Skip to content

Add InceptionResnetV1 model. #430

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
zacharydenton wants to merge 2 commits into from
Closed

Add InceptionResnetV1 model. #430

Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d12912a
Add InceptionResnetV1 model.
zacharydenton Dec 18, 2025
eaed6d1
mix format
zacharydenton Dec 18, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
369 changes: 369 additions & 0 deletions lib/bumblebee/vision/inception_resnet_v1.ex
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,369 @@
defmodule Bumblebee.Vision.InceptionResnetV1 do
alias Bumblebee.Shared

options =
[
num_channels: [
default: 3,
doc: "the number of channels in the input"
],
image_size: [
default: 160,
doc: "the size of the input spatial dimensions"
],
dropout_prob: [
default: 0.6,
doc: "the dropout probability"
],
embedding_size: [
default: 512,
doc: "the dimensionality of the output embeddings"
]
] ++ Shared.common_options([:num_labels, :id_to_label])

@moduledoc """
Inception-ResNet-V1 model family.

## Architectures

* `:base` - plain InceptionResnetV1 without any head on top

* `:for_image_classification` - InceptionResnetV1 with a classification head.
The head consists of a single dense layer on top of the pooled
features and it returns logits corresponding to possible classes

## Inputs

* `"pixel_values"` - `{batch_size, height, width, num_channels}`

Featurized image pixel values (160x160).

## Global layer options

#{Shared.global_layer_options_doc([:output_hidden_states])}

## Configuration

#{Shared.options_doc(options)}

## References

* [Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning](https://arxiv.org/abs/1602.07261)
* [FaceNet: A Unified Embedding for Face Recognition and Clustering](https://arxiv.org/abs/1503.03832)

"""

defstruct [architecture: :base] ++ Shared.option_defaults(options)

@behaviour Bumblebee.ModelSpec
@behaviour Bumblebee.Configurable

import Bumblebee.Utils.Model, only: [join: 2]

alias Bumblebee.Layers

@impl true
def architectures(), do: [:base, :for_image_classification]

@impl true
def config(spec, opts) do
spec
|> Shared.put_config_attrs(opts)
|> Shared.validate_label_options()
end

@impl true
def input_template(spec) do
%{
"pixel_values" => Nx.template({1, spec.image_size, spec.image_size, spec.num_channels}, :f32)
}
end

@impl true
def model(%__MODULE__{architecture: :base} = spec) do
spec
|> core()
|> Layers.output()
end

def model(%__MODULE__{architecture: :for_image_classification} = spec) do
outputs = core(spec)

logits =
Axon.dense(outputs.pooled_state, spec.num_labels, name: "image_classification_head.output")

Layers.output(%{
logits: logits,
hidden_states: outputs.hidden_states
})
end

defp core(spec, opts \\ []) do
name = opts[:name]

input = Axon.input("pixel_values", shape: {nil, spec.image_size, spec.image_size, spec.num_channels})

pooled_state =
input
|> stem(spec, name: join(name, "stem"))
|> inception_resnet_blocks(spec, name: join(name, "blocks"))
|> head(spec, name: join(name, "head"))

%{
pooled_state: pooled_state,
hidden_states: Axon.container({input, pooled_state})
Copy link
Contributor Author

@zacharydenton zacharydenton Dec 18, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not sure what to do with hidden_state and hidden_states

Copy link
Member

@jonatanklosko jonatanklosko Dec 22, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

hidden_states is a tuple (effectively a list) of intermediate states between various blocks (see Bumblebee.Vision.ResNet for example). By default we actually prune it and only return if the user configures output_hidden_states: true. We conventionally do it for most models to mirror hf/transformers, but is not a requirement, and it should be fine to not return it.

