Merge pull request #145 from sovrasov/fix_interpolate

Fix funtional.interpolate hook

Author: sovrasov

CHANGELOG.md

@@ -1,6 +1,7 @@
 # ptflops versions log
 
 ## v 0.7.4
+- Fix hook for nn.functional.interpolate.
 - Switch to aten by default.
 - Add ignore and custom modules for aten.
 - Add an option to disable counting of functional-style operations in pytorch backend.

README.md

@@ -1,6 +1,5 @@
 # Flops counting tool for neural networks in pytorch framework
 [![Pypi version](https://img.shields.io/pypi/v/ptflops.svg)](https://pypi.org/project/ptflops/)
-[![Build Status](https://travis-ci.com/sovrasov/flops-counter.pytorch.svg?branch=master)](https://travis-ci.com/sovrasov/flops-counter.pytorch)
 
 This tool is designed to compute the theoretical amount of multiply-add operations
 in neural networks. It can also compute the number of parameters and

ptflops/pytorch_ops.py

@@ -347,15 +347,19 @@ def _interpolate_functional_flops_hook(*args, **kwargs):
     if input is None and len(args) > 0:
         input = args[0]
 
+    assert input.dim() - 2 > 0, "Input of interpolate should have NC... layout"
+
     size = kwargs.get('size', None)
     if size is None and len(args) > 1:
         size = args[1]
 
     if size is not None:
         if isinstance(size, tuple) or isinstance(size, list):
-            return int(np.prod(size, dtype=np.int64))
+            return int(np.prod(size, dtype=np.int64)) * \
+                np.prod(input.shape[:2], dtype=np.int64)
         else:
-            return int(size)
+            return int(size) ** (input.dim() - 2) * \
+                np.prod(input.shape[:2], dtype=np.int64)
 
     scale_factor = kwargs.get('scale_factor', None)
     if scale_factor is None and len(args) > 2:
@@ -364,10 +368,10 @@ def _interpolate_functional_flops_hook(*args, **kwargs):
     "should be passes to interpolate"
 
     flops = input.numel()
-    if isinstance(scale_factor, tuple) and len(scale_factor) == len(input):
+    if isinstance(scale_factor, tuple) and len(scale_factor) == len(input.shape) - 2:
         flops *= int(np.prod(scale_factor, dtype=np.int64))
-    else:
-        flops *= scale_factor**len(input)
+    else:  # NC... layout is assumed, see interpolate docs
+        flops *= scale_factor ** (input.dim() - 2)
 
     return flops
 

tests/common_test.py

@@ -90,24 +90,26 @@ def input_constructor(input_res):
 
         assert (macs, params) == (8, 8)
 
-    def test_func_interpolate_args(self):
+    @pytest.mark.parametrize("out_size", [(20, 20), 20])
+    def test_func_interpolate_args(self, out_size):
         class CustomModel(nn.Module):
             def __init__(self):
                 super().__init__()
 
             def forward(self, x):
-                return nn.functional.interpolate(input=x, size=(20, 20),
+                return nn.functional.interpolate(input=x, size=out_size,
                                                  mode='bilinear', align_corners=False)
 
         macs, params = \
             get_model_complexity_info(CustomModel(), (3, 10, 10),
                                       as_strings=False,
                                       print_per_layer_stat=False,
                                       backend=FLOPS_BACKEND.PYTORCH)
+
         assert params == 0
-        assert macs > 0
+        assert macs == 1200
 
-        CustomModel.forward = lambda self, x: nn.functional.interpolate(x, size=(20, 20),
+        CustomModel.forward = lambda self, x: nn.functional.interpolate(x, out_size,
                                                                         mode='bilinear')
 
         macs, params = \
@@ -116,7 +118,18 @@ def forward(self, x):
                                       print_per_layer_stat=False,
                                       backend=FLOPS_BACKEND.PYTORCH)
         assert params == 0
-        assert macs > 0
+        assert macs == 1200
+
+        CustomModel.forward = lambda self, x: nn.functional.interpolate(x, scale_factor=2,
+                                                                        mode='bilinear')
+
+        macs, params = \
+            get_model_complexity_info(CustomModel(), (3, 10, 10),
+                                      as_strings=False,
+                                      print_per_layer_stat=False,
+                                      backend=FLOPS_BACKEND.PYTORCH)
+        assert params == 0
+        assert macs == 1200
 
     def test_ten_matmul(self, simple_model_mm):
         macs, params = \