Merge pull request #1066 from pritesh2000/gram-1/06

06_pytorch_transfer_learning.ipynb

06_pytorch_transfer_learning.ipynb

@@ -68,11 +68,11 @@
     "\n",
     "Both research and practice support the use of transfer learning too.\n",
     "\n",
-    "A finding from a recent machine learning research paper recommended practioner's use transfer learning wherever possible.\n",
+    "A finding from a recent machine learning research paper recommended practitioners use transfer learning wherever possible.\n",
     "\n",
     "<img src=\"https://raw.githubusercontent.com/mrdbourke/pytorch-deep-learning/main/images/06-how-to-train-your-vit-section-6-transfer-learning-highlight.png\" width=900 alt=\"how to train your vision transformer paper section 6, advising to use transfer learning if you can\"/>\n",
     "\n",
-    "*A study into the effects of whether training from scratch or using transfer learning was better from a practioner's point of view, found transfer learning to be far more beneficial in terms of cost and time. **Source:** [How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers](https://arxiv.org/abs/2106.10270) paper section 6 (conclusion).*\n",
+    "*A study into the effects of whether training from scratch or using transfer learning was better from a practitioner's point of view, found transfer learning to be far more beneficial in terms of cost and time. **Source:** [How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers](https://arxiv.org/abs/2106.10270) paper section 6 (conclusion).*\n",
     "\n",
     "And Jeremy Howard (founder of [fastai](https://www.fast.ai/)) is a big proponent of transfer learning.\n",
     "\n",
@@ -439,11 +439,11 @@
     "\n",
     "We can create these using the `create_dataloaders` function from the [`data_setup.py`](https://github.com/mrdbourke/pytorch-deep-learning/blob/main/going_modular/going_modular/data_setup.py) script we created in [05. PyTorch Going Modular Part 2](https://www.learnpytorch.io/05_pytorch_going_modular/#2-create-datasets-and-dataloaders-data_setuppy).\n",
     "\n",
-    "We'll set `batch_size=32` so our model see's mini-batches of 32 samples at a time.\n",
+    "We'll set `batch_size=32` so our model sees mini-batches of 32 samples at a time.\n",
     "\n",
     "And we can transform our images using the transform pipeline we created above by setting `transform=manual_transforms`.\n",
     "\n",
-    "> **Note:** I've included this manual creation of transforms in this notebook because you may come across resources that use this style. It's also important to note that because these transforms are manually created, they're also infinitely customizable. So if you wanted to included data augmentation techniques in your transforms pipeline, you could."
+    "> **Note:** I've included this manual creation of transforms in this notebook because you may come across resources that use this style. It's also important to note that because these transforms are manually created, they're also infinitely customizable. So if you wanted to include data augmentation techniques in your transforms pipeline, you could."
    ]
   },
   {
@@ -622,7 +622,7 @@
     "\n",
     "That's where **transfer learning** comes in.\n",
     "\n",
-    "The whole idea of transfer learning is to **take an already well-performing model on a problem-space similar to yours and then customising it to your use case**.\n",
+    "The whole idea of transfer learning is to **take an already well-performing model on a problem-space similar to yours and then customise it to your use case**.\n",
     "\n",
     "Since we're working on a computer vision problem (image classification with FoodVision Mini), we can find pretrained classification models in [`torchvision.models`](https://pytorch.org/vision/stable/models.html#classification).\n",
     "\n",
@@ -1056,9 +1056,9 @@
     "Ho, ho! There's a fair few changes here!\n",
     "\n",
     "Let's go through them:\n",
-    "* **Trainable column** - You'll see that many of the base layers (the ones in the `features` portion) have their Trainable value as `False`. This is because we set their attribute `requires_grad=False`. Unless we change this, these layers won't be updated during furture training.\n",
+    "* **Trainable column** - You'll see that many of the base layers (the ones in the `features` portion) have their Trainable value as `False`. This is because we set their attribute `requires_grad=False`. Unless we change this, these layers won't be updated during future training.\n",
     "* **Output shape of `classifier`** - The `classifier` portion of the model now has an Output Shape value of `[32, 3]` instead of `[32, 1000]`. It's Trainable value is also `True`. This means its parameters will be updated during training. In essence, we're using the `features` portion to feed our `classifier` portion a base representation of an image and then our `classifier` layer is going to learn how to base representation aligns with our problem.\n",
-    "* **Less trainable parameters** - Previously there was 5,288,548 trainable parameters. But since we froze many of the layers of the model and only left the `classifier` as trainable, there's now only 3,843 trainable parameters (even less than our TinyVGG model). Though there's also 4,007,548 non-trainable parameters, these will create a base representation of our input images to feed into our `classifier` layer.\n",
+    "* **Less trainable parameters** - Previously there were 5,288,548 trainable parameters. But since we froze many of the layers of the model and only left the `classifier` as trainable, there's now only 3,843 trainable parameters (even less than our TinyVGG model). Though there's also 4,007,548 non-trainable parameters, these will create a base representation of our input images to feed into our `classifier` layer.\n",
     "\n",
     "> **Note:** The more trainable parameters a model has, the more compute power/longer it takes to train. Freezing the base layers of our model and leaving it with less trainable parameters means our model should train quite quickly. This is one huge benefit of transfer learning, taking the already learned parameters of a model trained on a problem similar to yours and only tweaking the outputs slightly to suit your problem."
    ]
@@ -1562,7 +1562,7 @@
     "## Extra-curriculum\n",
     "* Look up what \"model fine-tuning\" is and spend 30-minutes researching different methods to perform it with PyTorch. How would we change our code to fine-tune? Tip: fine-tuning usually works best if you have *lots* of custom data, where as, feature extraction is typically better if you have less custom data.\n",
     "* Check out the new/upcoming [PyTorch multi-weights API](https://pytorch.org/blog/introducing-torchvision-new-multi-weight-support-api/) (still in beta at time of writing, May 2022), it's a new way to perform transfer learning in PyTorch. What changes to our code would need to be made to use the new API?\n",
-    "* Try to create your own classifier on two classes of images, for example, you could collect 10 photos of your dog and your friends dog and train a model to classify the two dogs. This would be a good way to practice creating a dataset as well as building a model on that dataset."
+    "* Try to create your own classifier on two classes of images, for example, you could collect 10 photos of your dog and your friend's dog and train a model to classify the two dogs. This would be a good way to practice creating a dataset as well as building a model on that dataset."
    ]
   }
  ],