therealnaveenkamal

from keras.preprocessing.image import ImageDataGenerator

#To standardize the images, we use samplewise_center = True and samplewize_std_normalization = True
traingen = ImageDataGenerator(samplewise_center=True, samplewise_std_normalization= True)
traingenerator = traingen.flow_from_dataframe(
        dataframe=trainset,
        directory="images",
        x_col="Image",
        y_col= labels,
        class_mode="raw",

therealnaveenkamal

positive_freqs = np.mean(traingenerator.labels, axis = 0)
negative_freqs = 1 - positive_freqs
data = {
    'Class': labels,
    'Positive': positive_freqs,
    'Negative':negative_freqs
}

X_axis = np.arange(len(labels))
fig, ax = plt.subplots(figsize=(10, 5))

therealnaveenkamal

data = {
    'Class': labels,
    'Positive': positive_freqs*negative_freqs,
    'Negative':negative_freqs*positive_freqs
}

X_axis = np.arange(len(labels))
fig, ax = plt.subplots(figsize=(10, 5))
ax.bar(X_axis-0.2, data['Positive'], width=0.4, color='b', label = "Positive")
ax.bar(X_axis+0.2, data['Negative'], width=0.4, color='r', label = 'Negative')

therealnaveenkamal

def calcloss(positivewt, negativewt, al=1e-7):
    def weighted_loss(y_true, y_pred):
        loss = 0.0
        for i in range(len(positivewt)):
            loss += -((positivewt[i] * K.transpose(y_true)[i] * K.log(K.transpose(y_pred)[i] + al))+(negativewt[i]*(1 - K.transpose(y_true)[i])*K.log(1 - K.transpose(y_pred)[i] + al)))
        return K.mean(loss)
    return weighted_loss

therealnaveenkamal

model = DenseNet121(weights='densenet.hdf5', include_top=False)
model = Model(inputs=model.input, outputs=Dense(len(labels), activation="sigmoid")(GlobalAveragePooling2D()(model.output)))
model.compile(optimizer='adam', loss=calcloss(negative_freqs, positive_freqs))
fitter = model.fit(traingenerator, validation_data=valgenerator, steps_per_epoch = 1000, epochs = 50)
model.save_weights("cxr_naveen.h5")

plt.plot(fitter.history['loss'])
plt.ylabel("loss")
plt.xlabel("epoch")
plt.title("Training Loss Curve")

therealnaveenkamal

auc_vals = []
for i in range(len(labels)):
    try:
        gt = np.array(testgenerator.labels[:, i])
        pred = predicted_vals[:,i]
        gt = gt.astype('int64')
        auc_vals.append(roc_auc_score(gt, pred))
        fpr_rf, tpr_rf, _ = roc_curve(gt, pred)
        plt.figure(1, figsize=(10, 10))
        plt.plot([0, 1], [0, 1], 'k--')