Sketch out conj method #46

dallonasnes · 2023-10-09T23:01:28Z

Beginning of conj method

dallonasnes · 2023-10-09T23:01:44Z

qtensor/Simulate.py

+                if shared_indices.intersection(set(pair_node.get_all_dangling())):
+                    edge = tn.connect(node, pair_node)
+
+        # TODO: TNAdapter should support tensornetwork.Node


@danlkv this is my TODO for next week

dallonasnes · 2023-10-21T10:37:27Z

scratchpad/tn_api/tn.py


 class Array(np.ndarray):
-    shape: tuple[int]


removed all instances of tuple[... because was getting error that types can't have [

dallonasnes · 2023-10-21T10:38:22Z

scratchpad/tn_api/tn.py


    # slice not inplace
    def slice(self, slice_dict: dict) -> 'TensorNetwork':
-        ...
+        tn = self.copy()
+        sliced_buckets = np_framework.get_sliced_np_buckets(self.buckets, self.data_dict, slice_dict)


i imagine data_dict needs to be populated for this to be meaningful logic. is that something the callers should pass in? @danlkv

danlkv · 2023-10-21T16:19:45Z

The tensor class shouldn't contain buckets, it would be duplicating information, since we already have tensors property. The slice method should take the dict and for every index find those tensors that have the index. Then slice those tensors using the values in dict, and replace the original ones. Note also that the tensors are numpy arrays in this class

…

On Sat, Oct 21, 2023, 5:38 AM Dallon Asnes ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In scratchpad/tn_api/tn.py <#46 (comment)>: > # slice not inplace def slice(self, slice_dict: dict) -> 'TensorNetwork': - ... + tn = self.copy() + sliced_buckets = np_framework.get_sliced_np_buckets(self.buckets, self.data_dict, slice_dict) i imagine data_dict needs to be populated for this to be meaningful logic. is that something the callers should pass in? @danlkv <https://github.com/danlkv> — Reply to this email directly, view it on GitHub <#46 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADET2QO3G5X3AMHZEE5BB3LYAOQ2RAVCNFSM6AAAAAA5ZOEYEGVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMYTMOJRGIZTEOJYGI> . You are receiving this because you were mentioned.Message ID: ***@***.***>

dallonasnes · 2023-10-22T07:53:54Z

scratchpad/tn_api/tn.py


    # slice not inplace
    def slice(self, slice_dict: dict) -> 'TensorNetwork':
        tn = self.copy()
-        sliced_buckets = np_framework.get_sliced_np_buckets(self.buckets, self.data_dict, slice_dict)
-        tn.buckets = sliced_buckets
+        sliced_tns = []


@danlkv i think this implementation is better, wdyt?

Yes, this is much better. However, we need a way to refer to tensor indices globally. For example, what would happen if using this code we sliced a tensor with 3 tensors of shape (2, 2) and slice dict {1: 0}? The answer is that all tensors will be sliced along the 2nd dimension, but we want a behavior that is global to all TN

In addition, you have to consider how the edges change as well.

@dallonasnes Here's a sketch of what can be a solution.
First, let me clear out something about the edges structure. Edges in the edge array represent delta-tensors, which make all indices that are connected to them the same. For example for expression T_{abc}\delta_{abd} B_{edg}C_{mp} the equivalent expression is T_{xxc } B_{exg} C_{mp}. Note that you can have a one-to-one correspondence between each delta and an index (in the example index x).

Let's assume that all indices in our TN are represented by a delta tensor and its corresponding tuple in the edges attribute (This may seem excessive, but we can find overcome the overhead later). Now, a slice dict may refer to the element in the edges attribute by using just index of the edge.

So the slice dict is {index in edge dict: index value}

The algorithm would be
1 For each index, value pair of slice dict

Pop the edge from edges attr using the index

Get all tensors that are indexed by this edge (in the example above that would be T and B). Use Port.tensor_ix for that.

Using the information in ports in edges, construct the local (wrt each tensor) slicing dict for each tensor

Slice and update the tensor.

