Fix calculate_macs() for Linear layers. (#318)

* Fix calculate_macs() for Linear layers.

Fix MACs in lst.out and lstm_half.out.

* Add test for torch.nn.Linear.

* Change groud-truth Total mult-adds in flan_t5_small.out.

 MACs increased from 280.27M to 18.25G because of the Linear layer fix.

---------

Co-authored-by: Andrew Lavin <[email protected]>

Author: andravin

tests/test_output/flan_t5_small.out

@@ -37,7 +37,7 @@ T5ForConditionalGeneration                                   [3, 100, 512]
 Total params: 128,743,488
 Trainable params: 128,743,488
 Non-trainable params: 0
-Total mult-adds (M): 280.27
+Total mult-adds (G): 18.25
 ==============================================================================================================
 Input size (MB): 0.01
 Forward/backward pass size (MB): 326.28

tests/test_output/linear.out

@@ -0,0 +1,15 @@
+========================================================================================================================
+Layer (type:depth-idx)                   Input Shape          Output Shape         Param #              Mult-Adds
+========================================================================================================================
+Linear                                   [32, 16, 8]          [32, 16, 64]         576                  294,912
+========================================================================================================================
+Total params: 576
+Trainable params: 576
+Non-trainable params: 0
+Total mult-adds (M): 0.29
+========================================================================================================================
+Input size (MB): 0.02
+Forward/backward pass size (MB): 0.26
+Params size (MB): 0.00
+Estimated Total Size (MB): 0.28
+========================================================================================================================

tests/test_output/lstm.out

@@ -13,14 +13,14 @@ LSTMNet (LSTMNet)                        --                   [100, 20]
 │    └─weight_hh_l1                      [2048, 512]                               ├─1,048,576
 │    └─bias_ih_l1                        [2048]                                    ├─2,048
 │    └─bias_hh_l1                        [2048]                                    └─2,048
-├─Linear (decoder)                       --                   [1, 100, 20]         10,260               10,260
+├─Linear (decoder)                       --                   [1, 100, 20]         10,260               1,026,000
 │    └─weight                            [512, 20]                                 ├─10,240
 │    └─bias                              [20]                                      └─20
 ========================================================================================================================
 Total params: 3,784,580
 Trainable params: 3,784,580
 Non-trainable params: 0
-Total mult-adds (M): 376.85
+Total mult-adds (M): 377.86
 ========================================================================================================================
 Input size (MB): 0.00
 Forward/backward pass size (MB): 0.67

tests/test_output/lstm_half.out

@@ -4,12 +4,12 @@ Layer (type (var_name))                  Kernel Shape         Output Shape
 LSTMNet (LSTMNet)                        --                   [100, 20]            --                   --
 ├─Embedding (embedding)                  --                   [1, 100, 300]        6,000                6,000
 ├─LSTM (encoder)                         --                   [1, 100, 512]        3,768,320            376,832,000
-├─Linear (decoder)                       --                   [1, 100, 20]         10,260               10,260
+├─Linear (decoder)                       --                   [1, 100, 20]         10,260               1,026,000
 ========================================================================================================================
 Total params: 3,784,580
 Trainable params: 3,784,580
 Non-trainable params: 0
-Total mult-adds (M): 376.85
+Total mult-adds (M): 377.86
 ========================================================================================================================
 Input size (MB): 0.00
 Forward/backward pass size (MB): 0.33

tests/torchinfo_test.py

@@ -149,6 +149,20 @@ def test_groups() -> None:
     )
 
 
+def test_linear() -> None:
+    input_shape = (32, 16, 8)
+    module = nn.Linear(8, 64)
+    col_names = ("input_size", "output_size", "num_params", "mult_adds")
+    input_data = torch.randn(*input_shape)
+    summary(
+        module,
+        input_data=input_data,
+        depth=1,
+        col_names=col_names,
+        col_width=20,
+    )
+
+
 def test_single_input_batch_dim() -> None:
     model = SingleInputNet()
     col_names = ("kernel_size", "input_size", "output_size", "num_params", "mult_adds")

torchinfo/layer_info.py

@@ -244,6 +244,8 @@ def calculate_macs(self) -> None:
                     self.macs += int(
                         cur_params * prod(self.output_size[:1] + self.output_size[2:])
                     )
+                elif "Linear" in self.class_name:
+                    self.macs += int(cur_params * prod(self.output_size[:-1]))
                 else:
                     self.macs += self.output_size[0] * cur_params
             # RNN modules have inner weights such as weight_ih_l0