diff --git a/proto/inference.proto b/proto/inference.proto
index 39845dc5..59ae4d26 100644
--- a/proto/inference.proto
+++ b/proto/inference.proto
@@ -2,6 +2,7 @@ syntax = "proto3";
 
 service Inference {
   rpc CreateModelSession(CreateModelSessionRequest) returns (ModelSession) {}
+
   rpc CloseModelSession(ModelSession) returns (Empty) {}
 
   rpc CreateDatasetDescription(CreateDatasetDescriptionRequest) returns (DatasetDescription) {}
@@ -56,34 +57,6 @@ message NamedFloats {
 }
 
 
-/* InputShape will always be expected to have `shape` set.
- * For `ShapeType` PARAMETRIZED, also a `stepShape` has to be given.
- * ref: https://github.com/bioimage-io/spec-bioimage-io/blob/gh-pages/model_spec_latest.md */
-message InputShape {
-  enum ShapeType {
-    EXPLICIT = 0;
-    PARAMETRIZED = 1;
-  }
-
-  ShapeType shapeType = 1;
-  // shape is min, when PARAMETRIZED
-  NamedInts shape = 2;
-  NamedInts stepShape = 4;
-}
-
-message OutputShape {
-  enum ShapeType {
-    EXPLICIT = 0;
-    IMPLICIT = 1;
-  }
-  ShapeType shapeType = 1;
-  NamedInts shape = 2;
-  NamedInts halo = 3;
-  string referenceTensor = 4;
-  NamedFloats scale = 5;
-  NamedFloats offset = 6;
-}
-
 message ModelSession {
   string id = 1;
 }
@@ -136,16 +109,11 @@ message PredictResponse {
 
 message Empty {}
 
-
 service FlightControl {
   rpc Ping(Empty) returns (Empty) {}
   rpc Shutdown(Empty) returns (Empty) {}
 }
 
