mathai 2.5.2__tar.gz → 2.5.3__tar.gz

{mathai-2.5.2 → mathai-2.5.3}/PKG-INFO

@@ -1,13 +1,15 @@
 Metadata-Version: 2.4
 Name: mathai
-Version: 2.5.2
+Version: 2.5.3
 Summary: Mathematics solving Ai tailored to NCERT
 Home-page: https://github.com/infinity390/mathai
 Requires-Python: >=3.7
 Description-Content-Type: text/markdown
+Requires-Dist: numpy
 Dynamic: description
 Dynamic: description-content-type
 Dynamic: home-page
+Dynamic: requires-dist
 Dynamic: requires-python
 Dynamic: summary
 

{mathai-2.5.2 → mathai-2.5.3}/mathai/__init__.py

@@ -28,6 +28,7 @@ from .expand import expand
 
 from .ml import NeuralNetwork, diff_matrix_matrix
 from .mlscalar import NeuralNetworkScalar
+from .mldirect import NeuralNetworkDirect
 
 from .parser import parse, remove_extra_brackets
 

{mathai-2.5.2 → mathai-2.5.3}/mathai/ml.py

@@ -675,7 +675,7 @@ class NeuralNetwork:
                 A_list = A_list + flatten_list(item)
             exec(s, env2)
             return env2["fx"](A_list,flatten_list(given_x))
-    def train(self, train_x, train_y, learning_rate, epoch, batch_size=1):
+    def train(self, train_x, train_y, learning_rate=12, epoch=10000, batch_size=1):
         global hadamard, zeros, transpose, matadd, sigmoid, identity, power, vec, diag, flatten,\
                matmul, reshape, kronecker, commutation, broadcast, im2col, conv, col2im
         print("gradients calculated")

