In a few steps, you can use the mindee python library to perform benchmarks of receipt OCR API perfomances against your test dataset.
- Perform api calls for each file in your test set
mindee.Receiptobjects from your data (csv, json or whatever)
- Perform the benchmark and analyze the results
1. Perform api calls for each file in your test set
Before running the benchmark, we need to collect all the receipt predictions from Mindee API and store them so we can perform the benchmark later without calling the API again.
To do so, assuming that all your files are stored into a "./test_dataset" folder and that we want to save all the responses into a "./mindee_responses" folder, here is the code:
from mindee import Client import os TEST_DATA_PATH = "./test_dataset" RESPONSES_PATH = "./mindee_responses" mindee_client = Client(expense_receipts_token="your_expense_receipts_api_token_here") # Loop over all your test files for test_filename in os.listdir(TEST_DATA_PATH): # Get the current file path test_file_path = os.path.join(TEST_DATA_PATH, test_filename) # To make sure we don't stop the process if an error occurs try: # Parse the current file mindee_response = mindee_client.parse_receipt(test_file_path) # Store the response inside json file to be restored later # In this example we use the test file name in the json filename to # be able to retrieve the corresponding file response_filepath = os.path.join(RESPONSES_PATH, test_filename+".json") mindee_response.dump(response_filepath) except Exception as e: # In case of error, print the filename so you can understand later # what happened print(test_filename, e)
mindee.Receipt objects from your data (csv, json, whatever...)
mindee.Receipt class contains a static
compare() method that takes as inputs two
mindee.Receipt objects. Before running our final script, we need now to create a
mindee.Receipt object containing the true labels for each fields.
We'll use a csv file in this example and the pandas library.
To construct a Receipt object from this dummy csv example, you can simply do:
import pandas as pd from mindee import Receipt ground_truth_df = pd.read_csv("./ground_truth.csv") def receipt_from_csv_row(df_row): taxes_list = df_row["taxes"].split("|") return Receipt( total_incl=df_row["total_incl"], date=df_row["date"], taxes=[(t.split("-"), t.split("-")) for t in taxes_list] ) for index, df_row in ground_truth_df.iterrows(): receipt_truth = receipt_from_csv_row(df_row) print(receipt_truth)
Running this code should print receipt data from your csv file in your console.
3. Perform the benchmark and see the results
Last step, now we need to wrap it all up.
The Benchmark class has two methods
Benchmark.save for adding comparison between two receipt objects, and saving the final metrics.
import pandas as pd from mindee import Response, Receipt, Benchmark import os TEST_DATA_PATH = "./test_dataset" RESPONSES_PATH = "./mindee_responses" BENCHMARK_PATH = "./benchmark" ground_truth_df = pd.read_csv("./ground_truth.csv") benchmark = Benchmark(BENCHMARK_PATH) def receipt_from_csv_row(df_row): # Your method for creating the Receipt ground truth pass # Loop over each file in our csv for index, df_row in ground_truth_df.iterrows(): try: # Get test file path test_file_path = os.path.join(TEST_DATA_PATH, df_row["filename"]) # Create ground truth receipt object ground_truth_receipt = receipt_from_csv_row(df_row) # Load the mindee Response for the current file mindee_response = Response.load(os.path.join(RESPONSES_PATH, df_row["filename"] + ".json")) # Add the comparison between the two receipts to the benchmark benchmark.add( Receipt.compare(mindee_response.receipt, ground_truth=ground_truth_receipt), df_row["filename"] ) except Exception as e: print(df_row["filename"], e) benchmark.save()
Inside our benchmark folder, you should see a new directory was created, and a metrics.png file shoud have been created inside with the different metrics:
The receipt benchmark runs on the 7 fields as shown above. For each of them, you get an information of:
accuracy: the proportion of correct predictions
precision: The proportion of correct predictions among all the non null predictions