Eraly-ml/centraasia-ResNet-50

ResNet-50 Model for Central Asian Image Classification

Model Description

This is a pre-trained ResNet-50 model fine-tuned on the Central Asian Food Dataset. The model is used for image classification across multiple classes. The data was split into training, validation, and test sets. The model was trained using gradient descent with an SGD optimizer and CrossEntropyLoss as the loss function.

Training Parameters

• Epochs: 25
• Batch Size: 32
• Learning Rate: 0.001
• Optimizer: SGD with momentum of 0.9
• Loss Function: CrossEntropyLoss

Results

Training and Validation, F1

Stage	Loss (train)	Accuracy (train)	Loss (val)	Accuracy (val)
Epoch 1	2.1171	47.00%	0.8727	75.00%
Epoch 2	1.0462	69.00%	0.6721	78.00%
...	...	...	...	...
Epoch 25	0.4286	86.00%	0.4349	86.00%

Model was trained on two T4 GPUs in a Kaggle notebook trained 36m 7s
Best validation accuracy: 86,54%

                      precision    recall  f1-score   support

            achichuk       0.91      0.98      0.94        41
         airan-katyk       0.84      0.93      0.89        46
                asip       0.78      0.57      0.66        37
            bauyrsak       0.90      0.90      0.90        62
   beshbarmak-w-kazy       0.71      0.84      0.77        44
  beshbarmak-wo-kazy       0.86      0.69      0.76        61
           chak-chak       0.94      0.94      0.94        93
            cheburek       0.92      0.88      0.90        94
        doner-lavash       0.77      1.00      0.87        20
           doner-nan       0.86      0.82      0.84        22
             hvorost       0.98      0.86      0.91       141
            irimshik       0.96      0.94      0.95       175
         kattama-nan       0.84      0.88      0.86        66
          kazy-karta       0.72      0.78      0.75        46
                kurt       0.86      0.97      0.91        61
             kuyrdak       0.92      0.93      0.92        58
       kymyz-kymyran       0.93      0.82      0.87        49
        lagman-fried       0.86      0.95      0.90        38
       lagman-w-soup       0.90      0.80      0.85        75
      lagman-wo-soup       0.58      0.86      0.69        22
               manty       0.91      0.95      0.93        63
               naryn       0.97      0.99      0.98        84
        nauryz-kozhe       0.88      0.96      0.92        52
               orama       0.68      0.84      0.75        38
                plov       0.95      0.98      0.97       101
               samsa       0.91      0.93      0.92       106
    shashlyk-chicken       0.68      0.65      0.66        62
  shashlyk-chicken-v       0.74      0.76      0.75        33
   shashlyk-kuskovoi       0.75      0.75      0.75        71
 shashlyk-kuskovoi-v       0.53      0.79      0.64        29
shashlyk-minced-meat       0.74      0.69      0.72        42
          sheep-head       0.75      0.94      0.83        16
             shelpek       0.77      0.86      0.81        64
              shorpa       0.95      0.88      0.91        80
          soup-plain       0.96      0.94      0.95        71
              sushki       0.83      1.00      0.91        43
               suzbe       0.89      0.82      0.86        62
            taba-nan       0.92      0.80      0.86       136
        talkan-zhent       0.86      0.80      0.83        90
      tushpara-fried       0.79      0.74      0.76        46
     tushpara-w-soup       0.94      0.94      0.94        67
    tushpara-wo-soup       0.92      0.87      0.89        91

            accuracy                           0.87      2698
           macro avg       0.84      0.86      0.85      2698
        weighted avg       0.88      0.87      0.87      2698

Testing

After training, the model was tested on the test set:

• Test accuracy: 87%

Repository Structure

• main.py — Code for training and testing the model
• model/ — Saved model in SafeTensors format

Usage Instructions

from transformers import AutoModelForImageClassification from huggingface_hub import hf_hub_download from safetensors.torch import load_file

repo_id = "Eraly-ml/centraasia-ResNet-50" filename = "model.safetensors"

Load model

model_path = hf_hub_download(repo_id=repo_id, filename=filename)
model = AutoModelForImageClassification.from_pretrained(repo_id)
model.load_state_dict(load_file(model_path))

My telegram @eralyf