-message ModelInfo {
-    repeated string deviceIds = 1;
-}
-
 message CreateModelSessionChunkedRequest {
   oneof data {
     ModelInfo info = 1;
diff --git a/tests/data/__init__.py b/tests/data/__init__.py
deleted file mode 100644
index e69de29b..00000000
diff --git a/tests/data/dummy/Dummy.model.yaml b/tests/data/dummy/Dummy.model.yaml
deleted file mode 100644
index 1aa8fe90..00000000
--- a/tests/data/dummy/Dummy.model.yaml
+++ /dev/null
@@ -1,56 +0,0 @@
-format_version: 0.3.3
-language: python
-framework: pytorch
-
-name: UNet2DNucleiBroad
-description: A 2d U-Net pretrained on broad nucleus dataset.
-cite:
-    - text: "Ronneberger, Olaf et al. U-net: Convolutional networks for biomedical image segmentation. MICCAI 2015."
-      doi: https://doi.org/10.1007/978-3-319-24574-4_28
-authors:
-  - name: "ilastik-team"
-    affiliation: "EMBL Heidelberg"
-
-documentation: dummy.md
-tags: [pytorch, nucleus-segmentation]
-license: MIT
-git_repo: https://github.com/ilastik/tiktorch
-covers: []
-
-source: dummy.py::Dummy
-sha256: 00ffb1647cf7ec524892206dce6258d9da498fe040c62838f31b501a09bfd573
-timestamp: 2019-12-11T12:22:32Z  # ISO 8601
-
-test_inputs: [dummy_in.npy]
-test_outputs: [dummy_out.npy]
-
-weights:
-  pytorch_state_dict:
-    source: ./weights
-    sha256: 518cb80bad2eb3ec3dfbe6bab74920951391ce8fb24e15cf59b9b9f052a575a6
-    authors:
-     - name: "ilastik-team"
-       affiliation: "EMBL Heidelberg"
-
-
-# TODO double check inputs/outputs
-inputs:
-  - name: input
-    axes: bcyx
-    data_type: float32
-    data_range: [-inf, inf]
-    shape: [1, 1, 128, 128]
-
-
-outputs:
-  - name: output
-    axes: bcyx
-    data_type: float32
-    data_range: [0, 1]
-    shape:
-        reference_tensor: input
-        scale: [1, 1, 1, 1]
-        offset: [0, 0, 0, 0]
-    halo: [0, 0, 32, 32]
-
-type: model
diff --git a/tests/data/dummy/dummy.md b/tests/data/dummy/dummy.md
deleted file mode 100644
index e69de29b..00000000
diff --git a/tests/data/dummy/dummy.py b/tests/data/dummy/dummy.py
deleted file mode 100644
index 195e98e3..00000000
--- a/tests/data/dummy/dummy.py
+++ /dev/null
@@ -1,7 +0,0 @@
-from torch import nn
-
-
-class Dummy(nn.Module):
-    def forward(self, input):
-        x = input
-        return x + 1
diff --git a/tests/data/dummy/dummy_in.npy b/tests/data/dummy/dummy_in.npy
deleted file mode 100644
index 96a78a7b..00000000
Binary files a/tests/data/dummy/dummy_in.npy and /dev/null differ
diff --git a/tests/data/dummy/dummy_out.npy b/tests/data/dummy/dummy_out.npy
deleted file mode 100644
index 56f76ca7..00000000
Binary files a/tests/data/dummy/dummy_out.npy and /dev/null differ
diff --git a/tests/data/dummy/environment.yaml b/tests/data/dummy/environment.yaml
deleted file mode 100644
index e69de29b..00000000
diff --git a/tests/data/dummy/weights b/tests/data/dummy/weights
deleted file mode 100644
index da14f342..00000000
Binary files a/tests/data/dummy/weights and /dev/null differ
diff --git a/tests/data/dummy_param/Dummy.model_param.yaml b/tests/data/dummy_param/Dummy.model_param.yaml
deleted file mode 100644
index 87ddd885..00000000
--- a/tests/data/dummy_param/Dummy.model_param.yaml
+++ /dev/null
@@ -1,57 +0,0 @@
-format_version: 0.3.3
-language: python
-framework: pytorch
-
-name: UNet2DNucleiBroad
-description: A 2d U-Net pretrained on broad nucleus dataset.
-cite:
-    - text: "Ronneberger, Olaf et al. U-net: Convolutional networks for biomedical image segmentation. MICCAI 2015."
-      doi: https://doi.org/10.1007/978-3-319-24574-4_28
-authors:
-  - name: "ilastik-team"
-    affiliation: "EMBL Heidelberg"
-
-documentation: dummy.md
-tags: [pytorch, nucleus-segmentation]
-license: MIT
-git_repo: https://github.com/ilastik/tiktorch
-covers: []
-
-source: dummy.py::Dummy
-sha256: 00ffb1647cf7ec524892206dce6258d9da498fe040c62838f31b501a09bfd573
-timestamp: 2019-12-11T12:22:32Z  # ISO 8601
-
-test_inputs: [dummy_in.npy]
-test_outputs: [dummy_out.npy]
-
-weights:
-  pytorch_state_dict:
-    source: ./weights
-    sha256: 518cb80bad2eb3ec3dfbe6bab74920951391ce8fb24e15cf59b9b9f052a575a6
-    authors:
-     - name: "ilastik-team"
-       affiliation: "EMBL Heidelberg"
-
-
-# TODO double check inputs/outputs
-inputs:
-  - name: param
-    axes: bcyx
-    data_type: float32
-    data_range: [-inf, inf]
-    shape:
-      min: [1, 1, 64, 64]
-      step: [0, 0, 2, 1]
-
-outputs:
-  - name: output
-    axes: bcyx
-    data_type: float32
-    data_range: [0, 1]
-    shape:
-        reference_tensor: param
-        scale: [1, 1, 1, 1]
-        offset: [0, 0, 0, 0]
-    halo: [0, 0, 8, 8]
-
-type: model
diff --git a/tests/data/dummy_param/dummy.md b/tests/data/dummy_param/dummy.md
deleted file mode 100644
index e69de29b..00000000
diff --git a/tests/data/dummy_param/dummy.py b/tests/data/dummy_param/dummy.py
deleted file mode 100644
index 195e98e3..00000000
--- a/tests/data/dummy_param/dummy.py
+++ /dev/null
@@ -1,7 +0,0 @@
-from torch import nn
-
-
-class Dummy(nn.Module):
-    def forward(self, input):
-        x = input
-        return x + 1
diff --git a/tests/data/dummy_param/dummy_in.npy b/tests/data/dummy_param/dummy_in.npy
deleted file mode 100644
index 96a78a7b..00000000
Binary files a/tests/data/dummy_param/dummy_in.npy and /dev/null differ
diff --git a/tests/data/dummy_param/dummy_out.npy b/tests/data/dummy_param/dummy_out.npy
deleted file mode 100644
index 56f76ca7..00000000
Binary files a/tests/data/dummy_param/dummy_out.npy and /dev/null differ
diff --git a/tests/data/dummy_param/environment.yaml b/tests/data/dummy_param/environment.yaml
deleted file mode 100644
