model summary using torchsummary

[smidm]

After bit of a hassle I found a way how to get a model summary using the torchsummary package.

Input model:

type(model)

mmpose.models.detectors.top_down.TopDown

model.forward = model.forward_dummy  # forward_dummy() takes a person image and outputs a heatmap, it's forward() without pre and post-processing
input_dimensions = (3,) + tuple(model.cfg['data_cfg']['image_size'][::-1])  # default input dimensions
input_dimensions

(3, 384, 288)

import torchsummary
torchsummary.summary(model, input_dimensions)

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1         [-1, 64, 192, 144]           9,408
       BatchNorm2d-2         [-1, 64, 192, 144]             128
              ReLU-3         [-1, 64, 192, 144]               0
...
...
          Conv2d-183           [-1, 23, 96, 72]           5,911
TopdownHeatmapSimpleHead-184           [-1, 23, 96, 72]               0
================================================================
Total params: 34,001,239
Trainable params: 34,001,239
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 1.27
Forward/backward pass size (MB): 688.82
Params size (MB): 129.70
Estimated Total Size (MB): 819.79
----------------------------------------------------------------