TransferLearning Experiments on Fashion Dataset

Dependencies:

1. Pytorch : 1.0.1post2
2. Python : 3.6.9
3. Torchvision : 0.2.2

Pre-Training:

optional arguments:
  --lr                 learning rate for pretraining
  --batch_size         num_batch
  --epochs             training epochs 
  --add_sampler        1 if weighted sampler
  --dir_               image_dir where the images are present
  --resize             model input size [224, 224, 2(for Random resize)] 
  --exp_name           Name of the experiment
  --schedule           lr drop schedules [60, 80, 120]
  

Pre-training ResNet18:

python main.py --lr 0.1 --dir_ IMG_DIR --batchsize 128 --resize 224 224 2 --exp_name pretrain_best --schedule 80 60 120 --epochs 200 

Finetuning:

Additional optional arguments:
  --finetune_lr        learning rate for finetuning
  --batch_size         num_batch
  --epochs             training epochs 
  --add_sampler        1 if weighted sampler
  --dir_               image_dir where the images are present
  --finetune_freeze    True if freeze the layers
  --freeze_layers      [fc., layer4.] which layers to train
  --train_between      To train whole network in between
  --switch_all         unfreeze all layer at epoch - 40
  
  --resize             model input size [224, 224, 2(for Random resize)] 
  --exp_name           Name of the experiment
  --schedule           lr drop schedules [60, 80, 120]
  

Finetuning pretrained model

python main.py --finetune_lr 0.1 --dir_ IMG_DIR --batch_size 128 --resize 224 224 2  --model_weights PRETRAIN_MODEL_WEIGHTS --exp_name finetune_best --schedule  60 80 120 --epochs 200 --train_between  --switch_all 40 --freeze_layers fc. layer4. --finetune_freeze --schedule 60 80 120

Testing the pretrain/finetune models

python main.py --test pretrain_test/finetune_test --model_weights PRETRAIN/FINETUNE model  --dir_ IMG_DIR 

# outputs are stored as model_name.json

Notes

For plotting Top-1 and Top-5 accuracies use :

python inference/plots.py # Jsons obtained from testing the model

Results

• Results on Fashion Pretrain:

Model Name	Acc	Very Good	Good	Medium	Less	Top 5
ResNet_80px	84.3	91.11	82.23	89.03	68.52	93.43
CResNet_80px	87.25	95.23	84.29	91.21	70.2	96.72
ResNet_IN_224px	87.02	95.16	82.1	90.2	71.8	95.58
ResNet_50	86.8	94.71	81.74	89.1	72.31	95.21

Top-1/Top-5 Pretrain bar plot

|

• Results on Fashion Finetune:

Model Name	Acc
ResNet_224px with freeze till layer 4 pretrained*	46.72
ResNet_224px with freeze only FC pretrained	39.39
ResNet_224px with weight decay 0	37.09
ResNet_224px trained from scratch	45.82
ResNet_224px with weighted Sampler	12-11

Top-1/Top-5 bar Finetune plot

|