hwberry2/TensorFlowProject

1import gradio as gr2import tensorflow as tf3import numpy as np4import os5import PIL6import PIL.Image7 8# Create a Gradio App using Blocks    9with gr.Blocks() as demo:10    gr.Markdown(11    """12    # AI/ML Playground13    """14    )15    with gr.Accordion("Click for Instructions:"):16            gr.Markdown(17    """18    * uploading an image will engage the model in image classsification19    * trained on the following image types: 'T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat', 'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot'20    * Only accepts images 28x28. Trained on images with a black background.21    """)22 23    # Train, evaluate and test a ML24    # image classification model for25    # clothes images26 27    class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',28           'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']29    30    # clothing dataset31    mnist = tf.keras.datasets.fashion_mnist32 33    #split the training data in to a train/test sets34    (x_train, y_train), (x_test, y_test) = mnist.load_data()35    x_train, x_test = x_train / 255.0, x_test / 255.036 37    # create the neural net layers38    model = tf.keras.models.Sequential([39      tf.keras.layers.Flatten(input_shape=(28, 28)),40      tf.keras.layers.Dense(128, activation='relu'),41      tf.keras.layers.Dropout(0.2),42      tf.keras.layers.Dense(10)43    ])44 45    #make a post-training predition on the 46    #training set data47    predictions = model(x_train[:1]).numpy()48 49    # converts the logits into a probability50    tf.nn.softmax(predictions).numpy()51 52    #create and train the loss function53    loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)54    loss_fn(y_train[:1], predictions).numpy()55 56    # compile the model with the loss function57    model.compile(optimizer='adam',58                  loss=loss_fn,59                  metrics=['accuracy'])60    61    # train the model - 5 runs62    # evaluate the model on the test set63    model.fit(x_train, y_train, epochs=5, validation_split=0.3)64    test_loss, test_acc = model.evaluate(x_test,  y_test, verbose=2)65    post_train_results = f"Test accuracy: {test_acc} Test Loss: {test_loss}"66    print(post_train_results)67 68    # create the final model for production69    probability_model = tf.keras.Sequential([model, tf.keras.layers.Softmax()])70 71        72    def classifyImage(img):     73        # Normalize the pixel values74        img = np.array(img) / 255.075        76        input_array = np.expand_dims(img, axis=0) # add an extra dimension to represent the batch size77 78        # Make a prediction using the model79        prediction = probability_model.predict(input_array)80 81        # Postprocess the prediction and return it82        predicted_label = class_names[np.argmax(prediction)]83 84        return predicted_label85        86    def do_nothing():87        pass88        89    # Creates the Gradio interface objects90    with gr.Row():91        with gr.Column(scale=2):92            image_data = gr.Image(label="Upload Image", type="numpy", image_mode="L")93        with gr.Column(scale=1):94            model_prediction = gr.Text(label="Model Prediction", interactive=False)95        image_data.upload(classifyImage, image_data, model_prediction)96        image_data.clear(do_nothing, [], model_prediction)97    98    99# creates a local web s100# if share=True creates a public101# demo on huggingface.c102demo.launch(share=False)