diff --git a/python/__init__.py b/python/__init__.py
new file mode 100644
index 0000000..4228645
--- /dev/null
+++ b/python/__init__.py
@@ -0,0 +1 @@
+from .sparse import *
\ No newline at end of file
diff --git a/python/__pycache__/__init__.cpython-39.pyc b/python/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000..51c3946
Binary files /dev/null and b/python/__pycache__/__init__.cpython-39.pyc differ
diff --git a/python/__pycache__/sparse.cpython-39.pyc b/python/__pycache__/sparse.cpython-39.pyc
new file mode 100644
index 0000000..1a0ba05
Binary files /dev/null and b/python/__pycache__/sparse.cpython-39.pyc differ
diff --git a/python/__pycache__/utils.cpython-39.pyc b/python/__pycache__/utils.cpython-39.pyc
new file mode 100644
index 0000000..29217de
Binary files /dev/null and b/python/__pycache__/utils.cpython-39.pyc differ
diff --git a/python/indice_manager.py b/python/indice_manager.py
new file mode 100644
index 0000000..18e16a2
--- /dev/null
+++ b/python/indice_manager.py
@@ -0,0 +1,34 @@
+import os
+import numpy as np
+from typing import Union, List, Tuple
+
+import jittor as jt
+from jittor import Function
+
+class IndiceManager:
+    def __init__(
+        self,
+        ndim,
+        # indice_map_type,
+        # sparse_alorithm, # set m_hashtable_occupancy for concurrent_unordered_map 
+    ):
+        # if indice_map_type == 'GPU':
+        #     assert(jt.has_cuda)
+
+        self.ndim = ndim
+        # self.indice_map_type = indice_map_type
+        # self.sparse_algorithm = sparse_alorithm
+        self.stride_key_manager = {}
+        self.indice_map_manager = {}
+        self.kernel_map_manager = {}
+    
+    def insert(self, stride, indice_key, indice_hash):
+        self.stride_key_manager[stride] = indice_key
+        self.indice_map_manager[indice_key] = indice_hash
+
+
+# class IndiceMapManager
+        
+        
+
+        
\ No newline at end of file
diff --git a/python/nn/__init__.py b/python/nn/__init__.py
new file mode 100644
index 0000000..270dceb
--- /dev/null
+++ b/python/nn/__init__.py
@@ -0,0 +1 @@
+from .modules import *
diff --git a/python/nn/__pycache__/__init__.cpython-39.pyc b/python/nn/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000..3d661fc
Binary files /dev/null and b/python/nn/__pycache__/__init__.cpython-39.pyc differ
diff --git a/python/nn/functional/__init__.py b/python/nn/functional/__init__.py
index de6e4d9..240b7d0 100644
--- a/python/nn/functional/__init__.py
+++ b/python/nn/functional/__init__.py
@@ -1,10 +1,10 @@
-# from .activation import *
+from .activation import *
 from .conv import *
 # from .count import *
 # from .crop import *
 # from .devoxelize import *
 from .downsample import *
 from .hash import *
-# from .pooling import *
+from .pooling import *
 from .query import *
 # from .voxelize import *
diff --git a/python/nn/functional/__pycache__/__init__.cpython-39.pyc b/python/nn/functional/__pycache__/__init__.cpython-39.pyc
index f98f401..dafd0b9 100644
Binary files a/python/nn/functional/__pycache__/__init__.cpython-39.pyc and b/python/nn/functional/__pycache__/__init__.cpython-39.pyc differ
diff --git a/python/nn/functional/__pycache__/activation.cpython-39.pyc b/python/nn/functional/__pycache__/activation.cpython-39.pyc
new file mode 100644
index 0000000..6b15650
Binary files /dev/null and b/python/nn/functional/__pycache__/activation.cpython-39.pyc differ
diff --git a/python/nn/functional/__pycache__/conv.cpython-39.pyc b/python/nn/functional/__pycache__/conv.cpython-39.pyc
index 43dad00..0b42695 100644
Binary files a/python/nn/functional/__pycache__/conv.cpython-39.pyc and b/python/nn/functional/__pycache__/conv.cpython-39.pyc differ