index e69de29b..00000000
diff --git a/tests/data/dummy_param/weights b/tests/data/dummy_param/weights
deleted file mode 100644
index da14f342..00000000
Binary files a/tests/data/dummy_param/weights and /dev/null differ
diff --git a/tests/data/dummy_tensorflow/cover.png b/tests/data/dummy_tensorflow/cover.png
deleted file mode 100644
index b22b5124..00000000
Binary files a/tests/data/dummy_tensorflow/cover.png and /dev/null differ
diff --git a/tests/data/dummy_tensorflow/dummy-model-weights.zip b/tests/data/dummy_tensorflow/dummy-model-weights.zip
deleted file mode 100644
index 96df58e3..00000000
Binary files a/tests/data/dummy_tensorflow/dummy-model-weights.zip and /dev/null differ
diff --git a/tests/data/dummy_tensorflow/rdf.yaml b/tests/data/dummy_tensorflow/rdf.yaml
deleted file mode 100644
index e5a8597f..00000000
--- a/tests/data/dummy_tensorflow/rdf.yaml
+++ /dev/null
@@ -1,34 +0,0 @@
-authors:
-- {name: ilastik-team}
-cite:
-- {text: text, url: pass}
-covers: [cover.png]
-description: simple model that increments input by 1
-documentation: tensorflow-dummy.md
-format_version: 0.4.3
-inputs:
-- axes: bcyx
-  data_range: [0.0, 255.0]
-  data_type: uint8
-  name: input0
-  shape: [1, 1, 32, 32]
-license: CC-BY-4.0
-name: dummy-keras
-outputs:
-- axes: bcyx
-  data_range: [-.inf, .inf]
-  data_type: float32
-  halo: [0, 0, 0, 0]
-  name: output0
-  shape:
-    offset: [0.0, 0.0, 0.0, 0.0]
-    reference_tensor: input0
-    scale: [1.0, 1.0, 1.0, 1.0]
-tags: [testing]
-test_inputs: [test_input.npy]
-test_outputs: [test_output.npy]
-timestamp: '2022-02-02T11:19:07.303147'
-type: model
-weights:
-  tensorflow_saved_model_bundle: {sha256: a3c86e3b2ec3766d29b635ae4422e0c12912be5af571a88ad3e44ba6c93f737c,
-    source: dummy-model-weights.zip, tensorflow_version: '1.14'}
diff --git a/tests/data/dummy_tensorflow/tensorflow-dummy.md b/tests/data/dummy_tensorflow/tensorflow-dummy.md
deleted file mode 100644
index 8b76d7aa..00000000
--- a/tests/data/dummy_tensorflow/tensorflow-dummy.md
+++ /dev/null
@@ -1,74 +0,0 @@
-# Simple tensorflow test model
-
-## Code to generate the model
-
-```python
-# Simple tensorflow test model
-# inverts input.
-
-## Code to generate the model
-
-```python
-import tensorflow as tf
-from tensorflow.keras.models import Sequential
-from tensorflow.keras.layers import Input, Lambda
-import numpy
-import xarray
-import shutil
-
-shutil.rmtree("./dummy-model")
-
-model = Sequential()
-model.add(Input(shape=(1, 32, 32), name="input0"))
-model.add(Lambda(lambda x: x * -1, name="output0"))
-tf.keras.experimental.export_saved_model(model, "./dummy-model")
-
-input_data = numpy.random.randint(0, 255, (1, 1, 32, 32), dtype="uint8")
-input_ = xarray.DataArray(input_data, dims=("b", "c", "y", "x"))
-
-output_ = model.predict(input_)
-
-numpy.save("test_input.npy", input_)
-numpy.save("test_output.npy", output_)
-```
-
-
-```Python
-# generate a model zoo model out of it
-from bioimageio.core.build_spec import build_model
-import shutil
-
-weight_file = "./dummy-model-weights.zip"
-name = "dummy-keras"
-input_axes = ["bcyx"]
-output_axes = ["bcyx"]
-zip_path = "./dummy-model-tensorflow.zip"
-
-
-shutil.make_archive("dummy-model-weights", "zip", root_dir="dummy-model")
-
-
-new_mod_raw = build_model(
-    weight_uri=weight_file,
-    test_inputs=["./test_input.npy"],
-    test_outputs=["./test_output.npy"],
-    input_axes=input_axes,
-    output_axes=output_axes,
-    output_path=zip_path,
-    name=name,
-    input_names=["input0"],
-    description="simple model that increments input by 1",
-    authors=[{"name": "ilastik-team"}],
-    documentation="./tensorflow-dummy.md",
-    tags=["testing"],
-    cite={"text": "pass"},
-    tensorflow_version="1.14",
-    weight_type="tensorflow_saved_model_bundle",
-    halo=[[0, 0, 0, 0]],
-    output_offset=[[0, 0, 0, 0]],
-    output_reference=["input0"],
-    output_scale=[[1.0, 1.0, 1.0, 1.0]],
-    output_names=["output0"],
-)
-
-```
diff --git a/tests/data/dummy_tensorflow/test_input.npy b/tests/data/dummy_tensorflow/test_input.npy
deleted file mode 100644
index 048f67e5..00000000
Binary files a/tests/data/dummy_tensorflow/test_input.npy and /dev/null differ
diff --git a/tests/data/dummy_tensorflow/test_output.npy b/tests/data/dummy_tensorflow/test_output.npy
deleted file mode 100644
index 093872dc..00000000
Binary files a/tests/data/dummy_tensorflow/test_output.npy and /dev/null differ
diff --git a/tests/data/tiny_models.py b/tests/data/tiny_models.py
deleted file mode 100644
index 72ac02ec..00000000
--- a/tests/data/tiny_models.py
+++ /dev/null
@@ -1,102 +0,0 @@
-import logging
-
-import torch
-
-logger = logging.getLogger(__name__)
-
-
-class TestModel0(torch.nn.Module):
-    def __init__(self, N=7, D_in=3, H=4, D_out=2):
-        super().__init__()
-        self.linear1 = torch.nn.Linear(D_in, H)
-        self.linear2 = torch.nn.Linear(H, D_out)
-
-        torch.manual_seed(0)
-        self.x = torch.randn(N, D_in)
-        self.y = self.forward(self.x).detach()
-
-    def forward(self, x):
-        h_relu = self.linear1(x).clamp(min=0)
-        y_pred = self.linear2(h_relu)
-        return y_pred
-
-
-class TinyConvNet2d(torch.nn.Module):
-    def __init__(self, in_channels=1, out_channels=1):
-        super().__init__()
-        self.conv1 = torch.nn.Conv2d(in_channels, 16, 1)
-        self.nlin1 = torch.nn.ReLU()
-        self.conv2 = torch.nn.Conv2d(16, 64, 1)
-        self.nlin2 = torch.nn.ReLU()
-        self.conv3 = torch.nn.Conv2d(64, out_channels, 1)
-        self.nlin3 = torch.nn.Sigmoid()
-
-    def forward(self, x):
