[misc] update Python & PyTorch constraint (#371)

* [misc] update Python & PyTorch constraint

* update reqs

* update pillow

* use classic solver

* debugging

* debugging

* bump to pytorch 1.10

* update pillow

* update

* update ruff

* apply fixes for ruff

* fix style check

.github/workflows/full-test.yml

@@ -12,7 +12,7 @@ jobs:
       fail-fast: false
       matrix:
         os: ["ubuntu-latest", "macos-13", "windows-latest"]
-        vers: [ {pt_ver: "1.6.0", tv_ver: "0.7.0"}, {pt_ver: "1.7.0", tv_ver: "0.8.1"}, {pt_ver: "1.8.0", tv_ver: "0.9.0"}, {pt_ver: "1.9.0", tv_ver: "0.10.0"}, {pt_ver: "1.10.0", tv_ver: "0.11.1"}, {pt_ver: "1.11.0", tv_ver: "0.12.0"}, {pt_ver: "1.12.0", tv_ver: "0.13.0"} ]
+        vers: [ {pt_ver: "1.10.0", tv_ver: "0.11.1"}, {pt_ver: "1.11.0", tv_ver: "0.12.0"}, {pt_ver: "1.12.0", tv_ver: "0.13.0"} ]
         include:
           - os: macos-latest
             vers:
@@ -36,7 +36,7 @@ jobs:
       - uses: conda-incubator/setup-miniconda@v3
         with:
           auto-update-conda: true
-          python-version: 3.8
+          python-version: 3.9
       - name: Install PyTorch
         env:
           PYTORCH_VER: ${{ matrix.vers.pt_ver }}
@@ -45,19 +45,14 @@ jobs:
          if [ "$RUNNER_OS" == "macOS" ]; then
             if [[ "$TORCHVISION_VER" == "latest" && "$PYTORCH_VER" == "latest" ]]; then
               conda install pytorch::pytorch torchvision -c pytorch
-            elif [[ "$TORCHVISION_VER" == "0.9."* || "$TORCHVISION_VER" == "0.10."* ]]; then
-              conda install pytorch=$PYTORCH_VER torchvision=$TORCHVISION_VER pillow=6 -c pytorch
             else
               conda install pytorch=$PYTORCH_VER torchvision=$TORCHVISION_VER -c pytorch
             fi
          elif [ "$RUNNER_OS" == "Windows" ]; then
             if [[ "$TORCHVISION_VER" == "latest" && "$PYTORCH_VER" == "latest" ]]; then
-              conda install pytorch torchvision cpuonly pillow=6 -c pytorch
-            elif [[ "$TORCHVISION_VER" == "0.9."* || "$TORCHVISION_VER" == "0.10."* ]]; then
-              conda install pillow=6 -c conda-forge
-              conda install pytorch=$PYTORCH_VER torchvision=$TORCHVISION_VER cpuonly -c pytorch
+              conda install pytorch torchvision cpuonly pillow=8 -c pytorch
             else
-              conda install pytorch=$PYTORCH_VER torchvision=$TORCHVISION_VER cpuonly pillow=6 -c pytorch
+              conda install pytorch=$PYTORCH_VER torchvision=$TORCHVISION_VER cpuonly pillow=8 -c pytorch
             fi
          else
             if [[ "$TORCHVISION_VER" == "latest" && "$PYTORCH_VER" == "latest" ]]; then
@@ -72,14 +67,9 @@ jobs:
         env:
           PYTORCH_VER: ${{ matrix.vers.pt_ver }}
           TORCHVISION_VER: ${{ matrix.vers.tv_ver }}
-        run: |
-          if [[ "$RUNNER_OS" == "Linux" && "$TORCHVISION_VER" == "latest" && "$PYTORCH_VER" == "latest" ]]; then
-            pip install black 'ruff<0.0.234' 'tensorflow<2.12' pytest scipy interval
-          else
-            pip install black 'ruff<0.0.234' tensorflow pytest scipy interval
-          fi
+        run: pip install black==22.3.0 ruff tensorflow pytest scipy interval
       - name: Lint checks
-        run: python -m ruff .
+        run: python -m ruff check .
       - name: Run tests
         run: |
           source activate.sh

