Sketch out conj method

Dockerfile.dev

@@ -0,0 +1,18 @@
+FROM nvidia/cuda:11.0.3-runtime-ubuntu20.04
+
+ENV DEBIAN_FRONTEND=noninteractive 
+
+RUN apt update -y && \
+    apt install -y python3 python3-pip git htop vim
+
+# Make sure you first recursively clone down the git repo before building 
+WORKDIR /app
+RUN pip install quimb pyrofiler cartesian-explorer opt_einsum
+RUN pip install --no-binary pynauty pynauty
+# Run the below commands after the container opens - because volume hasn't mounted yet
+# RUN cd qtree && pip install . 
+# RUN pip install .
+RUN pip install pdbpp
+RUN pip install tensornetwork
+
+ENTRYPOINT ["bash"]

dev.sh

@@ -0,0 +1,4 @@
+#!/bin/bash
+
+docker build -f Dockerfile.dev -t dev .
+docker run -v $(pwd):/app -it dev

qtensor/Simulate.py

@@ -155,3 +155,77 @@ def simulate(self, qc, **params):
         sim = cirq.Simulator(**params)
         return sim.simulate(qc)
 
+if __name__=="__main__":
+    import networkx as nx
+    import numpy as np
+    import tensornetwork as tn
+
+    G = nx.random_regular_graph(3, 10)
+    gamma, beta = [np.pi/3], [np.pi/2]
+
+    from qtensor import QtreeQAOAComposer, QAOAQtreeSimulator
+    composer = QtreeQAOAComposer(graph=G, gamma=gamma, beta=beta)
+    composer.ansatz_state()
+
+    sim = QAOAQtreeSimulator(composer)
+
+    # now let's run the prepare buckets method to init the tn
+    sim.simulate_batch(composer.circuit)
+    buckets = sim.tn.buckets
+
+    # now let's use these buckets to square the TN
+    def conj(buckets):
+        # turn each bucket into a node
+        nodes = []
+        for bucket in buckets:
+            node = tn.Node(np.array(bucket))
+            nodes.append(node)
+
+        # now for each node, append its conjugate
+        conj_nodes = []
+        for node in nodes:
+            conj = np.conj(node.tensor)
+            conj_node = tn.Node(conj)
+            conj_nodes.append(conj_node)
+
+
+        indices = {}
+
+        for node in nodes:
+            for conj_node in conj_nodes:
+                # check if there is a shared index between a node and a conj_node
+                node_indices = node.get_all_dangling()
+                conj_indices = conj_node.get_all_dangling()
+
+                shared_indices = set(node_indices).intersection(set(conj_indices))
+                if shared_indices:
+                    if node not in indices:
+                        indices[node] = shared_indices
+                    else:
+                        indices[node].update(shared_indices)
+
+                    if conj_node not in indices:
+                        indices[conj_node] = shared_indices
+                    else:
+                        indices[conj_node].update(shared_indices)
+
+        for node, shared_indices in indices.items():
+            for pair_node in indices.keys():
+                if node == pair_node:
+                    continue
+                # if there are shared indices, connect an edge
+                if shared_indices.intersection(set(pair_node.get_all_dangling())):
+                    edge = tn.connect(node, pair_node)
+
+        # TODO: TNAdapter should support tensornetwork.Node
+        # So that we can contract this resulting tensor network directly
+        return []
+
+    tn_with_conj = conj(buckets)
+
+    # TODO: contract or sample using tn_with_conj based on method in other branch
+
+
+
+    log.debug('hello world')
+    import pdb; pdb.set_trace()

scratchpad/tn_api/test_tn_api.py

@@ -0,0 +1,51 @@
+import random
+import numpy as np
+from tn import TensorNetwork
+from functools import reduce
+
+def test_add_numpy_array():
+    a = TensorNetwork()
+    t = np.random.randn(2, 2)
+    a.add(t)
+    b = TensorNetwork()
+    b.add(a)
+    assert b == a
+
+
+def test_composition():
+    """
+    tensor network adding is associative
+    """
+    tns = [TensorNetwork.new_random_cpu(2, 3, 4) for _ in range(5)]
+    stack = TensorNetwork()
+    # (((0A)B)C)D
+    for tn in tns:
+        stack.add(tn)
+    # A(B(CD))
+    for i in range(len(tns)):
+        l = tns[len(tns)-2-i]
+        r = tns[len(tns)-1-i]
+        l.add(r)
+
+    assert stack == tns[0]
+
+def test_edges_consistent_ports():
+    tns = [TensorNetwork.new_random_cpu(2, 3, 4) for _ in range(5)]
+    tn = TensorNetwork()
+    # (((0A)B)C)D
+    for t in tns:
+        tn.add(t)
+
+    port_data = {}
+    for e in tn._edges:
+        for p in e:
+            port_data[p.tensor_ref] = port_data.get(p.tensor_ref, [])
+            port_data[p.tensor_ref].append(p.ix)
+    for i, t in enumerate(tn._tensors):
+        assert len(t.shape) == len(port_data[i])
+
+
+
+if __name__=="__main__":
+    test_edges_consistent_ports()
+    test_composition()

