Model save

README.md

@@ -0,0 +1,75 @@
+---
+tags:
+- generated_from_trainer
+datasets:
+- imagefolder
+metrics:
+- accuracy
+model-index:
+- name: MoE-leaf-disease-convnextv2-base-22k-224
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# MoE-leaf-disease-convnextv2-base-22k-224
+
+This model is a fine-tuned version of [](https://huggingface.co/) on the imagefolder dataset.
+It achieves the following results on the evaluation set:
+- Loss: 0.3415
+- Accuracy: 0.8827
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 5e-05
+- train_batch_size: 2000
+- eval_batch_size: 2000
+- seed: 42
+- gradient_accumulation_steps: 4
+- total_train_batch_size: 8000
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_ratio: 0.1
+- num_epochs: 16
+
+### Training results
+
+| Training Loss | Epoch | Step | Validation Loss | Accuracy |
+|:-------------:|:-----:|:----:|:---------------:|:--------:|
+| No log        | 0.8   | 2    | 0.4614          | 0.8808   |
+| No log        | 2.0   | 5    | 0.3440          | 0.8827   |
+| No log        | 2.8   | 7    | 0.3441          | 0.8827   |
+| 0.3155        | 4.0   | 10   | 0.3439          | 0.8827   |
+| 0.3155        | 4.8   | 12   | 0.3437          | 0.8827   |
+| 0.3155        | 6.0   | 15   | 0.3431          | 0.8827   |
+| 0.3155        | 6.8   | 17   | 0.3426          | 0.8827   |
+| 0.2577        | 8.0   | 20   | 0.3421          | 0.8827   |
+| 0.2577        | 8.8   | 22   | 0.3419          | 0.8827   |
+| 0.2577        | 10.0  | 25   | 0.3417          | 0.8827   |
+| 0.2577        | 10.8  | 27   | 0.3416          | 0.8827   |
+| 0.2601        | 12.0  | 30   | 0.3415          | 0.8827   |
+| 0.2601        | 12.8  | 32   | 0.3415          | 0.8827   |
+
+
+### Framework versions
+
+- Transformers 4.39.3
+- Pytorch 2.2.1
+- Datasets 2.18.0
+- Tokenizers 0.15.1

modeling_moe.py

@@ -0,0 +1,75 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PreTrainedModel, AutoModelForImageClassification
+from .configuration_moe import MoEConfig
+
+
+def subgate(num_out):
+    layers = nn.Sequential(
+        nn.Flatten(),
+        nn.Linear(224 * 224 * 3, 1024),
+        nn.ReLU(),
+        nn.Linear(1024, 512),
+        nn.ReLU(),
+        nn.Linear(512, num_out),
+    )
+    return layers
+
+
+class MoEModelForImageClassification(PreTrainedModel):
+    config_class = MoEConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_classes = config.num_classes
+        self.switch_gate_model = AutoModelForImageClassification.from_pretrained(
+            config.switch_gate
+        )
+        self.baseline_model = AutoModelForImageClassification.from_pretrained(
+            config.baseline_model
+        )
+        self.expert_model_1 = AutoModelForImageClassification.from_pretrained(
+            config.experts[0]
+        )
+        self.expert_model_2 = AutoModelForImageClassification.from_pretrained(
+            config.experts[1]
+        )
+
+        self.subgate = subgate(2)
+
+        # Freeze all params
+        for module in [
+            self.switch_gate_model,
+            self.baseline_model,
+            self.expert_model_1,
+            self.expert_model_2,
+        ]:
+            for param in module.parameters():
+                param.requires_grad = False
+
+    def forward(self, pixel_values, labels=None):
+        switch_gate_result = self.switch_gate_model(pixel_values).logits
+        expert1_result = self.expert_model_1(pixel_values).logits
+        expert2_result = self.expert_model_2(pixel_values).logits
+
+        # Gating Network
+        experts_result = torch.stack(
+            [expert1_result, expert2_result], dim=1
+        ) * switch_gate_result.unsqueeze(-1)
+
+        experts_result = experts_result.sum(dim=1)
+        baseline_model_result = self.baseline_model(pixel_values).logits
+
+        subgate_result = self.subgate(pixel_values)
+        subgate_prob = F.softmax(subgate_result, dim=-1)
+
+        experts_and_base_result = torch.stack(
+            [experts_result, baseline_model_result], dim=1
+        ) * subgate_prob.unsqueeze(-1)
+
+        logits = experts_and_base_result.sum(dim=1)
+        if labels is not None:
+            loss = F.cross_entropy(logits, labels)
+            return {"loss": loss, "logits": logits}
+        return {"logits": logits}