mathai-2.5.3/mathai/mldirect.py

@@ -0,0 +1,1018 @@
+from .formula_compiler import formula_list_compiler
+import itertools
+from .parser import parse, remove_extra_brackets
+from .simplify import simplify, addition_node_mat
+from .ml import diff_matrix_matrix
+from .base import *
+import random
+import copy
+import math
+import numpy as np
+class TreeNode2:
+    count = 0
+    count_x = 0
+    count_y = 0
+    def __new__(cls, name=None, children=None):
+        if name is None:
+            return super().__new__(cls)
+        if children and len(children) == 2:
+            c1, c2 = children[0], children[1]
+            if hasattr(c1, 'name') and c1.name.startswith("d_") and \
+               hasattr(c2, 'name') and c2.name.startswith("d_"):
+                try:
+                    val1 = int(c1.name[2:])
+                    val2 = int(c2.name[2:])
+                    res = None
+                    if name == "f_add":
+                        res = val1 + val2
+                    elif name == "f_sub":
+                        res = val1 - val2
+                    elif name == "f_mul":
+                        res = val1 * val2
+                    elif name == "f_pow":
+                        if val2 >= 0:
+                            res = val1 ** val2
+                        else:
+                            if val1 == 1:
+                                res = 1
+                            elif val1 == -1:
+                                res = 1 if val2 % 2 == 0 else -1
+                    if res is not None:
+                        obj = super().__new__(cls)
+                        obj.name = f"d_{res}"
+                        obj.children = []
+                        return obj
+                except ValueError:
+                    pass
+        if name == "f_mul" and children and any(hasattr(child, 'name') and child.name == "d_0" for child in children):
+            obj = super().__new__(cls)
+            obj.name = "d_0"
+            obj.children = []
+            return obj
+        if name == "f_mul" and children and len(children) == 2:
+            if hasattr(children[0], 'name') and children[0].name == "d_1":
+                return children[1]
+            if hasattr(children[1], 'name') and children[1].name == "d_1":
+                return children[0]
+        if name == "f_add" and children and len(children) == 2:
+            if hasattr(children[0], 'name') and children[0].name == "d_0":
+                return children[1]
+            if hasattr(children[1], 'name') and children[1].name == "d_0":
+                return children[0]
+        return super().__new__(cls)
+    def __init__(self, name=None, children=None):
+        if hasattr(self, 'name'):
+            return
+        if name is None:
+            return
+        self.name = name
+        self.children = children
+        self.compare = None
+    def normal(self):
+        commutative={"f_add","f_mul"}
+        stack=[(self,False)]
+        while stack:
+            node,visited=stack.pop()
+            if visited:
+                if node.children:
+                    if node.name in commutative:
+                        node.children.sort(key=lambda c:c.compare)
+                    node.compare=(node.name,*(c.compare for c in node.children))
+                else:
+                    node.compare=(node.name,)
+            else:
+                stack.append((node,True))
+                stack.extend((c,False) for c in node.children)
+    def __repr__(self):
+        action_stack = [(self, False)]
+        value_stack = []
+        while action_stack:
+            node, children_processed = action_stack.pop()
+            if node.name.startswith("v_"):
+                var_idx = node.name.split("_")[1]
+                value_stack.append(f"A[{var_idx}]")
+                continue
+            if node.name.startswith("x_"):
+                var_idx = node.name.split("_")[1]
+                value_stack.append(f"X[{var_idx}]")
+                continue
+            if node.name.startswith("y_"):
+                var_idx = node.name.split("_")[1]
+                value_stack.append(f"Y[{var_idx}]")
+                continue
+            if node.name.startswith("w_"):
+                value_stack.append(node.name.split("_")[1])
+                continue
+            if node.name.startswith("d_"):
+                value_stack.append(node.name.split("_")[1])
+                continue
+            if not children_processed:
+                action_stack.append((node, True))
+                for child in reversed(node.children):
+                    action_stack.append((child, False))
+                continue
+            num_children = len(node.children)
+            child_strs = []
+            for _ in range(num_children):
+                child_strs.append(value_stack.pop())
+            child_strs.reverse()
+            if node.name == "f_add":
+                value_stack.append(f"({'+'.join(child_strs)})")
+            elif node.name == "f_mul":
+                value_stack.append(f"({'*'.join(child_strs)})")
+            elif node.name == "f_pow":
+                value_stack.append(f"({'**'.join(child_strs)})")
+            elif node.name == "f_div":
+                value_stack.append(f"({'/'.join(child_strs)})")
+            elif node.name == "f_exp":
+                value_stack.append(f"exp({child_strs[0]})")
+            elif node.name == "f_neg":
+                value_stack.append(f"(-{child_strs[0]})")
+            elif node.name == "f_sigmoid":
+                value_stack.append(f"sigmoid({child_strs[0]})")
+            else:
+                value_stack.append(f"{node.name}({','.join(child_strs)})")
+        return value_stack[0]
+def zeros(*shape):
+    if len(shape) == 0:
+        return TreeNode2("d_0",[])
+    return [zeros(*shape[1:]) for _ in range(shape[0])]
+def flatten(A):
+    B = len(A)
+    H = len(A[0])
+    W = len(A[0][0])
+    C = len(A[0][0][0])
+    out = []
+    for b in range(B):
+        row = []
+        for i in range(H):
+            for j in range(W):
+                for c in range(C):
+                    row.append(A[b][i][j][c])
+        out.append(row)
+    return out
+def conv(X, K):
+    B = len(X)
+    H = len(X[0])
+    W = len(X[0][0])
+    C = len(X[0][0][0])
+    P = len(K[0])
+    OH = H - P + 1
+    OW = W - P + 1
+    out = []
+    for b in range(B):
+        batch = []
+        for i in range(OH):
+            row = []
+            for j in range(OW):
+                s = TreeNode2("d_0",[])
