test.py

# Disable tensorflow debugging output
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' 

import sys
from helpers import preprocess, plot
import numpy as np
import requests
from mlserver.types import InferenceRequest
from mlserver.codecs import NumpyCodec
import tensorflow_datasets as tfds
import tensorflow as tf


# Inference variables
if len(sys.argv) < 2:
    sys.exit("Please provide inference mode (--local or --remote)")
if sys.argv[1] == "--local":
    inference_url = 'http://localhost:8080/v2/models/cassava/infer'
elif sys.argv[1] == "--remote":
    inference_url = 'http://localhost:8080/seldon/default/cassava/v2/models/infer'
else:
    sys.exit("Please provide inference mode (--local or --remote)")
batch_size = 16

# Load the dataset and class names
print("Lodaing dataset...")
dataset, info = tfds.load('cassava', with_info=True)
class_names = info.features['label'].names + ['unknown']

# Shuffle the dataset with a buffer size equal to the number of examples in the 'validation' split
validation_dataset = dataset['validation']
buffer_size = info.splits['validation'].num_examples
shuffled_validation_dataset = validation_dataset.shuffle(buffer_size)

# Select a batch of examples from the validation dataset
batch = shuffled_validation_dataset.map(preprocess).batch(batch_size).as_numpy_iterator()
examples = next(batch)

# Convert the TensorFlow tensor to a numpy array
input_data = np.array(examples['image'])

# Build the inference request
inference_request = InferenceRequest(
    inputs=[
        NumpyCodec.encode_input(name="payload", payload=input_data)
    ]
)

# Send the inference request and capture response
print("Sending Inference Request...")
res = requests.post(inference_url, json=inference_request.dict())
print("Got Response...")

# Parse the JSON string into a Python dictionary
response_dict = res.json()

# Extract the data array and shape from the output, assuming only one output or the target output is at index 0
data_list = response_dict["outputs"][0]["data"]
data_shape = response_dict["outputs"][0]["shape"]

# Convert the data list to a numpy array and reshape it
data_array = np.array(data_list).reshape(data_shape)
print("Predictions:", data_array)

# Convert the numpy array to tf tensor
data_tensor = tf.convert_to_tensor(np.squeeze(data_array), dtype=tf.int64)

# Plot the examples with their predictions
plot(examples, class_names, data_tensor)