ultralytics-thop 2.0.15__tar.gz → 2.0.17__tar.gz

{ultralytics_thop-2.0.15 → ultralytics_thop-2.0.17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ultralytics-thop
-Version: 2.0.15
+Version: 2.0.17
 Summary: Ultralytics THOP package for fast computation of PyTorch model FLOPs and parameters.
 Author-email: Ligeng Zhu <ligeng.zhu+github@gmail.com>
 Maintainer-email: Ultralytics <hello@ultralytics.com>
@@ -56,7 +56,7 @@ THOP offers an intuitive API designed to profile PyTorch models by calculating t
 
 Get started with THOP quickly by installing it via pip:
 
-[![PyPI - Version](https://img.shields.io/pypi/v/ultralytics-thop?logo=pypi&logoColor=white)](https://pypi.org/project/ultralytics-thop/) [![Downloads](https://static.pepy.tech/badge/ultralytics-thop)](https://www.pepy.tech/projects/ultralytics-thop) [![PyPI - Python Version](https://img.shields.io/pypi/pyversions/ultralytics-thop?logo=python&logoColor=gold)](https://pypi.org/project/ultralytics-thop/)
+[![PyPI - Version](https://img.shields.io/pypi/v/ultralytics-thop?logo=pypi&logoColor=white)](https://pypi.org/project/ultralytics-thop/) [![Downloads](https://static.pepy.tech/badge/ultralytics-thop)](https://clickpy.clickhouse.com/dashboard/ultralytics-thop) [![PyPI - Python Version](https://img.shields.io/pypi/pyversions/ultralytics-thop?logo=python&logoColor=gold)](https://pypi.org/project/ultralytics-thop/)
 
 ```bash
 pip install ultralytics-thop
@@ -74,7 +74,7 @@ This ensures you have the most recent version, incorporating the latest improvem
 
 ### Basic Usage
 
-Profiling a standard PyTorch model like [ResNet50](https://pytorch.org/vision/main/models/generated/torchvision.models.resnet50.html) is straightforward. Import the necessary libraries, load your model and a sample input tensor, then use the `profile` function:
+Profiling a standard PyTorch model like [ResNet50](https://docs.pytorch.org/vision/main/models/generated/torchvision.models.resnet50.html) is straightforward. Import the necessary libraries, load your model and a sample input tensor, then use the `profile` function:
 
 ```python
 import torch
@@ -230,7 +230,7 @@ The table below showcases the parameters and MACs for several popular [computer
 
 ## 🙌 Contribute
 
-We actively welcome and encourage community contributions to make THOP even better! Whether it's adding support for new [PyTorch layers](https://pytorch.org/docs/stable/nn.html), improving existing calculations, enhancing documentation, or fixing bugs, your input is valuable. Please see our [Contributing Guide](https://docs.ultralytics.com/help/contributing/) for detailed instructions on how to participate. Together, we can ensure THOP remains a state-of-the-art tool for the [machine learning](https://www.ultralytics.com/glossary/machine-learning-ml) community. Don't hesitate to share your feedback and suggestions!
+We actively welcome and encourage community contributions to make THOP even better! Whether it's adding support for new [PyTorch layers](https://docs.pytorch.org/docs/stable/nn.html), improving existing calculations, enhancing documentation, or fixing bugs, your input is valuable. Please see our [Contributing Guide](https://docs.ultralytics.com/help/contributing/) for detailed instructions on how to participate. Together, we can ensure THOP remains a state-of-the-art tool for the [machine learning](https://www.ultralytics.com/glossary/machine-learning-ml) community. Don't hesitate to share your feedback and suggestions!
 
