From 41c7509dc3d2153da79ed09ecf4a3b592503f15e Mon Sep 17 00:00:00 2001
From: =?utf8?q?Fran=C3=A7ois=20Fleuret?= <francois@fleuret.org>
Date: Fri, 17 Mar 2023 17:22:40 +0100
Subject: [PATCH] Update

---
 beaver.py | 28 ++++++++++++++++++++++++----
 maze.py   | 37 ++++++++++++++++++++++++++++---------
 2 files changed, 52 insertions(+), 13 deletions(-)

diff --git a/beaver.py b/beaver.py
index 54510f0..2cc2140 100755
--- a/beaver.py
+++ b/beaver.py
@@ -178,7 +178,10 @@ def one_shot(gpt, task):
         nn.Linear(args.dim_model, 4),
     ).to(device)
 
+    print(f"{args.nb_epochs=}")
+
     for n_epoch in range(args.nb_epochs):
+        print(f"{n_epoch=}")
         learning_rate = learning_rate_schedule[n_epoch]
         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 
@@ -212,6 +215,23 @@ def one_shot(gpt, task):
             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
         )
 
+        # -------------------
+        input, targets = next(task.policy_batches(split="test"))
+        output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
+        output = model(output_gpt)
+        losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
+        losses = losses / losses.max()
+        print(f"{input.size()=} {losses.size()=} {losses.min()=} {losses.max()=}")
+        losses = losses * (input == 0)
+        losses = losses.reshape(-1, args.maze_height, args.maze_width)
+        input = input.reshape(-1, args.maze_height, args.maze_width)
+        maze.save_image(
+            os.path.join(args.result_dir, f"oneshot_{n_epoch:04d}.png"),
+            mazes=input,
+            score_paths=losses,
+        )
+        # -------------------
+
     gpt.train(t)
 
 
@@ -354,10 +374,10 @@ class TaskMaze(Task):
             _, predicted_paths = self.seq2map(result)
             maze.save_image(
                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
-                mazes,
-                paths,
-                predicted_paths,
-                maze.path_correctness(mazes, predicted_paths),
+                mazes=mazes,
+                target_paths=paths,
+                predicted_paths=predicted_paths,
+                path_correct=maze.path_correctness(mazes, predicted_paths),
             )
 
             model.train(t)
diff --git a/maze.py b/maze.py
index 6c3fe94..6e8e179 100755
--- a/maze.py
+++ b/maze.py
@@ -187,9 +187,14 @@ def create_maze_data(
 ######################################################################
 
 
-def save_image(name, mazes, target_paths, predicted_paths=None, path_correct=None):
-    mazes, target_paths = mazes.cpu(), target_paths.cpu()
-
+def save_image(
+    name,
+    mazes,
+    target_paths=None,
+    predicted_paths=None,
+    score_paths=None,
+    path_correct=None,
+):
     colors = torch.tensor(
         [
             [255, 255, 255],  # empty
@@ -200,13 +205,22 @@ def save_image(name, mazes, target_paths, predicted_paths=None, path_correct=Non
         ]
     )
 
+    mazes = mazes.cpu()
+
     mazes = colors[mazes.reshape(-1)].reshape(mazes.size() + (-1,)).permute(0, 3, 1, 2)
-    target_paths = (
-        colors[target_paths.reshape(-1)]
-        .reshape(target_paths.size() + (-1,))
-        .permute(0, 3, 1, 2)
-    )
-    imgs = torch.cat((mazes.unsqueeze(1), target_paths.unsqueeze(1)), 1)
+
+    imgs = mazes.unsqueeze(1)
+
+    if target_paths is not None:
+        target_paths = target_paths.cpu()
+
+        target_paths = (
+            colors[target_paths.reshape(-1)]
+            .reshape(target_paths.size() + (-1,))
+            .permute(0, 3, 1, 2)
+        )
+
+        imgs = torch.cat((imgs, target_paths.unsqueeze(1)), 1)
 
     if predicted_paths is not None:
         predicted_paths = predicted_paths.cpu()
@@ -217,6 +231,11 @@ def save_image(name, mazes, target_paths, predicted_paths=None, path_correct=Non
         )
         imgs = torch.cat((imgs, predicted_paths.unsqueeze(1)), 1)
 
+    if score_paths is not None:
+        score_paths = (score_paths.cpu() * 255.0).long()
+        score_paths = score_paths.unsqueeze(1).expand(-1, 3, -1, -1)
+        imgs = torch.cat((imgs, score_paths.unsqueeze(1)), 1)
+
     # NxKxCxHxW
     if path_correct is None:
         path_correct = torch.zeros(imgs.size(0)) <= 1
-- 
2.39.5