}
end

# Stem: Initial convolutional layers
defp stem(pixel_values, _spec, opts) do
name = opts[:name]

pixel_values
|> basic_conv2d(32, kernel_size: 3, strides: 2, name: join(name, "conv2d_1a"))
|> basic_conv2d(32, kernel_size: 3, strides: 1, name: join(name, "conv2d_2a"))
|> basic_conv2d(64, kernel_size: 3, strides: 1, padding: [{1, 1}, {1, 1}], name: join(name, "conv2d_2b"))
|> Axon.max_pool(kernel_size: 3, strides: 2, name: join(name, "maxpool_3a"))
|> basic_conv2d(80, kernel_size: 1, strides: 1, name: join(name, "conv2d_3b"))
|> basic_conv2d(192, kernel_size: 3, strides: 1, name: join(name, "conv2d_4a"))
|> basic_conv2d(256, kernel_size: 3, strides: 2, name: join(name, "conv2d_4b"))
end

# Inception-ResNet blocks
defp inception_resnet_blocks(hidden_state, _spec, opts) do
name = opts[:name]

hidden_state
# 5x Block35 (Inception-ResNet-A)
|> then(&Enum.reduce(0..4, &1, fn i, acc ->
block35(acc, 0.17, name: join(name, "repeat_1.#{i}"))
end))
# Mixed 6a (Reduction-A)
|> mixed_6a(name: join(name, "mixed_6a"))
# 10x Block17 (Inception-ResNet-B)
|> then(&Enum.reduce(0..9, &1, fn i, acc ->
block17(acc, 0.10, name: join(name, "repeat_2.#{i}"))
end))
# Mixed 7a (Reduction-B)
|> mixed_7a(name: join(name, "mixed_7a"))
# 5x Block8 (Inception-ResNet-C)
|> then(&Enum.reduce(0..4, &1, fn i, acc ->
block8(acc, 0.20, name: join(name, "repeat_3.#{i}"))
end))
# Final Block8 (scale=1.0, no activation)
|> block8(1.0, activation: :linear, name: join(name, "block8"))
end

# Head: pooling, dropout, and embedding projection
defp head(hidden_state, spec, opts) do
name = opts[:name]

hidden_state
|> Axon.adaptive_avg_pool(output_size: {1, 1}, name: join(name, "avgpool_1a"))
|> Axon.flatten()
|> Axon.dropout(rate: spec.dropout_prob, name: join(name, "dropout"))
|> Axon.dense(spec.embedding_size, use_bias: false, name: join(name, "last_linear"))
|> Axon.batch_norm(
epsilon: 0.001,
momentum: 0.1,
gamma_initializer: :ones,
name: join(name, "last_bn")
)
end

# BasicConv2d: Conv2d + BatchNorm + ReLU
defp basic_conv2d(x, out_channels, opts) do
opts = Keyword.validate!(opts, [:name, kernel_size: 3, strides: 1, padding: :valid])
name = opts[:name]
kernel_size = opts[:kernel_size]
strides = opts[:strides]
padding = opts[:padding]

x
|> Axon.conv(out_channels,
kernel_size: kernel_size,
strides: strides,
padding: padding,
use_bias: false,
name: join(name, "conv")
)
|> Axon.batch_norm(
epsilon: 0.001,
momentum: 0.1,
gamma_initializer: :ones,
name: join(name, "bn")
)
|> Axon.activation(:relu, name: join(name, "relu"))
end

# Block35: Inception-ResNet-A
defp block35(x, scale, opts) do
name = opts[:name]

branch0 = basic_conv2d(x, 32, kernel_size: 1, strides: 1, name: join(name, "branch0"))

branch1 =
x
|> basic_conv2d(32, kernel_size: 1, strides: 1, name: join(name, "branch1.0"))
|> basic_conv2d(32, kernel_size: 3, strides: 1, padding: [{1, 1}, {1, 1}], name: join(name, "branch1.1"))

branch2 =
x
|> basic_conv2d(32, kernel_size: 1, strides: 1, name: join(name, "branch2.0"))
|> basic_conv2d(32, kernel_size: 3, strides: 1, padding: [{1, 1}, {1, 1}], name: join(name, "branch2.1"))
|> basic_conv2d(32, kernel_size: 3, strides: 1, padding: [{1, 1}, {1, 1}], name: join(name, "branch2.2"))