diff --git a/python/nn/functional/__pycache__/pooling.cpython-39.pyc b/python/nn/functional/__pycache__/pooling.cpython-39.pyc
new file mode 100644
index 0000000..2e8e8e8
Binary files /dev/null and b/python/nn/functional/__pycache__/pooling.cpython-39.pyc differ
diff --git a/python/nn/functional/activation.py b/python/nn/functional/activation.py
new file mode 100644
index 0000000..52b3158
--- /dev/null
+++ b/python/nn/functional/activation.py
@@ -0,0 +1,24 @@
+import jittor as jt
+import jittor.nn as nn
+
+from python import SparseTensor
+from python.nn.utils import fapply
+
+__all__ = ['relu', 'leaky_relu']
+# __all__ = ['relu', 'leaky_relu', 'ReLU', 'LeakyReLU']
+
+def relu(input: SparseTensor) -> SparseTensor:
+    return fapply(input, nn.relu)
+
+
+def leaky_relu(input: SparseTensor,
+               scale: float = 0.01) -> SparseTensor:
+    return fapply(input,
+                  nn.leaky_relu,
+                  scale=scale)
+
+# Relu = jt.make_module(relu)
+# ReLU = Relu
+# Leaky_relu = jt.make_module(leaky_relu, 2)
+# LeakyReLU = Leaky_relu
+
diff --git a/python/nn/functional/conv.py b/python/nn/functional/conv.py
index 5812dc8..096210c 100644
--- a/python/nn/functional/conv.py
+++ b/python/nn/functional/conv.py
@@ -270,6 +270,7 @@ def conv3d(
     bias: Optional[jt.Var] = None,
     stride: Union[int, Tuple[int, ...]] = 1,
     dilation: Union[int, Tuple[int, ...]] = 1,
+    group: int = 1,
     transposed: bool = False,
 ) -> SparseTensor:
     # kernel_size = make_ntuple(kernel_size, ndim=3)
@@ -337,10 +338,12 @@ def conv3d(
     if bias is not None:
         output_values += bias
     
+    # size have to be set
     output = SparseTensor(
         indices=output_indices,
         values=output_values,
         stride=output_stride,
+        size=input.size
     )
     output.cmaps = input.cmaps
     output.cmaps.setdefault(output_stride, output_indices)
diff --git a/python/nn/functional/pooling.py b/python/nn/functional/pooling.py
new file mode 100644
index 0000000..3d84fce
--- /dev/null
+++ b/python/nn/functional/pooling.py
@@ -0,0 +1,26 @@
+import jittor as jt
+
+from python import SparseTensor
+
+__all__ = ['global_avg_pool', 'global_max_pool']
+
+def global_avg_pool(inputs: SparseTensor) -> jt.Var:
+    batch_size = jt.max(inputs.indices[:, 0]).item() + 1
+    outputs = []
+    for k in range(batch_size):
+        input = inputs.values[inputs.indices[:, 0] == k]
+        output = jt.mean(input, dim=0)
+        outputs.append(output)
+    outputs = jt.stack(outputs, dim=0)
+    return outputs
+
+
+def global_max_pool(inputs: SparseTensor) -> jt.Var:
+    batch_size = jt.max(inputs.indices[:, 0]).item() + 1
+    outputs = []
+    for k in range(batch_size):
+        input = inputs.values[inputs.indices[:, 0] == k]
+        output = jt.max(input, dim=0)[0]
+        outputs.append(output)
+    outputs = jt.stack(outputs, dim=0)
+    return outputs
\ No newline at end of file
diff --git a/python/nn/modules/__init__.py b/python/nn/modules/__init__.py
new file mode 100644
index 0000000..82665d0
--- /dev/null
+++ b/python/nn/modules/__init__.py
@@ -0,0 +1,6 @@
+from .activation import *
+# from .bev import *
+from .conv import *
+# from .crop import *
+from .norm import *
+from .pooling import *
diff --git a/python/nn/modules/__pycache__/__init__.cpython-39.pyc b/python/nn/modules/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000..bd3d116
Binary files /dev/null and b/python/nn/modules/__pycache__/__init__.cpython-39.pyc differ
diff --git a/python/nn/modules/__pycache__/activation.cpython-39.pyc b/python/nn/modules/__pycache__/activation.cpython-39.pyc
