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

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

Results with Basic Agumentation

Case-1 (Freezed all layers except last one)

Case-2 Finetuning the entire network

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