From: François Fleuret <francois@fleuret.org>
Date: Tue, 2 Jan 2024 06:01:48 +0000 (+0100)
Subject: Update.
X-Git-Url: https://www.fleuret.org/cgi-bin/gitweb/gitweb.cgi?p=picoclvr.git;a=commitdiff_plain;h=9d8e9d4fac19329a328dd33b26115792a8090c57

Update.
---

diff --git a/main.py b/main.py
index 1d52b6d..69731ff 100755
--- a/main.py
+++ b/main.py
@@ -5,7 +5,7 @@
 
 # Written by Francois Fleuret <francois@fleuret.org>
 
-import math, sys, argparse, time, tqdm, os
+import math, sys, argparse, time, tqdm, os, datetime
 
 import torch, torchvision
 from torch import nn
@@ -257,9 +257,9 @@ default_task_args = {
         "nb_test_samples": 10000,
     },
     "memory": {
-        "model": "4M",
+        "model": "37M",
         "batch_size": 100,
-        "nb_train_samples": 5000,
+        "nb_train_samples": 25000,
         "nb_test_samples": 1000,
     },
     "mixing": {
@@ -718,6 +718,8 @@ if nb_epochs_finished >= nb_epochs:
         deterministic_synthesis=args.deterministic_synthesis,
     )
 
+time_pred_result = None
+
 for n_epoch in range(nb_epochs_finished, nb_epochs):
     learning_rate = learning_rate_schedule[n_epoch]
 
@@ -776,6 +778,13 @@ for n_epoch in range(nb_epochs_finished, nb_epochs):
             deterministic_synthesis=args.deterministic_synthesis,
         )
 
+        time_current_result = datetime.datetime.now()
+        if time_pred_result is not None:
+            log_string(
+                f"next_result {time_current_result + (time_current_result - time_pred_result)}"
+            )
+        time_pred_result = time_current_result
+
     checkpoint = {
         "nb_epochs_finished": n_epoch + 1,
         "model_state": model.state_dict(),
diff --git a/tasks.py b/tasks.py
index f4be293..a53d213 100755
--- a/tasks.py
+++ b/tasks.py
@@ -14,10 +14,8 @@ from torch.nn import functional as F
 
 from mygpt import BracketedSequence
 
-try:
-    from graph import save_attention_image
-except ImportError:
-    save_attention_image = None
+# from graph import save_attention_image
+save_attention_image = None
 
 ######################################################################
 
@@ -202,9 +200,7 @@ class SandBox(Task):
 
         logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
 
-        if save_attention_image is None:
-            logger("no save_attention_image (is pycairo installed?)")
-        else:
+        if save_attention_image is not None:
             for k in range(10):
                 ns = torch.randint(self.test_input.size(0), (1,)).item()
                 input = self.test_input[ns : ns + 1].clone()
@@ -1567,6 +1563,42 @@ class Grid(Task):
         logger(f"test_performance {n_epoch} {nb_total=} {nb_correct=}")
         logger(f"main_test_accuracy {n_epoch} {nb_correct / nb_total}")
 
+        if n_epoch == 5 or n_epoch == 10 or n_epoch == 20:
+            if save_attention_image is None:
+                logger("no save_attention_image (is pycairo installed?)")
+            else:
+                for k in range(10):
+                    ns = k  # torch.randint(self.test_input.size(0), (1,)).item()
+                    input = self.test_input[ns : ns + 1].clone()
+                    with torch.autograd.no_grad():
+                        t = model.training
+                        model.eval()
+                        model.record_attention(True)
+                        model(BracketedSequence(input))
+                        model.train(t)
+                        ram = model.retrieve_attention()
+                        model.record_attention(False)
+
+                    tokens_output = [self.id2token[t.item()] for t in input[0]]
+                    tokens_input = ["n/a"] + tokens_output[:-1]
+                    for n_head in range(ram[0].size(1)):
+                        filename = os.path.join(
+                            result_dir,
+                            f"sandbox_attention_epoch_{n_epoch}_sample_{k}_head_{n_head}.pdf",
+                        )
+                        attention_matrices = [m[0, n_head] for m in ram]
+                        save_attention_image(
+                            filename,
+                            tokens_input,
+                            tokens_output,
+                            attention_matrices,
+                            k_top=10,
+                            # min_total_attention=0.9,
+                            token_gap=12,
+                            layer_gap=50,
+                        )
+                        logger(f"wrote {filename}")
+
 
 ######################################################################