Fix #955: Avoid TypeError in read_file

[Bond099]

I have added "all(isinstance(x, str) for x in df['geometry'])" which checks all elements in the geometry column are strings ,before applying gpd.GeoSeries.from_wkt.

[henrykironde]

Is there a way we can solve this issue with out specifically looping through the data frame?

[Bond099]

Hey @henrykironde, we have a couple of options :
1)if isinstance(df['geometry'].iloc[0], str): This only checks the first element in the column.
2)if pd.api.types.infer_dtype(df['geometry'])== 'string': This can analyse the whole column without looping.
If the column has mixed types (e.g: some strings, some 'polygon' objects) the first option could fail, while the second one is more robust, which one do you think we should go with?

[henrykironde]

Use the option you feel works best and less complexity than looping

[ethanwhite]

@Bond099 - what was the situation where you ended up with non-string data in the geometry column?

Sorry - I hadn't seen the initial issue.

[bw4sz]

@Bond099 can you show me the motivating example and confirm it no longer errors from the issue.

import pandas as pd
from deepforest import utilities
from deepforest import get_data 
import os

# Create a sample dataframe with annotations
df = pd.DataFrame({
    'xmin': [100],
    'ymin': [100], 
    'xmax': [200],
    'ymax': [200],
    'label': ['Tree'],
    'image_path': ['OSBS_029.tif']
})

root_dir = os.path.dirname(get_data("OSBS_029.tif"))
# First call to read_file works fine
df = utilities.read_file(df, root_dir=root_dir)

# Second call to read_file raises error if it's a dataframe
df = utilities.read_file(pd.DataFrame(df), root_dir=root_dir)

[Bond099]

Hey @bw4sz,

When you call read_file for the first time, it works perfectly because there’s no 'geometry' column in your DataFrame yet. The function takes your coordinates (like xmin, ymin, etc.), creates polygon shapes from them, and turns your DataFrame into a GeoDataFrame. That’s why the initial call succeeds without any issues.

But on the second call, it crashes with a TypeError. This happens because, by then, the 'geometry' column already exists and contains polygon objects (from the first call). The original function tries to process this column using gpd.GeoSeries.from_wkt, which expects text strings in WKT format—like "POLYGON ((0 0, 1 0, 1 1, 0 1, 0 0))". Since the column already has polygon objects instead of strings, it fails.

fix: I’ve added a simple check: pd.api.types.infer_dtype(df['geometry']) == 'string'. This looks at the entire 'geometry' column to determine its data type:

-If it’s strings (WKT format), the function converts them into polygon objects.
-If it’s already polygon objects, the function leaves them as is.

Hope this clears things up! Let me know if you have any questions or need further clarification

[bw4sz]

Great, just for reproducibility, please show that code passing.

[Bond099]

@bw4sz I have attached the screenshot of the code and its output .

[henrykironde]

rm changes debug_predictions.datagrid