2025.2 (#94)

* attempt to add cuda 12.4

* Update main.yml

* attempt to add cuda 12.4, update cudnn

* attempt to add cuda 12.4, update torch

* attempt to add cuda 12.4, update torch

* attempt to add cuda 12.4, downgrade tf to 2.17 and cudnn to 8.9

* attempt to add cuda 12.4, downgrade tf to 2.17 and cudnn to 9.1

* try cuda 12.6

* downgrade keras to 2.15

* downgrade keras to 2.15 + comment

* attempt to use tf-keras

* use tf-keras and keras, update tag, update tests for keras 3

Author: RockfordMankiniUCSD

.github/workflows/main.yml

@@ -142,7 +142,7 @@ jobs:
 
       - name: Archive artifacts and logs
         if: ${{ always() }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
           name: build-artifacts
           path: |

images/scipy-ml-notebook/Dockerfile

@@ -11,9 +11,11 @@ USER root
 ## Package versions
 ## tf 2.13 does not work with torch 2.2.1. Both require conflicting versions of typing-extensions
 ## CUDA 11: ARG CUDA_VERSION=11.8 CUDNN_VERSION=8.7.0.84 \
-ARG CUDA_VERSION=12.0 CUDNN_VERSION=8.9.2.26 LIBNVINFER=7.2.2 LIBNVINFER_MAJOR_VERSION=7 \
-  TENSORFLOW_VERSION=2.15.0 KERAS_VERSION=2.15.0 TENSORRT_VERSION=8.6.1 TORCH_VERSION=2.2.1 \
-  PROTOBUF_VERSION=3.20.3 
+
+## keras 3>= doesn't work with transformers as of 3-5-25. Check this when building future stable images.
+ARG CUDA_VERSION=12.6 CUDNN_VERSION=9.5.1.17 LIBNVINFER=7.2.2 LIBNVINFER_MAJOR_VERSION=7 \
+  TENSORFLOW_VERSION=2.18.0 KERAS_VERSION=3.9.0 TENSORRT_VERSION=10.7.0 TORCH_VERSION=2.6.0 \
+  PROTOBUF_VERSION=3.20.3 TF_KERAS_VERSION=2.18.0
 
 # apt deps
 RUN apt-get update && \
@@ -51,7 +53,7 @@ USER jovyan
 # CUDA setup w/mamba
 ## TODO: Investigate this command, seems to duplicate cuda packages for nvidia (pypi + conda-forge). 
 # cuda-toolkit is a skeleton package on CUDA 12, unlike CUDA <= 11
-RUN mamba install -c "nvidia/label/cuda-12.0.0" cuda-nvcc \
+RUN mamba install -c "nvidia/label/cuda-12.6.0" cuda-nvcc \
     cuda-toolkit=$CUDA_VERSION \
     # For CUDA 11: cudatoolkit=$CUDA_VERSION \
     cuda-version=$CUDA_VERSION \
@@ -84,8 +86,10 @@ RUN mamba install -c conda-forge pyqt pycocotools pillow scapy && \
 ## no purge required but no-cache-dir is used. pip purge will actually break the build here!
 ## Beware of potentially needing to update these if we update the drivers.
 ## Check tensorrt_env_vars.sh if you have to bump tensorrt!
-RUN pip install nvidia-cudnn-cu12==$CUDNN_VERSION torch==$TORCH_VERSION torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 && \
-  pip install tensorflow==$TENSORFLOW_VERSION tensorflow-datasets tensorrt==$TENSORRT_VERSION keras==$KERAS_VERSION && \
+
+## tf-keras is keras 2.x for higher versions of tensorflow that would normally require keras 3
+RUN pip install nvidia-cudnn-cu12==$CUDNN_VERSION torch==$TORCH_VERSION torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126 && \
+  pip install tensorflow==$TENSORFLOW_VERSION tensorflow-datasets tensorrt==$TENSORRT_VERSION keras==$KERAS_VERSION tf-keras==$TF_KERAS_VERSION && \
   fix-permissions $CONDA_DIR && \
   fix-permissions /home/$NB_USER && \
   mamba clean -a -y && \

…ebook/workflow_tests/test_huggingface.py …ook/workflow_tests/test_0_huggingface.pyimages/scipy-ml-notebook/workflow_tests/test_huggingface.py renamed to images/scipy-ml-notebook/workflow_tests/test_0_huggingface.py images/scipy-ml-notebook/workflow_tests/test_huggingface.py renamed to images/scipy-ml-notebook/workflow_tests/test_0_huggingface.py

@@ -15,8 +15,8 @@ def test_model_training(simple_model):
     x_train = np.random.random((100, 5))
     y_train = np.random.randint(2, size=(100, 1))
     simple_model.fit(x_train, y_train, epochs=1, batch_size=32, verbose=0)
-    assert simple_model.layers[0].input_shape == (None, 5)
-    assert simple_model.layers[1].output_shape == (None, 1)
+    assert simple_model.layers[0].input.shape == (None, 5)
+    assert simple_model.layers[1].output.shape == (None, 1)
 
 def test_model_evaluation(simple_model):
     x_test = np.random.random((20, 5))

…otebook/workflow_tests/test_0_pytorch.py …otebook/workflow_tests/test_1_pytorch.pyimages/scipy-ml-notebook/workflow_tests/test_0_pytorch.py renamed to images/scipy-ml-notebook/workflow_tests/test_1_pytorch.py images/scipy-ml-notebook/workflow_tests/test_0_pytorch.py renamed to images/scipy-ml-notebook/workflow_tests/test_1_pytorch.py

@@ -25,7 +25,7 @@ images:
     #prepull: false #-- uncomment to disable prepulling behavior for scipy-ml. gives you space on machine in exchange for build time.
 
 tag:
-  prefix: "2025.1"
+  prefix: "2025.2"
 
 all_info_cmds:
   PY_VER: