Update LFNA

exps/LFNA/lfna.py

@@ -107,11 +107,20 @@ def online_evaluate(env, meta_model, base_model, criterion, args, logger):
             base_model.eval()
             time_seqs = torch.Tensor(time_seqs).view(1, -1).to(args.device)
             [seq_containers], _ = meta_model(time_seqs, None)
-            future_container = seq_containers[-2]
-            _, (future_x, future_y) = env(time_seqs[0, -2].item())
-            future_x, future_y = future_x.to(args.device), future_y.to(args.device)
-            future_y_hat = base_model.forward_with_container(future_x, future_container)
-            future_loss = criterion(future_y_hat, future_y)
+            # For Debug
+            for idx in range(time_seqs.numel()):
+                future_container = seq_containers[idx]
+                _, (future_x, future_y) = env(time_seqs[0, idx].item())
+                future_x, future_y = future_x.to(args.device), future_y.to(args.device)
+                future_y_hat = base_model.forward_with_container(
+                    future_x, future_container
+                )
+                future_loss = criterion(future_y_hat, future_y)
+                logger.log(
+                    "--> time={:.4f} -> loss={:.4f}".format(
+                        time_seqs[0, idx].item(), future_loss.item()
+                    )
+                )
             logger.log(
                 "[ONLINE] [{:03d}/{:03d}] loss={:.4f}".format(
                     idx, len(env), future_loss.item()

exps/LFNA/lfna_meta_model.py

@@ -47,17 +47,17 @@ class LFNA_Meta(super_core.SuperModule):
         self._append_meta_timestamps = dict(fixed=None, learnt=None)
 
         self._tscalar_embed = super_core.SuperDynamicPositionE(
-            time_embedding, scale=100
+            time_embedding, scale=500
         )
 
         # build transformer
-        self._trans_att = super_core.SuperQKVAttention(
-            time_embedding,
-            time_embedding,
-            time_embedding,
-            time_embedding,
-            4,
-            True,
+        self._trans_att = super_core.SuperQKVAttentionV2(
+            qk_att_dim=time_embedding,
+            in_v_dim=time_embedding,
+            hidden_dim=time_embedding,
+            num_heads=4,
+            proj_dim=time_embedding,
+            qkv_bias=True,
             attn_drop=None,
             proj_drop=dropout,
         )
@@ -166,9 +166,12 @@ class LFNA_Meta(super_core.SuperModule):
         # timestamps is a batch of sequence of timestamps
         batch, seq = timestamps.shape
         meta_timestamps, meta_embeds = self.meta_timestamps, self.super_meta_embed
-        timestamp_q_embed = self._tscalar_embed(timestamps)
-        timestamp_k_embed = self._tscalar_embed(meta_timestamps.view(1, -1))
+        # timestamp_q_embed = self._tscalar_embed(timestamps)
+        # timestamp_k_embed = self._tscalar_embed(meta_timestamps.view(1, -1))
         timestamp_v_embed = meta_embeds.unsqueeze(dim=0)
+        timestamp_qk_att_embed = self._tscalar_embed(
+            torch.unsqueeze(timestamps, dim=-1) - meta_timestamps
+        )
         # create the mask
         mask = (
             torch.unsqueeze(timestamps, dim=-1) <= meta_timestamps.view(1, 1, -1)
@@ -179,11 +182,13 @@ class LFNA_Meta(super_core.SuperModule):
             > self._thresh
         )
         timestamp_embeds = self._trans_att(
-            timestamp_q_embed, timestamp_k_embed, timestamp_v_embed, mask
+            timestamp_qk_att_embed, timestamp_v_embed, mask
+        )
+        relative_timestamps = timestamps - timestamps[:, :1]
+        relative_pos_embeds = self._tscalar_embed(relative_timestamps)
+        init_timestamp_embeds = torch.cat(
+            (timestamp_embeds, relative_pos_embeds), dim=-1
         )
-        # relative_timestamps = timestamps - timestamps[:, :1]
-        # relative_pos_embeds = self._tscalar_embed(relative_timestamps)
-        init_timestamp_embeds = torch.cat((timestamp_q_embed, timestamp_embeds), dim=-1)
         corrected_embeds = self._meta_corrector(init_timestamp_embeds)
         return corrected_embeds
 

xautodl/xlayers/super_attention.py

@@ -238,7 +238,7 @@ class SuperQKVAttention(SuperModule):
         return root_node
 
     def apply_candidate(self, abstract_child: spaces.VirtualNode):
-        super(SuperQVKAttention, self).apply_candidate(abstract_child)
+        super(SuperQKVAttention, self).apply_candidate(abstract_child)
         if "q_fc" in abstract_child:
             self.q_fc.apply_candidate(abstract_child["q_fc"])
         if "k_fc" in abstract_child:

