differences = [abs(extracted_time - each_date) for each_date in date_list_converted]
minimum = min(differences)
closest_date_index = differences.index(minimum)
# Find index of relevant entry
relevant_index = patient_series.index[closest_date_index]
relevant_indices.append(relevant_index)
# Update dataframe: Only if mismatch is smaller! Otherwise might overwrite with worse image!
if minimum.days < original_labels.at[relevant_index, 'Time_Mismatch']:
original_labels.at[relevant_index, 'Filename'] = os.path.basename(img_name)
original_labels.at[relevant_index, 'Time_Mismatch'] = minimum.days
# else:
# original_labels.at[relevant_index, 'Filename'] = img_name
# original_labels.at[relevant_index, 'Time_Mismatch'] = minimum.days
print(relevant_index)
counts += 1
# Plot histogram of time mismatches
original_labels['Time_Mismatch'] = original_labels['Time_Mismatch'].astype(float)
original_labels['Time_Mismatch'].plot.hist(by='Time_Mismatch', bins=40)
plt.show()
# Value counts
viable_entries = original_labels.Time_Mismatch.value_counts()[0] + original_labels.Time_Mismatch.value_counts()[1]
print(f'Number of viable entries is {viable_entries}')
# Make invalid mismatches nulls
original_labels.loc[original_labels.Time_Mismatch == 500, 'Time_Mismatch'] = np.nan
original_labels.loc[original_labels.Filename == "", 'Filename'] = np.nan
original_labels.to_csv(os.path.join('/data/COVID/Labels', 'new_gstt.csv'), index=False)
# Better approach would find combination of dates that minimises overall difference
def train():
"""
function that trains the generator and discriminator
loss function based on (Ledig 2017)
optimzer - adam
"""
gz, gvl = generator(blur)
r_out, dvl = discriminator(real)
f_out, dvl = discriminator(gz)
with tf.name_scope("cost"):
fake_dloss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels = tf.zeros_like(f_out),logits = f_out))
real_dloss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels = tf.ones_like(r_out),logits = r_out))
tf.summary.scalar("fake_dloss",fake_dloss)
tf.summary.scalar("real_dloss",real_dloss)
dloss = fake_dloss + real_dloss
gloss = tf.reduce_mean(tf.math.square(real-gz)) + tf.math.pow(10.0,-3)*tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels = tf.ones_like(f_out),logits = f_out))
tf. summary.scalar("gloss",gloss)
with tf.name_scope("optimizer"):
dis_optimizer = tf.train.AdamOptimizer(learning_rate = 0.001, name = 'doptimizer')
gen_optimizer = tf.train.AdamOptimizer(learning_rate = 0.001, name = 'goptimizer')
dgrads = dis_optimizer.compute_gradients(dloss, var_list = dvl)
ggrads = gen_optimizer.compute_gradients(gloss, var_list = gvl)
for g in dgrads:
tf.summary.histogram("{} grad".format(g[1].name),g[0])
for g in ggrads: #plotting the gradients
tf.summary.histogram("{} grad".format(g[1].name), g[0])
dis_opt = dis_optimizer.apply_gradients(dgrads)
gen_opt = gen_optimizer.apply_gradients(ggrads)
merged = tf.summary.merge_all()
saver = tf.train.Saver(tf.global_variables(),max_to_keep = 5, keep_checkpoint_every_n_hours = 1)
nepochs = 30
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
try:
saver = tf.train.import_meta_graph('/tmp/model.ckpt.meta') #restoring model
saver.restore(sess,'/tmp/model.ckpt')
except:
print("no saved model")
writer = tf.summary.FileWriter('logs',graph = sess.graph)
for _ in range(nepochs):
i = 1
prev_sum = 0
while i<(n//batch_size):
start = time.time()
print("batch: ",i)
batch_real, batch_blur = create_batch(i)
_,dc = sess.run([dis_opt,dloss], feed_dict = {blur: np.array(batch_blur), real: np.array(batch_real)})
_,gc,summary,gout = sess.run([gen_opt,gloss,merged, gz], feed_dict = {blur:np.array(batch_blur), real: np.array(batch_real)})
writer.add_summary(summary,i)
saver.save(sess,'/tmp/model.ckpt',global_step = i)
if i%10==0:
j =0
for x,y in zip(batch_blur, batch_real):
fig,axs = plt.subplots(1,3, figsize = (15,15))
axs[0].imshow(cv.cvtColor(x, cv.COLOR_BGR2RGB))
axs[1].imshow(cv.cvtColor(gout[j,:,:,:], cv.COLOR_BGR2RGB))
axs[2].imshow(cv.cvtColor(y, cv.COLOR_BGR2RGB))
j+=1
plt.show()
end = time.time()
prev_sum+=(end-start)
print('\n')
print("Eta: ",str(datetime.timedelta(seconds =(prev_sum/i)*((n//batch_size)-i))))
i+=1
print("discriminator cost: ",dc)
print("generator cost: ",gc)
print('\n')
writer.close()