new file mode 100644
index 0000000..4666e13
Binary files /dev/null and b/python/nn/modules/__pycache__/activation.cpython-39.pyc differ
diff --git a/python/nn/modules/__pycache__/conv.cpython-39.pyc b/python/nn/modules/__pycache__/conv.cpython-39.pyc
new file mode 100644
index 0000000..fe3e238
Binary files /dev/null and b/python/nn/modules/__pycache__/conv.cpython-39.pyc differ
diff --git a/python/nn/modules/__pycache__/norm.cpython-39.pyc b/python/nn/modules/__pycache__/norm.cpython-39.pyc
new file mode 100644
index 0000000..334f604
Binary files /dev/null and b/python/nn/modules/__pycache__/norm.cpython-39.pyc differ
diff --git a/python/nn/modules/__pycache__/pooling.cpython-39.pyc b/python/nn/modules/__pycache__/pooling.cpython-39.pyc
new file mode 100644
index 0000000..9d8fc1e
Binary files /dev/null and b/python/nn/modules/__pycache__/pooling.cpython-39.pyc differ
diff --git a/python/nn/modules/activation.py b/python/nn/modules/activation.py
new file mode 100644
index 0000000..e751d58
--- /dev/null
+++ b/python/nn/modules/activation.py
@@ -0,0 +1,21 @@
+import jittor as jt
+from jittor import nn
+
+from python import SparseTensor
+from python.nn.functional import relu, leaky_relu
+# from python.nn.utils import fapply
+
+__all__ = ['ReLU', 'LeakyReLU']
+
+# class ReLU(nn.ReLU):
+#     def execute(self, input: SparseTensor) -> SparseTensor:
+#         return fapply(input, super().execute)
+
+# class LeakyReLU(nn.LeakyReLU):
+#     def execute(self, input: SparseTensor) -> SparseTensor:
+#         return fapply(input, super().execute)
+
+Relu = jt.make_module(relu)
+ReLU = Relu
+Leaky_relu = jt.make_module(leaky_relu, 2)
+LeakyReLU = Leaky_relu
\ No newline at end of file
diff --git a/python/nn/modules/conv3d.py b/python/nn/modules/conv.py
similarity index 99%
rename from python/nn/modules/conv3d.py
rename to python/nn/modules/conv.py
index d96ec97..5790979 100644
--- a/python/nn/modules/conv3d.py
+++ b/python/nn/modules/conv.py
@@ -9,7 +9,7 @@ from jittor.misc import _pair, _triple
 
 from python import SparseTensor
 from python.nn import functional as F
-from utils import make_ntuple
+# from utils import make_ntuple
 
 __all__ = ['Conv3d']
 
diff --git a/python/nn/modules/norm.py b/python/nn/modules/norm.py
new file mode 100644
index 0000000..de483f4
--- /dev/null
+++ b/python/nn/modules/norm.py
@@ -0,0 +1,34 @@
+import jittor as jt
+from jittor import nn
+from numpy import kaiser
+
+from python import SparseTensor
+from python.nn.utils import fapply
+
+__all__ = ['BatchNorm', 'GroupNorm']
+
+class BatchNorm(nn.BatchNorm):
+    def execute(self, input: SparseTensor) -> SparseTensor:
+        return fapply(input, super().execute)
+
+class GroupNorm(nn.GroupNorm):
+    def execute(self, input: SparseTensor) -> SparseTensor:
+        indices, values, stride, size = input.indices, input.values, input.stride, input.size
+
+        batch_size = jt.max(indices[:, 0]).item() + 1
+        num_channels = values.shape[1]
+
+        n_values = jt.zeros_like(values)
+        for k in range(batch_size):
+            idx = indices[:, 0] == k
+            b_values = values[idx]
+            b_values = b_values.t().reshape(1, num_channels, -1)
+            b_values = super().execute(b_values)
+            b_values = b_values.reshape(num_channels, -1).t()
+            n_values[idx] = b_values
+        
+        output = SparseTensor(indices=indices, values=n_values, stride=stride, size=size)
+        output.cmaps = input.cmaps
+        output.kmaps = input.kmaps
+        return output
+