-        return torch.nn.Sequential(self.conv1, self.nlin1, self.conv2, self.nlin2, self.conv3, self.nlin3)(x)
-
-
-class TinyConvNet3d(torch.nn.Module):
-    def __init__(self, in_channels=1, out_channels=1):
-        super().__init__()
-        self.conv1 = torch.nn.Conv3d(in_channels, 16, 1)
-        self.nlin1 = torch.nn.ReLU()
-        self.conv2 = torch.nn.Conv3d(16, 64, 1)
-        self.nlin2 = torch.nn.ReLU()
-        self.conv3 = torch.nn.Conv3d(64, out_channels, 1)
-        self.nlin3 = torch.nn.Sigmoid()
-
-    def forward(self, x):
-        return torch.nn.Sequential(self.conv1, self.nlin1, self.conv2, self.nlin2, self.conv3, self.nlin3)(x)
-
-
-def train(model, x, y):
-    loss_fn = torch.nn.MSELoss(reduction="sum")
-    optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
-
-    for t in range(10):
-        y_pred = model(x)
-
-        loss = loss_fn(y_pred, y)
-        logger.debug(t, loss.item())
-
-        optimizer.zero_grad()
-        loss.backward()
-        optimizer.step()
-
-
-if __name__ == "__main__":
-    test1 = TestModel0()
-    print("train", test1.training)
-    # test.to(dtype=torch.float)
-    y = test1(test1.x)
-    assert y.allclose(test1.y)
-
-    test2 = test1.__class__()
-    test2.eval()
-    test2.load_state_dict(test1.state_dict())
-    print("here train2", test2.training)
-    print(id(test1.state_dict()))
-    print(id(test2.state_dict()))
-
-    sd = test2.state_dict()
-    for key, value in sd.items():
-        sd[key] = torch.zeros_like(value)
-
-    print("train2", test2.training)
-
-    print(id(test1) == id(test2))
-    print(id(test1.state_dict()["linear1.weight"]) == id(test2.state_dict()["linear1.weight"]))
-    print(id(test1.state_dict()), test1.state_dict()["linear1.weight"])
-    print(id(test2.state_dict()), test2.state_dict()["linear1.weight"])
-    test2.load_state_dict(sd)
-    print(id(test1) == id(test2))
-    magic1 = test1.state_dict()
-    magic2 = test2.state_dict()
-    print(id(test1.state_dict()) == id(test2.state_dict()))
-    print(id(magic1) == id(magic2))
-    # print(id(magic1['linear1.weight']) == id(magic2['linear1.weight']))
-    # print(id(test1.state_dict()['linear1.weight']) == id(test2.state_dict()['linear1.weight']))
-    print(id(test1.state_dict()), test1.state_dict()["linear1.weight"])
-    print(id(test2.state_dict()), test2.state_dict()["linear1.weight"])
-
-    print("train2", test2.training)
diff --git a/tests/data/unet2d/UNet2DNucleiBroad.md b/tests/data/unet2d/UNet2DNucleiBroad.md
deleted file mode 100644
index e69de29b..00000000
diff --git a/tests/data/unet2d/UNet2DNucleiBroad.model.yaml b/tests/data/unet2d/UNet2DNucleiBroad.model.yaml
deleted file mode 100644
index 2da3d352..00000000
--- a/tests/data/unet2d/UNet2DNucleiBroad.model.yaml
+++ /dev/null
@@ -1,76 +0,0 @@
-# TODO physical scale of the data
-format_version: 0.3.3
-type: model
-
-name: UNet 2D Nuclei Broad
-description: A 2d U-Net trained on the nuclei broad dataset.
-authors:
-  - name: "Constantin Pape;@bioimage-io"
-    affiliation: "EMBL Heidelberg"
-    orcid: "0000-0001-6562-7187"
-  - name: "Fynn Beuttenmueller"
-    affiliation: "EMBL Heidelberg"
-    orcid: "0000-0002-8567-6389"
-
-# we allow for multiple citations. Each citation contains TEXT, DOI and URL. One of DOI or URL needs to be given.
-cite:
-  - text: "Ronneberger, Olaf et al. U-net: Convolutional networks for biomedical image segmentation. MICCAI 2015."
-    doi: https://doi.org/10.1007/978-3-319-24574-4_28
-  - text: "2018 Data Science Bowl"
-    url: https://www.kaggle.com/c/data-science-bowl-2018
-
-git_repo: https://github.com/bioimage-io/spec-bioimage-io/tree/main/example_specs/models/unet2d_nuclei_broad
-tags: [unet2d, pytorch, nucleus, segmentation, dsb2018]
-license: MIT
-
-documentation: UNet2DNucleiBroad.md
-covers: []
-attachments: {}
-timestamp: 2019-12-11T12:22:32Z  # ISO 8601
-
-inputs:
-  - name: raw
-    description: raw input
-    axes: bcyx  # letters of axes in btczyx
-    data_type: float32
-    data_range: [-.inf, .inf]
-    shape: [1, 1, 512, 512]
-    preprocessing:  # list of preprocessing steps
-      - name: zero_mean_unit_variance  # name of preprocessing step
-        kwargs:
-          mode: per_sample      # mode in [fixed, per_dataset, per_sample]
-          axes: yx              # subset of axes to normalize jointly, batch ('b') is not a valid axis key here!
-
-outputs:
-  - name: probability
-    description: probability in [0,1]
-    axes: bcyx
-    data_type: float32
-    data_range: [-.inf, .inf]
-    halo: [0, 0, 32, 32]
-    shape:
-      reference_tensor: raw
-      scale: [1.0, 1.0, 1.0, 1.0]
-      offset: [0, 0, 0, 0]
-
-language: python
-framework: pytorch
-source: unet2d.py:UNet2d
-sha256: cf42a6d86adeb4eb6e8e37b539a20e5413866b183bed88f4e2e26ad1639761ed
-kwargs: {input_channels: 1, output_channels: 1}
-dependencies: conda:./environment.yaml
-
-test_inputs: [test_input.npy]
-test_outputs: [test_output.npy]
-
-sample_inputs: [test_input.npy]
-sample_outputs: [test_output.npy]
-
-weights:
- pytorch_state_dict:
-   authors:
-     - name: "Constantin Pape;@bioimage-io"
-       affiliation: "EMBL Heidelberg"
-       orcid: "0000-0001-6562-7187"
-   sha256: e4d3885bccbe41cbf6c1d825f3cd2b707c7021ead5593156007e407a16b27cf2
-   source: https://zenodo.org/record/3446812/files/unet2d_weights.torch
\ No newline at end of file
diff --git a/tests/data/unet2d/environment.yaml b/tests/data/unet2d/environment.yaml
deleted file mode 100644
index 4297c6ca..00000000
--- a/tests/data/unet2d/environment.yaml
+++ /dev/null
@@ -1,8 +0,0 @@
-name:
-    unet2d_nuclei_broad
-channels:
-    - conda-forge
-    - defaults
-dependencies:
-    - pytorch
-    - numpy
diff --git a/tests/data/unet2d/test_input.npy b/tests/data/unet2d/test_input.npy
deleted file mode 100644
index 228057f8..00000000
Binary files a/tests/data/unet2d/test_input.npy and /dev/null differ
diff --git a/tests/data/unet2d/test_output.npy b/tests/data/unet2d/test_output.npy
deleted file mode 100644
index 4aea8257..00000000
Binary files a/tests/data/unet2d/test_output.npy and /dev/null differ
diff --git a/tests/data/unet2d/unet2d.py b/tests/data/unet2d/unet2d.py
deleted file mode 100644
index fedc9be1..00000000
--- a/tests/data/unet2d/unet2d.py
+++ /dev/null
@@ -1,73 +0,0 @@
-import torch
-import torch.nn as nn
-
-
-class Upsample(nn.Module):
-    def __init__(self, scale_factor, mode="bilinear"):
-        super().__init__()
-        self.scale_factor = scale_factor
-        self.mode = mode
-
-    def forward(self, input):
-        return nn.functional.interpolate(input, scale_factor=self.scale_factor, mode=self.mode, align_corners=False)
-
-
-class UNet2d(nn.Module):
-    def __init__(self, input_channels, output_channels, training=False):
-        super().__init__()
-        self.input_channels = input_channels
-        self.output_channels = output_channels
-        self.n_levels = 3
-
-        self.encoders = nn.ModuleList(
-            [self.conv_layer(self.input_channels, 16), self.conv_layer(16, 32), self.conv_layer(32, 64)]
-        )
-        self.downsamplers = nn.ModuleList([self.downsampler()] * self.n_levels)
-
-        self.base = self.conv_layer(64, 128)
-
-        self.decoders = nn.ModuleList([self.conv_layer(128, 64), self.conv_layer(64, 32), self.conv_layer(32, 16)])
-        self.upsamplers = nn.ModuleList([self.upsampler(128, 64), self.upsampler(64, 32), self.upsampler(32, 16)])
-
-        self.output = nn.Conv2d(16, self.output_channels, 1)
-        self.training = training
-
-    def conv_layer(self, in_channels, out_channels):
-        kernel_size = 3
-        padding = 1
-        return nn.Sequential(
-            nn.Conv2d(in_channels, out_channels, kernel_size, padding=padding),
-            nn.Conv2d(out_channels, out_channels, kernel_size, padding=padding),
-            nn.ReLU(inplace=True),
-        )
-
-    def downsampler(self):
-        return nn.MaxPool2d(2)
-
-    def upsampler(self, in_channels, out_channels):
-        return nn.Sequential(Upsample(2), nn.Conv2d(in_channels, out_channels, 1))
-
-    def forward(self, input):
-        x = input
-
-        from_encoder = []
-        for encoder, sampler in zip(self.encoders, self.downsamplers):
-            x = encoder(x)
-            from_encoder.append(x)
-            x = sampler(x)
-
-        x = self.base(x)
-
-        for decoder, sampler, enc in zip(self.decoders, self.upsamplers, from_encoder[::-1]):
-            x = sampler(x)
-            x = torch.cat([enc, x], dim=1)
-            x = decoder(x)
-
-        x = self.output(x)
-
-        # apply a sigmoid directly if we are in inference mode
-        if not self.training:
-            # postprocessing
-            x = torch.sigmoid(x)
-
-        return x
diff --git a/tests/data/unet2d_onnx/UNet2DNucleiBroad.md b/tests/data/unet2d_onnx/UNet2DNucleiBroad.md
deleted file mode 100644
index e69de29b..00000000
diff --git a/tests/data/unet2d_onnx/UNet2DNucleiBroad.model.yaml b/tests/data/unet2d_onnx/UNet2DNucleiBroad.model.yaml
deleted file mode 100644
index 7225c75d..00000000
--- a/tests/data/unet2d_onnx/UNet2DNucleiBroad.model.yaml
+++ /dev/null
@@ -1,73 +0,0 @@
-# TODO physical scale of the data
-format_version: 0.3.0
-
-name: UNet 2D Nuclei Broad
-description: A 2d U-Net trained on the nuclei broad dataset.
-authors:
-  - Constantin Pape;@bioimage-io
-  - Fynn Beuttenmüller
-
-# we allow for multiple citations. Each citation contains TEXT, DOI and URL. One of DOI or URL needs to be given.
-cite:
-  - text: "Ronneberger, Olaf et al. U-net: Convolutional networks for biomedical image segmentation. MICCAI 2015."
-    doi: https://doi.org/10.1007/978-3-319-24574-4_28
-  - text: "2018 Data Science Bowl"
-    url: https://www.kaggle.com/c/data-science-bowl-2018
-
-git_repo: https://github.com/bioimage-io/pytorch-bioimage-io/tree/master/specs/models/unet2d
-tags: [unet2d, pytorch, nucleus, segmentation, dsb2018]
-license: MIT
-
-documentation: UNet2DNucleiBroad.md
-covers: []  # todo unet 2d arch image
-attachments: {}
-timestamp: 2019-12-11T12:22:32Z  # ISO 8601
-
-inputs:
-  - name: raw
-    description: raw input
-    axes: bcyx  # letters of axes in btczyx
-    data_type: float32
-    data_range: [-inf, inf]
-    shape: [1, 1, 512, 512]
-    preprocessing:  # list of preprocessing steps
-      - name: zero_mean_unit_variance  # name of preprocessing step
-        kwargs:
-          mode: per_sample      # mode in [fixed, per_dataset, per_sample]
-          axes: yx              # subset of axes to normalize jointly, batch ('b') is not a valid axis key here!
-#      # example for zero_mean_unit_variance with mode 'fixed'
-#      - name: zero_mean_unit_variance
-#        kwargs:  # example kwargs for zero_mean_unit_variance
-#          mode: fixed
-#          axes: xy      # subset of axes to reduce
-#          mean: [262.3] # mean if mode == fixed. An array as nested lists. Its shape equals the input shape without the reduced axes.
-#          std: [220.2]  # standard deviation if mode == fixed analogously to mean
-
-outputs:
-  - name: probability
-    description: probability in [0,1]
-    axes: bcyx
-    data_type: float32
-    data_range: [-inf, inf]
-    halo: [0, 0, 32, 32]
-    shape:
-      reference_input: raw
-      scale: [1, 1, 1, 1]
-      offset: [0, 0, 0, 0]
-
-language: python
-framework: pytorch
-source: bioimageio.torch.models.unet2d.UNet2d
-kwargs: {input_channels: 1, output_channels: 1}
-dependencies: conda:../environment.yaml
-
-test_inputs: [test_input.npy]
-test_outputs: [test_output.npy]
-
-sample_inputs: [sample_input.npy]
-sample_outputs: [sample_output.npy]
-
-weights:
-  onnx:
-    sha256: 6d180e39527789d8a17b39a6eb78405bdeb58c3e6c2cf8d5a2c6dcc3ac5dcc7f
-    source: ./weights.onnx
diff --git a/tests/data/unet2d_onnx/cover0.png b/tests/data/unet2d_onnx/cover0.png
deleted file mode 100644
index cac369cb..00000000
Binary files a/tests/data/unet2d_onnx/cover0.png and /dev/null differ
diff --git a/tests/data/unet2d_onnx/sample_input.tif b/tests/data/unet2d_onnx/sample_input.tif
deleted file mode 100644
index 640f1a75..00000000
Binary files a/tests/data/unet2d_onnx/sample_input.tif and /dev/null differ
diff --git a/tests/data/unet2d_onnx/sample_output.tif b/tests/data/unet2d_onnx/sample_output.tif
deleted file mode 100644
index be0dfc2e..00000000
Binary files a/tests/data/unet2d_onnx/sample_output.tif and /dev/null differ
diff --git a/tests/data/unet2d_onnx/test_input.npy b/tests/data/unet2d_onnx/test_input.npy
deleted file mode 100644
index 228057f8..00000000
Binary files a/tests/data/unet2d_onnx/test_input.npy and /dev/null differ
diff --git a/tests/data/unet2d_onnx/test_output.npy b/tests/data/unet2d_onnx/test_output.npy
deleted file mode 100644
index 4aea8257..00000000
Binary files a/tests/data/unet2d_onnx/test_output.npy and /dev/null differ
diff --git a/tests/data/unet2d_onnx/weights.onnx b/tests/data/unet2d_onnx/weights.onnx
deleted file mode 100644
index 7ac57a89..00000000
Binary files a/tests/data/unet2d_onnx/weights.onnx and /dev/null differ
diff --git a/tests/data/unet2d_torchscript/UNet2DNucleiBroad.md b/tests/data/unet2d_torchscript/UNet2DNucleiBroad.md
deleted file mode 100644
index e69de29b..00000000
diff --git a/tests/data/unet2d_torchscript/UNet2DNucleiBroad.model.yaml b/tests/data/unet2d_torchscript/UNet2DNucleiBroad.model.yaml
deleted file mode 100644
index eeca5837..00000000
--- a/tests/data/unet2d_torchscript/UNet2DNucleiBroad.model.yaml
+++ /dev/null
@@ -1,66 +0,0 @@
-# TODO physical scale of the data
-format_version: 0.3.0
-
-name: UNet 2D Nuclei Broad
-description: A 2d U-Net trained on the nuclei broad dataset.
-authors:
-  - Constantin Pape;@bioimage-io
-  - Fynn Beuttenmüller
-
-# we allow for multiple citations. Each citation contains TEXT, DOI and URL. One of DOI or URL needs to be given.
-cite:
-  - text: "Ronneberger, Olaf et al. U-net: Convolutional networks for biomedical image segmentation. MICCAI 2015."
-    doi: https://doi.org/10.1007/978-3-319-24574-4_28
-  - text: "2018 Data Science Bowl"
-    url: https://www.kaggle.com/c/data-science-bowl-2018
-
-git_repo: https://github.com/bioimage-io/pytorch-bioimage-io/tree/master/specs/models/unet2d
-tags: [unet2d, pytorch, nucleus, segmentation, dsb2018]
-license: MIT
-
-documentation: UNet2DNucleiBroad.md
-covers: [cover0.png]
-attachments: {}
-timestamp: 2019-12-11T12:22:32Z  # ISO 8601
-
-inputs:
-  - name: raw
-    description: raw input
-    axes: bcyx  # letters of axes in btczyx
-    data_type: float32
-    data_range: [-inf, inf]
-    shape: [1, 1, 512, 512]
-    preprocessing:  # list of preprocessing steps
-      - name: zero_mean_unit_variance  # name of preprocessing step
-        kwargs:
-          mode: per_sample      # mode in [fixed, per_dataset, per_sample]
-          axes: yx              # subset of axes to normalize jointly, batch ('b') is not a valid axis key here!
-
-outputs:
-  - name: probability
-    description: probability in [0,1]
-    axes: bcyx
-    data_type: float32
-    data_range: [-inf, inf]
-    halo: [0, 0, 32, 32]
-    shape:
-      reference_input: raw
-      scale: [1, 1, 1, 1]
-      offset: [0, 0, 0, 0]
-
-language: python
-framework: pytorch
-source: bioimageio.torch.models.unet2d.UNet2d
-kwargs: {input_channels: 1, output_channels: 1}
-dependencies: conda:../environment.yaml
-
-test_inputs: [test_input.npy]
-test_outputs: [test_output.npy]
-
-sample_inputs: [sample_input.npy]
-sample_outputs: [sample_output.npy]
-
-weights:
-  pytorch_script:
-      sha256: b7f9dcf1da55a6d4cb29a0186d5558a86e4969916368479754517d00fa365848
-      source: ./weights.pt
diff --git a/tests/data/unet2d_torchscript/cover0.png b/tests/data/unet2d_torchscript/cover0.png
deleted file mode 100644
index cac369cb..00000000
Binary files a/tests/data/unet2d_torchscript/cover0.png and /dev/null differ
diff --git a/tests/data/unet2d_torchscript/test_input.npy b/tests/data/unet2d_torchscript/test_input.npy
deleted file mode 100644
index 228057f8..00000000
Binary files a/tests/data/unet2d_torchscript/test_input.npy and /dev/null differ
diff --git a/tests/data/unet2d_torchscript/test_output.npy b/tests/data/unet2d_torchscript/test_output.npy
deleted file mode 100644
index de18244e..00000000
Binary files a/tests/data/unet2d_torchscript/test_output.npy and /dev/null differ
diff --git a/tests/data/unet2d_torchscript/weights.pt b/tests/data/unet2d_torchscript/weights.pt
deleted file mode 100644
index abe68864..00000000
Binary files a/tests/data/unet2d_torchscript/weights.pt and /dev/null differ
diff --git a/tests/test_converters.py b/tests/test_converters.py
index be268e42..e112ede9 100644
--- a/tests/test_converters.py
+++ b/tests/test_converters.py
@@ -3,18 +3,7 @@
 import xarray as xr
 from numpy.testing import assert_array_equal
 
