- Add the model architecture in the console header #32

- Increase F1-Score precision to four digit floating point #25
- Change CPU / GPU device identifier for training and inference #28
- Input the file log to the Trainer for the train and dev accuracy / loss #29
- Print at the end the best model path and full f1-score #30
- Print the best model score and path for every epochs #31
- Add the support for Tensorboard #23

README.md

@@ -205,6 +205,7 @@ Build: `python setup.py sdist bdist_wheel`
 Upload: `twine upload dist/*`
 
 # Citation
+
 If you want to cite the tool you can use this:
 
 ```bibtex

build/lib/hugsvision/inference/TorchVisionClassifierInference.py

@@ -17,13 +17,14 @@ class TorchVisionClassifierInference:
   """
   def __init__(self, model_path: str, transform=transformTorchVision, device=None):
 
+    self.transform = transform
+
     if device == None:
       self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
     else:
       self.device = device
 
     self.model_path = model_path if model_path.endswith("/") else model_path + "/"
-    self.transform = transform
 
     self.model = torch.load(self.model_path + "best_model.pth", map_location=self.device)
 
@@ -42,8 +43,10 @@ def predict_image(self, img, save_preview=False):
       pil_img = transforms.ToPILImage()(img)
       pil_img.save("preview.jpg")
 
+    img = torch.unsqueeze(img, 0).to(self.device)
+
     # Predict and get the corresponding label identifier
-    pred = self.model(torch.unsqueeze(img, 0))
+    pred = self.model(img)
 
     # Get string label from index
     label = self.config["id2label"][str(torch.max(pred, 1)[1].item())]
@@ -53,5 +56,5 @@ def predict_image(self, img, save_preview=False):
   """
   🤔 Predict from one image path
   """
-  def predict(self, img_path: str):  
-    return self.predict_image(img=Image.open(img_path))
+  def predict(self, img_path: str, save_preview=False):  
+    return self.predict_image(img=Image.open(img_path), save_preview=save_preview)

build/lib/hugsvision/nnet/TorchVisionClassifierTrainer.py

@@ -52,6 +52,8 @@ def __init__(
     force_cpu     = False,
     requires_grad = False, # True: Only last layer of the classifier are updated
     load_best_model_at_end = True,
+    classification_report_digits = 4,
+    parallelized = False,
   ):
 
     self.train             = train
@@ -69,12 +71,15 @@ def __init__(
     self.ids2labels        = id2label
     self.labels2ids        = label2id
     self.best_acc          = 0
+    self.best_path         = ""
     self.logs_path         = self.output_dir + "logs/"
     self.config_path       = self.output_dir + "config.json"
     self.current_date      = datetime.today().strftime("%Y-%m-%d-%H-%M-%S")
     self.config            = {}
     self.requires_grad     = requires_grad
     self.load_best_model_at_end = load_best_model_at_end
+    self.classification_report_digits = classification_report_digits
+    self.parallelized = parallelized
 
     self.tensor_board = SummaryWriter()
 
@@ -96,8 +101,9 @@ def __init__(
     self.num_classes = len(self.ids2labels.keys())
     self.config["num_classes"] = self.num_classes
 
-    # Open the logs file
+    # Open the logs files
     self.__openLogs()
+    self.__openLogsLossAcc()
 
     # Load the model from the TorchVision Models Zoo 
     if pretrained:
@@ -114,7 +120,10 @@ def __init__(
     # """
     if self.model_name.startswith('alexnet') or self.model_name.startswith('vgg'):
       self.model.classifier[6] = nn.Linear(self.model.classifier[6].in_features, self.num_classes)
-      self.model.features = torch.nn.DataParallel(self.model.features)
+
+      if self.parallelized == True:
+        self.model.features = torch.nn.DataParallel(self.model.features)
+
       self.model.to(self.device)
       self.config["hidden_size"] = self.model.classifier[6].in_features
     else:
@@ -143,7 +152,10 @@ def __init__(
         self.model.classifier[3] = nn.Linear(self.model.classifier[3].in_features, self.num_classes)
         self.config["hidden_size"] = self.model.classifier[3].in_features
 
-      self.model = torch.nn.DataParallel(self.model).to(self.device)
+      if self.parallelized == True:
+        self.model = torch.nn.DataParallel(self.model)
+
+      self.model = self.model.to(self.device)
 
     if self.requires_grad:
       for param in self.model.parameters():
@@ -181,13 +193,15 @@ def __init__(
     self.training()
 
     if self.load_best_model_at_end:
-      print("Best model loaded!")
