pulsatrix 1.1.0
Loading...
Searching...
No Matches
vision_models.hpp
Go to the documentation of this file.
1
6#pragma once
7
8#include <cstdint>
9#include <memory>
10#include <optional>
11#include <string>
12#include <vector>
13
22#include "pulsatrix/module.hpp"
26
27namespace pulsatrix {
28
33 std::vector<int64_t> blocks = {2, 2, 2, 2};
35 int64_t width = 64;
36 int64_t num_classes = 1000;
37 int64_t in_channels = 3;
38};
39
55class TorchvisionResNet : public Module {
56public:
59
61 [[nodiscard]] static ResNetConfig ResNet18() { return {}; }
62
65 [[nodiscard]] std::vector<Module*> layers();
68 [[nodiscard]] std::vector<std::unique_ptr<BatchNormFold>> fold_batch_norms();
69
70 [[nodiscard]] Tensor backward(const Tensor& grad_output) override;
71 [[nodiscard]] Tensor propagate_relevance(const Tensor& relevance_out, const LRPRuleConfig& config) override;
72 [[nodiscard]] OpType op_type() const override { return OpType::Composite; }
73 [[nodiscard]] std::vector<NamedParamRef> named_parameters() override;
74 [[nodiscard]] std::vector<NamedBufferRef> named_buffers() override;
75 [[nodiscard]] std::optional<DeviceType> compute_device() const override { return backend_->device(); }
76 void set_training(bool training) override;
77 [[nodiscard]] const ResNetConfig& config() const { return config_; }
78
79protected:
80 [[nodiscard]] Tensor forward_impl(const Tensor& input) override;
81
82private:
83 struct Block {
84 std::string name;
85 std::unique_ptr<Conv2DModule> conv1, conv2, down_conv;
86 std::unique_ptr<BatchNormModule> bn1, bn2, down_bn;
87 std::unique_ptr<ReluModule> relu1, relu_out;
88 std::unique_ptr<SequentialModule> main, shortcut;
89 std::unique_ptr<ResidualModule> residual;
90 std::unique_ptr<SequentialModule> block;
91 };
92 ResNetConfig config_;
93 DeviceBackend* backend_;
94 std::unique_ptr<Conv2DModule> conv1_;
95 std::unique_ptr<BatchNormModule> bn1_;
96 std::unique_ptr<ReluModule> relu_;
97 std::unique_ptr<MaxPool2DModule> maxpool_;
98 std::vector<std::unique_ptr<Block>> blocks_;
99 std::unique_ptr<AdaptiveAvgPool2DModule> avgpool_;
100 std::unique_ptr<FlattenModule> flatten_;
101 std::unique_ptr<LinearModule> fc_;
102 std::unique_ptr<SequentialModule> net_;
103};
104
107struct VGGConfig {
110 std::vector<int64_t> features = {64, 64, 0, 128, 128, 0, 256, 256, 256, 0, 512, 512, 512, 0, 512, 512, 512, 0};
112 int64_t pool_size = 7;
114 int64_t hidden = 4096;
115 int64_t num_classes = 1000;
116 int64_t in_channels = 3;
117};
118
126class TorchvisionVGG : public Module {
127public:
130
132 [[nodiscard]] static VGGConfig VGG16() { return {}; }
133
134 [[nodiscard]] std::vector<Module*> layers();
135
136 [[nodiscard]] Tensor backward(const Tensor& grad_output) override;
137 [[nodiscard]] Tensor propagate_relevance(const Tensor& relevance_out, const LRPRuleConfig& config) override;
138 [[nodiscard]] OpType op_type() const override { return OpType::Composite; }
139 [[nodiscard]] std::vector<NamedParamRef> named_parameters() override;
140 [[nodiscard]] std::optional<DeviceType> compute_device() const override { return backend_->device(); }
141 void set_training(bool training) override;
142 [[nodiscard]] const VGGConfig& config() const { return config_; }
143
144protected:
145 [[nodiscard]] Tensor forward_impl(const Tensor& input) override;
146
147private:
148 VGGConfig config_;
149 DeviceBackend* backend_;
150 std::vector<std::unique_ptr<Module>> owned_;
151 std::vector<Module*> layers_;
152 std::vector<std::pair<std::string, Module*>> named_;
153 std::unique_ptr<SequentialModule> net_;
154};
155
158 int64_t loaded = 0;
161 std::vector<std::string> left_at_default;
162};
163
172TorchvisionLoadReport LoadTorchvisionWeights(Module& model, const std::string& safetensors_path);
173
174} // namespace pulsatrix
Average pooling to a fixed output size, as PyTorch's AdaptiveAvgPool2d (KS-9).
Folds an eval-mode BatchNorm into the Conv2D before it for the duration of an LRP explanation – Zenni...
Batch normalization (Ioffe & Szegedy, 2015) – per-channel statistics computed across the batch and sp...
Vendor-agnostic compute/memory backend. CPUBackend, CUDABackend (Phase 1.5), and HIPBackend (Phase 1....
Definition device_backend.hpp:248
virtual DeviceType device() const noexcept=0
Which device this backend's buffers reside on.
Base class for every layer type (LinearModule, Conv2DModule, activations, ...).
Definition module.hpp:78
N-dimensional tensor. Owns its data buffer exclusively; a DeviceBackend* is injected (not owned) – th...
Definition tensor.hpp:29
torchvision's ResNet(BasicBlock, ...): a 7x7 stride-2 convolution, BatchNorm, ReLU and 3x3 stride-2 m...
