from tensorflow.keras import Model from tensorflow.keras.layers import ( concatenate, Dense, MaxPooling2D, Flatten, Activation, Dropout, ) from .soft_attention import SoftAttention from .inception_v2 import get_inception_resnet_v2_layers import numpy as np from io import BytesIO, StringIO import matplotlib.pyplot as plt import cv2 from PIL import Image def create_model(): irv2, conv = get_inception_resnet_v2_layers() attention_layer, _ = SoftAttention( aggregate=True, m=16, concat_with_x=False, ch=int(conv.shape[-1]), name="soft_attention", )(conv) attention_layer = MaxPooling2D(pool_size=(2, 2), padding="same")(attention_layer) conv = MaxPooling2D(pool_size=(2, 2), padding="same")(conv) conv = concatenate([conv, attention_layer]) conv = Activation("relu")(conv) conv = Dropout(0.5)(conv) output = Flatten()(conv) output = Dense(7, activation="softmax")(output) model = Model(inputs=irv2.input, outputs=output) model.load_weights("./data/IRV2+SA.weights.h5") return model def get_friendly_output(prediction): idx = np.argmax(prediction, axis=1)[0] targetnames = ["akiec", "bcc", "bkl", "df", "mel", "nv", "vasc"] map_label = { "akiec": "Ceratoses actínicas e carcinoma intraepitelial / Doença de Bowen", "bcc": "Carcinoma basocelular", "bkl": "Lesões benignas semelhantes a ceratose / lentigos solares / ceratose seborreica e ceratose semelhante a líquen plano", "df": "Dermatofibroma - lesões benignas típicas de extremidades, como braços, antebraços e pernas.", "mel": "Melanoma - O tipo mais grave de câncer de pele.", "nv": "Nevos melanocíticos - Distúrbio geralmente não cancerígeno das células da pele produtoras de pigmento, comumente chamado de marcas de nascença ou pintas.", "vasc": "Lesões vasculares / angiomas / angioqueratomas / granulomas piogênicos e hemorragia", } return map_label[targetnames[idx]] def get_attention_map(image, model): sa_model = Model(model.inputs, model.get_layer("soft_attention").output) _, sa_maps = sa_model.predict(image) sum_attnmap = np.sum(sa_maps[0], 0) buffer_ = BytesIO() plt.imshow((image[0] + 1) / 2, alpha=1.0, aspect="auto") plt.imshow( cv2.resize(sum_attnmap, (299, 299), interpolation=cv2.INTER_CUBIC), cmap="jet", alpha=0.5, ) plt.axis('off') plt.show() plt.savefig(buffer_, format = "png", bbox_inches='tight', pad_inches=0) buffer_.seek(0) return buffer_