 ## 📜 License
 

{ultralytics_thop-2.0.15 → ultralytics_thop-2.0.17}/README.md

@@ -17,7 +17,7 @@ THOP offers an intuitive API designed to profile PyTorch models by calculating t
 
 Get started with THOP quickly by installing it via pip:
 
-[![PyPI - Version](https://img.shields.io/pypi/v/ultralytics-thop?logo=pypi&logoColor=white)](https://pypi.org/project/ultralytics-thop/) [![Downloads](https://static.pepy.tech/badge/ultralytics-thop)](https://www.pepy.tech/projects/ultralytics-thop) [![PyPI - Python Version](https://img.shields.io/pypi/pyversions/ultralytics-thop?logo=python&logoColor=gold)](https://pypi.org/project/ultralytics-thop/)
+[![PyPI - Version](https://img.shields.io/pypi/v/ultralytics-thop?logo=pypi&logoColor=white)](https://pypi.org/project/ultralytics-thop/) [![Downloads](https://static.pepy.tech/badge/ultralytics-thop)](https://clickpy.clickhouse.com/dashboard/ultralytics-thop) [![PyPI - Python Version](https://img.shields.io/pypi/pyversions/ultralytics-thop?logo=python&logoColor=gold)](https://pypi.org/project/ultralytics-thop/)
 
 ```bash
 pip install ultralytics-thop
@@ -35,7 +35,7 @@ This ensures you have the most recent version, incorporating the latest improvem
 
 ### Basic Usage
 
-Profiling a standard PyTorch model like [ResNet50](https://pytorch.org/vision/main/models/generated/torchvision.models.resnet50.html) is straightforward. Import the necessary libraries, load your model and a sample input tensor, then use the `profile` function:
+Profiling a standard PyTorch model like [ResNet50](https://docs.pytorch.org/vision/main/models/generated/torchvision.models.resnet50.html) is straightforward. Import the necessary libraries, load your model and a sample input tensor, then use the `profile` function:
 
 ```python
 import torch
@@ -191,7 +191,7 @@ The table below showcases the parameters and MACs for several popular [computer
 
 ## 🙌 Contribute
 
-We actively welcome and encourage community contributions to make THOP even better! Whether it's adding support for new [PyTorch layers](https://pytorch.org/docs/stable/nn.html), improving existing calculations, enhancing documentation, or fixing bugs, your input is valuable. Please see our [Contributing Guide](https://docs.ultralytics.com/help/contributing/) for detailed instructions on how to participate. Together, we can ensure THOP remains a state-of-the-art tool for the [machine learning](https://www.ultralytics.com/glossary/machine-learning-ml) community. Don't hesitate to share your feedback and suggestions!
+We actively welcome and encourage community contributions to make THOP even better! Whether it's adding support for new [PyTorch layers](https://docs.pytorch.org/docs/stable/nn.html), improving existing calculations, enhancing documentation, or fixing bugs, your input is valuable. Please see our [Contributing Guide](https://docs.ultralytics.com/help/contributing/) for detailed instructions on how to participate. Together, we can ensure THOP remains a state-of-the-art tool for the [machine learning](https://www.ultralytics.com/glossary/machine-learning-ml) community. Don't hesitate to share your feedback and suggestions!
 
