Demo_Parameters.py

# -*- coding: utf-8 -*-
"""
Created on Mon Dec 16 11:20:44 2019
Parameters for histogram layer experiments
Only change parameters in this file before running
demo.py
@author: jpeeples 
"""

######## ONLY CHANGE PARAMETERS BELOW ########
#Flag for if results are to be saved out
#Set to True to save results out and False to not save results
save_results = True

#Location to store trained models
#Always add slash (/) after folder name
folder = 'Saved_Models/'

#Flag to use histogram model or baseline global average pooling (GAP)
# Set to True to use histogram layer and False to use GAP model 
histogram = True

#Select histogram layer type: RBF or Piecewise Linear. 
#Recommended is RBF (implements histogram function in paper)
histogram_type = 'Linear'

#Flag to create separate validation set
#For paper, only used train/test split (default: False)
val_split = False

#Parallelize results
Parallelize_model = True

#Select dataset. Set to number of desired texture dataset
data_selection = 1
Dataset_names = { 1: 'DTD', 2: 'GTOS-mobile', 3: 'MINC_2500'}

#Set random state for K fold CV for repeatability of data
random_state = 1

#Number of bins for histogram layer. Recommended values are 4, 8 and 16.
#Set number of bins to powers of 2 (e.g., 2, 4, 8, etc.)
#For HistRes_B models using ResNet18 and ResNet50, do not set number of bins
#higher than 128 and 512 respectively. Note: a 1x1xK convolution is used to 
#downsample feature maps before binning process. If the bin values are set
#higher than 128 or 512 for HistRes_B models using ResNet18 or ResNet50 
#respectively, than an error will occur due to attempting to reduce the number of 
#features maps to values less than one
numBins = 4

#Flag for feature extraction. False, train whole model. True, only update 
#fully connected and histogram layers parameters (default: False)
#Flag to use pretrained model from ImageNet or train from scratch (default: True)
#Flag to add BN to convolutional features (default:True)
#Location/Scale at which to apply histogram layer (default: 5 (at the end))
feature_extraction = False
use_pretrained = True
add_bn = True
scale = 5

#Set learning rate for new and pretrained (pt) layers
#Recommended values are to have the new layers at 
#a learning rate 10 times larger than the pt learning rate.
#e.g., new_lr = .001 and pt_lr = .01
pt_lr = .001
new_lr = .01

#Set momentum for SGD optimizer. 
#Recommended value is .9 (used in paper)
alpha = .9

#Parameters of Histogram Layer
#For no padding, set 0. If padding is desired,
#enter amount of zero padding to add to each side of image 
#(did not use padding in paper, recommended value is 0 for padding)
padding = 0

#Apply rotation to test set (did not use in paper)
#Set rotation to True to add rotation, False if no rotation (used in paper)
#Recommend values are between 0 and 25 degrees
#Can use to test robustness of model to rotation transformation
rotation = False
degrees = 25

#Reduce dimensionality based on number of output feature maps from GAP layer
#Used to compute number of features from histogram layer
out_channels = {"resnet50": 2048, "resnet18": 512}

#Set whether to have the histogram layer inline or parallel (default: parallel)
#Set whether to use sum (unnormalized count) or average pooling (normalized count)
# (default: average pooling)
#Set whether to enforce sum to one constraint across bins (default: True)
parallel = True
normalize_count = True
normalize_bins = True

#Set step_size and decay rate for scheduler
#In paper, learning rate was decayed factor of .1 every ten epochs (recommended)
step_size = 10
gamma = .1

#Batch size for training and epochs. If running experiments on single GPU (e.g., 2080ti),
#training batch size is recommended to be 64. If using at least two GPUs, 
#the recommended training batch size is 128 (as done in paper)
#May need to reduce batch size if CUDA out of memory issue occurs
batch_size = {'train': 8, 'val': 256, 'test': 256}
num_epochs = 30

#Resize the image before center crop. Recommended values for resize is 256 (used in paper), 384,
#and 512 (from http://openaccess.thecvf.com/content_cvpr_2018/papers/Xue_Deep_Texture_Manifold_CVPR_2018_paper.pdf)
#Center crop size is recommended to be 256.
resize_size = 256
center_size = 256

