InceptionV4Config
ModelConfigInceptionV4Config(num_classes: int = 1000, in_channels: int = 3, dropout: float = 0.2)Frozen configuration for Inception-v4.
Inception-v4 is a single-size architecture: the paper (Figure 9) defines exactly one network, so this configuration carries only the input / output widths and the head dropout. Every channel count and repeat count inside the body is fixed by the paper.
Parameters
num_classesint= 1000in_channelsint= 3dropoutfloat= 0.2Attributes
model_typestr"inception_v4".Notes
This is the TF-Slim Inception-v4 — the network the released checkpoint was trained as — which follows the paper's figures in every channel width and repeat count. One detail is dictated by the released tensors rather than by the drawings: where a figure lists an asymmetric factorised pair ( then , or then ) in one order and TF-Slim builds the other, the TF-Slim order is used, because the checkpoint's weight shapes fix it.
Layout as built (299×299 input):
- Stem (Figure 3): 299×299×3 → 35×35×384
- 4× Inception-A: 35×35×384
- Reduction-A: 35×35×384 → 17×17×1024 (k, l, m, n = 192, 224, 256, 384)
- 7× Inception-B: 17×17×1024
- Reduction-B: 17×17×1024 → 8×8×1536
- 3× Inception-C: 8×8×1536
- AdaptiveAvgPool → Dropout(p=0.2) → Linear(1536, num_classes)
Every convolution is followed by BatchNorm2d(eps=1e-3) (the
TF-Slim default) and a ReLU.
Examples
>>> from lucid.models.vision.inception_v4 import InceptionV4Config
>>> cfg = InceptionV4Config()
>>> cfg.model_type
'inception_v4'
>>> cfg.num_classes, cfg.in_channels, cfg.dropout
(1000, 3, 0.2)