My Fast.ai Lesson 1 of Course 22

A walkthrough of the steps taken to run Fast.ai lesson 1 of Course 22

A walkthrough of my attempt to run the Fast.ai Lesson 1, Is it a bird? Creating a model from your own data from course 22 on Kaggle.

Create a copy of the notebook

The first step was to create a copy of the notebook. A click on the “Copy & Edit” button creates the copy.

Internet setup

The first codeblock checks if you’ve configured your kaggle account to access the internet. This is possible if you have linked a valid phone number. I had already done this step prior at some point before when I attempted the course.

 #NB: Kaggle requires phone verification to use the internet or a GPU. If you haven't done that yet, the cell below will fail
 #    This code is only here to check that your internet is enabled. It doesn't do anything else.
 #    Here's a help thread on getting your phone number verified: https://www.kaggle.com/product-feedback/135367

Upgrade latest version of the libraries

This step upgrades the latest version of the fastai and duckduckgo_search libraries. There’s also a note to indicate that warnings are expected. As suggested I see the following warnings which I ignore.

Broken function to search for images

The first section of code is to download bird pictures. This is done by defining a function search_images. This function fails. At first I follow the note mentioned above the call and try running the cell a couple of times.

#NB: `search_images` depends on duckduckgo.com, which doesn't always return correct responses.
#    If you get a JSON error, just try running it again (it may take a couple of tries).
urls = search_images('bird photos', max_images=1)
urls[0]

I keep getting the following error which leads me to believe the code is broken. The warning states that ddg_images is deprecated.

Searching for 'bird photos'
/opt/conda/lib/python3.7/site-packages/duckduckgo_search/compat.py:60: UserWarning: ddg_images is deprecated. Use DDGS().images() generator
  warnings.warn("ddg_images is deprecated. Use DDGS().images() generator")
/opt/conda/lib/python3.7/site-packages/duckduckgo_search/compat.py:64: UserWarning: parameter page is deprecated
  warnings.warn("parameter page is deprecated")
/opt/conda/lib/python3.7/site-packages/duckduckgo_search/compat.py:66: UserWarning: parameter max_results is deprecated
  warnings.warn("parameter max_results is deprecated")
---------------------------------------------------------------------------
HTTPStatusError                           Traceback (most recent call last)
/tmp/ipykernel_18/2432147335.py in <module>
      1 #NB: `search_images` depends on duckduckgo.com, which doesn't always return correct responses.
      2 #    If you get a JSON error, just try running it again (it may take a couple of tries).
----> 3 urls = search_images('bird photos', max_images=1)
      4 urls[0]

/tmp/ipykernel_18/1717929076.py in search_images(term, max_images)
      4 def search_images(term, max_images=30):
      5     print(f"Searching for '{term}'")
----> 6     return L(ddg_images(term, max_results=max_images)).itemgot('image')

/opt/conda/lib/python3.7/site-packages/duckduckgo_search/compat.py in ddg_images(keywords, region, safesearch, time, size, color, type_image, layout, license_image, max_results, page, output, download)
     80         type_image=type_image,
     81         layout=layout,
---> 82         license_image=license_image,
     83     ):
     84         results.append(r)

/opt/conda/lib/python3.7/site-packages/duckduckgo_search/duckduckgo_search.py in images(self, keywords, region, safesearch, timelimit, size, color, type_image, layout, license_image)
    425         cache = set()
    426         for _ in range(10):
--> 427             resp = self._get_url("GET", "https://duckduckgo.com/i.js", params=payload)
    428             if resp is None:
    429                 break

/opt/conda/lib/python3.7/site-packages/duckduckgo_search/duckduckgo_search.py in _get_url(self, method, url, **kwargs)
     87                 logger.warning(f"_get_url() {url} {type(ex).__name__} {ex}")
     88                 if i >= 2 or "418" in str(ex):
---> 89                     raise ex
     90             sleep(3)
     91         return None