+                for u in range(P):
+                    for v in range(P):
+                        for c in range(C):
+                            s = TreeNode2("f_add", [s, TreeNode2("f_mul",[X[b][i + u][j + v][c] , K[0][u][v][c]])])
+                row.append([s])
+            batch.append(row)
+        out.append(batch)
+    return out
+def im2col(image, kernel):
+    B = len(image)
+    H = len(image[0])
+    W = len(image[0][0])
+    C = len(image[0][0][0])
+    Kh = len(kernel[0])
+    Kw = len(kernel[0][0])
+    out_h = H - Kh + 1
+    out_w = W - Kw + 1
+    P = Kh * Kw * C
+    cols = []
+    for b in range(B):
+        for i in range(out_h):
+            for j in range(out_w):
+                row = [TreeNode2("d_0",[]) for _ in range(B * P)]
+                idx = b * P
+                for r in range(Kh):
+                    for c in range(Kw):
+                        for ch in range(C):
+                            row[idx] = image[b][i+r][j+c][ch]
+                            idx += 1
+                cols.append(row)
+    return cols
+def col2im(kernel, image):
+    B = len(image)
+    H = len(image[0])
+    W = len(image[0][0])
+    C = len(image[0][0][0])
+    K = len(kernel)
+    Kh = len(kernel[0])
+    Kw = len(kernel[0][0])
+    assert K == 1 or K == B
+    OH = H - Kh + 1
+    OW = W - Kw + 1
+    rows = B * OH * OW
+    cols = B * H * W * C
+    J = [[TreeNode2("d_0",[]) for _ in range(cols)] for _ in range(rows)]
+    row = 0
+    for b in range(B):
+        kb = 0 if K == 1 else b
+        for i in range(OH):
+            for j in range(OW):
+                for u in range(Kh):
+                    for v in range(Kw):
+                        for ch in range(C):
+                            col = (
+                                (((b * H + (i + u)) * W + (j + v)) * C)
+                                + ch
+                            )
+                            J[row][col] = kernel[kb][u][v][ch]
+                row += 1
+    return J
+def commutation(m, n):
+    size = m * n
+    K = [[TreeNode2("d_0", []) for _ in range(size)] for _ in range(size)]
+    for i in range(m):
+        for j in range(n):
+            src_idx = j * m + i
+            dest_idx = i * n + j
+            K[dest_idx][src_idx] = TreeNode2("d_1",[])
+    return K
+def hadamard_h(A, B):
+    if isinstance(A, list) and isinstance(B, list):
+        assert len(A) == len(B)
+        return [hadamard_h(a, b) for a, b in zip(A, B)]
+    return TreeNode2("f_mul",[A,B])
+def hadamard(*lst_prod):
+    result = lst_prod[0]
+    for x in lst_prod[1:]:
+        if isinstance(result, list) and isinstance(x, list):
+            result = hadamard_h(result, x)
+        elif isinstance(result, list):
+            result = apply(result, lambda y: TreeNode2("f_mul",[y,x]))
+        elif isinstance(x, list):
+            result = apply(x, lambda y: TreeNode2("f_mul",[result,y]))
+        else:
+            result = TreeNode2("f_mul",[result,x])
+    return result
+def matadd_h(A, B):
+    if isinstance(A, list) and isinstance(B, list):
+        assert len(A) == len(B)
+        return [matadd_h(a, b) for a, b in zip(A, B)]
+    return TreeNode2("f_add",[A,B])
+def matadd(*lst_prod):
+    result = lst_prod[0]
+    for x in lst_prod[1:]:
+        if isinstance(result, list) and isinstance(x, list):
+            result = matadd_h(result, x)
+        elif isinstance(result, list):
+            result = apply(result, lambda y: TreeNode2("f_add",[y,x]))
+        elif isinstance(x, list):
+            result = apply(x, lambda y: TreeNode2("f_add",[result,y]))
+        else:
+            result = TreeNode2("f_add",[result,x])
+    return result
+def matmul_h(A, B):
+    assert len(A[0])==len(B)
+    rows = len(A)
+    inner = len(A[0])
+    cols = len(B[0])
+    C = []
+    for i in range(rows):
+        row = []
+        for j in range(cols):
+            expr = TreeNode2("d_0",[])
+            for k in range(inner):
+                left = A[i][k]
+                right = B[k][j]
+                expr = TreeNode2("f_add",[expr,TreeNode2("f_mul",[left,right])])
+            row.append(expr)
+        C.append(row)
+    return C
+def depth(x):
+    d = 0
+    while isinstance(x, list):
+        d += 1
+        x = x[0]
+    return d
+def matmul(*lst_prod):
+    result = lst_prod[0]
+    for index, x in enumerate(lst_prod[1:]):
+        if not isinstance(result, list) or not isinstance(x, list):
+            if isinstance(result, list):
+                result = apply(result, lambda y: TreeNode2("f_mul",[y,x]))
+            elif isinstance(x, list):
+                result = apply(x, lambda y: TreeNode2("f_mul",[result,y]))
+            else:
+                result = TreeNode2("f_mul",[result,x])
+            continue
+        r1 = depth(result)
+        r2 = depth(x)
+        if r1 == 2 and r2 == 2:
+            result = matmul_h(result, x)
+        elif r1 == 3 and r2 == 2:
+            result = [matmul_h(A, x) for A in result]
+        elif r1 == 2 and r2 == 3:
+            result = [matmul_h(result, B) for B in x]
+        elif r1 == 3 and r2 == 3:
+            if len(result) == 1:
+                result = [matmul_h(result[0], B) for B in x]
+            elif len(x) == 1:
+                result = [matmul_h(A, x[0]) for A in result]
+            else:
+                assert len(result) == len(x)
+                result = [
+                    matmul_h(A, B)
+                    for A, B in zip(result, x)
+                ]
+        else:
+            raise ValueError(
+                f"Unsupported matmul ranks ({r1}, {r2})"
+            )
+    return result
+def kronecker_h(A, B):
+    m = len(A)
+    n = len(A[0])
+    p = len(B)
+    q = len(B[0])
+    result = [[TreeNode2("d_0", []) for _ in range(n * q)] for _ in range(m * p)]
+    for j in range(n):
+        for c in range(q):
+            col_target = (j*q)+c
+            for i in range(m):
+                for r in range(p):
+                    row_target = (i*p)+r
+                    result[row_target][col_target] = TreeNode2("f_mul",[A[i][j],B[r][c]])
+    return result
+def kronecker(*lst_prod):
+    result = lst_prod[0]
+    for x in lst_prod[1:]:
+        result = kronecker_h(result, x)
+    return result
+def transpose(A):
+    rows = len(A)
+    cols = len(A[0])
+    return [[A[i][j] for i in range(rows)] for j in range(cols)]
+def apply(arr, fx):
+    if isinstance(arr, list):
+        return [apply(item, fx) for item in arr]
+    return fx(arr)
+def sigmoid(arg):
+    def fx_apply(x):