.github/workflows/smoke-test.yml

@@ -12,7 +12,7 @@ jobs:
       fail-fast: false
       matrix:
         os: ["ubuntu-latest", "macos-13", "macos-latest", "windows-latest"]
-        python-version: ["3.8", "3.9", "3.10"]
+        python-version: ["3.9", "3.10", "3.11"]
     defaults:
       run:
         shell: bash -l {0}

.github/workflows/unit-test.yml

@@ -16,7 +16,7 @@ jobs:
       fail-fast: false
       matrix:
         os: ["ubuntu-latest", "macos-13", "windows-latest"]
-        vers: [ {pt_ver: "1.6.0", tv_ver: "0.7.0"}, {pt_ver: "latest", tv_ver: "latest"} ]
+        vers: [ {pt_ver: "1.10.0", tv_ver: "0.11.1"}, {pt_ver: "latest", tv_ver: "latest"} ]
         include:
           - os: macos-latest
             vers:
@@ -30,7 +30,7 @@ jobs:
       - uses: conda-incubator/setup-miniconda@v3
         with:
           auto-update-conda: true
-          python-version: 3.8
+          python-version: 3.9
       - name: Install PyTorch
         env:
           PYTORCH_VER: ${{ matrix.vers.pt_ver }}
@@ -39,19 +39,14 @@ jobs:
          if [ "$RUNNER_OS" == "macOS" ]; then
             if [[ "$TORCHVISION_VER" == "latest" && "$PYTORCH_VER" == "latest" ]]; then
               conda install pytorch::pytorch torchvision -c pytorch
-            elif [[ "$TORCHVISION_VER" == "0.9."* || "$TORCHVISION_VER" == "0.10."* ]]; then
-              conda install pytorch=$PYTORCH_VER torchvision=$TORCHVISION_VER pillow=6 -c pytorch
             else
               conda install pytorch=$PYTORCH_VER torchvision=$TORCHVISION_VER -c pytorch
             fi
          elif [ "$RUNNER_OS" == "Windows" ]; then
             if [[ "$TORCHVISION_VER" == "latest" && "$PYTORCH_VER" == "latest" ]]; then
-              conda install pytorch torchvision cpuonly pillow=6 -c pytorch
-            elif [[ "$TORCHVISION_VER" == "0.9."* || "$TORCHVISION_VER" == "0.10."* ]]; then
-              conda install pillow=6 -c conda-forge
-              conda install pytorch=$PYTORCH_VER torchvision=$TORCHVISION_VER cpuonly -c pytorch
+              conda install pytorch torchvision cpuonly pillow=8 -c pytorch
             else
-              conda install pytorch=$PYTORCH_VER torchvision=$TORCHVISION_VER cpuonly pillow=6 -c pytorch
+              conda install pytorch=$PYTORCH_VER torchvision=$TORCHVISION_VER cpuonly pillow=8 -c pytorch
             fi
          else
             if [[ "$TORCHVISION_VER" == "latest" && "$PYTORCH_VER" == "latest" ]]; then
@@ -66,16 +61,11 @@ jobs:
         env:
           PYTORCH_VER: ${{ matrix.vers.pt_ver }}
           TORCHVISION_VER: ${{ matrix.vers.tv_ver }}
-        run: |
-          if [[ "$RUNNER_OS" == "Linux" && "$TORCHVISION_VER" == "latest" && "$PYTORCH_VER" == "latest" ]]; then
-            pip install black 'ruff<0.0.234' 'tensorflow<2.12' scipy interval
-          else
-            pip install black 'ruff<0.0.234' tensorflow scipy interval
-          fi
+        run: pip install black==22.3.0 ruff tensorflow scipy interval
       - name: Lint checks
-        run: python -m ruff .
+        run: python -m ruff check .
       - name: Style checks
-        run: python -m black .
+        run: python -m black --check .
       - name: Run unit tests
         run: |
           cd tests
@@ -109,7 +99,7 @@ jobs:
       - uses: conda-incubator/setup-miniconda@v2
         with:
           auto-update-conda: true
-          python-version: 3.8
+          python-version: 3.9
       - name: Install PyTorch
         run: conda install pytorch torchvision cpuonly -c pytorch
       - name: Install TinyNeuralNetwork

