grad-cam

Abhishek Das ea86dc1c57 Merge pull request #5 from varunagrawal/master Cmd Option for custom output name		7 years ago
VQA_LSTM_CNN @ ffd1d75469	Adds VQA eval & grad-CAM code	8 years ago

images	Adds more example images	7 years ago

misc	Fixes typo	7 years ago

models	Removes `ques_feat.json`; Minor typos / changes in parameter definitions	7 years ago

neuraltalk2 @ 3c81602809	initial commit	8 years ago

.gitignore	Model download script	8 years ago

.gitmodules	initial commit	8 years ago

README.md	removes paper image and fixes font in approach figure	7 years ago

captioning.lua	Updates `captioning.lua` for CPU	7 years ago

classification.lua	added command line option to append custom output name to generated images	7 years ago

visual_question_answering.lua	Updates `visual_question_answering.lua` for CPU	7 years ago

README.md

Grad-CAM: Gradient-weighted Class Activation Mapping
- Usage
- License
- - 3rd-party

Grad-CAM: Gradient-weighted Class Activation Mapping

Code for the paper

Grad-CAM: Why did you say that? Visual Explanations from Deep Networks via Gradient-based Localization
Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, Dhruv Batra
https://arxiv.org/abs/1610.02391

Demo: gradcam.cloudcv.org

Usage

Download Caffe model(s) and prototxt for VGG-16/VGG-19/AlexNet using sh models/download_models.sh.

Classification

th classification.lua -input_image_path images/cat_dog.jpg -label 243 -gpuid 0
th classification.lua -input_image_path images/cat_dog.jpg -label 283 -gpuid 0

Options

proto_file: Path to the deploy.prototxt file for the CNN Caffe model. Default is models/VGG_ILSVRC_16_layers_deploy.prototxt
model_file: Path to the .caffemodel file for the CNN Caffe model. Default is models/VGG_ILSVRC_16_layers.caffemodel
input_image_path: Path to the input image. Default is images/cat_dog.jpg
input_sz: Input image size. Default is 224 (Change to 227 if using AlexNet)
layer_name: Layer to use for Grad-CAM. Default is relu5_3 (use relu5_4 for VGG-19 and relu5 for AlexNet)
label: Class label to generate grad-CAM for (-1 = use predicted class, 283 = Tiger cat, 243 = Boxer). Default is -1. These correspond to ILSVRC synset IDs
out_path: Path to save images in. Default is output/
gpuid: 0-indexed id of GPU to use. Default is -1 = CPU
backend: Backend to use with loadcaffe. Default is nn
save_as_heatmap: Whether to save heatmap or raw Grad-CAM. 1 = save heatmap, 0 = save raw Grad-CAM. Default is 1

Examples

'border collie' (233)

'tabby cat' (282)

'boxer' (243)

'tiger cat' (283)

Visual Question Answering

Clone the VQA (http://arxiv.org/abs/1505.00468) sub-repository (git submodule init && git submodule update), and download and unzip the provided extracted features and pretrained model.

th visual_question_answering.lua -input_image_path images/cat_dog.jpg -question 'What animal?' -answer 'dog' -gpuid 0
th visual_question_answering.lua -input_image_path images/cat_dog.jpg -question 'What animal?' -answer 'cat' -gpuid 0

Options

proto_file: Path to the deploy.prototxt file for the CNN Caffe model. Default is models/VGG_ILSVRC_19_layers_deploy.prototxt
model_file: Path to the .caffemodel file for the CNN Caffe model. Default is models/VGG_ILSVRC_19_layers.caffemodel
input_image_path: Path to the input image. Default is images/cat_dog.jpg
input_sz: Input image size. Default is 224 (Change to 227 if using AlexNet)
layer_name: Layer to use for Grad-CAM. Default is relu5_4 (use relu5_3 for VGG-16 and relu5 for AlexNet)
question: Input question. Default is What animal?
answer: Optional answer (For eg. "cat") to generate Grad-CAM for ('' = use predicted answer). Default is ''
out_path: Path to save images in. Default is output/
model_path: Path to VQA model checkpoint. Default is VQA_LSTM_CNN/lstm.t7
gpuid: 0-indexed id of GPU to use. Default is -1 = CPU
backend: Backend to use with loadcaffe. Default is cudnn
save_as_heatmap: Whether to save heatmap or raw Grad-CAM. 1 = save heatmap, 0 = save raw Grad-CAM. Default is 1

Examples

What animal? Dog

What animal? Cat

What color is the fire hydrant? Green

What color is the fire hydrant? Yellow

What color is the fire hydrant? Green and Yellow

What color is the fire hydrant? Red and Yellow

Image Captioning

Clone the neuraltalk2 sub-repository. Running sh models/download_models.sh will download the pretrained model and place it in the neuraltalk2 folder.

Change lines 2-4 of neuraltalk2/misc/LanguageModel.lua to the following:

local utils = require 'neuraltalk2.misc.utils'
local net_utils = require 'neuraltalk2.misc.net_utils'
local LSTM = require 'neuraltalk2.misc.LSTM'

th captioning.lua -input_image_path images/cat_dog.jpg -caption 'a dog and cat posing for a picture' -gpuid 0
th captioning.lua -input_image_path images/cat_dog.jpg -caption '' -gpuid 0

Options

input_image_path: Path to the input image. Default is images/cat_dog.jpg
input_sz: Input image size. Default is 224 (Change to 227 if using AlexNet)
layer: Layer to use for Grad-CAM. Default is 30 (relu5_3 for vgg16)
caption: Optional input caption. No input will use the generated caption as default
out_path: Path to save images in. Default is output/
model_path: Path to captioning model checkpoint. Default is neuraltalk2/model_id1-501-1448236541.t7
gpuid: 0-indexed id of GPU to use. Default is -1 = CPU
backend: Backend to use with loadcaffe. Default is cudnn
save_as_heatmap: Whether to save heatmap or raw Grad-CAM. 1 = save heatmap, 0 = save raw Grad-CAM. Default is 1