scratchpad/tn_api/tn.py

@@ -0,0 +1,173 @@
+import numpy as np
+import math
+from dataclasses import dataclass
+from typing import TypeVar, Generic, Iterable
+
+class Array(np.ndarray):
+    shape: tuple
+
+D = TypeVar('D') # tensor data type (numpy, torch, etc.)
+
+class ContractionInfo:
+    pass
+
+class TensorNetworkIFC(Generic[D]):
+    def __init__(self, *args, **kwargs):
+        ...
+
+    def optimize(self, out_indices: Iterable = []) -> ContractionInfo:
+        return ContractionInfo()
+
+    # slice not inplace
+    def slice(self, slice_dict: dict) -> 'TensorNetwork':
+        ...
+
+    # contract to produce a new tensor
+    def contract(self, contraction_info: ContractionInfo) -> D:
+        ...
+
+    # 
+    def copy(self):
+        ...
+
+    def add(self, other: "TensorNetworkIFC[D] | D"):
+        ...
+
+
+    @classmethod
+    def new_random_cpu(cls, dims: Iterable[int])-> 'TensorNetworkIFC[D]':
+        ...
+
+    def __eq__(a, b):
+        ...
+
+
+N = TypeVar('N', bound=np.ndarray)
+
+@dataclass
+class Port:
+    tensor_ref: int
+    ix: int
+
+class TensorNetwork(TensorNetworkIFC[np.ndarray]):
+    tensors: Iterable[np.ndarray]
+    shape: tuple
+    edges: tuple
+
+    def __init__(self, *args, **kwargs):
+        self._tensors = []
+        self._edges = tuple()
+        self.shape = tuple()
+
+    # slice not inplace
+    def slice(self, slice_dict: dict) -> 'TensorNetwork':
+        tn = self.copy()
+        sliced_tns = []
+        for tensor in tn._tensors:
+            slice_bounds = []
+            for idx in range(tensor.ndim):
+                try:
+                    slice_bounds.append(slice_dict[idx])
+                except KeyError:
+                    slice_bounds.append(slice(None))
+
+            sliced_tns.append(tensor[tuple(slice_bounds)])
+
+        tn._tensors = sliced_tns
+        return tn
+
+    def copy(self):
+        new = TensorNetwork()
+        new._tensors = self._tensors
+        new._edges = self._edges
+        new.shape = self.shape
+        return new
+
+    def add(self, other: "TensorNetwork | np.ndarray"):
+        if not isinstance(other, TensorNetwork):
+            self._tensors.append(other)
+            self.shape = self.shape + other.shape
+        else:
+            m = len(self._tensors)
+            n = len(self.shape)
+            # -- other's edges tensors will refer to shifted tensor location
+            enew = []
+            for e in other._edges:
+                e_ = []
+                for p in e:
+                    if p.tensor_ref == -1:
+                        e_.append(Port(tensor_ref=-1, ix=p.ix+n))
+                    else:
+                        e_.append(Port(tensor_ref=p.tensor_ref+m, ix=p.ix))
+                enew.append(tuple(e_))
+
+            self._edges += tuple(enew)
+            self._tensors += other._tensors
+            self.shape += other.shape
+
+    # contract to produce a new tensor
+    def contract(self, contraction_info: ContractionInfo) -> np.ndarray:
+        raise NotImplementedError()
+
+    def optimize(self, out_indices: Iterable = []) -> ContractionInfo:
+        raise NotImplementedError()
+
+
+    @classmethod
+    def new_random_cpu(cls, count, size, dim: int):
+        out = cls()
+        for i in range(count):
+            t: np.ndarray = np.random.random((dim, )*size)
+            out.add(t)
+        # arbitrary number of output indices
+        out_dims = np.random.randint(low=0, high=len(out.shape))
+        tensor_dims = len(out.shape)
+        out.shape = (dim, )*out_dims
+        # -- random connectivity
+        # A hypergraph can be generated as a partition into
+        # E parts where E is number of edges. The isolated vertices are equivalent
+        # to vertices with 1 edge that contains only them.
+        # arbitrary max number of edges, must be less than total indices
+        edges_cnt = np.random.randint(low=1, high=tensor_dims+out_dims)
+        # a partition can be implemented using a random function
+        partition_fn = lambda : np.random.randint(low=0, high=edges_cnt)
+        partition_dict = {}
+        for t_ref, t in enumerate(out._tensors):
+            for i in range(t.ndim):
+                eix = partition_fn()
+                new_port = Port(tensor_ref=t_ref, ix=i)
+                partition_dict[eix] = partition_dict.get(eix, [])
+                partition_dict[eix].append(new_port)
+
+        # add "self" tensor indices to partition
+        for i in range(len(out.shape)):
+            eix = partition_fn()
+            new_port = Port(tensor_ref=-1, ix=i)
+            partition_dict[eix] = partition_dict.get(eix, [])
+            partition_dict[eix].append(new_port)
+
+        edges = []
+        for i in range(edges_cnt):
+            p = partition_dict.get(i)
+            if p is not None:
+                edges.append(tuple(p))
+        out._edges = tuple(edges)
+        return out
+
+    def __eq__(self, other):
+        if self.shape != other.shape:
+            return False
+        if self._edges != other._edges:
+            return False
+        return all((a==b).all() for a, b in zip(self._tensors, other._tensors))
+
+    def __repr__(self):
+        return f"TensorNetwork({self.shape})<{self._tensors}, {self._edges}>"
+
+
+
+if __name__ == "__main__":
+    tn = TensorNetwork.new_random_cpu(2, 3, 4)
+    slice_dict = {0: slice(0, 2), 1: slice(1, 3)}
+    sliced_tn = tn.slice(slice_dict)
+    import pdb; pdb.set_trace()