 ## 📜 License
 

ultralytics_thop-2.0.17/tests/test_conv2d.py

@@ -0,0 +1,219 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+import torch
+import torch.nn as nn
+
+from thop import profile
+
+
+class TestUtils:
+    """Utility class for testing Conv2D layers with and without bias, profiling MACs and parameters using THOP."""
+
+    def test_conv2d_no_bias(self):
+        """Tests a 2D Conv layer without bias using THOP profiling on predefined input dimensions and parameters."""
+        n, in_c, ih, iw = 1, 3, 32, 32  # torch.randint(1, 10, (4,)).tolist()
+        out_c, kh, kw = 12, 5, 5
+        s, p, d, g = 1, 1, 1, 1
+
+        net = nn.Conv2d(in_c, out_c, kernel_size=(kh, kw), stride=s, padding=p, dilation=d, groups=g, bias=False)
+        data = torch.randn(n, in_c, ih, iw)
+        out = net(data)
+
+        _, _, oh, ow = out.shape
+        print(f"Conv2d: in={ih}x{iw}, kernel={kh}x{kw}, stride={s}, padding={p}, out={oh}x{ow}")
+        macs, params = profile(net, inputs=(data,))
+        assert macs == 810000, f"{macs} v.s. 810000"
+
+    def test_conv2d(self):
+        """Tests Conv2D layer with bias, profiling MACs and params for specific input dimensions and layer configs."""
+        n, in_c, ih, iw = 1, 3, 32, 32  # torch.randint(1, 10, (4,)).tolist()
+        out_c, kh, kw = 12, 5, 5
+        s, p, d, g = 1, 1, 1, 1
+
+        net = nn.Conv2d(in_c, out_c, kernel_size=(kh, kw), stride=s, padding=p, dilation=d, groups=g, bias=True)
+        data = torch.randn(n, in_c, ih, iw)
+        out = net(data)
+
+        _, _, oh, ow = out.shape
+        print(f"Conv2d: in={ih}x{iw}, kernel={kh}x{kw}, stride={s}, padding={p}, out={oh}x{ow}")
+        macs, params = profile(net, inputs=(data,))
+        assert macs == 810000, f"{macs} v.s. 810000"
+
+    def test_conv2d_random(self):
+        """Test Conv2D layer with random parameters and validate the computed MACs and parameters using 'profile'."""
+        for _ in range(10):
+            out_c, kh, kw = torch.randint(1, 20, (3,)).tolist()
+            n, in_c, ih, iw = torch.randint(1, 20, (4,)).tolist()  # torch.randint(1, 10, (4,)).tolist()
+            ih += kh
+            iw += kw
+            s, p, d, g = 1, 1, 1, 1
+
+            net = nn.Conv2d(in_c, out_c, kernel_size=(kh, kw), stride=s, padding=p, dilation=d, groups=g, bias=False)
+            data = torch.randn(n, in_c, ih, iw)
+            out = net(data)
+
+            _, _, oh, ow = out.shape
+            print(f"Conv2d: in={ih}x{iw}, kernel={kh}x{kw}, stride={s}, padding={p}, out={oh}x{ow}")
+            macs, params = profile(net, inputs=(data,))
+            assert macs == n * out_c * oh * ow // g * in_c * kh * kw, (
+                f"{macs} v.s. {n * out_c * oh * ow // g * in_c * kh * kw}"
+            )
+
+    def test_convtranspose2d_no_bias(self):
+        """Tests a 2D ConvTranspose layer without bias using THOP profiling on predefined input dimensions and
+        parameters.
+        """
+        n, in_c, ih, iw = 1, 3, 2, 2
+        out_c, kh, kw = 1, 2, 2
+        s, p, d, g = 2, 0, 1, 1
+
+        net = nn.ConvTranspose2d(
+            in_c, out_c, kernel_size=(kh, kw), stride=s, padding=p, dilation=d, groups=g, bias=False
+        )
+        data = torch.randn(n, in_c, ih, iw)
+        out = net(data)
+
+        _, _, oh, ow = out.shape
+
+        profile_result = profile(net, inputs=(data,))
+        macs = profile_result[0]
+        profile_result[1]
+        # For ConvTranspose: MACs = input_size * (output_channels / groups) * kernel_size
+        print(f"ConvTranspose2d: in={ih}x{iw}, kernel={kh}x{kw}, stride={s}, padding={p}, out={oh}x{ow}")
+        expected_macs = n * in_c * ih * iw * (out_c // g) * kh * kw
+        assert macs == expected_macs, f"{macs} v.s. {expected_macs}"
+
+    def test_convtranspose2d(self):
+        """Tests ConvTranspose2D layer with bias, profiling MACs and params for specific input dimensions and layer
+        configs.
+        """
+        n, in_c, ih, iw = 1, 3, 2, 2
+        out_c, kh, kw = 1, 2, 2
+        s, p, d, g = 2, 0, 1, 1
+
+        net = nn.ConvTranspose2d(
+            in_c, out_c, kernel_size=(kh, kw), stride=s, padding=p, dilation=d, groups=g, bias=True
+        )
+        data = torch.randn(n, in_c, ih, iw)
+        out = net(data)
+
+        _, _, oh, ow = out.shape
+
+        profile_result = profile(net, inputs=(data,))
+        macs = profile_result[0]
+        profile_result[1]
+        # For ConvTranspose: MACs = input_size * (output_channels / groups) * kernel_size
+        print(f"ConvTranspose2d: in={ih}x{iw}, kernel={kh}x{kw}, stride={s}, padding={p}, out={oh}x{ow}")
+        expected_macs = n * in_c * ih * iw * (out_c // g) * kh * kw
+        assert macs == expected_macs, f"{macs} v.s. {expected_macs}"
+
+    def test_convtranspose2d_groups(self):
+        """Tests ConvTranspose2D layer with groups, validating MAC calculation for grouped transposed convolution."""
+        n, in_c, ih, iw = 1, 8, 4, 4
+        out_c, kh, kw = 4, 3, 3
+        s, p, d, g = 1, 1, 1, 2
+
+        net = nn.ConvTranspose2d(