Definition vision_models.hpp:55
std::vector< std::unique_ptr< BatchNormFold > > fold_batch_norms()
Folds every BatchNorm into the convolution before it, until the returned objects are destroyed.
OpType op_type() const override
This module's operation-category tag, for ComputationGraph node tagging.
Definition vision_models.hpp:72
std::optional< DeviceType > compute_device() const override
The device this module computes on, so forward() can reject an input on another device before any ker...
Definition vision_models.hpp:75
Tensor backward(const Tensor &grad_output) override
Computes the gradient w.r.t. this module's input, given the gradient w.r.t. its output....
Tensor propagate_relevance(const Tensor &relevance_out, const LRPRuleConfig &config) override
Computes this module's contribution to LRP relevance propagation.
void set_training(bool training) override
Sets this module's training/eval mode. Defaults to training (matches every mainstream framework's Mod...
std::vector< NamedParamRef > named_parameters() override
This module's trainable parameters, each with its hierarchical name – the one place a module declares...
const ResNetConfig & config() const
Definition vision_models.hpp:77
TorchvisionResNet(const ResNetConfig &config, DeviceBackend *backend)
static ResNetConfig ResNet18()
ResNet18's config (11.7M parameters).
Definition vision_models.hpp:61
std::vector< Module * > layers()
The top-level layers, in order: conv1, bn1, relu, maxpool, one per block, avgpool,...
Tensor forward_impl(const Tensor &input) override
The actual forward computation. Called by forward() after precondition checks.
std::vector< NamedBufferRef > named_buffers() override
This module's buffers: state a checkpoint must save that no optimizer updates.
torchvision's VGG: 3x3 convolutions with bias and ReLU, 2x2 max pooling, adaptive average pooling to ...
Definition vision_models.hpp:126
TorchvisionVGG(const VGGConfig &config, DeviceBackend *backend)
std::vector< Module * > layers()
Tensor propagate_relevance(const Tensor &relevance_out, const LRPRuleConfig &config) override
Computes this module's contribution to LRP relevance propagation.
Tensor forward_impl(const Tensor &input) override
The actual forward computation. Called by forward() after precondition checks.
static VGGConfig VGG16()
VGG16's config (138M parameters).
Definition vision_models.hpp:132
const VGGConfig & config() const
Definition vision_models.hpp:142
std::vector< NamedParamRef > named_parameters() override
This module's trainable parameters, each with its hierarchical name – the one place a module declares...
Tensor backward(const Tensor &grad_output) override
Computes the gradient w.r.t. this module's input, given the gradient w.r.t. its output....
std::optional< DeviceType > compute_device() const override
The device this module computes on, so forward() can reject an input on another device before any ker...
Definition vision_models.hpp:140
OpType op_type() const override
This module's operation-category tag, for ComputationGraph node tagging.
Definition vision_models.hpp:138
void set_training(bool training) override
Sets this module's training/eval mode. Defaults to training (matches every mainstream framework's Mod...
2D convolution – implemented via im2col + DeviceBackend::gemm (no new backend primitive).
Inverted dropout – the first module whose forward behavior genuinely differs between training and inf...
Reshape-only Module – flattens every non-batch dim of a (N, ...) input to (N, flattened_features),...
Dense/fully-connected layer – the reference Module implementation.
2D max pooling with a stride and padding (non-overlapping windows by default).
Abstract base every layer subclasses – NVI forward(), pure-virtual LRP contract.
Definition acquisition_functions.hpp:16
TorchvisionLoadReport LoadTorchvisionWeights(Module &model, const std::string &safetensors_path)
Loads a torchvision state dict converted to safetensors (tools/convert/pickle_to_safetensors....
OpType
The op-type tag a Node carries. Charter Part 2 §3: nodes are tagged by a small closed set of op types...
Definition op_type.hpp:19
ReLU activation – the second Module subclass, following LinearModule's pattern.
Generic skip-connection wrapper – Phase 4's only mission, the ResNet-style residual block generalized...
Model container – chains a sequence of existing Modules.
Configuration for LRP relevance propagation: which rule a module applies and its hyperparameters....
Definition lrp_rule_config.hpp:72
A torchvision ResNet with basic blocks (ResNet18 and ResNet34; ResNet18 is the one tested on the publ...
Definition vision_models.hpp:31
int64_t in_channels
Definition vision_models.hpp:37
std::vector< int64_t > blocks
Blocks per stage: {2, 2, 2, 2} is ResNet18, {3, 4, 6, 3} ResNet34.
Definition vision_models.hpp:33
int64_t width
The first stage's channels, doubled at each later stage (torchvision: 64).
Definition vision_models.hpp:35
int64_t num_classes
Definition vision_models.hpp:36
What LoadTorchvisionWeights() did.
Definition vision_models.hpp:157
std::vector< std::string > left_at_default
Model parameters the file doesn't have, left as they were (torchvision's absent convolution biases).
Definition vision_models.hpp:161
int64_t loaded
Definition vision_models.hpp:158
A torchvision VGG without BatchNorm (VGG11 to VGG19; VGG16 is the one tested on the published weights...
Definition vision_models.hpp:107
int64_t pool_size
The adaptive average pool's output size (7 in torchvision).
Definition vision_models.hpp:112
int64_t in_channels
Definition vision_models.hpp:116
std::vector< int64_t > features
The convolution widths in order, 0 for a 2x2 max pool: torchvision's config "D" for VGG16.
Definition vision_models.hpp:110
int64_t hidden
The two hidden classifier widths (4096 in torchvision).
Definition vision_models.hpp:114
int64_t num_classes
Definition vision_models.hpp:115