+        return TreeNode2("f_sigmoid",[x])
+    return apply(arg, fx_apply)
+def identity(size):
+    arr = zeros(size,size)
+    for i in range(size):
+        for j in range(size):
+            if i == j:
+                arr[i][j] = TreeNode2("d_1",[])
+    return arr
+def vec(A):
+    out = []
+    def visit(x):
+        if isinstance(x, list):
+            for y in x:
+                visit(y)
+        else:
+            out.append([x])
+    visit(A)
+    return out
+def diag(v):
+    if len(v) > 0 and isinstance(v[0], list):
+        if len(v[0]) != 1:
+            raise ValueError("diag expects a column vector.")
+        v = [x[0] for x in v]
+    for x in v:
+        if isinstance(x, list):
+            raise ValueError("diag received a tensor instead of a vector.")
+    n = len(v)
+    result = [[TreeNode2("d_0", []) for _ in range(n)] for _ in range(n)]
+    for i in range(n):
+        result[i][i] = v[i]
+    return result
+def broadcast(M, rows):
+    if not isinstance(M, list):
+        raise ValueError("Cannot broadcast a scalar.")
+    r = len(M)
+    if r == rows:
+        return M
+    if r != 1:
+        raise ValueError(f"Cannot broadcast first dimension {r} to {rows}")
+    return [copy.deepcopy(M[0]) for _ in range(rows)]
+def reshape(data, *shape):
+    def get_shape(x):
+        if not isinstance(x, list):
+            return []
+        return [len(x)] + get_shape(x[0])
+    in_shape = get_shape(data)
+    def get_element(x, indices):
+        for idx in indices:
+            x = x[idx]
+        return x
+    flat = []
+    if in_shape:
+        num_elements = 1
+        for s in in_shape:
+            num_elements *= s
+        for i in range(num_elements):
+            indices = [0] * len(in_shape)
+            rem = i
+            for d in reversed(range(len(in_shape))):
+                indices[d] = rem % in_shape[d]
+                rem //= in_shape[d]
+            flat.append(get_element(data, indices))
+    else:
+        flat = [data]
+    total = 1
+    for s in shape:
+        total *= s
+    if len(flat) != total:
+        raise ValueError(f"Cannot reshape: incompatible dimensions. Total size {len(flat)} vs {total}")
+    def make_empty(shp):
+        if len(shp) == 1:
+            return [None] * shp[0]
+        return [make_empty(shp[1:]) for _ in range(shp[0])]
+    out_structure = make_empty(shape)
+    def set_element(x, indices, val):
+        for idx in indices[:-1]:
+            x = x[idx]
+        x[indices[-1]] = val
+    for i in range(total):
+        indices = [0] * len(shape)
+        rem = i
+        for d in reversed(range(len(shape))):
+            indices[d] = rem % shape[d]
+            rem //= shape[d]
+        set_element(out_structure, indices, flat[i])
+    return out_structure
+def gen2(eq, w):
+    def from_treenode(eq,keep_integer=False):
+        nonlocal w
+        alter = {"f_wadd":"matadd", "f_transpose":"transpose", "f_diag":"diag", "f_identity":"identity","f_wmul":"matmul",\
+                 "f_hadamard":"hadamard","f_commutation":"commutation","f_kronecker":"kronecker",\
+                 "f_conv":"conv", "f_im2col":"im2col","f_col2im":"col2im", "f_flatten":"flatten",\
+                 "f_sigmoid":"sigmoid", "f_vec":"vec"\
+                 }
+        if eq.name in alter.keys():
+            return alter[eq.name]+"("+",".join([from_treenode(child,keep_integer) for child in eq.children])+")"
+        if eq.name in ["f_reshape", "f_broadcast"]:
+            return eq.name[2:]+"("+",".join([from_treenode(eq.children[0])]+\
+                                            [from_treenode(child,True) for child in eq.children[1:]])+")"
+        if eq.name in ["f_zeros","f_len"]:
+            return eq.name[2:]+"("+",".join([from_treenode(child,True) for child in eq.children])+")"
+        alter2 = ["f_add", "f_pow", "f_mul"]
+        if eq.name in alter2:
+            return f"TreeNode2('{eq.name}',["+",".join([from_treenode(child,keep_integer) for child in eq.children])+"])"
+        if not keep_integer and eq.name in [parse("m").name, parse("n").name]:
+            return "TreeNode2(f'd_{m}',[])"
+        if eq.name == "f_index":
+            return from_treenode(eq.children[0],keep_integer)+"["+str(eq.children[1])+"]"
+        if eq in w:
+            return f"w[{w.index(eq)}]"
+        if not keep_integer and eq.name.startswith("d_"):
+            return f"TreeNode2('d_{eq.name[2:]}',[])"
+        return str(eq)
+    return from_treenode(eq)
+def gpu_var(*shape):
+    if len(shape) == 0:
+        TreeNode2.count += 1
+        return TreeNode2(f"v_{TreeNode2.count-1}", [])
+    return [gpu_var(*shape[1:]) for _ in range(shape[0])]
+def gpu_var_x(*shape):
+    if len(shape) == 0:
+        TreeNode2.count_x += 1
+        return TreeNode2(f"x_{TreeNode2.count_x-1}", [])
+    return [gpu_var_x(*shape[1:]) for _ in range(shape[0])]
+def gpu_var_y(*shape):
+    if len(shape) == 0:
+        TreeNode2.count_y += 1
+        return TreeNode2(f"y_{TreeNode2.count_y-1}", [])
+    return [gpu_var_y(*shape[1:]) for _ in range(shape[0])]
+def gpu_var_write(content,*shape):
+    def gpu_var_helper(*shape):
+        if len(shape) == 0:
+            return content.pop(0)
+        return [gpu_var_helper(*shape[1:]) for _ in range(shape[0])]
+    return gpu_var_helper(*shape)
+def flatten_list(nested_list):
+    flat = []
+    def _traverse(items):
+        for item in items:
+            if isinstance(item, list):
+                _traverse(item)
+            else:
+                flat.append(item)
+    _traverse(nested_list)
+    return flat
+def shape(nested_list):
+    if not isinstance(nested_list, list):
+        return []
+    return [len(nested_list)] + shape(nested_list[0])
+def flatten_list2(nested_list):
+    flat = []
+    def _traverse(items):
+        for item in items:
+            if isinstance(item, list):
+                _traverse(item)
+            else:
+                flat.append(item)
+    output = []
+    for item in nested_list:
+        flat = []
+        _traverse(item)
+        output += flat
+    return output
+
+def _t_const(k):
+    return TreeNode2(f"d_{k}", [])
+def _t_add(terms):
+    terms = [t for t in terms if t.name != "d_0"]
+    if not terms:
+        return _t_const(0)
+    r = terms[0]
+    for t in terms[1:]:
+        r = TreeNode2("f_add", [r, t])
+    return r
+def _t_mul(factors):
+    r = factors[0]
+    for f in factors[1:]:
+        r = TreeNode2("f_mul", [r, f])
+    return r
+def _diff_rule(node, dch, var):
+    name = node.name
+    ch = node.children or []
+    if not ch:
