[WIP] Analyse false predictions

analyse_false.py

@@ -0,0 +1,112 @@
+import argparse
+from keras.models import load_model
+from autowebcompat.network import contrastive_loss
+import tkinter as tk
+from PIL import ImageTk, Image
+import numpy as np
+
+from autowebcompat import utils
+
+parser = argparse.ArgumentParser()
+parser.add_argument('-c', '--classification_type', type=str, choices=utils.CLASSIFICATION_TYPES, default=utils.CLASSIFICATION_TYPES[0], help='Select the classification_type for training')
+
+args = parser.parse_args()
+
+bugs = utils.get_bugs()
+
+BATCH_SIZE = 32
+EPOCHS = 50
+
+labels = utils.read_labels()
+
+utils.prepare_images()
+all_image_names = [i for i in utils.get_images() if i in labels]
+all_images = sum([[i + '_firefox.png', i + '_chrome.png'] for i in all_image_names], [])
+image = utils.load_image(all_images[0])
+input_shape = image.shape
+
+def load_pair(fname):
+    return [fname + '_firefox.png', fname + '_chrome.png']
+
+def couples_generator(images):
+    for i in images:
+        yield load_pair(i), utils.to_categorical_label(labels[i], args.classification_type)
+
+def gen_func(images, coupleIterator):
+    index = 0
+    for x_batch, y_batch in coupleIterator:
+        yield images[index: index + BATCH_SIZE], x_batch, y_batch
+        index += BATCH_SIZE
+
+number_of_couples = sum(1 for e in couples_generator(all_image_names))
+
+print(input_shape)
+
+data_gen = utils.get_ImageDataGenerator(all_images, input_shape)
+all_iterator = gen_func(all_image_names, utils.CouplesIterator(couples_generator(all_image_names), input_shape, data_gen, BATCH_SIZE))
+
+model = load_model('best_train_model.hdf5', custom_objects={'contrastive_loss': contrastive_loss})
+
+images_for_analysis = []
+for image_names, x_batch, y_batch in all_iterator:
+    predictions = model.predict(x_batch)
+
+    for batch_index in range(BATCH_SIZE):
+        truth = y_batch[batch_index]
+        prediction = np.round(predictions[batch_index])
+        image_name = image_names[batch_index]
+
+        if truth != np.round(prediction):
+            images_for_analysis.append([ truth, prediction, image_name ])
+
+# Create UI
+window = tk.Tk()
+
+image_view = tk.Frame(window)
+
+image_view.pack(side="top")
+
+panel1 = tk.Label(image_view)
+panel1.grid(row=0, column=0)
+
+panel1_text = tk.Label(image_view)
+panel1_text.grid(row=1, column=0)
+
+panel2 = tk.Label(image_view)
+panel2.grid(row=0, column=1)
+
+panel2_text = tk.Label(image_view)
+panel2_text.grid(row=1, column=1)
+
+prediction_text = tk.Label(image_view)
+prediction_text.grid(row=2)
+
+image_index = 0
+
+# Fill with image
+def show_next_image():
+    global panel1, panel2, panel1_text, panel2_text, image_index
+
+    truth, prediction, image_name = images_for_analysis[image_index]
+
+    firefox_name =  image_name + "_firefox.png"
+    chrome_name = image_name + "_chrome.png"
+
+    img1 = ImageTk.PhotoImage(Image.open("data/" + firefox_name))
+    panel1.configure(image=img1)
+    panel1.image = img1
+    panel1_text.config(text=firefox_name)
+
+    img2 = ImageTk.PhotoImage(Image.open("data/" + chrome_name))
+    panel2.configure(image=img2)
+    panel2.image = img2
+    panel2_text.config(text=chrome_name)
+
+    prediction_text.config(text="Truth: {}, Prediction: {}".format(truth, prediction))
+
+    image_index += 1
+
+next_button = tk.Button(window, text="Next", command=show_next_image)
+next_button.pack(side="bottom", ipadx=30, ipady=10, pady=20)
+
+window.mainloop()