Loss: {loss:0.4f} {value} / {length}
\n \n \"\"\"\n \n def __init__(self, length):\n self.length = length\n self.count = 0\n self.display = display(self.html(0, 0), display_id=True)\n \n def html(self, count, loss):\n return HTML(self.tmpl.format(length=self.length, value=count, loss=loss))\n \n def update(self, count, loss):\n self.count += count\n self.display.update(self.html(self.count, loss))\n\ndef checkpoint_and_save(model, best_loss, epoch, optimizer, epoch_valid_loss, model_save_name):\n print('saving')\n print()\n state = {'model': model,'best_loss': best_loss,'epoch': epoch,'rng_state': torch.get_rng_state(), 'optimizer': optimizer.state_dict(),}\n torch.save(model, './{}-CL.pt'.format(model_save_name))\n torch.save(model.state_dict(),'./{}-statedict-CL.pt'.format(model_save_name))\n \ndef train_new(model,criterion,optimizer,num_epochs,dataloaders,dataset_sizes, model_save_name,first_epoch=1):\n since = time.time() \n best_loss = 999999\n best_epoch = -1\n last_train_loss = -1\n plot_train_loss = []\n plot_valid_loss = []\n plot_train_acc = []\n plot_valid_acc = []\n \n \n for epoch in range(first_epoch, first_epoch + num_epochs):\n print()\n print('Epoch', epoch)\n running_loss = 0.0\n valid_loss = 0.0\n training_accuracy = 0\n validation_accuracy = 0\n \n # train phase\n model.train(True)\n \n # create a progress bar\n progress = ProgressMonitor(length=dataset_sizes[\"train\"])\n \n for data in dataloaders[\"train\"]:\n inputs, labels = data # (Batch_size, width, height, channels)\n batch_size = inputs.shape[0]\n inputs = inputs.permute(0,3,1,2) # Batch_size, channels, width, height\n inputs = inputs.to(device)\n labels = labels.to(device,dtype=torch.long)\n inputs = Variable(inputs)\n labels = Variable(labels)\n optimizer.zero_grad() \n outputs = model(inputs) # batch, 2, 240, 240\n\n loss = criterion(outputs, labels) # comparing outputs with the actual labels\n \n loss.backward() \n optimizer.step() \n \n running_loss += loss.data * batch_size\n training_accuracy += (outputs.argmax(1) == labels).sum().item() # if(0 ==0, or 1 == 1)\n # update progress bar\n progress.update(batch_size, running_loss)\n \n epoch_loss = running_loss / dataset_sizes[\"train\"]\n \n print('Training Accuracy is {:.2f} and Training loss {:.2f}'.format(training_accuracy / dataset_sizes[\"train\"],epoch_loss.item()))\n print('Correctly predicted {} Training samples out of {}'.format(training_accuracy, dataset_sizes[\"train\"]))\n print()\n plot_train_loss.append(epoch_loss)\n plot_train_acc.append(training_accuracy / dataset_sizes[\"train\"])\n \n \n # validation phase\n model.eval()\n # no_grad to save memory\n with torch.no_grad():\n for data in dataloaders[\"valid\"]:\n inputs, labels = data\n batch_size = inputs.shape[0]\n inputs = inputs.permute(0,3,1,2)\n inputs = inputs.to(device)\n labels = labels.to(device,dtype=torch.long)\n inputs = Variable(inputs)\n labels = Variable(labels)\n \n outputs = model(inputs)\n \n # calculate the loss\n optimizer.zero_grad()\n loss = criterion(outputs, labels)\n \n # update running loss value\n valid_loss += loss.data * batch_size\n validation_accuracy += (outputs.argmax(1) == labels).sum().item()\n \n \n epoch_valid_loss = valid_loss / dataset_sizes[\"valid\"]\n scheduler.step(epoch_valid_loss)\n print('Validation Accuracy is {:.2f} and Validation loss {:.2f}'.format(validation_accuracy / dataset_sizes[\"valid\"],epoch_valid_loss.item()))\n print('Correctly predicted {} Validation samples out of {}'.format(validation_accuracy, dataset_sizes[\"valid\"]))\n print()\n plot_valid_loss.append(epoch_valid_loss)\n plot_valid_acc.append(validation_accuracy / dataset_sizes[\"valid\"])\n \n if epoch_valid_loss < best_loss:\n best_loss = epoch_valid_loss\n best_epoch = epoch\n checkpoint_and_save(model, best_loss, epoch, optimizer, epoch_valid_loss.item(), model_save_name)\n\n if ((epoch - best_epoch) >= 10):\n print(\"no improvement in 10 epochs, break\")\n break\n \n time_elapsed = time.time() - since\n print('Training complete in {:.0f}m {:.0f}s'.format(time_elapsed // 60, time_elapsed % 60))\n \n return plot_train_loss, plot_valid_loss, plot_train_acc, plot_valid_acc, model","execution_count":null,"outputs":[]},{"metadata":{"id":"cffFjW0Y0Ab2"},"cell_type":"markdown","source":"## Training Loop"},{"metadata":{"id":"FypxE1230A7N","outputId":"ce04abd8-924a-4ac6-96a3-c5afdcb0dc02","trusted":true},"cell_type":"code","source":"dataset_sizes = {x: len(transformed_datasets[x]) for x in ['train', 'valid']}\nepochs = 20\nif __name__==\"__main__\":\n train_losses, valid_losses, train_accuracy, valid_accuracy, model = train_new(model = model_ft ,criterion = criterion,optimizer = optimizer,\n num_epochs=epochs,dataloaders = dataloaders,\n dataset_sizes = dataset_sizes, model_save_name = model_name)","execution_count":null,"outputs":[]},{"metadata":{"id":"haDxg3o10FkF"},"cell_type":"markdown","source":"# Model's Metrics ~ Accuracy, F1 Score, Confusion Matrix"},{"metadata":{"id":"3forXqii0C2R","outputId":"dbdf12ec-16c4-4a95-c2ae-117723d2d0ff","trusted":true},"cell_type":"code","source":"# Plot Accuracy\nplt.title('Training and Validation Accuracy')\nplt.plot(train_accuracy)\nplt.plot(valid_accuracy)\nplt.legend(['Training_Accuracy','Validation_Accuracy'])","execution_count":null,"outputs":[]},{"metadata":{"id":"RKSPHl710HPQ","outputId":"2d46b18d-b182-4d50-e508-4192ea64acb0","trusted":true},"cell_type":"code","source":"plt.title('Training and Validation Loss')\nplt.plot(train_losses)\nplt.plot(valid_losses)\nplt.legend(['Training_loss','Validation_loss'])","execution_count":null,"outputs":[]},{"metadata":{"id":"AvQq_DWy0IoB","outputId":"b52a039c-9df2-41d2-b23f-4e93dc7e0eac","trusted":true},"cell_type":"code","source":"plt.figure(figsize=(12,8))\nplt.title('Training & Validation Accuracy and Loss')\nplt.plot(train_accuracy)\nplt.plot(valid_accuracy)\nplt.plot(train_losses)\nplt.plot(valid_losses)\nplt.legend(['Training_Accuracy','Validation_Accuracy','Training_loss','Validation_loss'])","execution_count":null,"outputs":[]},{"metadata":{"id":"pmUta1dX0Lo7"},"cell_type":"markdown","source":"## F1 - Score, Confusion Matrix"},{"metadata":{"id":"1MLkzcFz0MAs","outputId":"5b35692d-e62c-401f-8e03-225f08b5336b","trusted":true},"cell_type":"code","source":"from sklearn.metrics import f1_score\nwith torch.no_grad():\n complete_outputs = []\n complete_labels = []\n for data in dataloaders[\"valid\"]:\n inputs, labels = data\n inputs = inputs.permute(0,3,1,2)\n inputs = inputs.to(device)\n labels = labels.to(device,dtype=torch.long)\n inputs = Variable(inputs)\n labels = Variable(labels)\n outputs = model(inputs)\n complete_outputs.extend(list(outputs.argmax(1).cpu().data.numpy()))\n complete_labels.extend(list(labels.cpu().data.numpy()))\n\nassert len(complete_outputs) == len(complete_labels)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"from sklearn.metrics import confusion_matrix\nimport seaborn as sns\n\ncf_matrix = confusion_matrix(complete_outputs, complete_labels)\nsns.heatmap(cf_matrix, annot=True, fmt=\"d\", cmap=\"YlGnBu\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"from sklearn.metrics import classification_report\ntarget_names = ['Class_1', 'Class_2', 'Class_3', 'Class_4', 'Class_5']\nprint(classification_report(complete_outputs, complete_labels, target_names=target_names))","execution_count":null,"outputs":[]},{"metadata":{"id":"W7bNm0j10PEe"},"cell_type":"markdown","source":"# Inference"},{"metadata":{"id":"iT8fcSKH0PZu","outputId":"fa50e2ac-55da-4cfb-b291-b477289ae2e5","trusted":true},"cell_type":"code","source":"def tester(image, model):\n img = (image).to(device)\n img = img.unsqueeze(0)\n img = img.permute(0,3,1,2) # (bs, width, height, channels) --> (bs, channels, width, height)\n output = model(img)\n return output.argmax(1).item()\n\nNUM_SAMP=10\nfig = plt.figure(figsize=(25, 16))\ncount = 0\nfor jj in range(5):\n for i, (idx, row) in enumerate(df_valid.sample(NUM_SAMP,random_state=123+jj).iterrows()):\n ax = fig.add_subplot(5, NUM_SAMP, jj * NUM_SAMP + i + 1, xticks=[], yticks=[])\n path= row['file_path']\n orig_label = int(row['label'] )\n image = image_preprocess(path, img_size=input_size) #224,224,3\n pred_label = tester(image, model)\n if (orig_label == pred_label):\n count +=1\n plt.imshow(image)\n ax.set_title('%d - %d' % (orig_label, pred_label))\nprint()\nprint(\"Out of {} samples, model predicted {} samples correctly\".format((NUM_SAMP*5), count))\nprint()\nprint(\"Legends : Original Label Vs Predicted Label\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def Load_Models_Dynamic():\n num_gpus = torch.cuda.device_count()\n device, model = {}, {}\n models = []\n device = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\n model = torch.load(\"./efficientnet-b0-TS.pt\", map_location=device).eval()\n models.append(model)\n return models\n\nmodels = Load_Models_Dynamic()\nmodels","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"import os\ntry:\n os.mkdir(\"./deps\")\nexcept FileExistsError:\n pass\n!pip install efficientnet_pytorch -t \"./deps\" --no-deps","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"import shutil\nshutil.make_archive(\"./deps\",'zip','./')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]}],"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat":4,"nbformat_minor":4}