+        return _t_const(1) if name == var else _t_const(0)
+    if name == "f_add":
+        return _t_add(dch)
+    if name == "f_mul":
+        terms = []
+        for i in range(len(ch)):
+            if dch[i].name == "d_0":
+                continue
+            terms.append(_t_mul([dch[i]] + ch[:i] + ch[i+1:]))
+        return _t_add(terms)
+    if name == "f_pow":
+        base, expo = ch
+        if not expo.name.startswith("d_"):
+            raise NotImplementedError("derivative of a pow with a non-constant exponent is not supported")
+        k = int(expo.name[2:])
+        if k == 0 or dch[0].name == "d_0":
+            return _t_const(0)
+        if k == 1:
+            return dch[0]
+        return _t_mul([_t_const(k), TreeNode2("f_pow", [base, _t_const(k-1)]), dch[0]])
+    if name == "f_sigmoid":
+        if dch[0].name == "d_0":
+            return _t_const(0)
+        one_minus = TreeNode2("f_add", [_t_const(1), TreeNode2("f_mul", [_t_const(-1), node])])
+        return _t_mul([node, one_minus, dch[0]])
+    if name == "f_exp":
+        if dch[0].name == "d_0":
+            return _t_const(0)
+        return _t_mul([node, dch[0]])
+    if name == "f_neg":
+        return _t_mul([_t_const(-1), dch[0]])
+    if name == "f_div":
+        a, b = ch
+        t1 = _t_mul([dch[0], TreeNode2("f_pow", [b, _t_const(-1)])])
+        t2 = _t_mul([_t_const(-1), a, dch[1], TreeNode2("f_pow", [b, _t_const(-2)])])
+        return _t_add([t1, t2])
+    raise NotImplementedError(f"no derivative rule for {name}")
+def diff_node(root, var, memo):
+    stack = [root]
+    while stack:
+        node = stack[-1]
+        if id(node) in memo:
+            stack.pop()
+            continue
+        kids = node.children or []
+        pending = [c for c in kids if id(c) not in memo]
+        if pending:
+            stack.extend(pending)
+            continue
+        stack.pop()
+        memo[id(node)] = _diff_rule(node, [memo[id(c)] for c in kids], var)
+    return memo[id(root)]
+
+_DAG_OPS = {"f_add":"add", "f_mul":"mul", "f_pow":"pow", "f_div":"div",
+            "f_exp":"exp", "f_neg":"neg", "f_sigmoid":"sigmoid"}
+def compile_dag(roots):
+    index = {}
+    prog = []
+    for root in roots:
+        stack = [root]
+        while stack:
+            node = stack[-1]
+            if id(node) in index:
+                stack.pop()
+                continue
+            kids = node.children or []
+            pending = [c for c in kids if id(c) not in index]
+            if pending:
+                stack.extend(pending)
+                continue
+            stack.pop()
+            name = node.name
+            if not kids:
+                if name.startswith("d_"):
+                    entry = ("const", float(name[2:]))
+                elif name.startswith("v_"):
+                    entry = ("A", int(name[2:]))
+                elif name.startswith("x_"):
+                    entry = ("X", int(name[2:]))
+                elif name.startswith("y_"):
+                    entry = ("Y", int(name[2:]))
+                else:
+                    raise ValueError(f"cannot evaluate leaf {name}")
+            else:
+                if name not in _DAG_OPS:
+                    raise ValueError(f"cannot evaluate node {name}")
+                entry = (_DAG_OPS[name], tuple(index[id(c)] for c in kids))
+            index[id(node)] = len(prog)
+            prog.append(entry)
+    return prog, [index[id(r)] for r in roots]
+def sigmoid_num(x):
+    return 0.5 * (1.0 + np.tanh(0.5 * x))
+def run_dag(prog, outs, A, X, Y):
+    vals = [0.0] * len(prog)
+    with np.errstate(all="ignore"):
+        for i, (op, arg) in enumerate(prog):
+            if op == "const":
+                vals[i] = arg
+            elif op == "A":
+                vals[i] = A[arg]
+            elif op == "X":
+                vals[i] = X[arg]
+            elif op == "Y":
+                vals[i] = Y[arg]
+            elif op == "add":
+                s = 0.0
+                for k in arg:
+                    s += vals[k]
+                vals[i] = s
+            elif op == "mul":
+                p = 1.0
+                for k in arg:
+                    p *= vals[k]
+                vals[i] = p
+            elif op == "pow":
+                vals[i] = np.power(vals[arg[0]], vals[arg[1]])
+            elif op == "div":
+                vals[i] = np.divide(vals[arg[0]], vals[arg[1]])
+            elif op == "exp":
+                vals[i] = np.exp(vals[arg[0]])
+            elif op == "neg":
+                vals[i] = -vals[arg[0]]
+            elif op == "sigmoid":
+                vals[i] = sigmoid_num(vals[arg[0]])
+    return [float(vals[k]) for k in outs]
+
+class NeuralNetworkDirect:
+    def __init__(self, struct):
+        self.struct = struct
+        self.var_list = [tree_form(f"v_-{i}") for i in range(1,26+1-4)]
+        self.o = None
+        self.lst_w = None
+        self.gradient = None
+        self.weight_equation = None
+        self.hessian_equation = None
+        self.critical_points = None
+        self.critical_point_info = None
+        self.learn = None
+        self.learn_sym = None
+        self.learn_var = None
+        self.loss = None
+        self.model_type = None
+        self.init_mat = lambda *args: gpu_var(*args)
+        self._eq_key = None
+        self._flat_grad = None
+        self._grad_prog = None
+        self._grad_outs = None
+        self._hess_prog = None
+        self._hess_outs = None
+        self.loss_equation = None
+        self._loss_prog = None
+        self._loss_outs = None
+        self.minima = None
+        self.num_gradients = 0
+        self.diff_matrix_count = 0
+        self.weight_equation_count = 0
+        TreeNode2.count = 0
+        TreeNode2.count_x = 0
+        TreeNode2.count_y = 0
+    def model(self, t="image"):
+        self.model_type = t
+        if t == "image":
+            return self.model_image()
+    def model_image(self):
+        lst_z = []
+        lst_w = []
+        x = parse("X")
+        y = parse("Y")
+        lst_z.append(x)
+        flattening_needed = False
+        for i in range(1,len(self.struct)):
+            if self.struct[i]["type"] == "convolution":
+                flattening_needed = True
+                lst_w.append(self.var_list.pop(0))
+                eq = TreeNode("f_broadcast", [lst_w[-1], parse("m")])
+                tmp = TreeNode("f_conv", [lst_z[-1], eq])
+                tmp = replace(self.struct[i]["activation"], parse("Z"), tmp)
+                lst_z[-1] = tmp
