X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=inline;f=mygpt.py;h=d0fda7e4182878043e74a260f0676654fc12193f;hb=693af34e144cd20d2dde6a508a190d49c1a76c7f;hp=809f79032a55204abd8add007b025ca54b1ad227;hpb=60d829ba77c9769009d3d5a93a50d23c532d019a;p=culture.git

diff --git a/mygpt.py b/mygpt.py
index 809f790..d0fda7e 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -201,6 +201,26 @@ class QKVAttention(nn.Module):
 ##############################
 
 
+class NoiseInjector(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.noise_std = 0.0
+
+    def forward(self, x):
+        if self.noise_std > 0:
+            x = x + torch.randn(x.size(), device=x.device) * self.noise_std
+        return x
+
+
+def set_noise_injection(model, noise_std):
+    for m in model.modules():
+        if isinstance(m, NoiseInjector):
+            m.noise_std = noise_std
+
+
+##############################
+
+
 class MyGPT(nn.Module):
     def __init__(
         self,
@@ -228,7 +248,10 @@ class MyGPT(nn.Module):
         for b in range(nb_blocks):
             trunk_blocks += [
                 WithResidual(
-                    CacheWrapper(nn.LayerNorm((dim_model,))),
+                    CacheWrapper(
+                        nn.LayerNorm((dim_model,)),
+                        NoiseInjector(),
+                    ),
                     QKVAttention(
                         dim_in=dim_model,
                         dim_qk=dim_keys,
@@ -241,6 +264,7 @@ class MyGPT(nn.Module):
                 WithResidual(
                     CacheWrapper(
                         nn.LayerNorm((dim_model,)),
+                        NoiseInjector(),
                         nn.Linear(in_features=dim_model, out_features=dim_hidden),
                         nn.ReLU(),
                         nn.Linear(in_features=dim_hidden, out_features=dim_model),
@@ -271,51 +295,6 @@ class MyGPT(nn.Module):
         bs = self.readout(bs)
         return bs
 
-    # ar_mask is a tensor with 0s and 1s, of same shape as input, with
-    # 1s where tokens should be generated. The others are kept
-    # unchanged.
-
-    def masked_inplace_autoregression(
-        self,
-        input,
-        ar_mask,
-        seq_logproba,
-        temperature=1.0,
-        deterministic_synthesis=False,
-        forbidden_tokens=None,
-        forced_biases=None,
-    ):
-        to_generate = (ar_mask.sum(0) > 0).nonzero()
-
-        if to_generate.min() > 0:
-            self(
-                BracketedSequence(input, 0, to_generate.min())
-            )  # Needed to initialize the model's cache
-        for s in range(to_generate.min(), to_generate.max() + 1):
-            output = self(BracketedSequence(input, s, 1)).x
-
-            logits = output[:, s]
-
-            logits = (logits / temperature).log_softmax(dim=-1)
-
-            if forbidden_tokens is not None:
-                logits = logits.masked_fill(forbidden_tokens, float("-inf"))
-
-            if forced_biases is not None:
-                logits = logits + forced_biases[None, :]
-
-            if deterministic_synthesis:
-                t_next = logits.argmax(-1)
-            else:
-                dist = torch.distributions.categorical.Categorical(logits=logits)
-                t_next = dist.sample()
-
-            if seq_logproba is not None:
-                all_t = torch.arange(t_next.size(0))
-                seq_logproba += logits[all_t, t_next].sum(dim=-1)
-
-            input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
-
     def record_attention(self, v=True):
         for m in self.modules():
             if isinstance(m, QKVAttention):