initial commit

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+09_pretrained_effnetb2_feature_extractor_pizza_steak_sushi_20percent.pth filter=lfs diff=lfs merge=lfs -text

09_pretrained_effnetb2_feature_extractor_pizza_steak_sushi_20percent.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4aff8750b47940bc957d7ab187889f5ff6fece8890d2b9b1f47bfc9f0eb50129
+size 31314042

app.py

@@ -0,0 +1,79 @@
+
+### 1. Import and class names setup ###
+import gradio as gr
+import os
+from pathlib import Path
+import torch
+
+from model import create_effnetb2_model
+from time import perf_counter
+from typing import Tuple, Dict
+
+from PIL import Image
+import torchvision
+
+# Setup class names (hardcoded, these shall reside in a json file or sth like that...)
+class_names = ["pizza","steak","sushi"]
+
+### 2. Model and transforms preparation ###
+effnetb2_model, effnetb2_transforms = create_effnetb2_model(num_classes=len(class_names))
+
+# Load save weights
+effnetb2_model.load_state_dict(torch.load(f="09_pretrained_effnetb2_feature_extractor_pizza_steak_sushi_20percent.pth",
+                                          map_location=torch.device("cpu"))) # map location to cpu is a must, as we have trained our model in the GPU
+
+### 3. Predict function
+
+def predict(img) -> Tuple[Dict,float]:
+  # Start a timer
+  start_time = perf_counter()
+
+  # Transform the input image for use with EffNetB2
+  effnetb2_transforms = torchvision.models.EfficientNet_B2_Weights.DEFAULT.transforms()
+  img_tensor = effnetb2_transforms(img)
+
+  # Put model in eval and inference
+  effnetb2_model.eval()
+  with torch.inference_mode():
+    y_logits = effnetb2_model(img_tensor.unsqueeze(dim=0))
+    y_pred_probs = torch.softmax(y_logits,dim=1)
+    y_pred_probs_list = y_pred_probs.squeeze().tolist()
+  
+  # Creatae a prediction probability dictionary
+  pred_prob_dict = {class_names[i]:float(prob) for i,prob in enumerate(y_pred_probs_list)}
+
+  # End timer
+  end_time = perf_counter()
+
+  return pred_prob_dict, round(end_time-start_time,4)
+
+
+### 4. Launch app
+
+import gradio as gr
+
+foodvision_mini_examples_path = "examples"
+
+example_list = [str(path) for path in foodvision_mini_examples_path.rglob("*.jpg")]
+
+# Create title, description and article
+title = "FoodVisionMini V0 🍕🍖🍣"
+description = "An <a href='https://pytorch.org/vision/main/models/generated/torchvision.models.efficientnet_b2.html#torchvision.models.efficientnet_b2'>EfficientNetB2</a> feature extractor computer vision model to classify images into pizza, steak or sushi<br>I have yet to improve it to label non-food images. Paciencia amigos"
+article = "Created at <a href='#'>09_pytorch_model_deploy.ipynb</a> Google Colab notebook"
+
+# Create the Gradio demo
+demo = gr.Interface(fn=predict,
+                    inputs=gr.Image(type="pil"),
+                    outputs=[gr.Label(num_top_classes=3, label="predictions"),
+                             gr.Number(label="Prediction time (s)")],
+                    examples=example_list,
+                    title=title,
+                    description=description,
+                    article=article)
+
+# Launch the demo
+demo.launch(debug=False,# print errors locally in local colab debug=False
+            share=True) # run on public url
+
+
+# *** IMPORTANTE: The Flag button of the interface will create a folder named "flagged" that will contain the images and predictions of those images that someone has Flagged***

model.py

@@ -0,0 +1,37 @@
+
+import torch
+import torchvision
+from torch import nn
+from torchvision.models._api import WeightsEnum
+from torch.hub import load_state_dict_from_url
+
+def create_effnetb2_model(num_classes:int=3, seed:int=42):
+  # https://pytorch.org/vision/main/models/generated/torchvision.models.efficientnet_b2.html#torchvision.models.efficientnet_b2
+
+  def get_state_dict(self, *args, **kwargs):
+      kwargs.pop("check_hash")
+      return load_state_dict_from_url(self.url, *args, **kwargs)
+  WeightsEnum.get_state_dict = get_state_dict
+
+
+  # 1. Setup pretrained EffNetB2 weights
+  effnetb2_weights = torchvision.models.EfficientNet_B2_Weights.DEFAULT # DEFAULT = BEST
+
+  # 2. Get EffNetB2 transforms
+  effnetb2_transforms = effnetb2_weights.transforms()
+
+  # 3. Setup pretrained model instance
+  effnetb2_model = torchvision.models.efficientnet_b2(weights=effnetb2_weights)
+
+  # 4. Freeze the base layers in the model
+  for param in effnetb2_model.features.parameters():
+    param.requires_grad = False
+
+  # 5. Modify the classifier
+  torch.manual_seed(seed)
+  effnetb2_model.classifier = nn.Sequential(
+      nn.Dropout(p=0.3, inplace=True),
+      nn.Linear(in_features=1408,out_features=num_classes,bias=True)
+  )
+
+  return effnetb2_model, effnetb2_transforms

requirements.txt

@@ -0,0 +1,3 @@
+torch
+torchvision
+gradio