xautodl/xlayers/super_attention_v2.py

@@ -0,0 +1,117 @@
+#####################################################
+# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2021.03 #
+#####################################################
+from __future__ import division
+from __future__ import print_function
+
+import math
+from functools import partial
+from typing import Optional, Text
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+from xautodl import spaces
+from .super_module import SuperModule
+from .super_module import IntSpaceType
+from .super_module import BoolSpaceType
+from .super_linear import SuperLinear
+
+
+class SuperQKVAttentionV2(SuperModule):
+    """The super model for attention layer."""
+
+    def __init__(
+        self,
+        qk_att_dim: int,
+        in_v_dim: int,
+        hidden_dim: int,
+        num_heads: int,
+        proj_dim: int,
+        qkv_bias: bool = False,
+        attn_drop: Optional[float] = None,
+        proj_drop: Optional[float] = None,
+    ):
+        super(SuperQKVAttentionV2, self).__init__()
+        self._in_v_dim = in_v_dim
+        self._qk_att_dim = qk_att_dim
+        self._proj_dim = proj_dim
+        self._hidden_dim = hidden_dim
+        self._num_heads = num_heads
+        self._qkv_bias = qkv_bias
+
+        self.qk_fc = SuperLinear(qk_att_dim, num_heads, bias=qkv_bias)
+        self.v_fc = SuperLinear(in_v_dim, hidden_dim * num_heads, bias=qkv_bias)
+
+        self.attn_drop = nn.Dropout(attn_drop or 0.0)
+        self.proj = SuperLinear(hidden_dim * num_heads, proj_dim)
+        self.proj_drop = nn.Dropout(proj_drop or 0.0)
+        self._infinity = 1e9
+
+    @property
+    def num_heads(self):
+        return spaces.get_max(self._num_heads)
+
+    @property
+    def in_v_dim(self):
+        return spaces.get_max(self._in_v_dim)
+
+    @property
+    def qk_att_dim(self):
+        return spaces.get_max(self._qk_att_dim)
+
+    @property
+    def hidden_dim(self):
+        return spaces.get_max(self._hidden_dim)
+
+    @property
+    def proj_dim(self):
+        return spaces.get_max(self._proj_dim)
+
+    @property
+    def abstract_search_space(self):
+        root_node = spaces.VirtualNode(id(self))
+        raise NotImplementedError
+
+    def apply_candidate(self, abstract_child: spaces.VirtualNode):
+        super(SuperQKVAttentionV2, self).apply_candidate(abstract_child)
+        raise NotImplementedError
+
+    def forward_qkv(
+        self, qk_att_tensor, v_tensor, num_head: int, mask=None
+    ) -> torch.Tensor:
+        qk_att = self.qk_fc(qk_att_tensor)
+        B, N, S, _ = qk_att.shape
+        assert _ == num_head
+        attn_v1 = qk_att.permute(0, 3, 1, 2)
+        if mask is not None:
+            mask = torch.unsqueeze(mask, dim=1)
+            attn_v1 = attn_v1.masked_fill(mask, -self._infinity)
+        attn_v1 = attn_v1.softmax(dim=-1)  # B * #head * N * S
+        attn_v1 = self.attn_drop(attn_v1)
+
+        v = self.v_fc(v_tensor)
+        B0, _, _ = v.shape
+        v_v1 = v.reshape(B0, S, num_head, -1).permute(0, 2, 1, 3)
+        feats_v1 = (attn_v1 @ v_v1).permute(0, 2, 1, 3).reshape(B, N, -1)
+        return feats_v1
+
+    def forward_candidate(self, qk_att_tensor, v_tensor, mask=None) -> torch.Tensor:
+        return self.forward_raw(qk_att_tensor, v_tensor, mask)
+
+    def forward_raw(self, qk_att_tensor, v_tensor, mask=None) -> torch.Tensor:
+        feats = self.forward_qkv(qk_att_tensor, v_tensor, self.num_heads, mask)
+        outs = self.proj(feats)
+        outs = self.proj_drop(outs)
+        return outs
+
+    def extra_repr(self) -> str:
+        return "input_dim={:}, hidden_dim={:}, proj_dim={:}, num_heads={:}, infinity={:}".format(
+            (self.qk_att_dim, self.in_v_dim),
+            self._hidden_dim,
+            self._proj_dim,
+            self._num_heads,
+            self._infinity,
+        )

xautodl/xlayers/super_core.py

@@ -26,6 +26,7 @@ super_name2norm = {
 
 from .super_attention import SuperSelfAttention
 from .super_attention import SuperQKVAttention
+from .super_attention_v2 import SuperQKVAttentionV2
 from .super_transformer import SuperTransformerEncoderLayer
 
 from .super_activations import SuperReLU