change the architecture of the Code

README.md

@@ -1,3 +1,11 @@
 # JSparse
 
 JSparse is a high-performance auto-differentiation library for sparse tensor. 
+
+
+## How to Install
+
+```
+cd python
+python setup.py install
+```

examples/example.py

@@ -0,0 +1,73 @@
+from pickletools import optimize
+from statistics import mode
+import jittor
+from jittor import nn
+
+import JSparse 
+from JSparse import SparseTensor
+from JSparse import nn as spnn
+from JSparse.utils.quantize import sparse_quantize
+
+import numpy as np
+
+class RandomDataset(jittor.dataset.Dataset):
+
+    def __init__(self, input_size: int, voxel_size: float) -> None:
+        super().__init__()
+        self.set_attrs(total_len = input_size)
+        self.voxel_size = voxel_size
+
+    def __getitem__(self, _: int):
+        inputs = np.random.uniform(-100, 100, size=(self.total_len, 4))
+        labels = np.random.choice(10, size=self.total_len)
+
+        coords, feats = inputs[:, :], inputs
+        coords -= np.min(coords, axis=0, keepdims=True)
+#       coords, indices = sparse_quantize(coords, self.voxel_size, return_index=True)
+
+        coords = jittor.Var(coords)
+        feats = jittor.Var(feats)
+        labels = jittor.Var(labels)
+#       coords = jittor.Var(coords, dtype=jittor.int64)
+#       feats = jittor.Var(feats[indices], dtype=jittor.float64)
+#       labels = jittor.Var(labels[indices], dtype=jittor.int64)
+
+        print(type(coords))
+        inputs = SparseTensor(coords, feats, 1, 1)
+        labels = SparseTensor(coords, labels, 1, 1)
+        return inputs, labels
+
+if __name__ == '__main__':
+    np.random.seed(0)
+
+    dataset = RandomDataset(input_size=10000, voxel_size=0.2)
+
+    model = nn.Sequential(
+        spnn.Conv3d(4, 32, 3),
+        spnn.BatchNorm(32),
+        spnn.ReLU(True),
+        spnn.Conv3d(32, 64, 2, stride=2),
+        spnn.BatchNorm(64),
+        spnn.ReLU(True),
+        spnn.Conv3d(64, 64, 2, stride=2, transposed=True),
+        spnn.BatchNorm(64),
+        spnn.ReLU(True),
+        spnn.Conv3d(64, 32, 3),
+        spnn.BatchNorm(32),
+        spnn.ReLU(True),
+        spnn.Conv3d(32, 10, 1),
+    )
+
+    criterion = nn.CrossEntropyLoss()
+    optimizer = jittor.optim.Adam(model.parameters(), lr=1e-3)
+
+    model.train()
+    lens = len(dataset)
+    for batch_idx, (inputs, labels) in enumerate(dataset):
+        outputs = model(inputs)
+        loss = criterion(outputs, labels)
+        optimizer.setp(loss)
+
+        if batch_idx % 10 == 0:
+            print('Training: [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
+                batch_idx, lens , 100. * batch_idx / lens, loss.numpy()[0]))

python/JSparse/nn/functional/activation.py

@@ -1,8 +1,8 @@
 import jittor as jt
 import jittor.nn as nn
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
-from python.jsparse.nn.utils import fapply
+from JSparse.nn.utils import fapply
 
 __all__ = ['relu', 'leaky_relu']
 # __all__ = ['relu', 'leaky_relu', 'ReLU', 'LeakyReLU']

python/JSparse/nn/functional/conv.py

@@ -4,10 +4,10 @@ import jittor as jt
 from jittor import Function
 from jittor.misc import _pair, _triple
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
-from python.jsparse.nn import functional as F
+from JSparse.nn import functional as F
-from python.jsparse.nn.utils import get_kernel_offsets
+from JSparse.nn.utils import get_kernel_offsets
-from python.jsparse import make_ntuple
+from JSparse import make_ntuple
 
 __all__ = ['conv3d', 'Convolution']
 

python/JSparse/nn/functional/devoxelize.py

@@ -1,7 +1,7 @@
 import jittor as jt
 from jittor import Function
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
 
 __all__ = ['calc_ti_weights', 'spdevoxelize']
 

python/JSparse/nn/functional/downsample.py

@@ -3,8 +3,8 @@ from typing import Tuple, Union
 import jittor as jt
 from jittor.misc import _pair, _triple
 
-from python.jsparse.nn.utils import get_kernel_offsets
+from JSparse.nn.utils import get_kernel_offsets
-from python.jsparse.utils import make_ntuple, trunc
+from JSparse.utils import make_ntuple, trunc
 
