yaoqingyuan / pytorch_classifiers

Almost any Image classification problem using pytorch

This is an experimental setup to build code base for pytorch. Its main is to experiment faster using transfer learning on all available pre-trained models.

Dataset: Plant Seedlings classification

Classes present:

• Black-grass
• Charlock
• Cleavers
• Common Chickweed
• Common wheat
• Fat Hen
• Loose Silky-bent
• Maize
• Scentless Mayweed
• Shepherds Purse
• Small-flowered Cranesbill
• Sugar beet

Run the following commands

To train the models:

make -f Makefile.train

To test the models:

make -f Makefile.predict

Results

total number of train images: 4268 total number of val images: 482 total number of test images: 794

Results with Full-Agumentation strategy:

Trained the networks in three methods:

• Full Finetuning
• Freeze first few layers

Case-1 - Finetuning entire network

Models	Train Accuracy_score	Val Accuracy_score
resnet18	0.92783	0.93153
resnet34	0.9522	0.94190
resnet50	0.95665	0.94398
resnet101	0.96696	0.96265
resnet152	0.96555	0.95643
squeezenet1_0	0.94329	0.92738
squeezenet1_1	0.93955	0.93153
densenet121	0.95243	0.9336
densenet169	0.96626	0.93983
densenet201	0.96063	0.95020
inception_v3	0.94212	0.93568
vgg11	0.93814	0.93153
vgg13	0.94493	0.94190
vgg16	0.95665	0.93568
vgg19	0.95009	0.93775
vgg11_bn	0.94142	0.93775
vgg13_bn	0.94423	0.92738
vgg16_bn	0.94634	0.94190
vgg19_bn	0.94915	0.93360
alexnet	0.91260	0.90456
resnext101_64x4d	0.98055	0.96887
resnext101_32x4d	0.98172	0.96887
nasnetalarge	0.96907	0.96265
inceptionresnetv2	0.96134	0.95435
inceptionv4	0.96930	0.96473

Case-2 - Freezed first few layers(look at code)

Models	Train Accuracy_score	Val Accuracy_score
resnet18	0.9196	0.91493
resnet34	0.94845	0.93983
resnet50	0.9564	0.93983
resnet101	0.96790	0.96265
resnet152	0.96508	0.95643
squeezenet1_0	0.94048	0.92738
squeezenet1_1	0.93088	0.92116
densenet121	0.95173	0.95228
densenet169	0.96087	0.94813
densenet201	0.95384	0.95020
inception_v3	0.94025	0.93775
vgg11	0.93697	0.92946
vgg13	0.93533	0.92323
vgg16	0.94821	0.93983
vgg19	0.95243	0.94190
vgg11_bn	0.93416	0.92738
vgg13_bn	0.93322	0.92323
vgg16_bn	0.94728	0.93775
vgg19_bn	0.94798	0.94190
alexnet	0.89784	0.88589
resnext101_64x4d	0.98617	0.96887
resnext101_32x4d	0.98195	0.96473
nasnetalarge	0.95970	0.96265
inceptionresnetv2	0.95103	0.94813
inceptionv4	0.96251	0.93775

Results with Basic Agumentation

Case-1 (Freezed all layers except last one)

Models	Train Accuracy_score	Val Accuracy_score
resnet18	0.77553	0.75518
resnet152	0.82778	0.81535
resnet101	0.82333	0.80290
resnet50	0.79943	0.78630
resnet34	0.78655	0.74688
squeezenet1_0	0.91447	0.87966
squeezenet1_1	0.90089	0.87344
densenet121	0.80880	0.81120
densenet169	0.84746	0.82987
densenet201	0.86621	0.86514
inception_v3	0.76101	0.74688
vgg11	0.78209	0.78008
vgg13	0.75960	0.72821
vgg16	0.77038	0.71576
vgg19	0.71204	0.64522
vgg11_bn	0.76522	0.74481
vgg13_bn	0.76241	0.76348
vgg16_bn	0.76265	0.75726
vgg19_bn	0.75773	0.73858
alexnet	0.83153	0.76348

Case-2 Finetuning the entire network

Models	Train Accuracy_score	Val Accuracy_score
resnet18	0.98477	0.96058
resnet152	0.99273	0.97717
resnet101	0.99367	0.97717
resnet50	0.99133	0.97510
resnet34	0.98969	0.97095
squeezenet1_0	0.96274	0.94190
squeezenet1_1	0.96485	0.92738
densenet121	0.99086	0.96887
densenet169	0.99507	0.97510
densenet201	0.99390	0.97717
inception_v3	0.98898	0.97302
vgg11	0.98031	0.95020
vgg13	0.98078	0.95643
vgg16	0.98266	0.95435
vgg19	0.98430	0.95643
vgg11_bn	0.98570	0.96265
vgg13_bn	0.98687	0.97095
vgg16_bn	0.99179	0.96680
vgg19_bn	0.99297	0.96680
alexnet	0.95970	0.92946

Submissions:

• densenet201 LB - to 97.22
• ensemble1 - mode of all best performing models LB - 97.32

TO_DO

1. Adding mixup strategy to all the networks
2. Ensembling model outputs
3. Model stacking
4. Extracting bottleneck features and using - ML to train the model - Visualization using T-sne
5. Solve issue with bninception(Model is not training)
6. Train Vggm network