unfinished work on backtracking

app.py

@@ -1,5 +1,5 @@
 import streamlit as st
-from sudoku.train import SudokuTrialErrorLightning
+from sudoku.train import SudokuTrialErrorLightning, TrialEveryPosException
 from sudoku.helper import display_as_dataframe, get_grid_number_soluce
 import numpy as np
 import re
@@ -116,7 +116,14 @@ if n_sol==1:
     while new_X.sum()<729:
         i+=1
         st.markdown(f'iteration {i}')
-        new_X = model.predict(new_X)
+        try:
+            new_X = model.predict(new_X)
+        except TrialEveryPosException:
+            st.markdown('''## The grid is super evil!
+            please share it as A Discussion in the `Community` tab.
+            Except if it is this one: https://www.telegraph.co.uk/news/science/science-news/9359579/Worlds-hardest-sudoku-can-you-crack-it.html''')
+            is_valid, new_X = model.backtracking_predict(new_X)
+            assert is_valid
         st.html(display_as_dataframe(new_X).to_html(escape=False, index=False))
         new_X_sum = new_X.sum()
         assert new_X_sum> X_sum

sudoku/train.py

@@ -401,6 +401,9 @@ class SudokuLightning(pl.LightningModule):
 
 # TODO adapt training to something softer
 # 
+class TrialEveryPosException(Exception):
+    pass
+
 class SudokuTrialErrorLightning(SudokuLightning):
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
@@ -752,7 +755,7 @@ class SudokuTrialErrorLightning(SudokuLightning):
             mask_possibility[pos]=False
         if  mask_possibility.sum()==0:
             print('mask_possible=0')
-            raise ValueError()
+            raise TrialEveryPosException()
             
         with torch.no_grad():
             x_reg = self.sym_preprocess.forward(s_new_X.view(1,2,-1))
@@ -861,7 +864,36 @@ class SudokuTrialErrorLightning(SudokuLightning):
                         return x
                     X_tried = new_X
                     # if one of X_tried is complete (weird but possible) -> return x with tried_position mask_complet set to 1 (cause we still want a step by step resolution)
+    
+    def backtracking_predict(self, x):
+        """
+        return is_valid, new_x
+        """
+        try:
+            next_x = self.predict(x)
+        except TrialEveryPosException:
+            pos = self.search_trial(x.view(2,729), [])
+            new_x = deepcopy(x.view(2,729))
+            output = self.forward_layer(x.view(1,2,729))
+            pos_output = torch.argmax(output[0,:,pos])
+            if (pos_output[0].item()-self.threshold_abs)>(pos_output[1].item-self.threshold_pos):
+                trial_abs_pres = 0
+            else:
+                trial_abs_pres = 1
+            new_x[0,trial_abs_pres,pos]=1
+            is_valid, new_x = self.backtracking_predict(new_x)
+            if not is_valid:
+                new_x = deepcopy(x.view(2,729))
+                new_x[0,(trial_abs_pres+1)%2,pos]=1
+                return self.backtracking_predict(new_x)
+            return True, new_x
+        except ValueError:
+            return False, None
+        if next_x.sum()==729:
+            return True, next_x
+        return self.backtracking_predict(next_x)
             
+    
     def on_validation_epoch_start(self) -> None:
         # self.buffer = BufferArray(self.nets_number, self.batch_size)
         self.trial_error_buffer = Buffer(self.batch_size)