/opt/conda/lib/python3.7/site-packages/duckduckgo_search/duckduckgo_search.py in _get_url(self, method, url, **kwargs)
     81                 if self._is_500_in_url(str(resp.url)) or resp.status_code == 202:
     82                     raise httpx._exceptions.HTTPError("")
---> 83                 resp.raise_for_status()
     84                 if resp.status_code == 200:
     85                     return resp

/opt/conda/lib/python3.7/site-packages/httpx/_models.py in raise_for_status(self)
    747         error_type = error_types.get(status_class, "Invalid status code")
    748         message = message.format(self, error_type=error_type)
--> 749         raise HTTPStatusError(message, request=request, response=self)
    750 
    751     def json(self, **kwargs: typing.Any) -> typing.Any:

HTTPStatusError: Client error '403 Forbidden' for url 'https://duckduckgo.com/i.js?l=wt-wt&o=json&s=0&q=bird%20photos&vqd=4-120798375579229292297400223005060697795&f=%2C%2C%2C%2C%2C&p=1'
For more information check: https://httpstatuses.com/403

Ok, looks like there’s some help in the comments for the original kaggle notebook. This comment solves the issue by installing the fastbook package and redefining the search_images with a an implementation which uses search_images_ddg from the fastbook package. Thank you Kevin Graziani. So adding the following codeblock to install fastbook

if iskaggle:
    !pip install -Uqq fastbook

And redefining search_images with the following codeblock. (I wonder why the from fastai.vision.widgets import * statement is required? )

from fastbook import *
from fastai.vision.widgets import *

def search_images(term, max_images=30):
    print(f"Searching for '{term}'")
    return L(search_images_ddg(term, max_images=max_images))

The search_images_ddg function can be seen in the fastbook/utils.py file on github. There’s also a search_images_bing funtion, which I should try out in my assignment work.

With the issue solved I’m able to download the bird URL.

And download the image

The function download_url is from the fastdownload.core module. It downloads a single URL to a destination path.

For the forest photo search the download_url and search_images function is combined into a single statement followed by the Image.open function:

Using help

The help function can be used to get some documentation and information about the module in which a specific function is defined.

Download Images

After verifying that our search results are reasonable we can download the images for bird and forest. In order to do so the download_images function from the module fastai.vision.utils module is used. Each search is combined with lighting description like sun and shade to get images of birds and forest with different lighting conditions. The search_images has its max_images set to 30 so we get 90 images for bird and forest respectively.

Verify images

The second stage starts with removing images that can’t be opened. To do so we use the verify_images function from the fastai.vision.utils module. In my run 8 images were broken so it gets unlinked. This step is required to make sure our training set is valid.

DataBlocks and DataLoaders

Next we need an object based on DataLoaders which is obtained from the DataBlock object. Think of this as the set of data used for training and the set used for validation. The DataBlock object is defined in the fastai.data.block module. This object is important as it is used to define different types of data sets. The DataBlock is sort of a factory class that can be used to define the main things that change in each model.

The initialization of DataBlock has several parameters:

1. blocks, categorizes the type of inputs and outputs. In this example the inputs are images hence ImageBlock and the outputs are categories hence CategoryBlock.
2. get_items, specifies how the inputs are to be retrieved.
3. splitter, how the data is to be split into training and validation sets. Validation is important to determine the effectiveness of training.
4. get_y, defines how to get the categories for input items. In this example it is the parent directory i.e. bird or forest.
5. item_tfms, specifies what procedure to use to adjust the input data for training. Most computer vision models require the input to be of a fixed size. In this example we’re resizing the images to be 192x192 and if it doesn’t fit we want them to be “squished”.

Traiing of a model typically is done on a GPU. Pytorch uses DataLoaders to feed batches of the data set for training cycles.

To inspect the Dataloader we use the show_batch function. This should display the input data and categories if everything is defined correctly.

More information about construction of a DataBlock is available in the Data Block Tutorial or in the Data Block API docs

Train the model

In order to train the model the vision_learner function is used to create a learner. The resnet18 computer vision model is used which can be trained on a CPU. The fine_tune function is a helper function that uses the best practices to tune the model.

Test the model

The final step is to see if it can predict whether an image has a bird or not.