This code manually splits up the dataset into train, validation and test with 20:4:1 as their respective ratios.

cxr_multiclassclassification_datasetsplit.py

dataframe = pd.read_csv("Data_Entry_2017_v2020.csv")

#Enumerating all column names
columns = ["Image"]
for i in dataframe["Finding Labels"].values:
    for j in i.split("|"):
        if j not in columns:
            columns.append(j)
labels = columns.copy()
labels.remove("Image")

#Taking the first 10000 images from the master table as the train dataset
trainset = pd.DataFrame(columns = columns)
for i in range(10000):
    col = [0]*len(columns)
    col[0] = dataframe["Image Index"][i]
    count = 1
    for j in columns[1:]:
        if(j in dataframe["Finding Labels"][i]):
            col[count] = 1
        count+=1
    trainset.loc[len(trainset)] = col
    
#Taking the next 2000 images from the master table as the validation dataset
valset = pd.DataFrame(columns = columns)
for i in range(10000, 12000):
    col = [0]*len(columns)
    col[0] = dataframe["Image Index"][i]
    count = 1
    for j in columns[1:]:
        if(j in dataframe["Finding Labels"][i]):
            col[count] = 1
        count+=1
    valset.loc[len(valset)] = col

#Taking the next 500 images from the master table as the test dataset
testset = pd.DataFrame(columns = columns)
for i in range(15000, 15500):
    col = [0]*len(columns)
    col[0] = dataframe["Image Index"][i]
    count = 1
    for j in columns[1:]:
        if(j in dataframe["Finding Labels"][i]):
            col[count] = 1
        count+=1
    testset.loc[len(testset)] = col
   
#Plotting first 16 images with their disease labels  
img_dir = "images"
plt.figure(figsize = (15,15))
for i in range(16):
    plt.subplot(4, 4, i+1)
    plt.imshow(plt.imread(os.path.join(img_dir, trainset["Image"][i])), cmap = "gray")
    plt.title(dataframe[dataframe["Image Index"] == trainset["Image"][i]].values[0][1])
plt.tight_layout()