Fix railvision

.github/workflows/ci.yml

@@ -8,7 +8,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        os: [ubuntu-20.04]
+        os: [ubuntu-24.04]
     steps:
       - name: Check out repository
         uses: actions/checkout@v2
@@ -56,7 +56,7 @@ jobs:
     needs: python-setup
     strategy:
       matrix:
-        os: [ubuntu-20.04]
+        os: [ubuntu-24.04]
         pytest_target: [models, layers]
     name: test-${{ matrix.pytest_target }}
     steps:

onnx2kerastl/operation_layers.py

@@ -231,14 +231,19 @@ def convert_reduce_min(node, params, layers, lambda_func, node_name, keras_name)
         axes = params.get("axes")
     elif len(node.input) == 2:
         axes = layers.get(node.input[1])
+    else:
+        axes = None
     noop_with_empty_axes = bool(params.get("noop_with_empty_axes", False))
     keepdims = params.get("keepdims", True)
     if noop_with_empty_axes and params.get("axes") is None:
         layers[node_name] = layers[node.input[0]]
     else:
-        layers[node_name] = tf_math_reduce_min(layers[node.input[0]], axis=axes, keepdims=keepdims,
+        if axes is None:
+            layers[node_name] = tf_math_reduce_min(layers[node.input[0]], keepdims=keepdims,
+                                               tf_name=f"{params['cleaned_name']}_min")
+        else:
+            layers[node_name] = tf_math_reduce_min(layers[node.input[0]], axis=axes, keepdims=keepdims,
                                                tf_name=f"{params['cleaned_name']}_min")
-
 
 def convert_reduce_prod(node, params, layers, lambda_func, node_name, keras_name):
     """

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "onnx2kerastl"
-version = "0.0.165"
+version = "0.0.168"
 description = ""
 authors = ["dorhar <doron.harnoy@tensorleap.ai>"]
 license = "MIT"
@@ -17,7 +17,6 @@ fvcore = "^0.1.5.post20221221"
 boto3 = "^1.24.22"
 tensorflow-io-gcs-filesystem = "0.34.0"
 keras-data-format-converter = "0.1.22"
-optimum = "1.23.3"
 
 [tool.poetry.dev-dependencies]
 pytest = "^7.1.2"

test/models/private_tests/test_dinov2.py

@@ -0,0 +1,34 @@
+import numpy as np
+import onnx
+from onnx2kerastl import onnx_to_keras
+from keras_data_format_converter import convert_channels_first_to_last
+import onnxruntime as ort
+import torch
+import pytest
+from test.models.private_tests.aws_utils import aws_s3_download
+
+
+@pytest.mark.parametrize('aws_s3_download', [["dinov2/", "dinov2/", False]], indirect=True)
+def test_dinov2(aws_s3_download):
+    # This is commented out in case we'll upgrade python
+    # batch_size = 1
+    # dinov2_vits14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vits14')
+    # wm = wrapper_model(dinov2_vits14).to('cpu')
+    # wm.eval()
+    # dummy_input = torch.FloatTensor(np.random.uniform(0, 1, (batch_size, 3, 224, 224)))
+    # torch.onnx.export(wm, dummy_input, "dino-2-test.onnx", input_names=['img'],
+    #                   output_names=['vit_out'])
+    np_input = list(np.random.rand(1, 3, 224, 224))
+    onnx_path = f'{aws_s3_download}/dino-2-test.onnx'
+    onnx_model = onnx.load(onnx_path)
+    keras_model = onnx_to_keras(onnx_model, ['img', 'masks'], allow_partial_compilation=False)
+    flipped_model = convert_channels_first_to_last(keras_model.converted_model, should_transform_inputs_and_outputs=False)
+    ort_session = ort.InferenceSession(onnx_path)
+    keras_res = flipped_model(np.array(np_input))
+    res = ort_session.run(
+        ['vit_out'],
+        input_feed={"img": np.array(np_input).astype(np.float32)}
+    )
+    t_mean, t_max = (res[0]-keras_res).__abs__().numpy().mean(), (res[0]-keras_res).__abs__().numpy().max()
+    assert t_mean < 5e-2
+    assert t_max < 0.4

test/models/test_llama_sentiment_analysis.py

@@ -8,7 +8,7 @@
 from onnx2kerastl import onnx_to_keras
 from keras_data_format_converter import convert_channels_first_to_last
 from onnx2kerastl.customonnxlayer import onnx_custom_objects_map
-from test.utils import export_torch_to_onnx_optimum
+#from test.utils import export_torch_to_onnx_optimum
 
 
 @pytest.mark.skip(reason="Fails on CI but works locally (might be too big?)")

test/utils.py

@@ -7,7 +7,7 @@
 
 from onnx2kerastl import onnx_to_keras
 from onnx2kerastl.utils import check_torch_keras_error
-from optimum.exporters.onnx import main_export
+#from optimum.exporters.onnx import main_export
 
 NP_SEED = 42
 
@@ -58,31 +58,31 @@ def is_lambda_layers_exist(model: Model):
     return any(isinstance(layer, Lambda) for layer in model.layers)
 
 
-def export_torch_to_onnx_optimum(model_name: str, model_output_path: str, task="causal-lm"):
-    """
-    this function get a model as an input (Hugginface or local path), creates a folder and save the onnx model as output.
-    it uses the optimum library.
-    NOTE: For llama model the maximum absolute difference of the logits larget than 1e-5, it shouldnt be that important!
-    Args:
-        model_name: model path (local or HF name)
-        model_output_name: output folder path
-        task: model task
-
-    Returns:
-        creates the onnx model in the output folder path
-    """
-    main_export(
-        model_name_or_path=model_name,
-        task=task,
-        output=model_output_path,
-        opset=None,
-        device="cpu",
-        dtype=None,
-        pad_token_id=None,
-        trust_remote_code=False,
-        do_validation=True,
-        framework=None,
-        no_post_process=False,
-        model_kwargs=None,
-        atol = 1e-5
-    )
+# def export_torch_to_onnx_optimum(model_name: str, model_output_path: str, task="causal-lm"):
+#     """
+#     this function get a model as an input (Hugginface or local path), creates a folder and save the onnx model as output.
+#     it uses the optimum library.
+#     NOTE: For llama model the maximum absolute difference of the logits larget than 1e-5, it shouldnt be that important!
+#     Args:
+#         model_name: model path (local or HF name)
+#         model_output_name: output folder path
+#         task: model task
+#
+#     Returns:
+#         creates the onnx model in the output folder path
+#     """
+#     main_export(
+#         model_name_or_path=model_name,
+#         task=task,
+#         output=model_output_path,
+#         opset=None,
+#         device="cpu",
+#         dtype=None,
+#         pad_token_id=None,
+#         trust_remote_code=False,
+#         do_validation=True,
+#         framework=None,
+#         no_post_process=False,
+#         model_kwargs=None,
+#         atol = 1e-5
+#     )