.pre-commit-config.yaml

@@ -11,6 +11,6 @@ repos:
     -   id: black
         exclude: ^tinynn/converter/schemas
 -   repo: https://github.com/charliermarsh/ruff-pre-commit
-    rev: v0.0.150
+    rev: v0.7.0
     hooks:
     -   id: ruff

README.md

@@ -5,7 +5,7 @@ TinyNeuralNetwork is an efficient and easy-to-use deep learning model compressio
 
 ## Installation
 
-Python >= 3.8, PyTorch >= 1.4（ PyTorch >= 1.6 if quantization-aware training is involved, see [here](docs/quantization_support.md) for details ）
+Python >= 3.9, PyTorch >= 1.10
 
 ```shell
 # Install the TinyNeuralNetwork framework

README_zh-CN.md

@@ -5,7 +5,7 @@ TinyNeuralNetwork是一个高效、易用的深度学习模型压缩框架。它
 
 ## 安装
 
-python >= 3.8, pytorch >= 1.4（如果使用量化训练 pytorch >= 1.6，详细可见[这里](docs/quantization_support.md) ）
+python >= 3.9, pytorch >= 1.10
 
 ```shell
 # 安装TinyNeuralNetwork软件包

pyproject.toml

@@ -75,7 +75,7 @@ exclude = [
 # Allow unused variables when underscore-prefixed.
 dummy-variable-rgx = "^(_+|(_+[a-zA-Z0-9_]*[a-zA-Z0-9]+?))$"
 
-target-version = "py36"
+target-version = "py39"
 
 [tool.ruff.per-file-ignores]
 "tinynn/converter/schemas/torch/*.py" = ["E501"]
@@ -86,6 +86,7 @@ target-version = "py36"
 "examples/*.py" = ["E402"]
 "__init__.py" = ["F401", "F403"]
 "tests/import_test.py" = ["F401"]
+"tutorials/quantization/basic.ipynb" = ["F811", "F401"]
 
 [tool.ruff.mccabe]
 # Unlike Flake8, default to a complexity level of 10.

requirements.txt

@@ -1,5 +1,4 @@
-numpy>=1.18.5,<=1.24.4; python_version < '3.9'
-numpy>=1.18.5; python_version >= '3.9'
+numpy>=1.18.5
 PyYAML>=5.3.1
 ruamel.yaml>=0.16.12
 igraph>=0.9

tests/converter_op_test.py

@@ -565,7 +565,7 @@ def test_reduce_ops_single_dim(self):
 
             def model(x):
                 res = func(x, dim=1)
-                return res if type(res) == torch.Tensor else res[0]
+                return res if type(res) is torch.Tensor else res[0]
 
             model_path = get_model_path()
             converter = TFLiteConverter(model, dummy_input, model_path, nchw_transpose=False)
@@ -599,7 +599,7 @@ def test_reduce_ops_single_dim_keepdim(self):
 
             def model(x):
                 res = func(x, dim=1, keepdim=True)
-                return res if type(res) == torch.Tensor else res[0]
+                return res if type(res) is torch.Tensor else res[0]
 
             model_path = get_model_path()
             converter = TFLiteConverter(model, dummy_input, model_path, nchw_transpose=False)

tinynn/converter/operators/optimize.py

@@ -4661,21 +4661,21 @@ def elinimate_sequences(
         first_node = seq[0]
         last_node = seq[-1]
 
-        if type(skip_pred) == bool:
+        if type(skip_pred) is bool:
             skip = skip_pred
         elif skip_pred is not None:
             skip = skip_pred(seq)
 
         if skip:
             continue
 
-        if type(remove_first_pred) == bool:
+        if type(remove_first_pred) is bool:
             remove_first = remove_first_pred
             custom_data = None
         elif remove_first_pred is not None:
             remove_first, custom_data = remove_first_pred(seq)
 
