = single_label_result["score"]
+            output.append(res)
+        output = pd.DataFrame(output)
+        return output

python-client/giskard/scanner/llm/harmfulness_detector.py

@@ -1,7 +1,6 @@
 from pathlib import Path
 from typing import List, Sequence
 
-import numpy as np
 import pandas as pd
 
 from ...datasets import Dataset
@@ -73,13 +72,11 @@ def run(self, model: LangchainModel, dataset: Dataset) -> Sequence[Issue]:
 
     def _compute_harmfulness(self, sentences: List[str]):
         try:
-            from detoxify import Detoxify
+            from giskard.scanner.llm.detoxify import Detoxify
         except ImportError as err:
             raise LLMImportError() from err
 
         keys = ["toxicity", "severe_toxicity", "identity_attack", "insult", "threat"]
-        results = Detoxify("unbiased").predict(list(sentences))
-
-        harmfulness = np.vstack([results[k] for k in keys]).max(axis=0)
-
+        results = Detoxify().predict(list(sentences))
+        harmfulness = results[keys].max(axis="columns")
         return harmfulness

python-client/giskard/scanner/llm/minority_stereotype_detector.py

@@ -79,10 +79,9 @@ def run(self, model: LangchainModel, dataset: Dataset) -> Sequence[Issue]:
 
     def _compute_bias(self, sentences: List[str]):
         try:
-            from detoxify import Detoxify
+            from giskard.scanner.llm.detoxify import Detoxify
         except ImportError as err:
             raise LLMImportError() from err
 
-        results = Detoxify("unbiased").predict(list(sentences))
-
-        return results["identity_attack"]
+        results = Detoxify().predict(list(sentences))
+        return results["identity_attack"]

python-client/giskard/scanner/llm/toxicity_detector.py

@@ -131,11 +131,11 @@ def run(self, model: LangchainModel, dataset: Dataset) -> Sequence[Issue]:
 
     def _compute_toxicity_score(self, sentences: List[str]):
         try:
-            from detoxify import Detoxify
+            from giskard.scanner.llm.detoxify import Detoxify
         except ImportError as err:
             raise LLMImportError() from err
-
-        return Detoxify("unbiased").predict(list(sentences))["toxicity"]
+        results = Detoxify().predict(list(sentences))
+        return results["toxicity"]
 
 
 @dataclass