+      print("\033[95mBest model loaded!\033[0m")
       self.model = torch.load(self.output_dir + 'best_model.pth', map_location=self.device)
 
     self.evaluate_f1_score()
 
     # Close the logs file
     self.logs_file.close()
+    self.logs_loss.close()
+    self.logs_acc.close()
 
   """
   📜 Open the logs file
@@ -200,6 +214,21 @@ def __openLogs(self):
     # Open the logs file
     self.logs_file = open(self.logs_path + "logs_" + self.current_date + ".txt", "a")
 
+  """
+  📜 Open the logs file for loss and accuracy
+  """
+  def __openLogsLossAcc(self):
+
+    if not self.logs_path.endswith("/"):
+      self.logs_path += "/"
+
+    # Check if the directory already exist
+    os.makedirs(self.logs_path, exist_ok=True)
+
+    # Open the logs file
+    self.logs_loss = open(self.logs_path + "logs_loss_" + self.current_date + ".txt", "a")
+    self.logs_acc = open(self.logs_path + "logs_acc_" + self.current_date + ".txt", "a")
+
   """
   🧪 Evaluate the performances of the system of the test sub-dataset given a f1-score
   """
@@ -214,6 +243,7 @@ def evaluate_f1_score(self):
       labels = [int(a) for a in list(self.ids2labels.keys())],
       target_names = list(self.labels2ids.keys()),
       zero_division = 0,
+      digits=self.classification_report_digits,
     )
     print(table)
 
@@ -223,6 +253,7 @@ def evaluate_f1_score(self):
     self.logs_file.close()
 
     print("Logs saved at: \033[93m" + self.logs_path + "\033[0m")
+    print("\033[93m[" + self.model_name + "]\033[0m Best model saved at: \033[93m" + self.best_path + "\033[0m" + " - Accuracy " + "{:.2f}".format(self.best_acc*100))
 
     return all_target, all_preds
 
@@ -242,13 +273,19 @@ def training(self):
       all_targets, all_predictions = self.evaluate()
       total_acc = accuracy_score(all_targets, all_predictions)
 
-      f1_score = classification_report(all_targets, all_predictions, target_names=list(self.ids2labels.values()))
+      f1_score = classification_report(
+        all_targets,
+        all_predictions,
+        target_names=list(self.ids2labels.values()),
+        digits=self.classification_report_digits
+      )
       print(f1_score)
 
       os.makedirs(self.output_dir, exist_ok=True)
 
       if total_acc > self.best_acc:
         filename = self.output_dir + 'best_model.pth'
+        self.best_path = filename
       else:
         filename = self.output_dir + 'last_model.pth'
 
@@ -257,9 +294,12 @@ def training(self):
       torch.save(self.model, filename)
       saved_at = "Model saved at: \033[93m" + filename + "\033[0m"
       print(saved_at)
+      best_model_path = "\033[93m[" + self.model_name + "]\033[0m Best model saved at: \033[93m" + self.best_path + "\033[0m" + " - Accuracy " + "{:.2f}".format(self.best_acc*100) + "%"
+      print(best_model_path)
 
       self.logs_file.write(f1_score + "\n")
       self.logs_file.write(saved_at + "\n")
+      self.logs_file.write(best_model_path + "\n")
       self.logs_file.close()
 
       self.tensor_board.add_scalar('Loss/train', batches_loss, epoch)
@@ -280,7 +320,7 @@ def compute_batches(self, epoch):
     all_targets = []
 
     # For each batch
-    for i, (input, target) in enumerate(self.data_loader_train):
+    for i, (input, target) in tqdm(enumerate(self.data_loader_train)):
 
       input = input.to(self.device)
       target = target.to(self.device)
@@ -303,7 +343,6 @@ def compute_batches(self, epoch):
 
       if i % (len(self.data_loader_train) / 10) == 0:
         log_line = "[Epoch " + str(epoch) + "], [Batch " + str(i) + " / " + str(self.max_epochs) + "], [Loss " + str(loss.item()) + "]"
-        print(log_line)
         self.logs_file.write(log_line + "\n")
 
     self.scheduler.step()
@@ -312,6 +351,9 @@ def compute_batches(self, epoch):
     total_acc = accuracy_score(all_targets, all_preds)
     avg_loss = sum_loss / len(self.data_loader_train)
 
+    self.logs_loss.write(str(epoch) + "," + str(avg_loss.item()) + "\n")
+    self.logs_acc.write(str(epoch) + "," + str(total_acc.item()) + "\n")
+
     return total_acc, avg_loss
 
   """

build/lib/hugsvision/nnet/VisionClassifierTrainer.py

@@ -44,7 +44,8 @@ def __init__(
     batch_size    = 8,
     lr            = 2e-5,
     eval_metric   = "accuracy",
-    fp16          = True,
+    fp16          = False,
+    classification_report_digits = 4,
   ):
 
     self.model_name        = model_name
@@ -61,6 +62,7 @@ def __init__(
     self.eval_metric       = eval_metric
     self.ids2labels        = self.model.config.id2label
     self.labels2ids        = self.model.config.label2id
+    self.classification_report_digits = classification_report_digits
 
     print(self.ids2labels)
     print(self.labels2ids)
@@ -161,6 +163,7 @@ def evaluate_f1_score(self):
       labels = [int(a) for a in list(self.ids2labels.keys())],
       target_names = list(self.labels2ids.keys()),
       zero_division = 0,
+      digits=self.classification_report_digits,
     )
     print(table)
 

build/lib/hugsvision/version.txt

@@ -1 +1 @@
-0.70
+0.72

hugsvision.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: hugsvision
-Version: 0.70
+Version: 0.72
 Summary: A easy to use huggingface wrapper for computer vision.
