X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=beaver.py;h=1408f0b2b2b1a0705916b8ebb74676a47e7a4e7b;hb=1e6f089e67087e8cf1bcb6865e8d405b0a50f372;hp=c29dea59e368e642acc788de95535ccb11b7520b;hpb=91ed776af5968643864c4ccce2c8cc7f519a9d65;p=beaver.git

diff --git a/beaver.py b/beaver.py
index c29dea5..1408f0b 100755
--- a/beaver.py
+++ b/beaver.py
@@ -169,28 +169,27 @@ def compute_perplexity(model, split="train"):
 ######################################################################
 
 
-def nb_rank_error(output, targets):
-    output = output.reshape(-1, output.size(-1))
-    targets = targets.reshape(-1, targets.size(-1))
-    i = outputs.argmax(1)
-    # out=input.gather out[i][j]=input[i][index[i][j]]
-    # u[k]=targets[k][i[k]]
-    return output[targets.argmax(1)]
-
-
 def one_shot(gpt, task):
     t = gpt.training
     gpt.eval()
-    model = nn.Linear(args.dim_model, 4).to(device)
+    model = nn.Sequential(
+        nn.Linear(args.dim_model, args.dim_model),
+        nn.ReLU(),
+        nn.Linear(args.dim_model, 4)
+    ).to(device)
 
     for n_epoch in range(args.nb_epochs):
-        optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
+        learning_rate = learning_rate_schedule[n_epoch]
+        optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 
         acc_train_loss, nb_train_samples = 0, 0
         for input, targets in task.policy_batches(split="train"):
             output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
             output = model(output_gpt)
-            loss = -(output.log_softmax(-1) * targets).sum(-1).mean()
+            loss = (
+                -(output.log_softmax(-1) * targets).sum(-1).mean()
+                + targets.xlogy(targets).sum(-1).mean()
+            )
             acc_train_loss += loss.item() * input.size(0)
             nb_train_samples += input.size(0)
 
@@ -202,12 +201,15 @@ def one_shot(gpt, task):
         for input, targets in task.policy_batches(split="test"):
             output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
             output = model(output_gpt)
-            loss = -(output.log_softmax(-1) * targets).sum(-1).mean()
+            loss = (
+                -(output.log_softmax(-1) * targets).sum(-1).mean()
+                + targets.xlogy(targets).sum(-1).mean()
+            )
             acc_test_loss += loss.item() * input.size(0)
             nb_test_samples += input.size(0)
 
-        print(
-            f"{n_epoch=} {acc_train_loss/nb_train_samples=} {acc_test_loss/nb_test_samples=}"
+        log_string(
+            f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
         )
 
     gpt.train(t)