Note that after slicing the edge disappears.

dallonasnes · 2023-10-30T16:16:36Z

scratchpad/tn_api/tn.py

+        self._indices = {}
+
+    # slice not inplace
+    def slice(self, slice_dict: dict) -> 'TensorNetwork':


@danlkv thanks for the previous comment, i believed i've addressed that in this update

scratchpad/tn_api/tn.py

dallonasnes · 2023-10-30T22:42:08Z

scratchpad/tn_api/tn.py

+
+    # contract to produce a new tensor
+    def contract(self, contraction_info: ContractionInfo) -> np.ndarray:
+        raise NotImplementedError()


contraction_info here is indices i'd want in the output
einsum

test case is contracted A = contraction of sliceA[0] and sliceA[1]. this is true for any index so can select it at random

dallonasnes · 2023-12-04T12:50:12Z

scratchpad/tn_api/tn.py

+                sliced_tensor = tn._tensors[current_tensor_ref][tuple(slice_bounds)]
+                tn._tensors[current_tensor_ref] = sliced_tensor
+
+                # TODO: this is just a guess - but i am adding the ports from the popped edge back to the list of slices


@danlkv where we left off last time is that calling einsum is failing on a sliced tn. it's likely because slicing the tn removes all edges - i don't think that behavior is intended, so my proposal for a fix is that in this section of logic (starting at line 76) when we pop an edge out, at some point we need to add it back.

i've added some pdb loops so that when we err out, we can repro it in the same run

here's an example error output

aa,ba->ab [(4, 4, 2), (4, 4, 2)] ((Port(tensor_ref=0, ix=0), Port(tensor_ref=0, ix=1), Port(tensor_ref=1, ix=1)), (Port(tensor_ref=1, ix=0),)) 2 operand has more dimensions than subscripts given in einstein sum, but no '...' ellipsis provided to broadcast the extra dimensions. [1] > /app/scratchpad/tn_api/tn.py(172)contract() -> keep_going = True (Pdb++) c Traceback (most recent call last): File "tn_api/tn.py", line 168, in contract return np.einsum(einsum_expr, *self._tensors) File "<__array_function__ internals>", line 200, in einsum File "/usr/local/lib/python3.8/dist-packages/numpy/core/einsumfunc.py", line 1371, in einsum return c_einsum(*operands, **kwargs) ValueError: operand has more dimensions than subscripts given in einstein sum, but no '...' ellipsis provided to broadcast the extra dimensions.

Sketch out conj method

eeba989

dallonasnes requested a review from danlkv October 9, 2023 23:01

dallonasnes commented Oct 9, 2023

View reviewed changes

danlkv and others added 2 commits October 18, 2023 00:03

add draft of tn api

69d0341

Use np sliced buckets method from qtree

0d74462

dallonasnes commented Oct 21, 2023

View reviewed changes

Update

b0645a2

dallonasnes force-pushed the conj-api branch from 55dda4f to b0645a2 Compare October 22, 2023 07:53

dallonasnes commented Oct 22, 2023

View reviewed changes

Address comments

d7e24c2

dallonasnes commented Oct 30, 2023

View reviewed changes

scratchpad/tn_api/tn.py Outdated Show resolved Hide resolved

dallonasnes commented Oct 30, 2023

View reviewed changes

scratchpad/tn_api/tn.py Outdated Show resolved Hide resolved

dallonasnes commented Oct 30, 2023

View reviewed changes

scratchpad/tn_api/tn.py Show resolved Hide resolved

dallonasnes commented Oct 30, 2023

View reviewed changes

dallonasnes added 10 commits November 6, 2023 10:43

Cleanup

1aafd05

progress

3533289

Bugfixes

ea2a43c

Progress

0c91a8f

Progress

dbeee3d

Notes

1d7271b

Cleanup

0026283

Progress after meeting

b48ea18

Fix copy so that it is not in place

efbab98

Stuck

1631afb

dallonasnes commented Dec 4, 2023

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Sketch out conj method #46

Sketch out conj method #46

dallonasnes commented Oct 9, 2023

dallonasnes Oct 9, 2023

dallonasnes Oct 21, 2023

dallonasnes Oct 21, 2023

danlkv commented Oct 21, 2023 via email

dallonasnes Oct 22, 2023

danlkv Oct 23, 2023

danlkv Oct 23, 2023

danlkv Oct 23, 2023

dallonasnes Oct 30, 2023

dallonasnes Oct 30, 2023

dallonasnes Oct 30, 2023

dallonasnes Dec 4, 2023

Sketch out conj method #46

Are you sure you want to change the base?

Sketch out conj method #46

Conversation

dallonasnes commented Oct 9, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

danlkv commented Oct 21, 2023 via email

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment