Mobilenet v2 width multiplier incorrect

[yaysummeriscoming]

There seems to be a small issue with the width multiplier in mobilenet v2. The official implementation rounds filter channels to a multiple of 8.

For example, mobilenet v2 width 1.4. The first conv layer has 44 channels as opposed to 48 in the official implementation:
model = torchvision.models.mobilenet_v2(width_mult=1.4)

for module in model.modules():
if isinstance(module, nn.Conv2d):
print(module.weight.shape)

torch.Size([44, 3, 3, 3])
torch.Size([44, 1, 3, 3])
torch.Size([22, 44, 1, 1])
torch.Size([132, 22, 1, 1])
torch.Size([132, 1, 3, 3])
torch.Size([33, 132, 1, 1])

Corresponding tensorflow 2.0 keras code:
model = tf.keras.applications.MobileNetV2(
weights="imagenet", input_shape=(224, 224, 3), alpha=1.4)
model.summary()

---

Layer (type) Output Shape Param # Connected to

input_1 (InputLayer) [(None, 224, 224, 3) 0

---

Conv1_pad (ZeroPadding2D) (None, 225, 225, 3) 0 input_1[0][0]

---

Conv1 (Conv2D) (None, 112, 112, 48) 1296 Conv1_pad[0][0]

---

bn_Conv1 (BatchNormalization) (None, 112, 112, 48) 192 Conv1[0][0]

---

Conv1_relu (ReLU) (None, 112, 112, 48) 0 bn_Conv1[0][0]

---

expanded_conv_depthwise (Depthw (None, 112, 112, 48) 432 Conv1_relu[0][0]

---

expanded_conv_depthwise_BN (Bat (None, 112, 112, 48) 192 expanded_conv_depthwise[0][0]

---

expanded_conv_depthwise_relu (R (None, 112, 112, 48) 0 expanded_conv_depthwise_BN[0][0]

---

expanded_conv_project (Conv2D) (None, 112, 112, 24) 1152 expanded_conv_depthwise_relu[0][0

---

expanded_conv_project_BN (Batch (None, 112, 112, 24) 96 expanded_conv_project[0][0]

---

block_1_expand (Conv2D) (None, 112, 112, 144 3456 expanded_conv_project_BN[0][0]

---

block_1_expand_BN (BatchNormali (None, 112, 112, 144 576 block_1_expand[0][0]

---

block_1_expand_relu (ReLU) (None, 112, 112, 144 0 block_1_expand_BN[0][0]

---

block_1_pad (ZeroPadding2D) (None, 113, 113, 144 0 block_1_expand_relu[0][0]

---

block_1_depthwise (DepthwiseCon (None, 56, 56, 144) 1296 block_1_pad[0][0]

---

block_1_depthwise_BN (BatchNorm (None, 56, 56, 144) 576 block_1_depthwise[0][0]

---

block_1_depthwise_relu (ReLU) (None, 56, 56, 144) 0 block_1_depthwise_BN[0][0]

---

block_1_project (Conv2D) (None, 56, 56, 32) 4608 block_1_depthwise_relu[0][0]

---

I've implemented a fix here:
https://github.com/yaysummeriscoming/vision/blob/master/torchvision/models/mobilenet.py

Can I merge it in?

[fmassa]

Thanks for spotting this difference!

Yes, sending a PR would be great!

One question: does it change the behavior of the model for width_multiplier=1?

[yaysummeriscoming]

Ok cool, as the base network widths are already multiples of 8, width multiplier 1 is fine.