diff --git a/python/nn/modules/pooling.py b/python/nn/modules/pooling.py
new file mode 100644
index 0000000..871ea00
--- /dev/null
+++ b/python/nn/modules/pooling.py
@@ -0,0 +1,11 @@
+from ast import Global
+import jittor as jt
+from jittor import nn
+
+from python import SparseTensor
+from python.nn.functional import global_avg_pool, global_max_pool
+
+__all__ = ['GlobalAvgPool', 'GlobalMaxPool']
+
+GlobalAvgPool = jt.make_module(global_avg_pool)
+GlobalMaxPool = jt.make_module(global_max_pool)
\ No newline at end of file
diff --git a/python/nn/utils/__init__.py b/python/nn/utils/__init__.py
index 662c212..919f5b0 100644
--- a/python/nn/utils/__init__.py
+++ b/python/nn/utils/__init__.py
@@ -1 +1,2 @@
+from .apply import *
 from .kernel import *
\ No newline at end of file
diff --git a/python/nn/utils/__pycache__/__init__.cpython-39.pyc b/python/nn/utils/__pycache__/__init__.cpython-39.pyc
index 0b03769..6d461f4 100644
Binary files a/python/nn/utils/__pycache__/__init__.cpython-39.pyc and b/python/nn/utils/__pycache__/__init__.cpython-39.pyc differ
diff --git a/python/nn/utils/__pycache__/apply.cpython-39.pyc b/python/nn/utils/__pycache__/apply.cpython-39.pyc
new file mode 100644
index 0000000..722228e
Binary files /dev/null and b/python/nn/utils/__pycache__/apply.cpython-39.pyc differ
diff --git a/python/nn/utils/apply.py b/python/nn/utils/apply.py
new file mode 100644
index 0000000..2ca3432
--- /dev/null
+++ b/python/nn/utils/apply.py
@@ -0,0 +1,15 @@
+from typing import Callable
+
+import jittor as jt
+
+from python import SparseTensor
+
+__all__ = ['fapply']
+
+def fapply(input: SparseTensor, fn: Callable[..., jt.Var], *args,
+           **kwargs) -> SparseTensor:
+    values = fn(input.values, *args, **kwargs)
+    output = SparseTensor(indices=input.indices, values=values, stride=input.stride, size=input.size)
+    output.cmaps = input.cmaps
+    output.kmaps = input.kmaps
+    return output
\ No newline at end of file
diff --git a/python/sparse.py b/python/sparse.py
index 9baee94..3484ab7 100644
--- a/python/sparse.py
+++ b/python/sparse.py
@@ -24,11 +24,13 @@ class SparseTensor:
         assert (values.ndim == 2)
         # self.indices = indices
         # self.values = values
-        self.shape = size
+        self.size = size
         self.ndim = indices.shape[1] - 1
         self.stride = make_ntuple(stride, ndim=self.ndim)
         self.voxel_size = voxel_size
         self.coalesce_mode = coalesce_mode
+        self.cmaps = {}
+        self.kmaps = {}
 
         ##########################
         # Setup CoordsManager
diff --git a/python/sparse_dense_function.py b/python/sparse_dense_function.py
new file mode 100644
index 0000000..f8bb855
--- /dev/null
+++ b/python/sparse_dense_function.py
@@ -0,0 +1,42 @@
+import jittor as jt
+from jittor import Function
+
+
+
+
+def spmm(
+    rows: jt.Var,
+    cols: jt.Var,
+    vals: jt.Var,
+    size: jt.NanoVector,
+    mat: jt.Var,
+    spmm_mode='scatter',
+    is_sorted: bool = False,
+    cuda_spmm_alg: int = 1,
+) -> jt.Var:
+    assert len(rows) == len(cols), "Invalid length"
+    assert len(rows) == len(vals), "Invalid length"
+    assert vals.dtype == mat.dtype, "dtype mismatch"
+
+    if jt.flags.use_cuda > 1:
+        assert jt.has_cuda == 1, "No GPUs available"
+        rows = rows.int()
+        cols = cols.int()
+        ''' 
+        TODO: Using the coo_spmm of cuSPARSE on GPU
+        result = coo_spmm_int32(
+            rows, cols, vals, size[0], size[1], mat, cuda_spmm_alg, is_sorted
+        )
+        '''
+    else:
+        if (spmm_mode == 'scatter'):
+
+
+        
+        
+        
+
+
+
+
+class SPMM(Function):