@@ -17,7 +17,7 @@ This project implements a Convolutional Neural Network (CNN) for MNIST digit cla
17
17
- Parameter count verification (< 20,000 parameters)
18
18
- Input shape validation (28x28)
19
19
- Output shape validation (10 classes)
20
- - Accuracy threshold (> 95 % on test set)
20
+ - Accuracy threshold (> 99.3 % on test set)
21
21
- Automatic model versioning with timestamp and accuracy
22
22
- CPU-only training support for both local and CI environments
23
23
@@ -77,7 +77,7 @@ The pipeline is triggered on every push to the repository.
77
77
## Model Training Details
78
78
- Dataset: MNIST (60,000 training images, 10,000 test images)
79
79
- Batch Size: 128
80
- - Optimizer: Adam (lr=0.01)
80
+ - Optimizer: SGD (lr=0.01)
81
81
- Loss Function: Negative Log Likelihood
82
82
- Training: Single epoch with progress bar showing loss and accuracy
83
83
- Model Saving: Automatic with timestamp and accuracy in filename
@@ -93,13 +93,13 @@ The pipeline is triggered on every push to the repository.
93
93
## Testing
94
94
The test suite (` test_model.py ` ) verifies:
95
95
- Model architecture compatibility with 28x28 input images
96
- - Total parameter count (< 25 ,000)
97
- - Model accuracy on test set (> 95 %)
96
+ - Total parameter count (< 20 ,000)
97
+ - Model accuracy on test set (> 99.3 %)
98
98
- Output shape (10 classes)
99
99
100
100
## Model Artifacts
101
101
Trained models are saved in the ` models/ ` directory with the name mnist_model_YYYYMMDD_HHMMSS_accXX.X.pth
102
102
where:
103
103
- YYYYMMDD: Date
104
104
- HHMMSS: Time
105
- - XX.X: Achieved accuracy
105
+ - XX.X: Achieved accuracy
0 commit comments