+            in_c, out_c, kernel_size=(kh, kw), stride=s, padding=p, dilation=d, groups=g, bias=False
+        )
+        data = torch.randn(n, in_c, ih, iw)
+        out = net(data)
+
+        _, _, oh, ow = out.shape
+
+        profile_result = profile(net, inputs=(data,))
+        macs = profile_result[0]
+        profile_result[1]
+        # For ConvTranspose with groups: MACs = input_size * (output_channels / groups) * kernel_size
+        print(f"ConvTranspose2d: in={ih}x{iw}, kernel={kh}x{kw}, stride={s}, padding={p}, out={oh}x{ow}")
+        expected_macs = n * in_c * ih * iw * (out_c // g) * kh * kw
+        assert macs == expected_macs, f"{macs} v.s. {expected_macs}"
+
+    def test_convtranspose2d_random(self):
+        """Test ConvTranspose2D layer with random parameters and validate the computed MACs and parameters using
+        'profile'.
+        """
+        for _ in range(10):
+            # Generate random parameters ensuring valid ConvTranspose configurations
+            out_c, kh, kw = torch.randint(1, 10, (3,)).tolist()
+            n, in_c = torch.randint(1, 5, (2,)).tolist()
+            stride = int(torch.randint(1, 3, (1,)).item())  # stride
+            padding = int(torch.randint(0, 2, (1,)).item())  # padding
+            dilation, groups = 1, 1  # Keep dilation=1 and groups=1 for simplicity
+
+            # Ensure input size is large enough to produce valid output
+            # ConvTranspose output size formula: (input_size - 1) * stride - 2 * padding + kernel_size
+            # To ensure positive output, we need: input_size >= (2 * padding + 1) / stride + 1
+            min_input_size = max(3, (2 * padding + kh) // stride + 1, (2 * padding + kw) // stride + 1)
+            ih, iw = torch.randint(min_input_size, min_input_size + 10, (2,)).tolist()
+
+            net = nn.ConvTranspose2d(
+                in_c,
+                out_c,
+                kernel_size=(kh, kw),
+                stride=stride,
+                padding=padding,
+                dilation=dilation,
+                groups=groups,
+                bias=False,
+            )
+            data = torch.randn(n, in_c, ih, iw)
+            out = net(data)
+
+            _, _, oh, ow = out.shape
+
+            profile_result = profile(net, inputs=(data,))
+            macs = profile_result[0]
+            profile_result[1]
+            # For ConvTranspose: MACs = input_size * (output_channels / groups) * kernel_size
+            expected_macs = n * in_c * ih * iw * (out_c // groups) * kh * kw
+            print(
+                f"ConvTranspose2d: in={ih}x{iw}, kernel={kh}x{kw}, stride={stride}, padding={padding}, out={oh}x{ow}, macs={macs}"
+            )
+            assert macs == expected_macs, f"ConvTranspose2d MACs: {macs} v.s. {expected_macs}"
+
+    def test_conv_vs_convtranspose_symmetry(self):
+        """
+        Test that Conv2d and ConvTranspose2d with symmetric configurations have equal MAC counts.
+
+        Test case: Conv2d downsamples 4x4 -> 2x2, ConvTranspose2d upsamples 2x2 -> 4x4.
+        """
+        # Conv2d: 4x4 -> 2x2
+        conv_net = nn.Conv2d(1, 3, kernel_size=2, stride=2, bias=False)
+        conv_data = torch.randn(1, 1, 4, 4)
+        conv_out = conv_net(conv_data)
+        conv_profile_result = profile(conv_net, inputs=(conv_data,))
+        conv_macs = conv_profile_result[0]
+        conv_params = conv_profile_result[1]
+
+        # ConvTranspose2d: 2x2 -> 4x4 (symmetric operation)
+        convt_net = nn.ConvTranspose2d(3, 1, kernel_size=2, stride=2, bias=False)
+        convt_data = torch.randn(1, 3, 2, 2)
+        convt_out = convt_net(convt_data)
+        convt_profile_result = profile(convt_net, inputs=(convt_data,))
+        convt_macs = convt_profile_result[0]
+        convt_params = convt_profile_result[1]
+
+        # Verify symmetric operations have equal MAC counts
+        assert conv_macs == convt_macs, f"Symmetric operations should have equal MACs: {conv_macs} != {convt_macs}"
+
+        # Manual verification
+        # Conv: output_size * (input_channels / groups) * kernel_size
+        conv_expected = (
+            conv_out.numel()
+            * (conv_net.in_channels // conv_net.groups)
+            * (conv_net.kernel_size[0] * conv_net.kernel_size[1])
+        )
+        # ConvTranspose: input_size * (output_channels / groups) * kernel_size
+        convt_expected = (
+            convt_data.numel()
+            * (convt_net.out_channels // convt_net.groups)
+            * (convt_net.kernel_size[0] * convt_net.kernel_size[1])
+        )
+        print(f"Conv2d: {conv_data.shape} -> {conv_out.shape}, MACs: {conv_macs}, Params: {conv_params}")
+        print(f"ConvTranspose2d: {convt_data.shape} -> {convt_out.shape}, MACs: {convt_macs}, Params: {convt_params}")
+        print(f"Conv2d expected MACs: {conv_expected}, ConvTranspose2d expected MACs: {convt_expected}")
+        assert conv_macs == conv_expected, f"Conv2d MAC calculation incorrect: {conv_macs} != {conv_expected}"
+        assert convt_macs == convt_expected, (
+            f"ConvTranspose2d MAC calculation incorrect: {convt_macs} != {convt_expected}"
+        )