+            elif self.struct[i]["type"] == "dense":
+                lst_w.append(self.var_list.pop(0))
+                lst_w.append(self.var_list.pop(0))
+                if flattening_needed:
+                    flattening_needed = False
+                    lst_z[-1] = lst_z[-1].fx("flatten")
+                eq = TreeNode("f_wmul", [lst_z[-1], lst_w[-2]])
+                eq = TreeNode("f_wadd", [eq, TreeNode("f_broadcast", [lst_w[-1], parse("m")])])
+                lst_z.append(eq)
+                tmp = replace(self.struct[i]["activation"], parse("Z"), lst_z[-1])
+                lst_z[-1] = tmp
+        self.lst_w = lst_w
+        self.o = lst_z[-1]
+        eq = TreeNode("f_hadamard", [tree_form("d_-1"), y])
+        eq = TreeNode("f_wadd", [self.o, eq])
+        eq = eq.fx("vec")
+        eq = TreeNode("f_wmul", [eq.fx("transpose"), eq])
+        L = TreeNode("f_hadamard", [eq, (parse("m") * tree_form("d_2")) ** tree_form("d_-1")])
+        self.loss = L
+        gradient = []
+        index = 0
+        self.learn = []
+        for i in range(1,len(self.struct)):
+            if self.struct[i]["type"] == "dense":
+                prev = None
+                if self.struct[i-1]["type"] == "convolution":
+                    a,b,c = self.struct[0]["dim"]
+                    for j in range(1,i):
+                        d,e,_ = self.struct[j]["dim"]
+                        a,b = a-d + 1, b-e + 1
+                    prev = a*b
+                else:
+                    prev = self.struct[i-1]["dim"][0]
+                curr = self.struct[i]["dim"][0]
+                self.learn.append(self.init_mat(prev, curr))
+                self.learn.append(self.init_mat(1, curr))
+            elif self.struct[i]["type"] == "convolution":
+                self.learn.append(self.init_mat(*([1]+self.struct[i]["dim"])))
+        self.learn_sym = self.learn
+        self.num_gradients = len(lst_w)
+        self.diff_matrix_count = 0
+        print(f"gradients to calculate: {self.num_gradients}")
+        for j in range(len(lst_w)):
+            item = lst_w[j]
+            tmp = TreeNode("f_pdif", [L, item.fx("vec")])
+            tmp = diff_matrix_matrix(tmp)
+            self.diff_matrix_count += 1
+            print(f"diff_matrix_matrix {self.diff_matrix_count}/{self.num_gradients} done")
+            d = []
+            k = self.learn[j]
+            while isinstance(k, list):
+                d.append(len(k))
+                k = k[0]
+            tmp = TreeNode("f_reshape", [tmp]+[tree_form(f"d_{item}") for item in d])
+            gradient.append(tmp)
+        self.gradient = gradient
+        return self
+    def predict(self, given_x):
+        global hadamard, zeros, transpose, matadd, sigmoid, identity, vec, diag, flatten,\
+               matmul, reshape, kronecker, commutation, broadcast, im2col, conv, col2im, extract_values, shape
+        TreeNode2.count = 0
+        TreeNode2.count_x = 0
+        TreeNode2.count_y = 0
+        learn2 = []
+        for item in self.learn:
+            learn2.append(gpu_var(*shape(item)))
+        given_x = [given_x]
+        given2 = gpu_var_x(*shape(given_x))
+        env = {
+            "w": learn2,
+            "m": 1,
+            "identity":identity,
+            "X": given2,
+            "zeros":zeros,
+            "hadamard": hadamard,
+            "sigmoid":sigmoid,
+            "matadd": matadd,
+            "diag":diag,
+            "vec":vec,
+            "matmul":matmul,
+            "reshape":reshape,
+            "kronecker":kronecker,
+            "commutation":commutation,
+            "transpose":transpose,
+            "broadcast":broadcast,
+            "conv":conv,
+            "im2col":im2col,
+            "col2im": col2im,
+            "flatten":flatten
+        }
+        if self.model_type == "image":
+            tmp = gen2(self.o, self.lst_w)
+            eq = eval(tmp, {}, env)[0]
+            s = "fx = lambda A,X: "+str(eq)
+            env2 = {
+                "exp": np.exp,
+                "sigmoid": sigmoid_num
+            }
+            A_list = []
+            for item in self.learn:
+                A_list = A_list + flatten_list(item)
+            exec(s, env2)
+            return env2["fx"](A_list,flatten_list(given_x))
+    def build_equations(self, train_x, train_y, batch_size):
+        key = (batch_size, tuple(shape(train_x[:batch_size])), tuple(shape(train_y[:batch_size])))
+        if self._eq_key == key and self.weight_equation is not None:
+            return
+        TreeNode2.count_x = 0
+        TreeNode2.count_y = 0
+        data_x = gpu_var_x(*shape(train_x[:batch_size]))
+        data_y = gpu_var_y(*shape(train_y[:batch_size]))
+        env = {
+            "identity":identity,
+            "sigmoid":sigmoid,
+            "matadd": matadd,
+            "zeros":zeros,
+            "hadamard": hadamard,
+            "diag":diag,
+            "vec":vec,
+            "matmul":matmul,
+            "reshape":reshape,
+            "kronecker":kronecker,
+            "commutation":commutation,
+            "transpose":transpose,
+            "broadcast":broadcast,
+            "conv":conv,
+            "im2col":im2col,
+            "col2im":col2im,
+            "flatten":flatten,
+            "TreeNode2":TreeNode2,
+            "m":batch_size,
+            "w":self.learn_sym,
+            "X":data_x,
+            "Y":data_y
+        }
+        self.weight_equation_count = 0
+        print(f"gradients: {self.num_gradients}, diff_matrix_matrix done: {self.diff_matrix_count}/{self.num_gradients}")
+        weight_equation = []
+        for g in self.gradient:
+            weight_equation.append(eval(gen2(g, self.lst_w), {}, env))
+            self.weight_equation_count += 1
+            print(f"weight equation {self.weight_equation_count}/{self.num_gradients} done")
+        self.weight_equation = weight_equation
+        self._flat_grad = flatten_list(self.weight_equation)
+        n_weights = len(flatten_list(self.learn_sym))
+        assert len(self._flat_grad) == n_weights, "gradient size does not match number of weights"
+        self._grad_prog, self._grad_outs = compile_dag(self._flat_grad)
+        loss_flat = flatten_list(eval(gen2(self.loss, self.lst_w), {}, env))
+        assert len(loss_flat) == 1, "loss must be a scalar"
+        self.loss_equation = loss_flat[0]
+        self._loss_prog, self._loss_outs = compile_dag([self.loss_equation])
+        self.hessian_equation = None
+        self._hess_prog = None
+        self._hess_outs = None