-from tiktorch.converters import (
-    NamedExplicitOutputShape,
-    NamedImplicitOutputShape,
-    NamedParametrizedShape,
-    Sample,
-    input_shape_to_pb_input_shape,
-    numpy_to_pb_tensor,
-    output_shape_to_pb_output_shape,
-    pb_tensor_to_numpy,
-    pb_tensor_to_xarray,
-    xarray_to_pb_tensor,
-)
+from tiktorch.converters import Sample, numpy_to_pb_tensor, pb_tensor_to_numpy, pb_tensor_to_xarray, xarray_to_pb_tensor
 from tiktorch.proto import inference_pb2
 
 
@@ -177,99 +166,6 @@ def test_should_same_data(self, shape):
         assert_array_equal(arr, result_arr)
 
 
-class TestShapeConversions:
-    def to_named_explicit_shape(self, shape, axes, halo):
-        return NamedExplicitOutputShape(
-            halo=[(name, dim) for name, dim in zip(axes, halo)], shape=[(name, dim) for name, dim in zip(axes, shape)]
-        )
-
-    def to_named_implicit_shape(self, axes, halo, offset, scales, reference_tensor):
-        return NamedImplicitOutputShape(
-            halo=[(name, dim) for name, dim in zip(axes, halo)],
-            offset=[(name, dim) for name, dim in zip(axes, offset)],
-            scale=[(name, scale) for name, scale in zip(axes, scales)],
-            reference_tensor=reference_tensor,
-        )
-
-    def to_named_paramtrized_shape(self, min_shape, axes, step):
-        return NamedParametrizedShape(
-            min_shape=[(name, dim) for name, dim in zip(axes, min_shape)],
-            step_shape=[(name, dim) for name, dim in zip(axes, step)],
-        )
-
-    @pytest.mark.parametrize(
-        "shape,axes,halo",
-        [((42,), "x", (0,)), ((42, 128, 5), "abc", (1, 1, 1)), ((5, 4, 3, 2, 1, 42), "btzyxc", (1, 2, 3, 4, 5, 24))],
-    )
-    def test_explicit_output_shape(self, shape, axes, halo):
-        named_shape = self.to_named_explicit_shape(shape, axes, halo)
-        pb_shape = output_shape_to_pb_output_shape(named_shape)
-
-        assert pb_shape.shapeType == 0
-        assert pb_shape.referenceTensor == ""
-        assert len(pb_shape.scale.namedFloats) == 0
-        assert len(pb_shape.offset.namedFloats) == 0
-
-        assert [(d.name, d.size) for d in pb_shape.halo.namedInts] == [(name, size) for name, size in zip(axes, halo)]
-        assert [(d.name, d.size) for d in pb_shape.shape.namedInts] == [(name, size) for name, size in zip(axes, shape)]
-
-    @pytest.mark.parametrize(
-        "axes,halo,offset,scales,reference_tensor",
-        [("x", (0,), (10,), (1.0,), "forty-two"), ("abc", (1, 1, 1), (1, 2, 3), (1.0, 2.0, 3.0), "helloworld")],
-    )
-    def test_implicit_output_shape(self, axes, halo, offset, scales, reference_tensor):
-        named_shape = self.to_named_implicit_shape(axes, halo, offset, scales, reference_tensor)
-        pb_shape = output_shape_to_pb_output_shape(named_shape)
-
-        assert pb_shape.shapeType == 1
-        assert pb_shape.referenceTensor == reference_tensor
-        assert [(d.name, d.size) for d in pb_shape.scale.namedFloats] == [
-            (name, size) for name, size in zip(axes, scales)
-        ]
-        assert [(d.name, d.size) for d in pb_shape.offset.namedFloats] == [
-            (name, size) for name, size in zip(axes, offset)
-        ]
-
-        assert [(d.name, d.size) for d in pb_shape.halo.namedInts] == [(name, size) for name, size in zip(axes, halo)]
-        assert len(pb_shape.shape.namedInts) == 0
-
-    def test_output_shape_raises(self):
-        shape = [("a", 1)]
-        with pytest.raises(TypeError):
-            _ = output_shape_to_pb_output_shape(shape)
-
-    @pytest.mark.parametrize(
-        "shape,axes",
-        [((42,), "x"), ((42, 128, 5), "abc"), ((5, 4, 3, 2, 1, 42), "btzyxc")],
-    )
-    def test_explicit_input_shape(self, shape, axes):
-        named_shape = [(name, dim) for name, dim in zip(axes, shape)]
-        pb_shape = input_shape_to_pb_input_shape(named_shape)
-
-        assert pb_shape.shapeType == 0
-        assert [(d.name, d.size) for d in pb_shape.shape.namedInts] == [(name, size) for name, size in zip(axes, shape)]
-
-    @pytest.mark.parametrize(
-        "min_shape,axes,step",
-        [
-            ((42,), "x", (5,)),
-            ((42, 128, 5), "abc", (1, 2, 3)),
-            ((5, 4, 3, 2, 1, 42), "btzyxc", (15, 24, 33, 42, 51, 642)),
-        ],
-    )
-    def test_parametrized_input_shape(self, min_shape, axes, step):
-        named_shape = self.to_named_paramtrized_shape(min_shape, axes, step)
-        pb_shape = input_shape_to_pb_input_shape(named_shape)
-
-        assert pb_shape.shapeType == 1
-        assert [(d.name, d.size) for d in pb_shape.shape.namedInts] == [
-            (name, size) for name, size in zip(axes, min_shape)
-        ]
-        assert [(d.name, d.size) for d in pb_shape.stepShape.namedInts] == [
-            (name, size) for name, size in zip(axes, step)
-        ]
-
-
 class TestSample:
     def test_create_sample_from_pb_tensors(self):
         arr_1 = np.arange(32 * 32, dtype=np.int64).reshape(32, 32)
diff --git a/tiktorch/converters.py b/tiktorch/converters.py
index b7fee3fe..5ce890a7 100644
--- a/tiktorch/converters.py
+++ b/tiktorch/converters.py
@@ -1,39 +1,13 @@
 from __future__ import annotations
 