Axon.concatenate([branch0, branch1, branch2], axis: 3)
|> Axon.conv(256, kernel_size: 1, strides: 1, use_bias: true, name: join(name, "conv2d"))
|> Axon.nx(fn conv -> Nx.multiply(conv, scale) end)
|> Axon.add(x)
|> Axon.activation(:relu, name: join(name, "relu"))
end

# Block17: Inception-ResNet-B
defp block17(x, scale, opts) do
name = opts[:name]

branch0 = basic_conv2d(x, 128, kernel_size: 1, strides: 1, name: join(name, "branch0"))

branch1 =
x
|> basic_conv2d(128, kernel_size: 1, strides: 1, name: join(name, "branch1.0"))
|> basic_conv2d(128, kernel_size: {1, 7}, strides: 1, padding: [{0, 0}, {3, 3}], name: join(name, "branch1.1"))
|> basic_conv2d(128, kernel_size: {7, 1}, strides: 1, padding: [{3, 3}, {0, 0}], name: join(name, "branch1.2"))

Axon.concatenate([branch0, branch1], axis: 3)
|> Axon.conv(896, kernel_size: 1, strides: 1, use_bias: true, name: join(name, "conv2d"))
|> Axon.nx(fn conv -> Nx.multiply(conv, scale) end)
|> Axon.add(x)
|> Axon.activation(:relu, name: join(name, "relu"))
end

# Block8: Inception-ResNet-C
defp block8(x, scale, opts) do
opts = Keyword.validate!(opts, [:name, activation: :relu])
name = opts[:name]
activation = opts[:activation]

branch0 = basic_conv2d(x, 192, kernel_size: 1, strides: 1, name: join(name, "branch0"))

branch1 =
x
|> basic_conv2d(192, kernel_size: 1, strides: 1, name: join(name, "branch1.0"))
|> basic_conv2d(192, kernel_size: {1, 3}, strides: 1, padding: [{0, 0}, {1, 1}], name: join(name, "branch1.1"))
|> basic_conv2d(192, kernel_size: {3, 1}, strides: 1, padding: [{1, 1}, {0, 0}], name: join(name, "branch1.2"))

residual =
Axon.concatenate([branch0, branch1], axis: 3)
|> Axon.conv(1792, kernel_size: 1, strides: 1, use_bias: true, name: join(name, "conv2d"))
|> Axon.nx(fn conv -> Nx.multiply(conv, scale) end)
|> Axon.add(x)

if activation == :linear do
residual
else
Axon.activation(residual, activation, name: join(name, "relu"))
end
end

# Mixed_6a: Reduction-A (downsampling transition)
defp mixed_6a(x, opts) do
name = opts[:name]

branch0 = basic_conv2d(x, 384, kernel_size: 3, strides: 2, name: join(name, "branch0"))

branch1 =
x
|> basic_conv2d(192, kernel_size: 1, strides: 1, name: join(name, "branch1.0"))
|> basic_conv2d(192, kernel_size: 3, strides: 1, padding: [{1, 1}, {1, 1}], name: join(name, "branch1.1"))
|> basic_conv2d(256, kernel_size: 3, strides: 2, name: join(name, "branch1.2"))

branch2 = Axon.max_pool(x, kernel_size: 3, strides: 2, name: join(name, "branch2"))

Axon.concatenate([branch0, branch1, branch2], axis: 3)
end

# Mixed_7a: Reduction-B (downsampling transition)
defp mixed_7a(x, opts) do
name = opts[:name]

branch0 =
x
|> basic_conv2d(256, kernel_size: 1, strides: 1, name: join(name, "branch0.0"))
|> basic_conv2d(384, kernel_size: 3, strides: 2, name: join(name, "branch0.1"))

branch1 =
x
|> basic_conv2d(256, kernel_size: 1, strides: 1, name: join(name, "branch1.0"))
|> basic_conv2d(256, kernel_size: 3, strides: 2, name: join(name, "branch1.1"))

branch2 =
x
|> basic_conv2d(256, kernel_size: 1, strides: 1, name: join(name, "branch2.0"))
|> basic_conv2d(256, kernel_size: 3, strides: 1, padding: [{1, 1}, {1, 1}], name: join(name, "branch2.1"))
|> basic_conv2d(256, kernel_size: 3, strides: 2, name: join(name, "branch2.2"))

branch3 = Axon.max_pool(x, kernel_size: 3, strides: 2, name: join(name, "branch3"))