+        self.critical_points = None
+        self.critical_point_info = None
+        self._eq_key = key
+    def _eval_gradient(self, A, X, Y):
+        return run_dag(self._grad_prog, self._grad_outs, A, X, Y)
+    def loss_at(self, A, X, Y):
+        return run_dag(self._loss_prog, self._loss_outs, A, X, Y)[0]
+    def _unflatten(self, flat):
+        it = iter(flat)
+        def build(shp):
+            if not shp:
+                return next(it)
+            return [build(shp[1:]) for _ in range(shp[0])]
+        return [build(shape(item)) for item in self.learn_sym]
+    def build_hessian(self):
+        if self._flat_grad is None:
+            raise RuntimeError("call build_equations / find_critical_points / train first")
+        flat = self._flat_grad
+        N = len(flat)
+        H = [[None] * N for _ in range(N)]
+        step = max(1, N // 10)
+        print(f"building hessian: {N} x {N}")
+        for j in range(N):
+            if j % step == 0:
+                print(f"hessian column {j}/{N}")
+            memo = {}
+            var = f"v_{j}"
+            for i in range(j + 1):
+                H[i][j] = diff_node(flat[i], var, memo)
+                H[j][i] = H[i][j]
+        self.hessian_equation = H
+        roots = [H[i][j] for j in range(N) for i in range(j + 1)]
+        self._hess_prog, self._hess_outs = compile_dag(roots)
+        return H
+    def hessian_at(self, A, X, Y):
+        if self._hess_prog is None:
+            raise RuntimeError("call build_hessian() first")
+        vals = run_dag(self._hess_prog, self._hess_outs, A, X, Y)
+        N = len(self._flat_grad)
+        Hn = np.empty((N, N))
+        k = 0
+        for j in range(N):
+            for i in range(j + 1):
+                Hn[i, j] = Hn[j, i] = vals[k]
+                k += 1
+        return Hn
+    def _newton(self, x, X, Y, grad_tol, max_iter):
+        g = np.array(self._eval_gradient(x.tolist(), X, Y))
+        for it in range(max_iter):
+            if not np.all(np.isfinite(g)):
+                return x, it, False
+            gnorm = np.linalg.norm(g, np.inf)
+            if gnorm < grad_tol:
+                return x, it, True
+            Hn = self.hessian_at(x.tolist(), X, Y)
+            if not np.all(np.isfinite(Hn)):
+                return x, it, False
+            try:
+                dx = np.linalg.solve(Hn, -g)
+            except np.linalg.LinAlgError:
+                dx = np.linalg.lstsq(Hn, -g, rcond=None)[0]
+            t = 1.0
+            while True:
+                x_new = x + t * dx
+                g_new = np.array(self._eval_gradient(x_new.tolist(), X, Y))
+                if np.all(np.isfinite(g_new)) and np.linalg.norm(g_new, np.inf) < gnorm:
+                    break
+                t *= 0.5
+                if t < 1e-10:
+                    return x, it, False
+            x, g = x_new, g_new
+        return x, max_iter, bool(np.linalg.norm(g, np.inf) < grad_tol)
+    def find_critical_points(self, train_x, train_y, n_starts=50, init_scale=1.0, seed=None,
+                             grad_tol=1e-8, dedup_tol=1e-5, max_iter=100):
+        batch_size = len(train_x)
+        self.build_equations(train_x, train_y, batch_size)
+        if self._hess_prog is None:
+            self.build_hessian()
+        X = flatten_list(train_x)
+        Y = flatten_list(train_y)
+        N = len(self._flat_grad)
+        rng = np.random.default_rng(seed)
+        found = []
+        for s in range(n_starts):
+            x0 = rng.normal(0.0, init_scale, N)
+            x, iters, ok = self._newton(x0, X, Y, grad_tol, max_iter)
+            if s % n_starts//10 == 0:
+                print(f"newton start {s+1}/{n_starts}: {'converged' if ok else 'failed'} after {iters} iterations")
+            if not ok:
+                continue
+            if any(np.linalg.norm(x - p) <= dedup_tol * (1.0 + np.linalg.norm(p)) for p in found):
+                continue
+            found.append(x)
+        self.critical_points = found
+        return found
+    def classify_critical_points(self, train_x, train_y, points=None, eig_tol=1e-6):
+        batch_size = len(train_x)
+        self.build_equations(train_x, train_y, batch_size)
+        if self._hess_prog is None:
+            self.build_hessian()
+        if points is None:
+            points = self.critical_points
+        X = flatten_list(train_x)
+        Y = flatten_list(train_y)
+        results = []
+        for p in points:
+            Hn = self.hessian_at(np.asarray(p).tolist(), X, Y)
+            eig = np.linalg.eigvalsh(Hn)
+            thr = eig_tol * max(1.0, float(np.max(np.abs(eig))))
+            if eig.min() > thr:
+                kind = "local minimum"
+            elif eig.max() < -thr:
+                kind = "local maximum"
+            elif eig.min() < -thr and eig.max() > thr:
+                kind = "saddle point"
+            else:
+                kind = "inconclusive (singular Hessian)"
+            grad_norm = float(np.linalg.norm(self._eval_gradient(np.asarray(p).tolist(), X, Y), np.inf))
+            results.append({"point": np.asarray(p), "type": kind, "eigenvalues": eig,
+                            "grad_norm": grad_norm,
+                            "loss": self.loss_at(np.asarray(p).tolist(), X, Y)})
+        self.critical_point_info = results
+        return results
+    def analyze_critical_points(self, train_x, train_y, **find_kwargs):
+        points = self.find_critical_points(train_x, train_y, **find_kwargs)
+        return self.classify_critical_points(train_x, train_y, points)
+    def train(self, train_x, train_y, check_hessian=False, **find_kwargs):
+        points = self.find_critical_points(train_x, train_y, **find_kwargs)
+        if not points:
+            raise RuntimeError("no critical point found; try a larger n_starts or a different init_scale")
+        if check_hessian:
+            results = [r for r in self.classify_critical_points(train_x, train_y, points)
+                       if r["type"] == "local minimum"]
+        else:
+            X = flatten_list(train_x)
+            Y = flatten_list(train_y)
+            results = [{"point": p, "loss": self.loss_at(p.tolist(), X, Y)} for p in points]
+        results.sort(key=lambda r: r["loss"])
+        print(f"{len(points)} critical points found, {len(results)} kept")
+        if not results:
+            raise RuntimeError("no local minimum passed the Hessian test")
+        self.minima = results
+        self.learn = self._unflatten(results[0]["point"].tolist())
+        return results