#Pin memory for dataloader (set to True for experiments)
pin_memory = True

#Set number of workers, i.e., how many subprocesses to use for data loading.
#Usually set to 0 or 1. Can set to more if multiple machines are used.
#Number of workers for experiments for two GPUs was three
num_workers = 0

#Output feature map size after histogram layer
feat_map_size = 4

#Flag for TSNE visuals, set to True to create TSNE visual of features
#Set to false to not generate TSNE visuals
#Separate TSNE will visualize histogram and GAP features separately
#If set to True, TSNE of histogram and GAP features will be created
#Number of images to view for TSNE (defaults to all training imgs unless
#value is less than total training images).
TSNE_visual = True
Separate_TSNE = False
Num_TSNE_images = 10000

#Visualization parameters for figures
fig_size = 12
font_size = 16

#Set filter size and stride based on scale
# Current values will produce 2x2 local feature maps
if scale == 1:
    stride = [32, 32] 
    in_channels = {"resnet50": 64, "resnet18": 64}
    kernel_size = {"resnet50": [64,64],  "resnet18": [64,64]}
elif scale == 2:
    stride = [16, 16] 
    in_channels = {"resnet50": 256, "resnet18": 64}
    kernel_size = {"resnet50": [32,32],  "resnet18": [32,32]}
elif scale == 3:
    stride = [8, 8] 
    in_channels = {"resnet50": 512, "resnet18": 128}
    kernel_size = {"resnet50": [16,16],  "resnet18": [16,16]}
elif scale == 4: 
    stride = [4, 4] 
    in_channels = {"resnet50": 1024, "resnet18": 256}
    kernel_size = {"resnet50": [8,8],  "resnet18": [8,8]}
else:
    stride = [2, 2] 
    in_channels = {"resnet50": 2048, "resnet18": 512}
    kernel_size = {"resnet50": [4,4],  "resnet18": [4,4]}


######## ONLY CHANGE PARAMETERS ABOVE ########
if feature_extraction:
    mode = 'Feature_Extraction'
else:
    mode = 'Fine_Tuning'

#Location of texture datasets
Data_dirs = {'DTD': './Datasets/DTD/', 
             'MINC_2500': './Datasets/minc-2500/',
             'GTOS-mobile': './Datasets/gtos-mobile'}

#ResNet models to use for each dataset
Model_names = {'DTD': 'resnet50', 
               'MINC_2500': 'resnet50',
               'GTOS-mobile': 'resnet18'}

#Number of classes in each dataset
num_classes = {'DTD': 47, 
               'MINC_2500': 23,
               'GTOS-mobile': 31}

#Number of runs and/or splits for each dataset
Splits = {'DTD': 10, 
          'MINC_2500': 5,
          'GTOS-mobile': 5}

Dataset = Dataset_names[data_selection]
data_dir = Data_dirs[Dataset]

Hist_model_name = 'HistRes_' + str(numBins)

#Return dictionary of parameters
Network_parameters = {'save_results': save_results,'folder': folder, 
                      'histogram': histogram,'Dataset': Dataset, 'data_dir': data_dir,
                      'num_workers': num_workers, 'mode': mode,'new_lr': new_lr, 
                      'pt_lr': pt_lr,'momentum': alpha, 'step_size': step_size,
                      'gamma': gamma, 'batch_size' : batch_size, 
                      'num_epochs': num_epochs, 'resize_size': resize_size, 
                      'center_size': center_size, 'padding': padding, 
                      'stride': stride, 'kernel_size': kernel_size,
                      'in_channels': in_channels, 'out_channels': out_channels,
                      'normalize_count': normalize_count, 
                      'normalize_bins': normalize_bins,'parallel': parallel,
                      'numBins': numBins,'feat_map_size': feat_map_size,
                      'Model_names': Model_names, 'num_classes': num_classes, 
                      'Splits': Splits, 'feature_extraction': feature_extraction,
                      'hist_model': Hist_model_name, 'use_pretrained': use_pretrained,
                      'add_bn': add_bn, 'pin_memory': pin_memory, 'scale': scale,
                      'degrees': degrees, 'rotation': rotation, 
                      'histogram_type': histogram_type, 'random_state': random_state,
                      'TSNE_visual': TSNE_visual,
                      'Separate_TSNE': Separate_TSNE, 'Parallelize': Parallelize_model,
                      'Num_TSNE_images': Num_TSNE_images,'fig_size': fig_size,
                      'font_size': font_size, 'val_split': val_split}