Upload folder using huggingface_hub

README.md

@@ -0,0 +1,29 @@
+---
+license: apache-2.0
+pipeline_tag: image-classification
+tags:
+- pretrained
+---
+
+# Model Card for MNIST-MLP
+
+This is a simple MLP trained on the MNIST dataset.
+
+Its primary use is to be a very simple reference model to test quantization.
+
+## Inputs preprocessing
+
+The MNIST images must be normalized and flattened as follows:
+
+```
+from torchvision import datasets, transforms
+
+
+transform=transforms.Compose([
+        transforms.ToTensor(),
+        transforms.Normalize((0.1307,), (0.3081,)),
+        transforms.Lambda(lambda x: torch.flatten(x)),
+])
+test_set = datasets.MNIST('../data', train=False, download=True,
+                          transform=transform)
+```

config.json

@@ -0,0 +1,15 @@
+{
+  "architectures": [
+    "MLP"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_mlp.MLPConfig",
+    "AutoModel": "modeling_mlp.MLP"
+  },
+  "hidden_size": 256,
+  "input_size": 784,
+  "model_type": "mlp",
+  "output_size": 10,
+  "torch_dtype": "float32",
+  "transformers_version": "4.34.0"
+}

configuration_mlp.py

@@ -0,0 +1,20 @@
+from transformers import PretrainedConfig
+from typing import List
+
+
+class MLPConfig(PretrainedConfig):
+    model_type="mlp"
+
+    def __init__(
+        self,
+        input_size: int = 784,
+        output_size: int = 10,
+        hidden_size: int = 256,
+        **kwargs,
+    ):
+        self.input_size = input_size
+        self.output_size = output_size
+        self.hidden_size = hidden_size
+        super().__init__(**kwargs)
+
+MLPConfig.register_for_auto_class()

modeling_mlp.py

@@ -0,0 +1,22 @@
+import torch
+
+from transformers import PreTrainedModel
+
+from .configuration_mlp import MLPConfig
+
+
+class MLP(PreTrainedModel):
+    config_class = MLPConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.input_layer = torch.nn.Linear(config.input_size, config.hidden_size)
+        self.mid_layer = torch.nn.Linear(config.hidden_size, config.hidden_size)
+        self.output_layer = torch.nn.Linear(config.hidden_size, config.output_size)
+
+    def forward(self, inputs):
+        x = torch.nn.functional.relu(self.input_layer(inputs))
+        x = torch.nn.functional.relu(self.mid_layer(x))
+        return torch.nn.functional.softmax(self.output_layer(x), dim=-1)
+
+MLP.register_for_auto_class("AutoModel")

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26300c9e4b676ba94c7773ee82de3bb2831b9fd45cc1202b65bc94a6e4db4fa5
+size 1079431