{ultralytics_thop-2.0.15 → ultralytics_thop-2.0.17}/tests/test_matmul.py

@@ -17,7 +17,7 @@ class TestUtils:
         print(flops, params)
         assert flops == n * in_c * out_c
 
-    def test_matmul_case2(self):
+    def test_matmul_case3(self):  # Note renamed to case3 by Glenn Jocher as duplicated above function name
         """Tests matrix multiplication to profile FLOPs and parameters of nn.Linear layer using random dimensions."""
         for _ in range(10):
             n, in_c, out_c = torch.randint(1, 500, (3,)).tolist()

{ultralytics_thop-2.0.15 → ultralytics_thop-2.0.17}/thop/__init__.py

@@ -1,7 +1,6 @@
 # Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
 
-__version__ = "2.0.15"
-
+__version__ = "2.0.17"
 
 import torch
 
@@ -9,3 +8,5 @@ from .profile import profile, profile_origin
 from .utils import clever_format
 
 default_dtype = torch.float64
+
+__all__ = ["profile", "profile_origin", "clever_format", "default_dtype"]

{ultralytics_thop-2.0.15 → ultralytics_thop-2.0.17}/thop/fx_profile.py

@@ -6,6 +6,11 @@ from distutils.version import LooseVersion
 import torch
 import torch as th
 import torch.nn as nn
+from torch.fx import symbolic_trace
+from torch.fx.passes.shape_prop import ShapeProp
+
+from .utils import prRed, prYellow
+from .vision.calc_func import calculate_conv
 
 if LooseVersion(torch.__version__) < LooseVersion("1.8.0"):
     logging.warning(
@@ -40,9 +45,6 @@ def count_fn_linear(input_shapes, output_shapes, *args, **kwargs):
     return flops
 