{mathai-2.5.2 → mathai-2.5.3}/mathai.egg-info/PKG-INFO

@@ -1,13 +1,15 @@
 Metadata-Version: 2.4
 Name: mathai
-Version: 2.5.2
+Version: 2.5.3
 Summary: Mathematics solving Ai tailored to NCERT
 Home-page: https://github.com/infinity390/mathai
 Requires-Python: >=3.7
 Description-Content-Type: text/markdown
+Requires-Dist: numpy
 Dynamic: description
 Dynamic: description-content-type
 Dynamic: home-page
+Dynamic: requires-dist
 Dynamic: requires-python
 Dynamic: summary
 

{mathai-2.5.2 → mathai-2.5.3}/mathai.egg-info/SOURCES.txt

@@ -22,6 +22,7 @@ mathai/linear.py
 mathai/logic.py
 mathai/matrix.py
 mathai/ml.py
+mathai/mldirect.py
 mathai/mlscalar.py
 mathai/ncert_data.txt
 mathai/ncert_data_optimized.txt
@@ -40,4 +41,5 @@ mathai/univariate_inequality.py
 mathai.egg-info/PKG-INFO
 mathai.egg-info/SOURCES.txt
 mathai.egg-info/dependency_links.txt
+mathai.egg-info/requires.txt
 mathai.egg-info/top_level.txt

mathai-2.5.3/mathai.egg-info/requires.txt

@@ -0,0 +1 @@
+numpy

{mathai-2.5.2 → mathai-2.5.3}/setup.py

@@ -3,12 +3,15 @@ from setuptools import setup, find_packages
 
 setup(
     name="mathai",
-    version="2.5.2",
+    version="2.5.3",
     description="Mathematics solving Ai tailored to NCERT",
     long_description=Path("README.md").read_text(encoding="utf-8"),
     long_description_content_type="text/markdown",
     url="https://github.com/infinity390/mathai",
     packages=find_packages(),
+    install_requires=[
+        "numpy",
+    ],
     include_package_data=True,
     package_data={
         "mathai": ["ncert_data.txt", "ncert_data_optimized.txt"],