-        if type(remove_last_pred) == bool:
+        if type(remove_last_pred) is bool:
             remove_last = remove_last_pred
             custom_data_last = None
         elif remove_last_pred is not None:

tinynn/converter/operators/tflite/base.py

@@ -186,7 +186,7 @@ def __init__(
         self.index = 0
         self.is_variable = is_variable
 
-        if type(tensor) == FakeQuantTensor:
+        if type(tensor) is FakeQuantTensor:
             self.quantization = QuantizationParameters(tensor.scale, tensor.zero_point, tensor.dim)
             tensor = tensor.tensor
 
@@ -195,7 +195,7 @@ def __init__(
 
         if type(tensor).__module__ == 'numpy':
             self.tensor = tensor
-        elif type(tensor) == torch.Tensor:
+        elif type(tensor) is torch.Tensor:
             assert tensor.is_contiguous, "Tensor should be contiguous"
             if tensor.dtype == torch.quint8:
                 self.tensor = torch.int_repr(tensor.detach()).numpy()
@@ -253,7 +253,7 @@ def __init__(
                         self.quantization = QuantizationParameters(scales, zero_points, dim)
             else:
                 self.tensor = tensor.detach().numpy()
-        elif type(tensor) == torch.Size:
+        elif type(tensor) is torch.Size:
             self.tensor = np.asarray(tensor, dtype='int32')
         elif type(tensor) in (tuple, list):
             self.tensor = np.asarray(tensor, dtype=dtype)
@@ -390,7 +390,7 @@ def build(self, builder: flatbuffers.Builder) -> Offset:
 def create_offset_vector(builder: flatbuffers.Builder, prop: typing.Callable, vec: typing.Iterable):
     if type(vec) not in (tuple, list):
         assert False, "type of vec unexpected, expected: list or tuple"
-    elif type(vec) == tuple:
+    elif type(vec) is tuple:
         vec = list(vec)
 
     prop_name = prop.__name__
@@ -426,7 +426,7 @@ def create_numpy_array(builder: flatbuffers.Builder, prop: typing.Callable, vec:
 
 
 def create_string(builder: flatbuffers.Builder, prop: typing.Callable, val: str):
-    if type(val) != str:
+    if type(val) is not str:
         assert False, "type of val unexpected, expected: str"
 
     prop_name = prop.__name__

tinynn/converter/operators/torch/aten.py

@@ -1561,7 +1561,7 @@ def parse(self, node, attrs, args, graph_converter):
         self.run(node)
 
         dim = self.input_tensors[1]
-        assert type(dim) == int
+        assert type(dim) is int
 
         if dim < 0:
             dim += self.input_tensors[0][0].ndim + 1
@@ -1619,7 +1619,7 @@ def parse(self, node, attrs, args, graph_converter):
         self.run(node)
 
         dim = self.input_tensors[1]
-        assert type(dim) == int
+        assert type(dim) is int
 
         if dim < 0:
             dim += self.input_tensors[0][0].ndim
@@ -2067,8 +2067,8 @@ def parse(self, node, attrs, args, graph_converter):
         input_tensor = self.find_or_create_input(0, graph_converter)
         dim, index = self.input_tensors[1:]
 
-        assert type(dim) == int
-        assert type(index) == int
+        assert type(dim) is int
+        assert type(index) is int
 
         if dim < 0:
             dim += input_tensor.tensor.ndim
@@ -2166,11 +2166,11 @@ def parse(self, node, attrs, args, graph_converter):
         self.parse_common(node, attrs, args, graph_converter)
 
     def parse_common(self, node, attrs, args, graph_converter):
-        if type(self) == ATenClampOperator:
+        if type(self) is ATenClampOperator:
             min_value, max_value = self.input_tensors[1:]
-        elif type(self) == ATenClampMinOperator:
+        elif type(self) is ATenClampMinOperator:
             min_value, max_value = self.input_tensors[1], None
-        elif type(self) == ATenClampMaxOperator:
+        elif type(self) is ATenClampMaxOperator:
             min_value, max_value = None, self.input_tensors[1]
 