 __all__ = ['spdownsample']
 

python/JSparse/nn/functional/pooling.py

@@ -1,6 +1,6 @@
 import jittor as jt
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
 
 __all__ = ['global_avg_pool', 'global_max_pool']
 

python/JSparse/nn/functional/voxelize.py

@@ -1,7 +1,7 @@
 import jittor as jt
 from jittor import Function
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
 
 __all__ = ['spvoxelize']
 

python/JSparse/nn/modules/activation.py

@@ -1,9 +1,9 @@
 import jittor as jt
 from jittor import nn
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
-from python.jsparse.nn.functional import relu, leaky_relu
+from JSparse.nn.functional import relu, leaky_relu
-# from python.nn.utils import fapply
+# from nn.utils import fapply
 
 __all__ = ['ReLU', 'LeakyReLU']
 

python/JSparse/nn/modules/conv.py

@@ -7,8 +7,8 @@ from jittor import nn
 from jittor import init
 from jittor.misc import _pair, _triple
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
-from python.jsparse.nn import functional as F
+from JSparse.nn import functional as F
 # from utils import make_ntuple
 
 __all__ = ['Conv3d']

python/JSparse/nn/modules/norm.py

@@ -2,8 +2,8 @@ import jittor as jt
 from jittor import nn
 from numpy import kaiser
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
-from python.jsparse.nn.utils import fapply
+from JSparse.nn.utils import fapply
 
 __all__ = ['BatchNorm', 'GroupNorm']
 

python/JSparse/nn/modules/pooling.py

@@ -2,8 +2,8 @@ from ast import Global
 import jittor as jt
 from jittor import nn
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
-from python.jsparse.nn.functional import global_avg_pool, global_max_pool
+from JSparse.nn.functional import global_avg_pool, global_max_pool
 
 __all__ = ['GlobalAvgPool', 'GlobalMaxPool']
 

python/JSparse/nn/utils/apply.py

@@ -2,7 +2,7 @@ from typing import Callable
 
 import jittor as jt
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
 
 __all__ = ['fapply']
 

python/JSparse/nn/utils/kernel.py

@@ -3,7 +3,7 @@ from typing import Tuple, Union
 import numpy as np
 import jittor as jt
 
-from python.jsparse.utils import make_ntuple, trunc
+from JSparse.utils import make_ntuple, trunc
 
 __all__ = ['get_kernel_offsets']
 

python/JSparse/sparse.py

@@ -5,26 +5,27 @@ import jittor as jt
 from jittor.misc import _pair, _triple
 
 from typing import Any, Dict, Tuple, Union
+from type_check import type_check
 
-from python.jsparse.utils import make_ntuple, sparse_quantize, set_hash
+from JSparse.utils import make_ntuple, sparse_quantize, set_hash
 # from .utils.quantize import sparse_quantize
 # from indice_manager import IndiceManager
 
 class SparseTensor:
 
+    @type_check
     def __init__(
         self, 
         indices: jt.Var, 
         values: jt.Var,
         stride: Union[int, Tuple[int, ...]],
-        size,
+        size=None,
         quantize=True,
         voxel_size=1,
-        coalesce_mode='sum',
+        coalesce_mode:str='sum',
         indice_manager=None,
         device=None,
     ):
-        assert isinstance(indices, jt.Var) and isinstance(values, jt.Var)
         assert (values.ndim == 2)
         # self.indices = indices
         # self.values = values

python/setup.py

@@ -0,0 +1,15 @@
+from gettext import install
+
+from setuptools import find_packages, setup
+
+with open("../README.md", "r") as file:
+    description = file.read()
+
+print(find_packages())
+setup(
+    name='JSparse',
+    version='0.1',
+    description=description,
+    packages=find_packages(),
+    install_requires=["jittor", "type_check"]
+)

voxelize_test.py

@@ -11,12 +11,12 @@ from jittor.misc import _pair, _triple
 from itertools import repeat
 from typing import List, Tuple, Union
 
-from python.jsparse import SparseTensor
+from JSparse import SparseTensor
-from python.jsparse import PointTensor
+from JSparse import PointTensor
-from python.jsparse.utils import make_ntuple
+from JSparse.utils import make_ntuple
-from python.jsparse.nn import functional as F
+from JSparse.nn import functional as F
-from python.jsparse.nn.utils import get_kernel_offsets
+from JSparse.nn.utils import get_kernel_offsets
-from python.jsparse.nn.functional import Convolution
+from JSparse.nn.functional import Convolution
 
 import torchsparse
 from torchsparse import nn as spnn