 
-from .vision.calc_func import calculate_conv
-
-
 def count_fn_conv2d(input_shapes, output_shapes, *args, **kwargs):
     """Calculates total operations (FLOPs) for a 2D conv layer based on input and output shapes using
     `calculate_conv`.
@@ -118,11 +120,6 @@ for k in zero_ops:
 
 missing_maps = {}
 
-from torch.fx import symbolic_trace
-from torch.fx.passes.shape_prop import ShapeProp
-
-from .utils import prRed, prYellow
-
 
 def null_print(*args, **kwargs):
     """A no-op print function that takes any arguments without performing any actions."""
@@ -211,11 +208,15 @@ def fx_profile(mod: nn.Module, input: th.Tensor, verbose=False):
 if __name__ == "__main__":
 
     class MyOP(nn.Module):
+        """Custom operator that performs a simple forward pass dividing input by 1."""
+
         def forward(self, input):
             """Performs forward pass on given input data."""
             return input / 1
 
     class MyModule(torch.nn.Module):
+        """Neural network module with two linear layers and a custom MyOP operator."""
+
         def __init__(self):
             """Initializes MyModule with two linear layers and a custom MyOP operator."""
             super().__init__()

{ultralytics_thop-2.0.15 → ultralytics_thop-2.0.17}/thop/profile.py

@@ -1,7 +1,30 @@
 # Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
 
-from thop.rnn_hooks import *
-from thop.vision.basic_hooks import *
+from thop.rnn_hooks import (
+    count_gru,
+    count_gru_cell,
+    count_lstm,
+    count_lstm_cell,
+    count_rnn,
+    count_rnn_cell,
+    torch,
+)
+from thop.vision.basic_hooks import (
+    count_adap_avgpool,
+    count_avgpool,
+    count_convNd,
+    count_convtNd,
+    count_linear,
+    count_normalization,
+    count_parameters,
+    count_prelu,
+    count_relu,
+    count_softmax,
+    count_upsample,
+    logging,
+    nn,
+    zero_ops,
+)
 
 from .utils import prRed
 
@@ -12,9 +35,9 @@ register_hooks = {
     nn.Conv1d: count_convNd,
     nn.Conv2d: count_convNd,
     nn.Conv3d: count_convNd,
-    nn.ConvTranspose1d: count_convNd,
-    nn.ConvTranspose2d: count_convNd,
-    nn.ConvTranspose3d: count_convNd,
+    nn.ConvTranspose1d: count_convtNd,
+    nn.ConvTranspose2d: count_convtNd,
+    nn.ConvTranspose3d: count_convtNd,
     nn.BatchNorm1d: count_normalization,
     nn.BatchNorm2d: count_normalization,
     nn.BatchNorm3d: count_normalization,

{ultralytics_thop-2.0.15 → ultralytics_thop-2.0.17}/thop/vision/basic_hooks.py

@@ -2,10 +2,24 @@
 
 import logging
 
+import torch
 import torch.nn as nn
 from torch.nn.modules.conv import _ConvNd
 
-from .calc_func import *
+from thop.vision.calc_func import (
+    calculate_adaptive_avg,
+    calculate_avgpool,
+    calculate_conv,
+    calculate_conv2d_flops,
+    calculate_linear,
+    calculate_norm,
+    calculate_parameters,
+    calculate_relu,
+    calculate_relu_flops,
+    calculate_softmax,
+    calculate_upsample,
+    calculate_zero_ops,
+)
 
 multiply_adds = 1
 
@@ -30,6 +44,7 @@ def count_convNd(m: _ConvNd, x, y: torch.Tensor):
         kernel_size=list(m.weight.shape),
         groups=m.groups,
         bias=m.bias,
+        transpose=False,
     )
     # N x Cout x H x W x  (Cin x Kw x Kh + bias)
     # m.total_ops += calculate_conv(
@@ -41,6 +56,20 @@ def count_convNd(m: _ConvNd, x, y: torch.Tensor):
     # )
 