 Home-page: https://HugsVision.github.io/
 Author: Yanis Labrak & Others
@@ -14,7 +14,7 @@ Description: <p align="center">
         [![GitHub Issues](https://img.shields.io/github/issues/qanastek/HugsVision.svg)](https://github.com/qanastek/HugsVision/issues)
         [![Contributions welcome](https://img.shields.io/badge/contributions-welcome-brightgreen.svg)](CONTRIBUTING.md)
         [![License: MIT](https://img.shields.io/badge/License-MIT-brightgreen.svg)](https://opensource.org/licenses/MIT)
-
+        [![Downloads](https://static.pepy.tech/personalized-badge/hugsvision?period=total&units=international_system&left_color=grey&right_color=orange&left_text=Downloads)](https://pepy.tech/project/hugsvision)
 
         HugsVision is an open-source and easy to use all-in-one huggingface wrapper for computer vision.
 
@@ -161,6 +161,36 @@ Description: <p align="center">
                   </a>
               </td>
           </tr>
+
+          <!-- ------------------------------------------------------------------- -->
+
+          <tr>
+              <td rowspan="3" width="160">
+                <img src="https://raw.githubusercontent.com/qanastek/HugsVision/main/ressources/images/receipes/HAM10000.png" width="256">
+              </td>    
+              <td rowspan="3">  
+                <b>Training and using a TorchVision Image Classifier in 5 min to identify skin cancer:</b> A fast and easy tutorial to train a TorchVision Image Classifier that can help dermatologist in their identification procedures Melanoma cases with HugsVision and HAM10000 dataset.
+              </td>
+              <td align="center" width="80">
+                  <a href="https://nbviewer.jupyter.org/github/qanastek/HugsVision/blob/main/recipes/HAM10000/binary_classification/HAM10000_Image_Classifier.ipynb">
+                      <img src="https://raw.githubusercontent.com/qanastek/HugsVision/main/ressources/images/receipes/nbviewer_logo.svg" height="34">
+                  </a>
+              </td>
+          </tr>
+          <tr>
+              <td align="center">
+                  <a href="https://github.com/qanastek/HugsVision/blob/main/recipes/HAM10000/binary_classification/HAM10000_Image_Classifier.ipynb">
+                      <img src="https://raw.githubusercontent.com/qanastek/HugsVision/main/ressources/images/receipes/github_logo.png" height="32">
+                  </a>
+              </td>
+          </tr>
+          <tr>
+              <td align="center">
+                  <a href="https://colab.research.google.com/drive/1tfRpFTT1GJUgrcwHI0pYdAZ5_z0VSevJ?usp=sharing">
+                      <img src="https://raw.githubusercontent.com/qanastek/HugsVision/main/ressources/images/receipes/colab_logo.png" height="28">
+                  </a>
+              </td>
+          </tr>
         </table>
 
         # Model architectures
@@ -183,6 +213,7 @@ Description: <p align="center">
         Upload: `twine upload dist/*`
 
         # Citation
+
         If you want to cite the tool you can use this:
 
         ```bibtex

hugsvision/inference/TorchVisionClassifierInference.py

@@ -17,13 +17,14 @@ class TorchVisionClassifierInference:
   """
   def __init__(self, model_path: str, transform=transformTorchVision, device=None):
 
+    self.transform = transform
+
     if device == None:
       self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
     else:
       self.device = device
 
     self.model_path = model_path if model_path.endswith("/") else model_path + "/"
-    self.transform = transform
 
     self.model = torch.load(self.model_path + "best_model.pth", map_location=self.device)
 
@@ -42,8 +43,10 @@ def predict_image(self, img, save_preview=False):
       pil_img = transforms.ToPILImage()(img)
       pil_img.save("preview.jpg")
 
+    img = torch.unsqueeze(img, 0).to(self.device)
+
     # Predict and get the corresponding label identifier
-    pred = self.model(torch.unsqueeze(img, 0))
+    pred = self.model(img)
 
     # Get string label from index
     label = self.config["id2label"][str(torch.max(pred, 1)[1].item())]
@@ -53,5 +56,5 @@ def predict_image(self, img, save_preview=False):
   """
   🤔 Predict from one image path
   """
-  def predict(self, img_path: str):  
-    return self.predict_image(img=Image.open(img_path))
+  def predict(self, img_path: str, save_preview=False):  
+    return self.predict_image(img=Image.open(img_path), save_preview=save_preview)