Upload quantized model and evaluation results

README.md

@@ -0,0 +1,16 @@
+# 🧠 Model: Resnet18
+
+This is a quantized version of `Resnet18` using `W8A8 static` quantization.
+
+## 🧪 Evaluation Summary
+
+| Metric         | FP32             | Quantized  |
+|----------------|------------------|------------|
+| Top-1 Accuracy | 0.555 | 0.6577380895614624 |
+| Top-5 Accuracy | 0.777 | 0.8541666865348816 |
+
+- Dataset: `ImageNet`
+- Evaluation Date: `2025-07-20`
+
+## 🔍 Notes
+Quantized W8A8 Resnet18 model from torchvision

eval_results.json

@@ -0,0 +1 @@
+{"top1": 0.6577380895614624, "top5": 0.8541666865348816}

liteml_config.yaml

@@ -0,0 +1,18 @@
+QAT:
+  device: "cuda"
+  fully_quantized: True
+  data_quantization:
+    status: On
+    bits: 8
+    custom_bits: {}
+    symmetric: On
+    quantization_mode: static
+    observer: "MovingAverage"
+    per_channel: False
+
+  weights_quantization:
+    status: On
+    bits: 8
+    custom_bits: {}
+    symmetric: On
+    per_channel: False

metadata.yaml

@@ -0,0 +1,8 @@
+model_name: Resnet18
+liteml_config: 'liteml_config.yaml'
+quantization_type: W8A8 static
+dataset: ImageNet
+notes: Quantized W8A8 Resnet18 model from torchvision
+# TODO: move float accuracy somewhere else
+top1_float: 0.555
+top5_float: 0.777

model.py

@@ -0,0 +1,7 @@
+import torch
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+
+def get_model():
+    from torchvision.models import resnet18
+    return resnet18(weights='IMAGENET1K_V1').to(device)

quantized_model.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6a705f91a3c805bfb96fa9e960f6b3821d9423463a477a3e4be9f57afd92d06
+size 46844509