 
+def count_convtNd(m: _ConvNd, x, y: torch.Tensor):
+    """Calculate and add the number of convolutional operations (FLOPs) for a ConvNd layer to the model's total ops."""
+    x = x[0]
+
+    m.total_ops += calculate_conv2d_flops(
+        input_size=list(x.shape),
+        output_size=list(y.shape),
+        kernel_size=list(m.weight.shape),
+        groups=m.groups,
+        bias=m.bias,
+        transpose=True,
+    )
+
+
 def count_convNd_ver2(m: _ConvNd, x, y: torch.Tensor):
     """Calculates and updates total operations (FLOPs) for a convolutional layer in a PyTorch model."""
     x = x[0]

{ultralytics_thop-2.0.15 → ultralytics_thop-2.0.17}/thop/vision/calc_func.py

@@ -29,13 +29,18 @@ def calculate_zero_ops():
     return torch.DoubleTensor([0])
 
 
-def calculate_conv2d_flops(input_size: list, output_size: list, kernel_size: list, groups: int, bias: bool = False):
+def calculate_conv2d_flops(
+    input_size: list, output_size: list, kernel_size: list, groups: int, bias: bool = False, transpose: bool = False
+):
     """Calculate FLOPs for a Conv2D layer using input/output sizes, kernel size, groups, and the bias flag."""
     # n, in_c, ih, iw = input_size
     # out_c, in_c, kh, kw = kernel_size
-    in_c = input_size[1]
-    g = groups
-    return l_prod(output_size) * (in_c // g) * l_prod(kernel_size[2:])
+    if transpose:
+        out_c = output_size[1]
+        return l_prod(input_size) * (out_c // groups) * l_prod(kernel_size[2:])
+    else:
+        in_c = input_size[1]
+        return l_prod(output_size) * (in_c // groups) * l_prod(kernel_size[2:])
 
 
 def calculate_conv(bias, kernel_size, output_size, in_channel, group):

{ultralytics_thop-2.0.15 → ultralytics_thop-2.0.17}/ultralytics_thop.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ultralytics-thop
-Version: 2.0.15
+Version: 2.0.17
 Summary: Ultralytics THOP package for fast computation of PyTorch model FLOPs and parameters.
 Author-email: Ligeng Zhu <ligeng.zhu+github@gmail.com>
 Maintainer-email: Ultralytics <hello@ultralytics.com>
@@ -56,7 +56,7 @@ THOP offers an intuitive API designed to profile PyTorch models by calculating t
 
 Get started with THOP quickly by installing it via pip:
 
-[![PyPI - Version](https://img.shields.io/pypi/v/ultralytics-thop?logo=pypi&logoColor=white)](https://pypi.org/project/ultralytics-thop/) [![Downloads](https://static.pepy.tech/badge/ultralytics-thop)](https://www.pepy.tech/projects/ultralytics-thop) [![PyPI - Python Version](https://img.shields.io/pypi/pyversions/ultralytics-thop?logo=python&logoColor=gold)](https://pypi.org/project/ultralytics-thop/)
+[![PyPI - Version](https://img.shields.io/pypi/v/ultralytics-thop?logo=pypi&logoColor=white)](https://pypi.org/project/ultralytics-thop/) [![Downloads](https://static.pepy.tech/badge/ultralytics-thop)](https://clickpy.clickhouse.com/dashboard/ultralytics-thop) [![PyPI - Python Version](https://img.shields.io/pypi/pyversions/ultralytics-thop?logo=python&logoColor=gold)](https://pypi.org/project/ultralytics-thop/)
 
 ```bash
 pip install ultralytics-thop
@@ -74,7 +74,7 @@ This ensures you have the most recent version, incorporating the latest improvem
 
 ### Basic Usage
 
-Profiling a standard PyTorch model like [ResNet50](https://pytorch.org/vision/main/models/generated/torchvision.models.resnet50.html) is straightforward. Import the necessary libraries, load your model and a sample input tensor, then use the `profile` function:
+Profiling a standard PyTorch model like [ResNet50](https://docs.pytorch.org/vision/main/models/generated/torchvision.models.resnet50.html) is straightforward. Import the necessary libraries, load your model and a sample input tensor, then use the `profile` function:
 
 ```python
 import torch
@@ -230,7 +230,7 @@ The table below showcases the parameters and MACs for several popular [computer
 
