@@ -3336,8 +3336,7 @@ def check_reclayer_optimize_out(subnet_layer_dict, other_subnet_layers=None, sha
3336
3336
rec_layer_dict ["unit" ].update (other_subnet_layers )
3337
3337
config = Config ({
3338
3338
"debug_print_layer_output_template" : True ,
3339
- "num_inputs" : n_in ,
3340
- "num_outputs" : n_out
3339
+ "extern_data" : {"data" : {"dim" : n_in }},
3341
3340
})
3342
3341
from returnn .tf .layers .rec import _SubnetworkRecCell
3343
3342
with make_scope () as session :
@@ -3414,6 +3413,38 @@ def test_reclayer_optimize_out_selfatt_left():
3414
3413
"class" : "self_attention" , "attention_left_only" : True , "num_heads" : 2 , "total_key_dim" : 6 , "n_out" : 18 })
3415
3414
3416
3415
3416
+ def test_reclayer_optimize_out_cum_concat_gen_self_att ():
3417
+ new_dim = DimensionTag (kind = DimensionTag .Types .Spatial , description = "cum_concat_new_dim" )
3418
+ n_key = 5
3419
+ n_value = 7
3420
+ check_reclayer_optimize_out (
3421
+ {"class" : "linear" , "from" : "att" , "activation" : None },
3422
+ {
3423
+ # This is very much the vanilla self attention,
3424
+ # implemented via the new generic way.
3425
+ # See https://github.com/rwth-i6/returnn/issues/391 for a long discussion.
3426
+ # Commented shapes are always for the layers inside the loop (not optimized).
3427
+ "qkv" : {"class" : "linear" , "from" : "data:source" , "activation" : None , "n_out" : n_key * 2 + n_value }, # [B,2*K+V]
3428
+ "qkv_split" : {"class" : "split" , "from" : "qkv" , "size_splits" : [n_key , n_key , n_value ]},
3429
+ "q" : {"class" : "copy" , "from" : "qkv_split/0" }, # [B,K]
3430
+ "k" : {"class" : "copy" , "from" : "qkv_split/1" }, # [B,K]
3431
+ "v" : {"class" : "copy" , "from" : "qkv_split/2" }, # [B,V]
3432
+ # cum_concat here. Note that the optimized-out shape is not as you might expect [T,max(t),B,K],
3433
+ # but instead using the optimized format, with extended dyn size on the special dim tag.
3434
+ "k_accum" : {"class" : "cum_concat" , "new_dim" : new_dim , "from" : "k" }, # [t,B,K]
3435
+ "v_accum" : {"class" : "cum_concat" , "new_dim" : new_dim , "from" : "v" }, # [t,B,V]
3436
+ "energy" : {
3437
+ "class" : "dot" , "from" : ["q" , "k_accum" ],
3438
+ "red1" : "static:-1" , "red2" : "static:-1" ,
3439
+ "var1" : None , "var2" : new_dim }, # [B,t]
3440
+ "att_weights" : {"class" : "softmax_over_spatial" , "from" : "energy" , "axis" : new_dim }, # [B,t]
3441
+ "att" : {
3442
+ "class" : "dot" , "from" : ["att_weights" , "v_accum" ],
3443
+ "red1" : new_dim , "red2" : new_dim ,
3444
+ "var1" : None , "var2" : "static:-1" }, # [B,V]
3445
+ })
3446
+
3447
+
3417
3448
def test_reclayer_optimize_out_dot ():
3418
3449
# Used for multi-head dot-attention.
3419
3450
AttNumHeads = 4
0 commit comments