X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=d0fda7e4182878043e74a260f0676654fc12193f;hb=09c5eea203d5a2d8b1da84db0a336de151cf1c89;hp=8cd015297b3f29b037eea0c48d4449c2204b8180;hpb=f680fa1486b0a70c37f0951cedd7b5c56b5808bb;p=culture.git

diff --git a/mygpt.py b/mygpt.py
index 8cd0152..d0fda7e 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -45,6 +45,9 @@ class BracketedSequence:
     def slice(self):
         return self.x[:, self.first : self.first + self.nb]
 
+    def complete(self):
+        return self.first == 0 and self.nb == self.x.size(1)
+
 
 ######################################################################
 
@@ -113,16 +116,22 @@ class AddPositionalEncoding(nn.Module):
 
 class QKVAttention(nn.Module):
     def __init__(
-        self, dim_in, dim_qk, dim_v, nb_heads=1, causal=False, attention_dropout=0.0
+        self,
+        dim_in,
+        dim_qk,
+        dim_v,
+        nb_heads=1,
+        causal=False,
+        attention_dropout=0.0,
     ):
         super().__init__()
 
         def randw(*d):
             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 
-        assert causal, "TODO: Switch off the cache when non-causal!!!"
         self.causal = causal
         self.attention_dropout = attention_dropout
+        self.record_attention = False
 
         self.w_q = randw(nb_heads, dim_qk, dim_in)
         self.w_k = randw(nb_heads, dim_qk, dim_in)
@@ -132,6 +141,10 @@ class QKVAttention(nn.Module):
     def forward(self, bs_q):
         x_q = bs_q.x
 
+        assert (
+            self.causal or bs_q.complete()
+        ), "Partial evaluation is only possible for causal models"
+
         if bs_q.first == 0:
             self.cache_k = x_q.new_zeros(
                 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
@@ -170,6 +183,10 @@ class QKVAttention(nn.Module):
             )
 
         a = a.softmax(dim=3)
+
+        if self.record_attention:
+            self.a = a
+
         a = F.dropout(a, self.attention_dropout, self.training)
 
         y = torch.einsum(
@@ -184,6 +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,
@@ -211,7 +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,
@@ -224,6 +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),
@@ -247,35 +288,24 @@ class MyGPT(nn.Module):
                     m.weight.fill_(1.0)
 
     def forward(self, bs):
+        # print(f"GENERATE {bs.first} {bs.first+bs.nb}")
         bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb)
         bs = self.embedding(bs)
         bs = self.trunk(bs)
         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 record_attention(self, v=True):
+        for m in self.modules():
+            if isinstance(m, QKVAttention):
+                m.record_attention = v
 
-    def masked_inplace_autoregression(
-        self, input, ar_mask, forbidden_tokens=None, deterministic_synthesis=False
-    ):
-        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]
-            if forbidden_tokens is not None:
-                logits = logits.masked_fill(forbidden_tokens, float("-inf"))
-            if deterministic_synthesis:
-                t_next = logits.argmax(1)
-            else:
-                dist = torch.distributions.categorical.Categorical(logits=logits)
-                t_next = dist.sample()
-            input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
+    def retrieve_attention(self):
+        a = []
+        for m in self.modules():
+            if isinstance(m, QKVAttention):
+                a.append(m.a)
+        return a
 
 
 ######################################################################
@@ -292,13 +322,12 @@ if __name__ == "__main__":
         dim_keys=2,
         dim_hidden=2,
         nb_heads=2,
-        nb_blocks=1,
+        nb_blocks=2,
         dropout=0.1,
         causal=True,
     )
 
     model.eval()
-
     y1 = model(BracketedSequence(x)).x
     y2 = torch.randn_like(y1)
     for s in range(x.size(1)):