 ## 🙌 Contribute
 
-We actively welcome and encourage community contributions to make THOP even better! Whether it's adding support for new [PyTorch layers](https://pytorch.org/docs/stable/nn.html), improving existing calculations, enhancing documentation, or fixing bugs, your input is valuable. Please see our [Contributing Guide](https://docs.ultralytics.com/help/contributing/) for detailed instructions on how to participate. Together, we can ensure THOP remains a state-of-the-art tool for the [machine learning](https://www.ultralytics.com/glossary/machine-learning-ml) community. Don't hesitate to share your feedback and suggestions!
+We actively welcome and encourage community contributions to make THOP even better! Whether it's adding support for new [PyTorch layers](https://docs.pytorch.org/docs/stable/nn.html), improving existing calculations, enhancing documentation, or fixing bugs, your input is valuable. Please see our [Contributing Guide](https://docs.ultralytics.com/help/contributing/) for detailed instructions on how to participate. Together, we can ensure THOP remains a state-of-the-art tool for the [machine learning](https://www.ultralytics.com/glossary/machine-learning-ml) community. Don't hesitate to share your feedback and suggestions!
 
 ## 📜 License
 

ultralytics_thop-2.0.15/tests/test_conv2d.py

@@ -1,61 +0,0 @@
-# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
-
-import torch
-import torch.nn as nn
-
-from thop import profile
-
-
-class TestUtils:
-    """Utility class for testing Conv2D layers with and without bias, profiling FLOPs and parameters using THOP."""
-
-    def test_conv2d_no_bias(self):
-        """Tests a 2D Conv layer without bias using THOP profiling on predefined input dimensions and parameters."""
-        n, in_c, ih, iw = 1, 3, 32, 32  # torch.randint(1, 10, (4,)).tolist()
-        out_c, kh, kw = 12, 5, 5
-        s, p, d, g = 1, 1, 1, 1
-
-        net = nn.Conv2d(in_c, out_c, kernel_size=(kh, kw), stride=s, padding=p, dilation=d, groups=g, bias=False)
-        data = torch.randn(n, in_c, ih, iw)
-        out = net(data)
-
-        _, _, oh, ow = out.shape
-
-        flops, params = profile(net, inputs=(data,))
-        assert flops == 810000, f"{flops} v.s. 810000"
-
-    def test_conv2d(self):
-        """Tests Conv2D layer with bias, profiling FLOPs and params for specific input dimensions and layer configs."""
-        n, in_c, ih, iw = 1, 3, 32, 32  # torch.randint(1, 10, (4,)).tolist()
-        out_c, kh, kw = 12, 5, 5
-        s, p, d, g = 1, 1, 1, 1
-
-        net = nn.Conv2d(in_c, out_c, kernel_size=(kh, kw), stride=s, padding=p, dilation=d, groups=g, bias=True)
-        data = torch.randn(n, in_c, ih, iw)
-        out = net(data)
-
-        _, _, oh, ow = out.shape
-
-        flops, params = profile(net, inputs=(data,))
-        assert flops == 810000, f"{flops} v.s. 810000"
-
-    def test_conv2d_random(self):
-        """Test Conv2D layer with random parameters and validate the computed FLOPs and parameters using 'profile'."""
-        for _ in range(10):
-            out_c, kh, kw = torch.randint(1, 20, (3,)).tolist()
-            n, in_c, ih, iw = torch.randint(1, 20, (4,)).tolist()  # torch.randint(1, 10, (4,)).tolist()
-            ih += kh
-            iw += kw
-            s, p, d, g = 1, 1, 1, 1
-
-            net = nn.Conv2d(in_c, out_c, kernel_size=(kh, kw), stride=s, padding=p, dilation=d, groups=g, bias=False)
-            data = torch.randn(n, in_c, ih, iw)
-            out = net(data)
-
-            _, _, oh, ow = out.shape
-
-            flops, params = profile(net, inputs=(data,))
-            print(flops, params)
-            assert flops == n * out_c * oh * ow // g * in_c * kh * kw, (
-                f"{flops} v.s. {n * out_c * oh * ow // g * in_c * kh * kw}"
-            )