Axon.concatenate([branch0, branch1, branch2, branch3], axis: 3)
end

defimpl Bumblebee.HuggingFace.Transformers.Config do
def load(spec, data) do
import Shared.Converters

opts =
convert!(data,
num_channels: {"num_channels", number()},
image_size: {"image_size", number()},
dropout_prob: {"dropout_prob", number()},
embedding_size: {"embedding_size", number()}
) ++ Shared.common_options_from_transformers(data, spec)

@for.config(spec, opts)
end
end

defimpl Bumblebee.HuggingFace.Transformers.Model do
def params_mapping(_spec) do
%{
# Stem layers
"stem.conv2d_1a.{layer}" => "conv2d_1a.{layer}",
"stem.conv2d_2a.{layer}" => "conv2d_2a.{layer}",
"stem.conv2d_2b.{layer}" => "conv2d_2b.{layer}",
"stem.conv2d_3b.{layer}" => "conv2d_3b.{layer}",
"stem.conv2d_4a.{layer}" => "conv2d_4a.{layer}",
"stem.conv2d_4b.{layer}" => "conv2d_4b.{layer}",
# Block35 (repeat_1) - has branch0, branch1.0, branch1.1, branch2.0, branch2.1, branch2.2
"blocks.repeat_1.{n}.branch0.{layer}" => "repeat_1.{n}.branch0.{layer}",
"blocks.repeat_1.{n}.branch1.{m}.{layer}" => "repeat_1.{n}.branch1.{m}.{layer}",
"blocks.repeat_1.{n}.branch2.{m}.{layer}" => "repeat_1.{n}.branch2.{m}.{layer}",
"blocks.repeat_1.{n}.conv2d" => "repeat_1.{n}.conv2d",
# Mixed 6a - has branch0, branch1.0, branch1.1, branch1.2
"blocks.mixed_6a.branch0.{layer}" => "mixed_6a.branch0.{layer}",
"blocks.mixed_6a.branch1.{m}.{layer}" => "mixed_6a.branch1.{m}.{layer}",
# Block17 (repeat_2) - has branch0, branch1.0, branch1.1, branch1.2
"blocks.repeat_2.{n}.branch0.{layer}" => "repeat_2.{n}.branch0.{layer}",
"blocks.repeat_2.{n}.branch1.{m}.{layer}" => "repeat_2.{n}.branch1.{m}.{layer}",
"blocks.repeat_2.{n}.conv2d" => "repeat_2.{n}.conv2d",
# Mixed 7a - has branch0.0, branch0.1, branch1.0, branch1.1, branch2.0, branch2.1, branch2.2
"blocks.mixed_7a.branch0.{m}.{layer}" => "mixed_7a.branch0.{m}.{layer}",
"blocks.mixed_7a.branch1.{m}.{layer}" => "mixed_7a.branch1.{m}.{layer}",
"blocks.mixed_7a.branch2.{m}.{layer}" => "mixed_7a.branch2.{m}.{layer}",
# Block8 (repeat_3) - has branch0, branch1.0, branch1.1, branch1.2
"blocks.repeat_3.{n}.branch0.{layer}" => "repeat_3.{n}.branch0.{layer}",
"blocks.repeat_3.{n}.branch1.{m}.{layer}" => "repeat_3.{n}.branch1.{m}.{layer}",
"blocks.repeat_3.{n}.conv2d" => "repeat_3.{n}.conv2d",
# Final Block8 - has branch0, branch1.0, branch1.1, branch1.2
"blocks.block8.branch0.{layer}" => "block8.branch0.{layer}",
"blocks.block8.branch1.{m}.{layer}" => "block8.branch1.{m}.{layer}",
"blocks.block8.conv2d" => "block8.conv2d",
# Head
"head.last_linear" => "last_linear",
"head.last_bn" => "last_bn",
# Classification head
"image_classification_head.output" => "logits"
}
end
end
end
Loading