         has_min = min_value is not None
@@ -3808,7 +3808,7 @@ def parse(self, node, attrs, args, graph_converter):
     def parse_common(self, graph_converter, input_idx=0, mask_idx=1, other_idx=2, out_idx=0):
         for i in (input_idx, other_idx):
             t = self.input_tensors[i]
-            if type(t) == torch.Tensor:
+            if type(t) is torch.Tensor:
                 if t.dtype == torch.float64:
                     self.input_tensors[i] = t.to(dtype=torch.float32)
                 elif t.dtype == torch.int64:
@@ -3826,7 +3826,7 @@ def parse_common(self, graph_converter, input_idx=0, mask_idx=1, other_idx=2, ou
         input_tensor, mask_tensor = [self.find_or_create_input(i, graph_converter) for i in (input_idx, mask_idx)]
 
         ops = []
-        if type(other) == torch.Tensor:
+        if type(other) is torch.Tensor:
             other_t = self.find_or_create_input(other_idx, graph_converter)
             if out.dtype != other.dtype:
                 casted = other.clone().to(dtype=out.dtype)
@@ -4491,4 +4491,4 @@ def parse(self, node, attrs, args, graph_converter):
             ops.append(tfl.TileOperator([actual_input, repeat_tensor], [outp]))
 
         for op in ops:
-            graph_converter.add_operator(op)
+            graph_converter.add_operator(op)

tinynn/converter/operators/torch/base.py

@@ -190,7 +190,7 @@ def to_tfl_tensors(
         tfl_tensors = []
         if has_buffers is None:
             has_buffers = [None] * len(tensors)
-        elif type(has_buffers) == bool:
+        elif type(has_buffers) is bool:
             has_buffers = [has_buffers] * len(tensors)
         assert len(names) == len(tensors) == len(has_buffers)
         for n, t, b in zip(names, tensors, has_buffers):
@@ -491,7 +491,7 @@ def handle_padding(self, pad_h, pad_w, pad_op_index, ops, ceil_mode=False):
             input_size = [input_tensor.shape[2], input_tensor.shape[3]]
 
             if not all((i + 2 * p - k) % s == 0 for i, p, k, s in zip(input_size, padding, kernel_size, stride)):
-                assert type(ops[1]) == tfl.MaxPool2dOperator, 'ceil_mode=True for AvgPool not supported'
+                assert type(ops[1]) is tfl.MaxPool2dOperator, 'ceil_mode=True for AvgPool not supported'
                 fill_nan = True
                 ceil_pad = get_pool_ceil_padding(input_tensor, kernel_size, stride, padding)
                 ceil_pad = list(np.add(ceil_pad, padding))
@@ -503,7 +503,7 @@ def handle_padding(self, pad_h, pad_w, pad_op_index, ops, ceil_mode=False):
             pad_input = ops[pad_op_index - 1].outputs[0]
 
             inputs = [pad_input, pad_tensor]
-            if type(ops[1]) == tfl.MaxPool2dOperator:
+            if type(ops[1]) is tfl.MaxPool2dOperator:
                 constant_tensor = self.get_minimum_constant(pad_input)
                 inputs.append(constant_tensor)
                 pad_array = np.pad(pad_input.tensor, pad, constant_values=constant_tensor.tensor[0])

tinynn/converter/operators/torch/quantized.py

@@ -153,7 +153,7 @@ def parse(self, node, attrs, args, graph_converter):
         self.run(node)
 
         dim = self.input_tensors[1]
-        assert type(dim) == int
+        assert type(dim) is int
 
         if dim < 0:
             dim += self.input_tensors[0][0].ndim

tinynn/graph/configs/gen_creation_funcs_yml.py

@@ -17,7 +17,7 @@
     if k in block_list:
         continue
     c = getattr(torch, k)
-    if inspect.isclass(c) and k.endswith('Tensor') and c.__bases__[0] == object:
+    if inspect.isclass(c) and k.endswith('Tensor') and c.__bases__[0] is object:
         print(k)
         final_dict['torch'].append(k)
     elif inspect.isbuiltin(c):