 import dataclasses
-from typing import Dict, List, Tuple, Union
+from typing import Dict, List
 
 import numpy as np
 import xarray as xr
 
 from tiktorch.proto import inference_pb2
 
-# pairs of axis-shape for a single tensor
-NamedInt = Tuple[str, int]
-NamedFloat = Tuple[str, float]
-NamedShape = List[NamedInt]
-NamedVec = List[NamedFloat]
-
-
-@dataclasses.dataclass
-class NamedParametrizedShape:
-    min_shape: NamedShape
-    step_shape: NamedShape
-
-
-@dataclasses.dataclass
-class NamedExplicitOutputShape:
-    shape: NamedShape
-    halo: NamedShape
-
-
-@dataclasses.dataclass
-class NamedImplicitOutputShape:
-    reference_tensor: str
-    offset: NamedShape
-    scale: NamedVec
-    halo: NamedShape
-
 
 @dataclasses.dataclass(frozen=True)
 class Sample:
@@ -77,41 +51,6 @@ def name_float_tuples_to_pb_NamedFloats(name_float_tuples) -> inference_pb2.Name
     )
 
 
-def input_shape_to_pb_input_shape(input_shape: Union[NamedShape, NamedParametrizedShape]) -> inference_pb2.InputShape:
-    if isinstance(input_shape, NamedParametrizedShape):
-        return inference_pb2.InputShape(
-            shapeType=1,
-            shape=name_int_tuples_to_pb_NamedInts(input_shape.min_shape),
-            stepShape=name_int_tuples_to_pb_NamedInts(input_shape.step_shape),
-        )
-    else:
-        return inference_pb2.InputShape(
-            shapeType=0,
-            shape=name_int_tuples_to_pb_NamedInts(input_shape),
-        )
-
-
-def output_shape_to_pb_output_shape(
-    output_shape: Union[NamedExplicitOutputShape, NamedImplicitOutputShape]
-) -> inference_pb2.InputShape:
-    if isinstance(output_shape, NamedImplicitOutputShape):
-        return inference_pb2.OutputShape(
-            shapeType=1,
-            halo=name_int_tuples_to_pb_NamedInts(output_shape.halo),
-            referenceTensor=output_shape.reference_tensor,
-            scale=name_float_tuples_to_pb_NamedFloats(output_shape.scale),
-            offset=name_float_tuples_to_pb_NamedFloats(output_shape.offset),
-        )
-    elif isinstance(output_shape, NamedExplicitOutputShape):
-        return inference_pb2.OutputShape(
-            shapeType=0,
-            shape=name_int_tuples_to_pb_NamedInts(output_shape.shape),
-            halo=name_int_tuples_to_pb_NamedInts(output_shape.halo),
-        )
-    else:
-        raise TypeError(f"Conversion not supported for type {type(output_shape)}")
-
-
 def pb_tensor_to_xarray(tensor: inference_pb2.Tensor) -> inference_pb2.Tensor:
     if not tensor.dtype:
         raise ValueError("Tensor dtype is not specified")